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 " << pretty_si_t(target_bytes
)
807 << " meta/data ratios " << target_meta_ratio
808 << " + " << target_data_ratio
<< " ("
809 << pretty_si_t(target_meta
) << " + "
810 << pretty_si_t(target_buffer
) << "), "
811 << " current " << pretty_si_t(current
) << " ("
812 << pretty_si_t(current_meta
) << " + "
813 << pretty_si_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 " << pretty_si_t(target_bytes
)
836 << " ratio " << target_meta_ratio
<< " ("
837 << pretty_si_t(target_meta
) << " + "
838 << pretty_si_t(target_buffer
) << "), "
839 << " current " << pretty_si_t(current
) << " ("
840 << pretty_si_t(current_meta
) << " + "
841 << pretty_si_t(current_buffer
) << "),"
842 << " need_to_free " << pretty_si_t(need_to_free
) << " ("
843 << pretty_si_t(free_meta
) << " + "
844 << pretty_si_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 " << prettybyte_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 " << prettybyte_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 (get_cache()) { // the dummy instances have a nullptr
1654 std::lock_guard
<std::recursive_mutex
> l(get_cache()->lock
);
1655 bc
._clear(get_cache());
1656 get_cache()->rm_blob();
1658 if (loaded
&& persistent
) {
1663 void BlueStore::SharedBlob::put()
1666 ldout(coll
->store
->cct
, 20) << __func__
<< " " << this
1667 << " removing self from set " << get_parent()
1670 if (get_parent()->try_remove(this)) {
1673 ldout(coll
->store
->cct
, 20)
1674 << __func__
<< " " << this << " lost race to remove myself from set"
1683 void BlueStore::SharedBlob::get_ref(uint64_t offset
, uint32_t length
)
1686 persistent
->ref_map
.get(offset
, length
);
1689 void BlueStore::SharedBlob::put_ref(uint64_t offset
, uint32_t length
,
1691 set
<SharedBlob
*> *maybe_unshared
)
1695 persistent
->ref_map
.put(offset
, length
, r
, maybe_unshared
? &maybe
: nullptr);
1696 if (maybe_unshared
&& maybe
) {
1697 maybe_unshared
->insert(this);
1704 #define dout_prefix *_dout << "bluestore.sharedblobset(" << this << ") "
1706 void BlueStore::SharedBlobSet::dump(CephContext
*cct
, int lvl
)
1708 std::lock_guard
<std::mutex
> l(lock
);
1709 for (auto& i
: sb_map
) {
1710 ldout(cct
, lvl
) << i
.first
<< " : " << *i
.second
<< dendl
;
1717 #define dout_prefix *_dout << "bluestore.blob(" << this << ") "
1719 ostream
& operator<<(ostream
& out
, const BlueStore::Blob
& b
)
1721 out
<< "Blob(" << &b
;
1722 if (b
.is_spanning()) {
1723 out
<< " spanning " << b
.id
;
1725 out
<< " " << b
.get_blob() << " " << b
.get_blob_use_tracker();
1726 if (b
.shared_blob
) {
1727 out
<< " " << *b
.shared_blob
;
1729 out
<< " (shared_blob=NULL)";
1735 void BlueStore::Blob::discard_unallocated(Collection
*coll
)
1737 if (get_blob().is_shared()) {
1740 if (get_blob().is_compressed()) {
1741 bool discard
= false;
1742 bool all_invalid
= true;
1743 for (auto e
: get_blob().get_extents()) {
1744 if (!e
.is_valid()) {
1747 all_invalid
= false;
1750 assert(discard
== all_invalid
); // in case of compressed blob all
1751 // or none pextents are invalid.
1753 shared_blob
->bc
.discard(shared_blob
->get_cache(), 0,
1754 get_blob().get_logical_length());
1758 for (auto e
: get_blob().get_extents()) {
1759 if (!e
.is_valid()) {
1760 ldout(coll
->store
->cct
, 20) << __func__
<< " 0x" << std::hex
<< pos
1762 << std::dec
<< dendl
;
1763 shared_blob
->bc
.discard(shared_blob
->get_cache(), pos
, e
.length
);
1767 if (get_blob().can_prune_tail()) {
1768 dirty_blob().prune_tail();
1769 used_in_blob
.prune_tail(get_blob().get_ondisk_length());
1770 auto cct
= coll
->store
->cct
; //used by dout
1771 dout(20) << __func__
<< " pruned tail, now " << get_blob() << dendl
;
1776 void BlueStore::Blob::get_ref(
1781 // Caller has to initialize Blob's logical length prior to increment
1782 // references. Otherwise one is neither unable to determine required
1783 // amount of counters in case of per-au tracking nor obtain min_release_size
1784 // for single counter mode.
1785 assert(get_blob().get_logical_length() != 0);
1786 auto cct
= coll
->store
->cct
;
1787 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1788 << std::dec
<< " " << *this << dendl
;
1790 if (used_in_blob
.is_empty()) {
1791 uint32_t min_release_size
=
1792 get_blob().get_release_size(coll
->store
->min_alloc_size
);
1793 uint64_t l
= get_blob().get_logical_length();
1794 dout(20) << __func__
<< " init 0x" << std::hex
<< l
<< ", "
1795 << min_release_size
<< std::dec
<< dendl
;
1796 used_in_blob
.init(l
, min_release_size
);
1803 bool BlueStore::Blob::put_ref(
1809 PExtentVector logical
;
1811 auto cct
= coll
->store
->cct
;
1812 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1813 << std::dec
<< " " << *this << dendl
;
1815 bool empty
= used_in_blob
.put(
1820 // nothing to release
1821 if (!empty
&& logical
.empty()) {
1825 bluestore_blob_t
& b
= dirty_blob();
1826 return b
.release_extents(empty
, logical
, r
);
1829 bool BlueStore::Blob::can_reuse_blob(uint32_t min_alloc_size
,
1830 uint32_t target_blob_size
,
1832 uint32_t *length0
) {
1833 assert(min_alloc_size
);
1834 assert(target_blob_size
);
1835 if (!get_blob().is_mutable()) {
1839 uint32_t length
= *length0
;
1840 uint32_t end
= b_offset
+ length
;
1842 // Currently for the sake of simplicity we omit blob reuse if data is
1843 // unaligned with csum chunk. Later we can perform padding if needed.
1844 if (get_blob().has_csum() &&
1845 ((b_offset
% get_blob().get_csum_chunk_size()) != 0 ||
1846 (end
% get_blob().get_csum_chunk_size()) != 0)) {
1850 auto blen
= get_blob().get_logical_length();
1851 uint32_t new_blen
= blen
;
1853 // make sure target_blob_size isn't less than current blob len
1854 target_blob_size
= MAX(blen
, target_blob_size
);
1856 if (b_offset
>= blen
) {
1857 // new data totally stands out of the existing blob
1860 // new data overlaps with the existing blob
1861 new_blen
= MAX(blen
, end
);
1863 uint32_t overlap
= 0;
1864 if (new_blen
> blen
) {
1865 overlap
= blen
- b_offset
;
1870 if (!get_blob().is_unallocated(b_offset
, overlap
)) {
1871 // abort if any piece of the overlap has already been allocated
1876 if (new_blen
> blen
) {
1877 int64_t overflow
= int64_t(new_blen
) - target_blob_size
;
1878 // Unable to decrease the provided length to fit into max_blob_size
1879 if (overflow
>= length
) {
1883 // FIXME: in some cases we could reduce unused resolution
1884 if (get_blob().has_unused()) {
1889 new_blen
-= overflow
;
1894 if (new_blen
> blen
) {
1895 dirty_blob().add_tail(new_blen
);
1896 used_in_blob
.add_tail(new_blen
,
1897 get_blob().get_release_size(min_alloc_size
));
1903 void BlueStore::Blob::split(Collection
*coll
, uint32_t blob_offset
, Blob
*r
)
1905 auto cct
= coll
->store
->cct
; //used by dout
1906 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1907 << " start " << *this << dendl
;
1908 assert(blob
.can_split());
1909 assert(used_in_blob
.can_split());
1910 bluestore_blob_t
&lb
= dirty_blob();
1911 bluestore_blob_t
&rb
= r
->dirty_blob();
1915 &(r
->used_in_blob
));
1917 lb
.split(blob_offset
, rb
);
1918 shared_blob
->bc
.split(shared_blob
->get_cache(), blob_offset
, r
->shared_blob
->bc
);
1920 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1921 << " finish " << *this << dendl
;
1922 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1923 << " and " << *r
<< dendl
;
1926 #ifndef CACHE_BLOB_BL
1927 void BlueStore::Blob::decode(
1929 bufferptr::iterator
& p
,
1932 bool include_ref_map
)
1934 denc(blob
, p
, struct_v
);
1935 if (blob
.is_shared()) {
1938 if (include_ref_map
) {
1940 used_in_blob
.decode(p
);
1942 used_in_blob
.clear();
1943 bluestore_extent_ref_map_t legacy_ref_map
;
1944 legacy_ref_map
.decode(p
);
1945 for (auto r
: legacy_ref_map
.ref_map
) {
1949 r
.second
.refs
* r
.second
.length
);
1958 ostream
& operator<<(ostream
& out
, const BlueStore::Extent
& e
)
1960 return out
<< std::hex
<< "0x" << e
.logical_offset
<< "~" << e
.length
1961 << ": 0x" << e
.blob_offset
<< "~" << e
.length
<< std::dec
1966 BlueStore::OldExtent
* BlueStore::OldExtent::create(CollectionRef c
,
1971 OldExtent
* oe
= new OldExtent(lo
, o
, l
, b
);
1972 b
->put_ref(c
.get(), o
, l
, &(oe
->r
));
1973 oe
->blob_empty
= b
->get_referenced_bytes() == 0;
1980 #define dout_prefix *_dout << "bluestore.extentmap(" << this << ") "
1982 BlueStore::ExtentMap::ExtentMap(Onode
*o
)
1985 o
->c
->store
->cct
->_conf
->bluestore_extent_map_inline_shard_prealloc_size
) {
1988 void BlueStore::ExtentMap::update(KeyValueDB::Transaction t
,
1991 auto cct
= onode
->c
->store
->cct
; //used by dout
1992 dout(20) << __func__
<< " " << onode
->oid
<< (force
? " force" : "") << dendl
;
1993 if (onode
->onode
.extent_map_shards
.empty()) {
1994 if (inline_bl
.length() == 0) {
1996 // we need to encode inline_bl to measure encoded length
1997 bool never_happen
= encode_some(0, OBJECT_MAX_SIZE
, inline_bl
, &n
);
1998 inline_bl
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
1999 assert(!never_happen
);
2000 size_t len
= inline_bl
.length();
2001 dout(20) << __func__
<< " inline shard " << len
<< " bytes from " << n
2002 << " extents" << dendl
;
2003 if (!force
&& len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
2004 request_reshard(0, OBJECT_MAX_SIZE
);
2008 // will persist in the onode key.
2010 // pending shard update
2011 struct dirty_shard_t
{
2014 dirty_shard_t(Shard
*s
) : shard(s
) {}
2016 vector
<dirty_shard_t
> encoded_shards
;
2017 // allocate slots for all shards in a single call instead of
2018 // doing multiple allocations - one per each dirty shard
2019 encoded_shards
.reserve(shards
.size());
2021 auto p
= shards
.begin();
2023 while (p
!= shards
.end()) {
2024 assert(p
->shard_info
->offset
>= prev_p
->shard_info
->offset
);
2029 if (n
== shards
.end()) {
2030 endoff
= OBJECT_MAX_SIZE
;
2032 endoff
= n
->shard_info
->offset
;
2034 encoded_shards
.emplace_back(dirty_shard_t(&(*p
)));
2035 bufferlist
& bl
= encoded_shards
.back().bl
;
2036 if (encode_some(p
->shard_info
->offset
, endoff
- p
->shard_info
->offset
,
2039 derr
<< __func__
<< " encode_some needs reshard" << dendl
;
2043 size_t len
= bl
.length();
2045 dout(20) << __func__
<< " shard 0x" << std::hex
2046 << p
->shard_info
->offset
<< std::dec
<< " is " << len
2047 << " bytes (was " << p
->shard_info
->bytes
<< ") from "
2048 << p
->extents
<< " extents" << dendl
;
2051 if (len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
2052 // we are big; reshard ourselves
2053 request_reshard(p
->shard_info
->offset
, endoff
);
2055 // avoid resharding the trailing shard, even if it is small
2056 else if (n
!= shards
.end() &&
2057 len
< g_conf
->bluestore_extent_map_shard_min_size
) {
2058 assert(endoff
!= OBJECT_MAX_SIZE
);
2059 if (p
== shards
.begin()) {
2060 // we are the first shard, combine with next shard
2061 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2063 // combine either with the previous shard or the next,
2064 // whichever is smaller
2065 if (prev_p
->shard_info
->bytes
> n
->shard_info
->bytes
) {
2066 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2068 request_reshard(prev_p
->shard_info
->offset
, endoff
);
2077 if (needs_reshard()) {
2081 // schedule DB update for dirty shards
2083 for (auto& it
: encoded_shards
) {
2084 it
.shard
->dirty
= false;
2085 it
.shard
->shard_info
->bytes
= it
.bl
.length();
2086 generate_extent_shard_key_and_apply(
2088 it
.shard
->shard_info
->offset
,
2090 [&](const string
& final_key
) {
2091 t
->set(PREFIX_OBJ
, final_key
, it
.bl
);
2098 bid_t
BlueStore::ExtentMap::allocate_spanning_blob_id()
2100 if (spanning_blob_map
.empty())
2102 bid_t bid
= spanning_blob_map
.rbegin()->first
+ 1;
2103 // bid is valid and available.
2106 // Find next unused bid;
2107 bid
= rand() % (numeric_limits
<bid_t
>::max() + 1);
2108 const auto begin_bid
= bid
;
2110 if (!spanning_blob_map
.count(bid
))
2114 if (bid
< 0) bid
= 0;
2116 } while (bid
!= begin_bid
);
2117 assert(0 == "no available blob id");
2120 void BlueStore::ExtentMap::reshard(
2122 KeyValueDB::Transaction t
)
2124 auto cct
= onode
->c
->store
->cct
; // used by dout
2126 dout(10) << __func__
<< " 0x[" << std::hex
<< needs_reshard_begin
<< ","
2127 << needs_reshard_end
<< ")" << std::dec
2128 << " of " << onode
->onode
.extent_map_shards
.size()
2129 << " shards on " << onode
->oid
<< dendl
;
2130 for (auto& p
: spanning_blob_map
) {
2131 dout(20) << __func__
<< " spanning blob " << p
.first
<< " " << *p
.second
2134 // determine shard index range
2135 unsigned si_begin
= 0, si_end
= 0;
2136 if (!shards
.empty()) {
2137 while (si_begin
+ 1 < shards
.size() &&
2138 shards
[si_begin
+ 1].shard_info
->offset
<= needs_reshard_begin
) {
2141 needs_reshard_begin
= shards
[si_begin
].shard_info
->offset
;
2142 for (si_end
= si_begin
; si_end
< shards
.size(); ++si_end
) {
2143 if (shards
[si_end
].shard_info
->offset
>= needs_reshard_end
) {
2144 needs_reshard_end
= shards
[si_end
].shard_info
->offset
;
2148 if (si_end
== shards
.size()) {
2149 needs_reshard_end
= OBJECT_MAX_SIZE
;
2151 dout(20) << __func__
<< " shards [" << si_begin
<< "," << si_end
<< ")"
2152 << " over 0x[" << std::hex
<< needs_reshard_begin
<< ","
2153 << needs_reshard_end
<< ")" << std::dec
<< dendl
;
2156 fault_range(db
, needs_reshard_begin
, (needs_reshard_end
- needs_reshard_begin
));
2158 // we may need to fault in a larger interval later must have all
2159 // referring extents for spanning blobs loaded in order to have
2160 // accurate use_tracker values.
2161 uint32_t spanning_scan_begin
= needs_reshard_begin
;
2162 uint32_t spanning_scan_end
= needs_reshard_end
;
2166 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2167 generate_extent_shard_key_and_apply(
2168 onode
->key
, shards
[i
].shard_info
->offset
, &key
,
2169 [&](const string
& final_key
) {
2170 t
->rmkey(PREFIX_OBJ
, final_key
);
2175 // calculate average extent size
2177 unsigned extents
= 0;
2178 if (onode
->onode
.extent_map_shards
.empty()) {
2179 bytes
= inline_bl
.length();
2180 extents
= extent_map
.size();
2182 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2183 bytes
+= shards
[i
].shard_info
->bytes
;
2184 extents
+= shards
[i
].extents
;
2187 unsigned target
= cct
->_conf
->bluestore_extent_map_shard_target_size
;
2188 unsigned slop
= target
*
2189 cct
->_conf
->bluestore_extent_map_shard_target_size_slop
;
2190 unsigned extent_avg
= bytes
/ MAX(1, extents
);
2191 dout(20) << __func__
<< " extent_avg " << extent_avg
<< ", target " << target
2192 << ", slop " << slop
<< dendl
;
2195 unsigned estimate
= 0;
2196 unsigned offset
= needs_reshard_begin
;
2197 vector
<bluestore_onode_t::shard_info
> new_shard_info
;
2198 unsigned max_blob_end
= 0;
2199 Extent
dummy(needs_reshard_begin
);
2200 for (auto e
= extent_map
.lower_bound(dummy
);
2201 e
!= extent_map
.end();
2203 if (e
->logical_offset
>= needs_reshard_end
) {
2206 dout(30) << " extent " << *e
<< dendl
;
2208 // disfavor shard boundaries that span a blob
2209 bool would_span
= (e
->logical_offset
< max_blob_end
) || e
->blob_offset
;
2211 estimate
+ extent_avg
> target
+ (would_span
? slop
: 0)) {
2213 if (offset
== needs_reshard_begin
) {
2214 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2215 new_shard_info
.back().offset
= offset
;
2216 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2217 << std::dec
<< dendl
;
2219 offset
= e
->logical_offset
;
2220 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2221 new_shard_info
.back().offset
= offset
;
2222 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2223 << std::dec
<< dendl
;
2226 estimate
+= extent_avg
;
2227 unsigned bs
= e
->blob_start();
2228 if (bs
< spanning_scan_begin
) {
2229 spanning_scan_begin
= bs
;
2231 uint32_t be
= e
->blob_end();
2232 if (be
> max_blob_end
) {
2235 if (be
> spanning_scan_end
) {
2236 spanning_scan_end
= be
;
2239 if (new_shard_info
.empty() && (si_begin
> 0 ||
2240 si_end
< shards
.size())) {
2241 // we resharded a partial range; we must produce at least one output
2243 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2244 new_shard_info
.back().offset
= needs_reshard_begin
;
2245 dout(20) << __func__
<< " new shard 0x" << std::hex
<< needs_reshard_begin
2246 << std::dec
<< " (singleton degenerate case)" << dendl
;
2249 auto& sv
= onode
->onode
.extent_map_shards
;
2250 dout(20) << __func__
<< " new " << new_shard_info
<< dendl
;
2251 dout(20) << __func__
<< " old " << sv
<< dendl
;
2253 // no old shards to keep
2254 sv
.swap(new_shard_info
);
2255 init_shards(true, true);
2257 // splice in new shards
2258 sv
.erase(sv
.begin() + si_begin
, sv
.begin() + si_end
);
2259 shards
.erase(shards
.begin() + si_begin
, shards
.begin() + si_end
);
2261 sv
.begin() + si_begin
,
2262 new_shard_info
.begin(),
2263 new_shard_info
.end());
2264 shards
.insert(shards
.begin() + si_begin
, new_shard_info
.size(), Shard());
2265 si_end
= si_begin
+ new_shard_info
.size();
2267 assert(sv
.size() == shards
.size());
2269 // note that we need to update every shard_info of shards here,
2270 // as sv might have been totally re-allocated above
2271 for (unsigned i
= 0; i
< shards
.size(); i
++) {
2272 shards
[i
].shard_info
= &sv
[i
];
2275 // mark newly added shards as dirty
2276 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2277 shards
[i
].loaded
= true;
2278 shards
[i
].dirty
= true;
2281 dout(20) << __func__
<< " fin " << sv
<< dendl
;
2285 // no more shards; unspan all previously spanning blobs
2286 auto p
= spanning_blob_map
.begin();
2287 while (p
!= spanning_blob_map
.end()) {
2289 dout(30) << __func__
<< " un-spanning " << *p
->second
<< dendl
;
2290 p
= spanning_blob_map
.erase(p
);
2293 // identify new spanning blobs
2294 dout(20) << __func__
<< " checking spanning blobs 0x[" << std::hex
2295 << spanning_scan_begin
<< "," << spanning_scan_end
<< ")" << dendl
;
2296 if (spanning_scan_begin
< needs_reshard_begin
) {
2297 fault_range(db
, spanning_scan_begin
,
2298 needs_reshard_begin
- spanning_scan_begin
);
2300 if (spanning_scan_end
> needs_reshard_end
) {
2301 fault_range(db
, needs_reshard_end
,
2302 spanning_scan_end
- needs_reshard_end
);
2304 auto sp
= sv
.begin() + si_begin
;
2305 auto esp
= sv
.end();
2306 unsigned shard_start
= sp
->offset
;
2310 shard_end
= OBJECT_MAX_SIZE
;
2312 shard_end
= sp
->offset
;
2314 Extent
dummy(needs_reshard_begin
);
2315 for (auto e
= extent_map
.lower_bound(dummy
); e
!= extent_map
.end(); ++e
) {
2316 if (e
->logical_offset
>= needs_reshard_end
) {
2319 dout(30) << " extent " << *e
<< dendl
;
2320 while (e
->logical_offset
>= shard_end
) {
2321 shard_start
= shard_end
;
2325 shard_end
= OBJECT_MAX_SIZE
;
2327 shard_end
= sp
->offset
;
2329 dout(30) << __func__
<< " shard 0x" << std::hex
<< shard_start
2330 << " to 0x" << shard_end
<< std::dec
<< dendl
;
2332 if (e
->blob_escapes_range(shard_start
, shard_end
- shard_start
)) {
2333 if (!e
->blob
->is_spanning()) {
2334 // We have two options: (1) split the blob into pieces at the
2335 // shard boundaries (and adjust extents accordingly), or (2)
2336 // mark it spanning. We prefer to cut the blob if we can. Note that
2337 // we may have to split it multiple times--potentially at every
2339 bool must_span
= false;
2340 BlobRef b
= e
->blob
;
2341 if (b
->can_split()) {
2342 uint32_t bstart
= e
->blob_start();
2343 uint32_t bend
= e
->blob_end();
2344 for (const auto& sh
: shards
) {
2345 if (bstart
< sh
.shard_info
->offset
&&
2346 bend
> sh
.shard_info
->offset
) {
2347 uint32_t blob_offset
= sh
.shard_info
->offset
- bstart
;
2348 if (b
->can_split_at(blob_offset
)) {
2349 dout(20) << __func__
<< " splitting blob, bstart 0x"
2350 << std::hex
<< bstart
<< " blob_offset 0x"
2351 << blob_offset
<< std::dec
<< " " << *b
<< dendl
;
2352 b
= split_blob(b
, blob_offset
, sh
.shard_info
->offset
);
2353 // switch b to the new right-hand side, in case it
2354 // *also* has to get split.
2355 bstart
+= blob_offset
;
2356 onode
->c
->store
->logger
->inc(l_bluestore_blob_split
);
2367 auto bid
= allocate_spanning_blob_id();
2369 spanning_blob_map
[b
->id
] = b
;
2370 dout(20) << __func__
<< " adding spanning " << *b
<< dendl
;
2374 if (e
->blob
->is_spanning()) {
2375 spanning_blob_map
.erase(e
->blob
->id
);
2377 dout(30) << __func__
<< " un-spanning " << *e
->blob
<< dendl
;
2383 clear_needs_reshard();
2386 bool BlueStore::ExtentMap::encode_some(
2392 auto cct
= onode
->c
->store
->cct
; //used by dout
2393 Extent
dummy(offset
);
2394 auto start
= extent_map
.lower_bound(dummy
);
2395 uint32_t end
= offset
+ length
;
2397 __u8 struct_v
= 2; // Version 2 differs from v1 in blob's ref_map
2398 // serialization only. Hence there is no specific
2399 // handling at ExtentMap level.
2403 bool must_reshard
= false;
2404 for (auto p
= start
;
2405 p
!= extent_map
.end() && p
->logical_offset
< end
;
2407 assert(p
->logical_offset
>= offset
);
2408 p
->blob
->last_encoded_id
= -1;
2409 if (!p
->blob
->is_spanning() && p
->blob_escapes_range(offset
, length
)) {
2410 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2411 << std::dec
<< " hit new spanning blob " << *p
<< dendl
;
2412 request_reshard(p
->blob_start(), p
->blob_end());
2413 must_reshard
= true;
2415 if (!must_reshard
) {
2416 denc_varint(0, bound
); // blobid
2417 denc_varint(0, bound
); // logical_offset
2418 denc_varint(0, bound
); // len
2419 denc_varint(0, bound
); // blob_offset
2421 p
->blob
->bound_encode(
2424 p
->blob
->shared_blob
->get_sbid(),
2432 denc(struct_v
, bound
);
2433 denc_varint(0, bound
); // number of extents
2436 auto app
= bl
.get_contiguous_appender(bound
);
2437 denc(struct_v
, app
);
2438 denc_varint(n
, app
);
2445 uint64_t prev_len
= 0;
2446 for (auto p
= start
;
2447 p
!= extent_map
.end() && p
->logical_offset
< end
;
2450 bool include_blob
= false;
2451 if (p
->blob
->is_spanning()) {
2452 blobid
= p
->blob
->id
<< BLOBID_SHIFT_BITS
;
2453 blobid
|= BLOBID_FLAG_SPANNING
;
2454 } else if (p
->blob
->last_encoded_id
< 0) {
2455 p
->blob
->last_encoded_id
= n
+ 1; // so it is always non-zero
2456 include_blob
= true;
2457 blobid
= 0; // the decoder will infer the id from n
2459 blobid
= p
->blob
->last_encoded_id
<< BLOBID_SHIFT_BITS
;
2461 if (p
->logical_offset
== pos
) {
2462 blobid
|= BLOBID_FLAG_CONTIGUOUS
;
2464 if (p
->blob_offset
== 0) {
2465 blobid
|= BLOBID_FLAG_ZEROOFFSET
;
2467 if (p
->length
== prev_len
) {
2468 blobid
|= BLOBID_FLAG_SAMELENGTH
;
2470 prev_len
= p
->length
;
2472 denc_varint(blobid
, app
);
2473 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2474 denc_varint_lowz(p
->logical_offset
- pos
, app
);
2476 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2477 denc_varint_lowz(p
->blob_offset
, app
);
2479 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2480 denc_varint_lowz(p
->length
, app
);
2482 pos
= p
->logical_end();
2484 p
->blob
->encode(app
, struct_v
, p
->blob
->shared_blob
->get_sbid(), false);
2488 /*derr << __func__ << bl << dendl;
2489 derr << __func__ << ":";
2496 unsigned BlueStore::ExtentMap::decode_some(bufferlist
& bl
)
2498 auto cct
= onode
->c
->store
->cct
; //used by dout
2500 derr << __func__ << ":";
2505 assert(bl
.get_num_buffers() <= 1);
2506 auto p
= bl
.front().begin_deep();
2509 // Version 2 differs from v1 in blob's ref_map
2510 // serialization only. Hence there is no specific
2511 // handling at ExtentMap level below.
2512 assert(struct_v
== 1 || struct_v
== 2);
2515 denc_varint(num
, p
);
2516 vector
<BlobRef
> blobs(num
);
2518 uint64_t prev_len
= 0;
2522 Extent
*le
= new Extent();
2524 denc_varint(blobid
, p
);
2525 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2527 denc_varint_lowz(gap
, p
);
2530 le
->logical_offset
= pos
;
2531 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2532 denc_varint_lowz(le
->blob_offset
, p
);
2534 le
->blob_offset
= 0;
2536 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2537 denc_varint_lowz(prev_len
, p
);
2539 le
->length
= prev_len
;
2541 if (blobid
& BLOBID_FLAG_SPANNING
) {
2542 dout(30) << __func__
<< " getting spanning blob "
2543 << (blobid
>> BLOBID_SHIFT_BITS
) << dendl
;
2544 le
->assign_blob(get_spanning_blob(blobid
>> BLOBID_SHIFT_BITS
));
2546 blobid
>>= BLOBID_SHIFT_BITS
;
2548 le
->assign_blob(blobs
[blobid
- 1]);
2551 Blob
*b
= new Blob();
2553 b
->decode(onode
->c
, p
, struct_v
, &sbid
, false);
2555 onode
->c
->open_shared_blob(sbid
, b
);
2558 // we build ref_map dynamically for non-spanning blobs
2566 extent_map
.insert(*le
);
2573 void BlueStore::ExtentMap::bound_encode_spanning_blobs(size_t& p
)
2575 // Version 2 differs from v1 in blob's ref_map
2576 // serialization only. Hence there is no specific
2577 // handling at ExtentMap level.
2581 denc_varint((uint32_t)0, p
);
2582 size_t key_size
= 0;
2583 denc_varint((uint32_t)0, key_size
);
2584 p
+= spanning_blob_map
.size() * key_size
;
2585 for (const auto& i
: spanning_blob_map
) {
2586 i
.second
->bound_encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2590 void BlueStore::ExtentMap::encode_spanning_blobs(
2591 bufferlist::contiguous_appender
& p
)
2593 // Version 2 differs from v1 in blob's ref_map
2594 // serialization only. Hence there is no specific
2595 // handling at ExtentMap level.
2599 denc_varint(spanning_blob_map
.size(), p
);
2600 for (auto& i
: spanning_blob_map
) {
2601 denc_varint(i
.second
->id
, p
);
2602 i
.second
->encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2606 void BlueStore::ExtentMap::decode_spanning_blobs(
2607 bufferptr::iterator
& p
)
2611 // Version 2 differs from v1 in blob's ref_map
2612 // serialization only. Hence there is no specific
2613 // handling at ExtentMap level.
2614 assert(struct_v
== 1 || struct_v
== 2);
2619 BlobRef
b(new Blob());
2620 denc_varint(b
->id
, p
);
2621 spanning_blob_map
[b
->id
] = b
;
2623 b
->decode(onode
->c
, p
, struct_v
, &sbid
, true);
2624 onode
->c
->open_shared_blob(sbid
, b
);
2628 void BlueStore::ExtentMap::init_shards(bool loaded
, bool dirty
)
2630 shards
.resize(onode
->onode
.extent_map_shards
.size());
2632 for (auto &s
: onode
->onode
.extent_map_shards
) {
2633 shards
[i
].shard_info
= &s
;
2634 shards
[i
].loaded
= loaded
;
2635 shards
[i
].dirty
= dirty
;
2640 void BlueStore::ExtentMap::fault_range(
2645 auto cct
= onode
->c
->store
->cct
; //used by dout
2646 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2647 << std::dec
<< dendl
;
2648 auto start
= seek_shard(offset
);
2649 auto last
= seek_shard(offset
+ length
);
2654 assert(last
>= start
);
2656 while (start
<= last
) {
2657 assert((size_t)start
< shards
.size());
2658 auto p
= &shards
[start
];
2660 dout(30) << __func__
<< " opening shard 0x" << std::hex
2661 << p
->shard_info
->offset
<< std::dec
<< dendl
;
2663 generate_extent_shard_key_and_apply(
2664 onode
->key
, p
->shard_info
->offset
, &key
,
2665 [&](const string
& final_key
) {
2666 int r
= db
->get(PREFIX_OBJ
, final_key
, &v
);
2668 derr
<< __func__
<< " missing shard 0x" << std::hex
2669 << p
->shard_info
->offset
<< std::dec
<< " for " << onode
->oid
2675 p
->extents
= decode_some(v
);
2677 dout(20) << __func__
<< " open shard 0x" << std::hex
2678 << p
->shard_info
->offset
<< std::dec
2679 << " (" << v
.length() << " bytes)" << dendl
;
2680 assert(p
->dirty
== false);
2681 assert(v
.length() == p
->shard_info
->bytes
);
2682 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_misses
);
2684 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_hits
);
2690 void BlueStore::ExtentMap::dirty_range(
2694 auto cct
= onode
->c
->store
->cct
; //used by dout
2695 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2696 << std::dec
<< dendl
;
2697 if (shards
.empty()) {
2698 dout(20) << __func__
<< " mark inline shard dirty" << dendl
;
2702 auto start
= seek_shard(offset
);
2703 auto last
= seek_shard(offset
+ length
);
2707 assert(last
>= start
);
2708 while (start
<= last
) {
2709 assert((size_t)start
< shards
.size());
2710 auto p
= &shards
[start
];
2712 dout(20) << __func__
<< " shard 0x" << std::hex
<< p
->shard_info
->offset
2713 << std::dec
<< " is not loaded, can't mark dirty" << dendl
;
2714 assert(0 == "can't mark unloaded shard dirty");
2717 dout(20) << __func__
<< " mark shard 0x" << std::hex
2718 << p
->shard_info
->offset
<< std::dec
<< " dirty" << dendl
;
2725 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::find(
2728 Extent
dummy(offset
);
2729 return extent_map
.find(dummy
);
2732 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::seek_lextent(
2735 Extent
dummy(offset
);
2736 auto fp
= extent_map
.lower_bound(dummy
);
2737 if (fp
!= extent_map
.begin()) {
2739 if (fp
->logical_end() <= offset
) {
2746 BlueStore::extent_map_t::const_iterator
BlueStore::ExtentMap::seek_lextent(
2747 uint64_t offset
) const
2749 Extent
dummy(offset
);
2750 auto fp
= extent_map
.lower_bound(dummy
);
2751 if (fp
!= extent_map
.begin()) {
2753 if (fp
->logical_end() <= offset
) {
2760 bool BlueStore::ExtentMap::has_any_lextents(uint64_t offset
, uint64_t length
)
2762 auto fp
= seek_lextent(offset
);
2763 if (fp
== extent_map
.end() || fp
->logical_offset
>= offset
+ length
) {
2769 int BlueStore::ExtentMap::compress_extent_map(
2773 auto cct
= onode
->c
->store
->cct
; //used by dout
2774 if (extent_map
.empty())
2777 auto p
= seek_lextent(offset
);
2778 if (p
!= extent_map
.begin()) {
2779 --p
; // start to the left of offset
2781 // the caller should have just written to this region
2782 assert(p
!= extent_map
.end());
2784 // identify the *next* shard
2785 auto pshard
= shards
.begin();
2786 while (pshard
!= shards
.end() &&
2787 p
->logical_offset
>= pshard
->shard_info
->offset
) {
2791 if (pshard
!= shards
.end()) {
2792 shard_end
= pshard
->shard_info
->offset
;
2794 shard_end
= OBJECT_MAX_SIZE
;
2798 for (++n
; n
!= extent_map
.end(); p
= n
++) {
2799 if (n
->logical_offset
> offset
+ length
) {
2800 break; // stop after end
2802 while (n
!= extent_map
.end() &&
2803 p
->logical_end() == n
->logical_offset
&&
2804 p
->blob
== n
->blob
&&
2805 p
->blob_offset
+ p
->length
== n
->blob_offset
&&
2806 n
->logical_offset
< shard_end
) {
2807 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2808 << " next shard 0x" << shard_end
<< std::dec
2809 << " merging " << *p
<< " and " << *n
<< dendl
;
2810 p
->length
+= n
->length
;
2814 if (n
== extent_map
.end()) {
2817 if (n
->logical_offset
>= shard_end
) {
2818 assert(pshard
!= shards
.end());
2820 if (pshard
!= shards
.end()) {
2821 shard_end
= pshard
->shard_info
->offset
;
2823 shard_end
= OBJECT_MAX_SIZE
;
2827 if (removed
&& onode
) {
2828 onode
->c
->store
->logger
->inc(l_bluestore_extent_compress
, removed
);
2833 void BlueStore::ExtentMap::punch_hole(
2837 old_extent_map_t
*old_extents
)
2839 auto p
= seek_lextent(offset
);
2840 uint64_t end
= offset
+ length
;
2841 while (p
!= extent_map
.end()) {
2842 if (p
->logical_offset
>= end
) {
2845 if (p
->logical_offset
< offset
) {
2846 if (p
->logical_end() > end
) {
2847 // split and deref middle
2848 uint64_t front
= offset
- p
->logical_offset
;
2849 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ front
,
2851 old_extents
->push_back(*oe
);
2853 p
->blob_offset
+ front
+ length
,
2854 p
->length
- front
- length
,
2860 assert(p
->logical_end() > offset
); // else seek_lextent bug
2861 uint64_t keep
= offset
- p
->logical_offset
;
2862 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ keep
,
2863 p
->length
- keep
, p
->blob
);
2864 old_extents
->push_back(*oe
);
2870 if (p
->logical_offset
+ p
->length
<= end
) {
2871 // deref whole lextent
2872 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2873 p
->length
, p
->blob
);
2874 old_extents
->push_back(*oe
);
2879 uint64_t keep
= p
->logical_end() - end
;
2880 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2881 p
->length
- keep
, p
->blob
);
2882 old_extents
->push_back(*oe
);
2884 add(end
, p
->blob_offset
+ p
->length
- keep
, keep
, p
->blob
);
2890 BlueStore::Extent
*BlueStore::ExtentMap::set_lextent(
2892 uint64_t logical_offset
,
2893 uint64_t blob_offset
, uint64_t length
, BlobRef b
,
2894 old_extent_map_t
*old_extents
)
2896 // We need to have completely initialized Blob to increment its ref counters.
2897 assert(b
->get_blob().get_logical_length() != 0);
2899 // Do get_ref prior to punch_hole to prevent from putting reused blob into
2900 // old_extents list if we overwre the blob totally
2901 // This might happen during WAL overwrite.
2902 b
->get_ref(onode
->c
, blob_offset
, length
);
2905 punch_hole(c
, logical_offset
, length
, old_extents
);
2908 Extent
*le
= new Extent(logical_offset
, blob_offset
, length
, b
);
2909 extent_map
.insert(*le
);
2910 if (spans_shard(logical_offset
, length
)) {
2911 request_reshard(logical_offset
, logical_offset
+ length
);
2916 BlueStore::BlobRef
BlueStore::ExtentMap::split_blob(
2918 uint32_t blob_offset
,
2921 auto cct
= onode
->c
->store
->cct
; //used by dout
2923 uint32_t end_pos
= pos
+ lb
->get_blob().get_logical_length() - blob_offset
;
2924 dout(20) << __func__
<< " 0x" << std::hex
<< pos
<< " end 0x" << end_pos
2925 << " blob_offset 0x" << blob_offset
<< std::dec
<< " " << *lb
2927 BlobRef rb
= onode
->c
->new_blob();
2928 lb
->split(onode
->c
, blob_offset
, rb
.get());
2930 for (auto ep
= seek_lextent(pos
);
2931 ep
!= extent_map
.end() && ep
->logical_offset
< end_pos
;
2933 if (ep
->blob
!= lb
) {
2936 if (ep
->logical_offset
< pos
) {
2938 size_t left
= pos
- ep
->logical_offset
;
2939 Extent
*ne
= new Extent(pos
, 0, ep
->length
- left
, rb
);
2940 extent_map
.insert(*ne
);
2942 dout(30) << __func__
<< " split " << *ep
<< dendl
;
2943 dout(30) << __func__
<< " to " << *ne
<< dendl
;
2946 assert(ep
->blob_offset
>= blob_offset
);
2949 ep
->blob_offset
-= blob_offset
;
2950 dout(30) << __func__
<< " adjusted " << *ep
<< dendl
;
2959 #define dout_prefix *_dout << "bluestore.onode(" << this << ")." << __func__ << " "
2961 void BlueStore::Onode::flush()
2963 if (flushing_count
.load()) {
2964 ldout(c
->store
->cct
, 20) << __func__
<< " cnt:" << flushing_count
<< dendl
;
2965 std::unique_lock
<std::mutex
> l(flush_lock
);
2966 while (flushing_count
.load()) {
2970 ldout(c
->store
->cct
, 20) << __func__
<< " done" << dendl
;
2973 // =======================================================
2976 /// Checks for writes to the same pextent within a blob
2977 bool BlueStore::WriteContext::has_conflict(
2981 uint64_t min_alloc_size
)
2983 assert((loffs
% min_alloc_size
) == 0);
2984 assert((loffs_end
% min_alloc_size
) == 0);
2985 for (auto w
: writes
) {
2987 auto loffs2
= P2ALIGN(w
.logical_offset
, min_alloc_size
);
2988 auto loffs2_end
= P2ROUNDUP(w
.logical_offset
+ w
.length0
, min_alloc_size
);
2989 if ((loffs
<= loffs2
&& loffs_end
> loffs2
) ||
2990 (loffs
>= loffs2
&& loffs
< loffs2_end
)) {
2998 // =======================================================
3002 #define dout_prefix *_dout << "bluestore.DeferredBatch(" << this << ") "
3004 void BlueStore::DeferredBatch::prepare_write(
3006 uint64_t seq
, uint64_t offset
, uint64_t length
,
3007 bufferlist::const_iterator
& blp
)
3009 _discard(cct
, offset
, length
);
3010 auto i
= iomap
.insert(make_pair(offset
, deferred_io()));
3011 assert(i
.second
); // this should be a new insertion
3012 i
.first
->second
.seq
= seq
;
3013 blp
.copy(length
, i
.first
->second
.bl
);
3014 i
.first
->second
.bl
.reassign_to_mempool(
3015 mempool::mempool_bluestore_writing_deferred
);
3016 dout(20) << __func__
<< " seq " << seq
3017 << " 0x" << std::hex
<< offset
<< "~" << length
3018 << " crc " << i
.first
->second
.bl
.crc32c(-1)
3019 << std::dec
<< dendl
;
3020 seq_bytes
[seq
] += length
;
3021 #ifdef DEBUG_DEFERRED
3026 void BlueStore::DeferredBatch::_discard(
3027 CephContext
*cct
, uint64_t offset
, uint64_t length
)
3029 generic_dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
3030 << std::dec
<< dendl
;
3031 auto p
= iomap
.lower_bound(offset
);
3032 if (p
!= iomap
.begin()) {
3034 auto end
= p
->first
+ p
->second
.bl
.length();
3037 head
.substr_of(p
->second
.bl
, 0, offset
- p
->first
);
3038 dout(20) << __func__
<< " keep head " << p
->second
.seq
3039 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3040 << " -> 0x" << head
.length() << std::dec
<< dendl
;
3041 auto i
= seq_bytes
.find(p
->second
.seq
);
3042 assert(i
!= seq_bytes
.end());
3043 if (end
> offset
+ length
) {
3045 tail
.substr_of(p
->second
.bl
, offset
+ length
- p
->first
,
3046 end
- (offset
+ length
));
3047 dout(20) << __func__
<< " keep tail " << p
->second
.seq
3048 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3049 << " -> 0x" << tail
.length() << std::dec
<< dendl
;
3050 auto &n
= iomap
[offset
+ length
];
3052 n
.seq
= p
->second
.seq
;
3053 i
->second
-= length
;
3055 i
->second
-= end
- offset
;
3057 assert(i
->second
>= 0);
3058 p
->second
.bl
.swap(head
);
3062 while (p
!= iomap
.end()) {
3063 if (p
->first
>= offset
+ length
) {
3066 auto i
= seq_bytes
.find(p
->second
.seq
);
3067 assert(i
!= seq_bytes
.end());
3068 auto end
= p
->first
+ p
->second
.bl
.length();
3069 if (end
> offset
+ length
) {
3070 unsigned drop_front
= offset
+ length
- p
->first
;
3071 unsigned keep_tail
= end
- (offset
+ length
);
3072 dout(20) << __func__
<< " truncate front " << p
->second
.seq
3073 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3074 << " drop_front 0x" << drop_front
<< " keep_tail 0x" << keep_tail
3075 << " to 0x" << (offset
+ length
) << "~" << keep_tail
3076 << std::dec
<< dendl
;
3077 auto &s
= iomap
[offset
+ length
];
3078 s
.seq
= p
->second
.seq
;
3079 s
.bl
.substr_of(p
->second
.bl
, drop_front
, keep_tail
);
3080 i
->second
-= drop_front
;
3082 dout(20) << __func__
<< " drop " << p
->second
.seq
3083 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3084 << std::dec
<< dendl
;
3085 i
->second
-= p
->second
.bl
.length();
3087 assert(i
->second
>= 0);
3092 void BlueStore::DeferredBatch::_audit(CephContext
*cct
)
3094 map
<uint64_t,int> sb
;
3095 for (auto p
: seq_bytes
) {
3096 sb
[p
.first
] = 0; // make sure we have the same set of keys
3099 for (auto& p
: iomap
) {
3100 assert(p
.first
>= pos
);
3101 sb
[p
.second
.seq
] += p
.second
.bl
.length();
3102 pos
= p
.first
+ p
.second
.bl
.length();
3104 assert(sb
== seq_bytes
);
3111 #define dout_prefix *_dout << "bluestore(" << store->path << ").collection(" << cid << " " << this << ") "
3113 BlueStore::Collection::Collection(BlueStore
*ns
, Cache
*c
, coll_t cid
)
3117 lock("BlueStore::Collection::lock", true, false),
3123 void BlueStore::Collection::open_shared_blob(uint64_t sbid
, BlobRef b
)
3125 assert(!b
->shared_blob
);
3126 const bluestore_blob_t
& blob
= b
->get_blob();
3127 if (!blob
.is_shared()) {
3128 b
->shared_blob
= new SharedBlob(this);
3132 b
->shared_blob
= shared_blob_set
.lookup(sbid
);
3133 if (b
->shared_blob
) {
3134 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3135 << std::dec
<< " had " << *b
->shared_blob
<< dendl
;
3137 b
->shared_blob
= new SharedBlob(sbid
, this);
3138 shared_blob_set
.add(this, b
->shared_blob
.get());
3139 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3140 << std::dec
<< " opened " << *b
->shared_blob
3145 void BlueStore::Collection::load_shared_blob(SharedBlobRef sb
)
3147 if (!sb
->is_loaded()) {
3151 auto sbid
= sb
->get_sbid();
3152 get_shared_blob_key(sbid
, &key
);
3153 int r
= store
->db
->get(PREFIX_SHARED_BLOB
, key
, &v
);
3155 lderr(store
->cct
) << __func__
<< " sbid 0x" << std::hex
<< sbid
3156 << std::dec
<< " not found at key "
3157 << pretty_binary_string(key
) << dendl
;
3158 assert(0 == "uh oh, missing shared_blob");
3162 sb
->persistent
= new bluestore_shared_blob_t(sbid
);
3163 bufferlist::iterator p
= v
.begin();
3164 ::decode(*(sb
->persistent
), p
);
3165 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3166 << std::dec
<< " loaded shared_blob " << *sb
<< dendl
;
3170 void BlueStore::Collection::make_blob_shared(uint64_t sbid
, BlobRef b
)
3172 ldout(store
->cct
, 10) << __func__
<< " " << *b
<< dendl
;
3173 assert(!b
->shared_blob
->is_loaded());
3176 bluestore_blob_t
& blob
= b
->dirty_blob();
3177 blob
.set_flag(bluestore_blob_t::FLAG_SHARED
);
3179 // update shared blob
3180 b
->shared_blob
->loaded
= true;
3181 b
->shared_blob
->persistent
= new bluestore_shared_blob_t(sbid
);
3182 shared_blob_set
.add(this, b
->shared_blob
.get());
3183 for (auto p
: blob
.get_extents()) {
3185 b
->shared_blob
->get_ref(
3190 ldout(store
->cct
, 20) << __func__
<< " now " << *b
<< dendl
;
3193 uint64_t BlueStore::Collection::make_blob_unshared(SharedBlob
*sb
)
3195 ldout(store
->cct
, 10) << __func__
<< " " << *sb
<< dendl
;
3196 assert(sb
->is_loaded());
3198 uint64_t sbid
= sb
->get_sbid();
3199 shared_blob_set
.remove(sb
);
3201 delete sb
->persistent
;
3202 sb
->sbid_unloaded
= 0;
3203 ldout(store
->cct
, 20) << __func__
<< " now " << *sb
<< dendl
;
3207 BlueStore::OnodeRef
BlueStore::Collection::get_onode(
3208 const ghobject_t
& oid
,
3211 assert(create
? lock
.is_wlocked() : lock
.is_locked());
3214 if (cid
.is_pg(&pgid
)) {
3215 if (!oid
.match(cnode
.bits
, pgid
.ps())) {
3216 lderr(store
->cct
) << __func__
<< " oid " << oid
<< " not part of "
3217 << pgid
<< " bits " << cnode
.bits
<< dendl
;
3222 OnodeRef o
= onode_map
.lookup(oid
);
3226 mempool::bluestore_cache_other::string key
;
3227 get_object_key(store
->cct
, oid
, &key
);
3229 ldout(store
->cct
, 20) << __func__
<< " oid " << oid
<< " key "
3230 << pretty_binary_string(key
) << dendl
;
3233 int r
= store
->db
->get(PREFIX_OBJ
, key
.c_str(), key
.size(), &v
);
3234 ldout(store
->cct
, 20) << " r " << r
<< " v.len " << v
.length() << dendl
;
3236 if (v
.length() == 0) {
3237 assert(r
== -ENOENT
);
3238 if (!store
->cct
->_conf
->bluestore_debug_misc
&&
3242 // new object, new onode
3243 on
= new Onode(this, oid
, key
);
3247 on
= new Onode(this, oid
, key
);
3249 bufferptr::iterator p
= v
.front().begin_deep();
3250 on
->onode
.decode(p
);
3251 for (auto& i
: on
->onode
.attrs
) {
3252 i
.second
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
3255 // initialize extent_map
3256 on
->extent_map
.decode_spanning_blobs(p
);
3257 if (on
->onode
.extent_map_shards
.empty()) {
3258 denc(on
->extent_map
.inline_bl
, p
);
3259 on
->extent_map
.decode_some(on
->extent_map
.inline_bl
);
3260 on
->extent_map
.inline_bl
.reassign_to_mempool(
3261 mempool::mempool_bluestore_cache_other
);
3263 on
->extent_map
.init_shards(false, false);
3267 return onode_map
.add(oid
, o
);
3270 void BlueStore::Collection::split_cache(
3273 ldout(store
->cct
, 10) << __func__
<< " to " << dest
<< dendl
;
3275 // lock (one or both) cache shards
3276 std::lock(cache
->lock
, dest
->cache
->lock
);
3277 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
, std::adopt_lock
);
3278 std::lock_guard
<std::recursive_mutex
> l2(dest
->cache
->lock
, std::adopt_lock
);
3280 int destbits
= dest
->cnode
.bits
;
3282 bool is_pg
= dest
->cid
.is_pg(&destpg
);
3285 auto p
= onode_map
.onode_map
.begin();
3286 while (p
!= onode_map
.onode_map
.end()) {
3287 if (!p
->second
->oid
.match(destbits
, destpg
.pgid
.ps())) {
3288 // onode does not belong to this child
3291 OnodeRef o
= p
->second
;
3292 ldout(store
->cct
, 20) << __func__
<< " moving " << o
<< " " << o
->oid
3295 cache
->_rm_onode(p
->second
);
3296 p
= onode_map
.onode_map
.erase(p
);
3299 dest
->cache
->_add_onode(o
, 1);
3300 dest
->onode_map
.onode_map
[o
->oid
] = o
;
3301 dest
->onode_map
.cache
= dest
->cache
;
3303 // move over shared blobs and buffers. cover shared blobs from
3304 // both extent map and spanning blob map (the full extent map
3305 // may not be faulted in)
3306 vector
<SharedBlob
*> sbvec
;
3307 for (auto& e
: o
->extent_map
.extent_map
) {
3308 sbvec
.push_back(e
.blob
->shared_blob
.get());
3310 for (auto& b
: o
->extent_map
.spanning_blob_map
) {
3311 sbvec
.push_back(b
.second
->shared_blob
.get());
3313 for (auto sb
: sbvec
) {
3314 if (sb
->coll
== dest
) {
3315 ldout(store
->cct
, 20) << __func__
<< " already moved " << *sb
3319 ldout(store
->cct
, 20) << __func__
<< " moving " << *sb
<< dendl
;
3320 if (sb
->get_sbid()) {
3321 ldout(store
->cct
, 20) << __func__
3322 << " moving registration " << *sb
<< dendl
;
3323 shared_blob_set
.remove(sb
);
3324 dest
->shared_blob_set
.add(dest
, sb
);
3327 if (dest
->cache
!= cache
) {
3328 for (auto& i
: sb
->bc
.buffer_map
) {
3329 if (!i
.second
->is_writing()) {
3330 ldout(store
->cct
, 20) << __func__
<< " moving " << *i
.second
3332 dest
->cache
->_move_buffer(cache
, i
.second
.get());
3341 // =======================================================
3343 void *BlueStore::MempoolThread::entry()
3345 Mutex::Locker
l(lock
);
3347 uint64_t meta_bytes
=
3348 mempool::bluestore_cache_other::allocated_bytes() +
3349 mempool::bluestore_cache_onode::allocated_bytes();
3350 uint64_t onode_num
=
3351 mempool::bluestore_cache_onode::allocated_items();
3353 if (onode_num
< 2) {
3357 float bytes_per_onode
= (float)meta_bytes
/ (float)onode_num
;
3358 size_t num_shards
= store
->cache_shards
.size();
3359 float target_ratio
= store
->cache_meta_ratio
+ store
->cache_data_ratio
;
3360 // A little sloppy but should be close enough
3361 uint64_t shard_target
= target_ratio
* (store
->cache_size
/ num_shards
);
3363 for (auto i
: store
->cache_shards
) {
3364 i
->trim(shard_target
,
3365 store
->cache_meta_ratio
,
3366 store
->cache_data_ratio
,
3370 store
->_update_cache_logger();
3373 wait
+= store
->cct
->_conf
->bluestore_cache_trim_interval
;
3374 cond
.WaitInterval(lock
, wait
);
3380 // =======================================================
3385 #define dout_prefix *_dout << "bluestore.OmapIteratorImpl(" << this << ") "
3387 BlueStore::OmapIteratorImpl::OmapIteratorImpl(
3388 CollectionRef c
, OnodeRef o
, KeyValueDB::Iterator it
)
3389 : c(c
), o(o
), it(it
)
3391 RWLock::RLocker
l(c
->lock
);
3392 if (o
->onode
.has_omap()) {
3393 get_omap_key(o
->onode
.nid
, string(), &head
);
3394 get_omap_tail(o
->onode
.nid
, &tail
);
3395 it
->lower_bound(head
);
3399 int BlueStore::OmapIteratorImpl::seek_to_first()
3401 RWLock::RLocker
l(c
->lock
);
3402 if (o
->onode
.has_omap()) {
3403 it
->lower_bound(head
);
3405 it
= KeyValueDB::Iterator();
3410 int BlueStore::OmapIteratorImpl::upper_bound(const string
& after
)
3412 RWLock::RLocker
l(c
->lock
);
3413 if (o
->onode
.has_omap()) {
3415 get_omap_key(o
->onode
.nid
, after
, &key
);
3416 ldout(c
->store
->cct
,20) << __func__
<< " after " << after
<< " key "
3417 << pretty_binary_string(key
) << dendl
;
3418 it
->upper_bound(key
);
3420 it
= KeyValueDB::Iterator();
3425 int BlueStore::OmapIteratorImpl::lower_bound(const string
& to
)
3427 RWLock::RLocker
l(c
->lock
);
3428 if (o
->onode
.has_omap()) {
3430 get_omap_key(o
->onode
.nid
, to
, &key
);
3431 ldout(c
->store
->cct
,20) << __func__
<< " to " << to
<< " key "
3432 << pretty_binary_string(key
) << dendl
;
3433 it
->lower_bound(key
);
3435 it
= KeyValueDB::Iterator();
3440 bool BlueStore::OmapIteratorImpl::valid()
3442 RWLock::RLocker
l(c
->lock
);
3443 bool r
= o
->onode
.has_omap() && it
&& it
->valid() &&
3444 it
->raw_key().second
<= tail
;
3445 if (it
&& it
->valid()) {
3446 ldout(c
->store
->cct
,20) << __func__
<< " is at "
3447 << pretty_binary_string(it
->raw_key().second
)
3453 int BlueStore::OmapIteratorImpl::next(bool validate
)
3455 RWLock::RLocker
l(c
->lock
);
3456 if (o
->onode
.has_omap()) {
3464 string
BlueStore::OmapIteratorImpl::key()
3466 RWLock::RLocker
l(c
->lock
);
3467 assert(it
->valid());
3468 string db_key
= it
->raw_key().second
;
3470 decode_omap_key(db_key
, &user_key
);
3474 bufferlist
BlueStore::OmapIteratorImpl::value()
3476 RWLock::RLocker
l(c
->lock
);
3477 assert(it
->valid());
3482 // =====================================
3485 #define dout_prefix *_dout << "bluestore(" << path << ") "
3488 static void aio_cb(void *priv
, void *priv2
)
3490 BlueStore
*store
= static_cast<BlueStore
*>(priv
);
3491 BlueStore::AioContext
*c
= static_cast<BlueStore::AioContext
*>(priv2
);
3492 c
->aio_finish(store
);
3495 BlueStore::BlueStore(CephContext
*cct
, const string
& path
)
3496 : ObjectStore(cct
, path
),
3497 throttle_bytes(cct
, "bluestore_throttle_bytes",
3498 cct
->_conf
->bluestore_throttle_bytes
),
3499 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3500 cct
->_conf
->bluestore_throttle_bytes
+
3501 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3502 deferred_finisher(cct
, "defered_finisher", "dfin"),
3503 kv_sync_thread(this),
3504 kv_finalize_thread(this),
3505 mempool_thread(this)
3508 cct
->_conf
->add_observer(this);
3509 set_cache_shards(1);
3512 BlueStore::BlueStore(CephContext
*cct
,
3514 uint64_t _min_alloc_size
)
3515 : ObjectStore(cct
, path
),
3516 throttle_bytes(cct
, "bluestore_throttle_bytes",
3517 cct
->_conf
->bluestore_throttle_bytes
),
3518 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3519 cct
->_conf
->bluestore_throttle_bytes
+
3520 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3521 deferred_finisher(cct
, "defered_finisher", "dfin"),
3522 kv_sync_thread(this),
3523 kv_finalize_thread(this),
3524 min_alloc_size(_min_alloc_size
),
3525 min_alloc_size_order(ctz(_min_alloc_size
)),
3526 mempool_thread(this)
3529 cct
->_conf
->add_observer(this);
3530 set_cache_shards(1);
3533 BlueStore::~BlueStore()
3535 for (auto f
: finishers
) {
3540 cct
->_conf
->remove_observer(this);
3544 assert(bluefs
== NULL
);
3545 assert(fsid_fd
< 0);
3546 assert(path_fd
< 0);
3547 for (auto i
: cache_shards
) {
3550 cache_shards
.clear();
3553 const char **BlueStore::get_tracked_conf_keys() const
3555 static const char* KEYS
[] = {
3556 "bluestore_csum_type",
3557 "bluestore_compression_mode",
3558 "bluestore_compression_algorithm",
3559 "bluestore_compression_min_blob_size",
3560 "bluestore_compression_min_blob_size_ssd",
3561 "bluestore_compression_min_blob_size_hdd",
3562 "bluestore_compression_max_blob_size",
3563 "bluestore_compression_max_blob_size_ssd",
3564 "bluestore_compression_max_blob_size_hdd",
3565 "bluestore_compression_required_ratio",
3566 "bluestore_max_alloc_size",
3567 "bluestore_prefer_deferred_size",
3568 "bluestore_prefer_deferred_size_hdd",
3569 "bluestore_prefer_deferred_size_ssd",
3570 "bluestore_deferred_batch_ops",
3571 "bluestore_deferred_batch_ops_hdd",
3572 "bluestore_deferred_batch_ops_ssd",
3573 "bluestore_throttle_bytes",
3574 "bluestore_throttle_deferred_bytes",
3575 "bluestore_throttle_cost_per_io_hdd",
3576 "bluestore_throttle_cost_per_io_ssd",
3577 "bluestore_throttle_cost_per_io",
3578 "bluestore_max_blob_size",
3579 "bluestore_max_blob_size_ssd",
3580 "bluestore_max_blob_size_hdd",
3586 void BlueStore::handle_conf_change(const struct md_config_t
*conf
,
3587 const std::set
<std::string
> &changed
)
3589 if (changed
.count("bluestore_csum_type")) {
3592 if (changed
.count("bluestore_compression_mode") ||
3593 changed
.count("bluestore_compression_algorithm") ||
3594 changed
.count("bluestore_compression_min_blob_size") ||
3595 changed
.count("bluestore_compression_max_blob_size")) {
3600 if (changed
.count("bluestore_max_blob_size") ||
3601 changed
.count("bluestore_max_blob_size_ssd") ||
3602 changed
.count("bluestore_max_blob_size_hdd")) {
3604 // only after startup
3608 if (changed
.count("bluestore_prefer_deferred_size") ||
3609 changed
.count("bluestore_prefer_deferred_size_hdd") ||
3610 changed
.count("bluestore_prefer_deferred_size_ssd") ||
3611 changed
.count("bluestore_max_alloc_size") ||
3612 changed
.count("bluestore_deferred_batch_ops") ||
3613 changed
.count("bluestore_deferred_batch_ops_hdd") ||
3614 changed
.count("bluestore_deferred_batch_ops_ssd")) {
3616 // only after startup
3620 if (changed
.count("bluestore_throttle_cost_per_io") ||
3621 changed
.count("bluestore_throttle_cost_per_io_hdd") ||
3622 changed
.count("bluestore_throttle_cost_per_io_ssd")) {
3624 _set_throttle_params();
3627 if (changed
.count("bluestore_throttle_bytes")) {
3628 throttle_bytes
.reset_max(conf
->bluestore_throttle_bytes
);
3629 throttle_deferred_bytes
.reset_max(
3630 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3632 if (changed
.count("bluestore_throttle_deferred_bytes")) {
3633 throttle_deferred_bytes
.reset_max(
3634 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3638 void BlueStore::_set_compression()
3640 auto m
= Compressor::get_comp_mode_type(cct
->_conf
->bluestore_compression_mode
);
3644 derr
<< __func__
<< " unrecognized value '"
3645 << cct
->_conf
->bluestore_compression_mode
3646 << "' for bluestore_compression_mode, reverting to 'none'"
3648 comp_mode
= Compressor::COMP_NONE
;
3651 compressor
= nullptr;
3653 if (comp_mode
== Compressor::COMP_NONE
) {
3654 dout(10) << __func__
<< " compression mode set to 'none', "
3655 << "ignore other compression setttings" << dendl
;
3659 if (cct
->_conf
->bluestore_compression_min_blob_size
) {
3660 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size
;
3663 if (bdev
->is_rotational()) {
3664 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_hdd
;
3666 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_ssd
;
3670 if (cct
->_conf
->bluestore_compression_max_blob_size
) {
3671 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size
;
3674 if (bdev
->is_rotational()) {
3675 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_hdd
;
3677 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_ssd
;
3681 auto& alg_name
= cct
->_conf
->bluestore_compression_algorithm
;
3682 if (!alg_name
.empty()) {
3683 compressor
= Compressor::create(cct
, alg_name
);
3685 derr
<< __func__
<< " unable to initialize " << alg_name
.c_str() << " compressor"
3690 dout(10) << __func__
<< " mode " << Compressor::get_comp_mode_name(comp_mode
)
3691 << " alg " << (compressor
? compressor
->get_type_name() : "(none)")
3695 void BlueStore::_set_csum()
3697 csum_type
= Checksummer::CSUM_NONE
;
3698 int t
= Checksummer::get_csum_string_type(cct
->_conf
->bluestore_csum_type
);
3699 if (t
> Checksummer::CSUM_NONE
)
3702 dout(10) << __func__
<< " csum_type "
3703 << Checksummer::get_csum_type_string(csum_type
)
3707 void BlueStore::_set_throttle_params()
3709 if (cct
->_conf
->bluestore_throttle_cost_per_io
) {
3710 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io
;
3713 if (bdev
->is_rotational()) {
3714 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_hdd
;
3716 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_ssd
;
3720 dout(10) << __func__
<< " throttle_cost_per_io " << throttle_cost_per_io
3723 void BlueStore::_set_blob_size()
3725 if (cct
->_conf
->bluestore_max_blob_size
) {
3726 max_blob_size
= cct
->_conf
->bluestore_max_blob_size
;
3729 if (bdev
->is_rotational()) {
3730 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_hdd
;
3732 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_ssd
;
3735 dout(10) << __func__
<< " max_blob_size 0x" << std::hex
<< max_blob_size
3736 << std::dec
<< dendl
;
3739 int BlueStore::_set_cache_sizes()
3742 if (cct
->_conf
->bluestore_cache_size
) {
3743 cache_size
= cct
->_conf
->bluestore_cache_size
;
3745 // choose global cache size based on backend type
3746 if (bdev
->is_rotational()) {
3747 cache_size
= cct
->_conf
->bluestore_cache_size_hdd
;
3749 cache_size
= cct
->_conf
->bluestore_cache_size_ssd
;
3752 cache_meta_ratio
= cct
->_conf
->bluestore_cache_meta_ratio
;
3753 cache_kv_ratio
= cct
->_conf
->bluestore_cache_kv_ratio
;
3755 double cache_kv_max
= cct
->_conf
->bluestore_cache_kv_max
;
3756 double cache_kv_max_ratio
= 0;
3758 // if cache_kv_max is negative, disable it
3759 if (cache_size
> 0 && cache_kv_max
>= 0) {
3760 cache_kv_max_ratio
= (double) cache_kv_max
/ (double) cache_size
;
3761 if (cache_kv_max_ratio
< 1.0 && cache_kv_max_ratio
< cache_kv_ratio
) {
3762 dout(1) << __func__
<< " max " << cache_kv_max_ratio
3763 << " < ratio " << cache_kv_ratio
3765 cache_meta_ratio
= cache_meta_ratio
+ cache_kv_ratio
- cache_kv_max_ratio
;
3766 cache_kv_ratio
= cache_kv_max_ratio
;
3771 (double)1.0 - (double)cache_meta_ratio
- (double)cache_kv_ratio
;
3773 if (cache_meta_ratio
< 0 || cache_meta_ratio
> 1.0) {
3774 derr
<< __func__
<< " bluestore_cache_meta_ratio (" << cache_meta_ratio
3775 << ") must be in range [0,1.0]" << dendl
;
3778 if (cache_kv_ratio
< 0 || cache_kv_ratio
> 1.0) {
3779 derr
<< __func__
<< " bluestore_cache_kv_ratio (" << cache_kv_ratio
3780 << ") must be in range [0,1.0]" << dendl
;
3783 if (cache_meta_ratio
+ cache_kv_ratio
> 1.0) {
3784 derr
<< __func__
<< " bluestore_cache_meta_ratio (" << cache_meta_ratio
3785 << ") + bluestore_cache_kv_ratio (" << cache_kv_ratio
3786 << ") = " << cache_meta_ratio
+ cache_kv_ratio
<< "; must be <= 1.0"
3790 if (cache_data_ratio
< 0) {
3791 // deal with floating point imprecision
3792 cache_data_ratio
= 0;
3794 dout(1) << __func__
<< " cache_size " << cache_size
3795 << " meta " << cache_meta_ratio
3796 << " kv " << cache_kv_ratio
3797 << " data " << cache_data_ratio
3802 int BlueStore::write_meta(const std::string
& key
, const std::string
& value
)
3804 bluestore_bdev_label_t label
;
3805 string p
= path
+ "/block";
3806 int r
= _read_bdev_label(cct
, p
, &label
);
3808 return ObjectStore::write_meta(key
, value
);
3810 label
.meta
[key
] = value
;
3811 r
= _write_bdev_label(cct
, p
, label
);
3813 return ObjectStore::write_meta(key
, value
);
3816 int BlueStore::read_meta(const std::string
& key
, 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::read_meta(key
, value
);
3824 auto i
= label
.meta
.find(key
);
3825 if (i
== label
.meta
.end()) {
3826 return ObjectStore::read_meta(key
, value
);
3832 void BlueStore::_init_logger()
3834 PerfCountersBuilder
b(cct
, "bluestore",
3835 l_bluestore_first
, l_bluestore_last
);
3836 b
.add_time_avg(l_bluestore_kv_flush_lat
, "kv_flush_lat",
3837 "Average kv_thread flush latency",
3838 "fl_l", PerfCountersBuilder::PRIO_INTERESTING
);
3839 b
.add_time_avg(l_bluestore_kv_commit_lat
, "kv_commit_lat",
3840 "Average kv_thread commit latency");
3841 b
.add_time_avg(l_bluestore_kv_lat
, "kv_lat",
3842 "Average kv_thread sync latency",
3843 "k_l", PerfCountersBuilder::PRIO_INTERESTING
);
3844 b
.add_time_avg(l_bluestore_state_prepare_lat
, "state_prepare_lat",
3845 "Average prepare state latency");
3846 b
.add_time_avg(l_bluestore_state_aio_wait_lat
, "state_aio_wait_lat",
3847 "Average aio_wait state latency",
3848 "io_l", PerfCountersBuilder::PRIO_INTERESTING
);
3849 b
.add_time_avg(l_bluestore_state_io_done_lat
, "state_io_done_lat",
3850 "Average io_done state latency");
3851 b
.add_time_avg(l_bluestore_state_kv_queued_lat
, "state_kv_queued_lat",
3852 "Average kv_queued state latency");
3853 b
.add_time_avg(l_bluestore_state_kv_committing_lat
, "state_kv_commiting_lat",
3854 "Average kv_commiting state latency");
3855 b
.add_time_avg(l_bluestore_state_kv_done_lat
, "state_kv_done_lat",
3856 "Average kv_done state latency");
3857 b
.add_time_avg(l_bluestore_state_deferred_queued_lat
, "state_deferred_queued_lat",
3858 "Average deferred_queued state latency");
3859 b
.add_time_avg(l_bluestore_state_deferred_aio_wait_lat
, "state_deferred_aio_wait_lat",
3860 "Average aio_wait state latency");
3861 b
.add_time_avg(l_bluestore_state_deferred_cleanup_lat
, "state_deferred_cleanup_lat",
3862 "Average cleanup state latency");
3863 b
.add_time_avg(l_bluestore_state_finishing_lat
, "state_finishing_lat",
3864 "Average finishing state latency");
3865 b
.add_time_avg(l_bluestore_state_done_lat
, "state_done_lat",
3866 "Average done state latency");
3867 b
.add_time_avg(l_bluestore_throttle_lat
, "throttle_lat",
3868 "Average submit throttle latency",
3869 "th_l", PerfCountersBuilder::PRIO_CRITICAL
);
3870 b
.add_time_avg(l_bluestore_submit_lat
, "submit_lat",
3871 "Average submit latency",
3872 "s_l", PerfCountersBuilder::PRIO_CRITICAL
);
3873 b
.add_time_avg(l_bluestore_commit_lat
, "commit_lat",
3874 "Average commit latency",
3875 "c_l", PerfCountersBuilder::PRIO_CRITICAL
);
3876 b
.add_time_avg(l_bluestore_read_lat
, "read_lat",
3877 "Average read latency",
3878 "r_l", PerfCountersBuilder::PRIO_CRITICAL
);
3879 b
.add_time_avg(l_bluestore_read_onode_meta_lat
, "read_onode_meta_lat",
3880 "Average read onode metadata latency");
3881 b
.add_time_avg(l_bluestore_read_wait_aio_lat
, "read_wait_aio_lat",
3882 "Average read latency");
3883 b
.add_time_avg(l_bluestore_compress_lat
, "compress_lat",
3884 "Average compress latency");
3885 b
.add_time_avg(l_bluestore_decompress_lat
, "decompress_lat",
3886 "Average decompress latency");
3887 b
.add_time_avg(l_bluestore_csum_lat
, "csum_lat",
3888 "Average checksum latency");
3889 b
.add_u64_counter(l_bluestore_compress_success_count
, "compress_success_count",
3890 "Sum for beneficial compress ops");
3891 b
.add_u64_counter(l_bluestore_compress_rejected_count
, "compress_rejected_count",
3892 "Sum for compress ops rejected due to low net gain of space");
3893 b
.add_u64_counter(l_bluestore_write_pad_bytes
, "write_pad_bytes",
3894 "Sum for write-op padded bytes");
3895 b
.add_u64_counter(l_bluestore_deferred_write_ops
, "deferred_write_ops",
3896 "Sum for deferred write op");
3897 b
.add_u64_counter(l_bluestore_deferred_write_bytes
, "deferred_write_bytes",
3898 "Sum for deferred write bytes", "def");
3899 b
.add_u64_counter(l_bluestore_write_penalty_read_ops
, "write_penalty_read_ops",
3900 "Sum for write penalty read ops");
3901 b
.add_u64(l_bluestore_allocated
, "bluestore_allocated",
3902 "Sum for allocated bytes");
3903 b
.add_u64(l_bluestore_stored
, "bluestore_stored",
3904 "Sum for stored bytes");
3905 b
.add_u64(l_bluestore_compressed
, "bluestore_compressed",
3906 "Sum for stored compressed bytes");
3907 b
.add_u64(l_bluestore_compressed_allocated
, "bluestore_compressed_allocated",
3908 "Sum for bytes allocated for compressed data");
3909 b
.add_u64(l_bluestore_compressed_original
, "bluestore_compressed_original",
3910 "Sum for original bytes that were compressed");
3912 b
.add_u64(l_bluestore_onodes
, "bluestore_onodes",
3913 "Number of onodes in cache");
3914 b
.add_u64_counter(l_bluestore_onode_hits
, "bluestore_onode_hits",
3915 "Sum for onode-lookups hit in the cache");
3916 b
.add_u64_counter(l_bluestore_onode_misses
, "bluestore_onode_misses",
3917 "Sum for onode-lookups missed in the cache");
3918 b
.add_u64_counter(l_bluestore_onode_shard_hits
, "bluestore_onode_shard_hits",
3919 "Sum for onode-shard lookups hit in the cache");
3920 b
.add_u64_counter(l_bluestore_onode_shard_misses
,
3921 "bluestore_onode_shard_misses",
3922 "Sum for onode-shard lookups missed in the cache");
3923 b
.add_u64(l_bluestore_extents
, "bluestore_extents",
3924 "Number of extents in cache");
3925 b
.add_u64(l_bluestore_blobs
, "bluestore_blobs",
3926 "Number of blobs in cache");
3927 b
.add_u64(l_bluestore_buffers
, "bluestore_buffers",
3928 "Number of buffers in cache");
3929 b
.add_u64(l_bluestore_buffer_bytes
, "bluestore_buffer_bytes",
3930 "Number of buffer bytes in cache");
3931 b
.add_u64(l_bluestore_buffer_hit_bytes
, "bluestore_buffer_hit_bytes",
3932 "Sum for bytes of read hit in the cache");
3933 b
.add_u64(l_bluestore_buffer_miss_bytes
, "bluestore_buffer_miss_bytes",
3934 "Sum for bytes of read missed in the cache");
3936 b
.add_u64_counter(l_bluestore_write_big
, "bluestore_write_big",
3937 "Large aligned writes into fresh blobs");
3938 b
.add_u64_counter(l_bluestore_write_big_bytes
, "bluestore_write_big_bytes",
3939 "Large aligned writes into fresh blobs (bytes)");
3940 b
.add_u64_counter(l_bluestore_write_big_blobs
, "bluestore_write_big_blobs",
3941 "Large aligned writes into fresh blobs (blobs)");
3942 b
.add_u64_counter(l_bluestore_write_small
, "bluestore_write_small",
3943 "Small writes into existing or sparse small blobs");
3944 b
.add_u64_counter(l_bluestore_write_small_bytes
, "bluestore_write_small_bytes",
3945 "Small writes into existing or sparse small blobs (bytes)");
3946 b
.add_u64_counter(l_bluestore_write_small_unused
,
3947 "bluestore_write_small_unused",
3948 "Small writes into unused portion of existing blob");
3949 b
.add_u64_counter(l_bluestore_write_small_deferred
,
3950 "bluestore_write_small_deferred",
3951 "Small overwrites using deferred");
3952 b
.add_u64_counter(l_bluestore_write_small_pre_read
,
3953 "bluestore_write_small_pre_read",
3954 "Small writes that required we read some data (possibly "
3955 "cached) to fill out the block");
3956 b
.add_u64_counter(l_bluestore_write_small_new
, "bluestore_write_small_new",
3957 "Small write into new (sparse) blob");
3959 b
.add_u64_counter(l_bluestore_txc
, "bluestore_txc", "Transactions committed");
3960 b
.add_u64_counter(l_bluestore_onode_reshard
, "bluestore_onode_reshard",
3961 "Onode extent map reshard events");
3962 b
.add_u64_counter(l_bluestore_blob_split
, "bluestore_blob_split",
3963 "Sum for blob splitting due to resharding");
3964 b
.add_u64_counter(l_bluestore_extent_compress
, "bluestore_extent_compress",
3965 "Sum for extents that have been removed due to compression");
3966 b
.add_u64_counter(l_bluestore_gc_merged
, "bluestore_gc_merged",
3967 "Sum for extents that have been merged due to garbage "
3969 logger
= b
.create_perf_counters();
3970 cct
->get_perfcounters_collection()->add(logger
);
3973 int BlueStore::_reload_logger()
3975 struct store_statfs_t store_statfs
;
3977 int r
= statfs(&store_statfs
);
3979 logger
->set(l_bluestore_allocated
, store_statfs
.allocated
);
3980 logger
->set(l_bluestore_stored
, store_statfs
.stored
);
3981 logger
->set(l_bluestore_compressed
, store_statfs
.compressed
);
3982 logger
->set(l_bluestore_compressed_allocated
, store_statfs
.compressed_allocated
);
3983 logger
->set(l_bluestore_compressed_original
, store_statfs
.compressed_original
);
3988 void BlueStore::_shutdown_logger()
3990 cct
->get_perfcounters_collection()->remove(logger
);
3994 int BlueStore::get_block_device_fsid(CephContext
* cct
, const string
& path
,
3997 bluestore_bdev_label_t label
;
3998 int r
= _read_bdev_label(cct
, path
, &label
);
4001 *fsid
= label
.osd_uuid
;
4005 int BlueStore::_open_path()
4007 assert(path_fd
< 0);
4008 path_fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_DIRECTORY
));
4011 derr
<< __func__
<< " unable to open " << path
<< ": " << cpp_strerror(r
)
4018 void BlueStore::_close_path()
4020 VOID_TEMP_FAILURE_RETRY(::close(path_fd
));
4024 int BlueStore::_write_bdev_label(CephContext
*cct
,
4025 string path
, bluestore_bdev_label_t label
)
4027 dout(10) << __func__
<< " path " << path
<< " label " << label
<< dendl
;
4029 ::encode(label
, bl
);
4030 uint32_t crc
= bl
.crc32c(-1);
4032 assert(bl
.length() <= BDEV_LABEL_BLOCK_SIZE
);
4033 bufferptr
z(BDEV_LABEL_BLOCK_SIZE
- bl
.length());
4035 bl
.append(std::move(z
));
4037 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_WRONLY
));
4040 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
4044 int r
= bl
.write_fd(fd
);
4046 derr
<< __func__
<< " failed to write to " << path
4047 << ": " << cpp_strerror(r
) << dendl
;
4051 derr
<< __func__
<< " failed to fsync " << path
4052 << ": " << cpp_strerror(r
) << dendl
;
4054 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4058 int BlueStore::_read_bdev_label(CephContext
* cct
, string path
,
4059 bluestore_bdev_label_t
*label
)
4061 dout(10) << __func__
<< dendl
;
4062 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_RDONLY
));
4065 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
4070 int r
= bl
.read_fd(fd
, BDEV_LABEL_BLOCK_SIZE
);
4071 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4073 derr
<< __func__
<< " failed to read from " << path
4074 << ": " << cpp_strerror(r
) << dendl
;
4078 uint32_t crc
, expected_crc
;
4079 bufferlist::iterator p
= bl
.begin();
4081 ::decode(*label
, p
);
4083 t
.substr_of(bl
, 0, p
.get_off());
4085 ::decode(expected_crc
, p
);
4087 catch (buffer::error
& e
) {
4088 derr
<< __func__
<< " unable to decode label at offset " << p
.get_off()
4093 if (crc
!= expected_crc
) {
4094 derr
<< __func__
<< " bad crc on label, expected " << expected_crc
4095 << " != actual " << crc
<< dendl
;
4098 dout(10) << __func__
<< " got " << *label
<< dendl
;
4102 int BlueStore::_check_or_set_bdev_label(
4103 string path
, uint64_t size
, string desc
, bool create
)
4105 bluestore_bdev_label_t label
;
4107 label
.osd_uuid
= fsid
;
4109 label
.btime
= ceph_clock_now();
4110 label
.description
= desc
;
4111 int r
= _write_bdev_label(cct
, path
, label
);
4115 int r
= _read_bdev_label(cct
, path
, &label
);
4118 if (cct
->_conf
->bluestore_debug_permit_any_bdev_label
) {
4119 dout(20) << __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4120 << " and fsid " << fsid
<< " check bypassed" << dendl
;
4122 else if (label
.osd_uuid
!= fsid
) {
4123 derr
<< __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4124 << " does not match our fsid " << fsid
<< dendl
;
4131 void BlueStore::_set_alloc_sizes(void)
4133 max_alloc_size
= cct
->_conf
->bluestore_max_alloc_size
;
4135 if (cct
->_conf
->bluestore_prefer_deferred_size
) {
4136 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size
;
4139 if (bdev
->is_rotational()) {
4140 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_hdd
;
4142 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_ssd
;
4146 if (cct
->_conf
->bluestore_deferred_batch_ops
) {
4147 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops
;
4150 if (bdev
->is_rotational()) {
4151 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_hdd
;
4153 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_ssd
;
4157 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
4158 << std::dec
<< " order " << min_alloc_size_order
4159 << " max_alloc_size 0x" << std::hex
<< max_alloc_size
4160 << " prefer_deferred_size 0x" << prefer_deferred_size
4162 << " deferred_batch_ops " << deferred_batch_ops
4166 int BlueStore::_open_bdev(bool create
)
4168 assert(bdev
== NULL
);
4169 string p
= path
+ "/block";
4170 bdev
= BlockDevice::create(cct
, p
, aio_cb
, static_cast<void*>(this));
4171 int r
= bdev
->open(p
);
4175 if (bdev
->supported_bdev_label()) {
4176 r
= _check_or_set_bdev_label(p
, bdev
->get_size(), "main", create
);
4181 // initialize global block parameters
4182 block_size
= bdev
->get_block_size();
4183 block_mask
= ~(block_size
- 1);
4184 block_size_order
= ctz(block_size
);
4185 assert(block_size
== 1u << block_size_order
);
4186 // and set cache_size based on device type
4187 r
= _set_cache_sizes();
4201 void BlueStore::_close_bdev()
4209 int BlueStore::_open_fm(bool create
)
4212 fm
= FreelistManager::create(cct
, freelist_type
, db
, PREFIX_ALLOC
);
4215 // initialize freespace
4216 dout(20) << __func__
<< " initializing freespace" << dendl
;
4217 KeyValueDB::Transaction t
= db
->get_transaction();
4220 bl
.append(freelist_type
);
4221 t
->set(PREFIX_SUPER
, "freelist_type", bl
);
4223 fm
->create(bdev
->get_size(), min_alloc_size
, t
);
4225 // allocate superblock reserved space. note that we do not mark
4226 // bluefs space as allocated in the freelist; we instead rely on
4228 uint64_t reserved
= ROUND_UP_TO(MAX(SUPER_RESERVED
, min_alloc_size
),
4230 fm
->allocate(0, reserved
, t
);
4232 if (cct
->_conf
->bluestore_bluefs
) {
4233 assert(bluefs_extents
.num_intervals() == 1);
4234 interval_set
<uint64_t>::iterator p
= bluefs_extents
.begin();
4235 reserved
= ROUND_UP_TO(p
.get_start() + p
.get_len(), min_alloc_size
);
4236 dout(20) << __func__
<< " reserved 0x" << std::hex
<< reserved
<< std::dec
4237 << " for bluefs" << dendl
;
4239 ::encode(bluefs_extents
, bl
);
4240 t
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
4241 dout(20) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
4242 << std::dec
<< dendl
;
4245 if (cct
->_conf
->bluestore_debug_prefill
> 0) {
4246 uint64_t end
= bdev
->get_size() - reserved
;
4247 dout(1) << __func__
<< " pre-fragmenting freespace, using "
4248 << cct
->_conf
->bluestore_debug_prefill
<< " with max free extent "
4249 << cct
->_conf
->bluestore_debug_prefragment_max
<< dendl
;
4250 uint64_t start
= P2ROUNDUP(reserved
, min_alloc_size
);
4251 uint64_t max_b
= cct
->_conf
->bluestore_debug_prefragment_max
/ min_alloc_size
;
4252 float r
= cct
->_conf
->bluestore_debug_prefill
;
4256 while (!stop
&& start
< end
) {
4257 uint64_t l
= (rand() % max_b
+ 1) * min_alloc_size
;
4258 if (start
+ l
> end
) {
4260 l
= P2ALIGN(l
, min_alloc_size
);
4262 assert(start
+ l
<= end
);
4264 uint64_t u
= 1 + (uint64_t)(r
* (double)l
);
4265 u
= P2ROUNDUP(u
, min_alloc_size
);
4266 if (start
+ l
+ u
> end
) {
4267 u
= end
- (start
+ l
);
4268 // trim to align so we don't overflow again
4269 u
= P2ALIGN(u
, min_alloc_size
);
4272 assert(start
+ l
+ u
<= end
);
4274 dout(20) << " free 0x" << std::hex
<< start
<< "~" << l
4275 << " use 0x" << u
<< std::dec
<< dendl
;
4278 // break if u has been trimmed to nothing
4282 fm
->allocate(start
+ l
, u
, t
);
4286 db
->submit_transaction_sync(t
);
4289 int r
= fm
->init(bdev
->get_size());
4291 derr
<< __func__
<< " freelist init failed: " << cpp_strerror(r
) << dendl
;
4299 void BlueStore::_close_fm()
4301 dout(10) << __func__
<< dendl
;
4308 int BlueStore::_open_alloc()
4310 assert(alloc
== NULL
);
4311 assert(bdev
->get_size());
4312 alloc
= Allocator::create(cct
, cct
->_conf
->bluestore_allocator
,
4316 lderr(cct
) << __func__
<< " Allocator::unknown alloc type "
4317 << cct
->_conf
->bluestore_allocator
4322 uint64_t num
= 0, bytes
= 0;
4324 dout(1) << __func__
<< " opening allocation metadata" << dendl
;
4325 // initialize from freelist
4326 fm
->enumerate_reset();
4327 uint64_t offset
, length
;
4328 while (fm
->enumerate_next(&offset
, &length
)) {
4329 alloc
->init_add_free(offset
, length
);
4333 fm
->enumerate_reset();
4334 dout(1) << __func__
<< " loaded " << pretty_si_t(bytes
)
4335 << " in " << num
<< " extents"
4338 // also mark bluefs space as allocated
4339 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
4340 alloc
->init_rm_free(e
.get_start(), e
.get_len());
4342 dout(10) << __func__
<< " marked bluefs_extents 0x" << std::hex
4343 << bluefs_extents
<< std::dec
<< " as allocated" << dendl
;
4348 void BlueStore::_close_alloc()
4356 int BlueStore::_open_fsid(bool create
)
4358 assert(fsid_fd
< 0);
4362 fsid_fd
= ::openat(path_fd
, "fsid", flags
, 0644);
4365 derr
<< __func__
<< " " << cpp_strerror(err
) << dendl
;
4371 int BlueStore::_read_fsid(uuid_d
*uuid
)
4374 memset(fsid_str
, 0, sizeof(fsid_str
));
4375 int ret
= safe_read(fsid_fd
, fsid_str
, sizeof(fsid_str
));
4377 derr
<< __func__
<< " failed: " << cpp_strerror(ret
) << dendl
;
4384 if (!uuid
->parse(fsid_str
)) {
4385 derr
<< __func__
<< " unparsable uuid " << fsid_str
<< dendl
;
4391 int BlueStore::_write_fsid()
4393 int r
= ::ftruncate(fsid_fd
, 0);
4396 derr
<< __func__
<< " fsid truncate failed: " << cpp_strerror(r
) << dendl
;
4399 string str
= stringify(fsid
) + "\n";
4400 r
= safe_write(fsid_fd
, str
.c_str(), str
.length());
4402 derr
<< __func__
<< " fsid write failed: " << cpp_strerror(r
) << dendl
;
4405 r
= ::fsync(fsid_fd
);
4408 derr
<< __func__
<< " fsid fsync failed: " << cpp_strerror(r
) << dendl
;
4414 void BlueStore::_close_fsid()
4416 VOID_TEMP_FAILURE_RETRY(::close(fsid_fd
));
4420 int BlueStore::_lock_fsid()
4423 memset(&l
, 0, sizeof(l
));
4425 l
.l_whence
= SEEK_SET
;
4426 int r
= ::fcntl(fsid_fd
, F_SETLK
, &l
);
4429 derr
<< __func__
<< " failed to lock " << path
<< "/fsid"
4430 << " (is another ceph-osd still running?)"
4431 << cpp_strerror(err
) << dendl
;
4437 bool BlueStore::is_rotational()
4440 return bdev
->is_rotational();
4443 bool rotational
= true;
4444 int r
= _open_path();
4447 r
= _open_fsid(false);
4450 r
= _read_fsid(&fsid
);
4456 r
= _open_bdev(false);
4459 rotational
= bdev
->is_rotational();
4469 bool BlueStore::is_journal_rotational()
4472 dout(5) << __func__
<< " bluefs disabled, default to store media type"
4474 return is_rotational();
4476 dout(10) << __func__
<< " " << (int)bluefs
->wal_is_rotational() << dendl
;
4477 return bluefs
->wal_is_rotational();
4480 bool BlueStore::test_mount_in_use()
4482 // most error conditions mean the mount is not in use (e.g., because
4483 // it doesn't exist). only if we fail to lock do we conclude it is
4486 int r
= _open_path();
4489 r
= _open_fsid(false);
4494 ret
= true; // if we can't lock, it is in use
4501 int BlueStore::_open_db(bool create
)
4505 string fn
= path
+ "/db";
4508 ceph::shared_ptr
<Int64ArrayMergeOperator
> merge_op(new Int64ArrayMergeOperator
);
4512 kv_backend
= cct
->_conf
->bluestore_kvbackend
;
4514 r
= read_meta("kv_backend", &kv_backend
);
4516 derr
<< __func__
<< " unable to read 'kv_backend' meta" << dendl
;
4520 dout(10) << __func__
<< " kv_backend = " << kv_backend
<< dendl
;
4524 do_bluefs
= cct
->_conf
->bluestore_bluefs
;
4527 r
= read_meta("bluefs", &s
);
4529 derr
<< __func__
<< " unable to read 'bluefs' meta" << dendl
;
4534 } else if (s
== "0") {
4537 derr
<< __func__
<< " bluefs = " << s
<< " : not 0 or 1, aborting"
4542 dout(10) << __func__
<< " do_bluefs = " << do_bluefs
<< dendl
;
4544 rocksdb::Env
*env
= NULL
;
4546 dout(10) << __func__
<< " initializing bluefs" << dendl
;
4547 if (kv_backend
!= "rocksdb") {
4548 derr
<< " backend must be rocksdb to use bluefs" << dendl
;
4551 bluefs
= new BlueFS(cct
);
4556 if (read_meta("path_block.db", &bfn
) < 0) {
4557 bfn
= path
+ "/block.db";
4559 if (::stat(bfn
.c_str(), &st
) == 0) {
4560 r
= bluefs
->add_block_device(BlueFS::BDEV_DB
, bfn
);
4562 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4563 << cpp_strerror(r
) << dendl
;
4567 if (bluefs
->bdev_support_label(BlueFS::BDEV_DB
)) {
4568 r
= _check_or_set_bdev_label(
4570 bluefs
->get_block_device_size(BlueFS::BDEV_DB
),
4571 "bluefs db", create
);
4574 << " check block device(" << bfn
<< ") label returned: "
4575 << cpp_strerror(r
) << dendl
;
4580 bluefs
->add_block_extent(
4583 bluefs
->get_block_device_size(BlueFS::BDEV_DB
) - SUPER_RESERVED
);
4585 bluefs_shared_bdev
= BlueFS::BDEV_SLOW
;
4586 bluefs_single_shared_device
= false;
4587 } else if (::lstat(bfn
.c_str(), &st
) == -1) {
4588 bluefs_shared_bdev
= BlueFS::BDEV_DB
;
4590 //symlink exist is bug
4591 derr
<< __func__
<< " " << bfn
<< " link target doesn't exist" << dendl
;
4597 if (read_meta("path_block", &bfn
) < 0) {
4598 bfn
= path
+ "/block";
4600 r
= bluefs
->add_block_device(bluefs_shared_bdev
, bfn
);
4602 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4603 << cpp_strerror(r
) << dendl
;
4607 // note: we always leave the first SUPER_RESERVED (8k) of the device unused
4609 bdev
->get_size() * (cct
->_conf
->bluestore_bluefs_min_ratio
+
4610 cct
->_conf
->bluestore_bluefs_gift_ratio
);
4611 initial
= MAX(initial
, cct
->_conf
->bluestore_bluefs_min
);
4612 if (cct
->_conf
->bluefs_alloc_size
% min_alloc_size
) {
4613 derr
<< __func__
<< " bluefs_alloc_size 0x" << std::hex
4614 << cct
->_conf
->bluefs_alloc_size
<< " is not a multiple of "
4615 << "min_alloc_size 0x" << min_alloc_size
<< std::dec
<< dendl
;
4619 // align to bluefs's alloc_size
4620 initial
= P2ROUNDUP(initial
, cct
->_conf
->bluefs_alloc_size
);
4621 // put bluefs in the middle of the device in case it is an HDD
4622 uint64_t start
= P2ALIGN((bdev
->get_size() - initial
) / 2,
4623 cct
->_conf
->bluefs_alloc_size
);
4624 bluefs
->add_block_extent(bluefs_shared_bdev
, start
, initial
);
4625 bluefs_extents
.insert(start
, initial
);
4628 if (read_meta("path_block.wal", &bfn
) < 0) {
4629 bfn
= path
+ "/block.wal";
4631 if (::stat(bfn
.c_str(), &st
) == 0) {
4632 r
= bluefs
->add_block_device(BlueFS::BDEV_WAL
, bfn
);
4634 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4635 << cpp_strerror(r
) << dendl
;
4639 if (bluefs
->bdev_support_label(BlueFS::BDEV_WAL
)) {
4640 r
= _check_or_set_bdev_label(
4642 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
),
4643 "bluefs wal", create
);
4645 derr
<< __func__
<< " check block device(" << bfn
4646 << ") label returned: " << cpp_strerror(r
) << dendl
;
4652 bluefs
->add_block_extent(
4653 BlueFS::BDEV_WAL
, BDEV_LABEL_BLOCK_SIZE
,
4654 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
) -
4655 BDEV_LABEL_BLOCK_SIZE
);
4657 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "true");
4658 bluefs_single_shared_device
= false;
4659 } else if (::lstat(bfn
.c_str(), &st
) == -1) {
4660 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "false");
4662 //symlink exist is bug
4663 derr
<< __func__
<< " " << bfn
<< " link target doesn't exist" << dendl
;
4671 r
= bluefs
->mount();
4673 derr
<< __func__
<< " failed bluefs mount: " << cpp_strerror(r
) << dendl
;
4676 if (cct
->_conf
->bluestore_bluefs_env_mirror
) {
4677 rocksdb::Env
*a
= new BlueRocksEnv(bluefs
);
4678 rocksdb::Env
*b
= rocksdb::Env::Default();
4680 string cmd
= "rm -rf " + path
+ "/db " +
4681 path
+ "/db.slow " +
4683 int r
= system(cmd
.c_str());
4686 env
= new rocksdb::EnvMirror(b
, a
, false, true);
4688 env
= new BlueRocksEnv(bluefs
);
4690 // simplify the dir names, too, as "seen" by rocksdb
4694 if (bluefs_shared_bdev
== BlueFS::BDEV_SLOW
) {
4695 // we have both block.db and block; tell rocksdb!
4696 // note: the second (last) size value doesn't really matter
4697 ostringstream db_paths
;
4698 uint64_t db_size
= bluefs
->get_block_device_size(BlueFS::BDEV_DB
);
4699 uint64_t slow_size
= bluefs
->get_block_device_size(BlueFS::BDEV_SLOW
);
4700 db_paths
<< fn
<< ","
4701 << (uint64_t)(db_size
* 95 / 100) << " "
4702 << fn
+ ".slow" << ","
4703 << (uint64_t)(slow_size
* 95 / 100);
4704 cct
->_conf
->set_val("rocksdb_db_paths", db_paths
.str(), false);
4705 dout(10) << __func__
<< " set rocksdb_db_paths to "
4706 << cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths") << dendl
;
4711 if (cct
->_conf
->rocksdb_separate_wal_dir
)
4712 env
->CreateDir(fn
+ ".wal");
4713 if (cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths").length())
4714 env
->CreateDir(fn
+ ".slow");
4716 } else if (create
) {
4717 int r
= ::mkdir(fn
.c_str(), 0755);
4720 if (r
< 0 && r
!= -EEXIST
) {
4721 derr
<< __func__
<< " failed to create " << fn
<< ": " << cpp_strerror(r
)
4727 if (cct
->_conf
->rocksdb_separate_wal_dir
) {
4728 string walfn
= path
+ "/db.wal";
4729 r
= ::mkdir(walfn
.c_str(), 0755);
4732 if (r
< 0 && r
!= -EEXIST
) {
4733 derr
<< __func__
<< " failed to create " << walfn
4734 << ": " << cpp_strerror(r
)
4741 db
= KeyValueDB::create(cct
,
4744 static_cast<void*>(env
));
4746 derr
<< __func__
<< " error creating db" << dendl
;
4752 // delete env manually here since we can't depend on db to do this
4759 FreelistManager::setup_merge_operators(db
);
4760 db
->set_merge_operator(PREFIX_STAT
, merge_op
);
4762 db
->set_cache_size(cache_size
* cache_kv_ratio
);
4764 if (kv_backend
== "rocksdb")
4765 options
= cct
->_conf
->bluestore_rocksdb_options
;
4768 r
= db
->create_and_open(err
);
4772 derr
<< __func__
<< " erroring opening db: " << err
.str() << dendl
;
4782 dout(1) << __func__
<< " opened " << kv_backend
4783 << " path " << fn
<< " options " << options
<< dendl
;
4793 void BlueStore::_close_db()
4805 int BlueStore::_reconcile_bluefs_freespace()
4807 dout(10) << __func__
<< dendl
;
4808 interval_set
<uint64_t> bset
;
4809 int r
= bluefs
->get_block_extents(bluefs_shared_bdev
, &bset
);
4811 if (bset
== bluefs_extents
) {
4812 dout(10) << __func__
<< " we agree bluefs has 0x" << std::hex
<< bset
4813 << std::dec
<< dendl
;
4816 dout(10) << __func__
<< " bluefs says 0x" << std::hex
<< bset
<< std::dec
4818 dout(10) << __func__
<< " super says 0x" << std::hex
<< bluefs_extents
4819 << std::dec
<< dendl
;
4821 interval_set
<uint64_t> overlap
;
4822 overlap
.intersection_of(bset
, bluefs_extents
);
4824 bset
.subtract(overlap
);
4825 if (!bset
.empty()) {
4826 derr
<< __func__
<< " bluefs extra 0x" << std::hex
<< bset
<< std::dec
4831 interval_set
<uint64_t> super_extra
;
4832 super_extra
= bluefs_extents
;
4833 super_extra
.subtract(overlap
);
4834 if (!super_extra
.empty()) {
4835 // This is normal: it can happen if we commit to give extents to
4836 // bluefs and we crash before bluefs commits that it owns them.
4837 dout(10) << __func__
<< " super extra " << super_extra
<< dendl
;
4838 for (interval_set
<uint64_t>::iterator p
= super_extra
.begin();
4839 p
!= super_extra
.end();
4841 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.get_start(), p
.get_len());
4848 int BlueStore::_balance_bluefs_freespace(PExtentVector
*extents
)
4853 vector
<pair
<uint64_t,uint64_t>> bluefs_usage
; // <free, total> ...
4854 bluefs
->get_usage(&bluefs_usage
);
4855 assert(bluefs_usage
.size() > bluefs_shared_bdev
);
4857 // fixme: look at primary bdev only for now
4858 uint64_t bluefs_free
= bluefs_usage
[bluefs_shared_bdev
].first
;
4859 uint64_t bluefs_total
= bluefs_usage
[bluefs_shared_bdev
].second
;
4860 float bluefs_free_ratio
= (float)bluefs_free
/ (float)bluefs_total
;
4862 uint64_t my_free
= alloc
->get_free();
4863 uint64_t total
= bdev
->get_size();
4864 float my_free_ratio
= (float)my_free
/ (float)total
;
4866 uint64_t total_free
= bluefs_free
+ my_free
;
4868 float bluefs_ratio
= (float)bluefs_free
/ (float)total_free
;
4870 dout(10) << __func__
4871 << " bluefs " << pretty_si_t(bluefs_free
)
4872 << " free (" << bluefs_free_ratio
4873 << ") bluestore " << pretty_si_t(my_free
)
4874 << " free (" << my_free_ratio
4875 << "), bluefs_ratio " << bluefs_ratio
4879 uint64_t reclaim
= 0;
4880 if (bluefs_ratio
< cct
->_conf
->bluestore_bluefs_min_ratio
) {
4881 gift
= cct
->_conf
->bluestore_bluefs_gift_ratio
* total_free
;
4882 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4883 << " < min_ratio " << cct
->_conf
->bluestore_bluefs_min_ratio
4884 << ", should gift " << pretty_si_t(gift
) << dendl
;
4885 } else if (bluefs_ratio
> cct
->_conf
->bluestore_bluefs_max_ratio
) {
4886 reclaim
= cct
->_conf
->bluestore_bluefs_reclaim_ratio
* total_free
;
4887 if (bluefs_total
- reclaim
< cct
->_conf
->bluestore_bluefs_min
)
4888 reclaim
= bluefs_total
- cct
->_conf
->bluestore_bluefs_min
;
4889 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4890 << " > max_ratio " << cct
->_conf
->bluestore_bluefs_max_ratio
4891 << ", should reclaim " << pretty_si_t(reclaim
) << dendl
;
4894 // don't take over too much of the freespace
4895 uint64_t free_cap
= cct
->_conf
->bluestore_bluefs_max_ratio
* total_free
;
4896 if (bluefs_total
< cct
->_conf
->bluestore_bluefs_min
&&
4897 cct
->_conf
->bluestore_bluefs_min
< free_cap
) {
4898 uint64_t g
= cct
->_conf
->bluestore_bluefs_min
- bluefs_total
;
4899 dout(10) << __func__
<< " bluefs_total " << bluefs_total
4900 << " < min " << cct
->_conf
->bluestore_bluefs_min
4901 << ", should gift " << pretty_si_t(g
) << dendl
;
4906 uint64_t min_free
= cct
->_conf
->get_val
<uint64_t>("bluestore_bluefs_min_free");
4907 if (bluefs_free
< min_free
&&
4908 min_free
< free_cap
) {
4909 uint64_t g
= min_free
- bluefs_free
;
4910 dout(10) << __func__
<< " bluefs_free " << bluefs_total
4911 << " < min " << min_free
4912 << ", should gift " << pretty_si_t(g
) << dendl
;
4919 // round up to alloc size
4920 gift
= P2ROUNDUP(gift
, cct
->_conf
->bluefs_alloc_size
);
4922 // hard cap to fit into 32 bits
4923 gift
= MIN(gift
, 1ull<<31);
4924 dout(10) << __func__
<< " gifting " << gift
4925 << " (" << pretty_si_t(gift
) << ")" << dendl
;
4927 // fixme: just do one allocation to start...
4928 int r
= alloc
->reserve(gift
);
4931 AllocExtentVector exts
;
4932 int64_t alloc_len
= alloc
->allocate(gift
, cct
->_conf
->bluefs_alloc_size
,
4935 if (alloc_len
< (int64_t)gift
) {
4936 derr
<< __func__
<< " allocate failed on 0x" << std::hex
<< gift
4937 << " min_alloc_size 0x" << min_alloc_size
<< std::dec
<< dendl
;
4939 assert(0 == "allocate failed, wtf");
4942 for (auto& p
: exts
) {
4943 bluestore_pextent_t e
= bluestore_pextent_t(p
);
4944 dout(1) << __func__
<< " gifting " << e
<< " to bluefs" << dendl
;
4945 extents
->push_back(e
);
4952 // reclaim from bluefs?
4954 // round up to alloc size
4955 reclaim
= P2ROUNDUP(reclaim
, cct
->_conf
->bluefs_alloc_size
);
4957 // hard cap to fit into 32 bits
4958 reclaim
= MIN(reclaim
, 1ull<<31);
4959 dout(10) << __func__
<< " reclaiming " << reclaim
4960 << " (" << pretty_si_t(reclaim
) << ")" << dendl
;
4962 while (reclaim
> 0) {
4963 // NOTE: this will block and do IO.
4964 AllocExtentVector extents
;
4965 int r
= bluefs
->reclaim_blocks(bluefs_shared_bdev
, reclaim
,
4968 derr
<< __func__
<< " failed to reclaim space from bluefs"
4972 for (auto e
: extents
) {
4973 bluefs_extents
.erase(e
.offset
, e
.length
);
4974 bluefs_extents_reclaiming
.insert(e
.offset
, e
.length
);
4975 reclaim
-= e
.length
;
4985 void BlueStore::_commit_bluefs_freespace(
4986 const PExtentVector
& bluefs_gift_extents
)
4988 dout(10) << __func__
<< dendl
;
4989 for (auto& p
: bluefs_gift_extents
) {
4990 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.offset
, p
.length
);
4994 int BlueStore::_open_collections(int *errors
)
4996 assert(coll_map
.empty());
4997 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_COLL
);
4998 for (it
->upper_bound(string());
5002 if (cid
.parse(it
->key())) {
5006 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
5008 bufferlist bl
= it
->value();
5009 bufferlist::iterator p
= bl
.begin();
5011 ::decode(c
->cnode
, p
);
5012 } catch (buffer::error
& e
) {
5013 derr
<< __func__
<< " failed to decode cnode, key:"
5014 << pretty_binary_string(it
->key()) << dendl
;
5017 dout(20) << __func__
<< " opened " << cid
<< " " << c
<< dendl
;
5020 derr
<< __func__
<< " unrecognized collection " << it
->key() << dendl
;
5028 void BlueStore::_open_statfs()
5031 int r
= db
->get(PREFIX_STAT
, "bluestore_statfs", &bl
);
5033 if (size_t(bl
.length()) >= sizeof(vstatfs
.values
)) {
5034 auto it
= bl
.begin();
5037 dout(10) << __func__
<< " store_statfs is corrupt, using empty" << dendl
;
5041 dout(10) << __func__
<< " store_statfs missed, using empty" << dendl
;
5045 int BlueStore::_setup_block_symlink_or_file(
5051 dout(20) << __func__
<< " name " << name
<< " path " << epath
5052 << " size " << size
<< " create=" << (int)create
<< dendl
;
5057 if (epath
.length()) {
5058 r
= ::symlinkat(epath
.c_str(), path_fd
, name
.c_str());
5061 derr
<< __func__
<< " failed to create " << name
<< " symlink to "
5062 << epath
<< ": " << cpp_strerror(r
) << dendl
;
5066 if (!epath
.compare(0, strlen(SPDK_PREFIX
), SPDK_PREFIX
)) {
5067 int fd
= ::openat(path_fd
, epath
.c_str(), flags
, 0644);
5070 derr
<< __func__
<< " failed to open " << epath
<< " file: "
5071 << cpp_strerror(r
) << dendl
;
5074 string serial_number
= epath
.substr(strlen(SPDK_PREFIX
));
5075 r
= ::write(fd
, serial_number
.c_str(), serial_number
.size());
5076 assert(r
== (int)serial_number
.size());
5077 dout(1) << __func__
<< " created " << name
<< " symlink to "
5079 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5083 int fd
= ::openat(path_fd
, name
.c_str(), flags
, 0644);
5085 // block file is present
5087 int r
= ::fstat(fd
, &st
);
5089 S_ISREG(st
.st_mode
) && // if it is a regular file
5090 st
.st_size
== 0) { // and is 0 bytes
5091 r
= ::ftruncate(fd
, size
);
5094 derr
<< __func__
<< " failed to resize " << name
<< " file to "
5095 << size
<< ": " << cpp_strerror(r
) << dendl
;
5096 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5100 if (cct
->_conf
->bluestore_block_preallocate_file
) {
5101 #ifdef HAVE_POSIX_FALLOCATE
5102 r
= ::posix_fallocate(fd
, 0, size
);
5104 derr
<< __func__
<< " failed to prefallocate " << name
<< " file to "
5105 << size
<< ": " << cpp_strerror(r
) << dendl
;
5106 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5110 char data
[1024*128];
5111 for (uint64_t off
= 0; off
< size
; off
+= sizeof(data
)) {
5112 if (off
+ sizeof(data
) > size
)
5113 r
= ::write(fd
, data
, size
- off
);
5115 r
= ::write(fd
, data
, sizeof(data
));
5118 derr
<< __func__
<< " failed to prefallocate w/ write " << name
<< " file to "
5119 << size
<< ": " << cpp_strerror(r
) << dendl
;
5120 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5126 dout(1) << __func__
<< " resized " << name
<< " file to "
5127 << pretty_si_t(size
) << "B" << dendl
;
5129 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5133 derr
<< __func__
<< " failed to open " << name
<< " file: "
5134 << cpp_strerror(r
) << dendl
;
5142 int BlueStore::mkfs()
5144 dout(1) << __func__
<< " path " << path
<< dendl
;
5150 r
= read_meta("mkfs_done", &done
);
5152 dout(1) << __func__
<< " already created" << dendl
;
5153 if (cct
->_conf
->bluestore_fsck_on_mkfs
) {
5154 r
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5156 derr
<< __func__
<< " fsck found fatal error: " << cpp_strerror(r
)
5161 derr
<< __func__
<< " fsck found " << r
<< " errors" << dendl
;
5165 return r
; // idempotent
5171 r
= read_meta("type", &type
);
5173 if (type
!= "bluestore") {
5174 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5178 r
= write_meta("type", "bluestore");
5184 freelist_type
= "bitmap";
5190 r
= _open_fsid(true);
5196 goto out_close_fsid
;
5198 r
= _read_fsid(&old_fsid
);
5199 if (r
< 0 || old_fsid
.is_zero()) {
5200 if (fsid
.is_zero()) {
5201 fsid
.generate_random();
5202 dout(1) << __func__
<< " generated fsid " << fsid
<< dendl
;
5204 dout(1) << __func__
<< " using provided fsid " << fsid
<< dendl
;
5206 // we'll write it later.
5208 if (!fsid
.is_zero() && fsid
!= old_fsid
) {
5209 derr
<< __func__
<< " on-disk fsid " << old_fsid
5210 << " != provided " << fsid
<< dendl
;
5212 goto out_close_fsid
;
5217 r
= _setup_block_symlink_or_file("block", cct
->_conf
->bluestore_block_path
,
5218 cct
->_conf
->bluestore_block_size
,
5219 cct
->_conf
->bluestore_block_create
);
5221 goto out_close_fsid
;
5222 if (cct
->_conf
->bluestore_bluefs
) {
5223 r
= _setup_block_symlink_or_file("block.wal", cct
->_conf
->bluestore_block_wal_path
,
5224 cct
->_conf
->bluestore_block_wal_size
,
5225 cct
->_conf
->bluestore_block_wal_create
);
5227 goto out_close_fsid
;
5228 r
= _setup_block_symlink_or_file("block.db", cct
->_conf
->bluestore_block_db_path
,
5229 cct
->_conf
->bluestore_block_db_size
,
5230 cct
->_conf
->bluestore_block_db_create
);
5232 goto out_close_fsid
;
5235 r
= _open_bdev(true);
5237 goto out_close_fsid
;
5240 string wal_path
= cct
->_conf
->get_val
<string
>("bluestore_block_wal_path");
5241 if (wal_path
.size()) {
5242 write_meta("path_block.wal", wal_path
);
5244 string db_path
= cct
->_conf
->get_val
<string
>("bluestore_block_db_path");
5245 if (db_path
.size()) {
5246 write_meta("path_block.db", db_path
);
5250 // choose min_alloc_size
5251 if (cct
->_conf
->bluestore_min_alloc_size
) {
5252 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size
;
5255 if (bdev
->is_rotational()) {
5256 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_hdd
;
5258 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_ssd
;
5262 // make sure min_alloc_size is power of 2 aligned.
5263 if (!ISP2(min_alloc_size
)) {
5264 derr
<< __func__
<< " min_alloc_size 0x"
5265 << std::hex
<< min_alloc_size
<< std::dec
5266 << " is not power of 2 aligned!"
5269 goto out_close_bdev
;
5274 goto out_close_bdev
;
5281 KeyValueDB::Transaction t
= db
->get_transaction();
5284 ::encode((uint64_t)0, bl
);
5285 t
->set(PREFIX_SUPER
, "nid_max", bl
);
5286 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
5291 ::encode((uint64_t)min_alloc_size
, bl
);
5292 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
5295 ondisk_format
= latest_ondisk_format
;
5296 _prepare_ondisk_format_super(t
);
5297 db
->submit_transaction_sync(t
);
5301 r
= write_meta("kv_backend", cct
->_conf
->bluestore_kvbackend
);
5305 r
= write_meta("bluefs", stringify(bluefs
? 1 : 0));
5309 if (fsid
!= old_fsid
) {
5312 derr
<< __func__
<< " error writing fsid: " << cpp_strerror(r
) << dendl
;
5329 cct
->_conf
->bluestore_fsck_on_mkfs
) {
5330 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5334 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5340 // indicate success by writing the 'mkfs_done' file
5341 r
= write_meta("mkfs_done", "yes");
5345 derr
<< __func__
<< " failed, " << cpp_strerror(r
) << dendl
;
5347 dout(0) << __func__
<< " success" << dendl
;
5352 void BlueStore::set_cache_shards(unsigned num
)
5354 dout(10) << __func__
<< " " << num
<< dendl
;
5355 size_t old
= cache_shards
.size();
5357 cache_shards
.resize(num
);
5358 for (unsigned i
= old
; i
< num
; ++i
) {
5359 cache_shards
[i
] = Cache::create(cct
, cct
->_conf
->bluestore_cache_type
,
5364 int BlueStore::_mount(bool kv_only
)
5366 dout(1) << __func__
<< " path " << path
<< dendl
;
5372 int r
= read_meta("type", &type
);
5374 derr
<< __func__
<< " failed to load os-type: " << cpp_strerror(r
)
5379 if (type
!= "bluestore") {
5380 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5385 if (cct
->_conf
->bluestore_fsck_on_mount
) {
5386 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mount_deep
);
5390 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5395 int r
= _open_path();
5398 r
= _open_fsid(false);
5402 r
= _read_fsid(&fsid
);
5410 r
= _open_bdev(false);
5414 r
= _open_db(false);
5421 r
= _open_super_meta();
5425 r
= _open_fm(false);
5433 r
= _open_collections();
5437 r
= _reload_logger();
5442 r
= _reconcile_bluefs_freespace();
5449 r
= _deferred_replay();
5453 mempool_thread
.init();
5478 int BlueStore::umount()
5480 assert(_kv_only
|| mounted
);
5481 dout(1) << __func__
<< dendl
;
5484 _osr_unregister_all();
5488 mempool_thread
.shutdown();
5489 dout(20) << __func__
<< " stopping kv thread" << dendl
;
5491 _reap_collections();
5493 dout(20) << __func__
<< " closing" << dendl
;
5503 if (cct
->_conf
->bluestore_fsck_on_umount
) {
5504 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_umount_deep
);
5508 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5515 static void apply(uint64_t off
,
5517 uint64_t granularity
,
5518 BlueStore::mempool_dynamic_bitset
&bitset
,
5520 std::function
<void(uint64_t,
5521 BlueStore::mempool_dynamic_bitset
&)> f
) {
5522 auto end
= ROUND_UP_TO(off
+ len
, granularity
);
5524 uint64_t pos
= off
/ granularity
;
5530 int BlueStore::_fsck_check_extents(
5531 const ghobject_t
& oid
,
5532 const PExtentVector
& extents
,
5534 mempool_dynamic_bitset
&used_blocks
,
5535 store_statfs_t
& expected_statfs
)
5537 dout(30) << __func__
<< " oid " << oid
<< " extents " << extents
<< dendl
;
5539 for (auto e
: extents
) {
5542 expected_statfs
.allocated
+= e
.length
;
5544 expected_statfs
.compressed_allocated
+= e
.length
;
5546 bool already
= false;
5548 e
.offset
, e
.length
, min_alloc_size
, used_blocks
, __func__
,
5549 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5556 derr
<< " " << oid
<< " extent " << e
5557 << " or a subset is already allocated" << dendl
;
5560 if (e
.end() > bdev
->get_size()) {
5561 derr
<< " " << oid
<< " extent " << e
5562 << " past end of block device" << dendl
;
5569 int BlueStore::_fsck(bool deep
, bool repair
)
5572 << (repair
? " fsck" : " repair")
5573 << (deep
? " (deep)" : " (shallow)") << " start" << dendl
;
5577 typedef btree::btree_set
<
5578 uint64_t,std::less
<uint64_t>,
5579 mempool::bluestore_fsck::pool_allocator
<uint64_t>> uint64_t_btree_t
;
5580 uint64_t_btree_t used_nids
;
5581 uint64_t_btree_t used_omap_head
;
5582 uint64_t_btree_t used_sbids
;
5584 mempool_dynamic_bitset used_blocks
;
5585 KeyValueDB::Iterator it
;
5586 store_statfs_t expected_statfs
, actual_statfs
;
5588 list
<ghobject_t
> oids
;
5590 bluestore_extent_ref_map_t ref_map
;
5593 mempool::bluestore_fsck::map
<uint64_t,sb_info_t
> sb_info
;
5595 uint64_t num_objects
= 0;
5596 uint64_t num_extents
= 0;
5597 uint64_t num_blobs
= 0;
5598 uint64_t num_spanning_blobs
= 0;
5599 uint64_t num_shared_blobs
= 0;
5600 uint64_t num_sharded_objects
= 0;
5601 uint64_t num_object_shards
= 0;
5603 utime_t start
= ceph_clock_now();
5605 int r
= _open_path();
5608 r
= _open_fsid(false);
5612 r
= _read_fsid(&fsid
);
5620 r
= _open_bdev(false);
5624 r
= _open_db(false);
5628 r
= _open_super_meta();
5632 r
= _open_fm(false);
5640 r
= _open_collections(&errors
);
5644 mempool_thread
.init();
5646 // we need finishers and kv_{sync,finalize}_thread *just* for replay
5648 r
= _deferred_replay();
5653 used_blocks
.resize(bdev
->get_size() / min_alloc_size
);
5655 0, MAX(min_alloc_size
, SUPER_RESERVED
), min_alloc_size
, used_blocks
,
5657 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5663 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
5665 e
.get_start(), e
.get_len(), min_alloc_size
, used_blocks
, "bluefs",
5666 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5679 // get expected statfs; fill unaffected fields to be able to compare
5681 statfs(&actual_statfs
);
5682 expected_statfs
.total
= actual_statfs
.total
;
5683 expected_statfs
.available
= actual_statfs
.available
;
5686 dout(1) << __func__
<< " walking object keyspace" << dendl
;
5687 it
= db
->get_iterator(PREFIX_OBJ
);
5691 mempool::bluestore_fsck::list
<string
> expecting_shards
;
5692 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5693 if (g_conf
->bluestore_debug_fsck_abort
) {
5696 dout(30) << " key " << pretty_binary_string(it
->key()) << dendl
;
5697 if (is_extent_shard_key(it
->key())) {
5698 while (!expecting_shards
.empty() &&
5699 expecting_shards
.front() < it
->key()) {
5700 derr
<< "fsck error: missing shard key "
5701 << pretty_binary_string(expecting_shards
.front())
5704 expecting_shards
.pop_front();
5706 if (!expecting_shards
.empty() &&
5707 expecting_shards
.front() == it
->key()) {
5709 expecting_shards
.pop_front();
5715 get_key_extent_shard(it
->key(), &okey
, &offset
);
5716 derr
<< "fsck error: stray shard 0x" << std::hex
<< offset
5717 << std::dec
<< dendl
;
5718 if (expecting_shards
.empty()) {
5719 derr
<< "fsck error: " << pretty_binary_string(it
->key())
5720 << " is unexpected" << dendl
;
5724 while (expecting_shards
.front() > it
->key()) {
5725 derr
<< "fsck error: saw " << pretty_binary_string(it
->key())
5727 derr
<< "fsck error: exp "
5728 << pretty_binary_string(expecting_shards
.front()) << dendl
;
5730 expecting_shards
.pop_front();
5731 if (expecting_shards
.empty()) {
5739 int r
= get_key_object(it
->key(), &oid
);
5741 derr
<< "fsck error: bad object key "
5742 << pretty_binary_string(it
->key()) << dendl
;
5747 oid
.shard_id
!= pgid
.shard
||
5748 oid
.hobj
.pool
!= (int64_t)pgid
.pool() ||
5749 !c
->contains(oid
)) {
5751 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
=
5753 p
!= coll_map
.end();
5755 if (p
->second
->contains(oid
)) {
5761 derr
<< "fsck error: stray object " << oid
5762 << " not owned by any collection" << dendl
;
5766 c
->cid
.is_pg(&pgid
);
5767 dout(20) << __func__
<< " collection " << c
->cid
<< dendl
;
5770 if (!expecting_shards
.empty()) {
5771 for (auto &k
: expecting_shards
) {
5772 derr
<< "fsck error: missing shard key "
5773 << pretty_binary_string(k
) << dendl
;
5776 expecting_shards
.clear();
5779 dout(10) << __func__
<< " " << oid
<< dendl
;
5780 RWLock::RLocker
l(c
->lock
);
5781 OnodeRef o
= c
->get_onode(oid
, false);
5783 if (o
->onode
.nid
> nid_max
) {
5784 derr
<< "fsck error: " << oid
<< " nid " << o
->onode
.nid
5785 << " > nid_max " << nid_max
<< dendl
;
5788 if (used_nids
.count(o
->onode
.nid
)) {
5789 derr
<< "fsck error: " << oid
<< " nid " << o
->onode
.nid
5790 << " already in use" << dendl
;
5792 continue; // go for next object
5794 used_nids
.insert(o
->onode
.nid
);
5797 num_spanning_blobs
+= o
->extent_map
.spanning_blob_map
.size();
5798 o
->extent_map
.fault_range(db
, 0, OBJECT_MAX_SIZE
);
5801 if (!o
->extent_map
.shards
.empty()) {
5802 ++num_sharded_objects
;
5803 num_object_shards
+= o
->extent_map
.shards
.size();
5805 for (auto& s
: o
->extent_map
.shards
) {
5806 dout(20) << __func__
<< " shard " << *s
.shard_info
<< dendl
;
5807 expecting_shards
.push_back(string());
5808 get_extent_shard_key(o
->key
, s
.shard_info
->offset
,
5809 &expecting_shards
.back());
5810 if (s
.shard_info
->offset
>= o
->onode
.size
) {
5811 derr
<< "fsck error: " << oid
<< " shard 0x" << std::hex
5812 << s
.shard_info
->offset
<< " past EOF at 0x" << o
->onode
.size
5813 << std::dec
<< dendl
;
5818 map
<BlobRef
,bluestore_blob_t::unused_t
> referenced
;
5820 mempool::bluestore_fsck::map
<BlobRef
,
5821 bluestore_blob_use_tracker_t
> ref_map
;
5822 for (auto& l
: o
->extent_map
.extent_map
) {
5823 dout(20) << __func__
<< " " << l
<< dendl
;
5824 if (l
.logical_offset
< pos
) {
5825 derr
<< "fsck error: " << oid
<< " lextent at 0x"
5826 << std::hex
<< l
.logical_offset
5827 << " overlaps with the previous, which ends at 0x" << pos
5828 << std::dec
<< dendl
;
5831 if (o
->extent_map
.spans_shard(l
.logical_offset
, l
.length
)) {
5832 derr
<< "fsck error: " << oid
<< " lextent at 0x"
5833 << std::hex
<< l
.logical_offset
<< "~" << l
.length
5834 << " spans a shard boundary"
5835 << std::dec
<< dendl
;
5838 pos
= l
.logical_offset
+ l
.length
;
5839 expected_statfs
.stored
+= l
.length
;
5841 const bluestore_blob_t
& blob
= l
.blob
->get_blob();
5843 auto& ref
= ref_map
[l
.blob
];
5844 if (ref
.is_empty()) {
5845 uint32_t min_release_size
= blob
.get_release_size(min_alloc_size
);
5846 uint32_t l
= blob
.get_logical_length();
5847 ref
.init(l
, min_release_size
);
5853 if (blob
.has_unused()) {
5854 auto p
= referenced
.find(l
.blob
);
5855 bluestore_blob_t::unused_t
*pu
;
5856 if (p
== referenced
.end()) {
5857 pu
= &referenced
[l
.blob
];
5861 uint64_t blob_len
= blob
.get_logical_length();
5862 assert((blob_len
% (sizeof(*pu
)*8)) == 0);
5863 assert(l
.blob_offset
+ l
.length
<= blob_len
);
5864 uint64_t chunk_size
= blob_len
/ (sizeof(*pu
)*8);
5865 uint64_t start
= l
.blob_offset
/ chunk_size
;
5867 ROUND_UP_TO(l
.blob_offset
+ l
.length
, chunk_size
) / chunk_size
;
5868 for (auto i
= start
; i
< end
; ++i
) {
5873 for (auto &i
: referenced
) {
5874 dout(20) << __func__
<< " referenced 0x" << std::hex
<< i
.second
5875 << std::dec
<< " for " << *i
.first
<< dendl
;
5876 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5877 if (i
.second
& blob
.unused
) {
5878 derr
<< "fsck error: " << oid
<< " blob claims unused 0x"
5879 << std::hex
<< blob
.unused
5880 << " but extents reference 0x" << i
.second
5881 << " on blob " << *i
.first
<< dendl
;
5884 if (blob
.has_csum()) {
5885 uint64_t blob_len
= blob
.get_logical_length();
5886 uint64_t unused_chunk_size
= blob_len
/ (sizeof(blob
.unused
)*8);
5887 unsigned csum_count
= blob
.get_csum_count();
5888 unsigned csum_chunk_size
= blob
.get_csum_chunk_size();
5889 for (unsigned p
= 0; p
< csum_count
; ++p
) {
5890 unsigned pos
= p
* csum_chunk_size
;
5891 unsigned firstbit
= pos
/ unused_chunk_size
; // [firstbit,lastbit]
5892 unsigned lastbit
= (pos
+ csum_chunk_size
- 1) / unused_chunk_size
;
5893 unsigned mask
= 1u << firstbit
;
5894 for (unsigned b
= firstbit
+ 1; b
<= lastbit
; ++b
) {
5897 if ((blob
.unused
& mask
) == mask
) {
5898 // this csum chunk region is marked unused
5899 if (blob
.get_csum_item(p
) != 0) {
5900 derr
<< "fsck error: " << oid
5901 << " blob claims csum chunk 0x" << std::hex
<< pos
5902 << "~" << csum_chunk_size
5903 << " is unused (mask 0x" << mask
<< " of unused 0x"
5904 << blob
.unused
<< ") but csum is non-zero 0x"
5905 << blob
.get_csum_item(p
) << std::dec
<< " on blob "
5906 << *i
.first
<< dendl
;
5913 for (auto &i
: ref_map
) {
5915 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5916 bool equal
= i
.first
->get_blob_use_tracker().equal(i
.second
);
5918 derr
<< "fsck error: " << oid
<< " blob " << *i
.first
5919 << " doesn't match expected ref_map " << i
.second
<< dendl
;
5922 if (blob
.is_compressed()) {
5923 expected_statfs
.compressed
+= blob
.get_compressed_payload_length();
5924 expected_statfs
.compressed_original
+=
5925 i
.first
->get_referenced_bytes();
5927 if (blob
.is_shared()) {
5928 if (i
.first
->shared_blob
->get_sbid() > blobid_max
) {
5929 derr
<< "fsck error: " << oid
<< " blob " << blob
5930 << " sbid " << i
.first
->shared_blob
->get_sbid() << " > blobid_max "
5931 << blobid_max
<< dendl
;
5933 } else if (i
.first
->shared_blob
->get_sbid() == 0) {
5934 derr
<< "fsck error: " << oid
<< " blob " << blob
5935 << " marked as shared but has uninitialized sbid"
5939 sb_info_t
& sbi
= sb_info
[i
.first
->shared_blob
->get_sbid()];
5940 sbi
.sb
= i
.first
->shared_blob
;
5941 sbi
.oids
.push_back(oid
);
5942 sbi
.compressed
= blob
.is_compressed();
5943 for (auto e
: blob
.get_extents()) {
5945 sbi
.ref_map
.get(e
.offset
, e
.length
);
5949 errors
+= _fsck_check_extents(oid
, blob
.get_extents(),
5950 blob
.is_compressed(),
5957 int r
= _do_read(c
.get(), o
, 0, o
->onode
.size
, bl
, 0);
5960 derr
<< "fsck error: " << oid
<< " error during read: "
5961 << cpp_strerror(r
) << dendl
;
5965 if (o
->onode
.has_omap()) {
5966 if (used_omap_head
.count(o
->onode
.nid
)) {
5967 derr
<< "fsck error: " << oid
<< " omap_head " << o
->onode
.nid
5968 << " already in use" << dendl
;
5971 used_omap_head
.insert(o
->onode
.nid
);
5976 dout(1) << __func__
<< " checking shared_blobs" << dendl
;
5977 it
= db
->get_iterator(PREFIX_SHARED_BLOB
);
5979 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5980 string key
= it
->key();
5982 if (get_key_shared_blob(key
, &sbid
)) {
5983 derr
<< "fsck error: bad key '" << key
5984 << "' in shared blob namespace" << dendl
;
5988 auto p
= sb_info
.find(sbid
);
5989 if (p
== sb_info
.end()) {
5990 derr
<< "fsck error: found stray shared blob data for sbid 0x"
5991 << std::hex
<< sbid
<< std::dec
<< dendl
;
5995 sb_info_t
& sbi
= p
->second
;
5996 bluestore_shared_blob_t
shared_blob(sbid
);
5997 bufferlist bl
= it
->value();
5998 bufferlist::iterator blp
= bl
.begin();
5999 ::decode(shared_blob
, blp
);
6000 dout(20) << __func__
<< " " << *sbi
.sb
<< " " << shared_blob
<< dendl
;
6001 if (shared_blob
.ref_map
!= sbi
.ref_map
) {
6002 derr
<< "fsck error: shared blob 0x" << std::hex
<< sbid
6003 << std::dec
<< " ref_map " << shared_blob
.ref_map
6004 << " != expected " << sbi
.ref_map
<< dendl
;
6007 PExtentVector extents
;
6008 for (auto &r
: shared_blob
.ref_map
.ref_map
) {
6009 extents
.emplace_back(bluestore_pextent_t(r
.first
, r
.second
.length
));
6011 errors
+= _fsck_check_extents(p
->second
.oids
.front(),
6013 p
->second
.compressed
,
6014 used_blocks
, expected_statfs
);
6019 for (auto &p
: sb_info
) {
6020 derr
<< "fsck error: shared_blob 0x" << p
.first
6021 << " key is missing (" << *p
.second
.sb
<< ")" << dendl
;
6024 if (!(actual_statfs
== expected_statfs
)) {
6025 derr
<< "fsck error: actual " << actual_statfs
6026 << " != expected " << expected_statfs
<< dendl
;
6030 dout(1) << __func__
<< " checking for stray omap data" << dendl
;
6031 it
= db
->get_iterator(PREFIX_OMAP
);
6033 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
6035 _key_decode_u64(it
->key().c_str(), &omap_head
);
6036 if (used_omap_head
.count(omap_head
) == 0) {
6037 derr
<< "fsck error: found stray omap data on omap_head "
6038 << omap_head
<< dendl
;
6044 dout(1) << __func__
<< " checking deferred events" << dendl
;
6045 it
= db
->get_iterator(PREFIX_DEFERRED
);
6047 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
6048 bufferlist bl
= it
->value();
6049 bufferlist::iterator p
= bl
.begin();
6050 bluestore_deferred_transaction_t wt
;
6053 } catch (buffer::error
& e
) {
6054 derr
<< "fsck error: failed to decode deferred txn "
6055 << pretty_binary_string(it
->key()) << dendl
;
6059 dout(20) << __func__
<< " deferred " << wt
.seq
6060 << " ops " << wt
.ops
.size()
6061 << " released 0x" << std::hex
<< wt
.released
<< std::dec
<< dendl
;
6062 for (auto e
= wt
.released
.begin(); e
!= wt
.released
.end(); ++e
) {
6064 e
.get_start(), e
.get_len(), min_alloc_size
, used_blocks
, "deferred",
6065 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
6073 dout(1) << __func__
<< " checking freelist vs allocated" << dendl
;
6075 // remove bluefs_extents from used set since the freelist doesn't
6076 // know they are allocated.
6077 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
6079 e
.get_start(), e
.get_len(), min_alloc_size
, used_blocks
,
6081 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
6086 fm
->enumerate_reset();
6087 uint64_t offset
, length
;
6088 while (fm
->enumerate_next(&offset
, &length
)) {
6089 bool intersects
= false;
6091 offset
, length
, min_alloc_size
, used_blocks
, "free",
6092 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
6101 if (offset
== SUPER_RESERVED
&&
6102 length
== min_alloc_size
- SUPER_RESERVED
) {
6103 // this is due to the change just after luminous to min_alloc_size
6104 // granularity allocations, and our baked in assumption at the top
6105 // of _fsck that 0~ROUND_UP_TO(SUPER_RESERVED,min_alloc_size) is used
6106 // (vs luminous's ROUND_UP_TO(SUPER_RESERVED,block_size)). harmless,
6107 // since we will never allocate this region below min_alloc_size.
6108 dout(10) << __func__
<< " ignoring free extent between SUPER_RESERVED"
6109 << " and min_alloc_size, 0x" << std::hex
<< offset
<< "~"
6112 derr
<< "fsck error: free extent 0x" << std::hex
<< offset
6113 << "~" << length
<< std::dec
6114 << " intersects allocated blocks" << dendl
;
6119 fm
->enumerate_reset();
6120 size_t count
= used_blocks
.count();
6121 if (used_blocks
.size() != count
) {
6122 assert(used_blocks
.size() > count
);
6125 size_t start
= used_blocks
.find_first();
6126 while (start
!= decltype(used_blocks
)::npos
) {
6129 size_t next
= used_blocks
.find_next(cur
);
6130 if (next
!= cur
+ 1) {
6131 derr
<< "fsck error: leaked extent 0x" << std::hex
6132 << ((uint64_t)start
* min_alloc_size
) << "~"
6133 << ((cur
+ 1 - start
) * min_alloc_size
) << std::dec
6146 mempool_thread
.shutdown();
6153 it
.reset(); // before db is closed
6162 // fatal errors take precedence
6166 dout(2) << __func__
<< " " << num_objects
<< " objects, "
6167 << num_sharded_objects
<< " of them sharded. "
6169 dout(2) << __func__
<< " " << num_extents
<< " extents to "
6170 << num_blobs
<< " blobs, "
6171 << num_spanning_blobs
<< " spanning, "
6172 << num_shared_blobs
<< " shared."
6175 utime_t duration
= ceph_clock_now() - start
;
6176 dout(1) << __func__
<< " finish with " << errors
<< " errors, " << repaired
6177 << " repaired, " << (errors
- repaired
) << " remaining in "
6178 << duration
<< " seconds" << dendl
;
6179 return errors
- repaired
;
6182 void BlueStore::collect_metadata(map
<string
,string
> *pm
)
6184 dout(10) << __func__
<< dendl
;
6185 bdev
->collect_metadata("bluestore_bdev_", pm
);
6187 (*pm
)["bluefs"] = "1";
6188 (*pm
)["bluefs_single_shared_device"] = stringify((int)bluefs_single_shared_device
);
6189 bluefs
->collect_metadata(pm
);
6191 (*pm
)["bluefs"] = "0";
6195 int BlueStore::statfs(struct store_statfs_t
*buf
)
6198 buf
->total
= bdev
->get_size();
6199 buf
->available
= alloc
->get_free();
6202 // part of our shared device is "free" according to BlueFS
6203 // Don't include bluestore_bluefs_min because that space can't
6204 // be used for any other purpose.
6205 buf
->available
+= bluefs
->get_free(bluefs_shared_bdev
) - cct
->_conf
->bluestore_bluefs_min
;
6207 // include dedicated db, too, if that isn't the shared device.
6208 if (bluefs_shared_bdev
!= BlueFS::BDEV_DB
) {
6209 buf
->total
+= bluefs
->get_total(BlueFS::BDEV_DB
);
6214 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
6216 buf
->allocated
= vstatfs
.allocated();
6217 buf
->stored
= vstatfs
.stored();
6218 buf
->compressed
= vstatfs
.compressed();
6219 buf
->compressed_original
= vstatfs
.compressed_original();
6220 buf
->compressed_allocated
= vstatfs
.compressed_allocated();
6223 dout(20) << __func__
<< *buf
<< dendl
;
6230 BlueStore::CollectionRef
BlueStore::_get_collection(const coll_t
& cid
)
6232 RWLock::RLocker
l(coll_lock
);
6233 ceph::unordered_map
<coll_t
,CollectionRef
>::iterator cp
= coll_map
.find(cid
);
6234 if (cp
== coll_map
.end())
6235 return CollectionRef();
6239 void BlueStore::_queue_reap_collection(CollectionRef
& c
)
6241 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6242 std::lock_guard
<std::mutex
> l(reap_lock
);
6243 removed_collections
.push_back(c
);
6246 void BlueStore::_reap_collections()
6248 list
<CollectionRef
> removed_colls
;
6250 std::lock_guard
<std::mutex
> l(reap_lock
);
6251 removed_colls
.swap(removed_collections
);
6254 bool all_reaped
= true;
6256 for (list
<CollectionRef
>::iterator p
= removed_colls
.begin();
6257 p
!= removed_colls
.end();
6259 CollectionRef c
= *p
;
6260 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6261 if (c
->onode_map
.map_any([&](OnodeRef o
) {
6263 if (o
->flushing_count
.load()) {
6264 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " " << o
->oid
6265 << " flush_txns " << o
->flushing_count
<< dendl
;
6273 c
->onode_map
.clear();
6274 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " done" << dendl
;
6278 dout(10) << __func__
<< " all reaped" << dendl
;
6282 void BlueStore::_update_cache_logger()
6284 uint64_t num_onodes
= 0;
6285 uint64_t num_extents
= 0;
6286 uint64_t num_blobs
= 0;
6287 uint64_t num_buffers
= 0;
6288 uint64_t num_buffer_bytes
= 0;
6289 for (auto c
: cache_shards
) {
6290 c
->add_stats(&num_onodes
, &num_extents
, &num_blobs
,
6291 &num_buffers
, &num_buffer_bytes
);
6293 logger
->set(l_bluestore_onodes
, num_onodes
);
6294 logger
->set(l_bluestore_extents
, num_extents
);
6295 logger
->set(l_bluestore_blobs
, num_blobs
);
6296 logger
->set(l_bluestore_buffers
, num_buffers
);
6297 logger
->set(l_bluestore_buffer_bytes
, num_buffer_bytes
);
6303 ObjectStore::CollectionHandle
BlueStore::open_collection(const coll_t
& cid
)
6305 return _get_collection(cid
);
6308 bool BlueStore::exists(const coll_t
& cid
, const ghobject_t
& oid
)
6310 CollectionHandle c
= _get_collection(cid
);
6313 return exists(c
, oid
);
6316 bool BlueStore::exists(CollectionHandle
&c_
, const ghobject_t
& oid
)
6318 Collection
*c
= static_cast<Collection
*>(c_
.get());
6319 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
6326 RWLock::RLocker
l(c
->lock
);
6327 OnodeRef o
= c
->get_onode(oid
, false);
6328 if (!o
|| !o
->exists
)
6335 int BlueStore::stat(
6337 const ghobject_t
& oid
,
6341 CollectionHandle c
= _get_collection(cid
);
6344 return stat(c
, oid
, st
, allow_eio
);
6347 int BlueStore::stat(
6348 CollectionHandle
&c_
,
6349 const ghobject_t
& oid
,
6353 Collection
*c
= static_cast<Collection
*>(c_
.get());
6356 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
6359 RWLock::RLocker
l(c
->lock
);
6360 OnodeRef o
= c
->get_onode(oid
, false);
6361 if (!o
|| !o
->exists
)
6363 st
->st_size
= o
->onode
.size
;
6364 st
->st_blksize
= 4096;
6365 st
->st_blocks
= (st
->st_size
+ st
->st_blksize
- 1) / st
->st_blksize
;
6370 if (_debug_mdata_eio(oid
)) {
6372 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6376 int BlueStore::set_collection_opts(
6378 const pool_opts_t
& opts
)
6380 CollectionHandle ch
= _get_collection(cid
);
6383 Collection
*c
= static_cast<Collection
*>(ch
.get());
6384 dout(15) << __func__
<< " " << cid
<< " options " << opts
<< dendl
;
6387 RWLock::WLocker
l(c
->lock
);
6388 c
->pool_opts
= opts
;
6392 int BlueStore::read(
6394 const ghobject_t
& oid
,
6400 CollectionHandle c
= _get_collection(cid
);
6403 return read(c
, oid
, offset
, length
, bl
, op_flags
);
6406 int BlueStore::read(
6407 CollectionHandle
&c_
,
6408 const ghobject_t
& oid
,
6414 utime_t start
= ceph_clock_now();
6415 Collection
*c
= static_cast<Collection
*>(c_
.get());
6416 const coll_t
&cid
= c
->get_cid();
6417 dout(15) << __func__
<< " " << cid
<< " " << oid
6418 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6426 RWLock::RLocker
l(c
->lock
);
6427 utime_t start1
= ceph_clock_now();
6428 OnodeRef o
= c
->get_onode(oid
, false);
6429 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start1
);
6430 if (!o
|| !o
->exists
) {
6435 if (offset
== length
&& offset
== 0)
6436 length
= o
->onode
.size
;
6438 r
= _do_read(c
, o
, offset
, length
, bl
, op_flags
);
6442 if (r
== 0 && _debug_data_eio(oid
)) {
6444 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6445 } else if (cct
->_conf
->bluestore_debug_random_read_err
&&
6446 (rand() % (int)(cct
->_conf
->bluestore_debug_random_read_err
* 100.0)) == 0) {
6447 dout(0) << __func__
<< ": inject random EIO" << dendl
;
6450 dout(10) << __func__
<< " " << cid
<< " " << oid
6451 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6452 << " = " << r
<< dendl
;
6453 logger
->tinc(l_bluestore_read_lat
, ceph_clock_now() - start
);
6457 // --------------------------------------------------------
6458 // intermediate data structures used while reading
6460 uint64_t logical_offset
;
6461 uint64_t blob_xoffset
; //region offset within the blob
6465 // used later in read process
6469 region_t(uint64_t offset
, uint64_t b_offs
, uint64_t len
)
6470 : logical_offset(offset
),
6471 blob_xoffset(b_offs
),
6473 region_t(const region_t
& from
)
6474 : logical_offset(from
.logical_offset
),
6475 blob_xoffset(from
.blob_xoffset
),
6476 length(from
.length
){}
6478 friend ostream
& operator<<(ostream
& out
, const region_t
& r
) {
6479 return out
<< "0x" << std::hex
<< r
.logical_offset
<< ":"
6480 << r
.blob_xoffset
<< "~" << r
.length
<< std::dec
;
6484 typedef list
<region_t
> regions2read_t
;
6485 typedef map
<BlueStore::BlobRef
, regions2read_t
> blobs2read_t
;
6487 int BlueStore::_do_read(
6498 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6499 << " size 0x" << o
->onode
.size
<< " (" << std::dec
6500 << o
->onode
.size
<< ")" << dendl
;
6503 if (offset
>= o
->onode
.size
) {
6507 // generally, don't buffer anything, unless the client explicitly requests
6509 bool buffered
= false;
6510 if (op_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
6511 dout(20) << __func__
<< " will do buffered read" << dendl
;
6513 } else if (cct
->_conf
->bluestore_default_buffered_read
&&
6514 (op_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
6515 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
6516 dout(20) << __func__
<< " defaulting to buffered read" << dendl
;
6520 if (offset
+ length
> o
->onode
.size
) {
6521 length
= o
->onode
.size
- offset
;
6524 utime_t start
= ceph_clock_now();
6525 o
->extent_map
.fault_range(db
, offset
, length
);
6526 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start
);
6529 ready_regions_t ready_regions
;
6531 // build blob-wise list to of stuff read (that isn't cached)
6532 blobs2read_t blobs2read
;
6533 unsigned left
= length
;
6534 uint64_t pos
= offset
;
6535 unsigned num_regions
= 0;
6536 auto lp
= o
->extent_map
.seek_lextent(offset
);
6537 while (left
> 0 && lp
!= o
->extent_map
.extent_map
.end()) {
6538 if (pos
< lp
->logical_offset
) {
6539 unsigned hole
= lp
->logical_offset
- pos
;
6543 dout(30) << __func__
<< " hole 0x" << std::hex
<< pos
<< "~" << hole
6544 << std::dec
<< dendl
;
6548 BlobRef bptr
= lp
->blob
;
6549 unsigned l_off
= pos
- lp
->logical_offset
;
6550 unsigned b_off
= l_off
+ lp
->blob_offset
;
6551 unsigned b_len
= std::min(left
, lp
->length
- l_off
);
6553 ready_regions_t cache_res
;
6554 interval_set
<uint32_t> cache_interval
;
6555 bptr
->shared_blob
->bc
.read(
6556 bptr
->shared_blob
->get_cache(), b_off
, b_len
, cache_res
, cache_interval
);
6557 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6558 << " need 0x" << b_off
<< "~" << b_len
6559 << " cache has 0x" << cache_interval
6560 << std::dec
<< dendl
;
6562 auto pc
= cache_res
.begin();
6565 if (pc
!= cache_res
.end() &&
6566 pc
->first
== b_off
) {
6567 l
= pc
->second
.length();
6568 ready_regions
[pos
].claim(pc
->second
);
6569 dout(30) << __func__
<< " use cache 0x" << std::hex
<< pos
<< ": 0x"
6570 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6574 if (pc
!= cache_res
.end()) {
6575 assert(pc
->first
> b_off
);
6576 l
= pc
->first
- b_off
;
6578 dout(30) << __func__
<< " will read 0x" << std::hex
<< pos
<< ": 0x"
6579 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6580 blobs2read
[bptr
].emplace_back(region_t(pos
, b_off
, l
));
6591 // read raw blob data. use aio if we have >1 blobs to read.
6592 start
= ceph_clock_now(); // for the sake of simplicity
6593 // measure the whole block below.
6594 // The error isn't that much...
6595 vector
<bufferlist
> compressed_blob_bls
;
6596 IOContext
ioc(cct
, NULL
);
6597 for (auto& p
: blobs2read
) {
6598 BlobRef bptr
= p
.first
;
6599 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6600 << " need " << p
.second
<< std::dec
<< dendl
;
6601 if (bptr
->get_blob().is_compressed()) {
6602 // read the whole thing
6603 if (compressed_blob_bls
.empty()) {
6604 // ensure we avoid any reallocation on subsequent blobs
6605 compressed_blob_bls
.reserve(blobs2read
.size());
6607 compressed_blob_bls
.push_back(bufferlist());
6608 bufferlist
& bl
= compressed_blob_bls
.back();
6609 r
= bptr
->get_blob().map(
6610 0, bptr
->get_blob().get_ondisk_length(),
6611 [&](uint64_t offset
, uint64_t length
) {
6613 // use aio if there are more regions to read than those in this blob
6614 if (num_regions
> p
.second
.size()) {
6615 r
= bdev
->aio_read(offset
, length
, &bl
, &ioc
);
6617 r
= bdev
->read(offset
, length
, &bl
, &ioc
, false);
6626 for (auto& reg
: p
.second
) {
6627 // determine how much of the blob to read
6628 uint64_t chunk_size
= bptr
->get_blob().get_chunk_size(block_size
);
6629 reg
.r_off
= reg
.blob_xoffset
;
6630 uint64_t r_len
= reg
.length
;
6631 reg
.front
= reg
.r_off
% chunk_size
;
6633 reg
.r_off
-= reg
.front
;
6636 unsigned tail
= r_len
% chunk_size
;
6638 r_len
+= chunk_size
- tail
;
6640 dout(20) << __func__
<< " region 0x" << std::hex
6641 << reg
.logical_offset
6642 << ": 0x" << reg
.blob_xoffset
<< "~" << reg
.length
6643 << " reading 0x" << reg
.r_off
<< "~" << r_len
<< std::dec
6647 r
= bptr
->get_blob().map(
6649 [&](uint64_t offset
, uint64_t length
) {
6651 // use aio if there is more than one region to read
6652 if (num_regions
> 1) {
6653 r
= bdev
->aio_read(offset
, length
, ®
.bl
, &ioc
);
6655 r
= bdev
->read(offset
, length
, ®
.bl
, &ioc
, false);
6662 assert(reg
.bl
.length() == r_len
);
6666 if (ioc
.has_pending_aios()) {
6667 bdev
->aio_submit(&ioc
);
6668 dout(20) << __func__
<< " waiting for aio" << dendl
;
6671 logger
->tinc(l_bluestore_read_wait_aio_lat
, ceph_clock_now() - start
);
6673 // enumerate and decompress desired blobs
6674 auto p
= compressed_blob_bls
.begin();
6675 blobs2read_t::iterator b2r_it
= blobs2read
.begin();
6676 while (b2r_it
!= blobs2read
.end()) {
6677 BlobRef bptr
= b2r_it
->first
;
6678 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6679 << " need 0x" << b2r_it
->second
<< std::dec
<< dendl
;
6680 if (bptr
->get_blob().is_compressed()) {
6681 assert(p
!= compressed_blob_bls
.end());
6682 bufferlist
& compressed_bl
= *p
++;
6683 if (_verify_csum(o
, &bptr
->get_blob(), 0, compressed_bl
,
6684 b2r_it
->second
.front().logical_offset
) < 0) {
6688 r
= _decompress(compressed_bl
, &raw_bl
);
6692 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(), 0,
6695 for (auto& i
: b2r_it
->second
) {
6696 ready_regions
[i
.logical_offset
].substr_of(
6697 raw_bl
, i
.blob_xoffset
, i
.length
);
6700 for (auto& reg
: b2r_it
->second
) {
6701 if (_verify_csum(o
, &bptr
->get_blob(), reg
.r_off
, reg
.bl
,
6702 reg
.logical_offset
) < 0) {
6706 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(),
6710 // prune and keep result
6711 ready_regions
[reg
.logical_offset
].substr_of(
6712 reg
.bl
, reg
.front
, reg
.length
);
6718 // generate a resulting buffer
6719 auto pr
= ready_regions
.begin();
6720 auto pr_end
= ready_regions
.end();
6722 while (pos
< length
) {
6723 if (pr
!= pr_end
&& pr
->first
== pos
+ offset
) {
6724 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6725 << ": data from 0x" << pr
->first
<< "~" << pr
->second
.length()
6726 << std::dec
<< dendl
;
6727 pos
+= pr
->second
.length();
6728 bl
.claim_append(pr
->second
);
6731 uint64_t l
= length
- pos
;
6733 assert(pr
->first
> pos
+ offset
);
6734 l
= pr
->first
- (pos
+ offset
);
6736 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6737 << ": zeros for 0x" << (pos
+ offset
) << "~" << l
6738 << std::dec
<< dendl
;
6743 assert(bl
.length() == length
);
6744 assert(pos
== length
);
6745 assert(pr
== pr_end
);
6750 int BlueStore::_verify_csum(OnodeRef
& o
,
6751 const bluestore_blob_t
* blob
, uint64_t blob_xoffset
,
6752 const bufferlist
& bl
,
6753 uint64_t logical_offset
) const
6757 utime_t start
= ceph_clock_now();
6758 int r
= blob
->verify_csum(blob_xoffset
, bl
, &bad
, &bad_csum
);
6764 blob
->get_csum_chunk_size(),
6765 [&](uint64_t offset
, uint64_t length
) {
6766 pex
.emplace_back(bluestore_pextent_t(offset
, length
));
6769 derr
<< __func__
<< " bad "
6770 << Checksummer::get_csum_type_string(blob
->csum_type
)
6771 << "/0x" << std::hex
<< blob
->get_csum_chunk_size()
6772 << " checksum at blob offset 0x" << bad
6773 << ", got 0x" << bad_csum
<< ", expected 0x"
6774 << blob
->get_csum_item(bad
/ blob
->get_csum_chunk_size()) << std::dec
6775 << ", device location " << pex
6776 << ", logical extent 0x" << std::hex
6777 << (logical_offset
+ bad
- blob_xoffset
) << "~"
6778 << blob
->get_csum_chunk_size() << std::dec
6779 << ", object " << o
->oid
6782 derr
<< __func__
<< " failed with exit code: " << cpp_strerror(r
) << dendl
;
6785 logger
->tinc(l_bluestore_csum_lat
, ceph_clock_now() - start
);
6789 int BlueStore::_decompress(bufferlist
& source
, bufferlist
* result
)
6792 utime_t start
= ceph_clock_now();
6793 bufferlist::iterator i
= source
.begin();
6794 bluestore_compression_header_t chdr
;
6796 int alg
= int(chdr
.type
);
6797 CompressorRef cp
= compressor
;
6798 if (!cp
|| (int)cp
->get_type() != alg
) {
6799 cp
= Compressor::create(cct
, alg
);
6803 // if compressor isn't available - error, because cannot return
6804 // decompressed data?
6805 derr
<< __func__
<< " can't load decompressor " << alg
<< dendl
;
6808 r
= cp
->decompress(i
, chdr
.length
, *result
);
6810 derr
<< __func__
<< " decompression failed with exit code " << r
<< dendl
;
6814 logger
->tinc(l_bluestore_decompress_lat
, ceph_clock_now() - start
);
6818 // this stores fiemap into interval_set, other variations
6819 // use it internally
6820 int BlueStore::_fiemap(
6821 CollectionHandle
&c_
,
6822 const ghobject_t
& oid
,
6825 interval_set
<uint64_t>& destset
)
6827 Collection
*c
= static_cast<Collection
*>(c_
.get());
6831 RWLock::RLocker
l(c
->lock
);
6833 OnodeRef o
= c
->get_onode(oid
, false);
6834 if (!o
|| !o
->exists
) {
6839 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6840 << " size 0x" << o
->onode
.size
<< std::dec
<< dendl
;
6842 boost::intrusive::set
<Extent
>::iterator ep
, eend
;
6843 if (offset
>= o
->onode
.size
)
6846 if (offset
+ length
> o
->onode
.size
) {
6847 length
= o
->onode
.size
- offset
;
6850 o
->extent_map
.fault_range(db
, offset
, length
);
6851 eend
= o
->extent_map
.extent_map
.end();
6852 ep
= o
->extent_map
.seek_lextent(offset
);
6853 while (length
> 0) {
6854 dout(20) << __func__
<< " offset " << offset
<< dendl
;
6855 if (ep
!= eend
&& ep
->logical_offset
+ ep
->length
<= offset
) {
6860 uint64_t x_len
= length
;
6861 if (ep
!= eend
&& ep
->logical_offset
<= offset
) {
6862 uint64_t x_off
= offset
- ep
->logical_offset
;
6863 x_len
= MIN(x_len
, ep
->length
- x_off
);
6864 dout(30) << __func__
<< " lextent 0x" << std::hex
<< offset
<< "~"
6865 << x_len
<< std::dec
<< " blob " << ep
->blob
<< dendl
;
6866 destset
.insert(offset
, x_len
);
6869 if (x_off
+ x_len
== ep
->length
)
6874 ep
->logical_offset
> offset
&&
6875 ep
->logical_offset
- offset
< x_len
) {
6876 x_len
= ep
->logical_offset
- offset
;
6884 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6885 << " size = 0x(" << destset
<< ")" << std::dec
<< dendl
;
6889 int BlueStore::fiemap(
6891 const ghobject_t
& oid
,
6896 CollectionHandle c
= _get_collection(cid
);
6899 return fiemap(c
, oid
, offset
, len
, bl
);
6902 int BlueStore::fiemap(
6903 CollectionHandle
&c_
,
6904 const ghobject_t
& oid
,
6909 interval_set
<uint64_t> m
;
6910 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6917 int BlueStore::fiemap(
6919 const ghobject_t
& oid
,
6922 map
<uint64_t, uint64_t>& destmap
)
6924 CollectionHandle c
= _get_collection(cid
);
6927 return fiemap(c
, oid
, offset
, len
, destmap
);
6930 int BlueStore::fiemap(
6931 CollectionHandle
&c_
,
6932 const ghobject_t
& oid
,
6935 map
<uint64_t, uint64_t>& destmap
)
6937 interval_set
<uint64_t> m
;
6938 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6940 m
.move_into(destmap
);
6945 int BlueStore::getattr(
6947 const ghobject_t
& oid
,
6951 CollectionHandle c
= _get_collection(cid
);
6954 return getattr(c
, oid
, name
, value
);
6957 int BlueStore::getattr(
6958 CollectionHandle
&c_
,
6959 const ghobject_t
& oid
,
6963 Collection
*c
= static_cast<Collection
*>(c_
.get());
6964 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
<< dendl
;
6970 RWLock::RLocker
l(c
->lock
);
6971 mempool::bluestore_cache_other::string
k(name
);
6973 OnodeRef o
= c
->get_onode(oid
, false);
6974 if (!o
|| !o
->exists
) {
6979 if (!o
->onode
.attrs
.count(k
)) {
6983 value
= o
->onode
.attrs
[k
];
6987 if (r
== 0 && _debug_mdata_eio(oid
)) {
6989 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6991 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
6992 << " = " << r
<< dendl
;
6997 int BlueStore::getattrs(
6999 const ghobject_t
& oid
,
7000 map
<string
,bufferptr
>& aset
)
7002 CollectionHandle c
= _get_collection(cid
);
7005 return getattrs(c
, oid
, aset
);
7008 int BlueStore::getattrs(
7009 CollectionHandle
&c_
,
7010 const ghobject_t
& oid
,
7011 map
<string
,bufferptr
>& aset
)
7013 Collection
*c
= static_cast<Collection
*>(c_
.get());
7014 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
7020 RWLock::RLocker
l(c
->lock
);
7022 OnodeRef o
= c
->get_onode(oid
, false);
7023 if (!o
|| !o
->exists
) {
7027 for (auto& i
: o
->onode
.attrs
) {
7028 aset
.emplace(i
.first
.c_str(), i
.second
);
7034 if (r
== 0 && _debug_mdata_eio(oid
)) {
7036 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
7038 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
7039 << " = " << r
<< dendl
;
7043 int BlueStore::list_collections(vector
<coll_t
>& ls
)
7045 RWLock::RLocker
l(coll_lock
);
7046 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
= coll_map
.begin();
7047 p
!= coll_map
.end();
7049 ls
.push_back(p
->first
);
7053 bool BlueStore::collection_exists(const coll_t
& c
)
7055 RWLock::RLocker
l(coll_lock
);
7056 return coll_map
.count(c
);
7059 int BlueStore::collection_empty(const coll_t
& cid
, bool *empty
)
7061 dout(15) << __func__
<< " " << cid
<< dendl
;
7062 vector
<ghobject_t
> ls
;
7064 int r
= collection_list(cid
, ghobject_t(), ghobject_t::get_max(), 1,
7067 derr
<< __func__
<< " collection_list returned: " << cpp_strerror(r
)
7071 *empty
= ls
.empty();
7072 dout(10) << __func__
<< " " << cid
<< " = " << (int)(*empty
) << dendl
;
7076 int BlueStore::collection_bits(const coll_t
& cid
)
7078 dout(15) << __func__
<< " " << cid
<< dendl
;
7079 CollectionRef c
= _get_collection(cid
);
7082 RWLock::RLocker
l(c
->lock
);
7083 dout(10) << __func__
<< " " << cid
<< " = " << c
->cnode
.bits
<< dendl
;
7084 return c
->cnode
.bits
;
7087 int BlueStore::collection_list(
7088 const coll_t
& cid
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
7089 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
7091 CollectionHandle c
= _get_collection(cid
);
7094 return collection_list(c
, start
, end
, max
, ls
, pnext
);
7097 int BlueStore::collection_list(
7098 CollectionHandle
&c_
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
7099 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
7101 Collection
*c
= static_cast<Collection
*>(c_
.get());
7102 dout(15) << __func__
<< " " << c
->cid
7103 << " start " << start
<< " end " << end
<< " max " << max
<< dendl
;
7106 RWLock::RLocker
l(c
->lock
);
7107 r
= _collection_list(c
, start
, end
, max
, ls
, pnext
);
7110 dout(10) << __func__
<< " " << c
->cid
7111 << " start " << start
<< " end " << end
<< " max " << max
7112 << " = " << r
<< ", ls.size() = " << ls
->size()
7113 << ", next = " << (pnext
? *pnext
: ghobject_t()) << dendl
;
7117 int BlueStore::_collection_list(
7118 Collection
*c
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
7119 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
7126 ghobject_t static_next
;
7127 KeyValueDB::Iterator it
;
7128 string temp_start_key
, temp_end_key
;
7129 string start_key
, end_key
;
7130 bool set_next
= false;
7135 pnext
= &static_next
;
7137 if (start
== ghobject_t::get_max() ||
7138 start
.hobj
.is_max()) {
7141 get_coll_key_range(c
->cid
, c
->cnode
.bits
, &temp_start_key
, &temp_end_key
,
7142 &start_key
, &end_key
);
7143 dout(20) << __func__
7144 << " range " << pretty_binary_string(temp_start_key
)
7145 << " to " << pretty_binary_string(temp_end_key
)
7146 << " and " << pretty_binary_string(start_key
)
7147 << " to " << pretty_binary_string(end_key
)
7148 << " start " << start
<< dendl
;
7149 it
= db
->get_iterator(PREFIX_OBJ
);
7150 if (start
== ghobject_t() ||
7151 start
.hobj
== hobject_t() ||
7152 start
== c
->cid
.get_min_hobj()) {
7153 it
->upper_bound(temp_start_key
);
7157 get_object_key(cct
, start
, &k
);
7158 if (start
.hobj
.is_temp()) {
7160 assert(k
>= temp_start_key
&& k
< temp_end_key
);
7163 assert(k
>= start_key
&& k
< end_key
);
7165 dout(20) << " start from " << pretty_binary_string(k
)
7166 << " temp=" << (int)temp
<< dendl
;
7169 if (end
.hobj
.is_max()) {
7170 pend
= temp
? temp_end_key
: end_key
;
7172 get_object_key(cct
, end
, &end_key
);
7173 if (end
.hobj
.is_temp()) {
7179 pend
= temp
? temp_end_key
: end_key
;
7182 dout(20) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7184 if (!it
->valid() || it
->key() >= pend
) {
7186 dout(20) << __func__
<< " iterator not valid (end of db?)" << dendl
;
7188 dout(20) << __func__
<< " key " << pretty_binary_string(it
->key())
7189 << " >= " << end
<< dendl
;
7191 if (end
.hobj
.is_temp()) {
7194 dout(30) << __func__
<< " switch to non-temp namespace" << dendl
;
7196 it
->upper_bound(start_key
);
7198 dout(30) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7203 dout(30) << __func__
<< " key " << pretty_binary_string(it
->key()) << dendl
;
7204 if (is_extent_shard_key(it
->key())) {
7209 int r
= get_key_object(it
->key(), &oid
);
7211 dout(20) << __func__
<< " oid " << oid
<< " end " << end
<< dendl
;
7212 if (ls
->size() >= (unsigned)max
) {
7213 dout(20) << __func__
<< " reached max " << max
<< dendl
;
7223 *pnext
= ghobject_t::get_max();
7229 int BlueStore::omap_get(
7230 const coll_t
& cid
, ///< [in] Collection containing oid
7231 const ghobject_t
&oid
, ///< [in] Object containing omap
7232 bufferlist
*header
, ///< [out] omap header
7233 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7236 CollectionHandle c
= _get_collection(cid
);
7239 return omap_get(c
, oid
, header
, out
);
7242 int BlueStore::omap_get(
7243 CollectionHandle
&c_
, ///< [in] Collection containing oid
7244 const ghobject_t
&oid
, ///< [in] Object containing omap
7245 bufferlist
*header
, ///< [out] omap header
7246 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7249 Collection
*c
= static_cast<Collection
*>(c_
.get());
7250 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7253 RWLock::RLocker
l(c
->lock
);
7255 OnodeRef o
= c
->get_onode(oid
, false);
7256 if (!o
|| !o
->exists
) {
7260 if (!o
->onode
.has_omap())
7264 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7266 get_omap_header(o
->onode
.nid
, &head
);
7267 get_omap_tail(o
->onode
.nid
, &tail
);
7268 it
->lower_bound(head
);
7269 while (it
->valid()) {
7270 if (it
->key() == head
) {
7271 dout(30) << __func__
<< " got header" << dendl
;
7272 *header
= it
->value();
7273 } else if (it
->key() >= tail
) {
7274 dout(30) << __func__
<< " reached tail" << dendl
;
7278 decode_omap_key(it
->key(), &user_key
);
7279 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7280 << " -> " << user_key
<< dendl
;
7281 (*out
)[user_key
] = it
->value();
7287 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7292 int BlueStore::omap_get_header(
7293 const coll_t
& cid
, ///< [in] Collection containing oid
7294 const ghobject_t
&oid
, ///< [in] Object containing omap
7295 bufferlist
*header
, ///< [out] omap header
7296 bool allow_eio
///< [in] don't assert on eio
7299 CollectionHandle c
= _get_collection(cid
);
7302 return omap_get_header(c
, oid
, header
, allow_eio
);
7305 int BlueStore::omap_get_header(
7306 CollectionHandle
&c_
, ///< [in] Collection containing oid
7307 const ghobject_t
&oid
, ///< [in] Object containing omap
7308 bufferlist
*header
, ///< [out] omap header
7309 bool allow_eio
///< [in] don't assert on eio
7312 Collection
*c
= static_cast<Collection
*>(c_
.get());
7313 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7316 RWLock::RLocker
l(c
->lock
);
7318 OnodeRef o
= c
->get_onode(oid
, false);
7319 if (!o
|| !o
->exists
) {
7323 if (!o
->onode
.has_omap())
7328 get_omap_header(o
->onode
.nid
, &head
);
7329 if (db
->get(PREFIX_OMAP
, head
, header
) >= 0) {
7330 dout(30) << __func__
<< " got header" << dendl
;
7332 dout(30) << __func__
<< " no header" << dendl
;
7336 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7341 int BlueStore::omap_get_keys(
7342 const coll_t
& cid
, ///< [in] Collection containing oid
7343 const ghobject_t
&oid
, ///< [in] Object containing omap
7344 set
<string
> *keys
///< [out] Keys defined on oid
7347 CollectionHandle c
= _get_collection(cid
);
7350 return omap_get_keys(c
, oid
, keys
);
7353 int BlueStore::omap_get_keys(
7354 CollectionHandle
&c_
, ///< [in] Collection containing oid
7355 const ghobject_t
&oid
, ///< [in] Object containing omap
7356 set
<string
> *keys
///< [out] Keys defined on oid
7359 Collection
*c
= static_cast<Collection
*>(c_
.get());
7360 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7363 RWLock::RLocker
l(c
->lock
);
7365 OnodeRef o
= c
->get_onode(oid
, false);
7366 if (!o
|| !o
->exists
) {
7370 if (!o
->onode
.has_omap())
7374 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7376 get_omap_key(o
->onode
.nid
, string(), &head
);
7377 get_omap_tail(o
->onode
.nid
, &tail
);
7378 it
->lower_bound(head
);
7379 while (it
->valid()) {
7380 if (it
->key() >= tail
) {
7381 dout(30) << __func__
<< " reached tail" << dendl
;
7385 decode_omap_key(it
->key(), &user_key
);
7386 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7387 << " -> " << user_key
<< dendl
;
7388 keys
->insert(user_key
);
7393 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7398 int BlueStore::omap_get_values(
7399 const coll_t
& cid
, ///< [in] Collection containing oid
7400 const ghobject_t
&oid
, ///< [in] Object containing omap
7401 const set
<string
> &keys
, ///< [in] Keys to get
7402 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7405 CollectionHandle c
= _get_collection(cid
);
7408 return omap_get_values(c
, oid
, keys
, out
);
7411 int BlueStore::omap_get_values(
7412 CollectionHandle
&c_
, ///< [in] Collection containing oid
7413 const ghobject_t
&oid
, ///< [in] Object containing omap
7414 const set
<string
> &keys
, ///< [in] Keys to get
7415 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7418 Collection
*c
= static_cast<Collection
*>(c_
.get());
7419 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7422 RWLock::RLocker
l(c
->lock
);
7425 OnodeRef o
= c
->get_onode(oid
, false);
7426 if (!o
|| !o
->exists
) {
7430 if (!o
->onode
.has_omap())
7433 _key_encode_u64(o
->onode
.nid
, &final_key
);
7434 final_key
.push_back('.');
7435 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7436 final_key
.resize(9); // keep prefix
7439 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7440 dout(30) << __func__
<< " got " << pretty_binary_string(final_key
)
7441 << " -> " << *p
<< dendl
;
7442 out
->insert(make_pair(*p
, val
));
7446 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7451 int BlueStore::omap_check_keys(
7452 const coll_t
& cid
, ///< [in] Collection containing oid
7453 const ghobject_t
&oid
, ///< [in] Object containing omap
7454 const set
<string
> &keys
, ///< [in] Keys to check
7455 set
<string
> *out
///< [out] Subset of keys defined on oid
7458 CollectionHandle c
= _get_collection(cid
);
7461 return omap_check_keys(c
, oid
, keys
, out
);
7464 int BlueStore::omap_check_keys(
7465 CollectionHandle
&c_
, ///< [in] Collection containing oid
7466 const ghobject_t
&oid
, ///< [in] Object containing omap
7467 const set
<string
> &keys
, ///< [in] Keys to check
7468 set
<string
> *out
///< [out] Subset of keys defined on oid
7471 Collection
*c
= static_cast<Collection
*>(c_
.get());
7472 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7475 RWLock::RLocker
l(c
->lock
);
7478 OnodeRef o
= c
->get_onode(oid
, false);
7479 if (!o
|| !o
->exists
) {
7483 if (!o
->onode
.has_omap())
7486 _key_encode_u64(o
->onode
.nid
, &final_key
);
7487 final_key
.push_back('.');
7488 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7489 final_key
.resize(9); // keep prefix
7492 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7493 dout(30) << __func__
<< " have " << pretty_binary_string(final_key
)
7494 << " -> " << *p
<< dendl
;
7497 dout(30) << __func__
<< " miss " << pretty_binary_string(final_key
)
7498 << " -> " << *p
<< dendl
;
7502 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7507 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7508 const coll_t
& cid
, ///< [in] collection
7509 const ghobject_t
&oid
///< [in] object
7512 CollectionHandle c
= _get_collection(cid
);
7514 dout(10) << __func__
<< " " << cid
<< "doesn't exist" <<dendl
;
7515 return ObjectMap::ObjectMapIterator();
7517 return get_omap_iterator(c
, oid
);
7520 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7521 CollectionHandle
&c_
, ///< [in] collection
7522 const ghobject_t
&oid
///< [in] object
7525 Collection
*c
= static_cast<Collection
*>(c_
.get());
7526 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
7528 return ObjectMap::ObjectMapIterator();
7530 RWLock::RLocker
l(c
->lock
);
7531 OnodeRef o
= c
->get_onode(oid
, false);
7532 if (!o
|| !o
->exists
) {
7533 dout(10) << __func__
<< " " << oid
<< "doesn't exist" <<dendl
;
7534 return ObjectMap::ObjectMapIterator();
7537 dout(10) << __func__
<< " has_omap = " << (int)o
->onode
.has_omap() <<dendl
;
7538 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7539 return ObjectMap::ObjectMapIterator(new OmapIteratorImpl(c
, o
, it
));
7542 // -----------------
7545 void BlueStore::_prepare_ondisk_format_super(KeyValueDB::Transaction
& t
)
7547 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7548 << " min_compat_ondisk_format " << min_compat_ondisk_format
7550 assert(ondisk_format
== latest_ondisk_format
);
7553 ::encode(ondisk_format
, bl
);
7554 t
->set(PREFIX_SUPER
, "ondisk_format", bl
);
7558 ::encode(min_compat_ondisk_format
, bl
);
7559 t
->set(PREFIX_SUPER
, "min_compat_ondisk_format", bl
);
7563 int BlueStore::_open_super_meta()
7569 db
->get(PREFIX_SUPER
, "nid_max", &bl
);
7570 bufferlist::iterator p
= bl
.begin();
7575 } catch (buffer::error
& e
) {
7576 derr
<< __func__
<< " unable to read nid_max" << dendl
;
7579 dout(10) << __func__
<< " old nid_max " << nid_max
<< dendl
;
7580 nid_last
= nid_max
.load();
7587 db
->get(PREFIX_SUPER
, "blobid_max", &bl
);
7588 bufferlist::iterator p
= bl
.begin();
7593 } catch (buffer::error
& e
) {
7594 derr
<< __func__
<< " unable to read blobid_max" << dendl
;
7597 dout(10) << __func__
<< " old blobid_max " << blobid_max
<< dendl
;
7598 blobid_last
= blobid_max
.load();
7604 db
->get(PREFIX_SUPER
, "freelist_type", &bl
);
7606 freelist_type
= std::string(bl
.c_str(), bl
.length());
7607 dout(10) << __func__
<< " freelist_type " << freelist_type
<< dendl
;
7609 assert("Not Support extent freelist manager" == 0);
7614 if (cct
->_conf
->bluestore_bluefs
) {
7615 bluefs_extents
.clear();
7617 db
->get(PREFIX_SUPER
, "bluefs_extents", &bl
);
7618 bufferlist::iterator p
= bl
.begin();
7620 ::decode(bluefs_extents
, p
);
7622 catch (buffer::error
& e
) {
7623 derr
<< __func__
<< " unable to read bluefs_extents" << dendl
;
7626 dout(10) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
7627 << std::dec
<< dendl
;
7631 int32_t compat_ondisk_format
= 0;
7634 int r
= db
->get(PREFIX_SUPER
, "ondisk_format", &bl
);
7636 // base case: kraken bluestore is v1 and readable by v1
7637 dout(20) << __func__
<< " missing ondisk_format; assuming kraken"
7640 compat_ondisk_format
= 1;
7642 auto p
= bl
.begin();
7644 ::decode(ondisk_format
, p
);
7645 } catch (buffer::error
& e
) {
7646 derr
<< __func__
<< " unable to read ondisk_format" << dendl
;
7651 r
= db
->get(PREFIX_SUPER
, "min_compat_ondisk_format", &bl
);
7653 auto p
= bl
.begin();
7655 ::decode(compat_ondisk_format
, p
);
7656 } catch (buffer::error
& e
) {
7657 derr
<< __func__
<< " unable to read compat_ondisk_format" << dendl
;
7662 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7663 << " compat_ondisk_format " << compat_ondisk_format
7667 if (latest_ondisk_format
< compat_ondisk_format
) {
7668 derr
<< __func__
<< " compat_ondisk_format is "
7669 << compat_ondisk_format
<< " but we only understand version "
7670 << latest_ondisk_format
<< dendl
;
7673 if (ondisk_format
< latest_ondisk_format
) {
7674 int r
= _upgrade_super();
7682 db
->get(PREFIX_SUPER
, "min_alloc_size", &bl
);
7683 auto p
= bl
.begin();
7687 min_alloc_size
= val
;
7688 min_alloc_size_order
= ctz(val
);
7689 assert(min_alloc_size
== 1u << min_alloc_size_order
);
7690 } catch (buffer::error
& e
) {
7691 derr
<< __func__
<< " unable to read min_alloc_size" << dendl
;
7694 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
7695 << std::dec
<< dendl
;
7699 _set_throttle_params();
7708 int BlueStore::_upgrade_super()
7710 dout(1) << __func__
<< " from " << ondisk_format
<< ", latest "
7711 << latest_ondisk_format
<< dendl
;
7712 assert(ondisk_format
> 0);
7713 assert(ondisk_format
< latest_ondisk_format
);
7715 if (ondisk_format
== 1) {
7717 // - super: added ondisk_format
7718 // - super: added min_readable_ondisk_format
7719 // - super: added min_compat_ondisk_format
7720 // - super: added min_alloc_size
7721 // - super: removed min_min_alloc_size
7722 KeyValueDB::Transaction t
= db
->get_transaction();
7725 db
->get(PREFIX_SUPER
, "min_min_alloc_size", &bl
);
7726 auto p
= bl
.begin();
7730 min_alloc_size
= val
;
7731 } catch (buffer::error
& e
) {
7732 derr
<< __func__
<< " failed to read min_min_alloc_size" << dendl
;
7735 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
7736 t
->rmkey(PREFIX_SUPER
, "min_min_alloc_size");
7739 _prepare_ondisk_format_super(t
);
7740 int r
= db
->submit_transaction_sync(t
);
7745 dout(1) << __func__
<< " done" << dendl
;
7749 void BlueStore::_assign_nid(TransContext
*txc
, OnodeRef o
)
7755 uint64_t nid
= ++nid_last
;
7756 dout(20) << __func__
<< " " << nid
<< dendl
;
7758 txc
->last_nid
= nid
;
7762 uint64_t BlueStore::_assign_blobid(TransContext
*txc
)
7764 uint64_t bid
= ++blobid_last
;
7765 dout(20) << __func__
<< " " << bid
<< dendl
;
7766 txc
->last_blobid
= bid
;
7770 void BlueStore::get_db_statistics(Formatter
*f
)
7772 db
->get_statistics(f
);
7775 BlueStore::TransContext
*BlueStore::_txc_create(OpSequencer
*osr
)
7777 TransContext
*txc
= new TransContext(cct
, osr
);
7778 txc
->t
= db
->get_transaction();
7779 osr
->queue_new(txc
);
7780 dout(20) << __func__
<< " osr " << osr
<< " = " << txc
7781 << " seq " << txc
->seq
<< dendl
;
7785 void BlueStore::_txc_calc_cost(TransContext
*txc
)
7787 // this is about the simplest model for transaction cost you can
7788 // imagine. there is some fixed overhead cost by saying there is a
7789 // minimum of one "io". and then we have some cost per "io" that is
7790 // a configurable (with different hdd and ssd defaults), and add
7791 // that to the bytes value.
7792 int ios
= 1; // one "io" for the kv commit
7793 for (auto& p
: txc
->ioc
.pending_aios
) {
7794 ios
+= p
.iov
.size();
7796 auto cost
= throttle_cost_per_io
.load();
7797 txc
->cost
= ios
* cost
+ txc
->bytes
;
7798 dout(10) << __func__
<< " " << txc
<< " cost " << txc
->cost
<< " ("
7799 << ios
<< " ios * " << cost
<< " + " << txc
->bytes
7800 << " bytes)" << dendl
;
7803 void BlueStore::_txc_update_store_statfs(TransContext
*txc
)
7805 if (txc
->statfs_delta
.is_empty())
7808 logger
->inc(l_bluestore_allocated
, txc
->statfs_delta
.allocated());
7809 logger
->inc(l_bluestore_stored
, txc
->statfs_delta
.stored());
7810 logger
->inc(l_bluestore_compressed
, txc
->statfs_delta
.compressed());
7811 logger
->inc(l_bluestore_compressed_allocated
, txc
->statfs_delta
.compressed_allocated());
7812 logger
->inc(l_bluestore_compressed_original
, txc
->statfs_delta
.compressed_original());
7815 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
7816 vstatfs
+= txc
->statfs_delta
;
7820 txc
->statfs_delta
.encode(bl
);
7822 txc
->t
->merge(PREFIX_STAT
, "bluestore_statfs", bl
);
7823 txc
->statfs_delta
.reset();
7826 void BlueStore::_txc_state_proc(TransContext
*txc
)
7829 dout(10) << __func__
<< " txc " << txc
7830 << " " << txc
->get_state_name() << dendl
;
7831 switch (txc
->state
) {
7832 case TransContext::STATE_PREPARE
:
7833 txc
->log_state_latency(logger
, l_bluestore_state_prepare_lat
);
7834 if (txc
->ioc
.has_pending_aios()) {
7835 txc
->state
= TransContext::STATE_AIO_WAIT
;
7836 txc
->had_ios
= true;
7837 _txc_aio_submit(txc
);
7842 case TransContext::STATE_AIO_WAIT
:
7843 txc
->log_state_latency(logger
, l_bluestore_state_aio_wait_lat
);
7844 _txc_finish_io(txc
); // may trigger blocked txc's too
7847 case TransContext::STATE_IO_DONE
:
7848 //assert(txc->osr->qlock.is_locked()); // see _txc_finish_io
7850 ++txc
->osr
->txc_with_unstable_io
;
7852 txc
->log_state_latency(logger
, l_bluestore_state_io_done_lat
);
7853 txc
->state
= TransContext::STATE_KV_QUEUED
;
7854 if (cct
->_conf
->bluestore_sync_submit_transaction
) {
7855 if (txc
->last_nid
>= nid_max
||
7856 txc
->last_blobid
>= blobid_max
) {
7857 dout(20) << __func__
7858 << " last_{nid,blobid} exceeds max, submit via kv thread"
7860 } else if (txc
->osr
->kv_committing_serially
) {
7861 dout(20) << __func__
<< " prior txc submitted via kv thread, us too"
7863 // note: this is starvation-prone. once we have a txc in a busy
7864 // sequencer that is committing serially it is possible to keep
7865 // submitting new transactions fast enough that we get stuck doing
7866 // so. the alternative is to block here... fixme?
7867 } else if (txc
->osr
->txc_with_unstable_io
) {
7868 dout(20) << __func__
<< " prior txc(s) with unstable ios "
7869 << txc
->osr
->txc_with_unstable_io
.load() << dendl
;
7870 } else if (cct
->_conf
->bluestore_debug_randomize_serial_transaction
&&
7871 rand() % cct
->_conf
->bluestore_debug_randomize_serial_transaction
7873 dout(20) << __func__
<< " DEBUG randomly forcing submit via kv thread"
7876 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
7877 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
7879 _txc_applied_kv(txc
);
7883 std::lock_guard
<std::mutex
> l(kv_lock
);
7884 kv_queue
.push_back(txc
);
7885 kv_cond
.notify_one();
7886 if (txc
->state
!= TransContext::STATE_KV_SUBMITTED
) {
7887 kv_queue_unsubmitted
.push_back(txc
);
7888 ++txc
->osr
->kv_committing_serially
;
7892 kv_throttle_costs
+= txc
->cost
;
7895 case TransContext::STATE_KV_SUBMITTED
:
7896 txc
->log_state_latency(logger
, l_bluestore_state_kv_committing_lat
);
7897 txc
->state
= TransContext::STATE_KV_DONE
;
7898 _txc_committed_kv(txc
);
7901 case TransContext::STATE_KV_DONE
:
7902 txc
->log_state_latency(logger
, l_bluestore_state_kv_done_lat
);
7903 if (txc
->deferred_txn
) {
7904 txc
->state
= TransContext::STATE_DEFERRED_QUEUED
;
7905 _deferred_queue(txc
);
7908 txc
->state
= TransContext::STATE_FINISHING
;
7911 case TransContext::STATE_DEFERRED_CLEANUP
:
7912 txc
->log_state_latency(logger
, l_bluestore_state_deferred_cleanup_lat
);
7913 txc
->state
= TransContext::STATE_FINISHING
;
7916 case TransContext::STATE_FINISHING
:
7917 txc
->log_state_latency(logger
, l_bluestore_state_finishing_lat
);
7922 derr
<< __func__
<< " unexpected txc " << txc
7923 << " state " << txc
->get_state_name() << dendl
;
7924 assert(0 == "unexpected txc state");
7930 void BlueStore::_txc_finish_io(TransContext
*txc
)
7932 dout(20) << __func__
<< " " << txc
<< dendl
;
7935 * we need to preserve the order of kv transactions,
7936 * even though aio will complete in any order.
7939 OpSequencer
*osr
= txc
->osr
.get();
7940 std::lock_guard
<std::mutex
> l(osr
->qlock
);
7941 txc
->state
= TransContext::STATE_IO_DONE
;
7943 // release aio contexts (including pinned buffers).
7944 txc
->ioc
.running_aios
.clear();
7946 OpSequencer::q_list_t::iterator p
= osr
->q
.iterator_to(*txc
);
7947 while (p
!= osr
->q
.begin()) {
7949 if (p
->state
< TransContext::STATE_IO_DONE
) {
7950 dout(20) << __func__
<< " " << txc
<< " blocked by " << &*p
<< " "
7951 << p
->get_state_name() << dendl
;
7954 if (p
->state
> TransContext::STATE_IO_DONE
) {
7960 _txc_state_proc(&*p
++);
7961 } while (p
!= osr
->q
.end() &&
7962 p
->state
== TransContext::STATE_IO_DONE
);
7964 if (osr
->kv_submitted_waiters
&&
7965 osr
->_is_all_kv_submitted()) {
7966 osr
->qcond
.notify_all();
7970 void BlueStore::_txc_write_nodes(TransContext
*txc
, KeyValueDB::Transaction t
)
7972 dout(20) << __func__
<< " txc " << txc
7973 << " onodes " << txc
->onodes
7974 << " shared_blobs " << txc
->shared_blobs
7978 for (auto o
: txc
->onodes
) {
7979 // finalize extent_map shards
7980 o
->extent_map
.update(t
, false);
7981 if (o
->extent_map
.needs_reshard()) {
7982 o
->extent_map
.reshard(db
, t
);
7983 o
->extent_map
.update(t
, true);
7984 if (o
->extent_map
.needs_reshard()) {
7985 dout(20) << __func__
<< " warning: still wants reshard, check options?"
7987 o
->extent_map
.clear_needs_reshard();
7989 logger
->inc(l_bluestore_onode_reshard
);
7994 denc(o
->onode
, bound
);
7995 o
->extent_map
.bound_encode_spanning_blobs(bound
);
7996 if (o
->onode
.extent_map_shards
.empty()) {
7997 denc(o
->extent_map
.inline_bl
, bound
);
8002 unsigned onode_part
, blob_part
, extent_part
;
8004 auto p
= bl
.get_contiguous_appender(bound
, true);
8006 onode_part
= p
.get_logical_offset();
8007 o
->extent_map
.encode_spanning_blobs(p
);
8008 blob_part
= p
.get_logical_offset() - onode_part
;
8009 if (o
->onode
.extent_map_shards
.empty()) {
8010 denc(o
->extent_map
.inline_bl
, p
);
8012 extent_part
= p
.get_logical_offset() - onode_part
- blob_part
;
8015 dout(20) << " onode " << o
->oid
<< " is " << bl
.length()
8016 << " (" << onode_part
<< " bytes onode + "
8017 << blob_part
<< " bytes spanning blobs + "
8018 << extent_part
<< " bytes inline extents)"
8020 t
->set(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size(), bl
);
8021 o
->flushing_count
++;
8024 // objects we modified but didn't affect the onode
8025 auto p
= txc
->modified_objects
.begin();
8026 while (p
!= txc
->modified_objects
.end()) {
8027 if (txc
->onodes
.count(*p
) == 0) {
8028 (*p
)->flushing_count
++;
8031 // remove dups with onodes list to avoid problems in _txc_finish
8032 p
= txc
->modified_objects
.erase(p
);
8036 // finalize shared_blobs
8037 for (auto sb
: txc
->shared_blobs
) {
8039 auto sbid
= sb
->get_sbid();
8040 get_shared_blob_key(sbid
, &key
);
8041 if (sb
->persistent
->empty()) {
8042 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
8043 << " is empty" << dendl
;
8044 t
->rmkey(PREFIX_SHARED_BLOB
, key
);
8047 ::encode(*(sb
->persistent
), bl
);
8048 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
8049 << " is " << bl
.length() << " " << *sb
<< dendl
;
8050 t
->set(PREFIX_SHARED_BLOB
, key
, bl
);
8055 void BlueStore::BSPerfTracker::update_from_perfcounters(
8056 PerfCounters
&logger
)
8058 os_commit_latency
.consume_next(
8060 l_bluestore_commit_lat
));
8061 os_apply_latency
.consume_next(
8063 l_bluestore_commit_lat
));
8066 void BlueStore::_txc_finalize_kv(TransContext
*txc
, KeyValueDB::Transaction t
)
8068 dout(20) << __func__
<< " txc " << txc
<< std::hex
8069 << " allocated 0x" << txc
->allocated
8070 << " released 0x" << txc
->released
8071 << std::dec
<< dendl
;
8073 // We have to handle the case where we allocate *and* deallocate the
8074 // same region in this transaction. The freelist doesn't like that.
8075 // (Actually, the only thing that cares is the BitmapFreelistManager
8076 // debug check. But that's important.)
8077 interval_set
<uint64_t> tmp_allocated
, tmp_released
;
8078 interval_set
<uint64_t> *pallocated
= &txc
->allocated
;
8079 interval_set
<uint64_t> *preleased
= &txc
->released
;
8080 if (!txc
->allocated
.empty() && !txc
->released
.empty()) {
8081 interval_set
<uint64_t> overlap
;
8082 overlap
.intersection_of(txc
->allocated
, txc
->released
);
8083 if (!overlap
.empty()) {
8084 tmp_allocated
= txc
->allocated
;
8085 tmp_allocated
.subtract(overlap
);
8086 tmp_released
= txc
->released
;
8087 tmp_released
.subtract(overlap
);
8088 dout(20) << __func__
<< " overlap 0x" << std::hex
<< overlap
8089 << ", new allocated 0x" << tmp_allocated
8090 << " released 0x" << tmp_released
<< std::dec
8092 pallocated
= &tmp_allocated
;
8093 preleased
= &tmp_released
;
8097 // update freelist with non-overlap sets
8098 for (interval_set
<uint64_t>::iterator p
= pallocated
->begin();
8099 p
!= pallocated
->end();
8101 fm
->allocate(p
.get_start(), p
.get_len(), t
);
8103 for (interval_set
<uint64_t>::iterator p
= preleased
->begin();
8104 p
!= preleased
->end();
8106 dout(20) << __func__
<< " release 0x" << std::hex
<< p
.get_start()
8107 << "~" << p
.get_len() << std::dec
<< dendl
;
8108 fm
->release(p
.get_start(), p
.get_len(), t
);
8111 _txc_update_store_statfs(txc
);
8114 void BlueStore::_txc_applied_kv(TransContext
*txc
)
8116 for (auto ls
: { &txc
->onodes
, &txc
->modified_objects
}) {
8117 for (auto& o
: *ls
) {
8118 dout(20) << __func__
<< " onode " << o
<< " had " << o
->flushing_count
8120 if (--o
->flushing_count
== 0) {
8121 std::lock_guard
<std::mutex
> l(o
->flush_lock
);
8122 o
->flush_cond
.notify_all();
8128 void BlueStore::_txc_committed_kv(TransContext
*txc
)
8130 dout(20) << __func__
<< " txc " << txc
<< dendl
;
8132 // warning: we're calling onreadable_sync inside the sequencer lock
8133 if (txc
->onreadable_sync
) {
8134 txc
->onreadable_sync
->complete(0);
8135 txc
->onreadable_sync
= NULL
;
8137 unsigned n
= txc
->osr
->parent
->shard_hint
.hash_to_shard(m_finisher_num
);
8138 if (txc
->oncommit
) {
8139 logger
->tinc(l_bluestore_commit_lat
, ceph_clock_now() - txc
->start
);
8140 finishers
[n
]->queue(txc
->oncommit
);
8141 txc
->oncommit
= NULL
;
8143 if (txc
->onreadable
) {
8144 finishers
[n
]->queue(txc
->onreadable
);
8145 txc
->onreadable
= NULL
;
8148 if (!txc
->oncommits
.empty()) {
8149 finishers
[n
]->queue(txc
->oncommits
);
8153 void BlueStore::_txc_finish(TransContext
*txc
)
8155 dout(20) << __func__
<< " " << txc
<< " onodes " << txc
->onodes
<< dendl
;
8156 assert(txc
->state
== TransContext::STATE_FINISHING
);
8158 for (auto& sb
: txc
->shared_blobs_written
) {
8159 sb
->bc
.finish_write(sb
->get_cache(), txc
->seq
);
8161 txc
->shared_blobs_written
.clear();
8163 while (!txc
->removed_collections
.empty()) {
8164 _queue_reap_collection(txc
->removed_collections
.front());
8165 txc
->removed_collections
.pop_front();
8168 OpSequencerRef osr
= txc
->osr
;
8170 bool submit_deferred
= false;
8171 OpSequencer::q_list_t releasing_txc
;
8173 std::lock_guard
<std::mutex
> l(osr
->qlock
);
8174 txc
->state
= TransContext::STATE_DONE
;
8175 bool notify
= false;
8176 while (!osr
->q
.empty()) {
8177 TransContext
*txc
= &osr
->q
.front();
8178 dout(20) << __func__
<< " txc " << txc
<< " " << txc
->get_state_name()
8180 if (txc
->state
!= TransContext::STATE_DONE
) {
8181 if (txc
->state
== TransContext::STATE_PREPARE
&&
8182 deferred_aggressive
) {
8183 // for _osr_drain_preceding()
8186 if (txc
->state
== TransContext::STATE_DEFERRED_QUEUED
&&
8187 osr
->q
.size() > g_conf
->bluestore_max_deferred_txc
) {
8188 submit_deferred
= true;
8194 releasing_txc
.push_back(*txc
);
8198 osr
->qcond
.notify_all();
8200 if (osr
->q
.empty()) {
8201 dout(20) << __func__
<< " osr " << osr
<< " q now empty" << dendl
;
8205 while (!releasing_txc
.empty()) {
8206 // release to allocator only after all preceding txc's have also
8207 // finished any deferred writes that potentially land in these
8209 auto txc
= &releasing_txc
.front();
8210 _txc_release_alloc(txc
);
8211 releasing_txc
.pop_front();
8212 txc
->log_state_latency(logger
, l_bluestore_state_done_lat
);
8216 if (submit_deferred
) {
8217 // we're pinning memory; flush! we could be more fine-grained here but
8218 // i'm not sure it's worth the bother.
8219 deferred_try_submit();
8222 if (empty
&& osr
->zombie
) {
8223 dout(10) << __func__
<< " reaping empty zombie osr " << osr
<< dendl
;
8228 void BlueStore::_txc_release_alloc(TransContext
*txc
)
8230 // update allocator with full released set
8231 if (!cct
->_conf
->bluestore_debug_no_reuse_blocks
) {
8232 dout(10) << __func__
<< " " << txc
<< " " << txc
->released
<< dendl
;
8233 for (interval_set
<uint64_t>::iterator p
= txc
->released
.begin();
8234 p
!= txc
->released
.end();
8236 alloc
->release(p
.get_start(), p
.get_len());
8240 txc
->allocated
.clear();
8241 txc
->released
.clear();
8244 void BlueStore::_osr_drain_preceding(TransContext
*txc
)
8246 OpSequencer
*osr
= txc
->osr
.get();
8247 dout(10) << __func__
<< " " << txc
<< " osr " << osr
<< dendl
;
8248 ++deferred_aggressive
; // FIXME: maybe osr-local aggressive flag?
8250 // submit anything pending
8251 deferred_lock
.lock();
8252 if (osr
->deferred_pending
) {
8253 _deferred_submit_unlock(osr
);
8255 deferred_lock
.unlock();
8259 // wake up any previously finished deferred events
8260 std::lock_guard
<std::mutex
> l(kv_lock
);
8261 kv_cond
.notify_one();
8263 osr
->drain_preceding(txc
);
8264 --deferred_aggressive
;
8265 dout(10) << __func__
<< " " << osr
<< " done" << dendl
;
8268 void BlueStore::_osr_drain_all()
8270 dout(10) << __func__
<< dendl
;
8272 set
<OpSequencerRef
> s
;
8274 std::lock_guard
<std::mutex
> l(osr_lock
);
8277 dout(20) << __func__
<< " osr_set " << s
<< dendl
;
8279 ++deferred_aggressive
;
8281 // submit anything pending
8282 deferred_try_submit();
8285 // wake up any previously finished deferred events
8286 std::lock_guard
<std::mutex
> l(kv_lock
);
8287 kv_cond
.notify_one();
8290 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8291 kv_finalize_cond
.notify_one();
8293 for (auto osr
: s
) {
8294 dout(20) << __func__
<< " drain " << osr
<< dendl
;
8297 --deferred_aggressive
;
8299 dout(10) << __func__
<< " done" << dendl
;
8302 void BlueStore::_osr_unregister_all()
8304 set
<OpSequencerRef
> s
;
8306 std::lock_guard
<std::mutex
> l(osr_lock
);
8309 dout(10) << __func__
<< " " << s
<< dendl
;
8310 for (auto osr
: s
) {
8314 // break link from Sequencer to us so that this OpSequencer
8315 // instance can die with this mount/umount cycle. note that
8316 // we assume umount() will not race against ~Sequencer.
8317 assert(osr
->parent
);
8318 osr
->parent
->p
.reset();
8321 // nobody should be creating sequencers during umount either.
8323 std::lock_guard
<std::mutex
> l(osr_lock
);
8324 assert(osr_set
.empty());
8328 void BlueStore::_kv_start()
8330 dout(10) << __func__
<< dendl
;
8332 if (cct
->_conf
->bluestore_shard_finishers
) {
8333 if (cct
->_conf
->osd_op_num_shards
) {
8334 m_finisher_num
= cct
->_conf
->osd_op_num_shards
;
8337 if (bdev
->is_rotational()) {
8338 m_finisher_num
= cct
->_conf
->osd_op_num_shards_hdd
;
8340 m_finisher_num
= cct
->_conf
->osd_op_num_shards_ssd
;
8345 assert(m_finisher_num
!= 0);
8347 for (int i
= 0; i
< m_finisher_num
; ++i
) {
8349 oss
<< "finisher-" << i
;
8350 Finisher
*f
= new Finisher(cct
, oss
.str(), "finisher");
8351 finishers
.push_back(f
);
8354 deferred_finisher
.start();
8355 for (auto f
: finishers
) {
8358 kv_sync_thread
.create("bstore_kv_sync");
8359 kv_finalize_thread
.create("bstore_kv_final");
8362 void BlueStore::_kv_stop()
8364 dout(10) << __func__
<< dendl
;
8366 std::unique_lock
<std::mutex
> l(kv_lock
);
8367 while (!kv_sync_started
) {
8371 kv_cond
.notify_all();
8374 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8375 while (!kv_finalize_started
) {
8376 kv_finalize_cond
.wait(l
);
8378 kv_finalize_stop
= true;
8379 kv_finalize_cond
.notify_all();
8381 kv_sync_thread
.join();
8382 kv_finalize_thread
.join();
8384 std::lock_guard
<std::mutex
> l(kv_lock
);
8388 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8389 kv_finalize_stop
= false;
8391 dout(10) << __func__
<< " stopping finishers" << dendl
;
8392 deferred_finisher
.wait_for_empty();
8393 deferred_finisher
.stop();
8394 for (auto f
: finishers
) {
8395 f
->wait_for_empty();
8398 dout(10) << __func__
<< " stopped" << dendl
;
8401 void BlueStore::_kv_sync_thread()
8403 dout(10) << __func__
<< " start" << dendl
;
8404 std::unique_lock
<std::mutex
> l(kv_lock
);
8405 assert(!kv_sync_started
);
8406 kv_sync_started
= true;
8407 kv_cond
.notify_all();
8409 assert(kv_committing
.empty());
8410 if (kv_queue
.empty() &&
8411 ((deferred_done_queue
.empty() && deferred_stable_queue
.empty()) ||
8412 !deferred_aggressive
)) {
8415 dout(20) << __func__
<< " sleep" << dendl
;
8417 dout(20) << __func__
<< " wake" << dendl
;
8419 deque
<TransContext
*> kv_submitting
;
8420 deque
<DeferredBatch
*> deferred_done
, deferred_stable
;
8421 uint64_t aios
= 0, costs
= 0;
8423 dout(20) << __func__
<< " committing " << kv_queue
.size()
8424 << " submitting " << kv_queue_unsubmitted
.size()
8425 << " deferred done " << deferred_done_queue
.size()
8426 << " stable " << deferred_stable_queue
.size()
8428 kv_committing
.swap(kv_queue
);
8429 kv_submitting
.swap(kv_queue_unsubmitted
);
8430 deferred_done
.swap(deferred_done_queue
);
8431 deferred_stable
.swap(deferred_stable_queue
);
8433 costs
= kv_throttle_costs
;
8435 kv_throttle_costs
= 0;
8436 utime_t start
= ceph_clock_now();
8439 dout(30) << __func__
<< " committing " << kv_committing
<< dendl
;
8440 dout(30) << __func__
<< " submitting " << kv_submitting
<< dendl
;
8441 dout(30) << __func__
<< " deferred_done " << deferred_done
<< dendl
;
8442 dout(30) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8444 bool force_flush
= false;
8445 // if bluefs is sharing the same device as data (only), then we
8446 // can rely on the bluefs commit to flush the device and make
8447 // deferred aios stable. that means that if we do have done deferred
8448 // txcs AND we are not on a single device, we need to force a flush.
8449 if (bluefs_single_shared_device
&& bluefs
) {
8452 } else if (kv_committing
.empty() && kv_submitting
.empty() &&
8453 deferred_stable
.empty()) {
8454 force_flush
= true; // there's nothing else to commit!
8455 } else if (deferred_aggressive
) {
8462 dout(20) << __func__
<< " num_aios=" << aios
8463 << " force_flush=" << (int)force_flush
8464 << ", flushing, deferred done->stable" << dendl
;
8465 // flush/barrier on block device
8468 // if we flush then deferred done are now deferred stable
8469 deferred_stable
.insert(deferred_stable
.end(), deferred_done
.begin(),
8470 deferred_done
.end());
8471 deferred_done
.clear();
8473 utime_t after_flush
= ceph_clock_now();
8475 // we will use one final transaction to force a sync
8476 KeyValueDB::Transaction synct
= db
->get_transaction();
8478 // increase {nid,blobid}_max? note that this covers both the
8479 // case where we are approaching the max and the case we passed
8480 // it. in either case, we increase the max in the earlier txn
8482 uint64_t new_nid_max
= 0, new_blobid_max
= 0;
8483 if (nid_last
+ cct
->_conf
->bluestore_nid_prealloc
/2 > nid_max
) {
8484 KeyValueDB::Transaction t
=
8485 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8486 new_nid_max
= nid_last
+ cct
->_conf
->bluestore_nid_prealloc
;
8488 ::encode(new_nid_max
, bl
);
8489 t
->set(PREFIX_SUPER
, "nid_max", bl
);
8490 dout(10) << __func__
<< " new_nid_max " << new_nid_max
<< dendl
;
8492 if (blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
/2 > blobid_max
) {
8493 KeyValueDB::Transaction t
=
8494 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8495 new_blobid_max
= blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
;
8497 ::encode(new_blobid_max
, bl
);
8498 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
8499 dout(10) << __func__
<< " new_blobid_max " << new_blobid_max
<< dendl
;
8502 for (auto txc
: kv_committing
) {
8503 if (txc
->state
== TransContext::STATE_KV_QUEUED
) {
8504 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8505 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
8507 _txc_applied_kv(txc
);
8508 --txc
->osr
->kv_committing_serially
;
8509 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
8510 if (txc
->osr
->kv_submitted_waiters
) {
8511 std::lock_guard
<std::mutex
> l(txc
->osr
->qlock
);
8512 if (txc
->osr
->_is_all_kv_submitted()) {
8513 txc
->osr
->qcond
.notify_all();
8518 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8519 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8522 --txc
->osr
->txc_with_unstable_io
;
8526 // release throttle *before* we commit. this allows new ops
8527 // to be prepared and enter pipeline while we are waiting on
8528 // the kv commit sync/flush. then hopefully on the next
8529 // iteration there will already be ops awake. otherwise, we
8530 // end up going to sleep, and then wake up when the very first
8531 // transaction is ready for commit.
8532 throttle_bytes
.put(costs
);
8534 PExtentVector bluefs_gift_extents
;
8536 after_flush
- bluefs_last_balance
>
8537 cct
->_conf
->bluestore_bluefs_balance_interval
) {
8538 bluefs_last_balance
= after_flush
;
8539 int r
= _balance_bluefs_freespace(&bluefs_gift_extents
);
8542 for (auto& p
: bluefs_gift_extents
) {
8543 bluefs_extents
.insert(p
.offset
, p
.length
);
8546 ::encode(bluefs_extents
, bl
);
8547 dout(10) << __func__
<< " bluefs_extents now 0x" << std::hex
8548 << bluefs_extents
<< std::dec
<< dendl
;
8549 synct
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
8553 // cleanup sync deferred keys
8554 for (auto b
: deferred_stable
) {
8555 for (auto& txc
: b
->txcs
) {
8556 bluestore_deferred_transaction_t
& wt
= *txc
.deferred_txn
;
8557 if (!wt
.released
.empty()) {
8558 // kraken replay compat only
8559 txc
.released
= wt
.released
;
8560 dout(10) << __func__
<< " deferred txn has released "
8562 << " (we just upgraded from kraken) on " << &txc
<< dendl
;
8563 _txc_finalize_kv(&txc
, synct
);
8565 // cleanup the deferred
8567 get_deferred_key(wt
.seq
, &key
);
8568 synct
->rm_single_key(PREFIX_DEFERRED
, key
);
8572 // submit synct synchronously (block and wait for it to commit)
8573 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction_sync(synct
);
8577 nid_max
= new_nid_max
;
8578 dout(10) << __func__
<< " nid_max now " << nid_max
<< dendl
;
8580 if (new_blobid_max
) {
8581 blobid_max
= new_blobid_max
;
8582 dout(10) << __func__
<< " blobid_max now " << blobid_max
<< dendl
;
8586 utime_t finish
= ceph_clock_now();
8587 utime_t dur_flush
= after_flush
- start
;
8588 utime_t dur_kv
= finish
- after_flush
;
8589 utime_t dur
= finish
- start
;
8590 dout(20) << __func__
<< " committed " << kv_committing
.size()
8591 << " cleaned " << deferred_stable
.size()
8593 << " (" << dur_flush
<< " flush + " << dur_kv
<< " kv commit)"
8595 logger
->tinc(l_bluestore_kv_flush_lat
, dur_flush
);
8596 logger
->tinc(l_bluestore_kv_commit_lat
, dur_kv
);
8597 logger
->tinc(l_bluestore_kv_lat
, dur
);
8601 if (!bluefs_gift_extents
.empty()) {
8602 _commit_bluefs_freespace(bluefs_gift_extents
);
8604 for (auto p
= bluefs_extents_reclaiming
.begin();
8605 p
!= bluefs_extents_reclaiming
.end();
8607 dout(20) << __func__
<< " releasing old bluefs 0x" << std::hex
8608 << p
.get_start() << "~" << p
.get_len() << std::dec
8610 alloc
->release(p
.get_start(), p
.get_len());
8612 bluefs_extents_reclaiming
.clear();
8616 std::unique_lock
<std::mutex
> m(kv_finalize_lock
);
8617 if (kv_committing_to_finalize
.empty()) {
8618 kv_committing_to_finalize
.swap(kv_committing
);
8620 kv_committing_to_finalize
.insert(
8621 kv_committing_to_finalize
.end(),
8622 kv_committing
.begin(),
8623 kv_committing
.end());
8624 kv_committing
.clear();
8626 if (deferred_stable_to_finalize
.empty()) {
8627 deferred_stable_to_finalize
.swap(deferred_stable
);
8629 deferred_stable_to_finalize
.insert(
8630 deferred_stable_to_finalize
.end(),
8631 deferred_stable
.begin(),
8632 deferred_stable
.end());
8633 deferred_stable
.clear();
8635 kv_finalize_cond
.notify_one();
8639 // previously deferred "done" are now "stable" by virtue of this
8641 deferred_stable_queue
.swap(deferred_done
);
8644 dout(10) << __func__
<< " finish" << dendl
;
8645 kv_sync_started
= false;
8648 void BlueStore::_kv_finalize_thread()
8650 deque
<TransContext
*> kv_committed
;
8651 deque
<DeferredBatch
*> deferred_stable
;
8652 dout(10) << __func__
<< " start" << dendl
;
8653 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8654 assert(!kv_finalize_started
);
8655 kv_finalize_started
= true;
8656 kv_finalize_cond
.notify_all();
8658 assert(kv_committed
.empty());
8659 assert(deferred_stable
.empty());
8660 if (kv_committing_to_finalize
.empty() &&
8661 deferred_stable_to_finalize
.empty()) {
8662 if (kv_finalize_stop
)
8664 dout(20) << __func__
<< " sleep" << dendl
;
8665 kv_finalize_cond
.wait(l
);
8666 dout(20) << __func__
<< " wake" << dendl
;
8668 kv_committed
.swap(kv_committing_to_finalize
);
8669 deferred_stable
.swap(deferred_stable_to_finalize
);
8671 dout(20) << __func__
<< " kv_committed " << kv_committed
<< dendl
;
8672 dout(20) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8674 while (!kv_committed
.empty()) {
8675 TransContext
*txc
= kv_committed
.front();
8676 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8677 _txc_state_proc(txc
);
8678 kv_committed
.pop_front();
8681 for (auto b
: deferred_stable
) {
8682 auto p
= b
->txcs
.begin();
8683 while (p
!= b
->txcs
.end()) {
8684 TransContext
*txc
= &*p
;
8685 p
= b
->txcs
.erase(p
); // unlink here because
8686 _txc_state_proc(txc
); // this may destroy txc
8690 deferred_stable
.clear();
8692 if (!deferred_aggressive
) {
8693 if (deferred_queue_size
>= deferred_batch_ops
.load() ||
8694 throttle_deferred_bytes
.past_midpoint()) {
8695 deferred_try_submit();
8699 // this is as good a place as any ...
8700 _reap_collections();
8705 dout(10) << __func__
<< " finish" << dendl
;
8706 kv_finalize_started
= false;
8709 bluestore_deferred_op_t
*BlueStore::_get_deferred_op(
8710 TransContext
*txc
, OnodeRef o
)
8712 if (!txc
->deferred_txn
) {
8713 txc
->deferred_txn
= new bluestore_deferred_transaction_t
;
8715 txc
->deferred_txn
->ops
.push_back(bluestore_deferred_op_t());
8716 return &txc
->deferred_txn
->ops
.back();
8719 void BlueStore::_deferred_queue(TransContext
*txc
)
8721 dout(20) << __func__
<< " txc " << txc
<< " osr " << txc
->osr
<< dendl
;
8722 deferred_lock
.lock();
8723 if (!txc
->osr
->deferred_pending
&&
8724 !txc
->osr
->deferred_running
) {
8725 deferred_queue
.push_back(*txc
->osr
);
8727 if (!txc
->osr
->deferred_pending
) {
8728 txc
->osr
->deferred_pending
= new DeferredBatch(cct
, txc
->osr
.get());
8730 ++deferred_queue_size
;
8731 txc
->osr
->deferred_pending
->txcs
.push_back(*txc
);
8732 bluestore_deferred_transaction_t
& wt
= *txc
->deferred_txn
;
8733 for (auto opi
= wt
.ops
.begin(); opi
!= wt
.ops
.end(); ++opi
) {
8734 const auto& op
= *opi
;
8735 assert(op
.op
== bluestore_deferred_op_t::OP_WRITE
);
8736 bufferlist::const_iterator p
= op
.data
.begin();
8737 for (auto e
: op
.extents
) {
8738 txc
->osr
->deferred_pending
->prepare_write(
8739 cct
, wt
.seq
, e
.offset
, e
.length
, p
);
8742 if (deferred_aggressive
&&
8743 !txc
->osr
->deferred_running
) {
8744 _deferred_submit_unlock(txc
->osr
.get());
8746 deferred_lock
.unlock();
8750 void BlueStore::deferred_try_submit()
8752 dout(20) << __func__
<< " " << deferred_queue
.size() << " osrs, "
8753 << deferred_queue_size
<< " txcs" << dendl
;
8754 std::lock_guard
<std::mutex
> l(deferred_lock
);
8755 vector
<OpSequencerRef
> osrs
;
8756 osrs
.reserve(deferred_queue
.size());
8757 for (auto& osr
: deferred_queue
) {
8758 osrs
.push_back(&osr
);
8760 for (auto& osr
: osrs
) {
8761 if (osr
->deferred_pending
) {
8762 if (!osr
->deferred_running
) {
8763 _deferred_submit_unlock(osr
.get());
8764 deferred_lock
.lock();
8766 dout(20) << __func__
<< " osr " << osr
<< " already has running"
8770 dout(20) << __func__
<< " osr " << osr
<< " has no pending" << dendl
;
8775 void BlueStore::_deferred_submit_unlock(OpSequencer
*osr
)
8777 dout(10) << __func__
<< " osr " << osr
8778 << " " << osr
->deferred_pending
->iomap
.size() << " ios pending "
8780 assert(osr
->deferred_pending
);
8781 assert(!osr
->deferred_running
);
8783 auto b
= osr
->deferred_pending
;
8784 deferred_queue_size
-= b
->seq_bytes
.size();
8785 assert(deferred_queue_size
>= 0);
8787 osr
->deferred_running
= osr
->deferred_pending
;
8788 osr
->deferred_pending
= nullptr;
8790 uint64_t start
= 0, pos
= 0;
8792 auto i
= b
->iomap
.begin();
8794 if (i
== b
->iomap
.end() || i
->first
!= pos
) {
8796 dout(20) << __func__
<< " write 0x" << std::hex
8797 << start
<< "~" << bl
.length()
8798 << " crc " << bl
.crc32c(-1) << std::dec
<< dendl
;
8799 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
8800 logger
->inc(l_bluestore_deferred_write_ops
);
8801 logger
->inc(l_bluestore_deferred_write_bytes
, bl
.length());
8802 int r
= bdev
->aio_write(start
, bl
, &b
->ioc
, false);
8806 if (i
== b
->iomap
.end()) {
8813 dout(20) << __func__
<< " seq " << i
->second
.seq
<< " 0x"
8814 << std::hex
<< pos
<< "~" << i
->second
.bl
.length() << std::dec
8819 pos
+= i
->second
.bl
.length();
8820 bl
.claim_append(i
->second
.bl
);
8824 deferred_lock
.unlock();
8825 bdev
->aio_submit(&b
->ioc
);
8828 struct C_DeferredTrySubmit
: public Context
{
8830 C_DeferredTrySubmit(BlueStore
*s
) : store(s
) {}
8831 void finish(int r
) {
8832 store
->deferred_try_submit();
8836 void BlueStore::_deferred_aio_finish(OpSequencer
*osr
)
8838 dout(10) << __func__
<< " osr " << osr
<< dendl
;
8839 assert(osr
->deferred_running
);
8840 DeferredBatch
*b
= osr
->deferred_running
;
8843 std::lock_guard
<std::mutex
> l(deferred_lock
);
8844 assert(osr
->deferred_running
== b
);
8845 osr
->deferred_running
= nullptr;
8846 if (!osr
->deferred_pending
) {
8847 dout(20) << __func__
<< " dequeueing" << dendl
;
8848 auto q
= deferred_queue
.iterator_to(*osr
);
8849 deferred_queue
.erase(q
);
8850 } else if (deferred_aggressive
) {
8851 dout(20) << __func__
<< " queuing async deferred_try_submit" << dendl
;
8852 deferred_finisher
.queue(new C_DeferredTrySubmit(this));
8854 dout(20) << __func__
<< " leaving queued, more pending" << dendl
;
8860 std::lock_guard
<std::mutex
> l2(osr
->qlock
);
8861 for (auto& i
: b
->txcs
) {
8862 TransContext
*txc
= &i
;
8863 txc
->state
= TransContext::STATE_DEFERRED_CLEANUP
;
8866 osr
->qcond
.notify_all();
8867 throttle_deferred_bytes
.put(costs
);
8868 std::lock_guard
<std::mutex
> l(kv_lock
);
8869 deferred_done_queue
.emplace_back(b
);
8872 // in the normal case, do not bother waking up the kv thread; it will
8873 // catch us on the next commit anyway.
8874 if (deferred_aggressive
) {
8875 std::lock_guard
<std::mutex
> l(kv_lock
);
8876 kv_cond
.notify_one();
8880 int BlueStore::_deferred_replay()
8882 dout(10) << __func__
<< " start" << dendl
;
8883 OpSequencerRef osr
= new OpSequencer(cct
, this);
8886 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_DEFERRED
);
8887 for (it
->lower_bound(string()); it
->valid(); it
->next(), ++count
) {
8888 dout(20) << __func__
<< " replay " << pretty_binary_string(it
->key())
8890 bluestore_deferred_transaction_t
*deferred_txn
=
8891 new bluestore_deferred_transaction_t
;
8892 bufferlist bl
= it
->value();
8893 bufferlist::iterator p
= bl
.begin();
8895 ::decode(*deferred_txn
, p
);
8896 } catch (buffer::error
& e
) {
8897 derr
<< __func__
<< " failed to decode deferred txn "
8898 << pretty_binary_string(it
->key()) << dendl
;
8899 delete deferred_txn
;
8903 TransContext
*txc
= _txc_create(osr
.get());
8904 txc
->deferred_txn
= deferred_txn
;
8905 txc
->state
= TransContext::STATE_KV_DONE
;
8906 _txc_state_proc(txc
);
8909 dout(20) << __func__
<< " draining osr" << dendl
;
8912 dout(10) << __func__
<< " completed " << count
<< " events" << dendl
;
8916 // ---------------------------
8919 int BlueStore::queue_transactions(
8921 vector
<Transaction
>& tls
,
8923 ThreadPool::TPHandle
*handle
)
8926 Context
*onreadable
;
8928 Context
*onreadable_sync
;
8929 ObjectStore::Transaction::collect_contexts(
8930 tls
, &onreadable
, &ondisk
, &onreadable_sync
);
8932 if (cct
->_conf
->objectstore_blackhole
) {
8933 dout(0) << __func__
<< " objectstore_blackhole = TRUE, dropping transaction"
8937 delete onreadable_sync
;
8940 utime_t start
= ceph_clock_now();
8941 // set up the sequencer
8945 osr
= static_cast<OpSequencer
*>(posr
->p
.get());
8946 dout(10) << __func__
<< " existing " << osr
<< " " << *osr
<< dendl
;
8948 osr
= new OpSequencer(cct
, this);
8951 dout(10) << __func__
<< " new " << osr
<< " " << *osr
<< dendl
;
8955 TransContext
*txc
= _txc_create(osr
);
8956 txc
->onreadable
= onreadable
;
8957 txc
->onreadable_sync
= onreadable_sync
;
8958 txc
->oncommit
= ondisk
;
8960 for (vector
<Transaction
>::iterator p
= tls
.begin(); p
!= tls
.end(); ++p
) {
8962 txc
->bytes
+= (*p
).get_num_bytes();
8963 _txc_add_transaction(txc
, &(*p
));
8965 _txc_calc_cost(txc
);
8967 _txc_write_nodes(txc
, txc
->t
);
8969 // journal deferred items
8970 if (txc
->deferred_txn
) {
8971 txc
->deferred_txn
->seq
= ++deferred_seq
;
8973 ::encode(*txc
->deferred_txn
, bl
);
8975 get_deferred_key(txc
->deferred_txn
->seq
, &key
);
8976 txc
->t
->set(PREFIX_DEFERRED
, key
, bl
);
8979 _txc_finalize_kv(txc
, txc
->t
);
8981 handle
->suspend_tp_timeout();
8983 utime_t tstart
= ceph_clock_now();
8984 throttle_bytes
.get(txc
->cost
);
8985 if (txc
->deferred_txn
) {
8986 // ensure we do not block here because of deferred writes
8987 if (!throttle_deferred_bytes
.get_or_fail(txc
->cost
)) {
8988 dout(10) << __func__
<< " failed get throttle_deferred_bytes, aggressive"
8990 ++deferred_aggressive
;
8991 deferred_try_submit();
8993 // wake up any previously finished deferred events
8994 std::lock_guard
<std::mutex
> l(kv_lock
);
8995 kv_cond
.notify_one();
8997 throttle_deferred_bytes
.get(txc
->cost
);
8998 --deferred_aggressive
;
9001 utime_t tend
= ceph_clock_now();
9004 handle
->reset_tp_timeout();
9006 logger
->inc(l_bluestore_txc
);
9009 _txc_state_proc(txc
);
9011 logger
->tinc(l_bluestore_submit_lat
, ceph_clock_now() - start
);
9012 logger
->tinc(l_bluestore_throttle_lat
, tend
- tstart
);
9016 void BlueStore::_txc_aio_submit(TransContext
*txc
)
9018 dout(10) << __func__
<< " txc " << txc
<< dendl
;
9019 bdev
->aio_submit(&txc
->ioc
);
9022 void BlueStore::_txc_add_transaction(TransContext
*txc
, Transaction
*t
)
9024 Transaction::iterator i
= t
->begin();
9026 _dump_transaction(t
);
9028 vector
<CollectionRef
> cvec(i
.colls
.size());
9030 for (vector
<coll_t
>::iterator p
= i
.colls
.begin(); p
!= i
.colls
.end();
9032 cvec
[j
] = _get_collection(*p
);
9034 vector
<OnodeRef
> ovec(i
.objects
.size());
9036 for (int pos
= 0; i
.have_op(); ++pos
) {
9037 Transaction::Op
*op
= i
.decode_op();
9041 if (op
->op
== Transaction::OP_NOP
)
9044 // collection operations
9045 CollectionRef
&c
= cvec
[op
->cid
];
9047 case Transaction::OP_RMCOLL
:
9049 const coll_t
&cid
= i
.get_cid(op
->cid
);
9050 r
= _remove_collection(txc
, cid
, &c
);
9056 case Transaction::OP_MKCOLL
:
9059 const coll_t
&cid
= i
.get_cid(op
->cid
);
9060 r
= _create_collection(txc
, cid
, op
->split_bits
, &c
);
9066 case Transaction::OP_SPLIT_COLLECTION
:
9067 assert(0 == "deprecated");
9070 case Transaction::OP_SPLIT_COLLECTION2
:
9072 uint32_t bits
= op
->split_bits
;
9073 uint32_t rem
= op
->split_rem
;
9074 r
= _split_collection(txc
, c
, cvec
[op
->dest_cid
], bits
, rem
);
9080 case Transaction::OP_COLL_HINT
:
9082 uint32_t type
= op
->hint_type
;
9085 bufferlist::iterator hiter
= hint
.begin();
9086 if (type
== Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS
) {
9089 ::decode(pg_num
, hiter
);
9090 ::decode(num_objs
, hiter
);
9091 dout(10) << __func__
<< " collection hint objects is a no-op, "
9092 << " pg_num " << pg_num
<< " num_objects " << num_objs
9096 dout(10) << __func__
<< " unknown collection hint " << type
<< dendl
;
9102 case Transaction::OP_COLL_SETATTR
:
9106 case Transaction::OP_COLL_RMATTR
:
9110 case Transaction::OP_COLL_RENAME
:
9111 assert(0 == "not implemented");
9115 derr
<< __func__
<< " error " << cpp_strerror(r
)
9116 << " not handled on operation " << op
->op
9117 << " (op " << pos
<< ", counting from 0)" << dendl
;
9118 _dump_transaction(t
, 0);
9119 assert(0 == "unexpected error");
9122 // these operations implicity create the object
9123 bool create
= false;
9124 if (op
->op
== Transaction::OP_TOUCH
||
9125 op
->op
== Transaction::OP_WRITE
||
9126 op
->op
== Transaction::OP_ZERO
) {
9130 // object operations
9131 RWLock::WLocker
l(c
->lock
);
9132 OnodeRef
&o
= ovec
[op
->oid
];
9134 ghobject_t oid
= i
.get_oid(op
->oid
);
9135 o
= c
->get_onode(oid
, create
);
9137 if (!create
&& (!o
|| !o
->exists
)) {
9138 dout(10) << __func__
<< " op " << op
->op
<< " got ENOENT on "
9139 << i
.get_oid(op
->oid
) << dendl
;
9145 case Transaction::OP_TOUCH
:
9146 r
= _touch(txc
, c
, o
);
9149 case Transaction::OP_WRITE
:
9151 uint64_t off
= op
->off
;
9152 uint64_t len
= op
->len
;
9153 uint32_t fadvise_flags
= i
.get_fadvise_flags();
9156 r
= _write(txc
, c
, o
, off
, len
, bl
, fadvise_flags
);
9160 case Transaction::OP_ZERO
:
9162 uint64_t off
= op
->off
;
9163 uint64_t len
= op
->len
;
9164 r
= _zero(txc
, c
, o
, off
, len
);
9168 case Transaction::OP_TRIMCACHE
:
9170 // deprecated, no-op
9174 case Transaction::OP_TRUNCATE
:
9176 uint64_t off
= op
->off
;
9177 r
= _truncate(txc
, c
, o
, off
);
9181 case Transaction::OP_REMOVE
:
9183 r
= _remove(txc
, c
, o
);
9187 case Transaction::OP_SETATTR
:
9189 string name
= i
.decode_string();
9192 r
= _setattr(txc
, c
, o
, name
, bp
);
9196 case Transaction::OP_SETATTRS
:
9198 map
<string
, bufferptr
> aset
;
9199 i
.decode_attrset(aset
);
9200 r
= _setattrs(txc
, c
, o
, aset
);
9204 case Transaction::OP_RMATTR
:
9206 string name
= i
.decode_string();
9207 r
= _rmattr(txc
, c
, o
, name
);
9211 case Transaction::OP_RMATTRS
:
9213 r
= _rmattrs(txc
, c
, o
);
9217 case Transaction::OP_CLONE
:
9219 OnodeRef
& no
= ovec
[op
->dest_oid
];
9221 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9222 no
= c
->get_onode(noid
, true);
9224 r
= _clone(txc
, c
, o
, no
);
9228 case Transaction::OP_CLONERANGE
:
9229 assert(0 == "deprecated");
9232 case Transaction::OP_CLONERANGE2
:
9234 OnodeRef
& no
= ovec
[op
->dest_oid
];
9236 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9237 no
= c
->get_onode(noid
, true);
9239 uint64_t srcoff
= op
->off
;
9240 uint64_t len
= op
->len
;
9241 uint64_t dstoff
= op
->dest_off
;
9242 r
= _clone_range(txc
, c
, o
, no
, srcoff
, len
, dstoff
);
9246 case Transaction::OP_COLL_ADD
:
9247 assert(0 == "not implemented");
9250 case Transaction::OP_COLL_REMOVE
:
9251 assert(0 == "not implemented");
9254 case Transaction::OP_COLL_MOVE
:
9255 assert(0 == "deprecated");
9258 case Transaction::OP_COLL_MOVE_RENAME
:
9259 case Transaction::OP_TRY_RENAME
:
9261 assert(op
->cid
== op
->dest_cid
);
9262 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9263 OnodeRef
& no
= ovec
[op
->dest_oid
];
9265 no
= c
->get_onode(noid
, false);
9267 r
= _rename(txc
, c
, o
, no
, noid
);
9271 case Transaction::OP_OMAP_CLEAR
:
9273 r
= _omap_clear(txc
, c
, o
);
9276 case Transaction::OP_OMAP_SETKEYS
:
9279 i
.decode_attrset_bl(&aset_bl
);
9280 r
= _omap_setkeys(txc
, c
, o
, aset_bl
);
9283 case Transaction::OP_OMAP_RMKEYS
:
9286 i
.decode_keyset_bl(&keys_bl
);
9287 r
= _omap_rmkeys(txc
, c
, o
, keys_bl
);
9290 case Transaction::OP_OMAP_RMKEYRANGE
:
9293 first
= i
.decode_string();
9294 last
= i
.decode_string();
9295 r
= _omap_rmkey_range(txc
, c
, o
, first
, last
);
9298 case Transaction::OP_OMAP_SETHEADER
:
9302 r
= _omap_setheader(txc
, c
, o
, bl
);
9306 case Transaction::OP_SETALLOCHINT
:
9308 r
= _set_alloc_hint(txc
, c
, o
,
9309 op
->expected_object_size
,
9310 op
->expected_write_size
,
9311 op
->alloc_hint_flags
);
9316 derr
<< __func__
<< "bad op " << op
->op
<< dendl
;
9324 if (r
== -ENOENT
&& !(op
->op
== Transaction::OP_CLONERANGE
||
9325 op
->op
== Transaction::OP_CLONE
||
9326 op
->op
== Transaction::OP_CLONERANGE2
||
9327 op
->op
== Transaction::OP_COLL_ADD
||
9328 op
->op
== Transaction::OP_SETATTR
||
9329 op
->op
== Transaction::OP_SETATTRS
||
9330 op
->op
== Transaction::OP_RMATTR
||
9331 op
->op
== Transaction::OP_OMAP_SETKEYS
||
9332 op
->op
== Transaction::OP_OMAP_RMKEYS
||
9333 op
->op
== Transaction::OP_OMAP_RMKEYRANGE
||
9334 op
->op
== Transaction::OP_OMAP_SETHEADER
))
9335 // -ENOENT is usually okay
9341 const char *msg
= "unexpected error code";
9343 if (r
== -ENOENT
&& (op
->op
== Transaction::OP_CLONERANGE
||
9344 op
->op
== Transaction::OP_CLONE
||
9345 op
->op
== Transaction::OP_CLONERANGE2
))
9346 msg
= "ENOENT on clone suggests osd bug";
9349 // For now, if we hit _any_ ENOSPC, crash, before we do any damage
9350 // by partially applying transactions.
9351 msg
= "ENOSPC from bluestore, misconfigured cluster";
9353 if (r
== -ENOTEMPTY
) {
9354 msg
= "ENOTEMPTY suggests garbage data in osd data dir";
9357 derr
<< __func__
<< " error " << cpp_strerror(r
)
9358 << " not handled on operation " << op
->op
9359 << " (op " << pos
<< ", counting from 0)"
9361 derr
<< msg
<< dendl
;
9362 _dump_transaction(t
, 0);
9363 assert(0 == "unexpected error");
9371 // -----------------
9374 int BlueStore::_touch(TransContext
*txc
,
9378 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
9380 _assign_nid(txc
, o
);
9381 txc
->write_onode(o
);
9382 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
9386 void BlueStore::_dump_onode(OnodeRef o
, int log_level
)
9388 if (!cct
->_conf
->subsys
.should_gather(ceph_subsys_bluestore
, log_level
))
9390 dout(log_level
) << __func__
<< " " << o
<< " " << o
->oid
9391 << " nid " << o
->onode
.nid
9392 << " size 0x" << std::hex
<< o
->onode
.size
9393 << " (" << std::dec
<< o
->onode
.size
<< ")"
9394 << " expected_object_size " << o
->onode
.expected_object_size
9395 << " expected_write_size " << o
->onode
.expected_write_size
9396 << " in " << o
->onode
.extent_map_shards
.size() << " shards"
9397 << ", " << o
->extent_map
.spanning_blob_map
.size()
9398 << " spanning blobs"
9400 for (auto p
= o
->onode
.attrs
.begin();
9401 p
!= o
->onode
.attrs
.end();
9403 dout(log_level
) << __func__
<< " attr " << p
->first
9404 << " len " << p
->second
.length() << dendl
;
9406 _dump_extent_map(o
->extent_map
, log_level
);
9409 void BlueStore::_dump_extent_map(ExtentMap
&em
, int log_level
)
9412 for (auto& s
: em
.shards
) {
9413 dout(log_level
) << __func__
<< " shard " << *s
.shard_info
9414 << (s
.loaded
? " (loaded)" : "")
9415 << (s
.dirty
? " (dirty)" : "")
9418 for (auto& e
: em
.extent_map
) {
9419 dout(log_level
) << __func__
<< " " << e
<< dendl
;
9420 assert(e
.logical_offset
>= pos
);
9421 pos
= e
.logical_offset
+ e
.length
;
9422 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
9423 if (blob
.has_csum()) {
9425 unsigned n
= blob
.get_csum_count();
9426 for (unsigned i
= 0; i
< n
; ++i
)
9427 v
.push_back(blob
.get_csum_item(i
));
9428 dout(log_level
) << __func__
<< " csum: " << std::hex
<< v
<< std::dec
9431 std::lock_guard
<std::recursive_mutex
> l(e
.blob
->shared_blob
->get_cache()->lock
);
9432 for (auto& i
: e
.blob
->shared_blob
->bc
.buffer_map
) {
9433 dout(log_level
) << __func__
<< " 0x" << std::hex
<< i
.first
9434 << "~" << i
.second
->length
<< std::dec
9435 << " " << *i
.second
<< dendl
;
9440 void BlueStore::_dump_transaction(Transaction
*t
, int log_level
)
9442 dout(log_level
) << " transaction dump:\n";
9443 JSONFormatter
f(true);
9444 f
.open_object_section("transaction");
9451 void BlueStore::_pad_zeros(
9452 bufferlist
*bl
, uint64_t *offset
,
9453 uint64_t chunk_size
)
9455 auto length
= bl
->length();
9456 dout(30) << __func__
<< " 0x" << std::hex
<< *offset
<< "~" << length
9457 << " chunk_size 0x" << chunk_size
<< std::dec
<< dendl
;
9458 dout(40) << "before:\n";
9459 bl
->hexdump(*_dout
);
9462 size_t front_pad
= *offset
% chunk_size
;
9463 size_t back_pad
= 0;
9464 size_t pad_count
= 0;
9466 size_t front_copy
= MIN(chunk_size
- front_pad
, length
);
9467 bufferptr z
= buffer::create_page_aligned(chunk_size
);
9468 z
.zero(0, front_pad
, false);
9469 pad_count
+= front_pad
;
9470 bl
->copy(0, front_copy
, z
.c_str() + front_pad
);
9471 if (front_copy
+ front_pad
< chunk_size
) {
9472 back_pad
= chunk_size
- (length
+ front_pad
);
9473 z
.zero(front_pad
+ length
, back_pad
, false);
9474 pad_count
+= back_pad
;
9478 t
.substr_of(old
, front_copy
, length
- front_copy
);
9480 bl
->claim_append(t
);
9481 *offset
-= front_pad
;
9482 length
+= pad_count
;
9486 uint64_t end
= *offset
+ length
;
9487 unsigned back_copy
= end
% chunk_size
;
9489 assert(back_pad
== 0);
9490 back_pad
= chunk_size
- back_copy
;
9491 assert(back_copy
<= length
);
9492 bufferptr
tail(chunk_size
);
9493 bl
->copy(length
- back_copy
, back_copy
, tail
.c_str());
9494 tail
.zero(back_copy
, back_pad
, false);
9497 bl
->substr_of(old
, 0, length
- back_copy
);
9500 pad_count
+= back_pad
;
9502 dout(20) << __func__
<< " pad 0x" << std::hex
<< front_pad
<< " + 0x"
9503 << back_pad
<< " on front/back, now 0x" << *offset
<< "~"
9504 << length
<< std::dec
<< dendl
;
9505 dout(40) << "after:\n";
9506 bl
->hexdump(*_dout
);
9509 logger
->inc(l_bluestore_write_pad_bytes
, pad_count
);
9510 assert(bl
->length() == length
);
9513 void BlueStore::_do_write_small(
9517 uint64_t offset
, uint64_t length
,
9518 bufferlist::iterator
& blp
,
9521 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9522 << std::dec
<< dendl
;
9523 assert(length
< min_alloc_size
);
9524 uint64_t end_offs
= offset
+ length
;
9526 logger
->inc(l_bluestore_write_small
);
9527 logger
->inc(l_bluestore_write_small_bytes
, length
);
9530 blp
.copy(length
, bl
);
9532 // Look for an existing mutable blob we can use.
9533 auto begin
= o
->extent_map
.extent_map
.begin();
9534 auto end
= o
->extent_map
.extent_map
.end();
9535 auto ep
= o
->extent_map
.seek_lextent(offset
);
9538 if (ep
->blob_end() <= offset
) {
9543 if (prev_ep
!= begin
) {
9546 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9549 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9550 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9551 uint32_t alloc_len
= min_alloc_size
;
9552 auto offset0
= P2ALIGN(offset
, alloc_len
);
9556 // search suitable extent in both forward and reverse direction in
9557 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9558 // then check if blob can be reused via can_reuse_blob func or apply
9559 // direct/deferred write (the latter for extents including or higher
9560 // than 'offset' only).
9564 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9565 BlobRef b
= ep
->blob
;
9566 auto bstart
= ep
->blob_start();
9567 dout(20) << __func__
<< " considering " << *b
9568 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9569 if (bstart
>= end_offs
) {
9570 dout(20) << __func__
<< " ignoring distant " << *b
<< dendl
;
9571 } else if (!b
->get_blob().is_mutable()) {
9572 dout(20) << __func__
<< " ignoring immutable " << *b
<< dendl
;
9573 } else if (ep
->logical_offset
% min_alloc_size
!=
9574 ep
->blob_offset
% min_alloc_size
) {
9575 dout(20) << __func__
<< " ignoring offset-skewed " << *b
<< dendl
;
9577 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9578 // can we pad our head/tail out with zeros?
9579 uint64_t head_pad
, tail_pad
;
9580 head_pad
= P2PHASE(offset
, chunk_size
);
9581 tail_pad
= P2NPHASE(end_offs
, chunk_size
);
9582 if (head_pad
|| tail_pad
) {
9583 o
->extent_map
.fault_range(db
, offset
- head_pad
,
9584 end_offs
- offset
+ head_pad
+ tail_pad
);
9587 o
->extent_map
.has_any_lextents(offset
- head_pad
, chunk_size
)) {
9590 if (tail_pad
&& o
->extent_map
.has_any_lextents(end_offs
, tail_pad
)) {
9594 uint64_t b_off
= offset
- head_pad
- bstart
;
9595 uint64_t b_len
= length
+ head_pad
+ tail_pad
;
9597 // direct write into unused blocks of an existing mutable blob?
9598 if ((b_off
% chunk_size
== 0 && b_len
% chunk_size
== 0) &&
9599 b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9600 b
->get_blob().is_unused(b_off
, b_len
) &&
9601 b
->get_blob().is_allocated(b_off
, b_len
)) {
9602 _apply_padding(head_pad
, tail_pad
, bl
);
9604 dout(20) << __func__
<< " write to unused 0x" << std::hex
9605 << b_off
<< "~" << b_len
9606 << " pad 0x" << head_pad
<< " + 0x" << tail_pad
9607 << std::dec
<< " of mutable " << *b
<< dendl
;
9608 _buffer_cache_write(txc
, b
, b_off
, bl
,
9609 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9611 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
9612 if (b_len
<= prefer_deferred_size
) {
9613 dout(20) << __func__
<< " deferring small 0x" << std::hex
9614 << b_len
<< std::dec
<< " unused write via deferred" << dendl
;
9615 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9616 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9619 [&](uint64_t offset
, uint64_t length
) {
9620 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9625 b
->get_blob().map_bl(
9627 [&](uint64_t offset
, bufferlist
& t
) {
9628 bdev
->aio_write(offset
, t
,
9629 &txc
->ioc
, wctx
->buffered
);
9633 b
->dirty_blob().calc_csum(b_off
, bl
);
9634 dout(20) << __func__
<< " lex old " << *ep
<< dendl
;
9635 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, b_off
+ head_pad
, length
,
9637 &wctx
->old_extents
);
9638 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9639 txc
->statfs_delta
.stored() += le
->length
;
9640 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9641 logger
->inc(l_bluestore_write_small_unused
);
9644 // read some data to fill out the chunk?
9645 uint64_t head_read
= P2PHASE(b_off
, chunk_size
);
9646 uint64_t tail_read
= P2NPHASE(b_off
+ b_len
, chunk_size
);
9647 if ((head_read
|| tail_read
) &&
9648 (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
+ tail_read
) &&
9649 head_read
+ tail_read
< min_alloc_size
) {
9651 b_len
+= head_read
+ tail_read
;
9654 head_read
= tail_read
= 0;
9657 // chunk-aligned deferred overwrite?
9658 if (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9659 b_off
% chunk_size
== 0 &&
9660 b_len
% chunk_size
== 0 &&
9661 b
->get_blob().is_allocated(b_off
, b_len
)) {
9663 _apply_padding(head_pad
, tail_pad
, bl
);
9665 dout(20) << __func__
<< " reading head 0x" << std::hex
<< head_read
9666 << " and tail 0x" << tail_read
<< std::dec
<< dendl
;
9669 int r
= _do_read(c
.get(), o
, offset
- head_pad
- head_read
, head_read
,
9671 assert(r
>= 0 && r
<= (int)head_read
);
9672 size_t zlen
= head_read
- r
;
9674 head_bl
.append_zero(zlen
);
9675 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9677 bl
.claim_prepend(head_bl
);
9678 logger
->inc(l_bluestore_write_penalty_read_ops
);
9682 int r
= _do_read(c
.get(), o
, offset
+ length
+ tail_pad
, tail_read
,
9684 assert(r
>= 0 && r
<= (int)tail_read
);
9685 size_t zlen
= tail_read
- r
;
9687 tail_bl
.append_zero(zlen
);
9688 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9690 bl
.claim_append(tail_bl
);
9691 logger
->inc(l_bluestore_write_penalty_read_ops
);
9693 logger
->inc(l_bluestore_write_small_pre_read
);
9695 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9696 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9697 _buffer_cache_write(txc
, b
, b_off
, bl
,
9698 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9700 int r
= b
->get_blob().map(
9702 [&](uint64_t offset
, uint64_t length
) {
9703 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9707 if (b
->get_blob().csum_type
) {
9708 b
->dirty_blob().calc_csum(b_off
, bl
);
9711 dout(20) << __func__
<< " deferred write 0x" << std::hex
<< b_off
<< "~"
9712 << b_len
<< std::dec
<< " of mutable " << *b
9713 << " at " << op
->extents
<< dendl
;
9714 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, offset
- bstart
, length
,
9715 b
, &wctx
->old_extents
);
9716 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9717 txc
->statfs_delta
.stored() += le
->length
;
9718 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9719 logger
->inc(l_bluestore_write_small_deferred
);
9722 // try to reuse blob if we can
9723 if (b
->can_reuse_blob(min_alloc_size
,
9727 assert(alloc_len
== min_alloc_size
); // expecting data always
9728 // fit into reused blob
9729 // Need to check for pending writes desiring to
9730 // reuse the same pextent. The rationale is that during GC two chunks
9731 // from garbage blobs(compressed?) can share logical space within the same
9732 // AU. That's in turn might be caused by unaligned len in clone_range2.
9733 // Hence the second write will fail in an attempt to reuse blob at
9734 // do_alloc_write().
9735 if (!wctx
->has_conflict(b
,
9737 offset0
+ alloc_len
,
9740 // we can't reuse pad_head/pad_tail since they might be truncated
9741 // due to existent extents
9742 uint64_t b_off
= offset
- bstart
;
9743 uint64_t b_off0
= b_off
;
9744 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9746 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9747 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9748 << " (0x" << b_off
<< "~" << length
<< ")"
9749 << std::dec
<< dendl
;
9751 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9752 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9754 logger
->inc(l_bluestore_write_small_unused
);
9761 } // if (ep != end && ep->logical_offset < offset + max_bsize)
9763 // check extent for reuse in reverse order
9764 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9765 BlobRef b
= prev_ep
->blob
;
9766 auto bstart
= prev_ep
->blob_start();
9767 dout(20) << __func__
<< " considering " << *b
9768 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9769 if (b
->can_reuse_blob(min_alloc_size
,
9773 assert(alloc_len
== min_alloc_size
); // expecting data always
9774 // fit into reused blob
9775 // Need to check for pending writes desiring to
9776 // reuse the same pextent. The rationale is that during GC two chunks
9777 // from garbage blobs(compressed?) can share logical space within the same
9778 // AU. That's in turn might be caused by unaligned len in clone_range2.
9779 // Hence the second write will fail in an attempt to reuse blob at
9780 // do_alloc_write().
9781 if (!wctx
->has_conflict(b
,
9783 offset0
+ alloc_len
,
9786 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9787 uint64_t b_off
= offset
- bstart
;
9788 uint64_t b_off0
= b_off
;
9789 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9791 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9792 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9793 << " (0x" << b_off
<< "~" << length
<< ")"
9794 << std::dec
<< dendl
;
9796 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9797 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9799 logger
->inc(l_bluestore_write_small_unused
);
9803 if (prev_ep
!= begin
) {
9807 prev_ep
= end
; // to avoid useless first extent re-check
9809 } // if (prev_ep != end && prev_ep->logical_offset >= min_off)
9810 } while (any_change
);
9814 BlobRef b
= c
->new_blob();
9815 uint64_t b_off
= P2PHASE(offset
, alloc_len
);
9816 uint64_t b_off0
= b_off
;
9817 _pad_zeros(&bl
, &b_off0
, block_size
);
9818 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9819 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
, true, true);
9820 logger
->inc(l_bluestore_write_small_new
);
9825 void BlueStore::_do_write_big(
9829 uint64_t offset
, uint64_t length
,
9830 bufferlist::iterator
& blp
,
9833 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9834 << " target_blob_size 0x" << wctx
->target_blob_size
<< std::dec
9835 << " compress " << (int)wctx
->compress
9837 logger
->inc(l_bluestore_write_big
);
9838 logger
->inc(l_bluestore_write_big_bytes
, length
);
9839 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9840 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9841 while (length
> 0) {
9842 bool new_blob
= false;
9843 uint32_t l
= MIN(max_bsize
, length
);
9847 //attempting to reuse existing blob
9848 if (!wctx
->compress
) {
9849 // look for an existing mutable blob we can reuse
9850 auto begin
= o
->extent_map
.extent_map
.begin();
9851 auto end
= o
->extent_map
.extent_map
.end();
9852 auto ep
= o
->extent_map
.seek_lextent(offset
);
9854 if (prev_ep
!= begin
) {
9857 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9859 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9860 // search suitable extent in both forward and reverse direction in
9861 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9862 // then check if blob can be reused via can_reuse_blob func.
9866 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9867 if (offset
>= ep
->blob_start() &&
9868 ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9869 offset
- ep
->blob_start(),
9872 b_off
= offset
- ep
->blob_start();
9873 prev_ep
= end
; // to avoid check below
9874 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9875 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9882 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9883 if (prev_ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9884 offset
- prev_ep
->blob_start(),
9887 b_off
= offset
- prev_ep
->blob_start();
9888 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9889 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9890 } else if (prev_ep
!= begin
) {
9894 prev_ep
= end
; // to avoid useless first extent re-check
9897 } while (b
== nullptr && any_change
);
9907 wctx
->write(offset
, b
, l
, b_off
, t
, b_off
, l
, false, new_blob
);
9910 logger
->inc(l_bluestore_write_big_blobs
);
9914 int BlueStore::_do_alloc_write(
9920 dout(20) << __func__
<< " txc " << txc
9921 << " " << wctx
->writes
.size() << " blobs"
9923 if (wctx
->writes
.empty()) {
9929 if (wctx
->compress
) {
9931 "compression_algorithm",
9935 if (coll
->pool_opts
.get(pool_opts_t::COMPRESSION_ALGORITHM
, &val
)) {
9936 CompressorRef cp
= compressor
;
9937 if (!cp
|| cp
->get_type_name() != val
) {
9938 cp
= Compressor::create(cct
, val
);
9940 return boost::optional
<CompressorRef
>(cp
);
9942 return boost::optional
<CompressorRef
>();
9946 crr
= select_option(
9947 "compression_required_ratio",
9948 cct
->_conf
->bluestore_compression_required_ratio
,
9951 if (coll
->pool_opts
.get(pool_opts_t::COMPRESSION_REQUIRED_RATIO
, &val
)) {
9952 return boost::optional
<double>(val
);
9954 return boost::optional
<double>();
9960 int csum
= csum_type
.load();
9961 csum
= select_option(
9966 if (coll
->pool_opts
.get(pool_opts_t::CSUM_TYPE
, &val
)) {
9967 return boost::optional
<int>(val
);
9969 return boost::optional
<int>();
9973 // compress (as needed) and calc needed space
9975 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9976 for (auto& wi
: wctx
->writes
) {
9977 if (c
&& wi
.blob_length
> min_alloc_size
) {
9978 utime_t start
= ceph_clock_now();
9981 assert(wi
.b_off
== 0);
9982 assert(wi
.blob_length
== wi
.bl
.length());
9984 // FIXME: memory alignment here is bad
9986 int r
= c
->compress(wi
.bl
, t
);
9989 bluestore_compression_header_t chdr
;
9990 chdr
.type
= c
->get_type();
9991 chdr
.length
= t
.length();
9992 ::encode(chdr
, wi
.compressed_bl
);
9993 wi
.compressed_bl
.claim_append(t
);
9995 wi
.compressed_len
= wi
.compressed_bl
.length();
9996 uint64_t newlen
= P2ROUNDUP(wi
.compressed_len
, min_alloc_size
);
9997 uint64_t want_len_raw
= wi
.blob_length
* crr
;
9998 uint64_t want_len
= P2ROUNDUP(want_len_raw
, min_alloc_size
);
9999 if (newlen
<= want_len
&& newlen
< wi
.blob_length
) {
10000 // Cool. We compressed at least as much as we were hoping to.
10001 // pad out to min_alloc_size
10002 wi
.compressed_bl
.append_zero(newlen
- wi
.compressed_len
);
10003 logger
->inc(l_bluestore_write_pad_bytes
, newlen
- wi
.compressed_len
);
10004 dout(20) << __func__
<< std::hex
<< " compressed 0x" << wi
.blob_length
10005 << " -> 0x" << wi
.compressed_len
<< " => 0x" << newlen
10006 << " with " << c
->get_type()
10007 << std::dec
<< dendl
;
10008 txc
->statfs_delta
.compressed() += wi
.compressed_len
;
10009 txc
->statfs_delta
.compressed_original() += wi
.blob_length
;
10010 txc
->statfs_delta
.compressed_allocated() += newlen
;
10011 logger
->inc(l_bluestore_compress_success_count
);
10012 wi
.compressed
= true;
10015 dout(20) << __func__
<< std::hex
<< " 0x" << wi
.blob_length
10016 << " compressed to 0x" << wi
.compressed_len
<< " -> 0x" << newlen
10017 << " with " << c
->get_type()
10018 << ", which is more than required 0x" << want_len_raw
10019 << " -> 0x" << want_len
10020 << ", leaving uncompressed"
10021 << std::dec
<< dendl
;
10022 logger
->inc(l_bluestore_compress_rejected_count
);
10023 need
+= wi
.blob_length
;
10025 logger
->tinc(l_bluestore_compress_lat
,
10026 ceph_clock_now() - start
);
10028 need
+= wi
.blob_length
;
10031 int r
= alloc
->reserve(need
);
10033 derr
<< __func__
<< " failed to reserve 0x" << std::hex
<< need
<< std::dec
10037 AllocExtentVector prealloc
;
10038 prealloc
.reserve(2 * wctx
->writes
.size());;
10039 int prealloc_left
= 0;
10040 prealloc_left
= alloc
->allocate(
10041 need
, min_alloc_size
, need
,
10043 assert(prealloc_left
== (int64_t)need
);
10044 dout(20) << __func__
<< " prealloc " << prealloc
<< dendl
;
10045 auto prealloc_pos
= prealloc
.begin();
10047 for (auto& wi
: wctx
->writes
) {
10049 bluestore_blob_t
& dblob
= b
->dirty_blob();
10050 uint64_t b_off
= wi
.b_off
;
10051 bufferlist
*l
= &wi
.bl
;
10052 uint64_t final_length
= wi
.blob_length
;
10053 uint64_t csum_length
= wi
.blob_length
;
10054 unsigned csum_order
= block_size_order
;
10055 if (wi
.compressed
) {
10056 final_length
= wi
.compressed_bl
.length();
10057 csum_length
= final_length
;
10058 csum_order
= ctz(csum_length
);
10059 l
= &wi
.compressed_bl
;
10060 dblob
.set_compressed(wi
.blob_length
, wi
.compressed_len
);
10061 } else if (wi
.new_blob
) {
10062 // initialize newly created blob only
10063 assert(dblob
.is_mutable());
10064 if (l
->length() != wi
.blob_length
) {
10065 // hrm, maybe we could do better here, but let's not bother.
10066 dout(20) << __func__
<< " forcing csum_order to block_size_order "
10067 << block_size_order
<< dendl
;
10068 csum_order
= block_size_order
;
10070 csum_order
= std::min(wctx
->csum_order
, ctz(l
->length()));
10072 // try to align blob with max_blob_size to improve
10073 // its reuse ratio, e.g. in case of reverse write
10074 uint32_t suggested_boff
=
10075 (wi
.logical_offset
- (wi
.b_off0
- wi
.b_off
)) % max_bsize
;
10076 if ((suggested_boff
% (1 << csum_order
)) == 0 &&
10077 suggested_boff
+ final_length
<= max_bsize
&&
10078 suggested_boff
> b_off
) {
10079 dout(20) << __func__
<< " forcing blob_offset to 0x"
10080 << std::hex
<< suggested_boff
<< std::dec
<< dendl
;
10081 assert(suggested_boff
>= b_off
);
10082 csum_length
+= suggested_boff
- b_off
;
10083 b_off
= suggested_boff
;
10085 if (csum
!= Checksummer::CSUM_NONE
) {
10086 dout(20) << __func__
<< " initialize csum setting for new blob " << *b
10087 << " csum_type " << Checksummer::get_csum_type_string(csum
)
10088 << " csum_order " << csum_order
10089 << " csum_length 0x" << std::hex
<< csum_length
<< std::dec
10091 dblob
.init_csum(csum
, csum_order
, csum_length
);
10095 AllocExtentVector extents
;
10096 int64_t left
= final_length
;
10098 assert(prealloc_left
> 0);
10099 if (prealloc_pos
->length
<= left
) {
10100 prealloc_left
-= prealloc_pos
->length
;
10101 left
-= prealloc_pos
->length
;
10102 txc
->statfs_delta
.allocated() += prealloc_pos
->length
;
10103 extents
.push_back(*prealloc_pos
);
10106 extents
.emplace_back(prealloc_pos
->offset
, left
);
10107 prealloc_pos
->offset
+= left
;
10108 prealloc_pos
->length
-= left
;
10109 prealloc_left
-= left
;
10110 txc
->statfs_delta
.allocated() += left
;
10115 for (auto& p
: extents
) {
10116 txc
->allocated
.insert(p
.offset
, p
.length
);
10118 dblob
.allocated(P2ALIGN(b_off
, min_alloc_size
), final_length
, extents
);
10120 dout(20) << __func__
<< " blob " << *b
<< dendl
;
10121 if (dblob
.has_csum()) {
10122 dblob
.calc_csum(b_off
, *l
);
10125 if (wi
.mark_unused
) {
10126 auto b_end
= b_off
+ wi
.bl
.length();
10128 dblob
.add_unused(0, b_off
);
10130 if (b_end
< wi
.blob_length
) {
10131 dblob
.add_unused(b_end
, wi
.blob_length
- b_end
);
10135 Extent
*le
= o
->extent_map
.set_lextent(coll
, wi
.logical_offset
,
10136 b_off
+ (wi
.b_off0
- wi
.b_off
),
10140 wi
.b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
10141 txc
->statfs_delta
.stored() += le
->length
;
10142 dout(20) << __func__
<< " lex " << *le
<< dendl
;
10143 _buffer_cache_write(txc
, wi
.b
, b_off
, wi
.bl
,
10144 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
10147 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
10148 if (l
->length() <= prefer_deferred_size
.load()) {
10149 dout(20) << __func__
<< " deferring small 0x" << std::hex
10150 << l
->length() << std::dec
<< " write via deferred" << dendl
;
10151 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
10152 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
10153 int r
= b
->get_blob().map(
10154 b_off
, l
->length(),
10155 [&](uint64_t offset
, uint64_t length
) {
10156 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
10162 b
->get_blob().map_bl(
10164 [&](uint64_t offset
, bufferlist
& t
) {
10165 bdev
->aio_write(offset
, t
, &txc
->ioc
, false);
10170 assert(prealloc_pos
== prealloc
.end());
10171 assert(prealloc_left
== 0);
10175 void BlueStore::_wctx_finish(
10179 WriteContext
*wctx
,
10180 set
<SharedBlob
*> *maybe_unshared_blobs
)
10182 auto oep
= wctx
->old_extents
.begin();
10183 while (oep
!= wctx
->old_extents
.end()) {
10185 oep
= wctx
->old_extents
.erase(oep
);
10186 dout(20) << __func__
<< " lex_old " << lo
.e
<< dendl
;
10187 BlobRef b
= lo
.e
.blob
;
10188 const bluestore_blob_t
& blob
= b
->get_blob();
10189 if (blob
.is_compressed()) {
10190 if (lo
.blob_empty
) {
10191 txc
->statfs_delta
.compressed() -= blob
.get_compressed_payload_length();
10193 txc
->statfs_delta
.compressed_original() -= lo
.e
.length
;
10196 txc
->statfs_delta
.stored() -= lo
.e
.length
;
10198 dout(20) << __func__
<< " blob release " << r
<< dendl
;
10199 if (blob
.is_shared()) {
10200 PExtentVector final
;
10201 c
->load_shared_blob(b
->shared_blob
);
10203 b
->shared_blob
->put_ref(
10204 e
.offset
, e
.length
, &final
,
10205 b
->is_referenced() ? nullptr : maybe_unshared_blobs
);
10207 dout(20) << __func__
<< " shared_blob release " << final
10208 << " from " << *b
->shared_blob
<< dendl
;
10209 txc
->write_shared_blob(b
->shared_blob
);
10214 // we can't invalidate our logical extents as we drop them because
10215 // other lextents (either in our onode or others) may still
10216 // reference them. but we can throw out anything that is no
10217 // longer allocated. Note that this will leave behind edge bits
10218 // that are no longer referenced but not deallocated (until they
10219 // age out of the cache naturally).
10220 b
->discard_unallocated(c
.get());
10222 dout(20) << __func__
<< " release " << e
<< dendl
;
10223 txc
->released
.insert(e
.offset
, e
.length
);
10224 txc
->statfs_delta
.allocated() -= e
.length
;
10225 if (blob
.is_compressed()) {
10226 txc
->statfs_delta
.compressed_allocated() -= e
.length
;
10230 if (b
->is_spanning() && !b
->is_referenced()) {
10231 dout(20) << __func__
<< " spanning_blob_map removing empty " << *b
10233 o
->extent_map
.spanning_blob_map
.erase(b
->id
);
10238 void BlueStore::_do_write_data(
10245 WriteContext
*wctx
)
10247 uint64_t end
= offset
+ length
;
10248 bufferlist::iterator p
= bl
.begin();
10250 if (offset
/ min_alloc_size
== (end
- 1) / min_alloc_size
&&
10251 (length
!= min_alloc_size
)) {
10252 // we fall within the same block
10253 _do_write_small(txc
, c
, o
, offset
, length
, p
, wctx
);
10255 uint64_t head_offset
, head_length
;
10256 uint64_t middle_offset
, middle_length
;
10257 uint64_t tail_offset
, tail_length
;
10259 head_offset
= offset
;
10260 head_length
= P2NPHASE(offset
, min_alloc_size
);
10262 tail_offset
= P2ALIGN(end
, min_alloc_size
);
10263 tail_length
= P2PHASE(end
, min_alloc_size
);
10265 middle_offset
= head_offset
+ head_length
;
10266 middle_length
= length
- head_length
- tail_length
;
10269 _do_write_small(txc
, c
, o
, head_offset
, head_length
, p
, wctx
);
10272 if (middle_length
) {
10273 _do_write_big(txc
, c
, o
, middle_offset
, middle_length
, p
, wctx
);
10277 _do_write_small(txc
, c
, o
, tail_offset
, tail_length
, p
, wctx
);
10282 void BlueStore::_choose_write_options(
10285 uint32_t fadvise_flags
,
10286 WriteContext
*wctx
)
10288 if (fadvise_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
10289 dout(20) << __func__
<< " will do buffered write" << dendl
;
10290 wctx
->buffered
= true;
10291 } else if (cct
->_conf
->bluestore_default_buffered_write
&&
10292 (fadvise_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
10293 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
10294 dout(20) << __func__
<< " defaulting to buffered write" << dendl
;
10295 wctx
->buffered
= true;
10298 // apply basic csum block size
10299 wctx
->csum_order
= block_size_order
;
10301 // compression parameters
10302 unsigned alloc_hints
= o
->onode
.alloc_hint_flags
;
10303 auto cm
= select_option(
10304 "compression_mode",
10308 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MODE
, &val
)) {
10309 return boost::optional
<Compressor::CompressionMode
>(
10310 Compressor::get_comp_mode_type(val
));
10312 return boost::optional
<Compressor::CompressionMode
>();
10316 wctx
->compress
= (cm
!= Compressor::COMP_NONE
) &&
10317 ((cm
== Compressor::COMP_FORCE
) ||
10318 (cm
== Compressor::COMP_AGGRESSIVE
&&
10319 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE
) == 0) ||
10320 (cm
== Compressor::COMP_PASSIVE
&&
10321 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE
)));
10323 if ((alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_SEQUENTIAL_READ
) &&
10324 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_READ
) == 0 &&
10325 (alloc_hints
& (CEPH_OSD_ALLOC_HINT_FLAG_IMMUTABLE
|
10326 CEPH_OSD_ALLOC_HINT_FLAG_APPEND_ONLY
)) &&
10327 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_WRITE
) == 0) {
10329 dout(20) << __func__
<< " will prefer large blob and csum sizes" << dendl
;
10331 if (o
->onode
.expected_write_size
) {
10332 wctx
->csum_order
= std::max(min_alloc_size_order
,
10333 (uint8_t)ctz(o
->onode
.expected_write_size
));
10335 wctx
->csum_order
= min_alloc_size_order
;
10338 if (wctx
->compress
) {
10339 wctx
->target_blob_size
= select_option(
10340 "compression_max_blob_size",
10341 comp_max_blob_size
.load(),
10344 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MAX_BLOB_SIZE
, &val
)) {
10345 return boost::optional
<uint64_t>((uint64_t)val
);
10347 return boost::optional
<uint64_t>();
10352 if (wctx
->compress
) {
10353 wctx
->target_blob_size
= select_option(
10354 "compression_min_blob_size",
10355 comp_min_blob_size
.load(),
10358 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MIN_BLOB_SIZE
, &val
)) {
10359 return boost::optional
<uint64_t>((uint64_t)val
);
10361 return boost::optional
<uint64_t>();
10367 uint64_t max_bsize
= max_blob_size
.load();
10368 if (wctx
->target_blob_size
== 0 || wctx
->target_blob_size
> max_bsize
) {
10369 wctx
->target_blob_size
= max_bsize
;
10372 // set the min blob size floor at 2x the min_alloc_size, or else we
10373 // won't be able to allocate a smaller extent for the compressed
10375 if (wctx
->compress
&&
10376 wctx
->target_blob_size
< min_alloc_size
* 2) {
10377 wctx
->target_blob_size
= min_alloc_size
* 2;
10380 dout(20) << __func__
<< " prefer csum_order " << wctx
->csum_order
10381 << " target_blob_size 0x" << std::hex
<< wctx
->target_blob_size
10382 << std::dec
<< dendl
;
10385 int BlueStore::_do_gc(
10389 const GarbageCollector
& gc
,
10390 const WriteContext
& wctx
,
10391 uint64_t *dirty_start
,
10392 uint64_t *dirty_end
)
10394 auto& extents_to_collect
= gc
.get_extents_to_collect();
10396 WriteContext wctx_gc
;
10397 wctx_gc
.fork(wctx
); // make a clone for garbage collection
10399 for (auto it
= extents_to_collect
.begin();
10400 it
!= extents_to_collect
.end();
10403 int r
= _do_read(c
.get(), o
, it
->offset
, it
->length
, bl
, 0);
10404 assert(r
== (int)it
->length
);
10406 o
->extent_map
.fault_range(db
, it
->offset
, it
->length
);
10407 _do_write_data(txc
, c
, o
, it
->offset
, it
->length
, bl
, &wctx_gc
);
10408 logger
->inc(l_bluestore_gc_merged
, it
->length
);
10410 if (*dirty_start
> it
->offset
) {
10411 *dirty_start
= it
->offset
;
10414 if (*dirty_end
< it
->offset
+ it
->length
) {
10415 *dirty_end
= it
->offset
+ it
->length
;
10419 dout(30) << __func__
<< " alloc write" << dendl
;
10420 int r
= _do_alloc_write(txc
, c
, o
, &wctx_gc
);
10422 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10427 _wctx_finish(txc
, c
, o
, &wctx_gc
);
10431 int BlueStore::_do_write(
10438 uint32_t fadvise_flags
)
10442 dout(20) << __func__
10444 << " 0x" << std::hex
<< offset
<< "~" << length
10445 << " - have 0x" << o
->onode
.size
10446 << " (" << std::dec
<< o
->onode
.size
<< ")"
10448 << " fadvise_flags 0x" << std::hex
<< fadvise_flags
<< std::dec
10456 uint64_t end
= offset
+ length
;
10458 GarbageCollector
gc(c
->store
->cct
);
10460 auto dirty_start
= offset
;
10461 auto dirty_end
= end
;
10464 _choose_write_options(c
, o
, fadvise_flags
, &wctx
);
10465 o
->extent_map
.fault_range(db
, offset
, length
);
10466 _do_write_data(txc
, c
, o
, offset
, length
, bl
, &wctx
);
10467 r
= _do_alloc_write(txc
, c
, o
, &wctx
);
10469 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10474 // NB: _wctx_finish() will empty old_extents
10475 // so we must do gc estimation before that
10476 benefit
= gc
.estimate(offset
,
10482 _wctx_finish(txc
, c
, o
, &wctx
);
10483 if (end
> o
->onode
.size
) {
10484 dout(20) << __func__
<< " extending size to 0x" << std::hex
<< end
10485 << std::dec
<< dendl
;
10486 o
->onode
.size
= end
;
10489 if (benefit
>= g_conf
->bluestore_gc_enable_total_threshold
) {
10490 if (!gc
.get_extents_to_collect().empty()) {
10491 dout(20) << __func__
<< " perform garbage collection, "
10492 << "expected benefit = " << benefit
<< " AUs" << dendl
;
10493 r
= _do_gc(txc
, c
, o
, gc
, wctx
, &dirty_start
, &dirty_end
);
10495 derr
<< __func__
<< " _do_gc failed with " << cpp_strerror(r
)
10502 o
->extent_map
.compress_extent_map(dirty_start
, dirty_end
- dirty_start
);
10503 o
->extent_map
.dirty_range(dirty_start
, dirty_end
- dirty_start
);
10511 int BlueStore::_write(TransContext
*txc
,
10514 uint64_t offset
, size_t length
,
10516 uint32_t fadvise_flags
)
10518 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10519 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10522 if (offset
+ length
>= OBJECT_MAX_SIZE
) {
10525 _assign_nid(txc
, o
);
10526 r
= _do_write(txc
, c
, o
, offset
, length
, bl
, fadvise_flags
);
10527 txc
->write_onode(o
);
10529 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10530 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10531 << " = " << r
<< dendl
;
10535 int BlueStore::_zero(TransContext
*txc
,
10538 uint64_t offset
, size_t length
)
10540 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10541 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10544 if (offset
+ length
>= OBJECT_MAX_SIZE
) {
10547 _assign_nid(txc
, o
);
10548 r
= _do_zero(txc
, c
, o
, offset
, length
);
10550 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10551 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10552 << " = " << r
<< dendl
;
10556 int BlueStore::_do_zero(TransContext
*txc
,
10559 uint64_t offset
, size_t length
)
10561 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10562 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10569 o
->extent_map
.fault_range(db
, offset
, length
);
10570 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10571 o
->extent_map
.dirty_range(offset
, length
);
10572 _wctx_finish(txc
, c
, o
, &wctx
);
10574 if (offset
+ length
> o
->onode
.size
) {
10575 o
->onode
.size
= offset
+ length
;
10576 dout(20) << __func__
<< " extending size to " << offset
+ length
10579 txc
->write_onode(o
);
10581 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10582 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10583 << " = " << r
<< dendl
;
10587 void BlueStore::_do_truncate(
10588 TransContext
*txc
, CollectionRef
& c
, OnodeRef o
, uint64_t offset
,
10589 set
<SharedBlob
*> *maybe_unshared_blobs
)
10591 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10592 << " 0x" << std::hex
<< offset
<< std::dec
<< dendl
;
10594 _dump_onode(o
, 30);
10596 if (offset
== o
->onode
.size
)
10599 if (offset
< o
->onode
.size
) {
10601 uint64_t length
= o
->onode
.size
- offset
;
10602 o
->extent_map
.fault_range(db
, offset
, length
);
10603 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10604 o
->extent_map
.dirty_range(offset
, length
);
10605 _wctx_finish(txc
, c
, o
, &wctx
, maybe_unshared_blobs
);
10607 // if we have shards past EOF, ask for a reshard
10608 if (!o
->onode
.extent_map_shards
.empty() &&
10609 o
->onode
.extent_map_shards
.back().offset
>= offset
) {
10610 dout(10) << __func__
<< " request reshard past EOF" << dendl
;
10612 o
->extent_map
.request_reshard(offset
- 1, offset
+ length
);
10614 o
->extent_map
.request_reshard(0, length
);
10619 o
->onode
.size
= offset
;
10621 txc
->write_onode(o
);
10624 int BlueStore::_truncate(TransContext
*txc
,
10629 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10630 << " 0x" << std::hex
<< offset
<< std::dec
10633 if (offset
>= OBJECT_MAX_SIZE
) {
10636 _do_truncate(txc
, c
, o
, offset
);
10638 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10639 << " 0x" << std::hex
<< offset
<< std::dec
10640 << " = " << r
<< dendl
;
10644 int BlueStore::_do_remove(
10649 set
<SharedBlob
*> maybe_unshared_blobs
;
10650 bool is_gen
= !o
->oid
.is_no_gen();
10651 _do_truncate(txc
, c
, o
, 0, is_gen
? &maybe_unshared_blobs
: nullptr);
10652 if (o
->onode
.has_omap()) {
10654 _do_omap_clear(txc
, o
->onode
.nid
);
10658 for (auto &s
: o
->extent_map
.shards
) {
10659 dout(20) << __func__
<< " removing shard 0x" << std::hex
10660 << s
.shard_info
->offset
<< std::dec
<< dendl
;
10661 generate_extent_shard_key_and_apply(o
->key
, s
.shard_info
->offset
, &key
,
10662 [&](const string
& final_key
) {
10663 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
10667 txc
->t
->rmkey(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size());
10669 o
->extent_map
.clear();
10670 o
->onode
= bluestore_onode_t();
10671 _debug_obj_on_delete(o
->oid
);
10673 if (!is_gen
|| maybe_unshared_blobs
.empty()) {
10677 // see if we can unshare blobs still referenced by the head
10678 dout(10) << __func__
<< " gen and maybe_unshared_blobs "
10679 << maybe_unshared_blobs
<< dendl
;
10680 ghobject_t nogen
= o
->oid
;
10681 nogen
.generation
= ghobject_t::NO_GEN
;
10682 OnodeRef h
= c
->onode_map
.lookup(nogen
);
10684 if (!h
|| !h
->exists
) {
10688 dout(20) << __func__
<< " checking for unshareable blobs on " << h
10689 << " " << h
->oid
<< dendl
;
10690 map
<SharedBlob
*,bluestore_extent_ref_map_t
> expect
;
10691 for (auto& e
: h
->extent_map
.extent_map
) {
10692 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10693 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10694 if (b
.is_shared() &&
10696 maybe_unshared_blobs
.count(sb
)) {
10697 if (b
.is_compressed()) {
10698 expect
[sb
].get(0, b
.get_ondisk_length());
10700 b
.map(e
.blob_offset
, e
.length
, [&](uint64_t off
, uint64_t len
) {
10701 expect
[sb
].get(off
, len
);
10708 vector
<SharedBlob
*> unshared_blobs
;
10709 unshared_blobs
.reserve(maybe_unshared_blobs
.size());
10710 for (auto& p
: expect
) {
10711 dout(20) << " ? " << *p
.first
<< " vs " << p
.second
<< dendl
;
10712 if (p
.first
->persistent
->ref_map
== p
.second
) {
10713 SharedBlob
*sb
= p
.first
;
10714 dout(20) << __func__
<< " unsharing " << *sb
<< dendl
;
10715 unshared_blobs
.push_back(sb
);
10716 txc
->unshare_blob(sb
);
10717 uint64_t sbid
= c
->make_blob_unshared(sb
);
10719 get_shared_blob_key(sbid
, &key
);
10720 txc
->t
->rmkey(PREFIX_SHARED_BLOB
, key
);
10724 if (unshared_blobs
.empty()) {
10728 for (auto& e
: h
->extent_map
.extent_map
) {
10729 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10730 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10731 if (b
.is_shared() &&
10732 std::find(unshared_blobs
.begin(), unshared_blobs
.end(),
10733 sb
) != unshared_blobs
.end()) {
10734 dout(20) << __func__
<< " unsharing " << e
<< dendl
;
10735 bluestore_blob_t
& blob
= e
.blob
->dirty_blob();
10736 blob
.clear_flag(bluestore_blob_t::FLAG_SHARED
);
10737 h
->extent_map
.dirty_range(e
.logical_offset
, 1);
10740 txc
->write_onode(h
);
10745 int BlueStore::_remove(TransContext
*txc
,
10749 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10750 int r
= _do_remove(txc
, c
, o
);
10751 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10755 int BlueStore::_setattr(TransContext
*txc
,
10758 const string
& name
,
10761 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10762 << " " << name
<< " (" << val
.length() << " bytes)"
10765 if (val
.is_partial()) {
10766 auto& b
= o
->onode
.attrs
[name
.c_str()] = bufferptr(val
.c_str(),
10768 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10770 auto& b
= o
->onode
.attrs
[name
.c_str()] = val
;
10771 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10773 txc
->write_onode(o
);
10774 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10775 << " " << name
<< " (" << val
.length() << " bytes)"
10776 << " = " << r
<< dendl
;
10780 int BlueStore::_setattrs(TransContext
*txc
,
10783 const map
<string
,bufferptr
>& aset
)
10785 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10786 << " " << aset
.size() << " keys"
10789 for (map
<string
,bufferptr
>::const_iterator p
= aset
.begin();
10790 p
!= aset
.end(); ++p
) {
10791 if (p
->second
.is_partial()) {
10792 auto& b
= o
->onode
.attrs
[p
->first
.c_str()] =
10793 bufferptr(p
->second
.c_str(), p
->second
.length());
10794 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10796 auto& b
= o
->onode
.attrs
[p
->first
.c_str()] = p
->second
;
10797 b
.reassign_to_mempool(mempool::mempool_bluestore_cache_other
);
10800 txc
->write_onode(o
);
10801 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10802 << " " << aset
.size() << " keys"
10803 << " = " << r
<< dendl
;
10808 int BlueStore::_rmattr(TransContext
*txc
,
10811 const string
& name
)
10813 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10814 << " " << name
<< dendl
;
10816 auto it
= o
->onode
.attrs
.find(name
.c_str());
10817 if (it
== o
->onode
.attrs
.end())
10820 o
->onode
.attrs
.erase(it
);
10821 txc
->write_onode(o
);
10824 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10825 << " " << name
<< " = " << r
<< dendl
;
10829 int BlueStore::_rmattrs(TransContext
*txc
,
10833 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10836 if (o
->onode
.attrs
.empty())
10839 o
->onode
.attrs
.clear();
10840 txc
->write_onode(o
);
10843 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10847 void BlueStore::_do_omap_clear(TransContext
*txc
, uint64_t id
)
10849 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
10850 string prefix
, tail
;
10851 get_omap_header(id
, &prefix
);
10852 get_omap_tail(id
, &tail
);
10853 it
->lower_bound(prefix
);
10854 while (it
->valid()) {
10855 if (it
->key() >= tail
) {
10856 dout(30) << __func__
<< " stop at " << pretty_binary_string(tail
)
10860 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
10861 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
10866 int BlueStore::_omap_clear(TransContext
*txc
,
10870 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10872 if (o
->onode
.has_omap()) {
10874 _do_omap_clear(txc
, o
->onode
.nid
);
10875 o
->onode
.clear_omap_flag();
10876 txc
->write_onode(o
);
10878 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10882 int BlueStore::_omap_setkeys(TransContext
*txc
,
10887 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10889 bufferlist::iterator p
= bl
.begin();
10891 if (!o
->onode
.has_omap()) {
10892 o
->onode
.set_omap_flag();
10893 txc
->write_onode(o
);
10895 txc
->note_modified_object(o
);
10898 _key_encode_u64(o
->onode
.nid
, &final_key
);
10899 final_key
.push_back('.');
10905 ::decode(value
, p
);
10906 final_key
.resize(9); // keep prefix
10908 dout(30) << __func__
<< " " << pretty_binary_string(final_key
)
10909 << " <- " << key
<< dendl
;
10910 txc
->t
->set(PREFIX_OMAP
, final_key
, value
);
10913 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10917 int BlueStore::_omap_setheader(TransContext
*txc
,
10922 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10925 if (!o
->onode
.has_omap()) {
10926 o
->onode
.set_omap_flag();
10927 txc
->write_onode(o
);
10929 txc
->note_modified_object(o
);
10931 get_omap_header(o
->onode
.nid
, &key
);
10932 txc
->t
->set(PREFIX_OMAP
, key
, bl
);
10934 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10938 int BlueStore::_omap_rmkeys(TransContext
*txc
,
10943 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10945 bufferlist::iterator p
= bl
.begin();
10949 if (!o
->onode
.has_omap()) {
10952 _key_encode_u64(o
->onode
.nid
, &final_key
);
10953 final_key
.push_back('.');
10958 final_key
.resize(9); // keep prefix
10960 dout(30) << __func__
<< " rm " << pretty_binary_string(final_key
)
10961 << " <- " << key
<< dendl
;
10962 txc
->t
->rmkey(PREFIX_OMAP
, final_key
);
10964 txc
->note_modified_object(o
);
10967 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10971 int BlueStore::_omap_rmkey_range(TransContext
*txc
,
10974 const string
& first
, const string
& last
)
10976 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10977 KeyValueDB::Iterator it
;
10978 string key_first
, key_last
;
10980 if (!o
->onode
.has_omap()) {
10984 it
= db
->get_iterator(PREFIX_OMAP
);
10985 get_omap_key(o
->onode
.nid
, first
, &key_first
);
10986 get_omap_key(o
->onode
.nid
, last
, &key_last
);
10987 it
->lower_bound(key_first
);
10988 while (it
->valid()) {
10989 if (it
->key() >= key_last
) {
10990 dout(30) << __func__
<< " stop at " << pretty_binary_string(key_last
)
10994 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
10995 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
10998 txc
->note_modified_object(o
);
11001 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
11005 int BlueStore::_set_alloc_hint(
11009 uint64_t expected_object_size
,
11010 uint64_t expected_write_size
,
11013 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
11014 << " object_size " << expected_object_size
11015 << " write_size " << expected_write_size
11016 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
11019 o
->onode
.expected_object_size
= expected_object_size
;
11020 o
->onode
.expected_write_size
= expected_write_size
;
11021 o
->onode
.alloc_hint_flags
= flags
;
11022 txc
->write_onode(o
);
11023 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
11024 << " object_size " << expected_object_size
11025 << " write_size " << expected_write_size
11026 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
11027 << " = " << r
<< dendl
;
11031 int BlueStore::_clone(TransContext
*txc
,
11036 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11037 << newo
->oid
<< dendl
;
11039 if (oldo
->oid
.hobj
.get_hash() != newo
->oid
.hobj
.get_hash()) {
11040 derr
<< __func__
<< " mismatched hash on " << oldo
->oid
11041 << " and " << newo
->oid
<< dendl
;
11045 _assign_nid(txc
, newo
);
11049 _do_truncate(txc
, c
, newo
, 0);
11050 if (cct
->_conf
->bluestore_clone_cow
) {
11051 _do_clone_range(txc
, c
, oldo
, newo
, 0, oldo
->onode
.size
, 0);
11054 r
= _do_read(c
.get(), oldo
, 0, oldo
->onode
.size
, bl
, 0);
11057 r
= _do_write(txc
, c
, newo
, 0, oldo
->onode
.size
, bl
, 0);
11063 newo
->onode
.attrs
= oldo
->onode
.attrs
;
11066 if (newo
->onode
.has_omap()) {
11067 dout(20) << __func__
<< " clearing old omap data" << dendl
;
11069 _do_omap_clear(txc
, newo
->onode
.nid
);
11071 if (oldo
->onode
.has_omap()) {
11072 dout(20) << __func__
<< " copying omap data" << dendl
;
11073 if (!newo
->onode
.has_omap()) {
11074 newo
->onode
.set_omap_flag();
11076 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
11078 get_omap_header(oldo
->onode
.nid
, &head
);
11079 get_omap_tail(oldo
->onode
.nid
, &tail
);
11080 it
->lower_bound(head
);
11081 while (it
->valid()) {
11082 if (it
->key() >= tail
) {
11083 dout(30) << __func__
<< " reached tail" << dendl
;
11086 dout(30) << __func__
<< " got header/data "
11087 << pretty_binary_string(it
->key()) << dendl
;
11089 rewrite_omap_key(newo
->onode
.nid
, it
->key(), &key
);
11090 txc
->t
->set(PREFIX_OMAP
, key
, it
->value());
11095 newo
->onode
.clear_omap_flag();
11098 txc
->write_onode(newo
);
11102 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11103 << newo
->oid
<< " = " << r
<< dendl
;
11107 int BlueStore::_do_clone_range(
11116 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11118 << " 0x" << std::hex
<< srcoff
<< "~" << length
<< " -> "
11119 << " 0x" << dstoff
<< "~" << length
<< std::dec
<< dendl
;
11120 oldo
->extent_map
.fault_range(db
, srcoff
, length
);
11121 newo
->extent_map
.fault_range(db
, dstoff
, length
);
11125 // hmm, this could go into an ExtentMap::dup() method.
11126 vector
<BlobRef
> id_to_blob(oldo
->extent_map
.extent_map
.size());
11127 for (auto &e
: oldo
->extent_map
.extent_map
) {
11128 e
.blob
->last_encoded_id
= -1;
11131 uint64_t end
= srcoff
+ length
;
11132 uint32_t dirty_range_begin
= 0;
11133 uint32_t dirty_range_end
= 0;
11134 bool src_dirty
= false;
11135 for (auto ep
= oldo
->extent_map
.seek_lextent(srcoff
);
11136 ep
!= oldo
->extent_map
.extent_map
.end();
11139 if (e
.logical_offset
>= end
) {
11142 dout(20) << __func__
<< " src " << e
<< dendl
;
11144 bool blob_duped
= true;
11145 if (e
.blob
->last_encoded_id
>= 0) {
11146 // blob is already duped
11147 cb
= id_to_blob
[e
.blob
->last_encoded_id
];
11148 blob_duped
= false;
11151 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
11152 // make sure it is shared
11153 if (!blob
.is_shared()) {
11154 c
->make_blob_shared(_assign_blobid(txc
), e
.blob
);
11157 dirty_range_begin
= e
.logical_offset
;
11159 assert(e
.logical_end() > 0);
11160 // -1 to exclude next potential shard
11161 dirty_range_end
= e
.logical_end() - 1;
11163 c
->load_shared_blob(e
.blob
->shared_blob
);
11166 e
.blob
->last_encoded_id
= n
;
11167 id_to_blob
[n
] = cb
;
11169 // bump the extent refs on the copied blob's extents
11170 for (auto p
: blob
.get_extents()) {
11171 if (p
.is_valid()) {
11172 e
.blob
->shared_blob
->get_ref(p
.offset
, p
.length
);
11175 txc
->write_shared_blob(e
.blob
->shared_blob
);
11176 dout(20) << __func__
<< " new " << *cb
<< dendl
;
11179 int skip_front
, skip_back
;
11180 if (e
.logical_offset
< srcoff
) {
11181 skip_front
= srcoff
- e
.logical_offset
;
11185 if (e
.logical_end() > end
) {
11186 skip_back
= e
.logical_end() - end
;
11190 Extent
*ne
= new Extent(e
.logical_offset
+ skip_front
+ dstoff
- srcoff
,
11191 e
.blob_offset
+ skip_front
,
11192 e
.length
- skip_front
- skip_back
, cb
);
11193 newo
->extent_map
.extent_map
.insert(*ne
);
11194 ne
->blob
->get_ref(c
.get(), ne
->blob_offset
, ne
->length
);
11195 // fixme: we may leave parts of new blob unreferenced that could
11196 // be freed (relative to the shared_blob).
11197 txc
->statfs_delta
.stored() += ne
->length
;
11198 if (e
.blob
->get_blob().is_compressed()) {
11199 txc
->statfs_delta
.compressed_original() += ne
->length
;
11201 txc
->statfs_delta
.compressed() +=
11202 cb
->get_blob().get_compressed_payload_length();
11205 dout(20) << __func__
<< " dst " << *ne
<< dendl
;
11209 oldo
->extent_map
.dirty_range(dirty_range_begin
,
11210 dirty_range_end
- dirty_range_begin
);
11211 txc
->write_onode(oldo
);
11213 txc
->write_onode(newo
);
11215 if (dstoff
+ length
> newo
->onode
.size
) {
11216 newo
->onode
.size
= dstoff
+ length
;
11218 newo
->extent_map
.dirty_range(dstoff
, length
);
11224 int BlueStore::_clone_range(TransContext
*txc
,
11228 uint64_t srcoff
, uint64_t length
, uint64_t dstoff
)
11230 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11231 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11232 << " to offset 0x" << dstoff
<< std::dec
<< dendl
;
11235 if (srcoff
+ length
>= OBJECT_MAX_SIZE
||
11236 dstoff
+ length
>= OBJECT_MAX_SIZE
) {
11240 if (srcoff
+ length
> oldo
->onode
.size
) {
11245 _assign_nid(txc
, newo
);
11248 if (cct
->_conf
->bluestore_clone_cow
) {
11249 _do_zero(txc
, c
, newo
, dstoff
, length
);
11250 _do_clone_range(txc
, c
, oldo
, newo
, srcoff
, length
, dstoff
);
11253 r
= _do_read(c
.get(), oldo
, srcoff
, length
, bl
, 0);
11256 r
= _do_write(txc
, c
, newo
, dstoff
, bl
.length(), bl
, 0);
11262 txc
->write_onode(newo
);
11266 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11267 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11268 << " to offset 0x" << dstoff
<< std::dec
11269 << " = " << r
<< dendl
;
11273 int BlueStore::_rename(TransContext
*txc
,
11277 const ghobject_t
& new_oid
)
11279 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11280 << new_oid
<< dendl
;
11282 ghobject_t old_oid
= oldo
->oid
;
11283 mempool::bluestore_cache_other::string new_okey
;
11286 if (newo
->exists
) {
11290 assert(txc
->onodes
.count(newo
) == 0);
11293 txc
->t
->rmkey(PREFIX_OBJ
, oldo
->key
.c_str(), oldo
->key
.size());
11297 oldo
->extent_map
.fault_range(db
, 0, oldo
->onode
.size
);
11298 get_object_key(cct
, new_oid
, &new_okey
);
11300 for (auto &s
: oldo
->extent_map
.shards
) {
11301 generate_extent_shard_key_and_apply(oldo
->key
, s
.shard_info
->offset
, &key
,
11302 [&](const string
& final_key
) {
11303 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
11311 txc
->write_onode(newo
);
11313 // this adjusts oldo->{oid,key}, and reset oldo to a fresh empty
11314 // Onode in the old slot
11315 c
->onode_map
.rename(oldo
, old_oid
, new_oid
, new_okey
);
11319 dout(10) << __func__
<< " " << c
->cid
<< " " << old_oid
<< " -> "
11320 << new_oid
<< " = " << r
<< dendl
;
11326 int BlueStore::_create_collection(
11332 dout(15) << __func__
<< " " << cid
<< " bits " << bits
<< dendl
;
11337 RWLock::WLocker
l(coll_lock
);
11345 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
11347 (*c
)->cnode
.bits
= bits
;
11348 coll_map
[cid
] = *c
;
11350 ::encode((*c
)->cnode
, bl
);
11351 txc
->t
->set(PREFIX_COLL
, stringify(cid
), bl
);
11355 dout(10) << __func__
<< " " << cid
<< " bits " << bits
<< " = " << r
<< dendl
;
11359 int BlueStore::_remove_collection(TransContext
*txc
, const coll_t
&cid
,
11362 dout(15) << __func__
<< " " << cid
<< dendl
;
11366 RWLock::WLocker
l(coll_lock
);
11371 size_t nonexistent_count
= 0;
11372 assert((*c
)->exists
);
11373 if ((*c
)->onode_map
.map_any([&](OnodeRef o
) {
11375 dout(10) << __func__
<< " " << o
->oid
<< " " << o
11376 << " exists in onode_map" << dendl
;
11379 ++nonexistent_count
;
11386 vector
<ghobject_t
> ls
;
11388 // Enumerate onodes in db, up to nonexistent_count + 1
11389 // then check if all of them are marked as non-existent.
11390 // Bypass the check if returned number is greater than nonexistent_count
11391 r
= _collection_list(c
->get(), ghobject_t(), ghobject_t::get_max(),
11392 nonexistent_count
+ 1, &ls
, &next
);
11394 bool exists
= false; //ls.size() > nonexistent_count;
11395 for (auto it
= ls
.begin(); !exists
&& it
< ls
.end(); ++it
) {
11396 dout(10) << __func__
<< " oid " << *it
<< dendl
;
11397 auto onode
= (*c
)->onode_map
.lookup(*it
);
11398 exists
= !onode
|| onode
->exists
;
11400 dout(10) << __func__
<< " " << *it
11401 << " exists in db" << dendl
;
11405 coll_map
.erase(cid
);
11406 txc
->removed_collections
.push_back(*c
);
11407 (*c
)->exists
= false;
11409 txc
->t
->rmkey(PREFIX_COLL
, stringify(cid
));
11412 dout(10) << __func__
<< " " << cid
11413 << " is non-empty" << dendl
;
11420 dout(10) << __func__
<< " " << cid
<< " = " << r
<< dendl
;
11424 int BlueStore::_split_collection(TransContext
*txc
,
11427 unsigned bits
, int rem
)
11429 dout(15) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11430 << " bits " << bits
<< dendl
;
11431 RWLock::WLocker
l(c
->lock
);
11432 RWLock::WLocker
l2(d
->lock
);
11435 // flush all previous deferred writes on this sequencer. this is a bit
11436 // heavyweight, but we need to make sure all deferred writes complete
11437 // before we split as the new collection's sequencer may need to order
11438 // this after those writes, and we don't bother with the complexity of
11439 // moving those TransContexts over to the new osr.
11440 _osr_drain_preceding(txc
);
11442 // move any cached items (onodes and referenced shared blobs) that will
11443 // belong to the child collection post-split. leave everything else behind.
11444 // this may include things that don't strictly belong to the now-smaller
11445 // parent split, but the OSD will always send us a split for every new
11448 spg_t pgid
, dest_pgid
;
11449 bool is_pg
= c
->cid
.is_pg(&pgid
);
11451 is_pg
= d
->cid
.is_pg(&dest_pgid
);
11454 // the destination should initially be empty.
11455 assert(d
->onode_map
.empty());
11456 assert(d
->shared_blob_set
.empty());
11457 assert(d
->cnode
.bits
== bits
);
11459 c
->split_cache(d
.get());
11461 // adjust bits. note that this will be redundant for all but the first
11462 // split call for this parent (first child).
11463 c
->cnode
.bits
= bits
;
11464 assert(d
->cnode
.bits
== bits
);
11468 ::encode(c
->cnode
, bl
);
11469 txc
->t
->set(PREFIX_COLL
, stringify(c
->cid
), bl
);
11471 dout(10) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11472 << " bits " << bits
<< " = " << r
<< dendl
;
11476 // DB key value Histogram
11477 #define KEY_SLAB 32
11478 #define VALUE_SLAB 64
11480 const string prefix_onode
= "o";
11481 const string prefix_onode_shard
= "x";
11482 const string prefix_other
= "Z";
11484 int BlueStore::DBHistogram::get_key_slab(size_t sz
)
11486 return (sz
/KEY_SLAB
);
11489 string
BlueStore::DBHistogram::get_key_slab_to_range(int slab
)
11491 int lower_bound
= slab
* KEY_SLAB
;
11492 int upper_bound
= (slab
+ 1) * KEY_SLAB
;
11493 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11497 int BlueStore::DBHistogram::get_value_slab(size_t sz
)
11499 return (sz
/VALUE_SLAB
);
11502 string
BlueStore::DBHistogram::get_value_slab_to_range(int slab
)
11504 int lower_bound
= slab
* VALUE_SLAB
;
11505 int upper_bound
= (slab
+ 1) * VALUE_SLAB
;
11506 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11510 void BlueStore::DBHistogram::update_hist_entry(map
<string
, map
<int, struct key_dist
> > &key_hist
,
11511 const string
&prefix
, size_t key_size
, size_t value_size
)
11513 uint32_t key_slab
= get_key_slab(key_size
);
11514 uint32_t value_slab
= get_value_slab(value_size
);
11515 key_hist
[prefix
][key_slab
].count
++;
11516 key_hist
[prefix
][key_slab
].max_len
= MAX(key_size
, key_hist
[prefix
][key_slab
].max_len
);
11517 key_hist
[prefix
][key_slab
].val_map
[value_slab
].count
++;
11518 key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
=
11519 MAX(value_size
, key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
);
11522 void BlueStore::DBHistogram::dump(Formatter
*f
)
11524 f
->open_object_section("rocksdb_value_distribution");
11525 for (auto i
: value_hist
) {
11526 f
->dump_unsigned(get_value_slab_to_range(i
.first
).data(), i
.second
);
11528 f
->close_section();
11530 f
->open_object_section("rocksdb_key_value_histogram");
11531 for (auto i
: key_hist
) {
11532 f
->dump_string("prefix", i
.first
);
11533 f
->open_object_section("key_hist");
11534 for ( auto k
: i
.second
) {
11535 f
->dump_unsigned(get_key_slab_to_range(k
.first
).data(), k
.second
.count
);
11536 f
->dump_unsigned("max_len", k
.second
.max_len
);
11537 f
->open_object_section("value_hist");
11538 for ( auto j
: k
.second
.val_map
) {
11539 f
->dump_unsigned(get_value_slab_to_range(j
.first
).data(), j
.second
.count
);
11540 f
->dump_unsigned("max_len", j
.second
.max_len
);
11542 f
->close_section();
11544 f
->close_section();
11546 f
->close_section();
11549 //Itrerates through the db and collects the stats
11550 void BlueStore::generate_db_histogram(Formatter
*f
)
11553 uint64_t num_onodes
= 0;
11554 uint64_t num_shards
= 0;
11555 uint64_t num_super
= 0;
11556 uint64_t num_coll
= 0;
11557 uint64_t num_omap
= 0;
11558 uint64_t num_deferred
= 0;
11559 uint64_t num_alloc
= 0;
11560 uint64_t num_stat
= 0;
11561 uint64_t num_others
= 0;
11562 uint64_t num_shared_shards
= 0;
11563 size_t max_key_size
=0, max_value_size
= 0;
11564 uint64_t total_key_size
= 0, total_value_size
= 0;
11565 size_t key_size
= 0, value_size
= 0;
11568 utime_t start
= ceph_clock_now();
11570 KeyValueDB::WholeSpaceIterator iter
= db
->get_iterator();
11571 iter
->seek_to_first();
11572 while (iter
->valid()) {
11573 dout(30) << __func__
<< " Key: " << iter
->key() << dendl
;
11574 key_size
= iter
->key_size();
11575 value_size
= iter
->value_size();
11576 hist
.value_hist
[hist
.get_value_slab(value_size
)]++;
11577 max_key_size
= MAX(max_key_size
, key_size
);
11578 max_value_size
= MAX(max_value_size
, value_size
);
11579 total_key_size
+= key_size
;
11580 total_value_size
+= value_size
;
11582 pair
<string
,string
> key(iter
->raw_key());
11584 if (key
.first
== PREFIX_SUPER
) {
11585 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SUPER
, key_size
, value_size
);
11587 } else if (key
.first
== PREFIX_STAT
) {
11588 hist
.update_hist_entry(hist
.key_hist
, PREFIX_STAT
, key_size
, value_size
);
11590 } else if (key
.first
== PREFIX_COLL
) {
11591 hist
.update_hist_entry(hist
.key_hist
, PREFIX_COLL
, key_size
, value_size
);
11593 } else if (key
.first
== PREFIX_OBJ
) {
11594 if (key
.second
.back() == ONODE_KEY_SUFFIX
) {
11595 hist
.update_hist_entry(hist
.key_hist
, prefix_onode
, key_size
, value_size
);
11598 hist
.update_hist_entry(hist
.key_hist
, prefix_onode_shard
, key_size
, value_size
);
11601 } else if (key
.first
== PREFIX_OMAP
) {
11602 hist
.update_hist_entry(hist
.key_hist
, PREFIX_OMAP
, key_size
, value_size
);
11604 } else if (key
.first
== PREFIX_DEFERRED
) {
11605 hist
.update_hist_entry(hist
.key_hist
, PREFIX_DEFERRED
, key_size
, value_size
);
11607 } else if (key
.first
== PREFIX_ALLOC
|| key
.first
== "b" ) {
11608 hist
.update_hist_entry(hist
.key_hist
, PREFIX_ALLOC
, key_size
, value_size
);
11610 } else if (key
.first
== PREFIX_SHARED_BLOB
) {
11611 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SHARED_BLOB
, key_size
, value_size
);
11612 num_shared_shards
++;
11614 hist
.update_hist_entry(hist
.key_hist
, prefix_other
, key_size
, value_size
);
11620 utime_t duration
= ceph_clock_now() - start
;
11621 f
->open_object_section("rocksdb_key_value_stats");
11622 f
->dump_unsigned("num_onodes", num_onodes
);
11623 f
->dump_unsigned("num_shards", num_shards
);
11624 f
->dump_unsigned("num_super", num_super
);
11625 f
->dump_unsigned("num_coll", num_coll
);
11626 f
->dump_unsigned("num_omap", num_omap
);
11627 f
->dump_unsigned("num_deferred", num_deferred
);
11628 f
->dump_unsigned("num_alloc", num_alloc
);
11629 f
->dump_unsigned("num_stat", num_stat
);
11630 f
->dump_unsigned("num_shared_shards", num_shared_shards
);
11631 f
->dump_unsigned("num_others", num_others
);
11632 f
->dump_unsigned("max_key_size", max_key_size
);
11633 f
->dump_unsigned("max_value_size", max_value_size
);
11634 f
->dump_unsigned("total_key_size", total_key_size
);
11635 f
->dump_unsigned("total_value_size", total_value_size
);
11636 f
->close_section();
11640 dout(20) << __func__
<< " finished in " << duration
<< " seconds" << dendl
;
11644 void BlueStore::_flush_cache()
11646 dout(10) << __func__
<< dendl
;
11647 for (auto i
: cache_shards
) {
11649 assert(i
->empty());
11651 for (auto& p
: coll_map
) {
11652 if (!p
.second
->onode_map
.empty()) {
11653 derr
<< __func__
<< "stray onodes on " << p
.first
<< dendl
;
11654 p
.second
->onode_map
.dump(cct
, 0);
11656 if (!p
.second
->shared_blob_set
.empty()) {
11657 derr
<< __func__
<< " stray shared blobs on " << p
.first
<< dendl
;
11658 p
.second
->shared_blob_set
.dump(cct
, 0);
11660 assert(p
.second
->onode_map
.empty());
11661 assert(p
.second
->shared_blob_set
.empty());
11666 // For external caller.
11667 // We use a best-effort policy instead, e.g.,
11668 // we don't care if there are still some pinned onodes/data in the cache
11669 // after this command is completed.
11670 void BlueStore::flush_cache()
11672 dout(10) << __func__
<< dendl
;
11673 for (auto i
: cache_shards
) {
11678 void BlueStore::_apply_padding(uint64_t head_pad
,
11680 bufferlist
& padded
)
11683 padded
.prepend_zero(head_pad
);
11686 padded
.append_zero(tail_pad
);
11688 if (head_pad
|| tail_pad
) {
11689 dout(20) << __func__
<< " can pad head 0x" << std::hex
<< head_pad
11690 << " tail 0x" << tail_pad
<< std::dec
<< dendl
;
11691 logger
->inc(l_bluestore_write_pad_bytes
, head_pad
+ tail_pad
);
11695 // ===========================================