1 // vim: ts=8 sw=2 smarttab
3 * Ceph - scalable distributed file system
5 * Copyright (C) 2014 Red Hat
7 * This is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License version 2.1, as published by the Free Software
10 * Foundation. See file COPYING.
16 #include <sys/types.h>
20 #include "include/cpp-btree/btree_set.h"
22 #include "BlueStore.h"
24 #include "include/compat.h"
25 #include "include/intarith.h"
26 #include "include/stringify.h"
27 #include "common/errno.h"
28 #include "common/safe_io.h"
29 #include "Allocator.h"
30 #include "FreelistManager.h"
32 #include "BlueRocksEnv.h"
33 #include "auth/Crypto.h"
34 #include "common/EventTrace.h"
36 #define dout_context cct
37 #define dout_subsys ceph_subsys_bluestore
39 using bid_t
= decltype(BlueStore::Blob::id
);
41 // bluestore_cache_onode
42 MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Onode
, bluestore_onode
,
43 bluestore_cache_onode
);
45 // bluestore_cache_other
46 MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Buffer
, bluestore_buffer
,
47 bluestore_cache_other
);
48 MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Extent
, bluestore_extent
,
49 bluestore_cache_other
);
50 MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::Blob
, bluestore_blob
,
51 bluestore_cache_other
);
52 MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::SharedBlob
, bluestore_shared_blob
,
53 bluestore_cache_other
);
56 MEMPOOL_DEFINE_OBJECT_FACTORY(BlueStore::TransContext
, bluestore_transcontext
,
61 const string PREFIX_SUPER
= "S"; // field -> value
62 const string PREFIX_STAT
= "T"; // field -> value(int64 array)
63 const string PREFIX_COLL
= "C"; // collection name -> cnode_t
64 const string PREFIX_OBJ
= "O"; // object name -> onode_t
65 const string PREFIX_OMAP
= "M"; // u64 + keyname -> value
66 const string PREFIX_DEFERRED
= "L"; // id -> deferred_transaction_t
67 const string PREFIX_ALLOC
= "B"; // u64 offset -> u64 length (freelist)
68 const string PREFIX_SHARED_BLOB
= "X"; // u64 offset -> shared_blob_t
70 // write a label in the first block. always use this size. note that
71 // bluefs makes a matching assumption about the location of its
72 // superblock (always the second block of the device).
73 #define BDEV_LABEL_BLOCK_SIZE 4096
75 // reserve: label (4k) + bluefs super (4k), which means we start at 8k.
76 #define SUPER_RESERVED 8192
78 #define OBJECT_MAX_SIZE 0xffffffff // 32 bits
82 * extent map blob encoding
84 * we use the low bits of the blobid field to indicate some common scenarios
85 * and spanning vs local ids. See ExtentMap::{encode,decode}_some().
87 #define BLOBID_FLAG_CONTIGUOUS 0x1 // this extent starts at end of previous
88 #define BLOBID_FLAG_ZEROOFFSET 0x2 // blob_offset is 0
89 #define BLOBID_FLAG_SAMELENGTH 0x4 // length matches previous extent
90 #define BLOBID_FLAG_SPANNING 0x8 // has spanning blob id
91 #define BLOBID_SHIFT_BITS 4
94 * object name key structure
96 * encoded u8: shard + 2^7 (so that it sorts properly)
97 * encoded u64: poolid + 2^63 (so that it sorts properly)
98 * encoded u32: hash (bit reversed)
100 * escaped string: namespace
102 * escaped string: key or object name
103 * 1 char: '<', '=', or '>'. if =, then object key == object name, and
104 * we are done. otherwise, we are followed by the object name.
105 * escaped string: object name (unless '=' above)
108 * encoded u64: generation
111 #define ONODE_KEY_SUFFIX 'o'
120 #define EXTENT_SHARD_KEY_SUFFIX 'x'
123 * string encoding in the key
125 * The key string needs to lexicographically sort the same way that
126 * ghobject_t does. We do this by escaping anything <= to '#' with #
127 * plus a 2 digit hex string, and anything >= '~' with ~ plus the two
130 * We use ! as a terminator for strings; this works because it is < #
131 * and will get escaped if it is present in the string.
135 static void append_escaped(const string
&in
, S
*out
)
137 char hexbyte
[in
.length() * 3 + 1];
138 char* ptr
= &hexbyte
[0];
139 for (string::const_iterator i
= in
.begin(); i
!= in
.end(); ++i
) {
142 *ptr
++ = "0123456789abcdef"[(*i
>> 4) & 0x0f];
143 *ptr
++ = "0123456789abcdef"[*i
& 0x0f];
144 } else if (*i
>= '~') {
146 *ptr
++ = "0123456789abcdef"[(*i
>> 4) & 0x0f];
147 *ptr
++ = "0123456789abcdef"[*i
& 0x0f];
153 out
->append(hexbyte
, ptr
- &hexbyte
[0]);
156 inline unsigned h2i(char c
)
158 if ((c
>= '0') && (c
<= '9')) {
160 } else if ((c
>= 'a') && (c
<= 'f')) {
162 } else if ((c
>= 'A') && (c
<= 'F')) {
165 return 256; // make it always larger than 255
169 static int decode_escaped(const char *p
, string
*out
)
172 char* ptr
= &buff
[0];
173 char* max
= &buff
[252];
174 const char *orig_p
= p
;
175 while (*p
&& *p
!= '!') {
176 if (*p
== '#' || *p
== '~') {
179 hex
= h2i(*p
++) << 4;
192 out
->append(buff
, ptr
-buff
);
197 out
->append(buff
, ptr
-buff
);
202 // some things we encode in binary (as le32 or le64); print the
203 // resulting key strings nicely
205 static string
pretty_binary_string(const S
& in
)
209 out
.reserve(in
.length() * 3);
210 enum { NONE
, HEX
, STRING
} mode
= NONE
;
211 unsigned from
= 0, i
;
212 for (i
=0; i
< in
.length(); ++i
) {
213 if ((in
[i
] < 32 || (unsigned char)in
[i
] > 126) ||
214 (mode
== HEX
&& in
.length() - i
>= 4 &&
215 ((in
[i
] < 32 || (unsigned char)in
[i
] > 126) ||
216 (in
[i
+1] < 32 || (unsigned char)in
[i
+1] > 126) ||
217 (in
[i
+2] < 32 || (unsigned char)in
[i
+2] > 126) ||
218 (in
[i
+3] < 32 || (unsigned char)in
[i
+3] > 126)))) {
219 if (mode
== STRING
) {
220 out
.append(in
.c_str() + from
, i
- from
);
227 if (in
.length() - i
>= 4) {
228 // print a whole u32 at once
229 snprintf(buf
, sizeof(buf
), "%08x",
230 (uint32_t)(((unsigned char)in
[i
] << 24) |
231 ((unsigned char)in
[i
+1] << 16) |
232 ((unsigned char)in
[i
+2] << 8) |
233 ((unsigned char)in
[i
+3] << 0)));
236 snprintf(buf
, sizeof(buf
), "%02x", (int)(unsigned char)in
[i
]);
240 if (mode
!= STRING
) {
247 if (mode
== STRING
) {
248 out
.append(in
.c_str() + from
, i
- from
);
255 static void _key_encode_shard(shard_id_t shard
, T
*key
)
257 key
->push_back((char)((uint8_t)shard
.id
+ (uint8_t)0x80));
260 static const char *_key_decode_shard(const char *key
, shard_id_t
*pshard
)
262 pshard
->id
= (uint8_t)*key
- (uint8_t)0x80;
266 static void get_coll_key_range(const coll_t
& cid
, int bits
,
267 string
*temp_start
, string
*temp_end
,
268 string
*start
, string
*end
)
276 if (cid
.is_pg(&pgid
)) {
277 _key_encode_shard(pgid
.shard
, start
);
278 *temp_start
= *start
;
280 _key_encode_u64(pgid
.pool() + 0x8000000000000000ull
, start
);
281 _key_encode_u64((-2ll - pgid
.pool()) + 0x8000000000000000ull
, temp_start
);
284 *temp_end
= *temp_start
;
286 uint32_t reverse_hash
= hobject_t::_reverse_bits(pgid
.ps());
287 _key_encode_u32(reverse_hash
, start
);
288 _key_encode_u32(reverse_hash
, temp_start
);
290 uint64_t end_hash
= reverse_hash
+ (1ull << (32 - bits
));
291 if (end_hash
> 0xffffffffull
)
292 end_hash
= 0xffffffffull
;
294 _key_encode_u32(end_hash
, end
);
295 _key_encode_u32(end_hash
, temp_end
);
297 _key_encode_shard(shard_id_t::NO_SHARD
, start
);
298 _key_encode_u64(-1ull + 0x8000000000000000ull
, start
);
300 _key_encode_u32(0, start
);
301 _key_encode_u32(0xffffffff, end
);
303 // no separate temp section
309 static void get_shared_blob_key(uint64_t sbid
, string
*key
)
312 _key_encode_u64(sbid
, key
);
315 static int get_key_shared_blob(const string
& key
, uint64_t *sbid
)
317 const char *p
= key
.c_str();
318 if (key
.length() < sizeof(uint64_t))
320 _key_decode_u64(p
, sbid
);
325 static int get_key_object(const S
& key
, ghobject_t
*oid
)
328 const char *p
= key
.c_str();
330 if (key
.length() < 1 + 8 + 4)
332 p
= _key_decode_shard(p
, &oid
->shard_id
);
335 p
= _key_decode_u64(p
, &pool
);
336 oid
->hobj
.pool
= pool
- 0x8000000000000000ull
;
339 p
= _key_decode_u32(p
, &hash
);
341 oid
->hobj
.set_bitwise_key_u32(hash
);
343 r
= decode_escaped(p
, &oid
->hobj
.nspace
);
349 r
= decode_escaped(p
, &k
);
356 oid
->hobj
.oid
.name
= k
;
357 } else if (*p
== '<' || *p
== '>') {
360 r
= decode_escaped(p
, &oid
->hobj
.oid
.name
);
364 oid
->hobj
.set_key(k
);
370 p
= _key_decode_u64(p
, &oid
->hobj
.snap
.val
);
371 p
= _key_decode_u64(p
, &oid
->generation
);
373 if (*p
!= ONODE_KEY_SUFFIX
) {
378 // if we get something other than a null terminator here,
379 // something goes wrong.
387 static void get_object_key(CephContext
*cct
, const ghobject_t
& oid
, S
*key
)
391 size_t max_len
= 1 + 8 + 4 +
392 (oid
.hobj
.nspace
.length() * 3 + 1) +
393 (oid
.hobj
.get_key().length() * 3 + 1) +
394 1 + // for '<', '=', or '>'
395 (oid
.hobj
.oid
.name
.length() * 3 + 1) +
397 key
->reserve(max_len
);
399 _key_encode_shard(oid
.shard_id
, key
);
400 _key_encode_u64(oid
.hobj
.pool
+ 0x8000000000000000ull
, key
);
401 _key_encode_u32(oid
.hobj
.get_bitwise_key_u32(), key
);
403 append_escaped(oid
.hobj
.nspace
, key
);
405 if (oid
.hobj
.get_key().length()) {
406 // is a key... could be < = or >.
407 append_escaped(oid
.hobj
.get_key(), key
);
408 // (ASCII chars < = and > sort in that order, yay)
409 int r
= oid
.hobj
.get_key().compare(oid
.hobj
.oid
.name
);
411 key
->append(r
> 0 ? ">" : "<");
412 append_escaped(oid
.hobj
.oid
.name
, key
);
419 append_escaped(oid
.hobj
.oid
.name
, key
);
423 _key_encode_u64(oid
.hobj
.snap
, key
);
424 _key_encode_u64(oid
.generation
, key
);
426 key
->push_back(ONODE_KEY_SUFFIX
);
431 int r
= get_key_object(*key
, &t
);
433 derr
<< " r " << r
<< dendl
;
434 derr
<< "key " << pretty_binary_string(*key
) << dendl
;
435 derr
<< "oid " << oid
<< dendl
;
436 derr
<< " t " << t
<< dendl
;
437 assert(r
== 0 && t
== oid
);
443 // extent shard keys are the onode key, plus a u32, plus 'x'. the trailing
444 // char lets us quickly test whether it is a shard key without decoding any
445 // of the prefix bytes.
447 static void get_extent_shard_key(const S
& onode_key
, uint32_t offset
,
451 key
->reserve(onode_key
.length() + 4 + 1);
452 key
->append(onode_key
.c_str(), onode_key
.size());
453 _key_encode_u32(offset
, key
);
454 key
->push_back(EXTENT_SHARD_KEY_SUFFIX
);
457 static void rewrite_extent_shard_key(uint32_t offset
, string
*key
)
459 assert(key
->size() > sizeof(uint32_t) + 1);
460 assert(*key
->rbegin() == EXTENT_SHARD_KEY_SUFFIX
);
461 _key_encode_u32(offset
, key
->size() - sizeof(uint32_t) - 1, key
);
465 static void generate_extent_shard_key_and_apply(
469 std::function
<void(const string
& final_key
)> apply
)
471 if (key
->empty()) { // make full key
472 assert(!onode_key
.empty());
473 get_extent_shard_key(onode_key
, offset
, key
);
475 rewrite_extent_shard_key(offset
, key
);
480 int get_key_extent_shard(const string
& key
, string
*onode_key
, uint32_t *offset
)
482 assert(key
.size() > sizeof(uint32_t) + 1);
483 assert(*key
.rbegin() == EXTENT_SHARD_KEY_SUFFIX
);
484 int okey_len
= key
.size() - sizeof(uint32_t) - 1;
485 *onode_key
= key
.substr(0, okey_len
);
486 const char *p
= key
.data() + okey_len
;
487 _key_decode_u32(p
, offset
);
491 static bool is_extent_shard_key(const string
& key
)
493 return *key
.rbegin() == EXTENT_SHARD_KEY_SUFFIX
;
497 static void get_omap_header(uint64_t id
, string
*out
)
499 _key_encode_u64(id
, out
);
503 // hmm, I don't think there's any need to escape the user key since we
504 // have a clean prefix.
505 static void get_omap_key(uint64_t id
, const string
& key
, string
*out
)
507 _key_encode_u64(id
, out
);
512 static void rewrite_omap_key(uint64_t id
, string old
, string
*out
)
514 _key_encode_u64(id
, out
);
515 out
->append(old
.c_str() + out
->length(), old
.size() - out
->length());
518 static void decode_omap_key(const string
& key
, string
*user_key
)
520 *user_key
= key
.substr(sizeof(uint64_t) + 1);
523 static void get_omap_tail(uint64_t id
, string
*out
)
525 _key_encode_u64(id
, out
);
529 static void get_deferred_key(uint64_t seq
, string
*out
)
531 _key_encode_u64(seq
, out
);
537 struct Int64ArrayMergeOperator
: public KeyValueDB::MergeOperator
{
538 void merge_nonexistent(
539 const char *rdata
, size_t rlen
, std::string
*new_value
) override
{
540 *new_value
= std::string(rdata
, rlen
);
543 const char *ldata
, size_t llen
,
544 const char *rdata
, size_t rlen
,
545 std::string
*new_value
) override
{
546 assert(llen
== rlen
);
547 assert((rlen
% 8) == 0);
548 new_value
->resize(rlen
);
549 const __le64
* lv
= (const __le64
*)ldata
;
550 const __le64
* rv
= (const __le64
*)rdata
;
551 __le64
* nv
= &(__le64
&)new_value
->at(0);
552 for (size_t i
= 0; i
< rlen
>> 3; ++i
) {
553 nv
[i
] = lv
[i
] + rv
[i
];
556 // We use each operator name and each prefix to construct the
557 // overall RocksDB operator name for consistency check at open time.
558 string
name() const override
{
559 return "int64_array";
566 ostream
& operator<<(ostream
& out
, const BlueStore::Buffer
& b
)
568 out
<< "buffer(" << &b
<< " space " << b
.space
<< " 0x" << std::hex
569 << b
.offset
<< "~" << b
.length
<< std::dec
570 << " " << BlueStore::Buffer::get_state_name(b
.state
);
572 out
<< " " << BlueStore::Buffer::get_flag_name(b
.flags
);
578 void BlueStore::GarbageCollector::process_protrusive_extents(
579 const BlueStore::ExtentMap
& extent_map
,
580 uint64_t start_offset
,
582 uint64_t start_touch_offset
,
583 uint64_t end_touch_offset
,
584 uint64_t min_alloc_size
)
586 assert(start_offset
<= start_touch_offset
&& end_offset
>= end_touch_offset
);
588 uint64_t lookup_start_offset
= P2ALIGN(start_offset
, min_alloc_size
);
589 uint64_t lookup_end_offset
= ROUND_UP_TO(end_offset
, min_alloc_size
);
591 dout(30) << __func__
<< " (hex): [" << std::hex
592 << lookup_start_offset
<< ", " << lookup_end_offset
593 << ")" << std::dec
<< dendl
;
595 for (auto it
= extent_map
.seek_lextent(lookup_start_offset
);
596 it
!= extent_map
.extent_map
.end() &&
597 it
->logical_offset
< lookup_end_offset
;
599 uint64_t alloc_unit_start
= it
->logical_offset
/ min_alloc_size
;
600 uint64_t alloc_unit_end
= (it
->logical_end() - 1) / min_alloc_size
;
602 dout(30) << __func__
<< " " << *it
603 << "alloc_units: " << alloc_unit_start
<< ".." << alloc_unit_end
606 Blob
* b
= it
->blob
.get();
608 if (it
->logical_offset
>=start_touch_offset
&&
609 it
->logical_end() <= end_touch_offset
) {
610 // Process extents within the range affected by
611 // the current write request.
612 // Need to take into account if existing extents
613 // can be merged with them (uncompressed case)
614 if (!b
->get_blob().is_compressed()) {
615 if (blob_info_counted
&& used_alloc_unit
== alloc_unit_start
) {
616 --blob_info_counted
->expected_allocations
; // don't need to allocate
617 // new AU for compressed
618 // data since another
619 // collocated uncompressed
620 // blob already exists
621 dout(30) << __func__
<< " --expected:"
622 << alloc_unit_start
<< dendl
;
624 used_alloc_unit
= alloc_unit_end
;
625 blob_info_counted
= nullptr;
627 } else if (b
->get_blob().is_compressed()) {
629 // additionally we take compressed blobs that were not impacted
630 // by the write into account too
632 affected_blobs
.emplace(
633 b
, BlobInfo(b
->get_referenced_bytes())).first
->second
;
636 (used_alloc_unit
&& used_alloc_unit
== alloc_unit_start
) ? 0 : 1;
637 bi
.expected_allocations
+= alloc_unit_end
- alloc_unit_start
+ adjust
;
638 dout(30) << __func__
<< " expected_allocations="
639 << bi
.expected_allocations
<< " end_au:"
640 << alloc_unit_end
<< dendl
;
642 blob_info_counted
= &bi
;
643 used_alloc_unit
= alloc_unit_end
;
645 assert(it
->length
<= bi
.referenced_bytes
);
646 bi
.referenced_bytes
-= it
->length
;
647 dout(30) << __func__
<< " affected_blob:" << *b
648 << " unref 0x" << std::hex
<< it
->length
649 << " referenced = 0x" << bi
.referenced_bytes
650 << std::dec
<< dendl
;
651 // NOTE: we can't move specific blob to resulting GC list here
652 // when reference counter == 0 since subsequent extents might
653 // decrement its expected_allocation.
654 // Hence need to enumerate all the extents first.
655 if (!bi
.collect_candidate
) {
656 bi
.first_lextent
= it
;
657 bi
.collect_candidate
= true;
659 bi
.last_lextent
= it
;
661 if (blob_info_counted
&& used_alloc_unit
== alloc_unit_start
) {
662 // don't need to allocate new AU for compressed data since another
663 // collocated uncompressed blob already exists
664 --blob_info_counted
->expected_allocations
;
665 dout(30) << __func__
<< " --expected_allocations:"
666 << alloc_unit_start
<< dendl
;
668 used_alloc_unit
= alloc_unit_end
;
669 blob_info_counted
= nullptr;
673 for (auto b_it
= affected_blobs
.begin();
674 b_it
!= affected_blobs
.end();
676 Blob
* b
= b_it
->first
;
677 BlobInfo
& bi
= b_it
->second
;
678 if (bi
.referenced_bytes
== 0) {
679 uint64_t len_on_disk
= b_it
->first
->get_blob().get_ondisk_length();
680 int64_t blob_expected_for_release
=
681 ROUND_UP_TO(len_on_disk
, min_alloc_size
) / min_alloc_size
;
683 dout(30) << __func__
<< " " << *(b_it
->first
)
684 << " expected4release=" << blob_expected_for_release
685 << " expected_allocations=" << bi
.expected_allocations
687 int64_t benefit
= blob_expected_for_release
- bi
.expected_allocations
;
688 if (benefit
>= g_conf
->bluestore_gc_enable_blob_threshold
) {
689 if (bi
.collect_candidate
) {
690 auto it
= bi
.first_lextent
;
693 if (it
->blob
.get() == b
) {
694 extents_to_collect
.emplace_back(it
->logical_offset
, it
->length
);
696 bExit
= it
== bi
.last_lextent
;
700 expected_for_release
+= blob_expected_for_release
;
701 expected_allocations
+= bi
.expected_allocations
;
707 int64_t BlueStore::GarbageCollector::estimate(
708 uint64_t start_offset
,
710 const BlueStore::ExtentMap
& extent_map
,
711 const BlueStore::old_extent_map_t
& old_extents
,
712 uint64_t min_alloc_size
)
715 affected_blobs
.clear();
716 extents_to_collect
.clear();
717 used_alloc_unit
= boost::optional
<uint64_t >();
718 blob_info_counted
= nullptr;
720 gc_start_offset
= start_offset
;
721 gc_end_offset
= start_offset
+ length
;
723 uint64_t end_offset
= start_offset
+ length
;
725 for (auto it
= old_extents
.begin(); it
!= old_extents
.end(); ++it
) {
726 Blob
* b
= it
->e
.blob
.get();
727 if (b
->get_blob().is_compressed()) {
729 // update gc_start_offset/gc_end_offset if needed
730 gc_start_offset
= min(gc_start_offset
, (uint64_t)it
->e
.blob_start());
731 gc_end_offset
= max(gc_end_offset
, (uint64_t)it
->e
.blob_end());
733 auto o
= it
->e
.logical_offset
;
734 auto l
= it
->e
.length
;
736 uint64_t ref_bytes
= b
->get_referenced_bytes();
737 // micro optimization to bypass blobs that have no more references
738 if (ref_bytes
!= 0) {
739 dout(30) << __func__
<< " affected_blob:" << *b
740 << " unref 0x" << std::hex
<< o
<< "~" << l
741 << std::dec
<< dendl
;
742 affected_blobs
.emplace(b
, BlobInfo(ref_bytes
));
746 dout(30) << __func__
<< " gc range(hex): [" << std::hex
747 << gc_start_offset
<< ", " << gc_end_offset
748 << ")" << std::dec
<< dendl
;
750 // enumerate preceeding extents to check if they reference affected blobs
751 if (gc_start_offset
< start_offset
|| gc_end_offset
> end_offset
) {
752 process_protrusive_extents(extent_map
,
759 return expected_for_release
- expected_allocations
;
764 BlueStore::Cache
*BlueStore::Cache::create(CephContext
* cct
, string type
,
765 PerfCounters
*logger
)
770 c
= new LRUCache(cct
);
771 else if (type
== "2q")
772 c
= new TwoQCache(cct
);
774 assert(0 == "unrecognized cache type");
780 void BlueStore::Cache::trim_all()
782 std::lock_guard
<std::recursive_mutex
> l(lock
);
786 void BlueStore::Cache::trim(
787 uint64_t target_bytes
,
788 float target_meta_ratio
,
789 float target_data_ratio
,
790 float bytes_per_onode
)
792 std::lock_guard
<std::recursive_mutex
> l(lock
);
793 uint64_t current_meta
= _get_num_onodes() * bytes_per_onode
;
794 uint64_t current_buffer
= _get_buffer_bytes();
795 uint64_t current
= current_meta
+ current_buffer
;
797 uint64_t target_meta
= target_bytes
* target_meta_ratio
;
798 uint64_t target_buffer
= target_bytes
* target_data_ratio
;
800 // correct for overflow or float imprecision
801 target_meta
= min(target_bytes
, target_meta
);
802 target_buffer
= min(target_bytes
- target_meta
, target_buffer
);
804 if (current
<= target_bytes
) {
806 << " shard target " << byte_u_t(target_bytes
)
807 << " meta/data ratios " << target_meta_ratio
808 << " + " << target_data_ratio
<< " ("
809 << byte_u_t(target_meta
) << " + "
810 << byte_u_t(target_buffer
) << "), "
811 << " current " << byte_u_t(current
) << " ("
812 << byte_u_t(current_meta
) << " + "
813 << byte_u_t(current_buffer
) << ")"
818 uint64_t need_to_free
= current
- target_bytes
;
819 uint64_t free_buffer
= 0;
820 uint64_t free_meta
= 0;
821 if (current_buffer
> target_buffer
) {
822 free_buffer
= current_buffer
- target_buffer
;
823 if (free_buffer
> need_to_free
) {
824 free_buffer
= need_to_free
;
827 free_meta
= need_to_free
- free_buffer
;
829 // start bounds at what we have now
830 uint64_t max_buffer
= current_buffer
- free_buffer
;
831 uint64_t max_meta
= current_meta
- free_meta
;
832 uint64_t max_onodes
= max_meta
/ bytes_per_onode
;
835 << " shard target " << byte_u_t(target_bytes
)
836 << " ratio " << target_meta_ratio
<< " ("
837 << byte_u_t(target_meta
) << " + "
838 << byte_u_t(target_buffer
) << "), "
839 << " current " << byte_u_t(current
) << " ("
840 << byte_u_t(current_meta
) << " + "
841 << byte_u_t(current_buffer
) << "),"
842 << " need_to_free " << byte_u_t(need_to_free
) << " ("
843 << byte_u_t(free_meta
) << " + "
844 << byte_u_t(free_buffer
) << ")"
845 << " -> max " << max_onodes
<< " onodes + "
846 << max_buffer
<< " buffer"
848 _trim(max_onodes
, max_buffer
);
854 #define dout_prefix *_dout << "bluestore.LRUCache(" << this << ") "
856 void BlueStore::LRUCache::_touch_onode(OnodeRef
& o
)
858 auto p
= onode_lru
.iterator_to(*o
);
860 onode_lru
.push_front(*o
);
863 void BlueStore::LRUCache::_trim(uint64_t onode_max
, uint64_t buffer_max
)
865 dout(20) << __func__
<< " onodes " << onode_lru
.size() << " / " << onode_max
866 << " buffers " << buffer_size
<< " / " << buffer_max
869 _audit("trim start");
872 while (buffer_size
> buffer_max
) {
873 auto i
= buffer_lru
.rbegin();
874 if (i
== buffer_lru
.rend()) {
875 // stop if buffer_lru is now empty
880 assert(b
->is_clean());
881 dout(20) << __func__
<< " rm " << *b
<< dendl
;
882 b
->space
->_rm_buffer(this, b
);
886 int num
= onode_lru
.size() - onode_max
;
888 return; // don't even try
890 auto p
= onode_lru
.end();
891 assert(p
!= onode_lru
.begin());
894 int max_skipped
= g_conf
->bluestore_cache_trim_max_skip_pinned
;
897 int refs
= o
->nref
.load();
899 dout(20) << __func__
<< " " << o
->oid
<< " has " << refs
900 << " refs, skipping" << dendl
;
901 if (++skipped
>= max_skipped
) {
902 dout(20) << __func__
<< " maximum skip pinned reached; stopping with "
903 << num
<< " left to trim" << dendl
;
907 if (p
== onode_lru
.begin()) {
915 dout(30) << __func__
<< " rm " << o
->oid
<< dendl
;
916 if (p
!= onode_lru
.begin()) {
917 onode_lru
.erase(p
--);
922 o
->get(); // paranoia
923 o
->c
->onode_map
.remove(o
->oid
);
930 void BlueStore::LRUCache::_audit(const char *when
)
932 dout(10) << __func__
<< " " << when
<< " start" << dendl
;
934 for (auto i
= buffer_lru
.begin(); i
!= buffer_lru
.end(); ++i
) {
937 if (s
!= buffer_size
) {
938 derr
<< __func__
<< " buffer_size " << buffer_size
<< " actual " << s
940 for (auto i
= buffer_lru
.begin(); i
!= buffer_lru
.end(); ++i
) {
941 derr
<< __func__
<< " " << *i
<< dendl
;
943 assert(s
== buffer_size
);
945 dout(20) << __func__
<< " " << when
<< " buffer_size " << buffer_size
952 #define dout_prefix *_dout << "bluestore.2QCache(" << this << ") "
955 void BlueStore::TwoQCache::_touch_onode(OnodeRef
& o
)
957 auto p
= onode_lru
.iterator_to(*o
);
959 onode_lru
.push_front(*o
);
962 void BlueStore::TwoQCache::_add_buffer(Buffer
*b
, int level
, Buffer
*near
)
964 dout(20) << __func__
<< " level " << level
<< " near " << near
966 << " which has cache_private " << b
->cache_private
<< dendl
;
968 b
->cache_private
= near
->cache_private
;
969 switch (b
->cache_private
) {
971 buffer_warm_in
.insert(buffer_warm_in
.iterator_to(*near
), *b
);
973 case BUFFER_WARM_OUT
:
974 assert(b
->is_empty());
975 buffer_warm_out
.insert(buffer_warm_out
.iterator_to(*near
), *b
);
978 buffer_hot
.insert(buffer_hot
.iterator_to(*near
), *b
);
981 assert(0 == "bad cache_private");
983 } else if (b
->cache_private
== BUFFER_NEW
) {
984 b
->cache_private
= BUFFER_WARM_IN
;
986 buffer_warm_in
.push_front(*b
);
988 // take caller hint to start at the back of the warm queue
989 buffer_warm_in
.push_back(*b
);
992 // we got a hint from discard
993 switch (b
->cache_private
) {
995 // stay in warm_in. move to front, even though 2Q doesn't actually
997 dout(20) << __func__
<< " move to front of warm " << *b
<< dendl
;
998 buffer_warm_in
.push_front(*b
);
1000 case BUFFER_WARM_OUT
:
1001 b
->cache_private
= BUFFER_HOT
;
1002 // move to hot. fall-thru
1004 dout(20) << __func__
<< " move to front of hot " << *b
<< dendl
;
1005 buffer_hot
.push_front(*b
);
1008 assert(0 == "bad cache_private");
1011 if (!b
->is_empty()) {
1012 buffer_bytes
+= b
->length
;
1013 buffer_list_bytes
[b
->cache_private
] += b
->length
;
1017 void BlueStore::TwoQCache::_rm_buffer(Buffer
*b
)
1019 dout(20) << __func__
<< " " << *b
<< dendl
;
1020 if (!b
->is_empty()) {
1021 assert(buffer_bytes
>= b
->length
);
1022 buffer_bytes
-= b
->length
;
1023 assert(buffer_list_bytes
[b
->cache_private
] >= b
->length
);
1024 buffer_list_bytes
[b
->cache_private
] -= b
->length
;
1026 switch (b
->cache_private
) {
1027 case BUFFER_WARM_IN
:
1028 buffer_warm_in
.erase(buffer_warm_in
.iterator_to(*b
));
1030 case BUFFER_WARM_OUT
:
1031 buffer_warm_out
.erase(buffer_warm_out
.iterator_to(*b
));
1034 buffer_hot
.erase(buffer_hot
.iterator_to(*b
));
1037 assert(0 == "bad cache_private");
1041 void BlueStore::TwoQCache::_move_buffer(Cache
*srcc
, Buffer
*b
)
1043 TwoQCache
*src
= static_cast<TwoQCache
*>(srcc
);
1046 // preserve which list we're on (even if we can't preserve the order!)
1047 switch (b
->cache_private
) {
1048 case BUFFER_WARM_IN
:
1049 assert(!b
->is_empty());
1050 buffer_warm_in
.push_back(*b
);
1052 case BUFFER_WARM_OUT
:
1053 assert(b
->is_empty());
1054 buffer_warm_out
.push_back(*b
);
1057 assert(!b
->is_empty());
1058 buffer_hot
.push_back(*b
);
1061 assert(0 == "bad cache_private");
1063 if (!b
->is_empty()) {
1064 buffer_bytes
+= b
->length
;
1065 buffer_list_bytes
[b
->cache_private
] += b
->length
;
1069 void BlueStore::TwoQCache::_adjust_buffer_size(Buffer
*b
, int64_t delta
)
1071 dout(20) << __func__
<< " delta " << delta
<< " on " << *b
<< dendl
;
1072 if (!b
->is_empty()) {
1073 assert((int64_t)buffer_bytes
+ delta
>= 0);
1074 buffer_bytes
+= delta
;
1075 assert((int64_t)buffer_list_bytes
[b
->cache_private
] + delta
>= 0);
1076 buffer_list_bytes
[b
->cache_private
] += delta
;
1080 void BlueStore::TwoQCache::_trim(uint64_t onode_max
, uint64_t buffer_max
)
1082 dout(20) << __func__
<< " onodes " << onode_lru
.size() << " / " << onode_max
1083 << " buffers " << buffer_bytes
<< " / " << buffer_max
1086 _audit("trim start");
1089 if (buffer_bytes
> buffer_max
) {
1090 uint64_t kin
= buffer_max
* cct
->_conf
->bluestore_2q_cache_kin_ratio
;
1091 uint64_t khot
= buffer_max
- kin
;
1093 // pre-calculate kout based on average buffer size too,
1094 // which is typical(the warm_in and hot lists may change later)
1096 uint64_t buffer_num
= buffer_hot
.size() + buffer_warm_in
.size();
1098 uint64_t buffer_avg_size
= buffer_bytes
/ buffer_num
;
1099 assert(buffer_avg_size
);
1100 uint64_t calculated_buffer_num
= buffer_max
/ buffer_avg_size
;
1101 kout
= calculated_buffer_num
* cct
->_conf
->bluestore_2q_cache_kout_ratio
;
1104 if (buffer_list_bytes
[BUFFER_HOT
] < khot
) {
1105 // hot is small, give slack to warm_in
1106 kin
+= khot
- buffer_list_bytes
[BUFFER_HOT
];
1107 } else if (buffer_list_bytes
[BUFFER_WARM_IN
] < kin
) {
1108 // warm_in is small, give slack to hot
1109 khot
+= kin
- buffer_list_bytes
[BUFFER_WARM_IN
];
1112 // adjust warm_in list
1113 int64_t to_evict_bytes
= buffer_list_bytes
[BUFFER_WARM_IN
] - kin
;
1114 uint64_t evicted
= 0;
1116 while (to_evict_bytes
> 0) {
1117 auto p
= buffer_warm_in
.rbegin();
1118 if (p
== buffer_warm_in
.rend()) {
1119 // stop if warm_in list is now empty
1124 assert(b
->is_clean());
1125 dout(20) << __func__
<< " buffer_warm_in -> out " << *b
<< dendl
;
1126 assert(buffer_bytes
>= b
->length
);
1127 buffer_bytes
-= b
->length
;
1128 assert(buffer_list_bytes
[BUFFER_WARM_IN
] >= b
->length
);
1129 buffer_list_bytes
[BUFFER_WARM_IN
] -= b
->length
;
1130 to_evict_bytes
-= b
->length
;
1131 evicted
+= b
->length
;
1132 b
->state
= Buffer::STATE_EMPTY
;
1134 buffer_warm_in
.erase(buffer_warm_in
.iterator_to(*b
));
1135 buffer_warm_out
.push_front(*b
);
1136 b
->cache_private
= BUFFER_WARM_OUT
;
1140 dout(20) << __func__
<< " evicted " << byte_u_t(evicted
)
1141 << " from warm_in list, done evicting warm_in buffers"
1146 to_evict_bytes
= buffer_list_bytes
[BUFFER_HOT
] - khot
;
1149 while (to_evict_bytes
> 0) {
1150 auto p
= buffer_hot
.rbegin();
1151 if (p
== buffer_hot
.rend()) {
1152 // stop if hot list is now empty
1157 dout(20) << __func__
<< " buffer_hot rm " << *b
<< dendl
;
1158 assert(b
->is_clean());
1159 // adjust evict size before buffer goes invalid
1160 to_evict_bytes
-= b
->length
;
1161 evicted
+= b
->length
;
1162 b
->space
->_rm_buffer(this, b
);
1166 dout(20) << __func__
<< " evicted " << byte_u_t(evicted
)
1167 << " from hot list, done evicting hot buffers"
1171 // adjust warm out list too, if necessary
1172 int64_t num
= buffer_warm_out
.size() - kout
;
1174 Buffer
*b
= &*buffer_warm_out
.rbegin();
1175 assert(b
->is_empty());
1176 dout(20) << __func__
<< " buffer_warm_out rm " << *b
<< dendl
;
1177 b
->space
->_rm_buffer(this, b
);
1182 int num
= onode_lru
.size() - onode_max
;
1184 return; // don't even try
1186 auto p
= onode_lru
.end();
1187 assert(p
!= onode_lru
.begin());
1190 int max_skipped
= g_conf
->bluestore_cache_trim_max_skip_pinned
;
1193 dout(20) << __func__
<< " considering " << o
<< dendl
;
1194 int refs
= o
->nref
.load();
1196 dout(20) << __func__
<< " " << o
->oid
<< " has " << refs
1197 << " refs; skipping" << dendl
;
1198 if (++skipped
>= max_skipped
) {
1199 dout(20) << __func__
<< " maximum skip pinned reached; stopping with "
1200 << num
<< " left to trim" << dendl
;
1204 if (p
== onode_lru
.begin()) {
1212 dout(30) << __func__
<< " " << o
->oid
<< " num=" << num
<<" lru size="<<onode_lru
.size()<< dendl
;
1213 if (p
!= onode_lru
.begin()) {
1214 onode_lru
.erase(p
--);
1219 o
->get(); // paranoia
1220 o
->c
->onode_map
.remove(o
->oid
);
1227 void BlueStore::TwoQCache::_audit(const char *when
)
1229 dout(10) << __func__
<< " " << when
<< " start" << dendl
;
1231 for (auto i
= buffer_hot
.begin(); i
!= buffer_hot
.end(); ++i
) {
1235 uint64_t hot_bytes
= s
;
1236 if (hot_bytes
!= buffer_list_bytes
[BUFFER_HOT
]) {
1237 derr
<< __func__
<< " hot_list_bytes "
1238 << buffer_list_bytes
[BUFFER_HOT
]
1239 << " != actual " << hot_bytes
1241 assert(hot_bytes
== buffer_list_bytes
[BUFFER_HOT
]);
1244 for (auto i
= buffer_warm_in
.begin(); i
!= buffer_warm_in
.end(); ++i
) {
1248 uint64_t warm_in_bytes
= s
- hot_bytes
;
1249 if (warm_in_bytes
!= buffer_list_bytes
[BUFFER_WARM_IN
]) {
1250 derr
<< __func__
<< " warm_in_list_bytes "
1251 << buffer_list_bytes
[BUFFER_WARM_IN
]
1252 << " != actual " << warm_in_bytes
1254 assert(warm_in_bytes
== buffer_list_bytes
[BUFFER_WARM_IN
]);
1257 if (s
!= buffer_bytes
) {
1258 derr
<< __func__
<< " buffer_bytes " << buffer_bytes
<< " actual " << s
1260 assert(s
== buffer_bytes
);
1263 dout(20) << __func__
<< " " << when
<< " buffer_bytes " << buffer_bytes
1272 #define dout_prefix *_dout << "bluestore.BufferSpace(" << this << " in " << cache << ") "
1274 void BlueStore::BufferSpace::_clear(Cache
* cache
)
1276 // note: we already hold cache->lock
1277 ldout(cache
->cct
, 20) << __func__
<< dendl
;
1278 while (!buffer_map
.empty()) {
1279 _rm_buffer(cache
, buffer_map
.begin());
1283 int BlueStore::BufferSpace::_discard(Cache
* cache
, uint32_t offset
, uint32_t length
)
1285 // note: we already hold cache->lock
1286 ldout(cache
->cct
, 20) << __func__
<< std::hex
<< " 0x" << offset
<< "~" << length
1287 << std::dec
<< dendl
;
1288 int cache_private
= 0;
1289 cache
->_audit("discard start");
1290 auto i
= _data_lower_bound(offset
);
1291 uint32_t end
= offset
+ length
;
1292 while (i
!= buffer_map
.end()) {
1293 Buffer
*b
= i
->second
.get();
1294 if (b
->offset
>= end
) {
1297 if (b
->cache_private
> cache_private
) {
1298 cache_private
= b
->cache_private
;
1300 if (b
->offset
< offset
) {
1301 int64_t front
= offset
- b
->offset
;
1302 if (b
->end() > end
) {
1303 // drop middle (split)
1304 uint32_t tail
= b
->end() - end
;
1305 if (b
->data
.length()) {
1307 bl
.substr_of(b
->data
, b
->length
- tail
, tail
);
1308 Buffer
*nb
= new Buffer(this, b
->state
, b
->seq
, end
, bl
);
1309 nb
->maybe_rebuild();
1310 _add_buffer(cache
, nb
, 0, b
);
1312 _add_buffer(cache
, new Buffer(this, b
->state
, b
->seq
, end
, tail
),
1315 if (!b
->is_writing()) {
1316 cache
->_adjust_buffer_size(b
, front
- (int64_t)b
->length
);
1320 cache
->_audit("discard end 1");
1324 if (!b
->is_writing()) {
1325 cache
->_adjust_buffer_size(b
, front
- (int64_t)b
->length
);
1333 if (b
->end() <= end
) {
1334 // drop entire buffer
1335 _rm_buffer(cache
, i
++);
1339 uint32_t keep
= b
->end() - end
;
1340 if (b
->data
.length()) {
1342 bl
.substr_of(b
->data
, b
->length
- keep
, keep
);
1343 Buffer
*nb
= new Buffer(this, b
->state
, b
->seq
, end
, bl
);
1344 nb
->maybe_rebuild();
1345 _add_buffer(cache
, nb
, 0, b
);
1347 _add_buffer(cache
, new Buffer(this, b
->state
, b
->seq
, end
, keep
), 0, b
);
1349 _rm_buffer(cache
, i
);
1350 cache
->_audit("discard end 2");
1353 return cache_private
;
1356 void BlueStore::BufferSpace::read(
1360 BlueStore::ready_regions_t
& res
,
1361 interval_set
<uint32_t>& res_intervals
)
1364 res_intervals
.clear();
1365 uint32_t want_bytes
= length
;
1366 uint32_t end
= offset
+ length
;
1369 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1370 for (auto i
= _data_lower_bound(offset
);
1371 i
!= buffer_map
.end() && offset
< end
&& i
->first
< end
;
1373 Buffer
*b
= i
->second
.get();
1374 assert(b
->end() > offset
);
1375 if (b
->is_writing() || b
->is_clean()) {
1376 if (b
->offset
< offset
) {
1377 uint32_t skip
= offset
- b
->offset
;
1378 uint32_t l
= MIN(length
, b
->length
- skip
);
1379 res
[offset
].substr_of(b
->data
, skip
, l
);
1380 res_intervals
.insert(offset
, l
);
1383 if (!b
->is_writing()) {
1384 cache
->_touch_buffer(b
);
1388 if (b
->offset
> offset
) {
1389 uint32_t gap
= b
->offset
- offset
;
1390 if (length
<= gap
) {
1396 if (!b
->is_writing()) {
1397 cache
->_touch_buffer(b
);
1399 if (b
->length
> length
) {
1400 res
[offset
].substr_of(b
->data
, 0, length
);
1401 res_intervals
.insert(offset
, length
);
1404 res
[offset
].append(b
->data
);
1405 res_intervals
.insert(offset
, b
->length
);
1406 if (b
->length
== length
)
1408 offset
+= b
->length
;
1409 length
-= b
->length
;
1415 uint64_t hit_bytes
= res_intervals
.size();
1416 assert(hit_bytes
<= want_bytes
);
1417 uint64_t miss_bytes
= want_bytes
- hit_bytes
;
1418 cache
->logger
->inc(l_bluestore_buffer_hit_bytes
, hit_bytes
);
1419 cache
->logger
->inc(l_bluestore_buffer_miss_bytes
, miss_bytes
);
1422 void BlueStore::BufferSpace::finish_write(Cache
* cache
, uint64_t seq
)
1424 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1426 auto i
= writing
.begin();
1427 while (i
!= writing
.end()) {
1437 assert(b
->is_writing());
1439 if (b
->flags
& Buffer::FLAG_NOCACHE
) {
1441 ldout(cache
->cct
, 20) << __func__
<< " discard " << *b
<< dendl
;
1442 buffer_map
.erase(b
->offset
);
1444 b
->state
= Buffer::STATE_CLEAN
;
1447 b
->data
.reassign_to_mempool(mempool::mempool_bluestore_cache_data
);
1448 cache
->_add_buffer(b
, 1, nullptr);
1449 ldout(cache
->cct
, 20) << __func__
<< " added " << *b
<< dendl
;
1453 cache
->_audit("finish_write end");
1456 void BlueStore::BufferSpace::split(Cache
* cache
, size_t pos
, BlueStore::BufferSpace
&r
)
1458 std::lock_guard
<std::recursive_mutex
> lk(cache
->lock
);
1459 if (buffer_map
.empty())
1462 auto p
= --buffer_map
.end();
1464 if (p
->second
->end() <= pos
)
1467 if (p
->second
->offset
< pos
) {
1468 ldout(cache
->cct
, 30) << __func__
<< " cut " << *p
->second
<< dendl
;
1469 size_t left
= pos
- p
->second
->offset
;
1470 size_t right
= p
->second
->length
- left
;
1471 if (p
->second
->data
.length()) {
1473 bl
.substr_of(p
->second
->data
, left
, right
);
1474 r
._add_buffer(cache
, new Buffer(&r
, p
->second
->state
, p
->second
->seq
, 0, bl
),
1475 0, p
->second
.get());
1477 r
._add_buffer(cache
, new Buffer(&r
, p
->second
->state
, p
->second
->seq
, 0, right
),
1478 0, p
->second
.get());
1480 cache
->_adjust_buffer_size(p
->second
.get(), -right
);
1481 p
->second
->truncate(left
);
1485 assert(p
->second
->end() > pos
);
1486 ldout(cache
->cct
, 30) << __func__
<< " move " << *p
->second
<< dendl
;
1487 if (p
->second
->data
.length()) {
1488 r
._add_buffer(cache
, new Buffer(&r
, p
->second
->state
, p
->second
->seq
,
1489 p
->second
->offset
- pos
, p
->second
->data
),
1490 0, p
->second
.get());
1492 r
._add_buffer(cache
, new Buffer(&r
, p
->second
->state
, p
->second
->seq
,
1493 p
->second
->offset
- pos
, p
->second
->length
),
1494 0, p
->second
.get());
1496 if (p
== buffer_map
.begin()) {
1497 _rm_buffer(cache
, p
);
1500 _rm_buffer(cache
, p
--);
1503 assert(writing
.empty());
1509 #define dout_prefix *_dout << "bluestore.OnodeSpace(" << this << " in " << cache << ") "
1511 BlueStore::OnodeRef
BlueStore::OnodeSpace::add(const ghobject_t
& oid
, OnodeRef o
)
1513 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1514 auto p
= onode_map
.find(oid
);
1515 if (p
!= onode_map
.end()) {
1516 ldout(cache
->cct
, 30) << __func__
<< " " << oid
<< " " << o
1517 << " raced, returning existing " << p
->second
1521 ldout(cache
->cct
, 30) << __func__
<< " " << oid
<< " " << o
<< dendl
;
1523 cache
->_add_onode(o
, 1);
1527 BlueStore::OnodeRef
BlueStore::OnodeSpace::lookup(const ghobject_t
& oid
)
1529 ldout(cache
->cct
, 30) << __func__
<< dendl
;
1534 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1535 ceph::unordered_map
<ghobject_t
,OnodeRef
>::iterator p
= onode_map
.find(oid
);
1536 if (p
== onode_map
.end()) {
1537 ldout(cache
->cct
, 30) << __func__
<< " " << oid
<< " miss" << dendl
;
1539 ldout(cache
->cct
, 30) << __func__
<< " " << oid
<< " hit " << p
->second
1541 cache
->_touch_onode(p
->second
);
1548 cache
->logger
->inc(l_bluestore_onode_hits
);
1550 cache
->logger
->inc(l_bluestore_onode_misses
);
1555 void BlueStore::OnodeSpace::clear()
1557 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1558 ldout(cache
->cct
, 10) << __func__
<< dendl
;
1559 for (auto &p
: onode_map
) {
1560 cache
->_rm_onode(p
.second
);
1565 bool BlueStore::OnodeSpace::empty()
1567 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1568 return onode_map
.empty();
1571 void BlueStore::OnodeSpace::rename(
1573 const ghobject_t
& old_oid
,
1574 const ghobject_t
& new_oid
,
1575 const mempool::bluestore_cache_other::string
& new_okey
)
1577 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1578 ldout(cache
->cct
, 30) << __func__
<< " " << old_oid
<< " -> " << new_oid
1580 ceph::unordered_map
<ghobject_t
,OnodeRef
>::iterator po
, pn
;
1581 po
= onode_map
.find(old_oid
);
1582 pn
= onode_map
.find(new_oid
);
1585 assert(po
!= onode_map
.end());
1586 if (pn
!= onode_map
.end()) {
1587 ldout(cache
->cct
, 30) << __func__
<< " removing target " << pn
->second
1589 cache
->_rm_onode(pn
->second
);
1590 onode_map
.erase(pn
);
1592 OnodeRef o
= po
->second
;
1594 // install a non-existent onode at old location
1595 oldo
.reset(new Onode(o
->c
, old_oid
, o
->key
));
1597 cache
->_add_onode(po
->second
, 1);
1599 // add at new position and fix oid, key
1600 onode_map
.insert(make_pair(new_oid
, o
));
1601 cache
->_touch_onode(o
);
1606 bool BlueStore::OnodeSpace::map_any(std::function
<bool(OnodeRef
)> f
)
1608 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
);
1609 ldout(cache
->cct
, 20) << __func__
<< dendl
;
1610 for (auto& i
: onode_map
) {
1618 void BlueStore::OnodeSpace::dump(CephContext
*cct
, int lvl
)
1620 for (auto& i
: onode_map
) {
1621 ldout(cct
, lvl
) << i
.first
<< " : " << i
.second
<< dendl
;
1628 #define dout_prefix *_dout << "bluestore.sharedblob(" << this << ") "
1630 ostream
& operator<<(ostream
& out
, const BlueStore::SharedBlob
& sb
)
1632 out
<< "SharedBlob(" << &sb
;
1635 out
<< " loaded " << *sb
.persistent
;
1637 out
<< " sbid 0x" << std::hex
<< sb
.sbid_unloaded
<< std::dec
;
1642 BlueStore::SharedBlob::SharedBlob(uint64_t i
, Collection
*_coll
)
1643 : coll(_coll
), sbid_unloaded(i
)
1645 assert(sbid_unloaded
> 0);
1647 get_cache()->add_blob();
1651 BlueStore::SharedBlob::~SharedBlob()
1653 if (loaded
&& persistent
) {
1658 void BlueStore::SharedBlob::put()
1661 ldout(coll
->store
->cct
, 20) << __func__
<< " " << this
1662 << " removing self from set " << get_parent()
1665 auto coll_snap
= coll
;
1667 std::lock_guard
<std::recursive_mutex
> l(coll_snap
->cache
->lock
);
1668 if (coll_snap
!= coll
) {
1671 coll_snap
->shared_blob_set
.remove(this);
1673 bc
._clear(coll_snap
->cache
);
1674 coll_snap
->cache
->rm_blob();
1680 void BlueStore::SharedBlob::get_ref(uint64_t offset
, uint32_t length
)
1683 persistent
->ref_map
.get(offset
, length
);
1686 void BlueStore::SharedBlob::put_ref(uint64_t offset
, uint32_t length
,
1688 set
<SharedBlob
*> *maybe_unshared
)
1692 persistent
->ref_map
.put(offset
, length
, r
, maybe_unshared
? &maybe
: nullptr);
1693 if (maybe_unshared
&& maybe
) {
1694 maybe_unshared
->insert(this);
1701 #define dout_prefix *_dout << "bluestore.sharedblobset(" << this << ") "
1703 void BlueStore::SharedBlobSet::dump(CephContext
*cct
, int lvl
)
1705 std::lock_guard
<std::mutex
> l(lock
);
1706 for (auto& i
: sb_map
) {
1707 ldout(cct
, lvl
) << i
.first
<< " : " << *i
.second
<< dendl
;
1714 #define dout_prefix *_dout << "bluestore.blob(" << this << ") "
1716 ostream
& operator<<(ostream
& out
, const BlueStore::Blob
& b
)
1718 out
<< "Blob(" << &b
;
1719 if (b
.is_spanning()) {
1720 out
<< " spanning " << b
.id
;
1722 out
<< " " << b
.get_blob() << " " << b
.get_blob_use_tracker();
1723 if (b
.shared_blob
) {
1724 out
<< " " << *b
.shared_blob
;
1726 out
<< " (shared_blob=NULL)";
1732 void BlueStore::Blob::discard_unallocated(Collection
*coll
)
1734 if (get_blob().is_shared()) {
1737 if (get_blob().is_compressed()) {
1738 bool discard
= false;
1739 bool all_invalid
= true;
1740 for (auto e
: get_blob().get_extents()) {
1741 if (!e
.is_valid()) {
1744 all_invalid
= false;
1747 assert(discard
== all_invalid
); // in case of compressed blob all
1748 // or none pextents are invalid.
1750 shared_blob
->bc
.discard(shared_blob
->get_cache(), 0,
1751 get_blob().get_logical_length());
1755 for (auto e
: get_blob().get_extents()) {
1756 if (!e
.is_valid()) {
1757 ldout(coll
->store
->cct
, 20) << __func__
<< " 0x" << std::hex
<< pos
1759 << std::dec
<< dendl
;
1760 shared_blob
->bc
.discard(shared_blob
->get_cache(), pos
, e
.length
);
1764 if (get_blob().can_prune_tail()) {
1765 dirty_blob().prune_tail();
1766 used_in_blob
.prune_tail(get_blob().get_ondisk_length());
1767 auto cct
= coll
->store
->cct
; //used by dout
1768 dout(20) << __func__
<< " pruned tail, now " << get_blob() << dendl
;
1773 void BlueStore::Blob::get_ref(
1778 // Caller has to initialize Blob's logical length prior to increment
1779 // references. Otherwise one is neither unable to determine required
1780 // amount of counters in case of per-au tracking nor obtain min_release_size
1781 // for single counter mode.
1782 assert(get_blob().get_logical_length() != 0);
1783 auto cct
= coll
->store
->cct
;
1784 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1785 << std::dec
<< " " << *this << dendl
;
1787 if (used_in_blob
.is_empty()) {
1788 uint32_t min_release_size
=
1789 get_blob().get_release_size(coll
->store
->min_alloc_size
);
1790 uint64_t l
= get_blob().get_logical_length();
1791 dout(20) << __func__
<< " init 0x" << std::hex
<< l
<< ", "
1792 << min_release_size
<< std::dec
<< dendl
;
1793 used_in_blob
.init(l
, min_release_size
);
1800 bool BlueStore::Blob::put_ref(
1806 PExtentVector logical
;
1808 auto cct
= coll
->store
->cct
;
1809 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1810 << std::dec
<< " " << *this << dendl
;
1812 bool empty
= used_in_blob
.put(
1817 // nothing to release
1818 if (!empty
&& logical
.empty()) {
1822 bluestore_blob_t
& b
= dirty_blob();
1823 return b
.release_extents(empty
, logical
, r
);
1826 bool BlueStore::Blob::can_reuse_blob(uint32_t min_alloc_size
,
1827 uint32_t target_blob_size
,
1829 uint32_t *length0
) {
1830 assert(min_alloc_size
);
1831 assert(target_blob_size
);
1832 if (!get_blob().is_mutable()) {
1836 uint32_t length
= *length0
;
1837 uint32_t end
= b_offset
+ length
;
1839 // Currently for the sake of simplicity we omit blob reuse if data is
1840 // unaligned with csum chunk. Later we can perform padding if needed.
1841 if (get_blob().has_csum() &&
1842 ((b_offset
% get_blob().get_csum_chunk_size()) != 0 ||
1843 (end
% get_blob().get_csum_chunk_size()) != 0)) {
1847 auto blen
= get_blob().get_logical_length();
1848 uint32_t new_blen
= blen
;
1850 // make sure target_blob_size isn't less than current blob len
1851 target_blob_size
= MAX(blen
, target_blob_size
);
1853 if (b_offset
>= blen
) {
1854 // new data totally stands out of the existing blob
1857 // new data overlaps with the existing blob
1858 new_blen
= MAX(blen
, end
);
1860 uint32_t overlap
= 0;
1861 if (new_blen
> blen
) {
1862 overlap
= blen
- b_offset
;
1867 if (!get_blob().is_unallocated(b_offset
, overlap
)) {
1868 // abort if any piece of the overlap has already been allocated
1873 if (new_blen
> blen
) {
1874 int64_t overflow
= int64_t(new_blen
) - target_blob_size
;
1875 // Unable to decrease the provided length to fit into max_blob_size
1876 if (overflow
>= length
) {
1880 // FIXME: in some cases we could reduce unused resolution
1881 if (get_blob().has_unused()) {
1886 new_blen
-= overflow
;
1891 if (new_blen
> blen
) {
1892 dirty_blob().add_tail(new_blen
);
1893 used_in_blob
.add_tail(new_blen
,
1894 get_blob().get_release_size(min_alloc_size
));
1900 void BlueStore::Blob::split(Collection
*coll
, uint32_t blob_offset
, Blob
*r
)
1902 auto cct
= coll
->store
->cct
; //used by dout
1903 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1904 << " start " << *this << dendl
;
1905 assert(blob
.can_split());
1906 assert(used_in_blob
.can_split());
1907 bluestore_blob_t
&lb
= dirty_blob();
1908 bluestore_blob_t
&rb
= r
->dirty_blob();
1912 &(r
->used_in_blob
));
1914 lb
.split(blob_offset
, rb
);
1915 shared_blob
->bc
.split(shared_blob
->get_cache(), blob_offset
, r
->shared_blob
->bc
);
1917 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1918 << " finish " << *this << dendl
;
1919 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1920 << " and " << *r
<< dendl
;
1923 #ifndef CACHE_BLOB_BL
1924 void BlueStore::Blob::decode(
1926 bufferptr::iterator
& p
,
1929 bool include_ref_map
)
1931 denc(blob
, p
, struct_v
);
1932 if (blob
.is_shared()) {
1935 if (include_ref_map
) {
1937 used_in_blob
.decode(p
);
1939 used_in_blob
.clear();
1940 bluestore_extent_ref_map_t legacy_ref_map
;
1941 legacy_ref_map
.decode(p
);
1942 for (auto r
: legacy_ref_map
.ref_map
) {
1946 r
.second
.refs
* r
.second
.length
);
1955 ostream
& operator<<(ostream
& out
, const BlueStore::Extent
& e
)
1957 return out
<< std::hex
<< "0x" << e
.logical_offset
<< "~" << e
.length
1958 << ": 0x" << e
.blob_offset
<< "~" << e
.length
<< std::dec
1963 BlueStore::OldExtent
* BlueStore::OldExtent::create(CollectionRef c
,
1968 OldExtent
* oe
= new OldExtent(lo
, o
, l
, b
);
1969 b
->put_ref(c
.get(), o
, l
, &(oe
->r
));
1970 oe
->blob_empty
= b
->get_referenced_bytes() == 0;
1977 #define dout_prefix *_dout << "bluestore.extentmap(" << this << ") "
1979 BlueStore::ExtentMap::ExtentMap(Onode
*o
)
1982 o
->c
->store
->cct
->_conf
->bluestore_extent_map_inline_shard_prealloc_size
) {
1985 void BlueStore::ExtentMap::update(KeyValueDB::Transaction t
,
1988 auto cct
= onode
->c
->store
->cct
; //used by dout
1989 dout(20) << __func__
<< " " << onode
->oid
<< (force
? " force" : "") << dendl
;
1990 if (onode
->onode
.extent_map_shards
.empty()) {
1991 if (inline_bl
.length() == 0) {
1993 // we need to encode inline_bl to measure encoded length
1994 bool never_happen
= encode_some(0, OBJECT_MAX_SIZE
, inline_bl
, &n
);
1995 inline_bl
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
1996 assert(!never_happen
);
1997 size_t len
= inline_bl
.length();
1998 dout(20) << __func__
<< " inline shard " << len
<< " bytes from " << n
1999 << " extents" << dendl
;
2000 if (!force
&& len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
2001 request_reshard(0, OBJECT_MAX_SIZE
);
2005 // will persist in the onode key.
2007 // pending shard update
2008 struct dirty_shard_t
{
2011 dirty_shard_t(Shard
*s
) : shard(s
) {}
2013 vector
<dirty_shard_t
> encoded_shards
;
2014 // allocate slots for all shards in a single call instead of
2015 // doing multiple allocations - one per each dirty shard
2016 encoded_shards
.reserve(shards
.size());
2018 auto p
= shards
.begin();
2020 while (p
!= shards
.end()) {
2021 assert(p
->shard_info
->offset
>= prev_p
->shard_info
->offset
);
2026 if (n
== shards
.end()) {
2027 endoff
= OBJECT_MAX_SIZE
;
2029 endoff
= n
->shard_info
->offset
;
2031 encoded_shards
.emplace_back(dirty_shard_t(&(*p
)));
2032 bufferlist
& bl
= encoded_shards
.back().bl
;
2033 if (encode_some(p
->shard_info
->offset
, endoff
- p
->shard_info
->offset
,
2036 derr
<< __func__
<< " encode_some needs reshard" << dendl
;
2040 size_t len
= bl
.length();
2042 dout(20) << __func__
<< " shard 0x" << std::hex
2043 << p
->shard_info
->offset
<< std::dec
<< " is " << len
2044 << " bytes (was " << p
->shard_info
->bytes
<< ") from "
2045 << p
->extents
<< " extents" << dendl
;
2048 if (len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
2049 // we are big; reshard ourselves
2050 request_reshard(p
->shard_info
->offset
, endoff
);
2052 // avoid resharding the trailing shard, even if it is small
2053 else if (n
!= shards
.end() &&
2054 len
< g_conf
->bluestore_extent_map_shard_min_size
) {
2055 assert(endoff
!= OBJECT_MAX_SIZE
);
2056 if (p
== shards
.begin()) {
2057 // we are the first shard, combine with next shard
2058 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2060 // combine either with the previous shard or the next,
2061 // whichever is smaller
2062 if (prev_p
->shard_info
->bytes
> n
->shard_info
->bytes
) {
2063 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2065 request_reshard(prev_p
->shard_info
->offset
, endoff
);
2074 if (needs_reshard()) {
2078 // schedule DB update for dirty shards
2080 for (auto& it
: encoded_shards
) {
2081 it
.shard
->dirty
= false;
2082 it
.shard
->shard_info
->bytes
= it
.bl
.length();
2083 generate_extent_shard_key_and_apply(
2085 it
.shard
->shard_info
->offset
,
2087 [&](const string
& final_key
) {
2088 t
->set(PREFIX_OBJ
, final_key
, it
.bl
);
2095 bid_t
BlueStore::ExtentMap::allocate_spanning_blob_id()
2097 if (spanning_blob_map
.empty())
2099 bid_t bid
= spanning_blob_map
.rbegin()->first
+ 1;
2100 // bid is valid and available.
2103 // Find next unused bid;
2104 bid
= rand() % (numeric_limits
<bid_t
>::max() + 1);
2105 const auto begin_bid
= bid
;
2107 if (!spanning_blob_map
.count(bid
))
2111 if (bid
< 0) bid
= 0;
2113 } while (bid
!= begin_bid
);
2114 assert(0 == "no available blob id");
2117 void BlueStore::ExtentMap::reshard(
2119 KeyValueDB::Transaction t
)
2121 auto cct
= onode
->c
->store
->cct
; // used by dout
2123 dout(10) << __func__
<< " 0x[" << std::hex
<< needs_reshard_begin
<< ","
2124 << needs_reshard_end
<< ")" << std::dec
2125 << " of " << onode
->onode
.extent_map_shards
.size()
2126 << " shards on " << onode
->oid
<< dendl
;
2127 for (auto& p
: spanning_blob_map
) {
2128 dout(20) << __func__
<< " spanning blob " << p
.first
<< " " << *p
.second
2131 // determine shard index range
2132 unsigned si_begin
= 0, si_end
= 0;
2133 if (!shards
.empty()) {
2134 while (si_begin
+ 1 < shards
.size() &&
2135 shards
[si_begin
+ 1].shard_info
->offset
<= needs_reshard_begin
) {
2138 needs_reshard_begin
= shards
[si_begin
].shard_info
->offset
;
2139 for (si_end
= si_begin
; si_end
< shards
.size(); ++si_end
) {
2140 if (shards
[si_end
].shard_info
->offset
>= needs_reshard_end
) {
2141 needs_reshard_end
= shards
[si_end
].shard_info
->offset
;
2145 if (si_end
== shards
.size()) {
2146 needs_reshard_end
= OBJECT_MAX_SIZE
;
2148 dout(20) << __func__
<< " shards [" << si_begin
<< "," << si_end
<< ")"
2149 << " over 0x[" << std::hex
<< needs_reshard_begin
<< ","
2150 << needs_reshard_end
<< ")" << std::dec
<< dendl
;
2153 fault_range(db
, needs_reshard_begin
, (needs_reshard_end
- needs_reshard_begin
));
2155 // we may need to fault in a larger interval later must have all
2156 // referring extents for spanning blobs loaded in order to have
2157 // accurate use_tracker values.
2158 uint32_t spanning_scan_begin
= needs_reshard_begin
;
2159 uint32_t spanning_scan_end
= needs_reshard_end
;
2163 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2164 generate_extent_shard_key_and_apply(
2165 onode
->key
, shards
[i
].shard_info
->offset
, &key
,
2166 [&](const string
& final_key
) {
2167 t
->rmkey(PREFIX_OBJ
, final_key
);
2172 // calculate average extent size
2174 unsigned extents
= 0;
2175 if (onode
->onode
.extent_map_shards
.empty()) {
2176 bytes
= inline_bl
.length();
2177 extents
= extent_map
.size();
2179 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2180 bytes
+= shards
[i
].shard_info
->bytes
;
2181 extents
+= shards
[i
].extents
;
2184 unsigned target
= cct
->_conf
->bluestore_extent_map_shard_target_size
;
2185 unsigned slop
= target
*
2186 cct
->_conf
->bluestore_extent_map_shard_target_size_slop
;
2187 unsigned extent_avg
= bytes
/ MAX(1, extents
);
2188 dout(20) << __func__
<< " extent_avg " << extent_avg
<< ", target " << target
2189 << ", slop " << slop
<< dendl
;
2192 unsigned estimate
= 0;
2193 unsigned offset
= needs_reshard_begin
;
2194 vector
<bluestore_onode_t::shard_info
> new_shard_info
;
2195 unsigned max_blob_end
= 0;
2196 Extent
dummy(needs_reshard_begin
);
2197 for (auto e
= extent_map
.lower_bound(dummy
);
2198 e
!= extent_map
.end();
2200 if (e
->logical_offset
>= needs_reshard_end
) {
2203 dout(30) << " extent " << *e
<< dendl
;
2205 // disfavor shard boundaries that span a blob
2206 bool would_span
= (e
->logical_offset
< max_blob_end
) || e
->blob_offset
;
2208 estimate
+ extent_avg
> target
+ (would_span
? slop
: 0)) {
2210 if (offset
== needs_reshard_begin
) {
2211 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2212 new_shard_info
.back().offset
= offset
;
2213 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2214 << std::dec
<< dendl
;
2216 offset
= e
->logical_offset
;
2217 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2218 new_shard_info
.back().offset
= offset
;
2219 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2220 << std::dec
<< dendl
;
2223 estimate
+= extent_avg
;
2224 unsigned bs
= e
->blob_start();
2225 if (bs
< spanning_scan_begin
) {
2226 spanning_scan_begin
= bs
;
2228 uint32_t be
= e
->blob_end();
2229 if (be
> max_blob_end
) {
2232 if (be
> spanning_scan_end
) {
2233 spanning_scan_end
= be
;
2236 if (new_shard_info
.empty() && (si_begin
> 0 ||
2237 si_end
< shards
.size())) {
2238 // we resharded a partial range; we must produce at least one output
2240 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2241 new_shard_info
.back().offset
= needs_reshard_begin
;
2242 dout(20) << __func__
<< " new shard 0x" << std::hex
<< needs_reshard_begin
2243 << std::dec
<< " (singleton degenerate case)" << dendl
;
2246 auto& sv
= onode
->onode
.extent_map_shards
;
2247 dout(20) << __func__
<< " new " << new_shard_info
<< dendl
;
2248 dout(20) << __func__
<< " old " << sv
<< dendl
;
2250 // no old shards to keep
2251 sv
.swap(new_shard_info
);
2252 init_shards(true, true);
2254 // splice in new shards
2255 sv
.erase(sv
.begin() + si_begin
, sv
.begin() + si_end
);
2256 shards
.erase(shards
.begin() + si_begin
, shards
.begin() + si_end
);
2258 sv
.begin() + si_begin
,
2259 new_shard_info
.begin(),
2260 new_shard_info
.end());
2261 shards
.insert(shards
.begin() + si_begin
, new_shard_info
.size(), Shard());
2262 si_end
= si_begin
+ new_shard_info
.size();
2264 assert(sv
.size() == shards
.size());
2266 // note that we need to update every shard_info of shards here,
2267 // as sv might have been totally re-allocated above
2268 for (unsigned i
= 0; i
< shards
.size(); i
++) {
2269 shards
[i
].shard_info
= &sv
[i
];
2272 // mark newly added shards as dirty
2273 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2274 shards
[i
].loaded
= true;
2275 shards
[i
].dirty
= true;
2278 dout(20) << __func__
<< " fin " << sv
<< dendl
;
2282 // no more shards; unspan all previously spanning blobs
2283 auto p
= spanning_blob_map
.begin();
2284 while (p
!= spanning_blob_map
.end()) {
2286 dout(30) << __func__
<< " un-spanning " << *p
->second
<< dendl
;
2287 p
= spanning_blob_map
.erase(p
);
2290 // identify new spanning blobs
2291 dout(20) << __func__
<< " checking spanning blobs 0x[" << std::hex
2292 << spanning_scan_begin
<< "," << spanning_scan_end
<< ")" << dendl
;
2293 if (spanning_scan_begin
< needs_reshard_begin
) {
2294 fault_range(db
, spanning_scan_begin
,
2295 needs_reshard_begin
- spanning_scan_begin
);
2297 if (spanning_scan_end
> needs_reshard_end
) {
2298 fault_range(db
, needs_reshard_end
,
2299 spanning_scan_end
- needs_reshard_end
);
2301 auto sp
= sv
.begin() + si_begin
;
2302 auto esp
= sv
.end();
2303 unsigned shard_start
= sp
->offset
;
2307 shard_end
= OBJECT_MAX_SIZE
;
2309 shard_end
= sp
->offset
;
2311 Extent
dummy(needs_reshard_begin
);
2312 for (auto e
= extent_map
.lower_bound(dummy
); e
!= extent_map
.end(); ++e
) {
2313 if (e
->logical_offset
>= needs_reshard_end
) {
2316 dout(30) << " extent " << *e
<< dendl
;
2317 while (e
->logical_offset
>= shard_end
) {
2318 shard_start
= shard_end
;
2322 shard_end
= OBJECT_MAX_SIZE
;
2324 shard_end
= sp
->offset
;
2326 dout(30) << __func__
<< " shard 0x" << std::hex
<< shard_start
2327 << " to 0x" << shard_end
<< std::dec
<< dendl
;
2329 if (e
->blob_escapes_range(shard_start
, shard_end
- shard_start
)) {
2330 if (!e
->blob
->is_spanning()) {
2331 // We have two options: (1) split the blob into pieces at the
2332 // shard boundaries (and adjust extents accordingly), or (2)
2333 // mark it spanning. We prefer to cut the blob if we can. Note that
2334 // we may have to split it multiple times--potentially at every
2336 bool must_span
= false;
2337 BlobRef b
= e
->blob
;
2338 if (b
->can_split()) {
2339 uint32_t bstart
= e
->blob_start();
2340 uint32_t bend
= e
->blob_end();
2341 for (const auto& sh
: shards
) {
2342 if (bstart
< sh
.shard_info
->offset
&&
2343 bend
> sh
.shard_info
->offset
) {
2344 uint32_t blob_offset
= sh
.shard_info
->offset
- bstart
;
2345 if (b
->can_split_at(blob_offset
)) {
2346 dout(20) << __func__
<< " splitting blob, bstart 0x"
2347 << std::hex
<< bstart
<< " blob_offset 0x"
2348 << blob_offset
<< std::dec
<< " " << *b
<< dendl
;
2349 b
= split_blob(b
, blob_offset
, sh
.shard_info
->offset
);
2350 // switch b to the new right-hand side, in case it
2351 // *also* has to get split.
2352 bstart
+= blob_offset
;
2353 onode
->c
->store
->logger
->inc(l_bluestore_blob_split
);
2364 auto bid
= allocate_spanning_blob_id();
2366 spanning_blob_map
[b
->id
] = b
;
2367 dout(20) << __func__
<< " adding spanning " << *b
<< dendl
;
2371 if (e
->blob
->is_spanning()) {
2372 spanning_blob_map
.erase(e
->blob
->id
);
2374 dout(30) << __func__
<< " un-spanning " << *e
->blob
<< dendl
;
2380 clear_needs_reshard();
2383 bool BlueStore::ExtentMap::encode_some(
2389 auto cct
= onode
->c
->store
->cct
; //used by dout
2390 Extent
dummy(offset
);
2391 auto start
= extent_map
.lower_bound(dummy
);
2392 uint32_t end
= offset
+ length
;
2394 __u8 struct_v
= 2; // Version 2 differs from v1 in blob's ref_map
2395 // serialization only. Hence there is no specific
2396 // handling at ExtentMap level.
2400 bool must_reshard
= false;
2401 for (auto p
= start
;
2402 p
!= extent_map
.end() && p
->logical_offset
< end
;
2404 assert(p
->logical_offset
>= offset
);
2405 p
->blob
->last_encoded_id
= -1;
2406 if (!p
->blob
->is_spanning() && p
->blob_escapes_range(offset
, length
)) {
2407 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2408 << std::dec
<< " hit new spanning blob " << *p
<< dendl
;
2409 request_reshard(p
->blob_start(), p
->blob_end());
2410 must_reshard
= true;
2412 if (!must_reshard
) {
2413 denc_varint(0, bound
); // blobid
2414 denc_varint(0, bound
); // logical_offset
2415 denc_varint(0, bound
); // len
2416 denc_varint(0, bound
); // blob_offset
2418 p
->blob
->bound_encode(
2421 p
->blob
->shared_blob
->get_sbid(),
2429 denc(struct_v
, bound
);
2430 denc_varint(0, bound
); // number of extents
2433 auto app
= bl
.get_contiguous_appender(bound
);
2434 denc(struct_v
, app
);
2435 denc_varint(n
, app
);
2442 uint64_t prev_len
= 0;
2443 for (auto p
= start
;
2444 p
!= extent_map
.end() && p
->logical_offset
< end
;
2447 bool include_blob
= false;
2448 if (p
->blob
->is_spanning()) {
2449 blobid
= p
->blob
->id
<< BLOBID_SHIFT_BITS
;
2450 blobid
|= BLOBID_FLAG_SPANNING
;
2451 } else if (p
->blob
->last_encoded_id
< 0) {
2452 p
->blob
->last_encoded_id
= n
+ 1; // so it is always non-zero
2453 include_blob
= true;
2454 blobid
= 0; // the decoder will infer the id from n
2456 blobid
= p
->blob
->last_encoded_id
<< BLOBID_SHIFT_BITS
;
2458 if (p
->logical_offset
== pos
) {
2459 blobid
|= BLOBID_FLAG_CONTIGUOUS
;
2461 if (p
->blob_offset
== 0) {
2462 blobid
|= BLOBID_FLAG_ZEROOFFSET
;
2464 if (p
->length
== prev_len
) {
2465 blobid
|= BLOBID_FLAG_SAMELENGTH
;
2467 prev_len
= p
->length
;
2469 denc_varint(blobid
, app
);
2470 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2471 denc_varint_lowz(p
->logical_offset
- pos
, app
);
2473 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2474 denc_varint_lowz(p
->blob_offset
, app
);
2476 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2477 denc_varint_lowz(p
->length
, app
);
2479 pos
= p
->logical_end();
2481 p
->blob
->encode(app
, struct_v
, p
->blob
->shared_blob
->get_sbid(), false);
2485 /*derr << __func__ << bl << dendl;
2486 derr << __func__ << ":";
2493 unsigned BlueStore::ExtentMap::decode_some(bufferlist
& bl
)
2495 auto cct
= onode
->c
->store
->cct
; //used by dout
2497 derr << __func__ << ":";
2502 assert(bl
.get_num_buffers() <= 1);
2503 auto p
= bl
.front().begin_deep();
2506 // Version 2 differs from v1 in blob's ref_map
2507 // serialization only. Hence there is no specific
2508 // handling at ExtentMap level below.
2509 assert(struct_v
== 1 || struct_v
== 2);
2512 denc_varint(num
, p
);
2513 vector
<BlobRef
> blobs(num
);
2515 uint64_t prev_len
= 0;
2519 Extent
*le
= new Extent();
2521 denc_varint(blobid
, p
);
2522 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2524 denc_varint_lowz(gap
, p
);
2527 le
->logical_offset
= pos
;
2528 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2529 denc_varint_lowz(le
->blob_offset
, p
);
2531 le
->blob_offset
= 0;
2533 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2534 denc_varint_lowz(prev_len
, p
);
2536 le
->length
= prev_len
;
2538 if (blobid
& BLOBID_FLAG_SPANNING
) {
2539 dout(30) << __func__
<< " getting spanning blob "
2540 << (blobid
>> BLOBID_SHIFT_BITS
) << dendl
;
2541 le
->assign_blob(get_spanning_blob(blobid
>> BLOBID_SHIFT_BITS
));
2543 blobid
>>= BLOBID_SHIFT_BITS
;
2545 le
->assign_blob(blobs
[blobid
- 1]);
2548 Blob
*b
= new Blob();
2550 b
->decode(onode
->c
, p
, struct_v
, &sbid
, false);
2552 onode
->c
->open_shared_blob(sbid
, b
);
2555 // we build ref_map dynamically for non-spanning blobs
2563 extent_map
.insert(*le
);
2570 void BlueStore::ExtentMap::bound_encode_spanning_blobs(size_t& p
)
2572 // Version 2 differs from v1 in blob's ref_map
2573 // serialization only. Hence there is no specific
2574 // handling at ExtentMap level.
2578 denc_varint((uint32_t)0, p
);
2579 size_t key_size
= 0;
2580 denc_varint((uint32_t)0, key_size
);
2581 p
+= spanning_blob_map
.size() * key_size
;
2582 for (const auto& i
: spanning_blob_map
) {
2583 i
.second
->bound_encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2587 void BlueStore::ExtentMap::encode_spanning_blobs(
2588 bufferlist::contiguous_appender
& p
)
2590 // Version 2 differs from v1 in blob's ref_map
2591 // serialization only. Hence there is no specific
2592 // handling at ExtentMap level.
2596 denc_varint(spanning_blob_map
.size(), p
);
2597 for (auto& i
: spanning_blob_map
) {
2598 denc_varint(i
.second
->id
, p
);
2599 i
.second
->encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2603 void BlueStore::ExtentMap::decode_spanning_blobs(
2604 bufferptr::iterator
& p
)
2608 // Version 2 differs from v1 in blob's ref_map
2609 // serialization only. Hence there is no specific
2610 // handling at ExtentMap level.
2611 assert(struct_v
== 1 || struct_v
== 2);
2616 BlobRef
b(new Blob());
2617 denc_varint(b
->id
, p
);
2618 spanning_blob_map
[b
->id
] = b
;
2620 b
->decode(onode
->c
, p
, struct_v
, &sbid
, true);
2621 onode
->c
->open_shared_blob(sbid
, b
);
2625 void BlueStore::ExtentMap::init_shards(bool loaded
, bool dirty
)
2627 shards
.resize(onode
->onode
.extent_map_shards
.size());
2629 for (auto &s
: onode
->onode
.extent_map_shards
) {
2630 shards
[i
].shard_info
= &s
;
2631 shards
[i
].loaded
= loaded
;
2632 shards
[i
].dirty
= dirty
;
2637 void BlueStore::ExtentMap::fault_range(
2642 auto cct
= onode
->c
->store
->cct
; //used by dout
2643 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2644 << std::dec
<< dendl
;
2645 auto start
= seek_shard(offset
);
2646 auto last
= seek_shard(offset
+ length
);
2651 assert(last
>= start
);
2653 while (start
<= last
) {
2654 assert((size_t)start
< shards
.size());
2655 auto p
= &shards
[start
];
2657 dout(30) << __func__
<< " opening shard 0x" << std::hex
2658 << p
->shard_info
->offset
<< std::dec
<< dendl
;
2660 generate_extent_shard_key_and_apply(
2661 onode
->key
, p
->shard_info
->offset
, &key
,
2662 [&](const string
& final_key
) {
2663 int r
= db
->get(PREFIX_OBJ
, final_key
, &v
);
2665 derr
<< __func__
<< " missing shard 0x" << std::hex
2666 << p
->shard_info
->offset
<< std::dec
<< " for " << onode
->oid
2672 p
->extents
= decode_some(v
);
2674 dout(20) << __func__
<< " open shard 0x" << std::hex
2675 << p
->shard_info
->offset
<< std::dec
2676 << " (" << v
.length() << " bytes)" << dendl
;
2677 assert(p
->dirty
== false);
2678 assert(v
.length() == p
->shard_info
->bytes
);
2679 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_misses
);
2681 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_hits
);
2687 void BlueStore::ExtentMap::dirty_range(
2691 auto cct
= onode
->c
->store
->cct
; //used by dout
2692 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2693 << std::dec
<< dendl
;
2694 if (shards
.empty()) {
2695 dout(20) << __func__
<< " mark inline shard dirty" << dendl
;
2699 auto start
= seek_shard(offset
);
2700 auto last
= seek_shard(offset
+ length
);
2704 assert(last
>= start
);
2705 while (start
<= last
) {
2706 assert((size_t)start
< shards
.size());
2707 auto p
= &shards
[start
];
2709 dout(20) << __func__
<< " shard 0x" << std::hex
<< p
->shard_info
->offset
2710 << std::dec
<< " is not loaded, can't mark dirty" << dendl
;
2711 assert(0 == "can't mark unloaded shard dirty");
2714 dout(20) << __func__
<< " mark shard 0x" << std::hex
2715 << p
->shard_info
->offset
<< std::dec
<< " dirty" << dendl
;
2722 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::find(
2725 Extent
dummy(offset
);
2726 return extent_map
.find(dummy
);
2729 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::seek_lextent(
2732 Extent
dummy(offset
);
2733 auto fp
= extent_map
.lower_bound(dummy
);
2734 if (fp
!= extent_map
.begin()) {
2736 if (fp
->logical_end() <= offset
) {
2743 BlueStore::extent_map_t::const_iterator
BlueStore::ExtentMap::seek_lextent(
2744 uint64_t offset
) const
2746 Extent
dummy(offset
);
2747 auto fp
= extent_map
.lower_bound(dummy
);
2748 if (fp
!= extent_map
.begin()) {
2750 if (fp
->logical_end() <= offset
) {
2757 bool BlueStore::ExtentMap::has_any_lextents(uint64_t offset
, uint64_t length
)
2759 auto fp
= seek_lextent(offset
);
2760 if (fp
== extent_map
.end() || fp
->logical_offset
>= offset
+ length
) {
2766 int BlueStore::ExtentMap::compress_extent_map(
2770 auto cct
= onode
->c
->store
->cct
; //used by dout
2771 if (extent_map
.empty())
2774 auto p
= seek_lextent(offset
);
2775 if (p
!= extent_map
.begin()) {
2776 --p
; // start to the left of offset
2778 // the caller should have just written to this region
2779 assert(p
!= extent_map
.end());
2781 // identify the *next* shard
2782 auto pshard
= shards
.begin();
2783 while (pshard
!= shards
.end() &&
2784 p
->logical_offset
>= pshard
->shard_info
->offset
) {
2788 if (pshard
!= shards
.end()) {
2789 shard_end
= pshard
->shard_info
->offset
;
2791 shard_end
= OBJECT_MAX_SIZE
;
2795 for (++n
; n
!= extent_map
.end(); p
= n
++) {
2796 if (n
->logical_offset
> offset
+ length
) {
2797 break; // stop after end
2799 while (n
!= extent_map
.end() &&
2800 p
->logical_end() == n
->logical_offset
&&
2801 p
->blob
== n
->blob
&&
2802 p
->blob_offset
+ p
->length
== n
->blob_offset
&&
2803 n
->logical_offset
< shard_end
) {
2804 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2805 << " next shard 0x" << shard_end
<< std::dec
2806 << " merging " << *p
<< " and " << *n
<< dendl
;
2807 p
->length
+= n
->length
;
2811 if (n
== extent_map
.end()) {
2814 if (n
->logical_offset
>= shard_end
) {
2815 assert(pshard
!= shards
.end());
2817 if (pshard
!= shards
.end()) {
2818 shard_end
= pshard
->shard_info
->offset
;
2820 shard_end
= OBJECT_MAX_SIZE
;
2824 if (removed
&& onode
) {
2825 onode
->c
->store
->logger
->inc(l_bluestore_extent_compress
, removed
);
2830 void BlueStore::ExtentMap::punch_hole(
2834 old_extent_map_t
*old_extents
)
2836 auto p
= seek_lextent(offset
);
2837 uint64_t end
= offset
+ length
;
2838 while (p
!= extent_map
.end()) {
2839 if (p
->logical_offset
>= end
) {
2842 if (p
->logical_offset
< offset
) {
2843 if (p
->logical_end() > end
) {
2844 // split and deref middle
2845 uint64_t front
= offset
- p
->logical_offset
;
2846 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ front
,
2848 old_extents
->push_back(*oe
);
2850 p
->blob_offset
+ front
+ length
,
2851 p
->length
- front
- length
,
2857 assert(p
->logical_end() > offset
); // else seek_lextent bug
2858 uint64_t keep
= offset
- p
->logical_offset
;
2859 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ keep
,
2860 p
->length
- keep
, p
->blob
);
2861 old_extents
->push_back(*oe
);
2867 if (p
->logical_offset
+ p
->length
<= end
) {
2868 // deref whole lextent
2869 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2870 p
->length
, p
->blob
);
2871 old_extents
->push_back(*oe
);
2876 uint64_t keep
= p
->logical_end() - end
;
2877 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2878 p
->length
- keep
, p
->blob
);
2879 old_extents
->push_back(*oe
);
2881 add(end
, p
->blob_offset
+ p
->length
- keep
, keep
, p
->blob
);
2887 BlueStore::Extent
*BlueStore::ExtentMap::set_lextent(
2889 uint64_t logical_offset
,
2890 uint64_t blob_offset
, uint64_t length
, BlobRef b
,
2891 old_extent_map_t
*old_extents
)
2893 // We need to have completely initialized Blob to increment its ref counters.
2894 assert(b
->get_blob().get_logical_length() != 0);
2896 // Do get_ref prior to punch_hole to prevent from putting reused blob into
2897 // old_extents list if we overwre the blob totally
2898 // This might happen during WAL overwrite.
2899 b
->get_ref(onode
->c
, blob_offset
, length
);
2902 punch_hole(c
, logical_offset
, length
, old_extents
);
2905 Extent
*le
= new Extent(logical_offset
, blob_offset
, length
, b
);
2906 extent_map
.insert(*le
);
2907 if (spans_shard(logical_offset
, length
)) {
2908 request_reshard(logical_offset
, logical_offset
+ length
);
2913 BlueStore::BlobRef
BlueStore::ExtentMap::split_blob(
2915 uint32_t blob_offset
,
2918 auto cct
= onode
->c
->store
->cct
; //used by dout
2920 uint32_t end_pos
= pos
+ lb
->get_blob().get_logical_length() - blob_offset
;
2921 dout(20) << __func__
<< " 0x" << std::hex
<< pos
<< " end 0x" << end_pos
2922 << " blob_offset 0x" << blob_offset
<< std::dec
<< " " << *lb
2924 BlobRef rb
= onode
->c
->new_blob();
2925 lb
->split(onode
->c
, blob_offset
, rb
.get());
2927 for (auto ep
= seek_lextent(pos
);
2928 ep
!= extent_map
.end() && ep
->logical_offset
< end_pos
;
2930 if (ep
->blob
!= lb
) {
2933 if (ep
->logical_offset
< pos
) {
2935 size_t left
= pos
- ep
->logical_offset
;
2936 Extent
*ne
= new Extent(pos
, 0, ep
->length
- left
, rb
);
2937 extent_map
.insert(*ne
);
2939 dout(30) << __func__
<< " split " << *ep
<< dendl
;
2940 dout(30) << __func__
<< " to " << *ne
<< dendl
;
2943 assert(ep
->blob_offset
>= blob_offset
);
2946 ep
->blob_offset
-= blob_offset
;
2947 dout(30) << __func__
<< " adjusted " << *ep
<< dendl
;
2956 #define dout_prefix *_dout << "bluestore.onode(" << this << ")." << __func__ << " "
2958 void BlueStore::Onode::flush()
2960 if (flushing_count
.load()) {
2961 ldout(c
->store
->cct
, 20) << __func__
<< " cnt:" << flushing_count
<< dendl
;
2962 std::unique_lock
<std::mutex
> l(flush_lock
);
2963 while (flushing_count
.load()) {
2967 ldout(c
->store
->cct
, 20) << __func__
<< " done" << dendl
;
2970 // =======================================================
2973 /// Checks for writes to the same pextent within a blob
2974 bool BlueStore::WriteContext::has_conflict(
2978 uint64_t min_alloc_size
)
2980 assert((loffs
% min_alloc_size
) == 0);
2981 assert((loffs_end
% min_alloc_size
) == 0);
2982 for (auto w
: writes
) {
2984 auto loffs2
= P2ALIGN(w
.logical_offset
, min_alloc_size
);
2985 auto loffs2_end
= P2ROUNDUP(w
.logical_offset
+ w
.length0
, min_alloc_size
);
2986 if ((loffs
<= loffs2
&& loffs_end
> loffs2
) ||
2987 (loffs
>= loffs2
&& loffs
< loffs2_end
)) {
2995 // =======================================================
2999 #define dout_prefix *_dout << "bluestore.DeferredBatch(" << this << ") "
3001 void BlueStore::DeferredBatch::prepare_write(
3003 uint64_t seq
, uint64_t offset
, uint64_t length
,
3004 bufferlist::const_iterator
& blp
)
3006 _discard(cct
, offset
, length
);
3007 auto i
= iomap
.insert(make_pair(offset
, deferred_io()));
3008 assert(i
.second
); // this should be a new insertion
3009 i
.first
->second
.seq
= seq
;
3010 blp
.copy(length
, i
.first
->second
.bl
);
3011 i
.first
->second
.bl
.reassign_to_mempool(
3012 mempool::mempool_bluestore_writing_deferred
);
3013 dout(20) << __func__
<< " seq " << seq
3014 << " 0x" << std::hex
<< offset
<< "~" << length
3015 << " crc " << i
.first
->second
.bl
.crc32c(-1)
3016 << std::dec
<< dendl
;
3017 seq_bytes
[seq
] += length
;
3018 #ifdef DEBUG_DEFERRED
3023 void BlueStore::DeferredBatch::_discard(
3024 CephContext
*cct
, uint64_t offset
, uint64_t length
)
3026 generic_dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
3027 << std::dec
<< dendl
;
3028 auto p
= iomap
.lower_bound(offset
);
3029 if (p
!= iomap
.begin()) {
3031 auto end
= p
->first
+ p
->second
.bl
.length();
3034 head
.substr_of(p
->second
.bl
, 0, offset
- p
->first
);
3035 dout(20) << __func__
<< " keep head " << p
->second
.seq
3036 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3037 << " -> 0x" << head
.length() << std::dec
<< dendl
;
3038 auto i
= seq_bytes
.find(p
->second
.seq
);
3039 assert(i
!= seq_bytes
.end());
3040 if (end
> offset
+ length
) {
3042 tail
.substr_of(p
->second
.bl
, offset
+ length
- p
->first
,
3043 end
- (offset
+ length
));
3044 dout(20) << __func__
<< " keep tail " << p
->second
.seq
3045 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3046 << " -> 0x" << tail
.length() << std::dec
<< dendl
;
3047 auto &n
= iomap
[offset
+ length
];
3049 n
.seq
= p
->second
.seq
;
3050 i
->second
-= length
;
3052 i
->second
-= end
- offset
;
3054 assert(i
->second
>= 0);
3055 p
->second
.bl
.swap(head
);
3059 while (p
!= iomap
.end()) {
3060 if (p
->first
>= offset
+ length
) {
3063 auto i
= seq_bytes
.find(p
->second
.seq
);
3064 assert(i
!= seq_bytes
.end());
3065 auto end
= p
->first
+ p
->second
.bl
.length();
3066 if (end
> offset
+ length
) {
3067 unsigned drop_front
= offset
+ length
- p
->first
;
3068 unsigned keep_tail
= end
- (offset
+ length
);
3069 dout(20) << __func__
<< " truncate front " << p
->second
.seq
3070 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3071 << " drop_front 0x" << drop_front
<< " keep_tail 0x" << keep_tail
3072 << " to 0x" << (offset
+ length
) << "~" << keep_tail
3073 << std::dec
<< dendl
;
3074 auto &s
= iomap
[offset
+ length
];
3075 s
.seq
= p
->second
.seq
;
3076 s
.bl
.substr_of(p
->second
.bl
, drop_front
, keep_tail
);
3077 i
->second
-= drop_front
;
3079 dout(20) << __func__
<< " drop " << p
->second
.seq
3080 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3081 << std::dec
<< dendl
;
3082 i
->second
-= p
->second
.bl
.length();
3084 assert(i
->second
>= 0);
3089 void BlueStore::DeferredBatch::_audit(CephContext
*cct
)
3091 map
<uint64_t,int> sb
;
3092 for (auto p
: seq_bytes
) {
3093 sb
[p
.first
] = 0; // make sure we have the same set of keys
3096 for (auto& p
: iomap
) {
3097 assert(p
.first
>= pos
);
3098 sb
[p
.second
.seq
] += p
.second
.bl
.length();
3099 pos
= p
.first
+ p
.second
.bl
.length();
3101 assert(sb
== seq_bytes
);
3108 #define dout_prefix *_dout << "bluestore(" << store->path << ").collection(" << cid << " " << this << ") "
3110 BlueStore::Collection::Collection(BlueStore
*ns
, Cache
*c
, coll_t cid
)
3114 lock("BlueStore::Collection::lock", true, false),
3120 void BlueStore::Collection::open_shared_blob(uint64_t sbid
, BlobRef b
)
3122 assert(!b
->shared_blob
);
3123 const bluestore_blob_t
& blob
= b
->get_blob();
3124 if (!blob
.is_shared()) {
3125 b
->shared_blob
= new SharedBlob(this);
3129 b
->shared_blob
= shared_blob_set
.lookup(sbid
);
3130 if (b
->shared_blob
) {
3131 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3132 << std::dec
<< " had " << *b
->shared_blob
<< dendl
;
3134 b
->shared_blob
= new SharedBlob(sbid
, this);
3135 shared_blob_set
.add(this, b
->shared_blob
.get());
3136 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3137 << std::dec
<< " opened " << *b
->shared_blob
3142 void BlueStore::Collection::load_shared_blob(SharedBlobRef sb
)
3144 if (!sb
->is_loaded()) {
3148 auto sbid
= sb
->get_sbid();
3149 get_shared_blob_key(sbid
, &key
);
3150 int r
= store
->db
->get(PREFIX_SHARED_BLOB
, key
, &v
);
3152 lderr(store
->cct
) << __func__
<< " sbid 0x" << std::hex
<< sbid
3153 << std::dec
<< " not found at key "
3154 << pretty_binary_string(key
) << dendl
;
3155 assert(0 == "uh oh, missing shared_blob");
3159 sb
->persistent
= new bluestore_shared_blob_t(sbid
);
3160 bufferlist::iterator p
= v
.begin();
3161 ::decode(*(sb
->persistent
), p
);
3162 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3163 << std::dec
<< " loaded shared_blob " << *sb
<< dendl
;
3167 void BlueStore::Collection::make_blob_shared(uint64_t sbid
, BlobRef b
)
3169 ldout(store
->cct
, 10) << __func__
<< " " << *b
<< dendl
;
3170 assert(!b
->shared_blob
->is_loaded());
3173 bluestore_blob_t
& blob
= b
->dirty_blob();
3174 blob
.set_flag(bluestore_blob_t::FLAG_SHARED
);
3176 // update shared blob
3177 b
->shared_blob
->loaded
= true;
3178 b
->shared_blob
->persistent
= new bluestore_shared_blob_t(sbid
);
3179 shared_blob_set
.add(this, b
->shared_blob
.get());
3180 for (auto p
: blob
.get_extents()) {
3182 b
->shared_blob
->get_ref(
3187 ldout(store
->cct
, 20) << __func__
<< " now " << *b
<< dendl
;
3190 uint64_t BlueStore::Collection::make_blob_unshared(SharedBlob
*sb
)
3192 ldout(store
->cct
, 10) << __func__
<< " " << *sb
<< dendl
;
3193 assert(sb
->is_loaded());
3195 uint64_t sbid
= sb
->get_sbid();
3196 shared_blob_set
.remove(sb
);
3198 delete sb
->persistent
;
3199 sb
->sbid_unloaded
= 0;
3200 ldout(store
->cct
, 20) << __func__
<< " now " << *sb
<< dendl
;
3204 BlueStore::OnodeRef
BlueStore::Collection::get_onode(
3205 const ghobject_t
& oid
,
3208 assert(create
? lock
.is_wlocked() : lock
.is_locked());
3211 if (cid
.is_pg(&pgid
)) {
3212 if (!oid
.match(cnode
.bits
, pgid
.ps())) {
3213 lderr(store
->cct
) << __func__
<< " oid " << oid
<< " not part of "
3214 << pgid
<< " bits " << cnode
.bits
<< dendl
;
3219 OnodeRef o
= onode_map
.lookup(oid
);
3223 mempool::bluestore_cache_other::string key
;
3224 get_object_key(store
->cct
, oid
, &key
);
3226 ldout(store
->cct
, 20) << __func__
<< " oid " << oid
<< " key "
3227 << pretty_binary_string(key
) << dendl
;
3230 int r
= store
->db
->get(PREFIX_OBJ
, key
.c_str(), key
.size(), &v
);
3231 ldout(store
->cct
, 20) << " r " << r
<< " v.len " << v
.length() << dendl
;
3233 if (v
.length() == 0) {
3234 assert(r
== -ENOENT
);
3235 if (!store
->cct
->_conf
->bluestore_debug_misc
&&
3239 // new object, new onode
3240 on
= new Onode(this, oid
, key
);
3244 on
= new Onode(this, oid
, key
);
3246 bufferptr::iterator p
= v
.front().begin_deep();
3247 on
->onode
.decode(p
);
3248 for (auto& i
: on
->onode
.attrs
) {
3249 i
.second
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
3252 // initialize extent_map
3253 on
->extent_map
.decode_spanning_blobs(p
);
3254 if (on
->onode
.extent_map_shards
.empty()) {
3255 denc(on
->extent_map
.inline_bl
, p
);
3256 on
->extent_map
.decode_some(on
->extent_map
.inline_bl
);
3257 on
->extent_map
.inline_bl
.reassign_to_mempool(
3258 mempool::mempool_bluestore_cache_other
);
3260 on
->extent_map
.init_shards(false, false);
3264 return onode_map
.add(oid
, o
);
3267 void BlueStore::Collection::split_cache(
3270 ldout(store
->cct
, 10) << __func__
<< " to " << dest
<< dendl
;
3272 // lock (one or both) cache shards
3273 std::lock(cache
->lock
, dest
->cache
->lock
);
3274 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
, std::adopt_lock
);
3275 std::lock_guard
<std::recursive_mutex
> l2(dest
->cache
->lock
, std::adopt_lock
);
3277 int destbits
= dest
->cnode
.bits
;
3279 bool is_pg
= dest
->cid
.is_pg(&destpg
);
3282 auto p
= onode_map
.onode_map
.begin();
3283 while (p
!= onode_map
.onode_map
.end()) {
3284 if (!p
->second
->oid
.match(destbits
, destpg
.pgid
.ps())) {
3285 // onode does not belong to this child
3288 OnodeRef o
= p
->second
;
3289 ldout(store
->cct
, 20) << __func__
<< " moving " << o
<< " " << o
->oid
3292 cache
->_rm_onode(p
->second
);
3293 p
= onode_map
.onode_map
.erase(p
);
3296 dest
->cache
->_add_onode(o
, 1);
3297 dest
->onode_map
.onode_map
[o
->oid
] = o
;
3298 dest
->onode_map
.cache
= dest
->cache
;
3300 // move over shared blobs and buffers. cover shared blobs from
3301 // both extent map and spanning blob map (the full extent map
3302 // may not be faulted in)
3303 vector
<SharedBlob
*> sbvec
;
3304 for (auto& e
: o
->extent_map
.extent_map
) {
3305 sbvec
.push_back(e
.blob
->shared_blob
.get());
3307 for (auto& b
: o
->extent_map
.spanning_blob_map
) {
3308 sbvec
.push_back(b
.second
->shared_blob
.get());
3310 for (auto sb
: sbvec
) {
3311 if (sb
->coll
== dest
) {
3312 ldout(store
->cct
, 20) << __func__
<< " already moved " << *sb
3316 ldout(store
->cct
, 20) << __func__
<< " moving " << *sb
<< dendl
;
3317 if (sb
->get_sbid()) {
3318 ldout(store
->cct
, 20) << __func__
3319 << " moving registration " << *sb
<< dendl
;
3320 shared_blob_set
.remove(sb
);
3321 dest
->shared_blob_set
.add(dest
, sb
);
3324 if (dest
->cache
!= cache
) {
3325 for (auto& i
: sb
->bc
.buffer_map
) {
3326 if (!i
.second
->is_writing()) {
3327 ldout(store
->cct
, 20) << __func__
<< " moving " << *i
.second
3329 dest
->cache
->_move_buffer(cache
, i
.second
.get());
3338 // =======================================================
3340 void *BlueStore::MempoolThread::entry()
3342 Mutex::Locker
l(lock
);
3344 uint64_t meta_bytes
=
3345 mempool::bluestore_cache_other::allocated_bytes() +
3346 mempool::bluestore_cache_onode::allocated_bytes();
3347 uint64_t onode_num
=
3348 mempool::bluestore_cache_onode::allocated_items();
3350 if (onode_num
< 2) {
3354 float bytes_per_onode
= (float)meta_bytes
/ (float)onode_num
;
3355 size_t num_shards
= store
->cache_shards
.size();
3356 float target_ratio
= store
->cache_meta_ratio
+ store
->cache_data_ratio
;
3357 // A little sloppy but should be close enough
3358 uint64_t shard_target
= target_ratio
* (store
->cache_size
/ num_shards
);
3360 for (auto i
: store
->cache_shards
) {
3361 i
->trim(shard_target
,
3362 store
->cache_meta_ratio
,
3363 store
->cache_data_ratio
,
3367 store
->_update_cache_logger();
3370 wait
+= store
->cct
->_conf
->bluestore_cache_trim_interval
;
3371 cond
.WaitInterval(lock
, wait
);
3377 // =======================================================
3382 #define dout_prefix *_dout << "bluestore.OmapIteratorImpl(" << this << ") "
3384 BlueStore::OmapIteratorImpl::OmapIteratorImpl(
3385 CollectionRef c
, OnodeRef o
, KeyValueDB::Iterator it
)
3386 : c(c
), o(o
), it(it
)
3388 RWLock::RLocker
l(c
->lock
);
3389 if (o
->onode
.has_omap()) {
3390 get_omap_key(o
->onode
.nid
, string(), &head
);
3391 get_omap_tail(o
->onode
.nid
, &tail
);
3392 it
->lower_bound(head
);
3396 int BlueStore::OmapIteratorImpl::seek_to_first()
3398 RWLock::RLocker
l(c
->lock
);
3399 if (o
->onode
.has_omap()) {
3400 it
->lower_bound(head
);
3402 it
= KeyValueDB::Iterator();
3407 int BlueStore::OmapIteratorImpl::upper_bound(const string
& after
)
3409 RWLock::RLocker
l(c
->lock
);
3410 if (o
->onode
.has_omap()) {
3412 get_omap_key(o
->onode
.nid
, after
, &key
);
3413 ldout(c
->store
->cct
,20) << __func__
<< " after " << after
<< " key "
3414 << pretty_binary_string(key
) << dendl
;
3415 it
->upper_bound(key
);
3417 it
= KeyValueDB::Iterator();
3422 int BlueStore::OmapIteratorImpl::lower_bound(const string
& to
)
3424 RWLock::RLocker
l(c
->lock
);
3425 if (o
->onode
.has_omap()) {
3427 get_omap_key(o
->onode
.nid
, to
, &key
);
3428 ldout(c
->store
->cct
,20) << __func__
<< " to " << to
<< " key "
3429 << pretty_binary_string(key
) << dendl
;
3430 it
->lower_bound(key
);
3432 it
= KeyValueDB::Iterator();
3437 bool BlueStore::OmapIteratorImpl::valid()
3439 RWLock::RLocker
l(c
->lock
);
3440 bool r
= o
->onode
.has_omap() && it
&& it
->valid() &&
3441 it
->raw_key().second
<= tail
;
3442 if (it
&& it
->valid()) {
3443 ldout(c
->store
->cct
,20) << __func__
<< " is at "
3444 << pretty_binary_string(it
->raw_key().second
)
3450 int BlueStore::OmapIteratorImpl::next(bool validate
)
3452 RWLock::RLocker
l(c
->lock
);
3453 if (o
->onode
.has_omap()) {
3461 string
BlueStore::OmapIteratorImpl::key()
3463 RWLock::RLocker
l(c
->lock
);
3464 assert(it
->valid());
3465 string db_key
= it
->raw_key().second
;
3467 decode_omap_key(db_key
, &user_key
);
3471 bufferlist
BlueStore::OmapIteratorImpl::value()
3473 RWLock::RLocker
l(c
->lock
);
3474 assert(it
->valid());
3479 // =====================================
3482 #define dout_prefix *_dout << "bluestore(" << path << ") "
3485 static void aio_cb(void *priv
, void *priv2
)
3487 BlueStore
*store
= static_cast<BlueStore
*>(priv
);
3488 BlueStore::AioContext
*c
= static_cast<BlueStore::AioContext
*>(priv2
);
3489 c
->aio_finish(store
);
3492 BlueStore::BlueStore(CephContext
*cct
, const string
& path
)
3493 : ObjectStore(cct
, path
),
3494 throttle_bytes(cct
, "bluestore_throttle_bytes",
3495 cct
->_conf
->bluestore_throttle_bytes
),
3496 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3497 cct
->_conf
->bluestore_throttle_bytes
+
3498 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3499 deferred_finisher(cct
, "defered_finisher", "dfin"),
3500 kv_sync_thread(this),
3501 kv_finalize_thread(this),
3502 mempool_thread(this)
3505 cct
->_conf
->add_observer(this);
3506 set_cache_shards(1);
3509 BlueStore::BlueStore(CephContext
*cct
,
3511 uint64_t _min_alloc_size
)
3512 : ObjectStore(cct
, path
),
3513 throttle_bytes(cct
, "bluestore_throttle_bytes",
3514 cct
->_conf
->bluestore_throttle_bytes
),
3515 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3516 cct
->_conf
->bluestore_throttle_bytes
+
3517 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3518 deferred_finisher(cct
, "defered_finisher", "dfin"),
3519 kv_sync_thread(this),
3520 kv_finalize_thread(this),
3521 min_alloc_size(_min_alloc_size
),
3522 min_alloc_size_order(ctz(_min_alloc_size
)),
3523 mempool_thread(this)
3526 cct
->_conf
->add_observer(this);
3527 set_cache_shards(1);
3530 BlueStore::~BlueStore()
3532 for (auto f
: finishers
) {
3537 cct
->_conf
->remove_observer(this);
3541 assert(bluefs
== NULL
);
3542 assert(fsid_fd
< 0);
3543 assert(path_fd
< 0);
3544 for (auto i
: cache_shards
) {
3547 cache_shards
.clear();
3550 const char **BlueStore::get_tracked_conf_keys() const
3552 static const char* KEYS
[] = {
3553 "bluestore_csum_type",
3554 "bluestore_compression_mode",
3555 "bluestore_compression_algorithm",
3556 "bluestore_compression_min_blob_size",
3557 "bluestore_compression_min_blob_size_ssd",
3558 "bluestore_compression_min_blob_size_hdd",
3559 "bluestore_compression_max_blob_size",
3560 "bluestore_compression_max_blob_size_ssd",
3561 "bluestore_compression_max_blob_size_hdd",
3562 "bluestore_compression_required_ratio",
3563 "bluestore_max_alloc_size",
3564 "bluestore_prefer_deferred_size",
3565 "bluestore_prefer_deferred_size_hdd",
3566 "bluestore_prefer_deferred_size_ssd",
3567 "bluestore_deferred_batch_ops",
3568 "bluestore_deferred_batch_ops_hdd",
3569 "bluestore_deferred_batch_ops_ssd",
3570 "bluestore_throttle_bytes",
3571 "bluestore_throttle_deferred_bytes",
3572 "bluestore_throttle_cost_per_io_hdd",
3573 "bluestore_throttle_cost_per_io_ssd",
3574 "bluestore_throttle_cost_per_io",
3575 "bluestore_max_blob_size",
3576 "bluestore_max_blob_size_ssd",
3577 "bluestore_max_blob_size_hdd",
3583 void BlueStore::handle_conf_change(const struct md_config_t
*conf
,
3584 const std::set
<std::string
> &changed
)
3586 if (changed
.count("bluestore_csum_type")) {
3589 if (changed
.count("bluestore_compression_mode") ||
3590 changed
.count("bluestore_compression_algorithm") ||
3591 changed
.count("bluestore_compression_min_blob_size") ||
3592 changed
.count("bluestore_compression_max_blob_size")) {
3597 if (changed
.count("bluestore_max_blob_size") ||
3598 changed
.count("bluestore_max_blob_size_ssd") ||
3599 changed
.count("bluestore_max_blob_size_hdd")) {
3601 // only after startup
3605 if (changed
.count("bluestore_prefer_deferred_size") ||
3606 changed
.count("bluestore_prefer_deferred_size_hdd") ||
3607 changed
.count("bluestore_prefer_deferred_size_ssd") ||
3608 changed
.count("bluestore_max_alloc_size") ||
3609 changed
.count("bluestore_deferred_batch_ops") ||
3610 changed
.count("bluestore_deferred_batch_ops_hdd") ||
3611 changed
.count("bluestore_deferred_batch_ops_ssd")) {
3613 // only after startup
3617 if (changed
.count("bluestore_throttle_cost_per_io") ||
3618 changed
.count("bluestore_throttle_cost_per_io_hdd") ||
3619 changed
.count("bluestore_throttle_cost_per_io_ssd")) {
3621 _set_throttle_params();
3624 if (changed
.count("bluestore_throttle_bytes")) {
3625 throttle_bytes
.reset_max(conf
->bluestore_throttle_bytes
);
3626 throttle_deferred_bytes
.reset_max(
3627 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3629 if (changed
.count("bluestore_throttle_deferred_bytes")) {
3630 throttle_deferred_bytes
.reset_max(
3631 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3635 void BlueStore::_set_compression()
3637 auto m
= Compressor::get_comp_mode_type(cct
->_conf
->bluestore_compression_mode
);
3641 derr
<< __func__
<< " unrecognized value '"
3642 << cct
->_conf
->bluestore_compression_mode
3643 << "' for bluestore_compression_mode, reverting to 'none'"
3645 comp_mode
= Compressor::COMP_NONE
;
3648 compressor
= nullptr;
3650 if (comp_mode
== Compressor::COMP_NONE
) {
3651 dout(10) << __func__
<< " compression mode set to 'none', "
3652 << "ignore other compression setttings" << dendl
;
3656 if (cct
->_conf
->bluestore_compression_min_blob_size
) {
3657 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size
;
3660 if (bdev
->is_rotational()) {
3661 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_hdd
;
3663 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_ssd
;
3667 if (cct
->_conf
->bluestore_compression_max_blob_size
) {
3668 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size
;
3671 if (bdev
->is_rotational()) {
3672 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_hdd
;
3674 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_ssd
;
3678 auto& alg_name
= cct
->_conf
->bluestore_compression_algorithm
;
3679 if (!alg_name
.empty()) {
3680 compressor
= Compressor::create(cct
, alg_name
);
3682 derr
<< __func__
<< " unable to initialize " << alg_name
.c_str() << " compressor"
3687 dout(10) << __func__
<< " mode " << Compressor::get_comp_mode_name(comp_mode
)
3688 << " alg " << (compressor
? compressor
->get_type_name() : "(none)")
3692 void BlueStore::_set_csum()
3694 csum_type
= Checksummer::CSUM_NONE
;
3695 int t
= Checksummer::get_csum_string_type(cct
->_conf
->bluestore_csum_type
);
3696 if (t
> Checksummer::CSUM_NONE
)
3699 dout(10) << __func__
<< " csum_type "
3700 << Checksummer::get_csum_type_string(csum_type
)
3704 void BlueStore::_set_throttle_params()
3706 if (cct
->_conf
->bluestore_throttle_cost_per_io
) {
3707 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io
;
3710 if (bdev
->is_rotational()) {
3711 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_hdd
;
3713 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_ssd
;
3717 dout(10) << __func__
<< " throttle_cost_per_io " << throttle_cost_per_io
3720 void BlueStore::_set_blob_size()
3722 if (cct
->_conf
->bluestore_max_blob_size
) {
3723 max_blob_size
= cct
->_conf
->bluestore_max_blob_size
;
3726 if (bdev
->is_rotational()) {
3727 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_hdd
;
3729 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_ssd
;
3732 dout(10) << __func__
<< " max_blob_size 0x" << std::hex
<< max_blob_size
3733 << std::dec
<< dendl
;
3736 void BlueStore::_set_finisher_num()
3738 if (cct
->_conf
->bluestore_shard_finishers
) {
3739 if (cct
->_conf
->osd_op_num_shards
) {
3740 m_finisher_num
= cct
->_conf
->osd_op_num_shards
;
3743 if (bdev
->is_rotational()) {
3744 m_finisher_num
= cct
->_conf
->osd_op_num_shards_hdd
;
3746 m_finisher_num
= cct
->_conf
->osd_op_num_shards_ssd
;
3750 assert(m_finisher_num
!= 0);
3753 int BlueStore::_set_cache_sizes()
3756 if (cct
->_conf
->bluestore_cache_size
) {
3757 cache_size
= cct
->_conf
->bluestore_cache_size
;
3759 // choose global cache size based on backend type
3760 if (bdev
->is_rotational()) {
3761 cache_size
= cct
->_conf
->bluestore_cache_size_hdd
;
3763 cache_size
= cct
->_conf
->bluestore_cache_size_ssd
;
3766 cache_meta_ratio
= cct
->_conf
->bluestore_cache_meta_ratio
;
3767 cache_kv_ratio
= cct
->_conf
->bluestore_cache_kv_ratio
;
3769 double cache_kv_max
= cct
->_conf
->bluestore_cache_kv_max
;
3770 double cache_kv_max_ratio
= 0;
3772 // if cache_kv_max is negative, disable it
3773 if (cache_size
> 0 && cache_kv_max
>= 0) {
3774 cache_kv_max_ratio
= (double) cache_kv_max
/ (double) cache_size
;
3775 if (cache_kv_max_ratio
< 1.0 && cache_kv_max_ratio
< cache_kv_ratio
) {
3776 dout(1) << __func__
<< " max " << cache_kv_max_ratio
3777 << " < ratio " << cache_kv_ratio
3779 cache_meta_ratio
= cache_meta_ratio
+ cache_kv_ratio
- cache_kv_max_ratio
;
3780 cache_kv_ratio
= cache_kv_max_ratio
;
3785 (double)1.0 - (double)cache_meta_ratio
- (double)cache_kv_ratio
;
3787 if (cache_meta_ratio
< 0 || cache_meta_ratio
> 1.0) {
3788 derr
<< __func__
<< " bluestore_cache_meta_ratio (" << cache_meta_ratio
3789 << ") must be in range [0,1.0]" << dendl
;
3792 if (cache_kv_ratio
< 0 || cache_kv_ratio
> 1.0) {
3793 derr
<< __func__
<< " bluestore_cache_kv_ratio (" << cache_kv_ratio
3794 << ") must be in range [0,1.0]" << dendl
;
3797 if (cache_meta_ratio
+ cache_kv_ratio
> 1.0) {
3798 derr
<< __func__
<< " bluestore_cache_meta_ratio (" << cache_meta_ratio
3799 << ") + bluestore_cache_kv_ratio (" << cache_kv_ratio
3800 << ") = " << cache_meta_ratio
+ cache_kv_ratio
<< "; must be <= 1.0"
3804 if (cache_data_ratio
< 0) {
3805 // deal with floating point imprecision
3806 cache_data_ratio
= 0;
3808 dout(1) << __func__
<< " cache_size " << cache_size
3809 << " meta " << cache_meta_ratio
3810 << " kv " << cache_kv_ratio
3811 << " data " << cache_data_ratio
3816 int BlueStore::write_meta(const std::string
& key
, const std::string
& value
)
3818 bluestore_bdev_label_t label
;
3819 string p
= path
+ "/block";
3820 int r
= _read_bdev_label(cct
, p
, &label
);
3822 return ObjectStore::write_meta(key
, value
);
3824 label
.meta
[key
] = value
;
3825 r
= _write_bdev_label(cct
, p
, label
);
3827 return ObjectStore::write_meta(key
, value
);
3830 int BlueStore::read_meta(const std::string
& key
, std::string
*value
)
3832 bluestore_bdev_label_t label
;
3833 string p
= path
+ "/block";
3834 int r
= _read_bdev_label(cct
, p
, &label
);
3836 return ObjectStore::read_meta(key
, value
);
3838 auto i
= label
.meta
.find(key
);
3839 if (i
== label
.meta
.end()) {
3840 return ObjectStore::read_meta(key
, value
);
3846 void BlueStore::_init_logger()
3848 PerfCountersBuilder
b(cct
, "bluestore",
3849 l_bluestore_first
, l_bluestore_last
);
3850 b
.add_time_avg(l_bluestore_kv_flush_lat
, "kv_flush_lat",
3851 "Average kv_thread flush latency",
3852 "fl_l", PerfCountersBuilder::PRIO_INTERESTING
);
3853 b
.add_time_avg(l_bluestore_kv_commit_lat
, "kv_commit_lat",
3854 "Average kv_thread commit latency");
3855 b
.add_time_avg(l_bluestore_kv_lat
, "kv_lat",
3856 "Average kv_thread sync latency",
3857 "k_l", PerfCountersBuilder::PRIO_INTERESTING
);
3858 b
.add_time_avg(l_bluestore_state_prepare_lat
, "state_prepare_lat",
3859 "Average prepare state latency");
3860 b
.add_time_avg(l_bluestore_state_aio_wait_lat
, "state_aio_wait_lat",
3861 "Average aio_wait state latency",
3862 "io_l", PerfCountersBuilder::PRIO_INTERESTING
);
3863 b
.add_time_avg(l_bluestore_state_io_done_lat
, "state_io_done_lat",
3864 "Average io_done state latency");
3865 b
.add_time_avg(l_bluestore_state_kv_queued_lat
, "state_kv_queued_lat",
3866 "Average kv_queued state latency");
3867 b
.add_time_avg(l_bluestore_state_kv_committing_lat
, "state_kv_commiting_lat",
3868 "Average kv_commiting state latency");
3869 b
.add_time_avg(l_bluestore_state_kv_done_lat
, "state_kv_done_lat",
3870 "Average kv_done state latency");
3871 b
.add_time_avg(l_bluestore_state_deferred_queued_lat
, "state_deferred_queued_lat",
3872 "Average deferred_queued state latency");
3873 b
.add_time_avg(l_bluestore_state_deferred_aio_wait_lat
, "state_deferred_aio_wait_lat",
3874 "Average aio_wait state latency");
3875 b
.add_time_avg(l_bluestore_state_deferred_cleanup_lat
, "state_deferred_cleanup_lat",
3876 "Average cleanup state latency");
3877 b
.add_time_avg(l_bluestore_state_finishing_lat
, "state_finishing_lat",
3878 "Average finishing state latency");
3879 b
.add_time_avg(l_bluestore_state_done_lat
, "state_done_lat",
3880 "Average done state latency");
3881 b
.add_time_avg(l_bluestore_throttle_lat
, "throttle_lat",
3882 "Average submit throttle latency",
3883 "th_l", PerfCountersBuilder::PRIO_CRITICAL
);
3884 b
.add_time_avg(l_bluestore_submit_lat
, "submit_lat",
3885 "Average submit latency",
3886 "s_l", PerfCountersBuilder::PRIO_CRITICAL
);
3887 b
.add_time_avg(l_bluestore_commit_lat
, "commit_lat",
3888 "Average commit latency",
3889 "c_l", PerfCountersBuilder::PRIO_CRITICAL
);
3890 b
.add_time_avg(l_bluestore_read_lat
, "read_lat",
3891 "Average read latency",
3892 "r_l", PerfCountersBuilder::PRIO_CRITICAL
);
3893 b
.add_time_avg(l_bluestore_read_onode_meta_lat
, "read_onode_meta_lat",
3894 "Average read onode metadata latency");
3895 b
.add_time_avg(l_bluestore_read_wait_aio_lat
, "read_wait_aio_lat",
3896 "Average read latency");
3897 b
.add_time_avg(l_bluestore_compress_lat
, "compress_lat",
3898 "Average compress latency");
3899 b
.add_time_avg(l_bluestore_decompress_lat
, "decompress_lat",
3900 "Average decompress latency");
3901 b
.add_time_avg(l_bluestore_csum_lat
, "csum_lat",
3902 "Average checksum latency");
3903 b
.add_u64_counter(l_bluestore_compress_success_count
, "compress_success_count",
3904 "Sum for beneficial compress ops");
3905 b
.add_u64_counter(l_bluestore_compress_rejected_count
, "compress_rejected_count",
3906 "Sum for compress ops rejected due to low net gain of space");
3907 b
.add_u64_counter(l_bluestore_write_pad_bytes
, "write_pad_bytes",
3908 "Sum for write-op padded bytes", NULL
, 0, unit_t(BYTES
));
3909 b
.add_u64_counter(l_bluestore_deferred_write_ops
, "deferred_write_ops",
3910 "Sum for deferred write op");
3911 b
.add_u64_counter(l_bluestore_deferred_write_bytes
, "deferred_write_bytes",
3912 "Sum for deferred write bytes", "def", 0, unit_t(BYTES
));
3913 b
.add_u64_counter(l_bluestore_write_penalty_read_ops
, "write_penalty_read_ops",
3914 "Sum for write penalty read ops");
3915 b
.add_u64(l_bluestore_allocated
, "bluestore_allocated",
3916 "Sum for allocated bytes");
3917 b
.add_u64(l_bluestore_stored
, "bluestore_stored",
3918 "Sum for stored bytes");
3919 b
.add_u64(l_bluestore_compressed
, "bluestore_compressed",
3920 "Sum for stored compressed bytes");
3921 b
.add_u64(l_bluestore_compressed_allocated
, "bluestore_compressed_allocated",
3922 "Sum for bytes allocated for compressed data");
3923 b
.add_u64(l_bluestore_compressed_original
, "bluestore_compressed_original",
3924 "Sum for original bytes that were compressed");
3926 b
.add_u64(l_bluestore_onodes
, "bluestore_onodes",
3927 "Number of onodes in cache");
3928 b
.add_u64_counter(l_bluestore_onode_hits
, "bluestore_onode_hits",
3929 "Sum for onode-lookups hit in the cache");
3930 b
.add_u64_counter(l_bluestore_onode_misses
, "bluestore_onode_misses",
3931 "Sum for onode-lookups missed in the cache");
3932 b
.add_u64_counter(l_bluestore_onode_shard_hits
, "bluestore_onode_shard_hits",
3933 "Sum for onode-shard lookups hit in the cache");
3934 b
.add_u64_counter(l_bluestore_onode_shard_misses
,
3935 "bluestore_onode_shard_misses",
3936 "Sum for onode-shard lookups missed in the cache");
3937 b
.add_u64(l_bluestore_extents
, "bluestore_extents",
3938 "Number of extents in cache");
3939 b
.add_u64(l_bluestore_blobs
, "bluestore_blobs",
3940 "Number of blobs in cache");
3941 b
.add_u64(l_bluestore_buffers
, "bluestore_buffers",
3942 "Number of buffers in cache");
3943 b
.add_u64(l_bluestore_buffer_bytes
, "bluestore_buffer_bytes",
3944 "Number of buffer bytes in cache", NULL
, 0, unit_t(BYTES
));
3945 b
.add_u64(l_bluestore_buffer_hit_bytes
, "bluestore_buffer_hit_bytes",
3946 "Sum for bytes of read hit in the cache", NULL
, 0, unit_t(BYTES
));
3947 b
.add_u64(l_bluestore_buffer_miss_bytes
, "bluestore_buffer_miss_bytes",
3948 "Sum for bytes of read missed in the cache", NULL
, 0, unit_t(BYTES
));
3950 b
.add_u64_counter(l_bluestore_write_big
, "bluestore_write_big",
3951 "Large aligned writes into fresh blobs");
3952 b
.add_u64_counter(l_bluestore_write_big_bytes
, "bluestore_write_big_bytes",
3953 "Large aligned writes into fresh blobs (bytes)", NULL
, 0, unit_t(BYTES
));
3954 b
.add_u64_counter(l_bluestore_write_big_blobs
, "bluestore_write_big_blobs",
3955 "Large aligned writes into fresh blobs (blobs)");
3956 b
.add_u64_counter(l_bluestore_write_small
, "bluestore_write_small",
3957 "Small writes into existing or sparse small blobs");
3958 b
.add_u64_counter(l_bluestore_write_small_bytes
, "bluestore_write_small_bytes",
3959 "Small writes into existing or sparse small blobs (bytes)", NULL
, 0, unit_t(BYTES
));
3960 b
.add_u64_counter(l_bluestore_write_small_unused
,
3961 "bluestore_write_small_unused",
3962 "Small writes into unused portion of existing blob");
3963 b
.add_u64_counter(l_bluestore_write_small_deferred
,
3964 "bluestore_write_small_deferred",
3965 "Small overwrites using deferred");
3966 b
.add_u64_counter(l_bluestore_write_small_pre_read
,
3967 "bluestore_write_small_pre_read",
3968 "Small writes that required we read some data (possibly "
3969 "cached) to fill out the block");
3970 b
.add_u64_counter(l_bluestore_write_small_new
, "bluestore_write_small_new",
3971 "Small write into new (sparse) blob");
3973 b
.add_u64_counter(l_bluestore_txc
, "bluestore_txc", "Transactions committed");
3974 b
.add_u64_counter(l_bluestore_onode_reshard
, "bluestore_onode_reshard",
3975 "Onode extent map reshard events");
3976 b
.add_u64_counter(l_bluestore_blob_split
, "bluestore_blob_split",
3977 "Sum for blob splitting due to resharding");
3978 b
.add_u64_counter(l_bluestore_extent_compress
, "bluestore_extent_compress",
3979 "Sum for extents that have been removed due to compression");
3980 b
.add_u64_counter(l_bluestore_gc_merged
, "bluestore_gc_merged",
3981 "Sum for extents that have been merged due to garbage "
3983 b
.add_u64_counter(l_bluestore_read_eio
, "bluestore_read_eio",
3984 "Read EIO errors propagated to high level callers");
3985 logger
= b
.create_perf_counters();
3986 cct
->get_perfcounters_collection()->add(logger
);
3989 int BlueStore::_reload_logger()
3991 struct store_statfs_t store_statfs
;
3993 int r
= statfs(&store_statfs
);
3995 logger
->set(l_bluestore_allocated
, store_statfs
.allocated
);
3996 logger
->set(l_bluestore_stored
, store_statfs
.stored
);
3997 logger
->set(l_bluestore_compressed
, store_statfs
.compressed
);
3998 logger
->set(l_bluestore_compressed_allocated
, store_statfs
.compressed_allocated
);
3999 logger
->set(l_bluestore_compressed_original
, store_statfs
.compressed_original
);
4004 void BlueStore::_shutdown_logger()
4006 cct
->get_perfcounters_collection()->remove(logger
);
4010 int BlueStore::get_block_device_fsid(CephContext
* cct
, const string
& path
,
4013 bluestore_bdev_label_t label
;
4014 int r
= _read_bdev_label(cct
, path
, &label
);
4017 *fsid
= label
.osd_uuid
;
4021 int BlueStore::_open_path()
4024 if (cct
->_conf
->get_val
<uint64_t>("osd_max_object_size") >=
4025 4*1024*1024*1024ull) {
4026 derr
<< __func__
<< " osd_max_object_size >= 4GB; BlueStore has hard limit of 4GB." << dendl
;
4029 assert(path_fd
< 0);
4030 path_fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_DIRECTORY
));
4033 derr
<< __func__
<< " unable to open " << path
<< ": " << cpp_strerror(r
)
4040 void BlueStore::_close_path()
4042 VOID_TEMP_FAILURE_RETRY(::close(path_fd
));
4046 int BlueStore::_write_bdev_label(CephContext
*cct
,
4047 string path
, bluestore_bdev_label_t label
)
4049 dout(10) << __func__
<< " path " << path
<< " label " << label
<< dendl
;
4051 ::encode(label
, bl
);
4052 uint32_t crc
= bl
.crc32c(-1);
4054 assert(bl
.length() <= BDEV_LABEL_BLOCK_SIZE
);
4055 bufferptr
z(BDEV_LABEL_BLOCK_SIZE
- bl
.length());
4057 bl
.append(std::move(z
));
4059 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_WRONLY
));
4062 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
4066 int r
= bl
.write_fd(fd
);
4068 derr
<< __func__
<< " failed to write to " << path
4069 << ": " << cpp_strerror(r
) << dendl
;
4073 derr
<< __func__
<< " failed to fsync " << path
4074 << ": " << cpp_strerror(r
) << dendl
;
4076 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4080 int BlueStore::_read_bdev_label(CephContext
* cct
, string path
,
4081 bluestore_bdev_label_t
*label
)
4083 dout(10) << __func__
<< dendl
;
4084 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_RDONLY
));
4087 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
4092 int r
= bl
.read_fd(fd
, BDEV_LABEL_BLOCK_SIZE
);
4093 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4095 derr
<< __func__
<< " failed to read from " << path
4096 << ": " << cpp_strerror(r
) << dendl
;
4100 uint32_t crc
, expected_crc
;
4101 bufferlist::iterator p
= bl
.begin();
4103 ::decode(*label
, p
);
4105 t
.substr_of(bl
, 0, p
.get_off());
4107 ::decode(expected_crc
, p
);
4109 catch (buffer::error
& e
) {
4110 dout(2) << __func__
<< " unable to decode label at offset " << p
.get_off()
4115 if (crc
!= expected_crc
) {
4116 derr
<< __func__
<< " bad crc on label, expected " << expected_crc
4117 << " != actual " << crc
<< dendl
;
4120 dout(10) << __func__
<< " got " << *label
<< dendl
;
4124 int BlueStore::_check_or_set_bdev_label(
4125 string path
, uint64_t size
, string desc
, bool create
)
4127 bluestore_bdev_label_t label
;
4129 label
.osd_uuid
= fsid
;
4131 label
.btime
= ceph_clock_now();
4132 label
.description
= desc
;
4133 int r
= _write_bdev_label(cct
, path
, label
);
4137 int r
= _read_bdev_label(cct
, path
, &label
);
4140 if (cct
->_conf
->bluestore_debug_permit_any_bdev_label
) {
4141 dout(20) << __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4142 << " and fsid " << fsid
<< " check bypassed" << dendl
;
4144 else if (label
.osd_uuid
!= fsid
) {
4145 derr
<< __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4146 << " does not match our fsid " << fsid
<< dendl
;
4153 void BlueStore::_set_alloc_sizes(void)
4155 max_alloc_size
= cct
->_conf
->bluestore_max_alloc_size
;
4157 if (cct
->_conf
->bluestore_prefer_deferred_size
) {
4158 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size
;
4161 if (bdev
->is_rotational()) {
4162 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_hdd
;
4164 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_ssd
;
4168 if (cct
->_conf
->bluestore_deferred_batch_ops
) {
4169 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops
;
4172 if (bdev
->is_rotational()) {
4173 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_hdd
;
4175 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_ssd
;
4179 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
4180 << std::dec
<< " order " << min_alloc_size_order
4181 << " max_alloc_size 0x" << std::hex
<< max_alloc_size
4182 << " prefer_deferred_size 0x" << prefer_deferred_size
4184 << " deferred_batch_ops " << deferred_batch_ops
4188 int BlueStore::_open_bdev(bool create
)
4190 assert(bdev
== NULL
);
4191 string p
= path
+ "/block";
4192 bdev
= BlockDevice::create(cct
, p
, aio_cb
, static_cast<void*>(this));
4193 int r
= bdev
->open(p
);
4197 if (bdev
->supported_bdev_label()) {
4198 r
= _check_or_set_bdev_label(p
, bdev
->get_size(), "main", create
);
4203 // initialize global block parameters
4204 block_size
= bdev
->get_block_size();
4205 block_mask
= ~(block_size
- 1);
4206 block_size_order
= ctz(block_size
);
4207 assert(block_size
== 1u << block_size_order
);
4208 // and set cache_size based on device type
4209 r
= _set_cache_sizes();
4223 void BlueStore::_close_bdev()
4231 int BlueStore::_open_fm(bool create
)
4234 fm
= FreelistManager::create(cct
, freelist_type
, db
, PREFIX_ALLOC
);
4237 // initialize freespace
4238 dout(20) << __func__
<< " initializing freespace" << dendl
;
4239 KeyValueDB::Transaction t
= db
->get_transaction();
4242 bl
.append(freelist_type
);
4243 t
->set(PREFIX_SUPER
, "freelist_type", bl
);
4245 // being able to allocate in units less than bdev block size
4246 // seems to be a bad idea.
4247 assert( cct
->_conf
->bdev_block_size
<= (int64_t)min_alloc_size
);
4248 fm
->create(bdev
->get_size(), (int64_t)min_alloc_size
, t
);
4250 // allocate superblock reserved space. note that we do not mark
4251 // bluefs space as allocated in the freelist; we instead rely on
4253 uint64_t reserved
= ROUND_UP_TO(MAX(SUPER_RESERVED
, min_alloc_size
),
4255 fm
->allocate(0, reserved
, t
);
4257 if (cct
->_conf
->bluestore_bluefs
) {
4258 assert(bluefs_extents
.num_intervals() == 1);
4259 interval_set
<uint64_t>::iterator p
= bluefs_extents
.begin();
4260 reserved
= ROUND_UP_TO(p
.get_start() + p
.get_len(), min_alloc_size
);
4261 dout(20) << __func__
<< " reserved 0x" << std::hex
<< reserved
<< std::dec
4262 << " for bluefs" << dendl
;
4264 ::encode(bluefs_extents
, bl
);
4265 t
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
4266 dout(20) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
4267 << std::dec
<< dendl
;
4270 if (cct
->_conf
->bluestore_debug_prefill
> 0) {
4271 uint64_t end
= bdev
->get_size() - reserved
;
4272 dout(1) << __func__
<< " pre-fragmenting freespace, using "
4273 << cct
->_conf
->bluestore_debug_prefill
<< " with max free extent "
4274 << cct
->_conf
->bluestore_debug_prefragment_max
<< dendl
;
4275 uint64_t start
= P2ROUNDUP(reserved
, min_alloc_size
);
4276 uint64_t max_b
= cct
->_conf
->bluestore_debug_prefragment_max
/ min_alloc_size
;
4277 float r
= cct
->_conf
->bluestore_debug_prefill
;
4281 while (!stop
&& start
< end
) {
4282 uint64_t l
= (rand() % max_b
+ 1) * min_alloc_size
;
4283 if (start
+ l
> end
) {
4285 l
= P2ALIGN(l
, min_alloc_size
);
4287 assert(start
+ l
<= end
);
4289 uint64_t u
= 1 + (uint64_t)(r
* (double)l
);
4290 u
= P2ROUNDUP(u
, min_alloc_size
);
4291 if (start
+ l
+ u
> end
) {
4292 u
= end
- (start
+ l
);
4293 // trim to align so we don't overflow again
4294 u
= P2ALIGN(u
, min_alloc_size
);
4297 assert(start
+ l
+ u
<= end
);
4299 dout(20) << " free 0x" << std::hex
<< start
<< "~" << l
4300 << " use 0x" << u
<< std::dec
<< dendl
;
4303 // break if u has been trimmed to nothing
4307 fm
->allocate(start
+ l
, u
, t
);
4311 db
->submit_transaction_sync(t
);
4314 int r
= fm
->init(bdev
->get_size());
4316 derr
<< __func__
<< " freelist init failed: " << cpp_strerror(r
) << dendl
;
4324 void BlueStore::_close_fm()
4326 dout(10) << __func__
<< dendl
;
4333 int BlueStore::_open_alloc()
4335 assert(alloc
== NULL
);
4336 assert(bdev
->get_size());
4337 alloc
= Allocator::create(cct
, cct
->_conf
->bluestore_allocator
,
4341 lderr(cct
) << __func__
<< " Allocator::unknown alloc type "
4342 << cct
->_conf
->bluestore_allocator
4347 uint64_t num
= 0, bytes
= 0;
4349 dout(1) << __func__
<< " opening allocation metadata" << dendl
;
4350 // initialize from freelist
4351 fm
->enumerate_reset();
4352 uint64_t offset
, length
;
4353 while (fm
->enumerate_next(&offset
, &length
)) {
4354 alloc
->init_add_free(offset
, length
);
4358 fm
->enumerate_reset();
4359 dout(1) << __func__
<< " loaded " << byte_u_t(bytes
)
4360 << " in " << num
<< " extents"
4363 // also mark bluefs space as allocated
4364 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
4365 alloc
->init_rm_free(e
.get_start(), e
.get_len());
4367 dout(10) << __func__
<< " marked bluefs_extents 0x" << std::hex
4368 << bluefs_extents
<< std::dec
<< " as allocated" << dendl
;
4373 void BlueStore::_close_alloc()
4381 int BlueStore::_open_fsid(bool create
)
4383 assert(fsid_fd
< 0);
4387 fsid_fd
= ::openat(path_fd
, "fsid", flags
, 0644);
4390 derr
<< __func__
<< " " << cpp_strerror(err
) << dendl
;
4396 int BlueStore::_read_fsid(uuid_d
*uuid
)
4399 memset(fsid_str
, 0, sizeof(fsid_str
));
4400 int ret
= safe_read(fsid_fd
, fsid_str
, sizeof(fsid_str
));
4402 derr
<< __func__
<< " failed: " << cpp_strerror(ret
) << dendl
;
4409 if (!uuid
->parse(fsid_str
)) {
4410 derr
<< __func__
<< " unparsable uuid " << fsid_str
<< dendl
;
4416 int BlueStore::_write_fsid()
4418 int r
= ::ftruncate(fsid_fd
, 0);
4421 derr
<< __func__
<< " fsid truncate failed: " << cpp_strerror(r
) << dendl
;
4424 string str
= stringify(fsid
) + "\n";
4425 r
= safe_write(fsid_fd
, str
.c_str(), str
.length());
4427 derr
<< __func__
<< " fsid write failed: " << cpp_strerror(r
) << dendl
;
4430 r
= ::fsync(fsid_fd
);
4433 derr
<< __func__
<< " fsid fsync failed: " << cpp_strerror(r
) << dendl
;
4439 void BlueStore::_close_fsid()
4441 VOID_TEMP_FAILURE_RETRY(::close(fsid_fd
));
4445 int BlueStore::_lock_fsid()
4448 memset(&l
, 0, sizeof(l
));
4450 l
.l_whence
= SEEK_SET
;
4451 int r
= ::fcntl(fsid_fd
, F_SETLK
, &l
);
4454 derr
<< __func__
<< " failed to lock " << path
<< "/fsid"
4455 << " (is another ceph-osd still running?)"
4456 << cpp_strerror(err
) << dendl
;
4462 bool BlueStore::is_rotational()
4465 return bdev
->is_rotational();
4468 bool rotational
= true;
4469 int r
= _open_path();
4472 r
= _open_fsid(false);
4475 r
= _read_fsid(&fsid
);
4481 r
= _open_bdev(false);
4484 rotational
= bdev
->is_rotational();
4494 bool BlueStore::is_journal_rotational()
4497 dout(5) << __func__
<< " bluefs disabled, default to store media type"
4499 return is_rotational();
4501 dout(10) << __func__
<< " " << (int)bluefs
->wal_is_rotational() << dendl
;
4502 return bluefs
->wal_is_rotational();
4505 bool BlueStore::test_mount_in_use()
4507 // most error conditions mean the mount is not in use (e.g., because
4508 // it doesn't exist). only if we fail to lock do we conclude it is
4511 int r
= _open_path();
4514 r
= _open_fsid(false);
4519 ret
= true; // if we can't lock, it is in use
4526 int BlueStore::_open_db(bool create
)
4530 string fn
= path
+ "/db";
4533 ceph::shared_ptr
<Int64ArrayMergeOperator
> merge_op(new Int64ArrayMergeOperator
);
4537 kv_backend
= cct
->_conf
->bluestore_kvbackend
;
4539 r
= read_meta("kv_backend", &kv_backend
);
4541 derr
<< __func__
<< " unable to read 'kv_backend' meta" << dendl
;
4545 dout(10) << __func__
<< " kv_backend = " << kv_backend
<< dendl
;
4549 do_bluefs
= cct
->_conf
->bluestore_bluefs
;
4552 r
= read_meta("bluefs", &s
);
4554 derr
<< __func__
<< " unable to read 'bluefs' meta" << dendl
;
4559 } else if (s
== "0") {
4562 derr
<< __func__
<< " bluefs = " << s
<< " : not 0 or 1, aborting"
4567 dout(10) << __func__
<< " do_bluefs = " << do_bluefs
<< dendl
;
4569 rocksdb::Env
*env
= NULL
;
4571 dout(10) << __func__
<< " initializing bluefs" << dendl
;
4572 if (kv_backend
!= "rocksdb") {
4573 derr
<< " backend must be rocksdb to use bluefs" << dendl
;
4576 bluefs
= new BlueFS(cct
);
4581 bfn
= path
+ "/block.db";
4582 if (::stat(bfn
.c_str(), &st
) == 0) {
4583 r
= bluefs
->add_block_device(BlueFS::BDEV_DB
, bfn
);
4585 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4586 << cpp_strerror(r
) << dendl
;
4590 if (bluefs
->bdev_support_label(BlueFS::BDEV_DB
)) {
4591 r
= _check_or_set_bdev_label(
4593 bluefs
->get_block_device_size(BlueFS::BDEV_DB
),
4594 "bluefs db", create
);
4597 << " check block device(" << bfn
<< ") label returned: "
4598 << cpp_strerror(r
) << dendl
;
4603 bluefs
->add_block_extent(
4606 bluefs
->get_block_device_size(BlueFS::BDEV_DB
) - SUPER_RESERVED
);
4608 bluefs_shared_bdev
= BlueFS::BDEV_SLOW
;
4609 bluefs_single_shared_device
= false;
4612 if (::lstat(bfn
.c_str(), &st
) == -1) {
4614 bluefs_shared_bdev
= BlueFS::BDEV_DB
;
4616 derr
<< __func__
<< " " << bfn
<< " symlink exists but target unusable: "
4617 << cpp_strerror(r
) << dendl
;
4623 bfn
= path
+ "/block";
4624 r
= bluefs
->add_block_device(bluefs_shared_bdev
, bfn
);
4626 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4627 << cpp_strerror(r
) << dendl
;
4631 // note: we always leave the first SUPER_RESERVED (8k) of the device unused
4633 bdev
->get_size() * (cct
->_conf
->bluestore_bluefs_min_ratio
+
4634 cct
->_conf
->bluestore_bluefs_gift_ratio
);
4635 initial
= MAX(initial
, cct
->_conf
->bluestore_bluefs_min
);
4636 if (cct
->_conf
->bluefs_alloc_size
% min_alloc_size
) {
4637 derr
<< __func__
<< " bluefs_alloc_size 0x" << std::hex
4638 << cct
->_conf
->bluefs_alloc_size
<< " is not a multiple of "
4639 << "min_alloc_size 0x" << min_alloc_size
<< std::dec
<< dendl
;
4643 // align to bluefs's alloc_size
4644 initial
= P2ROUNDUP(initial
, cct
->_conf
->bluefs_alloc_size
);
4645 // put bluefs in the middle of the device in case it is an HDD
4646 uint64_t start
= P2ALIGN((bdev
->get_size() - initial
) / 2,
4647 cct
->_conf
->bluefs_alloc_size
);
4648 bluefs
->add_block_extent(bluefs_shared_bdev
, start
, initial
);
4649 bluefs_extents
.insert(start
, initial
);
4652 bfn
= path
+ "/block.wal";
4653 if (::stat(bfn
.c_str(), &st
) == 0) {
4654 r
= bluefs
->add_block_device(BlueFS::BDEV_WAL
, bfn
);
4656 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4657 << cpp_strerror(r
) << dendl
;
4661 if (bluefs
->bdev_support_label(BlueFS::BDEV_WAL
)) {
4662 r
= _check_or_set_bdev_label(
4664 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
),
4665 "bluefs wal", create
);
4667 derr
<< __func__
<< " check block device(" << bfn
4668 << ") label returned: " << cpp_strerror(r
) << dendl
;
4674 bluefs
->add_block_extent(
4675 BlueFS::BDEV_WAL
, BDEV_LABEL_BLOCK_SIZE
,
4676 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
) -
4677 BDEV_LABEL_BLOCK_SIZE
);
4679 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "true");
4680 bluefs_single_shared_device
= false;
4683 if (::lstat(bfn
.c_str(), &st
) == -1) {
4685 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "false");
4687 derr
<< __func__
<< " " << bfn
<< " symlink exists but target unusable: "
4688 << cpp_strerror(r
) << dendl
;
4696 r
= bluefs
->mount();
4698 derr
<< __func__
<< " failed bluefs mount: " << cpp_strerror(r
) << dendl
;
4701 if (cct
->_conf
->bluestore_bluefs_env_mirror
) {
4702 rocksdb::Env
*a
= new BlueRocksEnv(bluefs
);
4703 rocksdb::Env
*b
= rocksdb::Env::Default();
4705 string cmd
= "rm -rf " + path
+ "/db " +
4706 path
+ "/db.slow " +
4708 int r
= system(cmd
.c_str());
4711 env
= new rocksdb::EnvMirror(b
, a
, false, true);
4713 env
= new BlueRocksEnv(bluefs
);
4715 // simplify the dir names, too, as "seen" by rocksdb
4719 if (bluefs_shared_bdev
== BlueFS::BDEV_SLOW
) {
4720 // we have both block.db and block; tell rocksdb!
4721 // note: the second (last) size value doesn't really matter
4722 ostringstream db_paths
;
4723 uint64_t db_size
= bluefs
->get_block_device_size(BlueFS::BDEV_DB
);
4724 uint64_t slow_size
= bluefs
->get_block_device_size(BlueFS::BDEV_SLOW
);
4725 db_paths
<< fn
<< ","
4726 << (uint64_t)(db_size
* 95 / 100) << " "
4727 << fn
+ ".slow" << ","
4728 << (uint64_t)(slow_size
* 95 / 100);
4729 cct
->_conf
->set_val("rocksdb_db_paths", db_paths
.str(), false);
4730 dout(10) << __func__
<< " set rocksdb_db_paths to "
4731 << cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths") << dendl
;
4736 if (cct
->_conf
->rocksdb_separate_wal_dir
)
4737 env
->CreateDir(fn
+ ".wal");
4738 if (cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths").length())
4739 env
->CreateDir(fn
+ ".slow");
4741 } else if (create
) {
4742 int r
= ::mkdir(fn
.c_str(), 0755);
4745 if (r
< 0 && r
!= -EEXIST
) {
4746 derr
<< __func__
<< " failed to create " << fn
<< ": " << cpp_strerror(r
)
4752 if (cct
->_conf
->rocksdb_separate_wal_dir
) {
4753 string walfn
= path
+ "/db.wal";
4754 r
= ::mkdir(walfn
.c_str(), 0755);
4757 if (r
< 0 && r
!= -EEXIST
) {
4758 derr
<< __func__
<< " failed to create " << walfn
4759 << ": " << cpp_strerror(r
)
4766 db
= KeyValueDB::create(cct
,
4769 static_cast<void*>(env
));
4771 derr
<< __func__
<< " error creating db" << dendl
;
4777 // delete env manually here since we can't depend on db to do this
4784 FreelistManager::setup_merge_operators(db
);
4785 db
->set_merge_operator(PREFIX_STAT
, merge_op
);
4787 db
->set_cache_size(cache_size
* cache_kv_ratio
);
4789 if (kv_backend
== "rocksdb")
4790 options
= cct
->_conf
->bluestore_rocksdb_options
;
4793 r
= db
->create_and_open(err
);
4797 derr
<< __func__
<< " erroring opening db: " << err
.str() << dendl
;
4807 dout(1) << __func__
<< " opened " << kv_backend
4808 << " path " << fn
<< " options " << options
<< dendl
;
4818 void BlueStore::_close_db()
4830 int BlueStore::_reconcile_bluefs_freespace()
4832 dout(10) << __func__
<< dendl
;
4833 interval_set
<uint64_t> bset
;
4834 int r
= bluefs
->get_block_extents(bluefs_shared_bdev
, &bset
);
4836 if (bset
== bluefs_extents
) {
4837 dout(10) << __func__
<< " we agree bluefs has 0x" << std::hex
<< bset
4838 << std::dec
<< dendl
;
4841 dout(10) << __func__
<< " bluefs says 0x" << std::hex
<< bset
<< std::dec
4843 dout(10) << __func__
<< " super says 0x" << std::hex
<< bluefs_extents
4844 << std::dec
<< dendl
;
4846 interval_set
<uint64_t> overlap
;
4847 overlap
.intersection_of(bset
, bluefs_extents
);
4849 bset
.subtract(overlap
);
4850 if (!bset
.empty()) {
4851 derr
<< __func__
<< " bluefs extra 0x" << std::hex
<< bset
<< std::dec
4856 interval_set
<uint64_t> super_extra
;
4857 super_extra
= bluefs_extents
;
4858 super_extra
.subtract(overlap
);
4859 if (!super_extra
.empty()) {
4860 // This is normal: it can happen if we commit to give extents to
4861 // bluefs and we crash before bluefs commits that it owns them.
4862 dout(10) << __func__
<< " super extra " << super_extra
<< dendl
;
4863 for (interval_set
<uint64_t>::iterator p
= super_extra
.begin();
4864 p
!= super_extra
.end();
4866 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.get_start(), p
.get_len());
4873 int BlueStore::_balance_bluefs_freespace(PExtentVector
*extents
)
4878 vector
<pair
<uint64_t,uint64_t>> bluefs_usage
; // <free, total> ...
4879 bluefs
->get_usage(&bluefs_usage
);
4880 assert(bluefs_usage
.size() > bluefs_shared_bdev
);
4882 // fixme: look at primary bdev only for now
4883 uint64_t bluefs_free
= bluefs_usage
[bluefs_shared_bdev
].first
;
4884 uint64_t bluefs_total
= bluefs_usage
[bluefs_shared_bdev
].second
;
4885 float bluefs_free_ratio
= (float)bluefs_free
/ (float)bluefs_total
;
4887 uint64_t my_free
= alloc
->get_free();
4888 uint64_t total
= bdev
->get_size();
4889 float my_free_ratio
= (float)my_free
/ (float)total
;
4891 uint64_t total_free
= bluefs_free
+ my_free
;
4893 float bluefs_ratio
= (float)bluefs_free
/ (float)total_free
;
4895 dout(10) << __func__
4896 << " bluefs " << byte_u_t(bluefs_free
)
4897 << " free (" << bluefs_free_ratio
4898 << ") bluestore " << byte_u_t(my_free
)
4899 << " free (" << my_free_ratio
4900 << "), bluefs_ratio " << bluefs_ratio
4904 uint64_t reclaim
= 0;
4905 if (bluefs_ratio
< cct
->_conf
->bluestore_bluefs_min_ratio
) {
4906 gift
= cct
->_conf
->bluestore_bluefs_gift_ratio
* total_free
;
4907 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4908 << " < min_ratio " << cct
->_conf
->bluestore_bluefs_min_ratio
4909 << ", should gift " << byte_u_t(gift
) << dendl
;
4910 } else if (bluefs_ratio
> cct
->_conf
->bluestore_bluefs_max_ratio
) {
4911 reclaim
= cct
->_conf
->bluestore_bluefs_reclaim_ratio
* total_free
;
4912 if (bluefs_total
- reclaim
< cct
->_conf
->bluestore_bluefs_min
)
4913 reclaim
= bluefs_total
- cct
->_conf
->bluestore_bluefs_min
;
4914 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4915 << " > max_ratio " << cct
->_conf
->bluestore_bluefs_max_ratio
4916 << ", should reclaim " << byte_u_t(reclaim
) << dendl
;
4919 // don't take over too much of the freespace
4920 uint64_t free_cap
= cct
->_conf
->bluestore_bluefs_max_ratio
* total_free
;
4921 if (bluefs_total
< cct
->_conf
->bluestore_bluefs_min
&&
4922 cct
->_conf
->bluestore_bluefs_min
< free_cap
) {
4923 uint64_t g
= cct
->_conf
->bluestore_bluefs_min
- bluefs_total
;
4924 dout(10) << __func__
<< " bluefs_total " << bluefs_total
4925 << " < min " << cct
->_conf
->bluestore_bluefs_min
4926 << ", should gift " << byte_u_t(g
) << dendl
;
4931 uint64_t min_free
= cct
->_conf
->get_val
<uint64_t>("bluestore_bluefs_min_free");
4932 if (bluefs_free
< min_free
&&
4933 min_free
< free_cap
) {
4934 uint64_t g
= min_free
- bluefs_free
;
4935 dout(10) << __func__
<< " bluefs_free " << bluefs_total
4936 << " < min " << min_free
4937 << ", should gift " << byte_u_t(g
) << dendl
;
4944 // round up to alloc size
4945 gift
= P2ROUNDUP(gift
, cct
->_conf
->bluefs_alloc_size
);
4947 // hard cap to fit into 32 bits
4948 gift
= MIN(gift
, 1ull<<31);
4949 dout(10) << __func__
<< " gifting " << gift
4950 << " (" << byte_u_t(gift
) << ")" << dendl
;
4952 // fixme: just do one allocation to start...
4953 int r
= alloc
->reserve(gift
);
4956 AllocExtentVector exts
;
4957 int64_t alloc_len
= alloc
->allocate(gift
, cct
->_conf
->bluefs_alloc_size
,
4960 if (alloc_len
<= 0) {
4961 dout(1) << __func__
<< " no allocate on 0x" << std::hex
<< gift
4962 << " min_alloc_size 0x" << min_alloc_size
<< std::dec
<< dendl
;
4963 alloc
->unreserve(gift
);
4966 } else if (alloc_len
< (int64_t)gift
) {
4967 dout(1) << __func__
<< " insufficient allocate on 0x" << std::hex
<< gift
4968 << " min_alloc_size 0x" << min_alloc_size
4969 << " allocated 0x" << alloc_len
4970 << std::dec
<< dendl
;
4971 alloc
->unreserve(gift
- alloc_len
);
4974 for (auto& p
: exts
) {
4975 bluestore_pextent_t e
= bluestore_pextent_t(p
);
4976 dout(1) << __func__
<< " gifting " << e
<< " to bluefs" << dendl
;
4977 extents
->push_back(e
);
4984 // reclaim from bluefs?
4986 // round up to alloc size
4987 reclaim
= P2ROUNDUP(reclaim
, cct
->_conf
->bluefs_alloc_size
);
4989 // hard cap to fit into 32 bits
4990 reclaim
= MIN(reclaim
, 1ull<<31);
4991 dout(10) << __func__
<< " reclaiming " << reclaim
4992 << " (" << byte_u_t(reclaim
) << ")" << dendl
;
4994 while (reclaim
> 0) {
4995 // NOTE: this will block and do IO.
4996 AllocExtentVector extents
;
4997 int r
= bluefs
->reclaim_blocks(bluefs_shared_bdev
, reclaim
,
5000 derr
<< __func__
<< " failed to reclaim space from bluefs"
5004 for (auto e
: extents
) {
5005 bluefs_extents
.erase(e
.offset
, e
.length
);
5006 bluefs_extents_reclaiming
.insert(e
.offset
, e
.length
);
5007 reclaim
-= e
.length
;
5017 void BlueStore::_commit_bluefs_freespace(
5018 const PExtentVector
& bluefs_gift_extents
)
5020 dout(10) << __func__
<< dendl
;
5021 for (auto& p
: bluefs_gift_extents
) {
5022 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.offset
, p
.length
);
5026 int BlueStore::_open_collections(int *errors
)
5028 dout(10) << __func__
<< dendl
;
5029 assert(coll_map
.empty());
5030 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_COLL
);
5031 for (it
->upper_bound(string());
5035 if (cid
.parse(it
->key())) {
5039 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
5041 bufferlist bl
= it
->value();
5042 bufferlist::iterator p
= bl
.begin();
5044 ::decode(c
->cnode
, p
);
5045 } catch (buffer::error
& e
) {
5046 derr
<< __func__
<< " failed to decode cnode, key:"
5047 << pretty_binary_string(it
->key()) << dendl
;
5050 dout(20) << __func__
<< " opened " << cid
<< " " << c
5051 << " " << c
->cnode
<< dendl
;
5054 derr
<< __func__
<< " unrecognized collection " << it
->key() << dendl
;
5062 void BlueStore::_open_statfs()
5065 int r
= db
->get(PREFIX_STAT
, "bluestore_statfs", &bl
);
5067 if (size_t(bl
.length()) >= sizeof(vstatfs
.values
)) {
5068 auto it
= bl
.begin();
5071 dout(10) << __func__
<< " store_statfs is corrupt, using empty" << dendl
;
5075 dout(10) << __func__
<< " store_statfs missed, using empty" << dendl
;
5079 int BlueStore::_setup_block_symlink_or_file(
5085 dout(20) << __func__
<< " name " << name
<< " path " << epath
5086 << " size " << size
<< " create=" << (int)create
<< dendl
;
5091 if (epath
.length()) {
5092 r
= ::symlinkat(epath
.c_str(), path_fd
, name
.c_str());
5095 derr
<< __func__
<< " failed to create " << name
<< " symlink to "
5096 << epath
<< ": " << cpp_strerror(r
) << dendl
;
5100 if (!epath
.compare(0, strlen(SPDK_PREFIX
), SPDK_PREFIX
)) {
5101 int fd
= ::openat(path_fd
, epath
.c_str(), flags
, 0644);
5104 derr
<< __func__
<< " failed to open " << epath
<< " file: "
5105 << cpp_strerror(r
) << dendl
;
5108 string serial_number
= epath
.substr(strlen(SPDK_PREFIX
));
5109 r
= ::write(fd
, serial_number
.c_str(), serial_number
.size());
5110 assert(r
== (int)serial_number
.size());
5111 dout(1) << __func__
<< " created " << name
<< " symlink to "
5113 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5117 int fd
= ::openat(path_fd
, name
.c_str(), flags
, 0644);
5119 // block file is present
5121 int r
= ::fstat(fd
, &st
);
5123 S_ISREG(st
.st_mode
) && // if it is a regular file
5124 st
.st_size
== 0) { // and is 0 bytes
5125 r
= ::ftruncate(fd
, size
);
5128 derr
<< __func__
<< " failed to resize " << name
<< " file to "
5129 << size
<< ": " << cpp_strerror(r
) << dendl
;
5130 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5134 if (cct
->_conf
->bluestore_block_preallocate_file
) {
5135 r
= ::ceph_posix_fallocate(fd
, 0, size
);
5137 derr
<< __func__
<< " failed to prefallocate " << name
<< " file to "
5138 << size
<< ": " << cpp_strerror(r
) << dendl
;
5139 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5143 dout(1) << __func__
<< " resized " << name
<< " file to "
5144 << byte_u_t(size
) << dendl
;
5146 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5150 derr
<< __func__
<< " failed to open " << name
<< " file: "
5151 << cpp_strerror(r
) << dendl
;
5159 int BlueStore::mkfs()
5161 dout(1) << __func__
<< " path " << path
<< dendl
;
5167 r
= read_meta("mkfs_done", &done
);
5169 dout(1) << __func__
<< " already created" << dendl
;
5170 if (cct
->_conf
->bluestore_fsck_on_mkfs
) {
5171 r
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5173 derr
<< __func__
<< " fsck found fatal error: " << cpp_strerror(r
)
5178 derr
<< __func__
<< " fsck found " << r
<< " errors" << dendl
;
5182 return r
; // idempotent
5188 r
= read_meta("type", &type
);
5190 if (type
!= "bluestore") {
5191 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5195 r
= write_meta("type", "bluestore");
5201 freelist_type
= "bitmap";
5207 r
= _open_fsid(true);
5213 goto out_close_fsid
;
5215 r
= _read_fsid(&old_fsid
);
5216 if (r
< 0 || old_fsid
.is_zero()) {
5217 if (fsid
.is_zero()) {
5218 fsid
.generate_random();
5219 dout(1) << __func__
<< " generated fsid " << fsid
<< dendl
;
5221 dout(1) << __func__
<< " using provided fsid " << fsid
<< dendl
;
5223 // we'll write it later.
5225 if (!fsid
.is_zero() && fsid
!= old_fsid
) {
5226 derr
<< __func__
<< " on-disk fsid " << old_fsid
5227 << " != provided " << fsid
<< dendl
;
5229 goto out_close_fsid
;
5234 r
= _setup_block_symlink_or_file("block", cct
->_conf
->bluestore_block_path
,
5235 cct
->_conf
->bluestore_block_size
,
5236 cct
->_conf
->bluestore_block_create
);
5238 goto out_close_fsid
;
5239 if (cct
->_conf
->bluestore_bluefs
) {
5240 r
= _setup_block_symlink_or_file("block.wal", cct
->_conf
->bluestore_block_wal_path
,
5241 cct
->_conf
->bluestore_block_wal_size
,
5242 cct
->_conf
->bluestore_block_wal_create
);
5244 goto out_close_fsid
;
5245 r
= _setup_block_symlink_or_file("block.db", cct
->_conf
->bluestore_block_db_path
,
5246 cct
->_conf
->bluestore_block_db_size
,
5247 cct
->_conf
->bluestore_block_db_create
);
5249 goto out_close_fsid
;
5252 r
= _open_bdev(true);
5254 goto out_close_fsid
;
5256 // choose min_alloc_size
5257 if (cct
->_conf
->bluestore_min_alloc_size
) {
5258 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size
;
5261 if (bdev
->is_rotational()) {
5262 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_hdd
;
5264 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_ssd
;
5268 // make sure min_alloc_size is power of 2 aligned.
5269 if (!ISP2(min_alloc_size
)) {
5270 derr
<< __func__
<< " min_alloc_size 0x"
5271 << std::hex
<< min_alloc_size
<< std::dec
5272 << " is not power of 2 aligned!"
5275 goto out_close_bdev
;
5280 goto out_close_bdev
;
5287 KeyValueDB::Transaction t
= db
->get_transaction();
5290 ::encode((uint64_t)0, bl
);
5291 t
->set(PREFIX_SUPER
, "nid_max", bl
);
5292 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
5297 ::encode((uint64_t)min_alloc_size
, bl
);
5298 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
5301 ondisk_format
= latest_ondisk_format
;
5302 _prepare_ondisk_format_super(t
);
5303 db
->submit_transaction_sync(t
);
5307 r
= write_meta("kv_backend", cct
->_conf
->bluestore_kvbackend
);
5311 r
= write_meta("bluefs", stringify(bluefs
? 1 : 0));
5315 if (fsid
!= old_fsid
) {
5318 derr
<< __func__
<< " error writing fsid: " << cpp_strerror(r
) << dendl
;
5335 cct
->_conf
->bluestore_fsck_on_mkfs
) {
5336 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5340 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5346 // indicate success by writing the 'mkfs_done' file
5347 r
= write_meta("mkfs_done", "yes");
5351 derr
<< __func__
<< " failed, " << cpp_strerror(r
) << dendl
;
5353 dout(0) << __func__
<< " success" << dendl
;
5358 void BlueStore::set_cache_shards(unsigned num
)
5360 dout(10) << __func__
<< " " << num
<< dendl
;
5361 size_t old
= cache_shards
.size();
5363 cache_shards
.resize(num
);
5364 for (unsigned i
= old
; i
< num
; ++i
) {
5365 cache_shards
[i
] = Cache::create(cct
, cct
->_conf
->bluestore_cache_type
,
5370 int BlueStore::_mount(bool kv_only
)
5372 dout(1) << __func__
<< " path " << path
<< dendl
;
5378 int r
= read_meta("type", &type
);
5380 derr
<< __func__
<< " failed to load os-type: " << cpp_strerror(r
)
5385 if (type
!= "bluestore") {
5386 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5391 if (cct
->_conf
->bluestore_fsck_on_mount
) {
5392 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mount_deep
);
5396 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5401 int r
= _open_path();
5404 r
= _open_fsid(false);
5408 r
= _read_fsid(&fsid
);
5416 r
= _open_bdev(false);
5420 r
= _open_db(false);
5427 r
= _open_super_meta();
5431 r
= _open_fm(false);
5439 r
= _open_collections();
5443 r
= _reload_logger();
5448 r
= _reconcile_bluefs_freespace();
5455 r
= _deferred_replay();
5459 mempool_thread
.init();
5483 int BlueStore::umount()
5485 assert(_kv_only
|| mounted
);
5486 dout(1) << __func__
<< dendl
;
5489 _osr_unregister_all();
5493 mempool_thread
.shutdown();
5494 dout(20) << __func__
<< " stopping kv thread" << dendl
;
5497 dout(20) << __func__
<< " closing" << dendl
;
5507 if (cct
->_conf
->bluestore_fsck_on_umount
) {
5508 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_umount_deep
);
5512 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5519 static void apply(uint64_t off
,
5521 uint64_t granularity
,
5522 BlueStore::mempool_dynamic_bitset
&bitset
,
5523 std::function
<void(uint64_t,
5524 BlueStore::mempool_dynamic_bitset
&)> f
) {
5525 auto end
= ROUND_UP_TO(off
+ len
, granularity
);
5527 uint64_t pos
= off
/ granularity
;
5533 int BlueStore::_fsck_check_extents(
5534 const ghobject_t
& oid
,
5535 const PExtentVector
& extents
,
5537 mempool_dynamic_bitset
&used_blocks
,
5538 uint64_t granularity
,
5539 store_statfs_t
& expected_statfs
)
5541 dout(30) << __func__
<< " oid " << oid
<< " extents " << extents
<< dendl
;
5543 for (auto e
: extents
) {
5546 expected_statfs
.allocated
+= e
.length
;
5548 expected_statfs
.compressed_allocated
+= e
.length
;
5550 bool already
= false;
5552 e
.offset
, e
.length
, granularity
, used_blocks
,
5553 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5554 assert(pos
< bs
.size());
5561 derr
<< " " << oid
<< " extent " << e
5562 << " or a subset is already allocated" << dendl
;
5565 if (e
.end() > bdev
->get_size()) {
5566 derr
<< " " << oid
<< " extent " << e
5567 << " past end of block device" << dendl
;
5574 int BlueStore::_fsck(bool deep
, bool repair
)
5577 << (repair
? " fsck" : " repair")
5578 << (deep
? " (deep)" : " (shallow)") << " start" << dendl
;
5582 typedef btree::btree_set
<
5583 uint64_t,std::less
<uint64_t>,
5584 mempool::bluestore_fsck::pool_allocator
<uint64_t>> uint64_t_btree_t
;
5585 uint64_t_btree_t used_nids
;
5586 uint64_t_btree_t used_omap_head
;
5587 uint64_t_btree_t used_sbids
;
5589 mempool_dynamic_bitset used_blocks
;
5590 KeyValueDB::Iterator it
;
5591 store_statfs_t expected_statfs
, actual_statfs
;
5593 list
<ghobject_t
> oids
;
5595 bluestore_extent_ref_map_t ref_map
;
5598 mempool::bluestore_fsck::map
<uint64_t,sb_info_t
> sb_info
;
5600 uint64_t num_objects
= 0;
5601 uint64_t num_extents
= 0;
5602 uint64_t num_blobs
= 0;
5603 uint64_t num_spanning_blobs
= 0;
5604 uint64_t num_shared_blobs
= 0;
5605 uint64_t num_sharded_objects
= 0;
5606 uint64_t num_object_shards
= 0;
5608 utime_t start
= ceph_clock_now();
5610 int r
= _open_path();
5613 r
= _open_fsid(false);
5617 r
= _read_fsid(&fsid
);
5625 r
= _open_bdev(false);
5629 r
= _open_db(false);
5633 r
= _open_super_meta();
5637 r
= _open_fm(false);
5645 r
= _open_collections(&errors
);
5649 mempool_thread
.init();
5651 // we need finishers and kv_{sync,finalize}_thread *just* for replay
5653 r
= _deferred_replay();
5658 used_blocks
.resize(fm
->get_alloc_units());
5660 0, MAX(min_alloc_size
, SUPER_RESERVED
), fm
->get_alloc_size(), used_blocks
,
5661 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5662 assert(pos
< bs
.size());
5668 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
5670 e
.get_start(), e
.get_len(), fm
->get_alloc_size(), used_blocks
,
5671 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5672 assert(pos
< bs
.size());
5685 // get expected statfs; fill unaffected fields to be able to compare
5687 statfs(&actual_statfs
);
5688 expected_statfs
.total
= actual_statfs
.total
;
5689 expected_statfs
.available
= actual_statfs
.available
;
5692 dout(1) << __func__
<< " walking object keyspace" << dendl
;
5693 it
= db
->get_iterator(PREFIX_OBJ
);
5697 mempool::bluestore_fsck::list
<string
> expecting_shards
;
5698 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5699 if (g_conf
->bluestore_debug_fsck_abort
) {
5702 dout(30) << " key " << pretty_binary_string(it
->key()) << dendl
;
5703 if (is_extent_shard_key(it
->key())) {
5704 while (!expecting_shards
.empty() &&
5705 expecting_shards
.front() < it
->key()) {
5706 derr
<< "fsck error: missing shard key "
5707 << pretty_binary_string(expecting_shards
.front())
5710 expecting_shards
.pop_front();
5712 if (!expecting_shards
.empty() &&
5713 expecting_shards
.front() == it
->key()) {
5715 expecting_shards
.pop_front();
5721 get_key_extent_shard(it
->key(), &okey
, &offset
);
5722 derr
<< "fsck error: stray shard 0x" << std::hex
<< offset
5723 << std::dec
<< dendl
;
5724 if (expecting_shards
.empty()) {
5725 derr
<< "fsck error: " << pretty_binary_string(it
->key())
5726 << " is unexpected" << dendl
;
5730 while (expecting_shards
.front() > it
->key()) {
5731 derr
<< "fsck error: saw " << pretty_binary_string(it
->key())
5733 derr
<< "fsck error: exp "
5734 << pretty_binary_string(expecting_shards
.front()) << dendl
;
5736 expecting_shards
.pop_front();
5737 if (expecting_shards
.empty()) {
5745 int r
= get_key_object(it
->key(), &oid
);
5747 derr
<< "fsck error: bad object key "
5748 << pretty_binary_string(it
->key()) << dendl
;
5753 oid
.shard_id
!= pgid
.shard
||
5754 oid
.hobj
.pool
!= (int64_t)pgid
.pool() ||
5755 !c
->contains(oid
)) {
5757 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
=
5759 p
!= coll_map
.end();
5761 if (p
->second
->contains(oid
)) {
5767 derr
<< "fsck error: stray object " << oid
5768 << " not owned by any collection" << dendl
;
5772 c
->cid
.is_pg(&pgid
);
5773 dout(20) << __func__
<< " collection " << c
->cid
<< " " << c
->cnode
5777 if (!expecting_shards
.empty()) {
5778 for (auto &k
: expecting_shards
) {
5779 derr
<< "fsck error: missing shard key "
5780 << pretty_binary_string(k
) << dendl
;
5783 expecting_shards
.clear();
5786 dout(10) << __func__
<< " " << oid
<< dendl
;
5787 RWLock::RLocker
l(c
->lock
);
5788 OnodeRef o
= c
->get_onode(oid
, false);
5790 if (o
->onode
.nid
> nid_max
) {
5791 derr
<< "fsck error: " << oid
<< " nid " << o
->onode
.nid
5792 << " > nid_max " << nid_max
<< dendl
;
5795 if (used_nids
.count(o
->onode
.nid
)) {
5796 derr
<< "fsck error: " << oid
<< " nid " << o
->onode
.nid
5797 << " already in use" << dendl
;
5799 continue; // go for next object
5801 used_nids
.insert(o
->onode
.nid
);
5804 num_spanning_blobs
+= o
->extent_map
.spanning_blob_map
.size();
5805 o
->extent_map
.fault_range(db
, 0, OBJECT_MAX_SIZE
);
5808 if (!o
->extent_map
.shards
.empty()) {
5809 ++num_sharded_objects
;
5810 num_object_shards
+= o
->extent_map
.shards
.size();
5812 for (auto& s
: o
->extent_map
.shards
) {
5813 dout(20) << __func__
<< " shard " << *s
.shard_info
<< dendl
;
5814 expecting_shards
.push_back(string());
5815 get_extent_shard_key(o
->key
, s
.shard_info
->offset
,
5816 &expecting_shards
.back());
5817 if (s
.shard_info
->offset
>= o
->onode
.size
) {
5818 derr
<< "fsck error: " << oid
<< " shard 0x" << std::hex
5819 << s
.shard_info
->offset
<< " past EOF at 0x" << o
->onode
.size
5820 << std::dec
<< dendl
;
5825 map
<BlobRef
,bluestore_blob_t::unused_t
> referenced
;
5827 mempool::bluestore_fsck::map
<BlobRef
,
5828 bluestore_blob_use_tracker_t
> ref_map
;
5829 for (auto& l
: o
->extent_map
.extent_map
) {
5830 dout(20) << __func__
<< " " << l
<< dendl
;
5831 if (l
.logical_offset
< pos
) {
5832 derr
<< "fsck error: " << oid
<< " lextent at 0x"
5833 << std::hex
<< l
.logical_offset
5834 << " overlaps with the previous, which ends at 0x" << pos
5835 << std::dec
<< dendl
;
5838 if (o
->extent_map
.spans_shard(l
.logical_offset
, l
.length
)) {
5839 derr
<< "fsck error: " << oid
<< " lextent at 0x"
5840 << std::hex
<< l
.logical_offset
<< "~" << l
.length
5841 << " spans a shard boundary"
5842 << std::dec
<< dendl
;
5845 pos
= l
.logical_offset
+ l
.length
;
5846 expected_statfs
.stored
+= l
.length
;
5848 const bluestore_blob_t
& blob
= l
.blob
->get_blob();
5850 auto& ref
= ref_map
[l
.blob
];
5851 if (ref
.is_empty()) {
5852 uint32_t min_release_size
= blob
.get_release_size(min_alloc_size
);
5853 uint32_t l
= blob
.get_logical_length();
5854 ref
.init(l
, min_release_size
);
5860 if (blob
.has_unused()) {
5861 auto p
= referenced
.find(l
.blob
);
5862 bluestore_blob_t::unused_t
*pu
;
5863 if (p
== referenced
.end()) {
5864 pu
= &referenced
[l
.blob
];
5868 uint64_t blob_len
= blob
.get_logical_length();
5869 assert((blob_len
% (sizeof(*pu
)*8)) == 0);
5870 assert(l
.blob_offset
+ l
.length
<= blob_len
);
5871 uint64_t chunk_size
= blob_len
/ (sizeof(*pu
)*8);
5872 uint64_t start
= l
.blob_offset
/ chunk_size
;
5874 ROUND_UP_TO(l
.blob_offset
+ l
.length
, chunk_size
) / chunk_size
;
5875 for (auto i
= start
; i
< end
; ++i
) {
5880 for (auto &i
: referenced
) {
5881 dout(20) << __func__
<< " referenced 0x" << std::hex
<< i
.second
5882 << std::dec
<< " for " << *i
.first
<< dendl
;
5883 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5884 if (i
.second
& blob
.unused
) {
5885 derr
<< "fsck error: " << oid
<< " blob claims unused 0x"
5886 << std::hex
<< blob
.unused
5887 << " but extents reference 0x" << i
.second
5888 << " on blob " << *i
.first
<< dendl
;
5891 if (blob
.has_csum()) {
5892 uint64_t blob_len
= blob
.get_logical_length();
5893 uint64_t unused_chunk_size
= blob_len
/ (sizeof(blob
.unused
)*8);
5894 unsigned csum_count
= blob
.get_csum_count();
5895 unsigned csum_chunk_size
= blob
.get_csum_chunk_size();
5896 for (unsigned p
= 0; p
< csum_count
; ++p
) {
5897 unsigned pos
= p
* csum_chunk_size
;
5898 unsigned firstbit
= pos
/ unused_chunk_size
; // [firstbit,lastbit]
5899 unsigned lastbit
= (pos
+ csum_chunk_size
- 1) / unused_chunk_size
;
5900 unsigned mask
= 1u << firstbit
;
5901 for (unsigned b
= firstbit
+ 1; b
<= lastbit
; ++b
) {
5904 if ((blob
.unused
& mask
) == mask
) {
5905 // this csum chunk region is marked unused
5906 if (blob
.get_csum_item(p
) != 0) {
5907 derr
<< "fsck error: " << oid
5908 << " blob claims csum chunk 0x" << std::hex
<< pos
5909 << "~" << csum_chunk_size
5910 << " is unused (mask 0x" << mask
<< " of unused 0x"
5911 << blob
.unused
<< ") but csum is non-zero 0x"
5912 << blob
.get_csum_item(p
) << std::dec
<< " on blob "
5913 << *i
.first
<< dendl
;
5920 for (auto &i
: ref_map
) {
5922 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5923 bool equal
= i
.first
->get_blob_use_tracker().equal(i
.second
);
5925 derr
<< "fsck error: " << oid
<< " blob " << *i
.first
5926 << " doesn't match expected ref_map " << i
.second
<< dendl
;
5929 if (blob
.is_compressed()) {
5930 expected_statfs
.compressed
+= blob
.get_compressed_payload_length();
5931 expected_statfs
.compressed_original
+=
5932 i
.first
->get_referenced_bytes();
5934 if (blob
.is_shared()) {
5935 if (i
.first
->shared_blob
->get_sbid() > blobid_max
) {
5936 derr
<< "fsck error: " << oid
<< " blob " << blob
5937 << " sbid " << i
.first
->shared_blob
->get_sbid() << " > blobid_max "
5938 << blobid_max
<< dendl
;
5940 } else if (i
.first
->shared_blob
->get_sbid() == 0) {
5941 derr
<< "fsck error: " << oid
<< " blob " << blob
5942 << " marked as shared but has uninitialized sbid"
5946 sb_info_t
& sbi
= sb_info
[i
.first
->shared_blob
->get_sbid()];
5947 sbi
.sb
= i
.first
->shared_blob
;
5948 sbi
.oids
.push_back(oid
);
5949 sbi
.compressed
= blob
.is_compressed();
5950 for (auto e
: blob
.get_extents()) {
5952 sbi
.ref_map
.get(e
.offset
, e
.length
);
5956 errors
+= _fsck_check_extents(oid
, blob
.get_extents(),
5957 blob
.is_compressed(),
5959 fm
->get_alloc_size(),
5965 int r
= _do_read(c
.get(), o
, 0, o
->onode
.size
, bl
, 0);
5968 derr
<< "fsck error: " << oid
<< " error during read: "
5969 << cpp_strerror(r
) << dendl
;
5973 if (o
->onode
.has_omap()) {
5974 if (used_omap_head
.count(o
->onode
.nid
)) {
5975 derr
<< "fsck error: " << oid
<< " omap_head " << o
->onode
.nid
5976 << " already in use" << dendl
;
5979 used_omap_head
.insert(o
->onode
.nid
);
5984 dout(1) << __func__
<< " checking shared_blobs" << dendl
;
5985 it
= db
->get_iterator(PREFIX_SHARED_BLOB
);
5987 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5988 string key
= it
->key();
5990 if (get_key_shared_blob(key
, &sbid
)) {
5991 derr
<< "fsck error: bad key '" << key
5992 << "' in shared blob namespace" << dendl
;
5996 auto p
= sb_info
.find(sbid
);
5997 if (p
== sb_info
.end()) {
5998 derr
<< "fsck error: found stray shared blob data for sbid 0x"
5999 << std::hex
<< sbid
<< std::dec
<< dendl
;
6003 sb_info_t
& sbi
= p
->second
;
6004 bluestore_shared_blob_t
shared_blob(sbid
);
6005 bufferlist bl
= it
->value();
6006 bufferlist::iterator blp
= bl
.begin();
6007 ::decode(shared_blob
, blp
);
6008 dout(20) << __func__
<< " " << *sbi
.sb
<< " " << shared_blob
<< dendl
;
6009 if (shared_blob
.ref_map
!= sbi
.ref_map
) {
6010 derr
<< "fsck error: shared blob 0x" << std::hex
<< sbid
6011 << std::dec
<< " ref_map " << shared_blob
.ref_map
6012 << " != expected " << sbi
.ref_map
<< dendl
;
6015 PExtentVector extents
;
6016 for (auto &r
: shared_blob
.ref_map
.ref_map
) {
6017 extents
.emplace_back(bluestore_pextent_t(r
.first
, r
.second
.length
));
6019 errors
+= _fsck_check_extents(p
->second
.oids
.front(),
6021 p
->second
.compressed
,
6023 fm
->get_alloc_size(),
6029 for (auto &p
: sb_info
) {
6030 derr
<< "fsck error: shared_blob 0x" << p
.first
6031 << " key is missing (" << *p
.second
.sb
<< ")" << dendl
;
6034 if (!(actual_statfs
== expected_statfs
)) {
6035 derr
<< "fsck error: actual " << actual_statfs
6036 << " != expected " << expected_statfs
<< dendl
;
6040 dout(1) << __func__
<< " checking for stray omap data" << dendl
;
6041 it
= db
->get_iterator(PREFIX_OMAP
);
6043 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
6045 _key_decode_u64(it
->key().c_str(), &omap_head
);
6046 if (used_omap_head
.count(omap_head
) == 0) {
6047 derr
<< "fsck error: found stray omap data on omap_head "
6048 << omap_head
<< dendl
;
6054 dout(1) << __func__
<< " checking deferred events" << dendl
;
6055 it
= db
->get_iterator(PREFIX_DEFERRED
);
6057 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
6058 bufferlist bl
= it
->value();
6059 bufferlist::iterator p
= bl
.begin();
6060 bluestore_deferred_transaction_t wt
;
6063 } catch (buffer::error
& e
) {
6064 derr
<< "fsck error: failed to decode deferred txn "
6065 << pretty_binary_string(it
->key()) << dendl
;
6069 dout(20) << __func__
<< " deferred " << wt
.seq
6070 << " ops " << wt
.ops
.size()
6071 << " released 0x" << std::hex
<< wt
.released
<< std::dec
<< dendl
;
6072 for (auto e
= wt
.released
.begin(); e
!= wt
.released
.end(); ++e
) {
6074 e
.get_start(), e
.get_len(), fm
->get_alloc_size(), used_blocks
,
6075 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
6076 assert(pos
< bs
.size());
6084 dout(1) << __func__
<< " checking freelist vs allocated" << dendl
;
6086 // remove bluefs_extents from used set since the freelist doesn't
6087 // know they are allocated.
6088 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
6090 e
.get_start(), e
.get_len(), fm
->get_alloc_size(), used_blocks
,
6091 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
6092 assert(pos
< bs
.size());
6097 fm
->enumerate_reset();
6098 uint64_t offset
, length
;
6099 while (fm
->enumerate_next(&offset
, &length
)) {
6100 bool intersects
= false;
6102 offset
, length
, fm
->get_alloc_size(), used_blocks
,
6103 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
6104 assert(pos
< bs
.size());
6113 if (offset
== SUPER_RESERVED
&&
6114 length
== min_alloc_size
- SUPER_RESERVED
) {
6115 // this is due to the change just after luminous to min_alloc_size
6116 // granularity allocations, and our baked in assumption at the top
6117 // of _fsck that 0~ROUND_UP_TO(SUPER_RESERVED,min_alloc_size) is used
6118 // (vs luminous's ROUND_UP_TO(SUPER_RESERVED,block_size)). harmless,
6119 // since we will never allocate this region below min_alloc_size.
6120 dout(10) << __func__
<< " ignoring free extent between SUPER_RESERVED"
6121 << " and min_alloc_size, 0x" << std::hex
<< offset
<< "~"
6124 derr
<< "fsck error: free extent 0x" << std::hex
<< offset
6125 << "~" << length
<< std::dec
6126 << " intersects allocated blocks" << dendl
;
6131 fm
->enumerate_reset();
6132 size_t count
= used_blocks
.count();
6133 if (used_blocks
.size() != count
) {
6134 assert(used_blocks
.size() > count
);
6137 size_t start
= used_blocks
.find_first();
6138 while (start
!= decltype(used_blocks
)::npos
) {
6141 size_t next
= used_blocks
.find_next(cur
);
6142 if (next
!= cur
+ 1) {
6143 derr
<< "fsck error: leaked extent 0x" << std::hex
6144 << ((uint64_t)start
* fm
->get_alloc_size()) << "~"
6145 << ((cur
+ 1 - start
) * fm
->get_alloc_size()) << std::dec
6158 mempool_thread
.shutdown();
6165 it
.reset(); // before db is closed
6174 // fatal errors take precedence
6178 dout(2) << __func__
<< " " << num_objects
<< " objects, "
6179 << num_sharded_objects
<< " of them sharded. "
6181 dout(2) << __func__
<< " " << num_extents
<< " extents to "
6182 << num_blobs
<< " blobs, "
6183 << num_spanning_blobs
<< " spanning, "
6184 << num_shared_blobs
<< " shared."
6187 utime_t duration
= ceph_clock_now() - start
;
6188 dout(1) << __func__
<< " finish with " << errors
<< " errors, " << repaired
6189 << " repaired, " << (errors
- repaired
) << " remaining in "
6190 << duration
<< " seconds" << dendl
;
6191 return errors
- repaired
;
6194 void BlueStore::collect_metadata(map
<string
,string
> *pm
)
6196 dout(10) << __func__
<< dendl
;
6197 bdev
->collect_metadata("bluestore_bdev_", pm
);
6199 (*pm
)["bluefs"] = "1";
6200 (*pm
)["bluefs_single_shared_device"] = stringify((int)bluefs_single_shared_device
);
6201 bluefs
->collect_metadata(pm
);
6203 (*pm
)["bluefs"] = "0";
6207 int BlueStore::statfs(struct store_statfs_t
*buf
)
6210 buf
->total
= bdev
->get_size();
6211 buf
->available
= alloc
->get_free();
6214 // part of our shared device is "free" according to BlueFS, but we
6215 // can't touch bluestore_bluefs_min of it.
6216 int64_t shared_available
= std::min(
6217 bluefs
->get_free(bluefs_shared_bdev
),
6218 bluefs
->get_total(bluefs_shared_bdev
) - cct
->_conf
->bluestore_bluefs_min
);
6219 if (shared_available
> 0) {
6220 buf
->available
+= shared_available
;
6225 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
6227 buf
->allocated
= vstatfs
.allocated();
6228 buf
->stored
= vstatfs
.stored();
6229 buf
->compressed
= vstatfs
.compressed();
6230 buf
->compressed_original
= vstatfs
.compressed_original();
6231 buf
->compressed_allocated
= vstatfs
.compressed_allocated();
6234 dout(20) << __func__
<< *buf
<< dendl
;
6241 BlueStore::CollectionRef
BlueStore::_get_collection(const coll_t
& cid
)
6243 RWLock::RLocker
l(coll_lock
);
6244 ceph::unordered_map
<coll_t
,CollectionRef
>::iterator cp
= coll_map
.find(cid
);
6245 if (cp
== coll_map
.end())
6246 return CollectionRef();
6250 void BlueStore::_queue_reap_collection(CollectionRef
& c
)
6252 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6253 // _reap_collections and this in the same thread,
6254 // so no need a lock.
6255 removed_collections
.push_back(c
);
6258 void BlueStore::_reap_collections()
6261 list
<CollectionRef
> removed_colls
;
6263 // _queue_reap_collection and this in the same thread.
6264 // So no need a lock.
6265 if (!removed_collections
.empty())
6266 removed_colls
.swap(removed_collections
);
6271 list
<CollectionRef
>::iterator p
= removed_colls
.begin();
6272 while (p
!= removed_colls
.end()) {
6273 CollectionRef c
= *p
;
6274 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6275 if (c
->onode_map
.map_any([&](OnodeRef o
) {
6277 if (o
->flushing_count
.load()) {
6278 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " " << o
->oid
6279 << " flush_txns " << o
->flushing_count
<< dendl
;
6287 c
->onode_map
.clear();
6288 p
= removed_colls
.erase(p
);
6289 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " done" << dendl
;
6291 if (removed_colls
.empty()) {
6292 dout(10) << __func__
<< " all reaped" << dendl
;
6294 removed_collections
.splice(removed_collections
.begin(), removed_colls
);
6298 void BlueStore::_update_cache_logger()
6300 uint64_t num_onodes
= 0;
6301 uint64_t num_extents
= 0;
6302 uint64_t num_blobs
= 0;
6303 uint64_t num_buffers
= 0;
6304 uint64_t num_buffer_bytes
= 0;
6305 for (auto c
: cache_shards
) {
6306 c
->add_stats(&num_onodes
, &num_extents
, &num_blobs
,
6307 &num_buffers
, &num_buffer_bytes
);
6309 logger
->set(l_bluestore_onodes
, num_onodes
);
6310 logger
->set(l_bluestore_extents
, num_extents
);
6311 logger
->set(l_bluestore_blobs
, num_blobs
);
6312 logger
->set(l_bluestore_buffers
, num_buffers
);
6313 logger
->set(l_bluestore_buffer_bytes
, num_buffer_bytes
);
6319 ObjectStore::CollectionHandle
BlueStore::open_collection(const coll_t
& cid
)
6321 return _get_collection(cid
);
6324 bool BlueStore::exists(const coll_t
& cid
, const ghobject_t
& oid
)
6326 CollectionHandle c
= _get_collection(cid
);
6329 return exists(c
, oid
);
6332 bool BlueStore::exists(CollectionHandle
&c_
, const ghobject_t
& oid
)
6334 Collection
*c
= static_cast<Collection
*>(c_
.get());
6335 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
6342 RWLock::RLocker
l(c
->lock
);
6343 OnodeRef o
= c
->get_onode(oid
, false);
6344 if (!o
|| !o
->exists
)
6351 int BlueStore::stat(
6353 const ghobject_t
& oid
,
6357 CollectionHandle c
= _get_collection(cid
);
6360 return stat(c
, oid
, st
, allow_eio
);
6363 int BlueStore::stat(
6364 CollectionHandle
&c_
,
6365 const ghobject_t
& oid
,
6369 Collection
*c
= static_cast<Collection
*>(c_
.get());
6372 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
6375 RWLock::RLocker
l(c
->lock
);
6376 OnodeRef o
= c
->get_onode(oid
, false);
6377 if (!o
|| !o
->exists
)
6379 st
->st_size
= o
->onode
.size
;
6380 st
->st_blksize
= 4096;
6381 st
->st_blocks
= (st
->st_size
+ st
->st_blksize
- 1) / st
->st_blksize
;
6386 if (_debug_mdata_eio(oid
)) {
6388 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6392 int BlueStore::set_collection_opts(
6394 const pool_opts_t
& opts
)
6396 CollectionHandle ch
= _get_collection(cid
);
6399 Collection
*c
= static_cast<Collection
*>(ch
.get());
6400 dout(15) << __func__
<< " " << cid
<< " options " << opts
<< dendl
;
6403 RWLock::WLocker
l(c
->lock
);
6404 c
->pool_opts
= opts
;
6408 int BlueStore::read(
6410 const ghobject_t
& oid
,
6416 CollectionHandle c
= _get_collection(cid
);
6419 return read(c
, oid
, offset
, length
, bl
, op_flags
);
6422 int BlueStore::read(
6423 CollectionHandle
&c_
,
6424 const ghobject_t
& oid
,
6430 utime_t start
= ceph_clock_now();
6431 Collection
*c
= static_cast<Collection
*>(c_
.get());
6432 const coll_t
&cid
= c
->get_cid();
6433 dout(15) << __func__
<< " " << cid
<< " " << oid
6434 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6442 RWLock::RLocker
l(c
->lock
);
6443 utime_t start1
= ceph_clock_now();
6444 OnodeRef o
= c
->get_onode(oid
, false);
6445 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start1
);
6446 if (!o
|| !o
->exists
) {
6451 if (offset
== length
&& offset
== 0)
6452 length
= o
->onode
.size
;
6454 r
= _do_read(c
, o
, offset
, length
, bl
, op_flags
);
6456 logger
->inc(l_bluestore_read_eio
);
6461 if (r
>= 0 && _debug_data_eio(oid
)) {
6463 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6464 } else if (cct
->_conf
->bluestore_debug_random_read_err
&&
6465 (rand() % (int)(cct
->_conf
->bluestore_debug_random_read_err
* 100.0)) == 0) {
6466 dout(0) << __func__
<< ": inject random EIO" << dendl
;
6469 dout(10) << __func__
<< " " << cid
<< " " << oid
6470 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6471 << " = " << r
<< dendl
;
6472 logger
->tinc(l_bluestore_read_lat
, ceph_clock_now() - start
);
6476 // --------------------------------------------------------
6477 // intermediate data structures used while reading
6479 uint64_t logical_offset
;
6480 uint64_t blob_xoffset
; //region offset within the blob
6484 // used later in read process
6488 region_t(uint64_t offset
, uint64_t b_offs
, uint64_t len
)
6489 : logical_offset(offset
),
6490 blob_xoffset(b_offs
),
6492 region_t(const region_t
& from
)
6493 : logical_offset(from
.logical_offset
),
6494 blob_xoffset(from
.blob_xoffset
),
6495 length(from
.length
){}
6497 friend ostream
& operator<<(ostream
& out
, const region_t
& r
) {
6498 return out
<< "0x" << std::hex
<< r
.logical_offset
<< ":"
6499 << r
.blob_xoffset
<< "~" << r
.length
<< std::dec
;
6503 typedef list
<region_t
> regions2read_t
;
6504 typedef map
<BlueStore::BlobRef
, regions2read_t
> blobs2read_t
;
6506 int BlueStore::_do_read(
6517 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6518 << " size 0x" << o
->onode
.size
<< " (" << std::dec
6519 << o
->onode
.size
<< ")" << dendl
;
6522 if (offset
>= o
->onode
.size
) {
6526 // generally, don't buffer anything, unless the client explicitly requests
6528 bool buffered
= false;
6529 if (op_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
6530 dout(20) << __func__
<< " will do buffered read" << dendl
;
6532 } else if (cct
->_conf
->bluestore_default_buffered_read
&&
6533 (op_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
6534 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
6535 dout(20) << __func__
<< " defaulting to buffered read" << dendl
;
6539 if (offset
+ length
> o
->onode
.size
) {
6540 length
= o
->onode
.size
- offset
;
6543 utime_t start
= ceph_clock_now();
6544 o
->extent_map
.fault_range(db
, offset
, length
);
6545 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start
);
6548 ready_regions_t ready_regions
;
6550 // build blob-wise list to of stuff read (that isn't cached)
6551 blobs2read_t blobs2read
;
6552 unsigned left
= length
;
6553 uint64_t pos
= offset
;
6554 unsigned num_regions
= 0;
6555 auto lp
= o
->extent_map
.seek_lextent(offset
);
6556 while (left
> 0 && lp
!= o
->extent_map
.extent_map
.end()) {
6557 if (pos
< lp
->logical_offset
) {
6558 unsigned hole
= lp
->logical_offset
- pos
;
6562 dout(30) << __func__
<< " hole 0x" << std::hex
<< pos
<< "~" << hole
6563 << std::dec
<< dendl
;
6567 BlobRef
& bptr
= lp
->blob
;
6568 unsigned l_off
= pos
- lp
->logical_offset
;
6569 unsigned b_off
= l_off
+ lp
->blob_offset
;
6570 unsigned b_len
= std::min(left
, lp
->length
- l_off
);
6572 ready_regions_t cache_res
;
6573 interval_set
<uint32_t> cache_interval
;
6574 bptr
->shared_blob
->bc
.read(
6575 bptr
->shared_blob
->get_cache(), b_off
, b_len
, cache_res
, cache_interval
);
6576 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6577 << " need 0x" << b_off
<< "~" << b_len
6578 << " cache has 0x" << cache_interval
6579 << std::dec
<< dendl
;
6581 auto pc
= cache_res
.begin();
6584 if (pc
!= cache_res
.end() &&
6585 pc
->first
== b_off
) {
6586 l
= pc
->second
.length();
6587 ready_regions
[pos
].claim(pc
->second
);
6588 dout(30) << __func__
<< " use cache 0x" << std::hex
<< pos
<< ": 0x"
6589 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6593 if (pc
!= cache_res
.end()) {
6594 assert(pc
->first
> b_off
);
6595 l
= pc
->first
- b_off
;
6597 dout(30) << __func__
<< " will read 0x" << std::hex
<< pos
<< ": 0x"
6598 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6599 blobs2read
[bptr
].emplace_back(region_t(pos
, b_off
, l
));
6610 // read raw blob data. use aio if we have >1 blobs to read.
6611 start
= ceph_clock_now(); // for the sake of simplicity
6612 // measure the whole block below.
6613 // The error isn't that much...
6614 vector
<bufferlist
> compressed_blob_bls
;
6615 IOContext
ioc(cct
, NULL
, true); // allow EIO
6616 for (auto& p
: blobs2read
) {
6617 const BlobRef
& bptr
= p
.first
;
6618 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6619 << " need " << p
.second
<< std::dec
<< dendl
;
6620 if (bptr
->get_blob().is_compressed()) {
6621 // read the whole thing
6622 if (compressed_blob_bls
.empty()) {
6623 // ensure we avoid any reallocation on subsequent blobs
6624 compressed_blob_bls
.reserve(blobs2read
.size());
6626 compressed_blob_bls
.push_back(bufferlist());
6627 bufferlist
& bl
= compressed_blob_bls
.back();
6628 r
= bptr
->get_blob().map(
6629 0, bptr
->get_blob().get_ondisk_length(),
6630 [&](uint64_t offset
, uint64_t length
) {
6632 // use aio if there are more regions to read than those in this blob
6633 if (num_regions
> p
.second
.size()) {
6634 r
= bdev
->aio_read(offset
, length
, &bl
, &ioc
);
6636 r
= bdev
->read(offset
, length
, &bl
, &ioc
, false);
6643 derr
<< __func__
<< " bdev-read failed: " << cpp_strerror(r
) << dendl
;
6645 // propagate EIO to caller
6652 for (auto& reg
: p
.second
) {
6653 // determine how much of the blob to read
6654 uint64_t chunk_size
= bptr
->get_blob().get_chunk_size(block_size
);
6655 reg
.r_off
= reg
.blob_xoffset
;
6656 uint64_t r_len
= reg
.length
;
6657 reg
.front
= reg
.r_off
% chunk_size
;
6659 reg
.r_off
-= reg
.front
;
6662 unsigned tail
= r_len
% chunk_size
;
6664 r_len
+= chunk_size
- tail
;
6666 dout(20) << __func__
<< " region 0x" << std::hex
6667 << reg
.logical_offset
6668 << ": 0x" << reg
.blob_xoffset
<< "~" << reg
.length
6669 << " reading 0x" << reg
.r_off
<< "~" << r_len
<< std::dec
6673 r
= bptr
->get_blob().map(
6675 [&](uint64_t offset
, uint64_t length
) {
6677 // use aio if there is more than one region to read
6678 if (num_regions
> 1) {
6679 r
= bdev
->aio_read(offset
, length
, ®
.bl
, &ioc
);
6681 r
= bdev
->read(offset
, length
, ®
.bl
, &ioc
, false);
6688 derr
<< __func__
<< " bdev-read failed: " << cpp_strerror(r
)
6691 // propagate EIO to caller
6696 assert(reg
.bl
.length() == r_len
);
6700 if (ioc
.has_pending_aios()) {
6701 bdev
->aio_submit(&ioc
);
6702 dout(20) << __func__
<< " waiting for aio" << dendl
;
6704 r
= ioc
.get_return_value();
6706 assert(r
== -EIO
); // no other errors allowed
6710 logger
->tinc(l_bluestore_read_wait_aio_lat
, ceph_clock_now() - start
);
6712 // enumerate and decompress desired blobs
6713 auto p
= compressed_blob_bls
.begin();
6714 blobs2read_t::iterator b2r_it
= blobs2read
.begin();
6715 while (b2r_it
!= blobs2read
.end()) {
6716 const BlobRef
& bptr
= b2r_it
->first
;
6717 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6718 << " need 0x" << b2r_it
->second
<< std::dec
<< dendl
;
6719 if (bptr
->get_blob().is_compressed()) {
6720 assert(p
!= compressed_blob_bls
.end());
6721 bufferlist
& compressed_bl
= *p
++;
6722 if (_verify_csum(o
, &bptr
->get_blob(), 0, compressed_bl
,
6723 b2r_it
->second
.front().logical_offset
) < 0) {
6727 r
= _decompress(compressed_bl
, &raw_bl
);
6731 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(), 0,
6734 for (auto& i
: b2r_it
->second
) {
6735 ready_regions
[i
.logical_offset
].substr_of(
6736 raw_bl
, i
.blob_xoffset
, i
.length
);
6739 for (auto& reg
: b2r_it
->second
) {
6740 if (_verify_csum(o
, &bptr
->get_blob(), reg
.r_off
, reg
.bl
,
6741 reg
.logical_offset
) < 0) {
6745 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(),
6749 // prune and keep result
6750 ready_regions
[reg
.logical_offset
].substr_of(
6751 reg
.bl
, reg
.front
, reg
.length
);
6757 // generate a resulting buffer
6758 auto pr
= ready_regions
.begin();
6759 auto pr_end
= ready_regions
.end();
6761 while (pos
< length
) {
6762 if (pr
!= pr_end
&& pr
->first
== pos
+ offset
) {
6763 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6764 << ": data from 0x" << pr
->first
<< "~" << pr
->second
.length()
6765 << std::dec
<< dendl
;
6766 pos
+= pr
->second
.length();
6767 bl
.claim_append(pr
->second
);
6770 uint64_t l
= length
- pos
;
6772 assert(pr
->first
> pos
+ offset
);
6773 l
= pr
->first
- (pos
+ offset
);
6775 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6776 << ": zeros for 0x" << (pos
+ offset
) << "~" << l
6777 << std::dec
<< dendl
;
6782 assert(bl
.length() == length
);
6783 assert(pos
== length
);
6784 assert(pr
== pr_end
);
6789 int BlueStore::_verify_csum(OnodeRef
& o
,
6790 const bluestore_blob_t
* blob
, uint64_t blob_xoffset
,
6791 const bufferlist
& bl
,
6792 uint64_t logical_offset
) const
6796 utime_t start
= ceph_clock_now();
6797 int r
= blob
->verify_csum(blob_xoffset
, bl
, &bad
, &bad_csum
);
6803 blob
->get_csum_chunk_size(),
6804 [&](uint64_t offset
, uint64_t length
) {
6805 pex
.emplace_back(bluestore_pextent_t(offset
, length
));
6808 derr
<< __func__
<< " bad "
6809 << Checksummer::get_csum_type_string(blob
->csum_type
)
6810 << "/0x" << std::hex
<< blob
->get_csum_chunk_size()
6811 << " checksum at blob offset 0x" << bad
6812 << ", got 0x" << bad_csum
<< ", expected 0x"
6813 << blob
->get_csum_item(bad
/ blob
->get_csum_chunk_size()) << std::dec
6814 << ", device location " << pex
6815 << ", logical extent 0x" << std::hex
6816 << (logical_offset
+ bad
- blob_xoffset
) << "~"
6817 << blob
->get_csum_chunk_size() << std::dec
6818 << ", object " << o
->oid
6821 derr
<< __func__
<< " failed with exit code: " << cpp_strerror(r
) << dendl
;
6824 logger
->tinc(l_bluestore_csum_lat
, ceph_clock_now() - start
);
6828 int BlueStore::_decompress(bufferlist
& source
, bufferlist
* result
)
6831 utime_t start
= ceph_clock_now();
6832 bufferlist::iterator i
= source
.begin();
6833 bluestore_compression_header_t chdr
;
6835 int alg
= int(chdr
.type
);
6836 CompressorRef cp
= compressor
;
6837 if (!cp
|| (int)cp
->get_type() != alg
) {
6838 cp
= Compressor::create(cct
, alg
);
6842 // if compressor isn't available - error, because cannot return
6843 // decompressed data?
6844 derr
<< __func__
<< " can't load decompressor " << alg
<< dendl
;
6847 r
= cp
->decompress(i
, chdr
.length
, *result
);
6849 derr
<< __func__
<< " decompression failed with exit code " << r
<< dendl
;
6853 logger
->tinc(l_bluestore_decompress_lat
, ceph_clock_now() - start
);
6857 // this stores fiemap into interval_set, other variations
6858 // use it internally
6859 int BlueStore::_fiemap(
6860 CollectionHandle
&c_
,
6861 const ghobject_t
& oid
,
6864 interval_set
<uint64_t>& destset
)
6866 Collection
*c
= static_cast<Collection
*>(c_
.get());
6870 RWLock::RLocker
l(c
->lock
);
6872 OnodeRef o
= c
->get_onode(oid
, false);
6873 if (!o
|| !o
->exists
) {
6878 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6879 << " size 0x" << o
->onode
.size
<< std::dec
<< dendl
;
6881 boost::intrusive::set
<Extent
>::iterator ep
, eend
;
6882 if (offset
>= o
->onode
.size
)
6885 if (offset
+ length
> o
->onode
.size
) {
6886 length
= o
->onode
.size
- offset
;
6889 o
->extent_map
.fault_range(db
, offset
, length
);
6890 eend
= o
->extent_map
.extent_map
.end();
6891 ep
= o
->extent_map
.seek_lextent(offset
);
6892 while (length
> 0) {
6893 dout(20) << __func__
<< " offset " << offset
<< dendl
;
6894 if (ep
!= eend
&& ep
->logical_offset
+ ep
->length
<= offset
) {
6899 uint64_t x_len
= length
;
6900 if (ep
!= eend
&& ep
->logical_offset
<= offset
) {
6901 uint64_t x_off
= offset
- ep
->logical_offset
;
6902 x_len
= MIN(x_len
, ep
->length
- x_off
);
6903 dout(30) << __func__
<< " lextent 0x" << std::hex
<< offset
<< "~"
6904 << x_len
<< std::dec
<< " blob " << ep
->blob
<< dendl
;
6905 destset
.insert(offset
, x_len
);
6908 if (x_off
+ x_len
== ep
->length
)
6913 ep
->logical_offset
> offset
&&
6914 ep
->logical_offset
- offset
< x_len
) {
6915 x_len
= ep
->logical_offset
- offset
;
6923 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6924 << " size = 0x(" << destset
<< ")" << std::dec
<< dendl
;
6928 int BlueStore::fiemap(
6930 const ghobject_t
& oid
,
6935 CollectionHandle c
= _get_collection(cid
);
6938 return fiemap(c
, oid
, offset
, len
, bl
);
6941 int BlueStore::fiemap(
6942 CollectionHandle
&c_
,
6943 const ghobject_t
& oid
,
6948 interval_set
<uint64_t> m
;
6949 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6956 int BlueStore::fiemap(
6958 const ghobject_t
& oid
,
6961 map
<uint64_t, uint64_t>& destmap
)
6963 CollectionHandle c
= _get_collection(cid
);
6966 return fiemap(c
, oid
, offset
, len
, destmap
);
6969 int BlueStore::fiemap(
6970 CollectionHandle
&c_
,
6971 const ghobject_t
& oid
,
6974 map
<uint64_t, uint64_t>& destmap
)
6976 interval_set
<uint64_t> m
;
6977 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6979 m
.move_into(destmap
);
6984 int BlueStore::getattr(
6986 const ghobject_t
& oid
,
6990 CollectionHandle c
= _get_collection(cid
);
6993 return getattr(c
, oid
, name
, value
);
6996 int BlueStore::getattr(
6997 CollectionHandle
&c_
,
6998 const ghobject_t
& oid
,
7002 Collection
*c
= static_cast<Collection
*>(c_
.get());
7003 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
<< dendl
;
7009 RWLock::RLocker
l(c
->lock
);
7010 mempool::bluestore_cache_other::string
k(name
);
7012 OnodeRef o
= c
->get_onode(oid
, false);
7013 if (!o
|| !o
->exists
) {
7018 if (!o
->onode
.attrs
.count(k
)) {
7022 value
= o
->onode
.attrs
[k
];
7026 if (r
== 0 && _debug_mdata_eio(oid
)) {
7028 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
7030 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
7031 << " = " << r
<< dendl
;
7036 int BlueStore::getattrs(
7038 const ghobject_t
& oid
,
7039 map
<string
,bufferptr
>& aset
)
7041 CollectionHandle c
= _get_collection(cid
);
7044 return getattrs(c
, oid
, aset
);
7047 int BlueStore::getattrs(
7048 CollectionHandle
&c_
,
7049 const ghobject_t
& oid
,
7050 map
<string
,bufferptr
>& aset
)
7052 Collection
*c
= static_cast<Collection
*>(c_
.get());
7053 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
7059 RWLock::RLocker
l(c
->lock
);
7061 OnodeRef o
= c
->get_onode(oid
, false);
7062 if (!o
|| !o
->exists
) {
7066 for (auto& i
: o
->onode
.attrs
) {
7067 aset
.emplace(i
.first
.c_str(), i
.second
);
7073 if (r
== 0 && _debug_mdata_eio(oid
)) {
7075 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
7077 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
7078 << " = " << r
<< dendl
;
7082 int BlueStore::list_collections(vector
<coll_t
>& ls
)
7084 RWLock::RLocker
l(coll_lock
);
7085 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
= coll_map
.begin();
7086 p
!= coll_map
.end();
7088 ls
.push_back(p
->first
);
7092 bool BlueStore::collection_exists(const coll_t
& c
)
7094 RWLock::RLocker
l(coll_lock
);
7095 return coll_map
.count(c
);
7098 int BlueStore::collection_empty(const coll_t
& cid
, bool *empty
)
7100 dout(15) << __func__
<< " " << cid
<< dendl
;
7101 vector
<ghobject_t
> ls
;
7103 int r
= collection_list(cid
, ghobject_t(), ghobject_t::get_max(), 1,
7106 derr
<< __func__
<< " collection_list returned: " << cpp_strerror(r
)
7110 *empty
= ls
.empty();
7111 dout(10) << __func__
<< " " << cid
<< " = " << (int)(*empty
) << dendl
;
7115 int BlueStore::collection_bits(const coll_t
& cid
)
7117 dout(15) << __func__
<< " " << cid
<< dendl
;
7118 CollectionRef c
= _get_collection(cid
);
7121 RWLock::RLocker
l(c
->lock
);
7122 dout(10) << __func__
<< " " << cid
<< " = " << c
->cnode
.bits
<< dendl
;
7123 return c
->cnode
.bits
;
7126 int BlueStore::collection_list(
7127 const coll_t
& cid
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
7128 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
7130 CollectionHandle c
= _get_collection(cid
);
7133 return collection_list(c
, start
, end
, max
, ls
, pnext
);
7136 int BlueStore::collection_list(
7137 CollectionHandle
&c_
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
7138 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
7140 Collection
*c
= static_cast<Collection
*>(c_
.get());
7141 dout(15) << __func__
<< " " << c
->cid
7142 << " start " << start
<< " end " << end
<< " max " << max
<< dendl
;
7145 RWLock::RLocker
l(c
->lock
);
7146 r
= _collection_list(c
, start
, end
, max
, ls
, pnext
);
7149 dout(10) << __func__
<< " " << c
->cid
7150 << " start " << start
<< " end " << end
<< " max " << max
7151 << " = " << r
<< ", ls.size() = " << ls
->size()
7152 << ", next = " << (pnext
? *pnext
: ghobject_t()) << dendl
;
7156 int BlueStore::_collection_list(
7157 Collection
*c
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
7158 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
7165 ghobject_t static_next
;
7166 KeyValueDB::Iterator it
;
7167 string temp_start_key
, temp_end_key
;
7168 string start_key
, end_key
;
7169 bool set_next
= false;
7174 pnext
= &static_next
;
7176 if (start
== ghobject_t::get_max() ||
7177 start
.hobj
.is_max()) {
7180 get_coll_key_range(c
->cid
, c
->cnode
.bits
, &temp_start_key
, &temp_end_key
,
7181 &start_key
, &end_key
);
7182 dout(20) << __func__
7183 << " range " << pretty_binary_string(temp_start_key
)
7184 << " to " << pretty_binary_string(temp_end_key
)
7185 << " and " << pretty_binary_string(start_key
)
7186 << " to " << pretty_binary_string(end_key
)
7187 << " start " << start
<< dendl
;
7188 it
= db
->get_iterator(PREFIX_OBJ
);
7189 if (start
== ghobject_t() ||
7190 start
.hobj
== hobject_t() ||
7191 start
== c
->cid
.get_min_hobj()) {
7192 it
->upper_bound(temp_start_key
);
7196 get_object_key(cct
, start
, &k
);
7197 if (start
.hobj
.is_temp()) {
7199 assert(k
>= temp_start_key
&& k
< temp_end_key
);
7202 assert(k
>= start_key
&& k
< end_key
);
7204 dout(20) << " start from " << pretty_binary_string(k
)
7205 << " temp=" << (int)temp
<< dendl
;
7208 if (end
.hobj
.is_max()) {
7209 pend
= temp
? temp_end_key
: end_key
;
7211 get_object_key(cct
, end
, &end_key
);
7212 if (end
.hobj
.is_temp()) {
7218 pend
= temp
? temp_end_key
: end_key
;
7221 dout(20) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7223 if (!it
->valid() || it
->key() >= pend
) {
7225 dout(20) << __func__
<< " iterator not valid (end of db?)" << dendl
;
7227 dout(20) << __func__
<< " key " << pretty_binary_string(it
->key())
7228 << " >= " << end
<< dendl
;
7230 if (end
.hobj
.is_temp()) {
7233 dout(30) << __func__
<< " switch to non-temp namespace" << dendl
;
7235 it
->upper_bound(start_key
);
7237 dout(30) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7242 dout(30) << __func__
<< " key " << pretty_binary_string(it
->key()) << dendl
;
7243 if (is_extent_shard_key(it
->key())) {
7248 int r
= get_key_object(it
->key(), &oid
);
7250 dout(20) << __func__
<< " oid " << oid
<< " end " << end
<< dendl
;
7251 if (ls
->size() >= (unsigned)max
) {
7252 dout(20) << __func__
<< " reached max " << max
<< dendl
;
7262 *pnext
= ghobject_t::get_max();
7268 int BlueStore::omap_get(
7269 const coll_t
& cid
, ///< [in] Collection containing oid
7270 const ghobject_t
&oid
, ///< [in] Object containing omap
7271 bufferlist
*header
, ///< [out] omap header
7272 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7275 CollectionHandle c
= _get_collection(cid
);
7278 return omap_get(c
, oid
, header
, out
);
7281 int BlueStore::omap_get(
7282 CollectionHandle
&c_
, ///< [in] Collection containing oid
7283 const ghobject_t
&oid
, ///< [in] Object containing omap
7284 bufferlist
*header
, ///< [out] omap header
7285 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7288 Collection
*c
= static_cast<Collection
*>(c_
.get());
7289 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7292 RWLock::RLocker
l(c
->lock
);
7294 OnodeRef o
= c
->get_onode(oid
, false);
7295 if (!o
|| !o
->exists
) {
7299 if (!o
->onode
.has_omap())
7303 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7305 get_omap_header(o
->onode
.nid
, &head
);
7306 get_omap_tail(o
->onode
.nid
, &tail
);
7307 it
->lower_bound(head
);
7308 while (it
->valid()) {
7309 if (it
->key() == head
) {
7310 dout(30) << __func__
<< " got header" << dendl
;
7311 *header
= it
->value();
7312 } else if (it
->key() >= tail
) {
7313 dout(30) << __func__
<< " reached tail" << dendl
;
7317 decode_omap_key(it
->key(), &user_key
);
7318 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7319 << " -> " << user_key
<< dendl
;
7320 (*out
)[user_key
] = it
->value();
7326 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7331 int BlueStore::omap_get_header(
7332 const coll_t
& cid
, ///< [in] Collection containing oid
7333 const ghobject_t
&oid
, ///< [in] Object containing omap
7334 bufferlist
*header
, ///< [out] omap header
7335 bool allow_eio
///< [in] don't assert on eio
7338 CollectionHandle c
= _get_collection(cid
);
7341 return omap_get_header(c
, oid
, header
, allow_eio
);
7344 int BlueStore::omap_get_header(
7345 CollectionHandle
&c_
, ///< [in] Collection containing oid
7346 const ghobject_t
&oid
, ///< [in] Object containing omap
7347 bufferlist
*header
, ///< [out] omap header
7348 bool allow_eio
///< [in] don't assert on eio
7351 Collection
*c
= static_cast<Collection
*>(c_
.get());
7352 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7355 RWLock::RLocker
l(c
->lock
);
7357 OnodeRef o
= c
->get_onode(oid
, false);
7358 if (!o
|| !o
->exists
) {
7362 if (!o
->onode
.has_omap())
7367 get_omap_header(o
->onode
.nid
, &head
);
7368 if (db
->get(PREFIX_OMAP
, head
, header
) >= 0) {
7369 dout(30) << __func__
<< " got header" << dendl
;
7371 dout(30) << __func__
<< " no header" << dendl
;
7375 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7380 int BlueStore::omap_get_keys(
7381 const coll_t
& cid
, ///< [in] Collection containing oid
7382 const ghobject_t
&oid
, ///< [in] Object containing omap
7383 set
<string
> *keys
///< [out] Keys defined on oid
7386 CollectionHandle c
= _get_collection(cid
);
7389 return omap_get_keys(c
, oid
, keys
);
7392 int BlueStore::omap_get_keys(
7393 CollectionHandle
&c_
, ///< [in] Collection containing oid
7394 const ghobject_t
&oid
, ///< [in] Object containing omap
7395 set
<string
> *keys
///< [out] Keys defined on oid
7398 Collection
*c
= static_cast<Collection
*>(c_
.get());
7399 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7402 RWLock::RLocker
l(c
->lock
);
7404 OnodeRef o
= c
->get_onode(oid
, false);
7405 if (!o
|| !o
->exists
) {
7409 if (!o
->onode
.has_omap())
7413 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7415 get_omap_key(o
->onode
.nid
, string(), &head
);
7416 get_omap_tail(o
->onode
.nid
, &tail
);
7417 it
->lower_bound(head
);
7418 while (it
->valid()) {
7419 if (it
->key() >= tail
) {
7420 dout(30) << __func__
<< " reached tail" << dendl
;
7424 decode_omap_key(it
->key(), &user_key
);
7425 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7426 << " -> " << user_key
<< dendl
;
7427 keys
->insert(user_key
);
7432 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7437 int BlueStore::omap_get_values(
7438 const coll_t
& cid
, ///< [in] Collection containing oid
7439 const ghobject_t
&oid
, ///< [in] Object containing omap
7440 const set
<string
> &keys
, ///< [in] Keys to get
7441 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7444 CollectionHandle c
= _get_collection(cid
);
7447 return omap_get_values(c
, oid
, keys
, out
);
7450 int BlueStore::omap_get_values(
7451 CollectionHandle
&c_
, ///< [in] Collection containing oid
7452 const ghobject_t
&oid
, ///< [in] Object containing omap
7453 const set
<string
> &keys
, ///< [in] Keys to get
7454 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7457 Collection
*c
= static_cast<Collection
*>(c_
.get());
7458 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7461 RWLock::RLocker
l(c
->lock
);
7464 OnodeRef o
= c
->get_onode(oid
, false);
7465 if (!o
|| !o
->exists
) {
7469 if (!o
->onode
.has_omap())
7472 _key_encode_u64(o
->onode
.nid
, &final_key
);
7473 final_key
.push_back('.');
7474 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7475 final_key
.resize(9); // keep prefix
7478 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7479 dout(30) << __func__
<< " got " << pretty_binary_string(final_key
)
7480 << " -> " << *p
<< dendl
;
7481 out
->insert(make_pair(*p
, val
));
7485 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7490 int BlueStore::omap_check_keys(
7491 const coll_t
& cid
, ///< [in] Collection containing oid
7492 const ghobject_t
&oid
, ///< [in] Object containing omap
7493 const set
<string
> &keys
, ///< [in] Keys to check
7494 set
<string
> *out
///< [out] Subset of keys defined on oid
7497 CollectionHandle c
= _get_collection(cid
);
7500 return omap_check_keys(c
, oid
, keys
, out
);
7503 int BlueStore::omap_check_keys(
7504 CollectionHandle
&c_
, ///< [in] Collection containing oid
7505 const ghobject_t
&oid
, ///< [in] Object containing omap
7506 const set
<string
> &keys
, ///< [in] Keys to check
7507 set
<string
> *out
///< [out] Subset of keys defined on oid
7510 Collection
*c
= static_cast<Collection
*>(c_
.get());
7511 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7514 RWLock::RLocker
l(c
->lock
);
7517 OnodeRef o
= c
->get_onode(oid
, false);
7518 if (!o
|| !o
->exists
) {
7522 if (!o
->onode
.has_omap())
7525 _key_encode_u64(o
->onode
.nid
, &final_key
);
7526 final_key
.push_back('.');
7527 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7528 final_key
.resize(9); // keep prefix
7531 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7532 dout(30) << __func__
<< " have " << pretty_binary_string(final_key
)
7533 << " -> " << *p
<< dendl
;
7536 dout(30) << __func__
<< " miss " << pretty_binary_string(final_key
)
7537 << " -> " << *p
<< dendl
;
7541 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7546 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7547 const coll_t
& cid
, ///< [in] collection
7548 const ghobject_t
&oid
///< [in] object
7551 CollectionHandle c
= _get_collection(cid
);
7553 dout(10) << __func__
<< " " << cid
<< "doesn't exist" <<dendl
;
7554 return ObjectMap::ObjectMapIterator();
7556 return get_omap_iterator(c
, oid
);
7559 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7560 CollectionHandle
&c_
, ///< [in] collection
7561 const ghobject_t
&oid
///< [in] object
7564 Collection
*c
= static_cast<Collection
*>(c_
.get());
7565 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
7567 return ObjectMap::ObjectMapIterator();
7569 RWLock::RLocker
l(c
->lock
);
7570 OnodeRef o
= c
->get_onode(oid
, false);
7571 if (!o
|| !o
->exists
) {
7572 dout(10) << __func__
<< " " << oid
<< "doesn't exist" <<dendl
;
7573 return ObjectMap::ObjectMapIterator();
7576 dout(10) << __func__
<< " has_omap = " << (int)o
->onode
.has_omap() <<dendl
;
7577 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7578 return ObjectMap::ObjectMapIterator(new OmapIteratorImpl(c
, o
, it
));
7581 // -----------------
7584 void BlueStore::_prepare_ondisk_format_super(KeyValueDB::Transaction
& t
)
7586 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7587 << " min_compat_ondisk_format " << min_compat_ondisk_format
7589 assert(ondisk_format
== latest_ondisk_format
);
7592 ::encode(ondisk_format
, bl
);
7593 t
->set(PREFIX_SUPER
, "ondisk_format", bl
);
7597 ::encode(min_compat_ondisk_format
, bl
);
7598 t
->set(PREFIX_SUPER
, "min_compat_ondisk_format", bl
);
7602 int BlueStore::_open_super_meta()
7608 db
->get(PREFIX_SUPER
, "nid_max", &bl
);
7609 bufferlist::iterator p
= bl
.begin();
7614 } catch (buffer::error
& e
) {
7615 derr
<< __func__
<< " unable to read nid_max" << dendl
;
7618 dout(10) << __func__
<< " old nid_max " << nid_max
<< dendl
;
7619 nid_last
= nid_max
.load();
7626 db
->get(PREFIX_SUPER
, "blobid_max", &bl
);
7627 bufferlist::iterator p
= bl
.begin();
7632 } catch (buffer::error
& e
) {
7633 derr
<< __func__
<< " unable to read blobid_max" << dendl
;
7636 dout(10) << __func__
<< " old blobid_max " << blobid_max
<< dendl
;
7637 blobid_last
= blobid_max
.load();
7643 db
->get(PREFIX_SUPER
, "freelist_type", &bl
);
7645 freelist_type
= std::string(bl
.c_str(), bl
.length());
7646 dout(10) << __func__
<< " freelist_type " << freelist_type
<< dendl
;
7648 assert("Not Support extent freelist manager" == 0);
7653 if (cct
->_conf
->bluestore_bluefs
) {
7654 bluefs_extents
.clear();
7656 db
->get(PREFIX_SUPER
, "bluefs_extents", &bl
);
7657 bufferlist::iterator p
= bl
.begin();
7659 ::decode(bluefs_extents
, p
);
7661 catch (buffer::error
& e
) {
7662 derr
<< __func__
<< " unable to read bluefs_extents" << dendl
;
7665 dout(10) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
7666 << std::dec
<< dendl
;
7670 int32_t compat_ondisk_format
= 0;
7673 int r
= db
->get(PREFIX_SUPER
, "ondisk_format", &bl
);
7675 // base case: kraken bluestore is v1 and readable by v1
7676 dout(20) << __func__
<< " missing ondisk_format; assuming kraken"
7679 compat_ondisk_format
= 1;
7681 auto p
= bl
.begin();
7683 ::decode(ondisk_format
, p
);
7684 } catch (buffer::error
& e
) {
7685 derr
<< __func__
<< " unable to read ondisk_format" << dendl
;
7690 r
= db
->get(PREFIX_SUPER
, "min_compat_ondisk_format", &bl
);
7692 auto p
= bl
.begin();
7694 ::decode(compat_ondisk_format
, p
);
7695 } catch (buffer::error
& e
) {
7696 derr
<< __func__
<< " unable to read compat_ondisk_format" << dendl
;
7701 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7702 << " compat_ondisk_format " << compat_ondisk_format
7706 if (latest_ondisk_format
< compat_ondisk_format
) {
7707 derr
<< __func__
<< " compat_ondisk_format is "
7708 << compat_ondisk_format
<< " but we only understand version "
7709 << latest_ondisk_format
<< dendl
;
7712 if (ondisk_format
< latest_ondisk_format
) {
7713 int r
= _upgrade_super();
7721 db
->get(PREFIX_SUPER
, "min_alloc_size", &bl
);
7722 auto p
= bl
.begin();
7726 min_alloc_size
= val
;
7727 min_alloc_size_order
= ctz(val
);
7728 assert(min_alloc_size
== 1u << min_alloc_size_order
);
7729 } catch (buffer::error
& e
) {
7730 derr
<< __func__
<< " unable to read min_alloc_size" << dendl
;
7733 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
7734 << std::dec
<< dendl
;
7738 _set_throttle_params();
7744 _set_finisher_num();
7749 int BlueStore::_upgrade_super()
7751 dout(1) << __func__
<< " from " << ondisk_format
<< ", latest "
7752 << latest_ondisk_format
<< dendl
;
7753 assert(ondisk_format
> 0);
7754 assert(ondisk_format
< latest_ondisk_format
);
7756 if (ondisk_format
== 1) {
7758 // - super: added ondisk_format
7759 // - super: added min_readable_ondisk_format
7760 // - super: added min_compat_ondisk_format
7761 // - super: added min_alloc_size
7762 // - super: removed min_min_alloc_size
7763 KeyValueDB::Transaction t
= db
->get_transaction();
7766 db
->get(PREFIX_SUPER
, "min_min_alloc_size", &bl
);
7767 auto p
= bl
.begin();
7771 min_alloc_size
= val
;
7772 } catch (buffer::error
& e
) {
7773 derr
<< __func__
<< " failed to read min_min_alloc_size" << dendl
;
7776 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
7777 t
->rmkey(PREFIX_SUPER
, "min_min_alloc_size");
7780 _prepare_ondisk_format_super(t
);
7781 int r
= db
->submit_transaction_sync(t
);
7786 dout(1) << __func__
<< " done" << dendl
;
7790 void BlueStore::_assign_nid(TransContext
*txc
, OnodeRef o
)
7796 uint64_t nid
= ++nid_last
;
7797 dout(20) << __func__
<< " " << nid
<< dendl
;
7799 txc
->last_nid
= nid
;
7803 uint64_t BlueStore::_assign_blobid(TransContext
*txc
)
7805 uint64_t bid
= ++blobid_last
;
7806 dout(20) << __func__
<< " " << bid
<< dendl
;
7807 txc
->last_blobid
= bid
;
7811 void BlueStore::get_db_statistics(Formatter
*f
)
7813 db
->get_statistics(f
);
7816 BlueStore::TransContext
*BlueStore::_txc_create(OpSequencer
*osr
)
7818 TransContext
*txc
= new TransContext(cct
, osr
);
7819 txc
->t
= db
->get_transaction();
7820 osr
->queue_new(txc
);
7821 dout(20) << __func__
<< " osr " << osr
<< " = " << txc
7822 << " seq " << txc
->seq
<< dendl
;
7826 void BlueStore::_txc_calc_cost(TransContext
*txc
)
7828 // this is about the simplest model for transaction cost you can
7829 // imagine. there is some fixed overhead cost by saying there is a
7830 // minimum of one "io". and then we have some cost per "io" that is
7831 // a configurable (with different hdd and ssd defaults), and add
7832 // that to the bytes value.
7833 int ios
= 1; // one "io" for the kv commit
7834 for (auto& p
: txc
->ioc
.pending_aios
) {
7835 ios
+= p
.iov
.size();
7837 auto cost
= throttle_cost_per_io
.load();
7838 txc
->cost
= ios
* cost
+ txc
->bytes
;
7839 dout(10) << __func__
<< " " << txc
<< " cost " << txc
->cost
<< " ("
7840 << ios
<< " ios * " << cost
<< " + " << txc
->bytes
7841 << " bytes)" << dendl
;
7844 void BlueStore::_txc_update_store_statfs(TransContext
*txc
)
7846 if (txc
->statfs_delta
.is_empty())
7849 logger
->inc(l_bluestore_allocated
, txc
->statfs_delta
.allocated());
7850 logger
->inc(l_bluestore_stored
, txc
->statfs_delta
.stored());
7851 logger
->inc(l_bluestore_compressed
, txc
->statfs_delta
.compressed());
7852 logger
->inc(l_bluestore_compressed_allocated
, txc
->statfs_delta
.compressed_allocated());
7853 logger
->inc(l_bluestore_compressed_original
, txc
->statfs_delta
.compressed_original());
7856 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
7857 vstatfs
+= txc
->statfs_delta
;
7861 txc
->statfs_delta
.encode(bl
);
7863 txc
->t
->merge(PREFIX_STAT
, "bluestore_statfs", bl
);
7864 txc
->statfs_delta
.reset();
7867 void BlueStore::_txc_state_proc(TransContext
*txc
)
7870 dout(10) << __func__
<< " txc " << txc
7871 << " " << txc
->get_state_name() << dendl
;
7872 switch (txc
->state
) {
7873 case TransContext::STATE_PREPARE
:
7874 txc
->log_state_latency(logger
, l_bluestore_state_prepare_lat
);
7875 if (txc
->ioc
.has_pending_aios()) {
7876 txc
->state
= TransContext::STATE_AIO_WAIT
;
7877 txc
->had_ios
= true;
7878 _txc_aio_submit(txc
);
7883 case TransContext::STATE_AIO_WAIT
:
7884 txc
->log_state_latency(logger
, l_bluestore_state_aio_wait_lat
);
7885 _txc_finish_io(txc
); // may trigger blocked txc's too
7888 case TransContext::STATE_IO_DONE
:
7889 //assert(txc->osr->qlock.is_locked()); // see _txc_finish_io
7891 ++txc
->osr
->txc_with_unstable_io
;
7893 txc
->log_state_latency(logger
, l_bluestore_state_io_done_lat
);
7894 txc
->state
= TransContext::STATE_KV_QUEUED
;
7895 if (cct
->_conf
->bluestore_sync_submit_transaction
) {
7896 if (txc
->last_nid
>= nid_max
||
7897 txc
->last_blobid
>= blobid_max
) {
7898 dout(20) << __func__
7899 << " last_{nid,blobid} exceeds max, submit via kv thread"
7901 } else if (txc
->osr
->kv_committing_serially
) {
7902 dout(20) << __func__
<< " prior txc submitted via kv thread, us too"
7904 // note: this is starvation-prone. once we have a txc in a busy
7905 // sequencer that is committing serially it is possible to keep
7906 // submitting new transactions fast enough that we get stuck doing
7907 // so. the alternative is to block here... fixme?
7908 } else if (txc
->osr
->txc_with_unstable_io
) {
7909 dout(20) << __func__
<< " prior txc(s) with unstable ios "
7910 << txc
->osr
->txc_with_unstable_io
.load() << dendl
;
7911 } else if (cct
->_conf
->bluestore_debug_randomize_serial_transaction
&&
7912 rand() % cct
->_conf
->bluestore_debug_randomize_serial_transaction
7914 dout(20) << __func__
<< " DEBUG randomly forcing submit via kv thread"
7917 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
7918 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
7920 _txc_applied_kv(txc
);
7924 std::lock_guard
<std::mutex
> l(kv_lock
);
7925 kv_queue
.push_back(txc
);
7926 kv_cond
.notify_one();
7927 if (txc
->state
!= TransContext::STATE_KV_SUBMITTED
) {
7928 kv_queue_unsubmitted
.push_back(txc
);
7929 ++txc
->osr
->kv_committing_serially
;
7933 kv_throttle_costs
+= txc
->cost
;
7936 case TransContext::STATE_KV_SUBMITTED
:
7937 _txc_committed_kv(txc
);
7940 case TransContext::STATE_KV_DONE
:
7941 txc
->log_state_latency(logger
, l_bluestore_state_kv_done_lat
);
7942 if (txc
->deferred_txn
) {
7943 txc
->state
= TransContext::STATE_DEFERRED_QUEUED
;
7944 _deferred_queue(txc
);
7947 txc
->state
= TransContext::STATE_FINISHING
;
7950 case TransContext::STATE_DEFERRED_CLEANUP
:
7951 txc
->log_state_latency(logger
, l_bluestore_state_deferred_cleanup_lat
);
7952 txc
->state
= TransContext::STATE_FINISHING
;
7955 case TransContext::STATE_FINISHING
:
7956 txc
->log_state_latency(logger
, l_bluestore_state_finishing_lat
);
7961 derr
<< __func__
<< " unexpected txc " << txc
7962 << " state " << txc
->get_state_name() << dendl
;
7963 assert(0 == "unexpected txc state");
7969 void BlueStore::_txc_finish_io(TransContext
*txc
)
7971 dout(20) << __func__
<< " " << txc
<< dendl
;
7974 * we need to preserve the order of kv transactions,
7975 * even though aio will complete in any order.
7978 OpSequencer
*osr
= txc
->osr
.get();
7979 std::lock_guard
<std::mutex
> l(osr
->qlock
);
7980 txc
->state
= TransContext::STATE_IO_DONE
;
7982 // release aio contexts (including pinned buffers).
7983 txc
->ioc
.running_aios
.clear();
7985 OpSequencer::q_list_t::iterator p
= osr
->q
.iterator_to(*txc
);
7986 while (p
!= osr
->q
.begin()) {
7988 if (p
->state
< TransContext::STATE_IO_DONE
) {
7989 dout(20) << __func__
<< " " << txc
<< " blocked by " << &*p
<< " "
7990 << p
->get_state_name() << dendl
;
7993 if (p
->state
> TransContext::STATE_IO_DONE
) {
7999 _txc_state_proc(&*p
++);
8000 } while (p
!= osr
->q
.end() &&
8001 p
->state
== TransContext::STATE_IO_DONE
);
8003 if (osr
->kv_submitted_waiters
&&
8004 osr
->_is_all_kv_submitted()) {
8005 osr
->qcond
.notify_all();
8009 void BlueStore::_txc_write_nodes(TransContext
*txc
, KeyValueDB::Transaction t
)
8011 dout(20) << __func__
<< " txc " << txc
8012 << " onodes " << txc
->onodes
8013 << " shared_blobs " << txc
->shared_blobs
8017 for (auto o
: txc
->onodes
) {
8018 // finalize extent_map shards
8019 o
->extent_map
.update(t
, false);
8020 if (o
->extent_map
.needs_reshard()) {
8021 o
->extent_map
.reshard(db
, t
);
8022 o
->extent_map
.update(t
, true);
8023 if (o
->extent_map
.needs_reshard()) {
8024 dout(20) << __func__
<< " warning: still wants reshard, check options?"
8026 o
->extent_map
.clear_needs_reshard();
8028 logger
->inc(l_bluestore_onode_reshard
);
8033 denc(o
->onode
, bound
);
8034 o
->extent_map
.bound_encode_spanning_blobs(bound
);
8035 if (o
->onode
.extent_map_shards
.empty()) {
8036 denc(o
->extent_map
.inline_bl
, bound
);
8041 unsigned onode_part
, blob_part
, extent_part
;
8043 auto p
= bl
.get_contiguous_appender(bound
, true);
8045 onode_part
= p
.get_logical_offset();
8046 o
->extent_map
.encode_spanning_blobs(p
);
8047 blob_part
= p
.get_logical_offset() - onode_part
;
8048 if (o
->onode
.extent_map_shards
.empty()) {
8049 denc(o
->extent_map
.inline_bl
, p
);
8051 extent_part
= p
.get_logical_offset() - onode_part
- blob_part
;
8054 dout(20) << " onode " << o
->oid
<< " is " << bl
.length()
8055 << " (" << onode_part
<< " bytes onode + "
8056 << blob_part
<< " bytes spanning blobs + "
8057 << extent_part
<< " bytes inline extents)"
8059 t
->set(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size(), bl
);
8060 o
->flushing_count
++;
8063 // objects we modified but didn't affect the onode
8064 auto p
= txc
->modified_objects
.begin();
8065 while (p
!= txc
->modified_objects
.end()) {
8066 if (txc
->onodes
.count(*p
) == 0) {
8067 (*p
)->flushing_count
++;
8070 // remove dups with onodes list to avoid problems in _txc_finish
8071 p
= txc
->modified_objects
.erase(p
);
8075 // finalize shared_blobs
8076 for (auto sb
: txc
->shared_blobs
) {
8078 auto sbid
= sb
->get_sbid();
8079 get_shared_blob_key(sbid
, &key
);
8080 if (sb
->persistent
->empty()) {
8081 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
8082 << " is empty" << dendl
;
8083 t
->rmkey(PREFIX_SHARED_BLOB
, key
);
8086 ::encode(*(sb
->persistent
), bl
);
8087 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
8088 << " is " << bl
.length() << " " << *sb
<< dendl
;
8089 t
->set(PREFIX_SHARED_BLOB
, key
, bl
);
8094 void BlueStore::BSPerfTracker::update_from_perfcounters(
8095 PerfCounters
&logger
)
8097 os_commit_latency
.consume_next(
8099 l_bluestore_commit_lat
));
8100 os_apply_latency
.consume_next(
8102 l_bluestore_commit_lat
));
8105 void BlueStore::_txc_finalize_kv(TransContext
*txc
, KeyValueDB::Transaction t
)
8107 dout(20) << __func__
<< " txc " << txc
<< std::hex
8108 << " allocated 0x" << txc
->allocated
8109 << " released 0x" << txc
->released
8110 << std::dec
<< dendl
;
8112 // We have to handle the case where we allocate *and* deallocate the
8113 // same region in this transaction. The freelist doesn't like that.
8114 // (Actually, the only thing that cares is the BitmapFreelistManager
8115 // debug check. But that's important.)
8116 interval_set
<uint64_t> tmp_allocated
, tmp_released
;
8117 interval_set
<uint64_t> *pallocated
= &txc
->allocated
;
8118 interval_set
<uint64_t> *preleased
= &txc
->released
;
8119 if (!txc
->allocated
.empty() && !txc
->released
.empty()) {
8120 interval_set
<uint64_t> overlap
;
8121 overlap
.intersection_of(txc
->allocated
, txc
->released
);
8122 if (!overlap
.empty()) {
8123 tmp_allocated
= txc
->allocated
;
8124 tmp_allocated
.subtract(overlap
);
8125 tmp_released
= txc
->released
;
8126 tmp_released
.subtract(overlap
);
8127 dout(20) << __func__
<< " overlap 0x" << std::hex
<< overlap
8128 << ", new allocated 0x" << tmp_allocated
8129 << " released 0x" << tmp_released
<< std::dec
8131 pallocated
= &tmp_allocated
;
8132 preleased
= &tmp_released
;
8136 // update freelist with non-overlap sets
8137 for (interval_set
<uint64_t>::iterator p
= pallocated
->begin();
8138 p
!= pallocated
->end();
8140 fm
->allocate(p
.get_start(), p
.get_len(), t
);
8142 for (interval_set
<uint64_t>::iterator p
= preleased
->begin();
8143 p
!= preleased
->end();
8145 dout(20) << __func__
<< " release 0x" << std::hex
<< p
.get_start()
8146 << "~" << p
.get_len() << std::dec
<< dendl
;
8147 fm
->release(p
.get_start(), p
.get_len(), t
);
8150 _txc_update_store_statfs(txc
);
8153 void BlueStore::_txc_applied_kv(TransContext
*txc
)
8155 for (auto ls
: { &txc
->onodes
, &txc
->modified_objects
}) {
8156 for (auto& o
: *ls
) {
8157 dout(20) << __func__
<< " onode " << o
<< " had " << o
->flushing_count
8159 if (--o
->flushing_count
== 0) {
8160 std::lock_guard
<std::mutex
> l(o
->flush_lock
);
8161 o
->flush_cond
.notify_all();
8167 void BlueStore::_txc_committed_kv(TransContext
*txc
)
8169 dout(20) << __func__
<< " txc " << txc
<< dendl
;
8171 // warning: we're calling onreadable_sync inside the sequencer lock
8172 if (txc
->onreadable_sync
) {
8173 txc
->onreadable_sync
->complete(0);
8174 txc
->onreadable_sync
= NULL
;
8176 unsigned n
= txc
->osr
->parent
->shard_hint
.hash_to_shard(m_finisher_num
);
8177 if (txc
->oncommit
) {
8178 logger
->tinc(l_bluestore_commit_lat
, ceph_clock_now() - txc
->start
);
8179 finishers
[n
]->queue(txc
->oncommit
);
8180 txc
->oncommit
= NULL
;
8182 if (txc
->onreadable
) {
8183 finishers
[n
]->queue(txc
->onreadable
);
8184 txc
->onreadable
= NULL
;
8188 std::lock_guard
<std::mutex
> l(txc
->osr
->qlock
);
8189 txc
->state
= TransContext::STATE_KV_DONE
;
8190 if (!txc
->oncommits
.empty()) {
8191 finishers
[n
]->queue(txc
->oncommits
);
8194 txc
->log_state_latency(logger
, l_bluestore_state_kv_committing_lat
);
8197 void BlueStore::_txc_finish(TransContext
*txc
)
8199 dout(20) << __func__
<< " " << txc
<< " onodes " << txc
->onodes
<< dendl
;
8200 assert(txc
->state
== TransContext::STATE_FINISHING
);
8202 for (auto& sb
: txc
->shared_blobs_written
) {
8203 sb
->bc
.finish_write(sb
->get_cache(), txc
->seq
);
8205 txc
->shared_blobs_written
.clear();
8207 while (!txc
->removed_collections
.empty()) {
8208 _queue_reap_collection(txc
->removed_collections
.front());
8209 txc
->removed_collections
.pop_front();
8212 OpSequencerRef osr
= txc
->osr
;
8214 bool submit_deferred
= false;
8215 OpSequencer::q_list_t releasing_txc
;
8217 std::lock_guard
<std::mutex
> l(osr
->qlock
);
8218 txc
->state
= TransContext::STATE_DONE
;
8219 bool notify
= false;
8220 while (!osr
->q
.empty()) {
8221 TransContext
*txc
= &osr
->q
.front();
8222 dout(20) << __func__
<< " txc " << txc
<< " " << txc
->get_state_name()
8224 if (txc
->state
!= TransContext::STATE_DONE
) {
8225 if (txc
->state
== TransContext::STATE_PREPARE
&&
8226 deferred_aggressive
) {
8227 // for _osr_drain_preceding()
8230 if (txc
->state
== TransContext::STATE_DEFERRED_QUEUED
&&
8231 osr
->q
.size() > g_conf
->bluestore_max_deferred_txc
) {
8232 submit_deferred
= true;
8238 releasing_txc
.push_back(*txc
);
8242 osr
->qcond
.notify_all();
8244 if (osr
->q
.empty()) {
8245 dout(20) << __func__
<< " osr " << osr
<< " q now empty" << dendl
;
8249 while (!releasing_txc
.empty()) {
8250 // release to allocator only after all preceding txc's have also
8251 // finished any deferred writes that potentially land in these
8253 auto txc
= &releasing_txc
.front();
8254 _txc_release_alloc(txc
);
8255 releasing_txc
.pop_front();
8256 txc
->log_state_latency(logger
, l_bluestore_state_done_lat
);
8260 if (submit_deferred
) {
8261 // we're pinning memory; flush! we could be more fine-grained here but
8262 // i'm not sure it's worth the bother.
8263 deferred_try_submit();
8266 if (empty
&& osr
->zombie
) {
8267 dout(10) << __func__
<< " reaping empty zombie osr " << osr
<< dendl
;
8272 void BlueStore::_txc_release_alloc(TransContext
*txc
)
8274 // update allocator with full released set
8275 if (!cct
->_conf
->bluestore_debug_no_reuse_blocks
) {
8276 dout(10) << __func__
<< " " << txc
<< " " << std::hex
8277 << txc
->released
<< std::dec
<< dendl
;
8278 for (interval_set
<uint64_t>::iterator p
= txc
->released
.begin();
8279 p
!= txc
->released
.end();
8281 alloc
->release(p
.get_start(), p
.get_len());
8285 txc
->allocated
.clear();
8286 txc
->released
.clear();
8289 void BlueStore::_osr_drain_preceding(TransContext
*txc
)
8291 OpSequencer
*osr
= txc
->osr
.get();
8292 dout(10) << __func__
<< " " << txc
<< " osr " << osr
<< dendl
;
8293 ++deferred_aggressive
; // FIXME: maybe osr-local aggressive flag?
8295 // submit anything pending
8296 deferred_lock
.lock();
8297 if (osr
->deferred_pending
) {
8298 _deferred_submit_unlock(osr
);
8300 deferred_lock
.unlock();
8304 // wake up any previously finished deferred events
8305 std::lock_guard
<std::mutex
> l(kv_lock
);
8306 kv_cond
.notify_one();
8308 osr
->drain_preceding(txc
);
8309 --deferred_aggressive
;
8310 dout(10) << __func__
<< " " << osr
<< " done" << dendl
;
8313 void BlueStore::_osr_drain_all()
8315 dout(10) << __func__
<< dendl
;
8317 set
<OpSequencerRef
> s
;
8319 std::lock_guard
<std::mutex
> l(osr_lock
);
8322 dout(20) << __func__
<< " osr_set " << s
<< dendl
;
8324 ++deferred_aggressive
;
8326 // submit anything pending
8327 deferred_try_submit();
8330 // wake up any previously finished deferred events
8331 std::lock_guard
<std::mutex
> l(kv_lock
);
8332 kv_cond
.notify_one();
8335 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8336 kv_finalize_cond
.notify_one();
8338 for (auto osr
: s
) {
8339 dout(20) << __func__
<< " drain " << osr
<< dendl
;
8342 --deferred_aggressive
;
8344 dout(10) << __func__
<< " done" << dendl
;
8347 void BlueStore::_osr_unregister_all()
8349 set
<OpSequencerRef
> s
;
8351 std::lock_guard
<std::mutex
> l(osr_lock
);
8354 dout(10) << __func__
<< " " << s
<< dendl
;
8355 for (auto osr
: s
) {
8359 // break link from Sequencer to us so that this OpSequencer
8360 // instance can die with this mount/umount cycle. note that
8361 // we assume umount() will not race against ~Sequencer.
8362 assert(osr
->parent
);
8363 osr
->parent
->p
.reset();
8366 // nobody should be creating sequencers during umount either.
8368 std::lock_guard
<std::mutex
> l(osr_lock
);
8369 assert(osr_set
.empty());
8373 void BlueStore::_kv_start()
8375 dout(10) << __func__
<< dendl
;
8377 for (int i
= 0; i
< m_finisher_num
; ++i
) {
8379 oss
<< "finisher-" << i
;
8380 Finisher
*f
= new Finisher(cct
, oss
.str(), "finisher");
8381 finishers
.push_back(f
);
8384 deferred_finisher
.start();
8385 for (auto f
: finishers
) {
8388 kv_sync_thread
.create("bstore_kv_sync");
8389 kv_finalize_thread
.create("bstore_kv_final");
8392 void BlueStore::_kv_stop()
8394 dout(10) << __func__
<< dendl
;
8396 std::unique_lock
<std::mutex
> l(kv_lock
);
8397 while (!kv_sync_started
) {
8401 kv_cond
.notify_all();
8404 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8405 while (!kv_finalize_started
) {
8406 kv_finalize_cond
.wait(l
);
8408 kv_finalize_stop
= true;
8409 kv_finalize_cond
.notify_all();
8411 kv_sync_thread
.join();
8412 kv_finalize_thread
.join();
8413 assert(removed_collections
.empty());
8415 std::lock_guard
<std::mutex
> l(kv_lock
);
8419 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8420 kv_finalize_stop
= false;
8422 dout(10) << __func__
<< " stopping finishers" << dendl
;
8423 deferred_finisher
.wait_for_empty();
8424 deferred_finisher
.stop();
8425 for (auto f
: finishers
) {
8426 f
->wait_for_empty();
8429 dout(10) << __func__
<< " stopped" << dendl
;
8432 void BlueStore::_kv_sync_thread()
8434 dout(10) << __func__
<< " start" << dendl
;
8435 std::unique_lock
<std::mutex
> l(kv_lock
);
8436 assert(!kv_sync_started
);
8437 kv_sync_started
= true;
8438 kv_cond
.notify_all();
8440 assert(kv_committing
.empty());
8441 if (kv_queue
.empty() &&
8442 ((deferred_done_queue
.empty() && deferred_stable_queue
.empty()) ||
8443 !deferred_aggressive
)) {
8446 dout(20) << __func__
<< " sleep" << dendl
;
8448 dout(20) << __func__
<< " wake" << dendl
;
8450 deque
<TransContext
*> kv_submitting
;
8451 deque
<DeferredBatch
*> deferred_done
, deferred_stable
;
8452 uint64_t aios
= 0, costs
= 0;
8454 dout(20) << __func__
<< " committing " << kv_queue
.size()
8455 << " submitting " << kv_queue_unsubmitted
.size()
8456 << " deferred done " << deferred_done_queue
.size()
8457 << " stable " << deferred_stable_queue
.size()
8459 kv_committing
.swap(kv_queue
);
8460 kv_submitting
.swap(kv_queue_unsubmitted
);
8461 deferred_done
.swap(deferred_done_queue
);
8462 deferred_stable
.swap(deferred_stable_queue
);
8464 costs
= kv_throttle_costs
;
8466 kv_throttle_costs
= 0;
8467 utime_t start
= ceph_clock_now();
8470 dout(30) << __func__
<< " committing " << kv_committing
<< dendl
;
8471 dout(30) << __func__
<< " submitting " << kv_submitting
<< dendl
;
8472 dout(30) << __func__
<< " deferred_done " << deferred_done
<< dendl
;
8473 dout(30) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8475 bool force_flush
= false;
8476 // if bluefs is sharing the same device as data (only), then we
8477 // can rely on the bluefs commit to flush the device and make
8478 // deferred aios stable. that means that if we do have done deferred
8479 // txcs AND we are not on a single device, we need to force a flush.
8480 if (bluefs_single_shared_device
&& bluefs
) {
8483 } else if (kv_committing
.empty() && kv_submitting
.empty() &&
8484 deferred_stable
.empty()) {
8485 force_flush
= true; // there's nothing else to commit!
8486 } else if (deferred_aggressive
) {
8493 dout(20) << __func__
<< " num_aios=" << aios
8494 << " force_flush=" << (int)force_flush
8495 << ", flushing, deferred done->stable" << dendl
;
8496 // flush/barrier on block device
8499 // if we flush then deferred done are now deferred stable
8500 deferred_stable
.insert(deferred_stable
.end(), deferred_done
.begin(),
8501 deferred_done
.end());
8502 deferred_done
.clear();
8504 utime_t after_flush
= ceph_clock_now();
8506 // we will use one final transaction to force a sync
8507 KeyValueDB::Transaction synct
= db
->get_transaction();
8509 // increase {nid,blobid}_max? note that this covers both the
8510 // case where we are approaching the max and the case we passed
8511 // it. in either case, we increase the max in the earlier txn
8513 uint64_t new_nid_max
= 0, new_blobid_max
= 0;
8514 if (nid_last
+ cct
->_conf
->bluestore_nid_prealloc
/2 > nid_max
) {
8515 KeyValueDB::Transaction t
=
8516 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8517 new_nid_max
= nid_last
+ cct
->_conf
->bluestore_nid_prealloc
;
8519 ::encode(new_nid_max
, bl
);
8520 t
->set(PREFIX_SUPER
, "nid_max", bl
);
8521 dout(10) << __func__
<< " new_nid_max " << new_nid_max
<< dendl
;
8523 if (blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
/2 > blobid_max
) {
8524 KeyValueDB::Transaction t
=
8525 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8526 new_blobid_max
= blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
;
8528 ::encode(new_blobid_max
, bl
);
8529 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
8530 dout(10) << __func__
<< " new_blobid_max " << new_blobid_max
<< dendl
;
8533 for (auto txc
: kv_committing
) {
8534 if (txc
->state
== TransContext::STATE_KV_QUEUED
) {
8535 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8536 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
8538 _txc_applied_kv(txc
);
8539 --txc
->osr
->kv_committing_serially
;
8540 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
8541 if (txc
->osr
->kv_submitted_waiters
) {
8542 std::lock_guard
<std::mutex
> l(txc
->osr
->qlock
);
8543 if (txc
->osr
->_is_all_kv_submitted()) {
8544 txc
->osr
->qcond
.notify_all();
8549 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8550 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8553 --txc
->osr
->txc_with_unstable_io
;
8557 // release throttle *before* we commit. this allows new ops
8558 // to be prepared and enter pipeline while we are waiting on
8559 // the kv commit sync/flush. then hopefully on the next
8560 // iteration there will already be ops awake. otherwise, we
8561 // end up going to sleep, and then wake up when the very first
8562 // transaction is ready for commit.
8563 throttle_bytes
.put(costs
);
8565 PExtentVector bluefs_gift_extents
;
8567 after_flush
- bluefs_last_balance
>
8568 cct
->_conf
->bluestore_bluefs_balance_interval
) {
8569 bluefs_last_balance
= after_flush
;
8570 int r
= _balance_bluefs_freespace(&bluefs_gift_extents
);
8573 for (auto& p
: bluefs_gift_extents
) {
8574 bluefs_extents
.insert(p
.offset
, p
.length
);
8577 ::encode(bluefs_extents
, bl
);
8578 dout(10) << __func__
<< " bluefs_extents now 0x" << std::hex
8579 << bluefs_extents
<< std::dec
<< dendl
;
8580 synct
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
8584 // cleanup sync deferred keys
8585 for (auto b
: deferred_stable
) {
8586 for (auto& txc
: b
->txcs
) {
8587 bluestore_deferred_transaction_t
& wt
= *txc
.deferred_txn
;
8588 if (!wt
.released
.empty()) {
8589 // kraken replay compat only
8590 txc
.released
= wt
.released
;
8591 dout(10) << __func__
<< " deferred txn has released "
8593 << " (we just upgraded from kraken) on " << &txc
<< dendl
;
8594 _txc_finalize_kv(&txc
, synct
);
8596 // cleanup the deferred
8598 get_deferred_key(wt
.seq
, &key
);
8599 synct
->rm_single_key(PREFIX_DEFERRED
, key
);
8603 // submit synct synchronously (block and wait for it to commit)
8604 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction_sync(synct
);
8608 nid_max
= new_nid_max
;
8609 dout(10) << __func__
<< " nid_max now " << nid_max
<< dendl
;
8611 if (new_blobid_max
) {
8612 blobid_max
= new_blobid_max
;
8613 dout(10) << __func__
<< " blobid_max now " << blobid_max
<< dendl
;
8617 utime_t finish
= ceph_clock_now();
8618 utime_t dur_flush
= after_flush
- start
;
8619 utime_t dur_kv
= finish
- after_flush
;
8620 utime_t dur
= finish
- start
;
8621 dout(20) << __func__
<< " committed " << kv_committing
.size()
8622 << " cleaned " << deferred_stable
.size()
8624 << " (" << dur_flush
<< " flush + " << dur_kv
<< " kv commit)"
8626 logger
->tinc(l_bluestore_kv_flush_lat
, dur_flush
);
8627 logger
->tinc(l_bluestore_kv_commit_lat
, dur_kv
);
8628 logger
->tinc(l_bluestore_kv_lat
, dur
);
8632 if (!bluefs_gift_extents
.empty()) {
8633 _commit_bluefs_freespace(bluefs_gift_extents
);
8635 for (auto p
= bluefs_extents_reclaiming
.begin();
8636 p
!= bluefs_extents_reclaiming
.end();
8638 dout(20) << __func__
<< " releasing old bluefs 0x" << std::hex
8639 << p
.get_start() << "~" << p
.get_len() << std::dec
8641 alloc
->release(p
.get_start(), p
.get_len());
8643 bluefs_extents_reclaiming
.clear();
8647 std::unique_lock
<std::mutex
> m(kv_finalize_lock
);
8648 if (kv_committing_to_finalize
.empty()) {
8649 kv_committing_to_finalize
.swap(kv_committing
);
8651 kv_committing_to_finalize
.insert(
8652 kv_committing_to_finalize
.end(),
8653 kv_committing
.begin(),
8654 kv_committing
.end());
8655 kv_committing
.clear();
8657 if (deferred_stable_to_finalize
.empty()) {
8658 deferred_stable_to_finalize
.swap(deferred_stable
);
8660 deferred_stable_to_finalize
.insert(
8661 deferred_stable_to_finalize
.end(),
8662 deferred_stable
.begin(),
8663 deferred_stable
.end());
8664 deferred_stable
.clear();
8666 kv_finalize_cond
.notify_one();
8670 // previously deferred "done" are now "stable" by virtue of this
8672 deferred_stable_queue
.swap(deferred_done
);
8675 dout(10) << __func__
<< " finish" << dendl
;
8676 kv_sync_started
= false;
8679 void BlueStore::_kv_finalize_thread()
8681 deque
<TransContext
*> kv_committed
;
8682 deque
<DeferredBatch
*> deferred_stable
;
8683 dout(10) << __func__
<< " start" << dendl
;
8684 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8685 assert(!kv_finalize_started
);
8686 kv_finalize_started
= true;
8687 kv_finalize_cond
.notify_all();
8689 assert(kv_committed
.empty());
8690 assert(deferred_stable
.empty());
8691 if (kv_committing_to_finalize
.empty() &&
8692 deferred_stable_to_finalize
.empty()) {
8693 if (kv_finalize_stop
)
8695 dout(20) << __func__
<< " sleep" << dendl
;
8696 kv_finalize_cond
.wait(l
);
8697 dout(20) << __func__
<< " wake" << dendl
;
8699 kv_committed
.swap(kv_committing_to_finalize
);
8700 deferred_stable
.swap(deferred_stable_to_finalize
);
8702 dout(20) << __func__
<< " kv_committed " << kv_committed
<< dendl
;
8703 dout(20) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8705 while (!kv_committed
.empty()) {
8706 TransContext
*txc
= kv_committed
.front();
8707 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8708 _txc_state_proc(txc
);
8709 kv_committed
.pop_front();
8712 for (auto b
: deferred_stable
) {
8713 auto p
= b
->txcs
.begin();
8714 while (p
!= b
->txcs
.end()) {
8715 TransContext
*txc
= &*p
;
8716 p
= b
->txcs
.erase(p
); // unlink here because
8717 _txc_state_proc(txc
); // this may destroy txc
8721 deferred_stable
.clear();
8723 if (!deferred_aggressive
) {
8724 if (deferred_queue_size
>= deferred_batch_ops
.load() ||
8725 throttle_deferred_bytes
.past_midpoint()) {
8726 deferred_try_submit();
8730 // this is as good a place as any ...
8731 _reap_collections();
8736 dout(10) << __func__
<< " finish" << dendl
;
8737 kv_finalize_started
= false;
8740 bluestore_deferred_op_t
*BlueStore::_get_deferred_op(
8741 TransContext
*txc
, OnodeRef o
)
8743 if (!txc
->deferred_txn
) {
8744 txc
->deferred_txn
= new bluestore_deferred_transaction_t
;
8746 txc
->deferred_txn
->ops
.push_back(bluestore_deferred_op_t());
8747 return &txc
->deferred_txn
->ops
.back();
8750 void BlueStore::_deferred_queue(TransContext
*txc
)
8752 dout(20) << __func__
<< " txc " << txc
<< " osr " << txc
->osr
<< dendl
;
8753 deferred_lock
.lock();
8754 if (!txc
->osr
->deferred_pending
&&
8755 !txc
->osr
->deferred_running
) {
8756 deferred_queue
.push_back(*txc
->osr
);
8758 if (!txc
->osr
->deferred_pending
) {
8759 txc
->osr
->deferred_pending
= new DeferredBatch(cct
, txc
->osr
.get());
8761 ++deferred_queue_size
;
8762 txc
->osr
->deferred_pending
->txcs
.push_back(*txc
);
8763 bluestore_deferred_transaction_t
& wt
= *txc
->deferred_txn
;
8764 for (auto opi
= wt
.ops
.begin(); opi
!= wt
.ops
.end(); ++opi
) {
8765 const auto& op
= *opi
;
8766 assert(op
.op
== bluestore_deferred_op_t::OP_WRITE
);
8767 bufferlist::const_iterator p
= op
.data
.begin();
8768 for (auto e
: op
.extents
) {
8769 txc
->osr
->deferred_pending
->prepare_write(
8770 cct
, wt
.seq
, e
.offset
, e
.length
, p
);
8773 if (deferred_aggressive
&&
8774 !txc
->osr
->deferred_running
) {
8775 _deferred_submit_unlock(txc
->osr
.get());
8777 deferred_lock
.unlock();
8781 void BlueStore::deferred_try_submit()
8783 dout(20) << __func__
<< " " << deferred_queue
.size() << " osrs, "
8784 << deferred_queue_size
<< " txcs" << dendl
;
8785 std::lock_guard
<std::mutex
> l(deferred_lock
);
8786 vector
<OpSequencerRef
> osrs
;
8787 osrs
.reserve(deferred_queue
.size());
8788 for (auto& osr
: deferred_queue
) {
8789 osrs
.push_back(&osr
);
8791 for (auto& osr
: osrs
) {
8792 if (osr
->deferred_pending
) {
8793 if (!osr
->deferred_running
) {
8794 _deferred_submit_unlock(osr
.get());
8795 deferred_lock
.lock();
8797 dout(20) << __func__
<< " osr " << osr
<< " already has running"
8801 dout(20) << __func__
<< " osr " << osr
<< " has no pending" << dendl
;
8806 void BlueStore::_deferred_submit_unlock(OpSequencer
*osr
)
8808 dout(10) << __func__
<< " osr " << osr
8809 << " " << osr
->deferred_pending
->iomap
.size() << " ios pending "
8811 assert(osr
->deferred_pending
);
8812 assert(!osr
->deferred_running
);
8814 auto b
= osr
->deferred_pending
;
8815 deferred_queue_size
-= b
->seq_bytes
.size();
8816 assert(deferred_queue_size
>= 0);
8818 osr
->deferred_running
= osr
->deferred_pending
;
8819 osr
->deferred_pending
= nullptr;
8821 uint64_t start
= 0, pos
= 0;
8823 auto i
= b
->iomap
.begin();
8825 if (i
== b
->iomap
.end() || i
->first
!= pos
) {
8827 dout(20) << __func__
<< " write 0x" << std::hex
8828 << start
<< "~" << bl
.length()
8829 << " crc " << bl
.crc32c(-1) << std::dec
<< dendl
;
8830 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
8831 logger
->inc(l_bluestore_deferred_write_ops
);
8832 logger
->inc(l_bluestore_deferred_write_bytes
, bl
.length());
8833 int r
= bdev
->aio_write(start
, bl
, &b
->ioc
, false);
8837 if (i
== b
->iomap
.end()) {
8844 dout(20) << __func__
<< " seq " << i
->second
.seq
<< " 0x"
8845 << std::hex
<< pos
<< "~" << i
->second
.bl
.length() << std::dec
8850 pos
+= i
->second
.bl
.length();
8851 bl
.claim_append(i
->second
.bl
);
8855 deferred_lock
.unlock();
8856 bdev
->aio_submit(&b
->ioc
);
8859 struct C_DeferredTrySubmit
: public Context
{
8861 C_DeferredTrySubmit(BlueStore
*s
) : store(s
) {}
8862 void finish(int r
) {
8863 store
->deferred_try_submit();
8867 void BlueStore::_deferred_aio_finish(OpSequencer
*osr
)
8869 dout(10) << __func__
<< " osr " << osr
<< dendl
;
8870 assert(osr
->deferred_running
);
8871 DeferredBatch
*b
= osr
->deferred_running
;
8874 std::lock_guard
<std::mutex
> l(deferred_lock
);
8875 assert(osr
->deferred_running
== b
);
8876 osr
->deferred_running
= nullptr;
8877 if (!osr
->deferred_pending
) {
8878 dout(20) << __func__
<< " dequeueing" << dendl
;
8879 auto q
= deferred_queue
.iterator_to(*osr
);
8880 deferred_queue
.erase(q
);
8881 } else if (deferred_aggressive
) {
8882 dout(20) << __func__
<< " queuing async deferred_try_submit" << dendl
;
8883 deferred_finisher
.queue(new C_DeferredTrySubmit(this));
8885 dout(20) << __func__
<< " leaving queued, more pending" << dendl
;
8891 std::lock_guard
<std::mutex
> l2(osr
->qlock
);
8892 for (auto& i
: b
->txcs
) {
8893 TransContext
*txc
= &i
;
8894 txc
->state
= TransContext::STATE_DEFERRED_CLEANUP
;
8897 osr
->qcond
.notify_all();
8898 throttle_deferred_bytes
.put(costs
);
8899 std::lock_guard
<std::mutex
> l(kv_lock
);
8900 deferred_done_queue
.emplace_back(b
);
8903 // in the normal case, do not bother waking up the kv thread; it will
8904 // catch us on the next commit anyway.
8905 if (deferred_aggressive
) {
8906 std::lock_guard
<std::mutex
> l(kv_lock
);
8907 kv_cond
.notify_one();
8911 int BlueStore::_deferred_replay()
8913 dout(10) << __func__
<< " start" << dendl
;
8914 OpSequencerRef osr
= new OpSequencer(cct
, this);
8917 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_DEFERRED
);
8918 for (it
->lower_bound(string()); it
->valid(); it
->next(), ++count
) {
8919 dout(20) << __func__
<< " replay " << pretty_binary_string(it
->key())
8921 bluestore_deferred_transaction_t
*deferred_txn
=
8922 new bluestore_deferred_transaction_t
;
8923 bufferlist bl
= it
->value();
8924 bufferlist::iterator p
= bl
.begin();
8926 ::decode(*deferred_txn
, p
);
8927 } catch (buffer::error
& e
) {
8928 derr
<< __func__
<< " failed to decode deferred txn "
8929 << pretty_binary_string(it
->key()) << dendl
;
8930 delete deferred_txn
;
8934 TransContext
*txc
= _txc_create(osr
.get());
8935 txc
->deferred_txn
= deferred_txn
;
8936 txc
->state
= TransContext::STATE_KV_DONE
;
8937 _txc_state_proc(txc
);
8940 dout(20) << __func__
<< " draining osr" << dendl
;
8943 dout(10) << __func__
<< " completed " << count
<< " events" << dendl
;
8947 // ---------------------------
8950 int BlueStore::queue_transactions(
8952 vector
<Transaction
>& tls
,
8954 ThreadPool::TPHandle
*handle
)
8957 Context
*onreadable
;
8959 Context
*onreadable_sync
;
8960 ObjectStore::Transaction::collect_contexts(
8961 tls
, &onreadable
, &ondisk
, &onreadable_sync
);
8963 if (cct
->_conf
->objectstore_blackhole
) {
8964 dout(0) << __func__
<< " objectstore_blackhole = TRUE, dropping transaction"
8968 delete onreadable_sync
;
8971 utime_t start
= ceph_clock_now();
8972 // set up the sequencer
8976 osr
= static_cast<OpSequencer
*>(posr
->p
.get());
8977 dout(10) << __func__
<< " existing " << osr
<< " " << *osr
<< dendl
;
8979 osr
= new OpSequencer(cct
, this);
8982 dout(10) << __func__
<< " new " << osr
<< " " << *osr
<< dendl
;
8986 TransContext
*txc
= _txc_create(osr
);
8987 txc
->onreadable
= onreadable
;
8988 txc
->onreadable_sync
= onreadable_sync
;
8989 txc
->oncommit
= ondisk
;
8991 for (vector
<Transaction
>::iterator p
= tls
.begin(); p
!= tls
.end(); ++p
) {
8993 txc
->bytes
+= (*p
).get_num_bytes();
8994 _txc_add_transaction(txc
, &(*p
));
8996 _txc_calc_cost(txc
);
8998 _txc_write_nodes(txc
, txc
->t
);
9000 // journal deferred items
9001 if (txc
->deferred_txn
) {
9002 txc
->deferred_txn
->seq
= ++deferred_seq
;
9004 ::encode(*txc
->deferred_txn
, bl
);
9006 get_deferred_key(txc
->deferred_txn
->seq
, &key
);
9007 txc
->t
->set(PREFIX_DEFERRED
, key
, bl
);
9010 _txc_finalize_kv(txc
, txc
->t
);
9012 handle
->suspend_tp_timeout();
9014 utime_t tstart
= ceph_clock_now();
9015 throttle_bytes
.get(txc
->cost
);
9016 if (txc
->deferred_txn
) {
9017 // ensure we do not block here because of deferred writes
9018 if (!throttle_deferred_bytes
.get_or_fail(txc
->cost
)) {
9019 dout(10) << __func__
<< " failed get throttle_deferred_bytes, aggressive"
9021 ++deferred_aggressive
;
9022 deferred_try_submit();
9024 // wake up any previously finished deferred events
9025 std::lock_guard
<std::mutex
> l(kv_lock
);
9026 kv_cond
.notify_one();
9028 throttle_deferred_bytes
.get(txc
->cost
);
9029 --deferred_aggressive
;
9032 utime_t tend
= ceph_clock_now();
9035 handle
->reset_tp_timeout();
9037 logger
->inc(l_bluestore_txc
);
9040 _txc_state_proc(txc
);
9042 logger
->tinc(l_bluestore_submit_lat
, ceph_clock_now() - start
);
9043 logger
->tinc(l_bluestore_throttle_lat
, tend
- tstart
);
9047 void BlueStore::_txc_aio_submit(TransContext
*txc
)
9049 dout(10) << __func__
<< " txc " << txc
<< dendl
;
9050 bdev
->aio_submit(&txc
->ioc
);
9053 void BlueStore::_txc_add_transaction(TransContext
*txc
, Transaction
*t
)
9055 Transaction::iterator i
= t
->begin();
9057 _dump_transaction(t
);
9059 vector
<CollectionRef
> cvec(i
.colls
.size());
9061 for (vector
<coll_t
>::iterator p
= i
.colls
.begin(); p
!= i
.colls
.end();
9063 cvec
[j
] = _get_collection(*p
);
9065 vector
<OnodeRef
> ovec(i
.objects
.size());
9067 for (int pos
= 0; i
.have_op(); ++pos
) {
9068 Transaction::Op
*op
= i
.decode_op();
9072 if (op
->op
== Transaction::OP_NOP
)
9075 // collection operations
9076 CollectionRef
&c
= cvec
[op
->cid
];
9078 case Transaction::OP_RMCOLL
:
9080 const coll_t
&cid
= i
.get_cid(op
->cid
);
9081 r
= _remove_collection(txc
, cid
, &c
);
9087 case Transaction::OP_MKCOLL
:
9090 const coll_t
&cid
= i
.get_cid(op
->cid
);
9091 r
= _create_collection(txc
, cid
, op
->split_bits
, &c
);
9097 case Transaction::OP_SPLIT_COLLECTION
:
9098 assert(0 == "deprecated");
9101 case Transaction::OP_SPLIT_COLLECTION2
:
9103 uint32_t bits
= op
->split_bits
;
9104 uint32_t rem
= op
->split_rem
;
9105 r
= _split_collection(txc
, c
, cvec
[op
->dest_cid
], bits
, rem
);
9111 case Transaction::OP_COLL_HINT
:
9113 uint32_t type
= op
->hint_type
;
9116 bufferlist::iterator hiter
= hint
.begin();
9117 if (type
== Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS
) {
9120 ::decode(pg_num
, hiter
);
9121 ::decode(num_objs
, hiter
);
9122 dout(10) << __func__
<< " collection hint objects is a no-op, "
9123 << " pg_num " << pg_num
<< " num_objects " << num_objs
9127 dout(10) << __func__
<< " unknown collection hint " << type
<< dendl
;
9133 case Transaction::OP_COLL_SETATTR
:
9137 case Transaction::OP_COLL_RMATTR
:
9141 case Transaction::OP_COLL_RENAME
:
9142 assert(0 == "not implemented");
9146 derr
<< __func__
<< " error " << cpp_strerror(r
)
9147 << " not handled on operation " << op
->op
9148 << " (op " << pos
<< ", counting from 0)" << dendl
;
9149 _dump_transaction(t
, 0);
9150 assert(0 == "unexpected error");
9153 // these operations implicity create the object
9154 bool create
= false;
9155 if (op
->op
== Transaction::OP_TOUCH
||
9156 op
->op
== Transaction::OP_WRITE
||
9157 op
->op
== Transaction::OP_ZERO
) {
9161 // object operations
9162 RWLock::WLocker
l(c
->lock
);
9163 OnodeRef
&o
= ovec
[op
->oid
];
9165 ghobject_t oid
= i
.get_oid(op
->oid
);
9166 o
= c
->get_onode(oid
, create
);
9168 if (!create
&& (!o
|| !o
->exists
)) {
9169 dout(10) << __func__
<< " op " << op
->op
<< " got ENOENT on "
9170 << i
.get_oid(op
->oid
) << dendl
;
9176 case Transaction::OP_TOUCH
:
9177 r
= _touch(txc
, c
, o
);
9180 case Transaction::OP_WRITE
:
9182 uint64_t off
= op
->off
;
9183 uint64_t len
= op
->len
;
9184 uint32_t fadvise_flags
= i
.get_fadvise_flags();
9187 r
= _write(txc
, c
, o
, off
, len
, bl
, fadvise_flags
);
9191 case Transaction::OP_ZERO
:
9193 uint64_t off
= op
->off
;
9194 uint64_t len
= op
->len
;
9195 r
= _zero(txc
, c
, o
, off
, len
);
9199 case Transaction::OP_TRIMCACHE
:
9201 // deprecated, no-op
9205 case Transaction::OP_TRUNCATE
:
9207 uint64_t off
= op
->off
;
9208 r
= _truncate(txc
, c
, o
, off
);
9212 case Transaction::OP_REMOVE
:
9214 r
= _remove(txc
, c
, o
);
9218 case Transaction::OP_SETATTR
:
9220 string name
= i
.decode_string();
9223 r
= _setattr(txc
, c
, o
, name
, bp
);
9227 case Transaction::OP_SETATTRS
:
9229 map
<string
, bufferptr
> aset
;
9230 i
.decode_attrset(aset
);
9231 r
= _setattrs(txc
, c
, o
, aset
);
9235 case Transaction::OP_RMATTR
:
9237 string name
= i
.decode_string();
9238 r
= _rmattr(txc
, c
, o
, name
);
9242 case Transaction::OP_RMATTRS
:
9244 r
= _rmattrs(txc
, c
, o
);
9248 case Transaction::OP_CLONE
:
9250 OnodeRef
& no
= ovec
[op
->dest_oid
];
9252 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9253 no
= c
->get_onode(noid
, true);
9255 r
= _clone(txc
, c
, o
, no
);
9259 case Transaction::OP_CLONERANGE
:
9260 assert(0 == "deprecated");
9263 case Transaction::OP_CLONERANGE2
:
9265 OnodeRef
& no
= ovec
[op
->dest_oid
];
9267 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9268 no
= c
->get_onode(noid
, true);
9270 uint64_t srcoff
= op
->off
;
9271 uint64_t len
= op
->len
;
9272 uint64_t dstoff
= op
->dest_off
;
9273 r
= _clone_range(txc
, c
, o
, no
, srcoff
, len
, dstoff
);
9277 case Transaction::OP_COLL_ADD
:
9278 assert(0 == "not implemented");
9281 case Transaction::OP_COLL_REMOVE
:
9282 assert(0 == "not implemented");
9285 case Transaction::OP_COLL_MOVE
:
9286 assert(0 == "deprecated");
9289 case Transaction::OP_COLL_MOVE_RENAME
:
9290 case Transaction::OP_TRY_RENAME
:
9292 assert(op
->cid
== op
->dest_cid
);
9293 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9294 OnodeRef
& no
= ovec
[op
->dest_oid
];
9296 no
= c
->get_onode(noid
, false);
9298 r
= _rename(txc
, c
, o
, no
, noid
);
9302 case Transaction::OP_OMAP_CLEAR
:
9304 r
= _omap_clear(txc
, c
, o
);
9307 case Transaction::OP_OMAP_SETKEYS
:
9310 i
.decode_attrset_bl(&aset_bl
);
9311 r
= _omap_setkeys(txc
, c
, o
, aset_bl
);
9314 case Transaction::OP_OMAP_RMKEYS
:
9317 i
.decode_keyset_bl(&keys_bl
);
9318 r
= _omap_rmkeys(txc
, c
, o
, keys_bl
);
9321 case Transaction::OP_OMAP_RMKEYRANGE
:
9324 first
= i
.decode_string();
9325 last
= i
.decode_string();
9326 r
= _omap_rmkey_range(txc
, c
, o
, first
, last
);
9329 case Transaction::OP_OMAP_SETHEADER
:
9333 r
= _omap_setheader(txc
, c
, o
, bl
);
9337 case Transaction::OP_SETALLOCHINT
:
9339 r
= _set_alloc_hint(txc
, c
, o
,
9340 op
->expected_object_size
,
9341 op
->expected_write_size
,
9342 op
->alloc_hint_flags
);
9347 derr
<< __func__
<< "bad op " << op
->op
<< dendl
;
9355 if (r
== -ENOENT
&& !(op
->op
== Transaction::OP_CLONERANGE
||
9356 op
->op
== Transaction::OP_CLONE
||
9357 op
->op
== Transaction::OP_CLONERANGE2
||
9358 op
->op
== Transaction::OP_COLL_ADD
||
9359 op
->op
== Transaction::OP_SETATTR
||
9360 op
->op
== Transaction::OP_SETATTRS
||
9361 op
->op
== Transaction::OP_RMATTR
||
9362 op
->op
== Transaction::OP_OMAP_SETKEYS
||
9363 op
->op
== Transaction::OP_OMAP_RMKEYS
||
9364 op
->op
== Transaction::OP_OMAP_RMKEYRANGE
||
9365 op
->op
== Transaction::OP_OMAP_SETHEADER
))
9366 // -ENOENT is usually okay
9372 const char *msg
= "unexpected error code";
9374 if (r
== -ENOENT
&& (op
->op
== Transaction::OP_CLONERANGE
||
9375 op
->op
== Transaction::OP_CLONE
||
9376 op
->op
== Transaction::OP_CLONERANGE2
))
9377 msg
= "ENOENT on clone suggests osd bug";
9380 // For now, if we hit _any_ ENOSPC, crash, before we do any damage
9381 // by partially applying transactions.
9382 msg
= "ENOSPC from bluestore, misconfigured cluster";
9384 if (r
== -ENOTEMPTY
) {
9385 msg
= "ENOTEMPTY suggests garbage data in osd data dir";
9388 derr
<< __func__
<< " error " << cpp_strerror(r
)
9389 << " not handled on operation " << op
->op
9390 << " (op " << pos
<< ", counting from 0)"
9392 derr
<< msg
<< dendl
;
9393 _dump_transaction(t
, 0);
9394 assert(0 == "unexpected error");
9402 // -----------------
9405 int BlueStore::_touch(TransContext
*txc
,
9409 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
9411 _assign_nid(txc
, o
);
9412 txc
->write_onode(o
);
9413 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
9417 void BlueStore::_dump_onode(const OnodeRef
& o
, int log_level
)
9419 if (!cct
->_conf
->subsys
.should_gather(ceph_subsys_bluestore
, log_level
))
9421 dout(log_level
) << __func__
<< " " << o
<< " " << o
->oid
9422 << " nid " << o
->onode
.nid
9423 << " size 0x" << std::hex
<< o
->onode
.size
9424 << " (" << std::dec
<< o
->onode
.size
<< ")"
9425 << " expected_object_size " << o
->onode
.expected_object_size
9426 << " expected_write_size " << o
->onode
.expected_write_size
9427 << " in " << o
->onode
.extent_map_shards
.size() << " shards"
9428 << ", " << o
->extent_map
.spanning_blob_map
.size()
9429 << " spanning blobs"
9431 for (auto p
= o
->onode
.attrs
.begin();
9432 p
!= o
->onode
.attrs
.end();
9434 dout(log_level
) << __func__
<< " attr " << p
->first
9435 << " len " << p
->second
.length() << dendl
;
9437 _dump_extent_map(o
->extent_map
, log_level
);
9440 void BlueStore::_dump_extent_map(ExtentMap
&em
, int log_level
)
9443 for (auto& s
: em
.shards
) {
9444 dout(log_level
) << __func__
<< " shard " << *s
.shard_info
9445 << (s
.loaded
? " (loaded)" : "")
9446 << (s
.dirty
? " (dirty)" : "")
9449 for (auto& e
: em
.extent_map
) {
9450 dout(log_level
) << __func__
<< " " << e
<< dendl
;
9451 assert(e
.logical_offset
>= pos
);
9452 pos
= e
.logical_offset
+ e
.length
;
9453 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
9454 if (blob
.has_csum()) {
9456 unsigned n
= blob
.get_csum_count();
9457 for (unsigned i
= 0; i
< n
; ++i
)
9458 v
.push_back(blob
.get_csum_item(i
));
9459 dout(log_level
) << __func__
<< " csum: " << std::hex
<< v
<< std::dec
9462 std::lock_guard
<std::recursive_mutex
> l(e
.blob
->shared_blob
->get_cache()->lock
);
9463 for (auto& i
: e
.blob
->shared_blob
->bc
.buffer_map
) {
9464 dout(log_level
) << __func__
<< " 0x" << std::hex
<< i
.first
9465 << "~" << i
.second
->length
<< std::dec
9466 << " " << *i
.second
<< dendl
;
9471 void BlueStore::_dump_transaction(Transaction
*t
, int log_level
)
9473 dout(log_level
) << " transaction dump:\n";
9474 JSONFormatter
f(true);
9475 f
.open_object_section("transaction");
9482 void BlueStore::_pad_zeros(
9483 bufferlist
*bl
, uint64_t *offset
,
9484 uint64_t chunk_size
)
9486 auto length
= bl
->length();
9487 dout(30) << __func__
<< " 0x" << std::hex
<< *offset
<< "~" << length
9488 << " chunk_size 0x" << chunk_size
<< std::dec
<< dendl
;
9489 dout(40) << "before:\n";
9490 bl
->hexdump(*_dout
);
9493 size_t front_pad
= *offset
% chunk_size
;
9494 size_t back_pad
= 0;
9495 size_t pad_count
= 0;
9497 size_t front_copy
= MIN(chunk_size
- front_pad
, length
);
9498 bufferptr z
= buffer::create_page_aligned(chunk_size
);
9499 z
.zero(0, front_pad
, false);
9500 pad_count
+= front_pad
;
9501 bl
->copy(0, front_copy
, z
.c_str() + front_pad
);
9502 if (front_copy
+ front_pad
< chunk_size
) {
9503 back_pad
= chunk_size
- (length
+ front_pad
);
9504 z
.zero(front_pad
+ length
, back_pad
, false);
9505 pad_count
+= back_pad
;
9509 t
.substr_of(old
, front_copy
, length
- front_copy
);
9511 bl
->claim_append(t
);
9512 *offset
-= front_pad
;
9513 length
+= pad_count
;
9517 uint64_t end
= *offset
+ length
;
9518 unsigned back_copy
= end
% chunk_size
;
9520 assert(back_pad
== 0);
9521 back_pad
= chunk_size
- back_copy
;
9522 assert(back_copy
<= length
);
9523 bufferptr
tail(chunk_size
);
9524 bl
->copy(length
- back_copy
, back_copy
, tail
.c_str());
9525 tail
.zero(back_copy
, back_pad
, false);
9528 bl
->substr_of(old
, 0, length
- back_copy
);
9531 pad_count
+= back_pad
;
9533 dout(20) << __func__
<< " pad 0x" << std::hex
<< front_pad
<< " + 0x"
9534 << back_pad
<< " on front/back, now 0x" << *offset
<< "~"
9535 << length
<< std::dec
<< dendl
;
9536 dout(40) << "after:\n";
9537 bl
->hexdump(*_dout
);
9540 logger
->inc(l_bluestore_write_pad_bytes
, pad_count
);
9541 assert(bl
->length() == length
);
9544 void BlueStore::_do_write_small(
9548 uint64_t offset
, uint64_t length
,
9549 bufferlist::iterator
& blp
,
9552 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9553 << std::dec
<< dendl
;
9554 assert(length
< min_alloc_size
);
9555 uint64_t end_offs
= offset
+ length
;
9557 logger
->inc(l_bluestore_write_small
);
9558 logger
->inc(l_bluestore_write_small_bytes
, length
);
9561 blp
.copy(length
, bl
);
9563 // Look for an existing mutable blob we can use.
9564 auto begin
= o
->extent_map
.extent_map
.begin();
9565 auto end
= o
->extent_map
.extent_map
.end();
9566 auto ep
= o
->extent_map
.seek_lextent(offset
);
9569 if (ep
->blob_end() <= offset
) {
9574 if (prev_ep
!= begin
) {
9577 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9580 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9581 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9582 uint32_t alloc_len
= min_alloc_size
;
9583 auto offset0
= P2ALIGN(offset
, alloc_len
);
9587 // search suitable extent in both forward and reverse direction in
9588 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9589 // then check if blob can be reused via can_reuse_blob func or apply
9590 // direct/deferred write (the latter for extents including or higher
9591 // than 'offset' only).
9595 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9596 BlobRef b
= ep
->blob
;
9597 auto bstart
= ep
->blob_start();
9598 dout(20) << __func__
<< " considering " << *b
9599 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9600 if (bstart
>= end_offs
) {
9601 dout(20) << __func__
<< " ignoring distant " << *b
<< dendl
;
9602 } else if (!b
->get_blob().is_mutable()) {
9603 dout(20) << __func__
<< " ignoring immutable " << *b
<< dendl
;
9604 } else if (ep
->logical_offset
% min_alloc_size
!=
9605 ep
->blob_offset
% min_alloc_size
) {
9606 dout(20) << __func__
<< " ignoring offset-skewed " << *b
<< dendl
;
9608 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9609 // can we pad our head/tail out with zeros?
9610 uint64_t head_pad
, tail_pad
;
9611 head_pad
= P2PHASE(offset
, chunk_size
);
9612 tail_pad
= P2NPHASE(end_offs
, chunk_size
);
9613 if (head_pad
|| tail_pad
) {
9614 o
->extent_map
.fault_range(db
, offset
- head_pad
,
9615 end_offs
- offset
+ head_pad
+ tail_pad
);
9618 o
->extent_map
.has_any_lextents(offset
- head_pad
, chunk_size
)) {
9621 if (tail_pad
&& o
->extent_map
.has_any_lextents(end_offs
, tail_pad
)) {
9625 uint64_t b_off
= offset
- head_pad
- bstart
;
9626 uint64_t b_len
= length
+ head_pad
+ tail_pad
;
9628 // direct write into unused blocks of an existing mutable blob?
9629 if ((b_off
% chunk_size
== 0 && b_len
% chunk_size
== 0) &&
9630 b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9631 b
->get_blob().is_unused(b_off
, b_len
) &&
9632 b
->get_blob().is_allocated(b_off
, b_len
)) {
9633 _apply_padding(head_pad
, tail_pad
, bl
);
9635 dout(20) << __func__
<< " write to unused 0x" << std::hex
9636 << b_off
<< "~" << b_len
9637 << " pad 0x" << head_pad
<< " + 0x" << tail_pad
9638 << std::dec
<< " of mutable " << *b
<< dendl
;
9639 _buffer_cache_write(txc
, b
, b_off
, bl
,
9640 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9642 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
9643 if (b_len
<= prefer_deferred_size
) {
9644 dout(20) << __func__
<< " deferring small 0x" << std::hex
9645 << b_len
<< std::dec
<< " unused write via deferred" << dendl
;
9646 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9647 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9650 [&](uint64_t offset
, uint64_t length
) {
9651 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9656 b
->get_blob().map_bl(
9658 [&](uint64_t offset
, bufferlist
& t
) {
9659 bdev
->aio_write(offset
, t
,
9660 &txc
->ioc
, wctx
->buffered
);
9664 b
->dirty_blob().calc_csum(b_off
, bl
);
9665 dout(20) << __func__
<< " lex old " << *ep
<< dendl
;
9666 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, b_off
+ head_pad
, length
,
9668 &wctx
->old_extents
);
9669 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9670 txc
->statfs_delta
.stored() += le
->length
;
9671 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9672 logger
->inc(l_bluestore_write_small_unused
);
9675 // read some data to fill out the chunk?
9676 uint64_t head_read
= P2PHASE(b_off
, chunk_size
);
9677 uint64_t tail_read
= P2NPHASE(b_off
+ b_len
, chunk_size
);
9678 if ((head_read
|| tail_read
) &&
9679 (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
+ tail_read
) &&
9680 head_read
+ tail_read
< min_alloc_size
) {
9682 b_len
+= head_read
+ tail_read
;
9685 head_read
= tail_read
= 0;
9688 // chunk-aligned deferred overwrite?
9689 if (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9690 b_off
% chunk_size
== 0 &&
9691 b_len
% chunk_size
== 0 &&
9692 b
->get_blob().is_allocated(b_off
, b_len
)) {
9694 _apply_padding(head_pad
, tail_pad
, bl
);
9696 dout(20) << __func__
<< " reading head 0x" << std::hex
<< head_read
9697 << " and tail 0x" << tail_read
<< std::dec
<< dendl
;
9700 int r
= _do_read(c
.get(), o
, offset
- head_pad
- head_read
, head_read
,
9702 assert(r
>= 0 && r
<= (int)head_read
);
9703 size_t zlen
= head_read
- r
;
9705 head_bl
.append_zero(zlen
);
9706 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9708 bl
.claim_prepend(head_bl
);
9709 logger
->inc(l_bluestore_write_penalty_read_ops
);
9713 int r
= _do_read(c
.get(), o
, offset
+ length
+ tail_pad
, tail_read
,
9715 assert(r
>= 0 && r
<= (int)tail_read
);
9716 size_t zlen
= tail_read
- r
;
9718 tail_bl
.append_zero(zlen
);
9719 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9721 bl
.claim_append(tail_bl
);
9722 logger
->inc(l_bluestore_write_penalty_read_ops
);
9724 logger
->inc(l_bluestore_write_small_pre_read
);
9726 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9727 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9728 _buffer_cache_write(txc
, b
, b_off
, bl
,
9729 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9731 int r
= b
->get_blob().map(
9733 [&](uint64_t offset
, uint64_t length
) {
9734 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9738 if (b
->get_blob().csum_type
) {
9739 b
->dirty_blob().calc_csum(b_off
, bl
);
9742 dout(20) << __func__
<< " deferred write 0x" << std::hex
<< b_off
<< "~"
9743 << b_len
<< std::dec
<< " of mutable " << *b
9744 << " at " << op
->extents
<< dendl
;
9745 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, offset
- bstart
, length
,
9746 b
, &wctx
->old_extents
);
9747 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9748 txc
->statfs_delta
.stored() += le
->length
;
9749 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9750 logger
->inc(l_bluestore_write_small_deferred
);
9753 // try to reuse blob if we can
9754 if (b
->can_reuse_blob(min_alloc_size
,
9758 assert(alloc_len
== min_alloc_size
); // expecting data always
9759 // fit into reused blob
9760 // Need to check for pending writes desiring to
9761 // reuse the same pextent. The rationale is that during GC two chunks
9762 // from garbage blobs(compressed?) can share logical space within the same
9763 // AU. That's in turn might be caused by unaligned len in clone_range2.
9764 // Hence the second write will fail in an attempt to reuse blob at
9765 // do_alloc_write().
9766 if (!wctx
->has_conflict(b
,
9768 offset0
+ alloc_len
,
9771 // we can't reuse pad_head/pad_tail since they might be truncated
9772 // due to existent extents
9773 uint64_t b_off
= offset
- bstart
;
9774 uint64_t b_off0
= b_off
;
9775 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9777 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9778 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9779 << " (0x" << b_off
<< "~" << length
<< ")"
9780 << std::dec
<< dendl
;
9782 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9783 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9785 logger
->inc(l_bluestore_write_small_unused
);
9792 } // if (ep != end && ep->logical_offset < offset + max_bsize)
9794 // check extent for reuse in reverse order
9795 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9796 BlobRef b
= prev_ep
->blob
;
9797 auto bstart
= prev_ep
->blob_start();
9798 dout(20) << __func__
<< " considering " << *b
9799 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9800 if (b
->can_reuse_blob(min_alloc_size
,
9804 assert(alloc_len
== min_alloc_size
); // expecting data always
9805 // fit into reused blob
9806 // Need to check for pending writes desiring to
9807 // reuse the same pextent. The rationale is that during GC two chunks
9808 // from garbage blobs(compressed?) can share logical space within the same
9809 // AU. That's in turn might be caused by unaligned len in clone_range2.
9810 // Hence the second write will fail in an attempt to reuse blob at
9811 // do_alloc_write().
9812 if (!wctx
->has_conflict(b
,
9814 offset0
+ alloc_len
,
9817 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9818 uint64_t b_off
= offset
- bstart
;
9819 uint64_t b_off0
= b_off
;
9820 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9822 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9823 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9824 << " (0x" << b_off
<< "~" << length
<< ")"
9825 << std::dec
<< dendl
;
9827 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9828 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9830 logger
->inc(l_bluestore_write_small_unused
);
9834 if (prev_ep
!= begin
) {
9838 prev_ep
= end
; // to avoid useless first extent re-check
9840 } // if (prev_ep != end && prev_ep->logical_offset >= min_off)
9841 } while (any_change
);
9845 BlobRef b
= c
->new_blob();
9846 uint64_t b_off
= P2PHASE(offset
, alloc_len
);
9847 uint64_t b_off0
= b_off
;
9848 _pad_zeros(&bl
, &b_off0
, block_size
);
9849 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9850 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
, true, true);
9851 logger
->inc(l_bluestore_write_small_new
);
9856 void BlueStore::_do_write_big(
9860 uint64_t offset
, uint64_t length
,
9861 bufferlist::iterator
& blp
,
9864 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9865 << " target_blob_size 0x" << wctx
->target_blob_size
<< std::dec
9866 << " compress " << (int)wctx
->compress
9868 logger
->inc(l_bluestore_write_big
);
9869 logger
->inc(l_bluestore_write_big_bytes
, length
);
9870 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9871 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9872 while (length
> 0) {
9873 bool new_blob
= false;
9874 uint32_t l
= MIN(max_bsize
, length
);
9878 //attempting to reuse existing blob
9879 if (!wctx
->compress
) {
9880 // look for an existing mutable blob we can reuse
9881 auto begin
= o
->extent_map
.extent_map
.begin();
9882 auto end
= o
->extent_map
.extent_map
.end();
9883 auto ep
= o
->extent_map
.seek_lextent(offset
);
9885 if (prev_ep
!= begin
) {
9888 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9890 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9891 // search suitable extent in both forward and reverse direction in
9892 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9893 // then check if blob can be reused via can_reuse_blob func.
9897 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9898 if (offset
>= ep
->blob_start() &&
9899 ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9900 offset
- ep
->blob_start(),
9903 b_off
= offset
- ep
->blob_start();
9904 prev_ep
= end
; // to avoid check below
9905 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9906 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9913 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9914 if (prev_ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9915 offset
- prev_ep
->blob_start(),
9918 b_off
= offset
- prev_ep
->blob_start();
9919 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9920 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9921 } else if (prev_ep
!= begin
) {
9925 prev_ep
= end
; // to avoid useless first extent re-check
9928 } while (b
== nullptr && any_change
);
9938 wctx
->write(offset
, b
, l
, b_off
, t
, b_off
, l
, false, new_blob
);
9941 logger
->inc(l_bluestore_write_big_blobs
);
9945 int BlueStore::_do_alloc_write(
9951 dout(20) << __func__
<< " txc " << txc
9952 << " " << wctx
->writes
.size() << " blobs"
9954 if (wctx
->writes
.empty()) {
9960 if (wctx
->compress
) {
9962 "compression_algorithm",
9966 if (coll
->pool_opts
.get(pool_opts_t::COMPRESSION_ALGORITHM
, &val
)) {
9967 CompressorRef cp
= compressor
;
9968 if (!cp
|| cp
->get_type_name() != val
) {
9969 cp
= Compressor::create(cct
, val
);
9971 return boost::optional
<CompressorRef
>(cp
);
9973 return boost::optional
<CompressorRef
>();
9977 crr
= select_option(
9978 "compression_required_ratio",
9979 cct
->_conf
->bluestore_compression_required_ratio
,
9982 if (coll
->pool_opts
.get(pool_opts_t::COMPRESSION_REQUIRED_RATIO
, &val
)) {
9983 return boost::optional
<double>(val
);
9985 return boost::optional
<double>();
9991 int csum
= csum_type
.load();
9992 csum
= select_option(
9997 if (coll
->pool_opts
.get(pool_opts_t::CSUM_TYPE
, &val
)) {
9998 return boost::optional
<int>(val
);
10000 return boost::optional
<int>();
10004 // compress (as needed) and calc needed space
10006 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
10007 for (auto& wi
: wctx
->writes
) {
10008 if (c
&& wi
.blob_length
> min_alloc_size
) {
10009 utime_t start
= ceph_clock_now();
10012 assert(wi
.b_off
== 0);
10013 assert(wi
.blob_length
== wi
.bl
.length());
10015 // FIXME: memory alignment here is bad
10017 int r
= c
->compress(wi
.bl
, t
);
10020 bluestore_compression_header_t chdr
;
10021 chdr
.type
= c
->get_type();
10022 chdr
.length
= t
.length();
10023 ::encode(chdr
, wi
.compressed_bl
);
10024 wi
.compressed_bl
.claim_append(t
);
10026 wi
.compressed_len
= wi
.compressed_bl
.length();
10027 uint64_t newlen
= P2ROUNDUP(wi
.compressed_len
, min_alloc_size
);
10028 uint64_t want_len_raw
= wi
.blob_length
* crr
;
10029 uint64_t want_len
= P2ROUNDUP(want_len_raw
, min_alloc_size
);
10030 if (newlen
<= want_len
&& newlen
< wi
.blob_length
) {
10031 // Cool. We compressed at least as much as we were hoping to.
10032 // pad out to min_alloc_size
10033 wi
.compressed_bl
.append_zero(newlen
- wi
.compressed_len
);
10034 logger
->inc(l_bluestore_write_pad_bytes
, newlen
- wi
.compressed_len
);
10035 dout(20) << __func__
<< std::hex
<< " compressed 0x" << wi
.blob_length
10036 << " -> 0x" << wi
.compressed_len
<< " => 0x" << newlen
10037 << " with " << c
->get_type()
10038 << std::dec
<< dendl
;
10039 txc
->statfs_delta
.compressed() += wi
.compressed_len
;
10040 txc
->statfs_delta
.compressed_original() += wi
.blob_length
;
10041 txc
->statfs_delta
.compressed_allocated() += newlen
;
10042 logger
->inc(l_bluestore_compress_success_count
);
10043 wi
.compressed
= true;
10046 dout(20) << __func__
<< std::hex
<< " 0x" << wi
.blob_length
10047 << " compressed to 0x" << wi
.compressed_len
<< " -> 0x" << newlen
10048 << " with " << c
->get_type()
10049 << ", which is more than required 0x" << want_len_raw
10050 << " -> 0x" << want_len
10051 << ", leaving uncompressed"
10052 << std::dec
<< dendl
;
10053 logger
->inc(l_bluestore_compress_rejected_count
);
10054 need
+= wi
.blob_length
;
10056 logger
->tinc(l_bluestore_compress_lat
,
10057 ceph_clock_now() - start
);
10059 need
+= wi
.blob_length
;
10062 int r
= alloc
->reserve(need
);
10064 derr
<< __func__
<< " failed to reserve 0x" << std::hex
<< need
<< std::dec
10068 AllocExtentVector prealloc
;
10069 prealloc
.reserve(2 * wctx
->writes
.size());;
10070 int prealloc_left
= 0;
10071 prealloc_left
= alloc
->allocate(
10072 need
, min_alloc_size
, need
,
10074 assert(prealloc_left
== (int64_t)need
);
10075 dout(20) << __func__
<< " prealloc " << prealloc
<< dendl
;
10076 auto prealloc_pos
= prealloc
.begin();
10078 for (auto& wi
: wctx
->writes
) {
10080 bluestore_blob_t
& dblob
= b
->dirty_blob();
10081 uint64_t b_off
= wi
.b_off
;
10082 bufferlist
*l
= &wi
.bl
;
10083 uint64_t final_length
= wi
.blob_length
;
10084 uint64_t csum_length
= wi
.blob_length
;
10085 unsigned csum_order
= block_size_order
;
10086 if (wi
.compressed
) {
10087 final_length
= wi
.compressed_bl
.length();
10088 csum_length
= final_length
;
10089 csum_order
= ctz(csum_length
);
10090 l
= &wi
.compressed_bl
;
10091 dblob
.set_compressed(wi
.blob_length
, wi
.compressed_len
);
10092 } else if (wi
.new_blob
) {
10093 // initialize newly created blob only
10094 assert(dblob
.is_mutable());
10095 if (l
->length() != wi
.blob_length
) {
10096 // hrm, maybe we could do better here, but let's not bother.
10097 dout(20) << __func__
<< " forcing csum_order to block_size_order "
10098 << block_size_order
<< dendl
;
10099 csum_order
= block_size_order
;
10101 csum_order
= std::min(wctx
->csum_order
, ctz(l
->length()));
10103 // try to align blob with max_blob_size to improve
10104 // its reuse ratio, e.g. in case of reverse write
10105 uint32_t suggested_boff
=
10106 (wi
.logical_offset
- (wi
.b_off0
- wi
.b_off
)) % max_bsize
;
10107 if ((suggested_boff
% (1 << csum_order
)) == 0 &&
10108 suggested_boff
+ final_length
<= max_bsize
&&
10109 suggested_boff
> b_off
) {
10110 dout(20) << __func__
<< " forcing blob_offset to 0x"
10111 << std::hex
<< suggested_boff
<< std::dec
<< dendl
;
10112 assert(suggested_boff
>= b_off
);
10113 csum_length
+= suggested_boff
- b_off
;
10114 b_off
= suggested_boff
;
10116 if (csum
!= Checksummer::CSUM_NONE
) {
10117 dout(20) << __func__
<< " initialize csum setting for new blob " << *b
10118 << " csum_type " << Checksummer::get_csum_type_string(csum
)
10119 << " csum_order " << csum_order
10120 << " csum_length 0x" << std::hex
<< csum_length
<< std::dec
10122 dblob
.init_csum(csum
, csum_order
, csum_length
);
10126 AllocExtentVector extents
;
10127 int64_t left
= final_length
;
10129 assert(prealloc_left
> 0);
10130 if (prealloc_pos
->length
<= left
) {
10131 prealloc_left
-= prealloc_pos
->length
;
10132 left
-= prealloc_pos
->length
;
10133 txc
->statfs_delta
.allocated() += prealloc_pos
->length
;
10134 extents
.push_back(*prealloc_pos
);
10137 extents
.emplace_back(prealloc_pos
->offset
, left
);
10138 prealloc_pos
->offset
+= left
;
10139 prealloc_pos
->length
-= left
;
10140 prealloc_left
-= left
;
10141 txc
->statfs_delta
.allocated() += left
;
10146 for (auto& p
: extents
) {
10147 txc
->allocated
.insert(p
.offset
, p
.length
);
10149 dblob
.allocated(P2ALIGN(b_off
, min_alloc_size
), final_length
, extents
);
10151 dout(20) << __func__
<< " blob " << *b
<< dendl
;
10152 if (dblob
.has_csum()) {
10153 dblob
.calc_csum(b_off
, *l
);
10156 if (wi
.mark_unused
) {
10157 auto b_end
= b_off
+ wi
.bl
.length();
10159 dblob
.add_unused(0, b_off
);
10161 if (b_end
< wi
.blob_length
) {
10162 dblob
.add_unused(b_end
, wi
.blob_length
- b_end
);
10166 Extent
*le
= o
->extent_map
.set_lextent(coll
, wi
.logical_offset
,
10167 b_off
+ (wi
.b_off0
- wi
.b_off
),
10171 wi
.b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
10172 txc
->statfs_delta
.stored() += le
->length
;
10173 dout(20) << __func__
<< " lex " << *le
<< dendl
;
10174 _buffer_cache_write(txc
, wi
.b
, b_off
, wi
.bl
,
10175 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
10178 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
10179 if (l
->length() <= prefer_deferred_size
.load()) {
10180 dout(20) << __func__
<< " deferring small 0x" << std::hex
10181 << l
->length() << std::dec
<< " write via deferred" << dendl
;
10182 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
10183 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
10184 int r
= b
->get_blob().map(
10185 b_off
, l
->length(),
10186 [&](uint64_t offset
, uint64_t length
) {
10187 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
10193 b
->get_blob().map_bl(
10195 [&](uint64_t offset
, bufferlist
& t
) {
10196 bdev
->aio_write(offset
, t
, &txc
->ioc
, false);
10201 assert(prealloc_pos
== prealloc
.end());
10202 assert(prealloc_left
== 0);
10206 void BlueStore::_wctx_finish(
10210 WriteContext
*wctx
,
10211 set
<SharedBlob
*> *maybe_unshared_blobs
)
10213 auto oep
= wctx
->old_extents
.begin();
10214 while (oep
!= wctx
->old_extents
.end()) {
10216 oep
= wctx
->old_extents
.erase(oep
);
10217 dout(20) << __func__
<< " lex_old " << lo
.e
<< dendl
;
10218 BlobRef b
= lo
.e
.blob
;
10219 const bluestore_blob_t
& blob
= b
->get_blob();
10220 if (blob
.is_compressed()) {
10221 if (lo
.blob_empty
) {
10222 txc
->statfs_delta
.compressed() -= blob
.get_compressed_payload_length();
10224 txc
->statfs_delta
.compressed_original() -= lo
.e
.length
;
10227 txc
->statfs_delta
.stored() -= lo
.e
.length
;
10229 dout(20) << __func__
<< " blob release " << r
<< dendl
;
10230 if (blob
.is_shared()) {
10231 PExtentVector final
;
10232 c
->load_shared_blob(b
->shared_blob
);
10234 b
->shared_blob
->put_ref(
10235 e
.offset
, e
.length
, &final
,
10236 b
->is_referenced() ? nullptr : maybe_unshared_blobs
);
10238 dout(20) << __func__
<< " shared_blob release " << final
10239 << " from " << *b
->shared_blob
<< dendl
;
10240 txc
->write_shared_blob(b
->shared_blob
);
10245 // we can't invalidate our logical extents as we drop them because
10246 // other lextents (either in our onode or others) may still
10247 // reference them. but we can throw out anything that is no
10248 // longer allocated. Note that this will leave behind edge bits
10249 // that are no longer referenced but not deallocated (until they
10250 // age out of the cache naturally).
10251 b
->discard_unallocated(c
.get());
10253 dout(20) << __func__
<< " release " << e
<< dendl
;
10254 txc
->released
.insert(e
.offset
, e
.length
);
10255 txc
->statfs_delta
.allocated() -= e
.length
;
10256 if (blob
.is_compressed()) {
10257 txc
->statfs_delta
.compressed_allocated() -= e
.length
;
10261 if (b
->is_spanning() && !b
->is_referenced()) {
10262 dout(20) << __func__
<< " spanning_blob_map removing empty " << *b
10264 o
->extent_map
.spanning_blob_map
.erase(b
->id
);
10269 void BlueStore::_do_write_data(
10276 WriteContext
*wctx
)
10278 uint64_t end
= offset
+ length
;
10279 bufferlist::iterator p
= bl
.begin();
10281 if (offset
/ min_alloc_size
== (end
- 1) / min_alloc_size
&&
10282 (length
!= min_alloc_size
)) {
10283 // we fall within the same block
10284 _do_write_small(txc
, c
, o
, offset
, length
, p
, wctx
);
10286 uint64_t head_offset
, head_length
;
10287 uint64_t middle_offset
, middle_length
;
10288 uint64_t tail_offset
, tail_length
;
10290 head_offset
= offset
;
10291 head_length
= P2NPHASE(offset
, min_alloc_size
);
10293 tail_offset
= P2ALIGN(end
, min_alloc_size
);
10294 tail_length
= P2PHASE(end
, min_alloc_size
);
10296 middle_offset
= head_offset
+ head_length
;
10297 middle_length
= length
- head_length
- tail_length
;
10300 _do_write_small(txc
, c
, o
, head_offset
, head_length
, p
, wctx
);
10303 if (middle_length
) {
10304 _do_write_big(txc
, c
, o
, middle_offset
, middle_length
, p
, wctx
);
10308 _do_write_small(txc
, c
, o
, tail_offset
, tail_length
, p
, wctx
);
10313 void BlueStore::_choose_write_options(
10316 uint32_t fadvise_flags
,
10317 WriteContext
*wctx
)
10319 if (fadvise_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
10320 dout(20) << __func__
<< " will do buffered write" << dendl
;
10321 wctx
->buffered
= true;
10322 } else if (cct
->_conf
->bluestore_default_buffered_write
&&
10323 (fadvise_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
10324 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
10325 dout(20) << __func__
<< " defaulting to buffered write" << dendl
;
10326 wctx
->buffered
= true;
10329 // apply basic csum block size
10330 wctx
->csum_order
= block_size_order
;
10332 // compression parameters
10333 unsigned alloc_hints
= o
->onode
.alloc_hint_flags
;
10334 auto cm
= select_option(
10335 "compression_mode",
10339 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MODE
, &val
)) {
10340 return boost::optional
<Compressor::CompressionMode
>(
10341 Compressor::get_comp_mode_type(val
));
10343 return boost::optional
<Compressor::CompressionMode
>();
10347 wctx
->compress
= (cm
!= Compressor::COMP_NONE
) &&
10348 ((cm
== Compressor::COMP_FORCE
) ||
10349 (cm
== Compressor::COMP_AGGRESSIVE
&&
10350 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE
) == 0) ||
10351 (cm
== Compressor::COMP_PASSIVE
&&
10352 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE
)));
10354 if ((alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_SEQUENTIAL_READ
) &&
10355 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_READ
) == 0 &&
10356 (alloc_hints
& (CEPH_OSD_ALLOC_HINT_FLAG_IMMUTABLE
|
10357 CEPH_OSD_ALLOC_HINT_FLAG_APPEND_ONLY
)) &&
10358 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_WRITE
) == 0) {
10360 dout(20) << __func__
<< " will prefer large blob and csum sizes" << dendl
;
10362 if (o
->onode
.expected_write_size
) {
10363 wctx
->csum_order
= std::max(min_alloc_size_order
,
10364 (uint8_t)ctz(o
->onode
.expected_write_size
));
10366 wctx
->csum_order
= min_alloc_size_order
;
10369 if (wctx
->compress
) {
10370 wctx
->target_blob_size
= select_option(
10371 "compression_max_blob_size",
10372 comp_max_blob_size
.load(),
10375 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MAX_BLOB_SIZE
, &val
)) {
10376 return boost::optional
<uint64_t>((uint64_t)val
);
10378 return boost::optional
<uint64_t>();
10383 if (wctx
->compress
) {
10384 wctx
->target_blob_size
= select_option(
10385 "compression_min_blob_size",
10386 comp_min_blob_size
.load(),
10389 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MIN_BLOB_SIZE
, &val
)) {
10390 return boost::optional
<uint64_t>((uint64_t)val
);
10392 return boost::optional
<uint64_t>();
10398 uint64_t max_bsize
= max_blob_size
.load();
10399 if (wctx
->target_blob_size
== 0 || wctx
->target_blob_size
> max_bsize
) {
10400 wctx
->target_blob_size
= max_bsize
;
10403 // set the min blob size floor at 2x the min_alloc_size, or else we
10404 // won't be able to allocate a smaller extent for the compressed
10406 if (wctx
->compress
&&
10407 wctx
->target_blob_size
< min_alloc_size
* 2) {
10408 wctx
->target_blob_size
= min_alloc_size
* 2;
10411 dout(20) << __func__
<< " prefer csum_order " << wctx
->csum_order
10412 << " target_blob_size 0x" << std::hex
<< wctx
->target_blob_size
10413 << std::dec
<< dendl
;
10416 int BlueStore::_do_gc(
10420 const GarbageCollector
& gc
,
10421 const WriteContext
& wctx
,
10422 uint64_t *dirty_start
,
10423 uint64_t *dirty_end
)
10425 auto& extents_to_collect
= gc
.get_extents_to_collect();
10427 bool dirty_range_updated
= false;
10428 WriteContext wctx_gc
;
10429 wctx_gc
.fork(wctx
); // make a clone for garbage collection
10431 for (auto it
= extents_to_collect
.begin();
10432 it
!= extents_to_collect
.end();
10435 int r
= _do_read(c
.get(), o
, it
->offset
, it
->length
, bl
, 0);
10436 assert(r
== (int)it
->length
);
10438 o
->extent_map
.fault_range(db
, it
->offset
, it
->length
);
10439 _do_write_data(txc
, c
, o
, it
->offset
, it
->length
, bl
, &wctx_gc
);
10440 logger
->inc(l_bluestore_gc_merged
, it
->length
);
10442 if (*dirty_start
> it
->offset
) {
10443 *dirty_start
= it
->offset
;
10444 dirty_range_updated
= true;
10447 if (*dirty_end
< it
->offset
+ it
->length
) {
10448 *dirty_end
= it
->offset
+ it
->length
;
10449 dirty_range_updated
= true;
10452 if (dirty_range_updated
) {
10453 o
->extent_map
.fault_range(db
, *dirty_start
, *dirty_end
);
10456 dout(30) << __func__
<< " alloc write" << dendl
;
10457 int r
= _do_alloc_write(txc
, c
, o
, &wctx_gc
);
10459 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10464 _wctx_finish(txc
, c
, o
, &wctx_gc
);
10468 int BlueStore::_do_write(
10475 uint32_t fadvise_flags
)
10479 dout(20) << __func__
10481 << " 0x" << std::hex
<< offset
<< "~" << length
10482 << " - have 0x" << o
->onode
.size
10483 << " (" << std::dec
<< o
->onode
.size
<< ")"
10485 << " fadvise_flags 0x" << std::hex
<< fadvise_flags
<< std::dec
10493 uint64_t end
= offset
+ length
;
10495 GarbageCollector
gc(c
->store
->cct
);
10497 auto dirty_start
= offset
;
10498 auto dirty_end
= end
;
10501 _choose_write_options(c
, o
, fadvise_flags
, &wctx
);
10502 o
->extent_map
.fault_range(db
, offset
, length
);
10503 _do_write_data(txc
, c
, o
, offset
, length
, bl
, &wctx
);
10504 r
= _do_alloc_write(txc
, c
, o
, &wctx
);
10506 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10511 // NB: _wctx_finish() will empty old_extents
10512 // so we must do gc estimation before that
10513 benefit
= gc
.estimate(offset
,
10519 _wctx_finish(txc
, c
, o
, &wctx
);
10520 if (end
> o
->onode
.size
) {
10521 dout(20) << __func__
<< " extending size to 0x" << std::hex
<< end
10522 << std::dec
<< dendl
;
10523 o
->onode
.size
= end
;
10526 if (benefit
>= g_conf
->bluestore_gc_enable_total_threshold
) {
10527 if (!gc
.get_extents_to_collect().empty()) {
10528 dout(20) << __func__
<< " perform garbage collection, "
10529 << "expected benefit = " << benefit
<< " AUs" << dendl
;
10530 r
= _do_gc(txc
, c
, o
, gc
, wctx
, &dirty_start
, &dirty_end
);
10532 derr
<< __func__
<< " _do_gc failed with " << cpp_strerror(r
)
10536 dout(20)<<__func__
<<" gc range is " << std::hex
<< dirty_start
10537 << "~" << dirty_end
- dirty_start
<< std::dec
<< dendl
;
10540 o
->extent_map
.compress_extent_map(dirty_start
, dirty_end
- dirty_start
);
10541 o
->extent_map
.dirty_range(dirty_start
, dirty_end
- dirty_start
);
10549 int BlueStore::_write(TransContext
*txc
,
10552 uint64_t offset
, size_t length
,
10554 uint32_t fadvise_flags
)
10556 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10557 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10560 if (offset
+ length
>= OBJECT_MAX_SIZE
) {
10563 _assign_nid(txc
, o
);
10564 r
= _do_write(txc
, c
, o
, offset
, length
, bl
, fadvise_flags
);
10565 txc
->write_onode(o
);
10567 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10568 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10569 << " = " << r
<< dendl
;
10573 int BlueStore::_zero(TransContext
*txc
,
10576 uint64_t offset
, size_t length
)
10578 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10579 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10582 if (offset
+ length
>= OBJECT_MAX_SIZE
) {
10585 _assign_nid(txc
, o
);
10586 r
= _do_zero(txc
, c
, o
, offset
, length
);
10588 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10589 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10590 << " = " << r
<< dendl
;
10594 int BlueStore::_do_zero(TransContext
*txc
,
10597 uint64_t offset
, size_t length
)
10599 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10600 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10607 o
->extent_map
.fault_range(db
, offset
, length
);
10608 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10609 o
->extent_map
.dirty_range(offset
, length
);
10610 _wctx_finish(txc
, c
, o
, &wctx
);
10612 if (length
> 0 && offset
+ length
> o
->onode
.size
) {
10613 o
->onode
.size
= offset
+ length
;
10614 dout(20) << __func__
<< " extending size to " << offset
+ length
10617 txc
->write_onode(o
);
10619 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10620 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10621 << " = " << r
<< dendl
;
10625 void BlueStore::_do_truncate(
10626 TransContext
*txc
, CollectionRef
& c
, OnodeRef o
, uint64_t offset
,
10627 set
<SharedBlob
*> *maybe_unshared_blobs
)
10629 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10630 << " 0x" << std::hex
<< offset
<< std::dec
<< dendl
;
10632 _dump_onode(o
, 30);
10634 if (offset
== o
->onode
.size
)
10637 if (offset
< o
->onode
.size
) {
10639 uint64_t length
= o
->onode
.size
- offset
;
10640 o
->extent_map
.fault_range(db
, offset
, length
);
10641 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10642 o
->extent_map
.dirty_range(offset
, length
);
10643 _wctx_finish(txc
, c
, o
, &wctx
, maybe_unshared_blobs
);
10645 // if we have shards past EOF, ask for a reshard
10646 if (!o
->onode
.extent_map_shards
.empty() &&
10647 o
->onode
.extent_map_shards
.back().offset
>= offset
) {
10648 dout(10) << __func__
<< " request reshard past EOF" << dendl
;
10650 o
->extent_map
.request_reshard(offset
- 1, offset
+ length
);
10652 o
->extent_map
.request_reshard(0, length
);
10657 o
->onode
.size
= offset
;
10659 txc
->write_onode(o
);
10662 int BlueStore::_truncate(TransContext
*txc
,
10667 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10668 << " 0x" << std::hex
<< offset
<< std::dec
10671 if (offset
>= OBJECT_MAX_SIZE
) {
10674 _do_truncate(txc
, c
, o
, offset
);
10676 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10677 << " 0x" << std::hex
<< offset
<< std::dec
10678 << " = " << r
<< dendl
;
10682 int BlueStore::_do_remove(
10687 set
<SharedBlob
*> maybe_unshared_blobs
;
10688 bool is_gen
= !o
->oid
.is_no_gen();
10689 _do_truncate(txc
, c
, o
, 0, is_gen
? &maybe_unshared_blobs
: nullptr);
10690 if (o
->onode
.has_omap()) {
10692 _do_omap_clear(txc
, o
->onode
.nid
);
10696 for (auto &s
: o
->extent_map
.shards
) {
10697 dout(20) << __func__
<< " removing shard 0x" << std::hex
10698 << s
.shard_info
->offset
<< std::dec
<< dendl
;
10699 generate_extent_shard_key_and_apply(o
->key
, s
.shard_info
->offset
, &key
,
10700 [&](const string
& final_key
) {
10701 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
10705 txc
->t
->rmkey(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size());
10707 o
->extent_map
.clear();
10708 o
->onode
= bluestore_onode_t();
10709 _debug_obj_on_delete(o
->oid
);
10711 if (!is_gen
|| maybe_unshared_blobs
.empty()) {
10715 // see if we can unshare blobs still referenced by the head
10716 dout(10) << __func__
<< " gen and maybe_unshared_blobs "
10717 << maybe_unshared_blobs
<< dendl
;
10718 ghobject_t nogen
= o
->oid
;
10719 nogen
.generation
= ghobject_t::NO_GEN
;
10720 OnodeRef h
= c
->onode_map
.lookup(nogen
);
10722 if (!h
|| !h
->exists
) {
10726 dout(20) << __func__
<< " checking for unshareable blobs on " << h
10727 << " " << h
->oid
<< dendl
;
10728 map
<SharedBlob
*,bluestore_extent_ref_map_t
> expect
;
10729 for (auto& e
: h
->extent_map
.extent_map
) {
10730 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10731 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10732 if (b
.is_shared() &&
10734 maybe_unshared_blobs
.count(sb
)) {
10735 if (b
.is_compressed()) {
10736 expect
[sb
].get(0, b
.get_ondisk_length());
10738 b
.map(e
.blob_offset
, e
.length
, [&](uint64_t off
, uint64_t len
) {
10739 expect
[sb
].get(off
, len
);
10746 vector
<SharedBlob
*> unshared_blobs
;
10747 unshared_blobs
.reserve(maybe_unshared_blobs
.size());
10748 for (auto& p
: expect
) {
10749 dout(20) << " ? " << *p
.first
<< " vs " << p
.second
<< dendl
;
10750 if (p
.first
->persistent
->ref_map
== p
.second
) {
10751 SharedBlob
*sb
= p
.first
;
10752 dout(20) << __func__
<< " unsharing " << *sb
<< dendl
;
10753 unshared_blobs
.push_back(sb
);
10754 txc
->unshare_blob(sb
);
10755 uint64_t sbid
= c
->make_blob_unshared(sb
);
10757 get_shared_blob_key(sbid
, &key
);
10758 txc
->t
->rmkey(PREFIX_SHARED_BLOB
, key
);
10762 if (unshared_blobs
.empty()) {
10766 for (auto& e
: h
->extent_map
.extent_map
) {
10767 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10768 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10769 if (b
.is_shared() &&
10770 std::find(unshared_blobs
.begin(), unshared_blobs
.end(),
10771 sb
) != unshared_blobs
.end()) {
10772 dout(20) << __func__
<< " unsharing " << e
<< dendl
;
10773 bluestore_blob_t
& blob
= e
.blob
->dirty_blob();
10774 blob
.clear_flag(bluestore_blob_t::FLAG_SHARED
);
10775 h
->extent_map
.dirty_range(e
.logical_offset
, 1);
10778 txc
->write_onode(h
);
10783 int BlueStore::_remove(TransContext
*txc
,
10787 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10788 int r
= _do_remove(txc
, c
, o
);
10789 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10793 int BlueStore::_setattr(TransContext
*txc
,
10796 const string
& name
,
10799 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10800 << " " << name
<< " (" << val
.length() << " bytes)"
10803 if (val
.is_partial()) {
10804 auto& b
= o
->onode
.attrs
[name
.c_str()] = bufferptr(val
.c_str(),
10806 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10808 auto& b
= o
->onode
.attrs
[name
.c_str()] = val
;
10809 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10811 txc
->write_onode(o
);
10812 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10813 << " " << name
<< " (" << val
.length() << " bytes)"
10814 << " = " << r
<< dendl
;
10818 int BlueStore::_setattrs(TransContext
*txc
,
10821 const map
<string
,bufferptr
>& aset
)
10823 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10824 << " " << aset
.size() << " keys"
10827 for (map
<string
,bufferptr
>::const_iterator p
= aset
.begin();
10828 p
!= aset
.end(); ++p
) {
10829 if (p
->second
.is_partial()) {
10830 auto& b
= o
->onode
.attrs
[p
->first
.c_str()] =
10831 bufferptr(p
->second
.c_str(), p
->second
.length());
10832 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10834 auto& b
= o
->onode
.attrs
[p
->first
.c_str()] = p
->second
;
10835 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10838 txc
->write_onode(o
);
10839 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10840 << " " << aset
.size() << " keys"
10841 << " = " << r
<< dendl
;
10846 int BlueStore::_rmattr(TransContext
*txc
,
10849 const string
& name
)
10851 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10852 << " " << name
<< dendl
;
10854 auto it
= o
->onode
.attrs
.find(name
.c_str());
10855 if (it
== o
->onode
.attrs
.end())
10858 o
->onode
.attrs
.erase(it
);
10859 txc
->write_onode(o
);
10862 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10863 << " " << name
<< " = " << r
<< dendl
;
10867 int BlueStore::_rmattrs(TransContext
*txc
,
10871 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10874 if (o
->onode
.attrs
.empty())
10877 o
->onode
.attrs
.clear();
10878 txc
->write_onode(o
);
10881 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10885 void BlueStore::_do_omap_clear(TransContext
*txc
, uint64_t id
)
10887 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
10888 string prefix
, tail
;
10889 get_omap_header(id
, &prefix
);
10890 get_omap_tail(id
, &tail
);
10891 it
->lower_bound(prefix
);
10892 while (it
->valid()) {
10893 if (it
->key() >= tail
) {
10894 dout(30) << __func__
<< " stop at " << pretty_binary_string(tail
)
10898 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
10899 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
10904 int BlueStore::_omap_clear(TransContext
*txc
,
10908 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10910 if (o
->onode
.has_omap()) {
10912 _do_omap_clear(txc
, o
->onode
.nid
);
10913 o
->onode
.clear_omap_flag();
10914 txc
->write_onode(o
);
10916 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10920 int BlueStore::_omap_setkeys(TransContext
*txc
,
10925 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10927 bufferlist::iterator p
= bl
.begin();
10929 if (!o
->onode
.has_omap()) {
10930 o
->onode
.set_omap_flag();
10931 txc
->write_onode(o
);
10933 txc
->note_modified_object(o
);
10936 _key_encode_u64(o
->onode
.nid
, &final_key
);
10937 final_key
.push_back('.');
10943 ::decode(value
, p
);
10944 final_key
.resize(9); // keep prefix
10946 dout(30) << __func__
<< " " << pretty_binary_string(final_key
)
10947 << " <- " << key
<< dendl
;
10948 txc
->t
->set(PREFIX_OMAP
, final_key
, value
);
10951 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10955 int BlueStore::_omap_setheader(TransContext
*txc
,
10960 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10963 if (!o
->onode
.has_omap()) {
10964 o
->onode
.set_omap_flag();
10965 txc
->write_onode(o
);
10967 txc
->note_modified_object(o
);
10969 get_omap_header(o
->onode
.nid
, &key
);
10970 txc
->t
->set(PREFIX_OMAP
, key
, bl
);
10972 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10976 int BlueStore::_omap_rmkeys(TransContext
*txc
,
10981 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10983 bufferlist::iterator p
= bl
.begin();
10987 if (!o
->onode
.has_omap()) {
10990 _key_encode_u64(o
->onode
.nid
, &final_key
);
10991 final_key
.push_back('.');
10996 final_key
.resize(9); // keep prefix
10998 dout(30) << __func__
<< " rm " << pretty_binary_string(final_key
)
10999 << " <- " << key
<< dendl
;
11000 txc
->t
->rmkey(PREFIX_OMAP
, final_key
);
11002 txc
->note_modified_object(o
);
11005 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
11009 int BlueStore::_omap_rmkey_range(TransContext
*txc
,
11012 const string
& first
, const string
& last
)
11014 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
11015 KeyValueDB::Iterator it
;
11016 string key_first
, key_last
;
11018 if (!o
->onode
.has_omap()) {
11022 it
= db
->get_iterator(PREFIX_OMAP
);
11023 get_omap_key(o
->onode
.nid
, first
, &key_first
);
11024 get_omap_key(o
->onode
.nid
, last
, &key_last
);
11025 it
->lower_bound(key_first
);
11026 while (it
->valid()) {
11027 if (it
->key() >= key_last
) {
11028 dout(30) << __func__
<< " stop at " << pretty_binary_string(key_last
)
11032 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
11033 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
11036 txc
->note_modified_object(o
);
11039 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
11043 int BlueStore::_set_alloc_hint(
11047 uint64_t expected_object_size
,
11048 uint64_t expected_write_size
,
11051 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
11052 << " object_size " << expected_object_size
11053 << " write_size " << expected_write_size
11054 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
11057 o
->onode
.expected_object_size
= expected_object_size
;
11058 o
->onode
.expected_write_size
= expected_write_size
;
11059 o
->onode
.alloc_hint_flags
= flags
;
11060 txc
->write_onode(o
);
11061 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
11062 << " object_size " << expected_object_size
11063 << " write_size " << expected_write_size
11064 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
11065 << " = " << r
<< dendl
;
11069 int BlueStore::_clone(TransContext
*txc
,
11074 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11075 << newo
->oid
<< dendl
;
11077 if (oldo
->oid
.hobj
.get_hash() != newo
->oid
.hobj
.get_hash()) {
11078 derr
<< __func__
<< " mismatched hash on " << oldo
->oid
11079 << " and " << newo
->oid
<< dendl
;
11083 _assign_nid(txc
, newo
);
11087 _do_truncate(txc
, c
, newo
, 0);
11088 if (cct
->_conf
->bluestore_clone_cow
) {
11089 _do_clone_range(txc
, c
, oldo
, newo
, 0, oldo
->onode
.size
, 0);
11092 r
= _do_read(c
.get(), oldo
, 0, oldo
->onode
.size
, bl
, 0);
11095 r
= _do_write(txc
, c
, newo
, 0, oldo
->onode
.size
, bl
, 0);
11101 newo
->onode
.attrs
= oldo
->onode
.attrs
;
11104 if (newo
->onode
.has_omap()) {
11105 dout(20) << __func__
<< " clearing old omap data" << dendl
;
11107 _do_omap_clear(txc
, newo
->onode
.nid
);
11109 if (oldo
->onode
.has_omap()) {
11110 dout(20) << __func__
<< " copying omap data" << dendl
;
11111 if (!newo
->onode
.has_omap()) {
11112 newo
->onode
.set_omap_flag();
11114 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
11116 get_omap_header(oldo
->onode
.nid
, &head
);
11117 get_omap_tail(oldo
->onode
.nid
, &tail
);
11118 it
->lower_bound(head
);
11119 while (it
->valid()) {
11120 if (it
->key() >= tail
) {
11121 dout(30) << __func__
<< " reached tail" << dendl
;
11124 dout(30) << __func__
<< " got header/data "
11125 << pretty_binary_string(it
->key()) << dendl
;
11127 rewrite_omap_key(newo
->onode
.nid
, it
->key(), &key
);
11128 txc
->t
->set(PREFIX_OMAP
, key
, it
->value());
11133 newo
->onode
.clear_omap_flag();
11136 txc
->write_onode(newo
);
11140 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11141 << newo
->oid
<< " = " << r
<< dendl
;
11145 int BlueStore::_do_clone_range(
11154 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11156 << " 0x" << std::hex
<< srcoff
<< "~" << length
<< " -> "
11157 << " 0x" << dstoff
<< "~" << length
<< std::dec
<< dendl
;
11158 oldo
->extent_map
.fault_range(db
, srcoff
, length
);
11159 newo
->extent_map
.fault_range(db
, dstoff
, length
);
11163 // hmm, this could go into an ExtentMap::dup() method.
11164 vector
<BlobRef
> id_to_blob(oldo
->extent_map
.extent_map
.size());
11165 for (auto &e
: oldo
->extent_map
.extent_map
) {
11166 e
.blob
->last_encoded_id
= -1;
11169 uint64_t end
= srcoff
+ length
;
11170 uint32_t dirty_range_begin
= 0;
11171 uint32_t dirty_range_end
= 0;
11172 bool src_dirty
= false;
11173 for (auto ep
= oldo
->extent_map
.seek_lextent(srcoff
);
11174 ep
!= oldo
->extent_map
.extent_map
.end();
11177 if (e
.logical_offset
>= end
) {
11180 dout(20) << __func__
<< " src " << e
<< dendl
;
11182 bool blob_duped
= true;
11183 if (e
.blob
->last_encoded_id
>= 0) {
11184 // blob is already duped
11185 cb
= id_to_blob
[e
.blob
->last_encoded_id
];
11186 blob_duped
= false;
11189 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
11190 // make sure it is shared
11191 if (!blob
.is_shared()) {
11192 c
->make_blob_shared(_assign_blobid(txc
), e
.blob
);
11195 dirty_range_begin
= e
.logical_offset
;
11197 assert(e
.logical_end() > 0);
11198 // -1 to exclude next potential shard
11199 dirty_range_end
= e
.logical_end() - 1;
11201 c
->load_shared_blob(e
.blob
->shared_blob
);
11204 e
.blob
->last_encoded_id
= n
;
11205 id_to_blob
[n
] = cb
;
11207 // bump the extent refs on the copied blob's extents
11208 for (auto p
: blob
.get_extents()) {
11209 if (p
.is_valid()) {
11210 e
.blob
->shared_blob
->get_ref(p
.offset
, p
.length
);
11213 txc
->write_shared_blob(e
.blob
->shared_blob
);
11214 dout(20) << __func__
<< " new " << *cb
<< dendl
;
11217 int skip_front
, skip_back
;
11218 if (e
.logical_offset
< srcoff
) {
11219 skip_front
= srcoff
- e
.logical_offset
;
11223 if (e
.logical_end() > end
) {
11224 skip_back
= e
.logical_end() - end
;
11228 Extent
*ne
= new Extent(e
.logical_offset
+ skip_front
+ dstoff
- srcoff
,
11229 e
.blob_offset
+ skip_front
,
11230 e
.length
- skip_front
- skip_back
, cb
);
11231 newo
->extent_map
.extent_map
.insert(*ne
);
11232 ne
->blob
->get_ref(c
.get(), ne
->blob_offset
, ne
->length
);
11233 // fixme: we may leave parts of new blob unreferenced that could
11234 // be freed (relative to the shared_blob).
11235 txc
->statfs_delta
.stored() += ne
->length
;
11236 if (e
.blob
->get_blob().is_compressed()) {
11237 txc
->statfs_delta
.compressed_original() += ne
->length
;
11239 txc
->statfs_delta
.compressed() +=
11240 cb
->get_blob().get_compressed_payload_length();
11243 dout(20) << __func__
<< " dst " << *ne
<< dendl
;
11247 oldo
->extent_map
.dirty_range(dirty_range_begin
,
11248 dirty_range_end
- dirty_range_begin
);
11249 txc
->write_onode(oldo
);
11251 txc
->write_onode(newo
);
11253 if (dstoff
+ length
> newo
->onode
.size
) {
11254 newo
->onode
.size
= dstoff
+ length
;
11256 newo
->extent_map
.dirty_range(dstoff
, length
);
11262 int BlueStore::_clone_range(TransContext
*txc
,
11266 uint64_t srcoff
, uint64_t length
, uint64_t dstoff
)
11268 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11269 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11270 << " to offset 0x" << dstoff
<< std::dec
<< dendl
;
11273 if (srcoff
+ length
>= OBJECT_MAX_SIZE
||
11274 dstoff
+ length
>= OBJECT_MAX_SIZE
) {
11278 if (srcoff
+ length
> oldo
->onode
.size
) {
11283 _assign_nid(txc
, newo
);
11286 if (cct
->_conf
->bluestore_clone_cow
) {
11287 _do_zero(txc
, c
, newo
, dstoff
, length
);
11288 _do_clone_range(txc
, c
, oldo
, newo
, srcoff
, length
, dstoff
);
11291 r
= _do_read(c
.get(), oldo
, srcoff
, length
, bl
, 0);
11294 r
= _do_write(txc
, c
, newo
, dstoff
, bl
.length(), bl
, 0);
11300 txc
->write_onode(newo
);
11304 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11305 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11306 << " to offset 0x" << dstoff
<< std::dec
11307 << " = " << r
<< dendl
;
11311 int BlueStore::_rename(TransContext
*txc
,
11315 const ghobject_t
& new_oid
)
11317 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11318 << new_oid
<< dendl
;
11320 ghobject_t old_oid
= oldo
->oid
;
11321 mempool::bluestore_cache_other::string new_okey
;
11324 if (newo
->exists
) {
11328 assert(txc
->onodes
.count(newo
) == 0);
11331 txc
->t
->rmkey(PREFIX_OBJ
, oldo
->key
.c_str(), oldo
->key
.size());
11335 oldo
->extent_map
.fault_range(db
, 0, oldo
->onode
.size
);
11336 get_object_key(cct
, new_oid
, &new_okey
);
11338 for (auto &s
: oldo
->extent_map
.shards
) {
11339 generate_extent_shard_key_and_apply(oldo
->key
, s
.shard_info
->offset
, &key
,
11340 [&](const string
& final_key
) {
11341 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
11349 txc
->write_onode(newo
);
11351 // this adjusts oldo->{oid,key}, and reset oldo to a fresh empty
11352 // Onode in the old slot
11353 c
->onode_map
.rename(oldo
, old_oid
, new_oid
, new_okey
);
11357 dout(10) << __func__
<< " " << c
->cid
<< " " << old_oid
<< " -> "
11358 << new_oid
<< " = " << r
<< dendl
;
11364 int BlueStore::_create_collection(
11370 dout(15) << __func__
<< " " << cid
<< " bits " << bits
<< dendl
;
11375 RWLock::WLocker
l(coll_lock
);
11383 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
11385 (*c
)->cnode
.bits
= bits
;
11386 coll_map
[cid
] = *c
;
11388 ::encode((*c
)->cnode
, bl
);
11389 txc
->t
->set(PREFIX_COLL
, stringify(cid
), bl
);
11393 dout(10) << __func__
<< " " << cid
<< " bits " << bits
<< " = " << r
<< dendl
;
11397 int BlueStore::_remove_collection(TransContext
*txc
, const coll_t
&cid
,
11400 dout(15) << __func__
<< " " << cid
<< dendl
;
11404 RWLock::WLocker
l(coll_lock
);
11409 size_t nonexistent_count
= 0;
11410 assert((*c
)->exists
);
11411 if ((*c
)->onode_map
.map_any([&](OnodeRef o
) {
11413 dout(10) << __func__
<< " " << o
->oid
<< " " << o
11414 << " exists in onode_map" << dendl
;
11417 ++nonexistent_count
;
11424 vector
<ghobject_t
> ls
;
11426 // Enumerate onodes in db, up to nonexistent_count + 1
11427 // then check if all of them are marked as non-existent.
11428 // Bypass the check if returned number is greater than nonexistent_count
11429 r
= _collection_list(c
->get(), ghobject_t(), ghobject_t::get_max(),
11430 nonexistent_count
+ 1, &ls
, &next
);
11432 bool exists
= false; //ls.size() > nonexistent_count;
11433 for (auto it
= ls
.begin(); !exists
&& it
< ls
.end(); ++it
) {
11434 dout(10) << __func__
<< " oid " << *it
<< dendl
;
11435 auto onode
= (*c
)->onode_map
.lookup(*it
);
11436 exists
= !onode
|| onode
->exists
;
11438 dout(10) << __func__
<< " " << *it
11439 << " exists in db" << dendl
;
11443 coll_map
.erase(cid
);
11444 txc
->removed_collections
.push_back(*c
);
11445 (*c
)->exists
= false;
11447 txc
->t
->rmkey(PREFIX_COLL
, stringify(cid
));
11450 dout(10) << __func__
<< " " << cid
11451 << " is non-empty" << dendl
;
11458 dout(10) << __func__
<< " " << cid
<< " = " << r
<< dendl
;
11462 int BlueStore::_split_collection(TransContext
*txc
,
11465 unsigned bits
, int rem
)
11467 dout(15) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11468 << " bits " << bits
<< dendl
;
11469 RWLock::WLocker
l(c
->lock
);
11470 RWLock::WLocker
l2(d
->lock
);
11473 // flush all previous deferred writes on this sequencer. this is a bit
11474 // heavyweight, but we need to make sure all deferred writes complete
11475 // before we split as the new collection's sequencer may need to order
11476 // this after those writes, and we don't bother with the complexity of
11477 // moving those TransContexts over to the new osr.
11478 _osr_drain_preceding(txc
);
11480 // move any cached items (onodes and referenced shared blobs) that will
11481 // belong to the child collection post-split. leave everything else behind.
11482 // this may include things that don't strictly belong to the now-smaller
11483 // parent split, but the OSD will always send us a split for every new
11486 spg_t pgid
, dest_pgid
;
11487 bool is_pg
= c
->cid
.is_pg(&pgid
);
11489 is_pg
= d
->cid
.is_pg(&dest_pgid
);
11492 // the destination should initially be empty.
11493 assert(d
->onode_map
.empty());
11494 assert(d
->shared_blob_set
.empty());
11495 assert(d
->cnode
.bits
== bits
);
11497 c
->split_cache(d
.get());
11499 // adjust bits. note that this will be redundant for all but the first
11500 // split call for this parent (first child).
11501 c
->cnode
.bits
= bits
;
11502 assert(d
->cnode
.bits
== bits
);
11506 ::encode(c
->cnode
, bl
);
11507 txc
->t
->set(PREFIX_COLL
, stringify(c
->cid
), bl
);
11509 dout(10) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11510 << " bits " << bits
<< " = " << r
<< dendl
;
11514 // DB key value Histogram
11515 #define KEY_SLAB 32
11516 #define VALUE_SLAB 64
11518 const string prefix_onode
= "o";
11519 const string prefix_onode_shard
= "x";
11520 const string prefix_other
= "Z";
11522 int BlueStore::DBHistogram::get_key_slab(size_t sz
)
11524 return (sz
/KEY_SLAB
);
11527 string
BlueStore::DBHistogram::get_key_slab_to_range(int slab
)
11529 int lower_bound
= slab
* KEY_SLAB
;
11530 int upper_bound
= (slab
+ 1) * KEY_SLAB
;
11531 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11535 int BlueStore::DBHistogram::get_value_slab(size_t sz
)
11537 return (sz
/VALUE_SLAB
);
11540 string
BlueStore::DBHistogram::get_value_slab_to_range(int slab
)
11542 int lower_bound
= slab
* VALUE_SLAB
;
11543 int upper_bound
= (slab
+ 1) * VALUE_SLAB
;
11544 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11548 void BlueStore::DBHistogram::update_hist_entry(map
<string
, map
<int, struct key_dist
> > &key_hist
,
11549 const string
&prefix
, size_t key_size
, size_t value_size
)
11551 uint32_t key_slab
= get_key_slab(key_size
);
11552 uint32_t value_slab
= get_value_slab(value_size
);
11553 key_hist
[prefix
][key_slab
].count
++;
11554 key_hist
[prefix
][key_slab
].max_len
= MAX(key_size
, key_hist
[prefix
][key_slab
].max_len
);
11555 key_hist
[prefix
][key_slab
].val_map
[value_slab
].count
++;
11556 key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
=
11557 MAX(value_size
, key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
);
11560 void BlueStore::DBHistogram::dump(Formatter
*f
)
11562 f
->open_object_section("rocksdb_value_distribution");
11563 for (auto i
: value_hist
) {
11564 f
->dump_unsigned(get_value_slab_to_range(i
.first
).data(), i
.second
);
11566 f
->close_section();
11568 f
->open_object_section("rocksdb_key_value_histogram");
11569 for (auto i
: key_hist
) {
11570 f
->dump_string("prefix", i
.first
);
11571 f
->open_object_section("key_hist");
11572 for ( auto k
: i
.second
) {
11573 f
->dump_unsigned(get_key_slab_to_range(k
.first
).data(), k
.second
.count
);
11574 f
->dump_unsigned("max_len", k
.second
.max_len
);
11575 f
->open_object_section("value_hist");
11576 for ( auto j
: k
.second
.val_map
) {
11577 f
->dump_unsigned(get_value_slab_to_range(j
.first
).data(), j
.second
.count
);
11578 f
->dump_unsigned("max_len", j
.second
.max_len
);
11580 f
->close_section();
11582 f
->close_section();
11584 f
->close_section();
11587 //Itrerates through the db and collects the stats
11588 void BlueStore::generate_db_histogram(Formatter
*f
)
11591 uint64_t num_onodes
= 0;
11592 uint64_t num_shards
= 0;
11593 uint64_t num_super
= 0;
11594 uint64_t num_coll
= 0;
11595 uint64_t num_omap
= 0;
11596 uint64_t num_deferred
= 0;
11597 uint64_t num_alloc
= 0;
11598 uint64_t num_stat
= 0;
11599 uint64_t num_others
= 0;
11600 uint64_t num_shared_shards
= 0;
11601 size_t max_key_size
=0, max_value_size
= 0;
11602 uint64_t total_key_size
= 0, total_value_size
= 0;
11603 size_t key_size
= 0, value_size
= 0;
11606 utime_t start
= ceph_clock_now();
11608 KeyValueDB::WholeSpaceIterator iter
= db
->get_iterator();
11609 iter
->seek_to_first();
11610 while (iter
->valid()) {
11611 dout(30) << __func__
<< " Key: " << iter
->key() << dendl
;
11612 key_size
= iter
->key_size();
11613 value_size
= iter
->value_size();
11614 hist
.value_hist
[hist
.get_value_slab(value_size
)]++;
11615 max_key_size
= MAX(max_key_size
, key_size
);
11616 max_value_size
= MAX(max_value_size
, value_size
);
11617 total_key_size
+= key_size
;
11618 total_value_size
+= value_size
;
11620 pair
<string
,string
> key(iter
->raw_key());
11622 if (key
.first
== PREFIX_SUPER
) {
11623 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SUPER
, key_size
, value_size
);
11625 } else if (key
.first
== PREFIX_STAT
) {
11626 hist
.update_hist_entry(hist
.key_hist
, PREFIX_STAT
, key_size
, value_size
);
11628 } else if (key
.first
== PREFIX_COLL
) {
11629 hist
.update_hist_entry(hist
.key_hist
, PREFIX_COLL
, key_size
, value_size
);
11631 } else if (key
.first
== PREFIX_OBJ
) {
11632 if (key
.second
.back() == ONODE_KEY_SUFFIX
) {
11633 hist
.update_hist_entry(hist
.key_hist
, prefix_onode
, key_size
, value_size
);
11636 hist
.update_hist_entry(hist
.key_hist
, prefix_onode_shard
, key_size
, value_size
);
11639 } else if (key
.first
== PREFIX_OMAP
) {
11640 hist
.update_hist_entry(hist
.key_hist
, PREFIX_OMAP
, key_size
, value_size
);
11642 } else if (key
.first
== PREFIX_DEFERRED
) {
11643 hist
.update_hist_entry(hist
.key_hist
, PREFIX_DEFERRED
, key_size
, value_size
);
11645 } else if (key
.first
== PREFIX_ALLOC
|| key
.first
== "b" ) {
11646 hist
.update_hist_entry(hist
.key_hist
, PREFIX_ALLOC
, key_size
, value_size
);
11648 } else if (key
.first
== PREFIX_SHARED_BLOB
) {
11649 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SHARED_BLOB
, key_size
, value_size
);
11650 num_shared_shards
++;
11652 hist
.update_hist_entry(hist
.key_hist
, prefix_other
, key_size
, value_size
);
11658 utime_t duration
= ceph_clock_now() - start
;
11659 f
->open_object_section("rocksdb_key_value_stats");
11660 f
->dump_unsigned("num_onodes", num_onodes
);
11661 f
->dump_unsigned("num_shards", num_shards
);
11662 f
->dump_unsigned("num_super", num_super
);
11663 f
->dump_unsigned("num_coll", num_coll
);
11664 f
->dump_unsigned("num_omap", num_omap
);
11665 f
->dump_unsigned("num_deferred", num_deferred
);
11666 f
->dump_unsigned("num_alloc", num_alloc
);
11667 f
->dump_unsigned("num_stat", num_stat
);
11668 f
->dump_unsigned("num_shared_shards", num_shared_shards
);
11669 f
->dump_unsigned("num_others", num_others
);
11670 f
->dump_unsigned("max_key_size", max_key_size
);
11671 f
->dump_unsigned("max_value_size", max_value_size
);
11672 f
->dump_unsigned("total_key_size", total_key_size
);
11673 f
->dump_unsigned("total_value_size", total_value_size
);
11674 f
->close_section();
11678 dout(20) << __func__
<< " finished in " << duration
<< " seconds" << dendl
;
11682 void BlueStore::_flush_cache()
11684 dout(10) << __func__
<< dendl
;
11685 for (auto i
: cache_shards
) {
11687 assert(i
->empty());
11689 for (auto& p
: coll_map
) {
11690 if (!p
.second
->onode_map
.empty()) {
11691 derr
<< __func__
<< "stray onodes on " << p
.first
<< dendl
;
11692 p
.second
->onode_map
.dump(cct
, 0);
11694 if (!p
.second
->shared_blob_set
.empty()) {
11695 derr
<< __func__
<< " stray shared blobs on " << p
.first
<< dendl
;
11696 p
.second
->shared_blob_set
.dump(cct
, 0);
11698 assert(p
.second
->onode_map
.empty());
11699 assert(p
.second
->shared_blob_set
.empty());
11704 // For external caller.
11705 // We use a best-effort policy instead, e.g.,
11706 // we don't care if there are still some pinned onodes/data in the cache
11707 // after this command is completed.
11708 void BlueStore::flush_cache()
11710 dout(10) << __func__
<< dendl
;
11711 for (auto i
: cache_shards
) {
11716 void BlueStore::_apply_padding(uint64_t head_pad
,
11718 bufferlist
& padded
)
11721 padded
.prepend_zero(head_pad
);
11724 padded
.append_zero(tail_pad
);
11726 if (head_pad
|| tail_pad
) {
11727 dout(20) << __func__
<< " can pad head 0x" << std::hex
<< head_pad
11728 << " tail 0x" << tail_pad
<< std::dec
<< dendl
;
11729 logger
->inc(l_bluestore_write_pad_bytes
, head_pad
+ tail_pad
);
11733 // ===========================================