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
) {
1622 #define dout_prefix *_dout << "bluestore.sharedblob(" << this << ") "
1624 ostream
& operator<<(ostream
& out
, const BlueStore::SharedBlob
& sb
)
1626 out
<< "SharedBlob(" << &sb
;
1629 out
<< " loaded " << *sb
.persistent
;
1631 out
<< " sbid 0x" << std::hex
<< sb
.sbid_unloaded
<< std::dec
;
1636 BlueStore::SharedBlob::SharedBlob(uint64_t i
, Collection
*_coll
)
1637 : coll(_coll
), sbid_unloaded(i
)
1639 assert(sbid_unloaded
> 0);
1641 get_cache()->add_blob();
1645 BlueStore::SharedBlob::~SharedBlob()
1647 if (get_cache()) { // the dummy instances have a nullptr
1648 std::lock_guard
<std::recursive_mutex
> l(get_cache()->lock
);
1649 bc
._clear(get_cache());
1650 get_cache()->rm_blob();
1652 if (loaded
&& persistent
) {
1657 void BlueStore::SharedBlob::put()
1660 ldout(coll
->store
->cct
, 20) << __func__
<< " " << this
1661 << " removing self from set " << get_parent()
1664 if (get_parent()->remove(this)) {
1667 ldout(coll
->store
->cct
, 20)
1668 << __func__
<< " " << this << " lost race to remove myself from set"
1677 void BlueStore::SharedBlob::get_ref(uint64_t offset
, uint32_t length
)
1680 persistent
->ref_map
.get(offset
, length
);
1683 void BlueStore::SharedBlob::put_ref(uint64_t offset
, uint32_t length
,
1685 set
<SharedBlob
*> *maybe_unshared
)
1689 persistent
->ref_map
.put(offset
, length
, r
, maybe_unshared
? &maybe
: nullptr);
1690 if (maybe_unshared
&& maybe
) {
1691 maybe_unshared
->insert(this);
1698 #define dout_prefix *_dout << "bluestore.blob(" << this << ") "
1700 ostream
& operator<<(ostream
& out
, const BlueStore::Blob
& b
)
1702 out
<< "Blob(" << &b
;
1703 if (b
.is_spanning()) {
1704 out
<< " spanning " << b
.id
;
1706 out
<< " " << b
.get_blob() << " " << b
.get_blob_use_tracker();
1707 if (b
.shared_blob
) {
1708 out
<< " " << *b
.shared_blob
;
1710 out
<< " (shared_blob=NULL)";
1716 void BlueStore::Blob::discard_unallocated(Collection
*coll
)
1718 if (get_blob().is_shared()) {
1721 if (get_blob().is_compressed()) {
1722 bool discard
= false;
1723 bool all_invalid
= true;
1724 for (auto e
: get_blob().get_extents()) {
1725 if (!e
.is_valid()) {
1728 all_invalid
= false;
1731 assert(discard
== all_invalid
); // in case of compressed blob all
1732 // or none pextents are invalid.
1734 shared_blob
->bc
.discard(shared_blob
->get_cache(), 0,
1735 get_blob().get_logical_length());
1739 for (auto e
: get_blob().get_extents()) {
1740 if (!e
.is_valid()) {
1741 ldout(coll
->store
->cct
, 20) << __func__
<< " 0x" << std::hex
<< pos
1743 << std::dec
<< dendl
;
1744 shared_blob
->bc
.discard(shared_blob
->get_cache(), pos
, e
.length
);
1748 if (get_blob().can_prune_tail()) {
1749 dirty_blob().prune_tail();
1750 used_in_blob
.prune_tail(get_blob().get_ondisk_length());
1751 auto cct
= coll
->store
->cct
; //used by dout
1752 dout(20) << __func__
<< " pruned tail, now " << get_blob() << dendl
;
1757 void BlueStore::Blob::get_ref(
1762 // Caller has to initialize Blob's logical length prior to increment
1763 // references. Otherwise one is neither unable to determine required
1764 // amount of counters in case of per-au tracking nor obtain min_release_size
1765 // for single counter mode.
1766 assert(get_blob().get_logical_length() != 0);
1767 auto cct
= coll
->store
->cct
;
1768 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1769 << std::dec
<< " " << *this << dendl
;
1771 if (used_in_blob
.is_empty()) {
1772 uint32_t min_release_size
=
1773 get_blob().get_release_size(coll
->store
->min_alloc_size
);
1774 uint64_t l
= get_blob().get_logical_length();
1775 dout(20) << __func__
<< " init 0x" << std::hex
<< l
<< ", "
1776 << min_release_size
<< std::dec
<< dendl
;
1777 used_in_blob
.init(l
, min_release_size
);
1784 bool BlueStore::Blob::put_ref(
1790 PExtentVector logical
;
1792 auto cct
= coll
->store
->cct
;
1793 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1794 << std::dec
<< " " << *this << dendl
;
1796 bool empty
= used_in_blob
.put(
1801 // nothing to release
1802 if (!empty
&& logical
.empty()) {
1806 bluestore_blob_t
& b
= dirty_blob();
1807 return b
.release_extents(empty
, logical
, r
);
1810 bool BlueStore::Blob::can_reuse_blob(uint32_t min_alloc_size
,
1811 uint32_t target_blob_size
,
1813 uint32_t *length0
) {
1814 assert(min_alloc_size
);
1815 assert(target_blob_size
);
1816 if (!get_blob().is_mutable()) {
1820 uint32_t length
= *length0
;
1821 uint32_t end
= b_offset
+ length
;
1823 // Currently for the sake of simplicity we omit blob reuse if data is
1824 // unaligned with csum chunk. Later we can perform padding if needed.
1825 if (get_blob().has_csum() &&
1826 ((b_offset
% get_blob().get_csum_chunk_size()) != 0 ||
1827 (end
% get_blob().get_csum_chunk_size()) != 0)) {
1831 auto blen
= get_blob().get_logical_length();
1832 uint32_t new_blen
= blen
;
1834 // make sure target_blob_size isn't less than current blob len
1835 target_blob_size
= MAX(blen
, target_blob_size
);
1837 if (b_offset
>= blen
) {
1838 // new data totally stands out of the existing blob
1841 // new data overlaps with the existing blob
1842 new_blen
= MAX(blen
, end
);
1844 uint32_t overlap
= 0;
1845 if (new_blen
> blen
) {
1846 overlap
= blen
- b_offset
;
1851 if (!get_blob().is_unallocated(b_offset
, overlap
)) {
1852 // abort if any piece of the overlap has already been allocated
1857 if (new_blen
> blen
) {
1858 int64_t overflow
= int64_t(new_blen
) - target_blob_size
;
1859 // Unable to decrease the provided length to fit into max_blob_size
1860 if (overflow
>= length
) {
1864 // FIXME: in some cases we could reduce unused resolution
1865 if (get_blob().has_unused()) {
1870 new_blen
-= overflow
;
1875 if (new_blen
> blen
) {
1876 dirty_blob().add_tail(new_blen
);
1877 used_in_blob
.add_tail(new_blen
,
1878 get_blob().get_release_size(min_alloc_size
));
1884 void BlueStore::Blob::split(Collection
*coll
, uint32_t blob_offset
, Blob
*r
)
1886 auto cct
= coll
->store
->cct
; //used by dout
1887 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1888 << " start " << *this << dendl
;
1889 assert(blob
.can_split());
1890 assert(used_in_blob
.can_split());
1891 bluestore_blob_t
&lb
= dirty_blob();
1892 bluestore_blob_t
&rb
= r
->dirty_blob();
1896 &(r
->used_in_blob
));
1898 lb
.split(blob_offset
, rb
);
1899 shared_blob
->bc
.split(shared_blob
->get_cache(), blob_offset
, r
->shared_blob
->bc
);
1901 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1902 << " finish " << *this << dendl
;
1903 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1904 << " and " << *r
<< dendl
;
1907 #ifndef CACHE_BLOB_BL
1908 void BlueStore::Blob::decode(
1910 bufferptr::iterator
& p
,
1913 bool include_ref_map
)
1915 denc(blob
, p
, struct_v
);
1916 if (blob
.is_shared()) {
1919 if (include_ref_map
) {
1921 used_in_blob
.decode(p
);
1923 used_in_blob
.clear();
1924 bluestore_extent_ref_map_t legacy_ref_map
;
1925 legacy_ref_map
.decode(p
);
1926 for (auto r
: legacy_ref_map
.ref_map
) {
1930 r
.second
.refs
* r
.second
.length
);
1939 ostream
& operator<<(ostream
& out
, const BlueStore::Extent
& e
)
1941 return out
<< std::hex
<< "0x" << e
.logical_offset
<< "~" << e
.length
1942 << ": 0x" << e
.blob_offset
<< "~" << e
.length
<< std::dec
1947 BlueStore::OldExtent
* BlueStore::OldExtent::create(CollectionRef c
,
1952 OldExtent
* oe
= new OldExtent(lo
, o
, l
, b
);
1953 b
->put_ref(c
.get(), o
, l
, &(oe
->r
));
1954 oe
->blob_empty
= b
->get_referenced_bytes() == 0;
1961 #define dout_prefix *_dout << "bluestore.extentmap(" << this << ") "
1963 BlueStore::ExtentMap::ExtentMap(Onode
*o
)
1966 o
->c
->store
->cct
->_conf
->bluestore_extent_map_inline_shard_prealloc_size
) {
1969 void BlueStore::ExtentMap::update(KeyValueDB::Transaction t
,
1972 auto cct
= onode
->c
->store
->cct
; //used by dout
1973 dout(20) << __func__
<< " " << onode
->oid
<< (force
? " force" : "") << dendl
;
1974 if (onode
->onode
.extent_map_shards
.empty()) {
1975 if (inline_bl
.length() == 0) {
1977 // we need to encode inline_bl to measure encoded length
1978 bool never_happen
= encode_some(0, OBJECT_MAX_SIZE
, inline_bl
, &n
);
1979 assert(!never_happen
);
1980 size_t len
= inline_bl
.length();
1981 dout(20) << __func__
<< " inline shard " << len
<< " bytes from " << n
1982 << " extents" << dendl
;
1983 if (!force
&& len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
1984 request_reshard(0, OBJECT_MAX_SIZE
);
1988 // will persist in the onode key.
1990 // pending shard update
1991 struct dirty_shard_t
{
1994 dirty_shard_t(Shard
*s
) : shard(s
) {}
1996 vector
<dirty_shard_t
> encoded_shards
;
1997 // allocate slots for all shards in a single call instead of
1998 // doing multiple allocations - one per each dirty shard
1999 encoded_shards
.reserve(shards
.size());
2001 auto p
= shards
.begin();
2003 while (p
!= shards
.end()) {
2004 assert(p
->shard_info
->offset
>= prev_p
->shard_info
->offset
);
2009 if (n
== shards
.end()) {
2010 endoff
= OBJECT_MAX_SIZE
;
2012 endoff
= n
->shard_info
->offset
;
2014 encoded_shards
.emplace_back(dirty_shard_t(&(*p
)));
2015 bufferlist
& bl
= encoded_shards
.back().bl
;
2016 if (encode_some(p
->shard_info
->offset
, endoff
- p
->shard_info
->offset
,
2019 derr
<< __func__
<< " encode_some needs reshard" << dendl
;
2023 size_t len
= bl
.length();
2025 dout(20) << __func__
<< " shard 0x" << std::hex
2026 << p
->shard_info
->offset
<< std::dec
<< " is " << len
2027 << " bytes (was " << p
->shard_info
->bytes
<< ") from "
2028 << p
->extents
<< " extents" << dendl
;
2031 if (len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
2032 // we are big; reshard ourselves
2033 request_reshard(p
->shard_info
->offset
, endoff
);
2035 // avoid resharding the trailing shard, even if it is small
2036 else if (n
!= shards
.end() &&
2037 len
< g_conf
->bluestore_extent_map_shard_min_size
) {
2038 assert(endoff
!= OBJECT_MAX_SIZE
);
2039 if (p
== shards
.begin()) {
2040 // we are the first shard, combine with next shard
2041 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2043 // combine either with the previous shard or the next,
2044 // whichever is smaller
2045 if (prev_p
->shard_info
->bytes
> n
->shard_info
->bytes
) {
2046 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2048 request_reshard(prev_p
->shard_info
->offset
, endoff
);
2057 if (needs_reshard()) {
2061 // schedule DB update for dirty shards
2063 for (auto& it
: encoded_shards
) {
2064 it
.shard
->dirty
= false;
2065 it
.shard
->shard_info
->bytes
= it
.bl
.length();
2066 generate_extent_shard_key_and_apply(
2068 it
.shard
->shard_info
->offset
,
2070 [&](const string
& final_key
) {
2071 t
->set(PREFIX_OBJ
, final_key
, it
.bl
);
2078 bid_t
BlueStore::ExtentMap::allocate_spanning_blob_id()
2080 if (spanning_blob_map
.empty())
2082 bid_t bid
= spanning_blob_map
.rbegin()->first
+ 1;
2083 // bid is valid and available.
2086 // Find next unused bid;
2087 bid
= rand() % (numeric_limits
<bid_t
>::max() + 1);
2088 const auto begin_bid
= bid
;
2090 if (!spanning_blob_map
.count(bid
))
2094 if (bid
< 0) bid
= 0;
2096 } while (bid
!= begin_bid
);
2097 assert(0 == "no available blob id");
2100 void BlueStore::ExtentMap::reshard(
2102 KeyValueDB::Transaction t
)
2104 auto cct
= onode
->c
->store
->cct
; // used by dout
2106 dout(10) << __func__
<< " 0x[" << std::hex
<< needs_reshard_begin
<< ","
2107 << needs_reshard_end
<< ")" << std::dec
2108 << " of " << onode
->onode
.extent_map_shards
.size()
2109 << " shards on " << onode
->oid
<< dendl
;
2110 for (auto& p
: spanning_blob_map
) {
2111 dout(20) << __func__
<< " spanning blob " << p
.first
<< " " << *p
.second
2114 // determine shard index range
2115 unsigned si_begin
= 0, si_end
= 0;
2116 if (!shards
.empty()) {
2117 while (si_begin
+ 1 < shards
.size() &&
2118 shards
[si_begin
+ 1].shard_info
->offset
<= needs_reshard_begin
) {
2121 needs_reshard_begin
= shards
[si_begin
].shard_info
->offset
;
2122 for (si_end
= si_begin
; si_end
< shards
.size(); ++si_end
) {
2123 if (shards
[si_end
].shard_info
->offset
>= needs_reshard_end
) {
2124 needs_reshard_end
= shards
[si_end
].shard_info
->offset
;
2128 if (si_end
== shards
.size()) {
2129 needs_reshard_end
= OBJECT_MAX_SIZE
;
2131 dout(20) << __func__
<< " shards [" << si_begin
<< "," << si_end
<< ")"
2132 << " over 0x[" << std::hex
<< needs_reshard_begin
<< ","
2133 << needs_reshard_end
<< ")" << std::dec
<< dendl
;
2136 fault_range(db
, needs_reshard_begin
, needs_reshard_end
);
2138 // we may need to fault in a larger interval later must have all
2139 // referring extents for spanning blobs loaded in order to have
2140 // accurate use_tracker values.
2141 uint32_t spanning_scan_begin
= needs_reshard_begin
;
2142 uint32_t spanning_scan_end
= needs_reshard_end
;
2146 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2147 generate_extent_shard_key_and_apply(
2148 onode
->key
, shards
[i
].shard_info
->offset
, &key
,
2149 [&](const string
& final_key
) {
2150 t
->rmkey(PREFIX_OBJ
, final_key
);
2155 // calculate average extent size
2157 unsigned extents
= 0;
2158 if (onode
->onode
.extent_map_shards
.empty()) {
2159 bytes
= inline_bl
.length();
2160 extents
= extent_map
.size();
2162 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2163 bytes
+= shards
[i
].shard_info
->bytes
;
2164 extents
+= shards
[i
].extents
;
2167 unsigned target
= cct
->_conf
->bluestore_extent_map_shard_target_size
;
2168 unsigned slop
= target
*
2169 cct
->_conf
->bluestore_extent_map_shard_target_size_slop
;
2170 unsigned extent_avg
= bytes
/ MAX(1, extents
);
2171 dout(20) << __func__
<< " extent_avg " << extent_avg
<< ", target " << target
2172 << ", slop " << slop
<< dendl
;
2175 unsigned estimate
= 0;
2176 unsigned offset
= needs_reshard_begin
;
2177 vector
<bluestore_onode_t::shard_info
> new_shard_info
;
2178 unsigned max_blob_end
= 0;
2179 Extent
dummy(needs_reshard_begin
);
2180 for (auto e
= extent_map
.lower_bound(dummy
);
2181 e
!= extent_map
.end();
2183 if (e
->logical_offset
>= needs_reshard_end
) {
2186 dout(30) << " extent " << *e
<< dendl
;
2188 // disfavor shard boundaries that span a blob
2189 bool would_span
= (e
->logical_offset
< max_blob_end
) || e
->blob_offset
;
2191 estimate
+ extent_avg
> target
+ (would_span
? slop
: 0)) {
2193 if (offset
== needs_reshard_begin
) {
2194 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2195 new_shard_info
.back().offset
= offset
;
2196 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2197 << std::dec
<< dendl
;
2199 offset
= e
->logical_offset
;
2200 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2201 new_shard_info
.back().offset
= offset
;
2202 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2203 << std::dec
<< dendl
;
2206 estimate
+= extent_avg
;
2207 unsigned bs
= e
->blob_start();
2208 if (bs
< spanning_scan_begin
) {
2209 spanning_scan_begin
= bs
;
2211 uint32_t be
= e
->blob_end();
2212 if (be
> max_blob_end
) {
2215 if (be
> spanning_scan_end
) {
2216 spanning_scan_end
= be
;
2219 if (new_shard_info
.empty() && (si_begin
> 0 ||
2220 si_end
< shards
.size())) {
2221 // we resharded a partial range; we must produce at least one output
2223 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2224 new_shard_info
.back().offset
= needs_reshard_begin
;
2225 dout(20) << __func__
<< " new shard 0x" << std::hex
<< needs_reshard_begin
2226 << std::dec
<< " (singleton degenerate case)" << dendl
;
2229 auto& sv
= onode
->onode
.extent_map_shards
;
2230 dout(20) << __func__
<< " new " << new_shard_info
<< dendl
;
2231 dout(20) << __func__
<< " old " << sv
<< dendl
;
2233 // no old shards to keep
2234 sv
.swap(new_shard_info
);
2235 init_shards(true, true);
2237 // splice in new shards
2238 sv
.erase(sv
.begin() + si_begin
, sv
.begin() + si_end
);
2239 shards
.erase(shards
.begin() + si_begin
, shards
.begin() + si_end
);
2241 sv
.begin() + si_begin
,
2242 new_shard_info
.begin(),
2243 new_shard_info
.end());
2244 shards
.insert(shards
.begin() + si_begin
, new_shard_info
.size(), Shard());
2245 si_end
= si_begin
+ new_shard_info
.size();
2247 assert(sv
.size() == shards
.size());
2249 // note that we need to update every shard_info of shards here,
2250 // as sv might have been totally re-allocated above
2251 for (unsigned i
= 0; i
< shards
.size(); i
++) {
2252 shards
[i
].shard_info
= &sv
[i
];
2255 // mark newly added shards as dirty
2256 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2257 shards
[i
].loaded
= true;
2258 shards
[i
].dirty
= true;
2261 dout(20) << __func__
<< " fin " << sv
<< dendl
;
2265 // no more shards; unspan all previously spanning blobs
2266 auto p
= spanning_blob_map
.begin();
2267 while (p
!= spanning_blob_map
.end()) {
2269 dout(30) << __func__
<< " un-spanning " << *p
->second
<< dendl
;
2270 p
= spanning_blob_map
.erase(p
);
2273 // identify new spanning blobs
2274 dout(20) << __func__
<< " checking spanning blobs 0x[" << std::hex
2275 << spanning_scan_begin
<< "," << spanning_scan_end
<< ")" << dendl
;
2276 if (spanning_scan_begin
< needs_reshard_begin
) {
2277 fault_range(db
, spanning_scan_begin
,
2278 needs_reshard_begin
- spanning_scan_begin
);
2280 if (spanning_scan_end
> needs_reshard_end
) {
2281 fault_range(db
, needs_reshard_end
,
2282 spanning_scan_end
- needs_reshard_end
);
2284 auto sp
= sv
.begin() + si_begin
;
2285 auto esp
= sv
.end();
2286 unsigned shard_start
= sp
->offset
;
2290 shard_end
= OBJECT_MAX_SIZE
;
2292 shard_end
= sp
->offset
;
2294 Extent
dummy(needs_reshard_begin
);
2295 for (auto e
= extent_map
.lower_bound(dummy
); e
!= extent_map
.end(); ++e
) {
2296 if (e
->logical_offset
>= needs_reshard_end
) {
2299 dout(30) << " extent " << *e
<< dendl
;
2300 while (e
->logical_offset
>= shard_end
) {
2301 shard_start
= shard_end
;
2305 shard_end
= OBJECT_MAX_SIZE
;
2307 shard_end
= sp
->offset
;
2309 dout(30) << __func__
<< " shard 0x" << std::hex
<< shard_start
2310 << " to 0x" << shard_end
<< std::dec
<< dendl
;
2312 if (e
->blob_escapes_range(shard_start
, shard_end
- shard_start
)) {
2313 if (!e
->blob
->is_spanning()) {
2314 // We have two options: (1) split the blob into pieces at the
2315 // shard boundaries (and adjust extents accordingly), or (2)
2316 // mark it spanning. We prefer to cut the blob if we can. Note that
2317 // we may have to split it multiple times--potentially at every
2319 bool must_span
= false;
2320 BlobRef b
= e
->blob
;
2321 if (b
->can_split()) {
2322 uint32_t bstart
= e
->blob_start();
2323 uint32_t bend
= e
->blob_end();
2324 for (const auto& sh
: shards
) {
2325 if (bstart
< sh
.shard_info
->offset
&&
2326 bend
> sh
.shard_info
->offset
) {
2327 uint32_t blob_offset
= sh
.shard_info
->offset
- bstart
;
2328 if (b
->can_split_at(blob_offset
)) {
2329 dout(20) << __func__
<< " splitting blob, bstart 0x"
2330 << std::hex
<< bstart
<< " blob_offset 0x"
2331 << blob_offset
<< std::dec
<< " " << *b
<< dendl
;
2332 b
= split_blob(b
, blob_offset
, sh
.shard_info
->offset
);
2333 // switch b to the new right-hand side, in case it
2334 // *also* has to get split.
2335 bstart
+= blob_offset
;
2336 onode
->c
->store
->logger
->inc(l_bluestore_blob_split
);
2347 auto bid
= allocate_spanning_blob_id();
2349 spanning_blob_map
[b
->id
] = b
;
2350 dout(20) << __func__
<< " adding spanning " << *b
<< dendl
;
2354 if (e
->blob
->is_spanning()) {
2355 spanning_blob_map
.erase(e
->blob
->id
);
2357 dout(30) << __func__
<< " un-spanning " << *e
->blob
<< dendl
;
2363 clear_needs_reshard();
2366 bool BlueStore::ExtentMap::encode_some(
2372 auto cct
= onode
->c
->store
->cct
; //used by dout
2373 Extent
dummy(offset
);
2374 auto start
= extent_map
.lower_bound(dummy
);
2375 uint32_t end
= offset
+ length
;
2377 __u8 struct_v
= 2; // Version 2 differs from v1 in blob's ref_map
2378 // serialization only. Hence there is no specific
2379 // handling at ExtentMap level.
2383 bool must_reshard
= false;
2384 for (auto p
= start
;
2385 p
!= extent_map
.end() && p
->logical_offset
< end
;
2387 assert(p
->logical_offset
>= offset
);
2388 p
->blob
->last_encoded_id
= -1;
2389 if (!p
->blob
->is_spanning() && p
->blob_escapes_range(offset
, length
)) {
2390 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2391 << std::dec
<< " hit new spanning blob " << *p
<< dendl
;
2392 request_reshard(p
->blob_start(), p
->blob_end());
2393 must_reshard
= true;
2395 if (!must_reshard
) {
2396 denc_varint(0, bound
); // blobid
2397 denc_varint(0, bound
); // logical_offset
2398 denc_varint(0, bound
); // len
2399 denc_varint(0, bound
); // blob_offset
2401 p
->blob
->bound_encode(
2404 p
->blob
->shared_blob
->get_sbid(),
2412 denc(struct_v
, bound
);
2413 denc_varint(0, bound
); // number of extents
2416 auto app
= bl
.get_contiguous_appender(bound
);
2417 denc(struct_v
, app
);
2418 denc_varint(n
, app
);
2425 uint64_t prev_len
= 0;
2426 for (auto p
= start
;
2427 p
!= extent_map
.end() && p
->logical_offset
< end
;
2430 bool include_blob
= false;
2431 if (p
->blob
->is_spanning()) {
2432 blobid
= p
->blob
->id
<< BLOBID_SHIFT_BITS
;
2433 blobid
|= BLOBID_FLAG_SPANNING
;
2434 } else if (p
->blob
->last_encoded_id
< 0) {
2435 p
->blob
->last_encoded_id
= n
+ 1; // so it is always non-zero
2436 include_blob
= true;
2437 blobid
= 0; // the decoder will infer the id from n
2439 blobid
= p
->blob
->last_encoded_id
<< BLOBID_SHIFT_BITS
;
2441 if (p
->logical_offset
== pos
) {
2442 blobid
|= BLOBID_FLAG_CONTIGUOUS
;
2444 if (p
->blob_offset
== 0) {
2445 blobid
|= BLOBID_FLAG_ZEROOFFSET
;
2447 if (p
->length
== prev_len
) {
2448 blobid
|= BLOBID_FLAG_SAMELENGTH
;
2450 prev_len
= p
->length
;
2452 denc_varint(blobid
, app
);
2453 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2454 denc_varint_lowz(p
->logical_offset
- pos
, app
);
2456 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2457 denc_varint_lowz(p
->blob_offset
, app
);
2459 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2460 denc_varint_lowz(p
->length
, app
);
2462 pos
= p
->logical_end();
2464 p
->blob
->encode(app
, struct_v
, p
->blob
->shared_blob
->get_sbid(), false);
2468 /*derr << __func__ << bl << dendl;
2469 derr << __func__ << ":";
2476 unsigned BlueStore::ExtentMap::decode_some(bufferlist
& bl
)
2478 auto cct
= onode
->c
->store
->cct
; //used by dout
2480 derr << __func__ << ":";
2485 assert(bl
.get_num_buffers() <= 1);
2486 auto p
= bl
.front().begin_deep();
2489 // Version 2 differs from v1 in blob's ref_map
2490 // serialization only. Hence there is no specific
2491 // handling at ExtentMap level below.
2492 assert(struct_v
== 1 || struct_v
== 2);
2495 denc_varint(num
, p
);
2496 vector
<BlobRef
> blobs(num
);
2498 uint64_t prev_len
= 0;
2502 Extent
*le
= new Extent();
2504 denc_varint(blobid
, p
);
2505 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2507 denc_varint_lowz(gap
, p
);
2510 le
->logical_offset
= pos
;
2511 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2512 denc_varint_lowz(le
->blob_offset
, p
);
2514 le
->blob_offset
= 0;
2516 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2517 denc_varint_lowz(prev_len
, p
);
2519 le
->length
= prev_len
;
2521 if (blobid
& BLOBID_FLAG_SPANNING
) {
2522 dout(30) << __func__
<< " getting spanning blob "
2523 << (blobid
>> BLOBID_SHIFT_BITS
) << dendl
;
2524 le
->assign_blob(get_spanning_blob(blobid
>> BLOBID_SHIFT_BITS
));
2526 blobid
>>= BLOBID_SHIFT_BITS
;
2528 le
->assign_blob(blobs
[blobid
- 1]);
2531 Blob
*b
= new Blob();
2533 b
->decode(onode
->c
, p
, struct_v
, &sbid
, false);
2535 onode
->c
->open_shared_blob(sbid
, b
);
2538 // we build ref_map dynamically for non-spanning blobs
2546 extent_map
.insert(*le
);
2553 void BlueStore::ExtentMap::bound_encode_spanning_blobs(size_t& p
)
2555 // Version 2 differs from v1 in blob's ref_map
2556 // serialization only. Hence there is no specific
2557 // handling at ExtentMap level.
2561 denc_varint((uint32_t)0, p
);
2562 size_t key_size
= 0;
2563 denc_varint((uint32_t)0, key_size
);
2564 p
+= spanning_blob_map
.size() * key_size
;
2565 for (const auto& i
: spanning_blob_map
) {
2566 i
.second
->bound_encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2570 void BlueStore::ExtentMap::encode_spanning_blobs(
2571 bufferlist::contiguous_appender
& p
)
2573 // Version 2 differs from v1 in blob's ref_map
2574 // serialization only. Hence there is no specific
2575 // handling at ExtentMap level.
2579 denc_varint(spanning_blob_map
.size(), p
);
2580 for (auto& i
: spanning_blob_map
) {
2581 denc_varint(i
.second
->id
, p
);
2582 i
.second
->encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2586 void BlueStore::ExtentMap::decode_spanning_blobs(
2587 bufferptr::iterator
& p
)
2591 // Version 2 differs from v1 in blob's ref_map
2592 // serialization only. Hence there is no specific
2593 // handling at ExtentMap level.
2594 assert(struct_v
== 1 || struct_v
== 2);
2599 BlobRef
b(new Blob());
2600 denc_varint(b
->id
, p
);
2601 spanning_blob_map
[b
->id
] = b
;
2603 b
->decode(onode
->c
, p
, struct_v
, &sbid
, true);
2604 onode
->c
->open_shared_blob(sbid
, b
);
2608 void BlueStore::ExtentMap::init_shards(bool loaded
, bool dirty
)
2610 shards
.resize(onode
->onode
.extent_map_shards
.size());
2612 for (auto &s
: onode
->onode
.extent_map_shards
) {
2613 shards
[i
].shard_info
= &s
;
2614 shards
[i
].loaded
= loaded
;
2615 shards
[i
].dirty
= dirty
;
2620 void BlueStore::ExtentMap::fault_range(
2625 auto cct
= onode
->c
->store
->cct
; //used by dout
2626 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2627 << std::dec
<< dendl
;
2628 auto start
= seek_shard(offset
);
2629 auto last
= seek_shard(offset
+ length
);
2634 assert(last
>= start
);
2636 while (start
<= last
) {
2637 assert((size_t)start
< shards
.size());
2638 auto p
= &shards
[start
];
2640 dout(30) << __func__
<< " opening shard 0x" << std::hex
2641 << p
->shard_info
->offset
<< std::dec
<< dendl
;
2643 generate_extent_shard_key_and_apply(
2644 onode
->key
, p
->shard_info
->offset
, &key
,
2645 [&](const string
& final_key
) {
2646 int r
= db
->get(PREFIX_OBJ
, final_key
, &v
);
2648 derr
<< __func__
<< " missing shard 0x" << std::hex
2649 << p
->shard_info
->offset
<< std::dec
<< " for " << onode
->oid
2655 p
->extents
= decode_some(v
);
2657 dout(20) << __func__
<< " open shard 0x" << std::hex
2658 << p
->shard_info
->offset
<< std::dec
2659 << " (" << v
.length() << " bytes)" << dendl
;
2660 assert(p
->dirty
== false);
2661 assert(v
.length() == p
->shard_info
->bytes
);
2662 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_misses
);
2664 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_hits
);
2670 void BlueStore::ExtentMap::dirty_range(
2674 auto cct
= onode
->c
->store
->cct
; //used by dout
2675 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2676 << std::dec
<< dendl
;
2677 if (shards
.empty()) {
2678 dout(20) << __func__
<< " mark inline shard dirty" << dendl
;
2682 auto start
= seek_shard(offset
);
2683 auto last
= seek_shard(offset
+ length
);
2687 assert(last
>= start
);
2688 while (start
<= last
) {
2689 assert((size_t)start
< shards
.size());
2690 auto p
= &shards
[start
];
2692 dout(20) << __func__
<< " shard 0x" << std::hex
<< p
->shard_info
->offset
2693 << std::dec
<< " is not loaded, can't mark dirty" << dendl
;
2694 assert(0 == "can't mark unloaded shard dirty");
2697 dout(20) << __func__
<< " mark shard 0x" << std::hex
2698 << p
->shard_info
->offset
<< std::dec
<< " dirty" << dendl
;
2705 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::find(
2708 Extent
dummy(offset
);
2709 return extent_map
.find(dummy
);
2712 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::seek_lextent(
2715 Extent
dummy(offset
);
2716 auto fp
= extent_map
.lower_bound(dummy
);
2717 if (fp
!= extent_map
.begin()) {
2719 if (fp
->logical_end() <= offset
) {
2726 BlueStore::extent_map_t::const_iterator
BlueStore::ExtentMap::seek_lextent(
2727 uint64_t offset
) const
2729 Extent
dummy(offset
);
2730 auto fp
= extent_map
.lower_bound(dummy
);
2731 if (fp
!= extent_map
.begin()) {
2733 if (fp
->logical_end() <= offset
) {
2740 bool BlueStore::ExtentMap::has_any_lextents(uint64_t offset
, uint64_t length
)
2742 auto fp
= seek_lextent(offset
);
2743 if (fp
== extent_map
.end() || fp
->logical_offset
>= offset
+ length
) {
2749 int BlueStore::ExtentMap::compress_extent_map(
2753 auto cct
= onode
->c
->store
->cct
; //used by dout
2754 if (extent_map
.empty())
2757 auto p
= seek_lextent(offset
);
2758 if (p
!= extent_map
.begin()) {
2759 --p
; // start to the left of offset
2761 // the caller should have just written to this region
2762 assert(p
!= extent_map
.end());
2764 // identify the *next* shard
2765 auto pshard
= shards
.begin();
2766 while (pshard
!= shards
.end() &&
2767 p
->logical_offset
>= pshard
->shard_info
->offset
) {
2771 if (pshard
!= shards
.end()) {
2772 shard_end
= pshard
->shard_info
->offset
;
2774 shard_end
= OBJECT_MAX_SIZE
;
2778 for (++n
; n
!= extent_map
.end(); p
= n
++) {
2779 if (n
->logical_offset
> offset
+ length
) {
2780 break; // stop after end
2782 while (n
!= extent_map
.end() &&
2783 p
->logical_end() == n
->logical_offset
&&
2784 p
->blob
== n
->blob
&&
2785 p
->blob_offset
+ p
->length
== n
->blob_offset
&&
2786 n
->logical_offset
< shard_end
) {
2787 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2788 << " next shard 0x" << shard_end
<< std::dec
2789 << " merging " << *p
<< " and " << *n
<< dendl
;
2790 p
->length
+= n
->length
;
2794 if (n
== extent_map
.end()) {
2797 if (n
->logical_offset
>= shard_end
) {
2798 assert(pshard
!= shards
.end());
2800 if (pshard
!= shards
.end()) {
2801 shard_end
= pshard
->shard_info
->offset
;
2803 shard_end
= OBJECT_MAX_SIZE
;
2807 if (removed
&& onode
) {
2808 onode
->c
->store
->logger
->inc(l_bluestore_extent_compress
, removed
);
2813 void BlueStore::ExtentMap::punch_hole(
2817 old_extent_map_t
*old_extents
)
2819 auto p
= seek_lextent(offset
);
2820 uint64_t end
= offset
+ length
;
2821 while (p
!= extent_map
.end()) {
2822 if (p
->logical_offset
>= end
) {
2825 if (p
->logical_offset
< offset
) {
2826 if (p
->logical_end() > end
) {
2827 // split and deref middle
2828 uint64_t front
= offset
- p
->logical_offset
;
2829 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ front
,
2831 old_extents
->push_back(*oe
);
2833 p
->blob_offset
+ front
+ length
,
2834 p
->length
- front
- length
,
2840 assert(p
->logical_end() > offset
); // else seek_lextent bug
2841 uint64_t keep
= offset
- p
->logical_offset
;
2842 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ keep
,
2843 p
->length
- keep
, p
->blob
);
2844 old_extents
->push_back(*oe
);
2850 if (p
->logical_offset
+ p
->length
<= end
) {
2851 // deref whole lextent
2852 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2853 p
->length
, p
->blob
);
2854 old_extents
->push_back(*oe
);
2859 uint64_t keep
= p
->logical_end() - end
;
2860 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2861 p
->length
- keep
, p
->blob
);
2862 old_extents
->push_back(*oe
);
2864 add(end
, p
->blob_offset
+ p
->length
- keep
, keep
, p
->blob
);
2870 BlueStore::Extent
*BlueStore::ExtentMap::set_lextent(
2872 uint64_t logical_offset
,
2873 uint64_t blob_offset
, uint64_t length
, BlobRef b
,
2874 old_extent_map_t
*old_extents
)
2876 // We need to have completely initialized Blob to increment its ref counters.
2877 assert(b
->get_blob().get_logical_length() != 0);
2879 // Do get_ref prior to punch_hole to prevent from putting reused blob into
2880 // old_extents list if we overwre the blob totally
2881 // This might happen during WAL overwrite.
2882 b
->get_ref(onode
->c
, blob_offset
, length
);
2885 punch_hole(c
, logical_offset
, length
, old_extents
);
2888 Extent
*le
= new Extent(logical_offset
, blob_offset
, length
, b
);
2889 extent_map
.insert(*le
);
2890 if (spans_shard(logical_offset
, length
)) {
2891 request_reshard(logical_offset
, logical_offset
+ length
);
2896 BlueStore::BlobRef
BlueStore::ExtentMap::split_blob(
2898 uint32_t blob_offset
,
2901 auto cct
= onode
->c
->store
->cct
; //used by dout
2903 uint32_t end_pos
= pos
+ lb
->get_blob().get_logical_length() - blob_offset
;
2904 dout(20) << __func__
<< " 0x" << std::hex
<< pos
<< " end 0x" << end_pos
2905 << " blob_offset 0x" << blob_offset
<< std::dec
<< " " << *lb
2907 BlobRef rb
= onode
->c
->new_blob();
2908 lb
->split(onode
->c
, blob_offset
, rb
.get());
2910 for (auto ep
= seek_lextent(pos
);
2911 ep
!= extent_map
.end() && ep
->logical_offset
< end_pos
;
2913 if (ep
->blob
!= lb
) {
2916 if (ep
->logical_offset
< pos
) {
2918 size_t left
= pos
- ep
->logical_offset
;
2919 Extent
*ne
= new Extent(pos
, 0, ep
->length
- left
, rb
);
2920 extent_map
.insert(*ne
);
2922 dout(30) << __func__
<< " split " << *ep
<< dendl
;
2923 dout(30) << __func__
<< " to " << *ne
<< dendl
;
2926 assert(ep
->blob_offset
>= blob_offset
);
2929 ep
->blob_offset
-= blob_offset
;
2930 dout(30) << __func__
<< " adjusted " << *ep
<< dendl
;
2939 #define dout_prefix *_dout << "bluestore.onode(" << this << ")." << __func__ << " "
2941 void BlueStore::Onode::flush()
2943 if (flushing_count
.load()) {
2944 ldout(c
->store
->cct
, 20) << __func__
<< " cnt:" << flushing_count
<< dendl
;
2945 std::unique_lock
<std::mutex
> l(flush_lock
);
2946 while (flushing_count
.load()) {
2950 ldout(c
->store
->cct
, 20) << __func__
<< " done" << dendl
;
2953 // =======================================================
2956 /// Checks for writes to the same pextent within a blob
2957 bool BlueStore::WriteContext::has_conflict(
2961 uint64_t min_alloc_size
)
2963 assert((loffs
% min_alloc_size
) == 0);
2964 assert((loffs_end
% min_alloc_size
) == 0);
2965 for (auto w
: writes
) {
2967 auto loffs2
= P2ALIGN(w
.logical_offset
, min_alloc_size
);
2968 auto loffs2_end
= P2ROUNDUP(w
.logical_offset
+ w
.length0
, min_alloc_size
);
2969 if ((loffs
<= loffs2
&& loffs_end
> loffs2
) ||
2970 (loffs
>= loffs2
&& loffs
< loffs2_end
)) {
2978 // =======================================================
2982 #define dout_prefix *_dout << "bluestore.DeferredBatch(" << this << ") "
2984 void BlueStore::DeferredBatch::prepare_write(
2986 uint64_t seq
, uint64_t offset
, uint64_t length
,
2987 bufferlist::const_iterator
& blp
)
2989 _discard(cct
, offset
, length
);
2990 auto i
= iomap
.insert(make_pair(offset
, deferred_io()));
2991 assert(i
.second
); // this should be a new insertion
2992 i
.first
->second
.seq
= seq
;
2993 blp
.copy(length
, i
.first
->second
.bl
);
2994 i
.first
->second
.bl
.reassign_to_mempool(
2995 mempool::mempool_bluestore_writing_deferred
);
2996 dout(20) << __func__
<< " seq " << seq
2997 << " 0x" << std::hex
<< offset
<< "~" << length
2998 << " crc " << i
.first
->second
.bl
.crc32c(-1)
2999 << std::dec
<< dendl
;
3000 seq_bytes
[seq
] += length
;
3001 #ifdef DEBUG_DEFERRED
3006 void BlueStore::DeferredBatch::_discard(
3007 CephContext
*cct
, uint64_t offset
, uint64_t length
)
3009 generic_dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
3010 << std::dec
<< dendl
;
3011 auto p
= iomap
.lower_bound(offset
);
3012 if (p
!= iomap
.begin()) {
3014 auto end
= p
->first
+ p
->second
.bl
.length();
3017 head
.substr_of(p
->second
.bl
, 0, offset
- p
->first
);
3018 dout(20) << __func__
<< " keep head " << p
->second
.seq
3019 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3020 << " -> 0x" << head
.length() << std::dec
<< dendl
;
3021 auto i
= seq_bytes
.find(p
->second
.seq
);
3022 assert(i
!= seq_bytes
.end());
3023 if (end
> offset
+ length
) {
3025 tail
.substr_of(p
->second
.bl
, offset
+ length
- p
->first
,
3026 end
- (offset
+ length
));
3027 dout(20) << __func__
<< " keep tail " << p
->second
.seq
3028 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3029 << " -> 0x" << tail
.length() << std::dec
<< dendl
;
3030 auto &n
= iomap
[offset
+ length
];
3032 n
.seq
= p
->second
.seq
;
3033 i
->second
-= length
;
3035 i
->second
-= end
- offset
;
3037 assert(i
->second
>= 0);
3038 p
->second
.bl
.swap(head
);
3042 while (p
!= iomap
.end()) {
3043 if (p
->first
>= offset
+ length
) {
3046 auto i
= seq_bytes
.find(p
->second
.seq
);
3047 assert(i
!= seq_bytes
.end());
3048 auto end
= p
->first
+ p
->second
.bl
.length();
3049 if (end
> offset
+ length
) {
3050 unsigned drop_front
= offset
+ length
- p
->first
;
3051 unsigned keep_tail
= end
- (offset
+ length
);
3052 dout(20) << __func__
<< " truncate front " << p
->second
.seq
3053 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3054 << " drop_front 0x" << drop_front
<< " keep_tail 0x" << keep_tail
3055 << " to 0x" << (offset
+ length
) << "~" << keep_tail
3056 << std::dec
<< dendl
;
3057 auto &s
= iomap
[offset
+ length
];
3058 s
.seq
= p
->second
.seq
;
3059 s
.bl
.substr_of(p
->second
.bl
, drop_front
, keep_tail
);
3060 i
->second
-= drop_front
;
3062 dout(20) << __func__
<< " drop " << p
->second
.seq
3063 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3064 << std::dec
<< dendl
;
3065 i
->second
-= p
->second
.bl
.length();
3067 assert(i
->second
>= 0);
3072 void BlueStore::DeferredBatch::_audit(CephContext
*cct
)
3074 map
<uint64_t,int> sb
;
3075 for (auto p
: seq_bytes
) {
3076 sb
[p
.first
] = 0; // make sure we have the same set of keys
3079 for (auto& p
: iomap
) {
3080 assert(p
.first
>= pos
);
3081 sb
[p
.second
.seq
] += p
.second
.bl
.length();
3082 pos
= p
.first
+ p
.second
.bl
.length();
3084 assert(sb
== seq_bytes
);
3091 #define dout_prefix *_dout << "bluestore(" << store->path << ").collection(" << cid << " " << this << ") "
3093 BlueStore::Collection::Collection(BlueStore
*ns
, Cache
*c
, coll_t cid
)
3097 lock("BlueStore::Collection::lock", true, false),
3103 void BlueStore::Collection::open_shared_blob(uint64_t sbid
, BlobRef b
)
3105 assert(!b
->shared_blob
);
3106 const bluestore_blob_t
& blob
= b
->get_blob();
3107 if (!blob
.is_shared()) {
3108 b
->shared_blob
= new SharedBlob(this);
3112 b
->shared_blob
= shared_blob_set
.lookup(sbid
);
3113 if (b
->shared_blob
) {
3114 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3115 << std::dec
<< " had " << *b
->shared_blob
<< dendl
;
3117 b
->shared_blob
= new SharedBlob(sbid
, this);
3118 shared_blob_set
.add(this, b
->shared_blob
.get());
3119 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3120 << std::dec
<< " opened " << *b
->shared_blob
3125 void BlueStore::Collection::load_shared_blob(SharedBlobRef sb
)
3127 if (!sb
->is_loaded()) {
3131 auto sbid
= sb
->get_sbid();
3132 get_shared_blob_key(sbid
, &key
);
3133 int r
= store
->db
->get(PREFIX_SHARED_BLOB
, key
, &v
);
3135 lderr(store
->cct
) << __func__
<< " sbid 0x" << std::hex
<< sbid
3136 << std::dec
<< " not found at key "
3137 << pretty_binary_string(key
) << dendl
;
3138 assert(0 == "uh oh, missing shared_blob");
3142 sb
->persistent
= new bluestore_shared_blob_t(sbid
);
3143 bufferlist::iterator p
= v
.begin();
3144 ::decode(*(sb
->persistent
), p
);
3145 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3146 << std::dec
<< " loaded shared_blob " << *sb
<< dendl
;
3150 void BlueStore::Collection::make_blob_shared(uint64_t sbid
, BlobRef b
)
3152 ldout(store
->cct
, 10) << __func__
<< " " << *b
<< dendl
;
3153 assert(!b
->shared_blob
->is_loaded());
3156 bluestore_blob_t
& blob
= b
->dirty_blob();
3157 blob
.set_flag(bluestore_blob_t::FLAG_SHARED
);
3159 // update shared blob
3160 b
->shared_blob
->loaded
= true;
3161 b
->shared_blob
->persistent
= new bluestore_shared_blob_t(sbid
);
3162 shared_blob_set
.add(this, b
->shared_blob
.get());
3163 for (auto p
: blob
.get_extents()) {
3165 b
->shared_blob
->get_ref(
3170 ldout(store
->cct
, 20) << __func__
<< " now " << *b
<< dendl
;
3173 uint64_t BlueStore::Collection::make_blob_unshared(SharedBlob
*sb
)
3175 ldout(store
->cct
, 10) << __func__
<< " " << *sb
<< dendl
;
3176 assert(sb
->is_loaded());
3178 uint64_t sbid
= sb
->get_sbid();
3179 shared_blob_set
.remove(sb
);
3181 delete sb
->persistent
;
3182 sb
->sbid_unloaded
= 0;
3183 ldout(store
->cct
, 20) << __func__
<< " now " << *sb
<< dendl
;
3187 BlueStore::OnodeRef
BlueStore::Collection::get_onode(
3188 const ghobject_t
& oid
,
3191 assert(create
? lock
.is_wlocked() : lock
.is_locked());
3194 if (cid
.is_pg(&pgid
)) {
3195 if (!oid
.match(cnode
.bits
, pgid
.ps())) {
3196 lderr(store
->cct
) << __func__
<< " oid " << oid
<< " not part of "
3197 << pgid
<< " bits " << cnode
.bits
<< dendl
;
3202 OnodeRef o
= onode_map
.lookup(oid
);
3206 mempool::bluestore_cache_other::string key
;
3207 get_object_key(store
->cct
, oid
, &key
);
3209 ldout(store
->cct
, 20) << __func__
<< " oid " << oid
<< " key "
3210 << pretty_binary_string(key
) << dendl
;
3213 int r
= store
->db
->get(PREFIX_OBJ
, key
.c_str(), key
.size(), &v
);
3214 ldout(store
->cct
, 20) << " r " << r
<< " v.len " << v
.length() << dendl
;
3216 if (v
.length() == 0) {
3217 assert(r
== -ENOENT
);
3218 if (!store
->cct
->_conf
->bluestore_debug_misc
&&
3222 // new object, new onode
3223 on
= new Onode(this, oid
, key
);
3227 on
= new Onode(this, oid
, key
);
3229 bufferptr::iterator p
= v
.front().begin_deep();
3230 on
->onode
.decode(p
);
3232 // initialize extent_map
3233 on
->extent_map
.decode_spanning_blobs(p
);
3234 if (on
->onode
.extent_map_shards
.empty()) {
3235 denc(on
->extent_map
.inline_bl
, p
);
3236 on
->extent_map
.decode_some(on
->extent_map
.inline_bl
);
3238 on
->extent_map
.init_shards(false, false);
3242 return onode_map
.add(oid
, o
);
3245 void BlueStore::Collection::split_cache(
3248 ldout(store
->cct
, 10) << __func__
<< " to " << dest
<< dendl
;
3250 // lock (one or both) cache shards
3251 std::lock(cache
->lock
, dest
->cache
->lock
);
3252 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
, std::adopt_lock
);
3253 std::lock_guard
<std::recursive_mutex
> l2(dest
->cache
->lock
, std::adopt_lock
);
3255 int destbits
= dest
->cnode
.bits
;
3257 bool is_pg
= dest
->cid
.is_pg(&destpg
);
3260 auto p
= onode_map
.onode_map
.begin();
3261 while (p
!= onode_map
.onode_map
.end()) {
3262 if (!p
->second
->oid
.match(destbits
, destpg
.pgid
.ps())) {
3263 // onode does not belong to this child
3266 OnodeRef o
= p
->second
;
3267 ldout(store
->cct
, 20) << __func__
<< " moving " << o
<< " " << o
->oid
3270 cache
->_rm_onode(p
->second
);
3271 p
= onode_map
.onode_map
.erase(p
);
3274 dest
->cache
->_add_onode(o
, 1);
3275 dest
->onode_map
.onode_map
[o
->oid
] = o
;
3276 dest
->onode_map
.cache
= dest
->cache
;
3278 // move over shared blobs and buffers. cover shared blobs from
3279 // both extent map and spanning blob map (the full extent map
3280 // may not be faulted in)
3281 vector
<SharedBlob
*> sbvec
;
3282 for (auto& e
: o
->extent_map
.extent_map
) {
3283 sbvec
.push_back(e
.blob
->shared_blob
.get());
3285 for (auto& b
: o
->extent_map
.spanning_blob_map
) {
3286 sbvec
.push_back(b
.second
->shared_blob
.get());
3288 for (auto sb
: sbvec
) {
3289 if (sb
->coll
== dest
) {
3290 ldout(store
->cct
, 20) << __func__
<< " already moved " << *sb
3294 ldout(store
->cct
, 20) << __func__
<< " moving " << *sb
<< dendl
;
3296 if (sb
->get_sbid()) {
3297 ldout(store
->cct
, 20) << __func__
3298 << " moving registration " << *sb
<< dendl
;
3299 shared_blob_set
.remove(sb
);
3300 dest
->shared_blob_set
.add(dest
, sb
);
3302 if (dest
->cache
!= cache
) {
3303 for (auto& i
: sb
->bc
.buffer_map
) {
3304 if (!i
.second
->is_writing()) {
3305 ldout(store
->cct
, 20) << __func__
<< " moving " << *i
.second
3307 dest
->cache
->_move_buffer(cache
, i
.second
.get());
3316 // =======================================================
3318 void *BlueStore::MempoolThread::entry()
3320 Mutex::Locker
l(lock
);
3322 uint64_t meta_bytes
=
3323 mempool::bluestore_cache_other::allocated_bytes() +
3324 mempool::bluestore_cache_onode::allocated_bytes();
3325 uint64_t onode_num
=
3326 mempool::bluestore_cache_onode::allocated_items();
3328 if (onode_num
< 2) {
3332 float bytes_per_onode
= (float)meta_bytes
/ (float)onode_num
;
3333 size_t num_shards
= store
->cache_shards
.size();
3334 float target_ratio
= store
->cache_meta_ratio
+ store
->cache_data_ratio
;
3335 // A little sloppy but should be close enough
3336 uint64_t shard_target
= target_ratio
* (store
->cache_size
/ num_shards
);
3338 for (auto i
: store
->cache_shards
) {
3339 i
->trim(shard_target
,
3340 store
->cache_meta_ratio
,
3341 store
->cache_data_ratio
,
3345 store
->_update_cache_logger();
3348 wait
+= store
->cct
->_conf
->bluestore_cache_trim_interval
;
3349 cond
.WaitInterval(lock
, wait
);
3355 // =======================================================
3360 #define dout_prefix *_dout << "bluestore.OmapIteratorImpl(" << this << ") "
3362 BlueStore::OmapIteratorImpl::OmapIteratorImpl(
3363 CollectionRef c
, OnodeRef o
, KeyValueDB::Iterator it
)
3364 : c(c
), o(o
), it(it
)
3366 RWLock::RLocker
l(c
->lock
);
3367 if (o
->onode
.has_omap()) {
3368 get_omap_key(o
->onode
.nid
, string(), &head
);
3369 get_omap_tail(o
->onode
.nid
, &tail
);
3370 it
->lower_bound(head
);
3374 int BlueStore::OmapIteratorImpl::seek_to_first()
3376 RWLock::RLocker
l(c
->lock
);
3377 if (o
->onode
.has_omap()) {
3378 it
->lower_bound(head
);
3380 it
= KeyValueDB::Iterator();
3385 int BlueStore::OmapIteratorImpl::upper_bound(const string
& after
)
3387 RWLock::RLocker
l(c
->lock
);
3388 if (o
->onode
.has_omap()) {
3390 get_omap_key(o
->onode
.nid
, after
, &key
);
3391 ldout(c
->store
->cct
,20) << __func__
<< " after " << after
<< " key "
3392 << pretty_binary_string(key
) << dendl
;
3393 it
->upper_bound(key
);
3395 it
= KeyValueDB::Iterator();
3400 int BlueStore::OmapIteratorImpl::lower_bound(const string
& to
)
3402 RWLock::RLocker
l(c
->lock
);
3403 if (o
->onode
.has_omap()) {
3405 get_omap_key(o
->onode
.nid
, to
, &key
);
3406 ldout(c
->store
->cct
,20) << __func__
<< " to " << to
<< " key "
3407 << pretty_binary_string(key
) << dendl
;
3408 it
->lower_bound(key
);
3410 it
= KeyValueDB::Iterator();
3415 bool BlueStore::OmapIteratorImpl::valid()
3417 RWLock::RLocker
l(c
->lock
);
3418 bool r
= o
->onode
.has_omap() && it
&& it
->valid() &&
3419 it
->raw_key().second
<= tail
;
3420 if (it
&& it
->valid()) {
3421 ldout(c
->store
->cct
,20) << __func__
<< " is at "
3422 << pretty_binary_string(it
->raw_key().second
)
3428 int BlueStore::OmapIteratorImpl::next(bool validate
)
3430 RWLock::RLocker
l(c
->lock
);
3431 if (o
->onode
.has_omap()) {
3439 string
BlueStore::OmapIteratorImpl::key()
3441 RWLock::RLocker
l(c
->lock
);
3442 assert(it
->valid());
3443 string db_key
= it
->raw_key().second
;
3445 decode_omap_key(db_key
, &user_key
);
3449 bufferlist
BlueStore::OmapIteratorImpl::value()
3451 RWLock::RLocker
l(c
->lock
);
3452 assert(it
->valid());
3457 // =====================================
3460 #define dout_prefix *_dout << "bluestore(" << path << ") "
3463 static void aio_cb(void *priv
, void *priv2
)
3465 BlueStore
*store
= static_cast<BlueStore
*>(priv
);
3466 BlueStore::AioContext
*c
= static_cast<BlueStore::AioContext
*>(priv2
);
3467 c
->aio_finish(store
);
3470 BlueStore::BlueStore(CephContext
*cct
, const string
& path
)
3471 : ObjectStore(cct
, path
),
3472 throttle_bytes(cct
, "bluestore_throttle_bytes",
3473 cct
->_conf
->bluestore_throttle_bytes
),
3474 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3475 cct
->_conf
->bluestore_throttle_bytes
+
3476 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3477 kv_sync_thread(this),
3478 kv_finalize_thread(this),
3479 mempool_thread(this)
3482 cct
->_conf
->add_observer(this);
3483 set_cache_shards(1);
3486 BlueStore::BlueStore(CephContext
*cct
,
3488 uint64_t _min_alloc_size
)
3489 : ObjectStore(cct
, path
),
3490 throttle_bytes(cct
, "bluestore_throttle_bytes",
3491 cct
->_conf
->bluestore_throttle_bytes
),
3492 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3493 cct
->_conf
->bluestore_throttle_bytes
+
3494 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3495 kv_sync_thread(this),
3496 kv_finalize_thread(this),
3497 min_alloc_size(_min_alloc_size
),
3498 min_alloc_size_order(ctz(_min_alloc_size
)),
3499 mempool_thread(this)
3502 cct
->_conf
->add_observer(this);
3503 set_cache_shards(1);
3506 BlueStore::~BlueStore()
3508 for (auto f
: finishers
) {
3513 cct
->_conf
->remove_observer(this);
3517 assert(bluefs
== NULL
);
3518 assert(fsid_fd
< 0);
3519 assert(path_fd
< 0);
3520 for (auto i
: cache_shards
) {
3523 cache_shards
.clear();
3526 const char **BlueStore::get_tracked_conf_keys() const
3528 static const char* KEYS
[] = {
3529 "bluestore_csum_type",
3530 "bluestore_compression_mode",
3531 "bluestore_compression_algorithm",
3532 "bluestore_compression_min_blob_size",
3533 "bluestore_compression_min_blob_size_ssd",
3534 "bluestore_compression_min_blob_size_hdd",
3535 "bluestore_compression_max_blob_size",
3536 "bluestore_compression_max_blob_size_ssd",
3537 "bluestore_compression_max_blob_size_hdd",
3538 "bluestore_compression_required_ratio",
3539 "bluestore_max_alloc_size",
3540 "bluestore_prefer_deferred_size",
3541 "bluestore_deferred_batch_ops",
3542 "bluestore_deferred_batch_ops_hdd",
3543 "bluestore_deferred_batch_ops_ssd",
3544 "bluestore_throttle_bytes",
3545 "bluestore_throttle_deferred_bytes",
3546 "bluestore_throttle_cost_per_io_hdd",
3547 "bluestore_throttle_cost_per_io_ssd",
3548 "bluestore_throttle_cost_per_io",
3549 "bluestore_max_blob_size",
3550 "bluestore_max_blob_size_ssd",
3551 "bluestore_max_blob_size_hdd",
3557 void BlueStore::handle_conf_change(const struct md_config_t
*conf
,
3558 const std::set
<std::string
> &changed
)
3560 if (changed
.count("bluestore_csum_type")) {
3563 if (changed
.count("bluestore_compression_mode") ||
3564 changed
.count("bluestore_compression_algorithm") ||
3565 changed
.count("bluestore_compression_min_blob_size") ||
3566 changed
.count("bluestore_compression_max_blob_size")) {
3571 if (changed
.count("bluestore_max_blob_size") ||
3572 changed
.count("bluestore_max_blob_size_ssd") ||
3573 changed
.count("bluestore_max_blob_size_hdd")) {
3575 // only after startup
3579 if (changed
.count("bluestore_prefer_deferred_size") ||
3580 changed
.count("bluestore_max_alloc_size") ||
3581 changed
.count("bluestore_deferred_batch_ops") ||
3582 changed
.count("bluestore_deferred_batch_ops_hdd") ||
3583 changed
.count("bluestore_deferred_batch_ops_ssd")) {
3585 // only after startup
3589 if (changed
.count("bluestore_throttle_cost_per_io") ||
3590 changed
.count("bluestore_throttle_cost_per_io_hdd") ||
3591 changed
.count("bluestore_throttle_cost_per_io_ssd")) {
3593 _set_throttle_params();
3596 if (changed
.count("bluestore_throttle_bytes")) {
3597 throttle_bytes
.reset_max(conf
->bluestore_throttle_bytes
);
3598 throttle_deferred_bytes
.reset_max(
3599 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3601 if (changed
.count("bluestore_throttle_deferred_bytes")) {
3602 throttle_deferred_bytes
.reset_max(
3603 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3607 void BlueStore::_set_compression()
3609 auto m
= Compressor::get_comp_mode_type(cct
->_conf
->bluestore_compression_mode
);
3613 derr
<< __func__
<< " unrecognized value '"
3614 << cct
->_conf
->bluestore_compression_mode
3615 << "' for bluestore_compression_mode, reverting to 'none'"
3617 comp_mode
= Compressor::COMP_NONE
;
3620 compressor
= nullptr;
3622 if (comp_mode
== Compressor::COMP_NONE
) {
3623 dout(10) << __func__
<< " compression mode set to 'none', "
3624 << "ignore other compression setttings" << dendl
;
3628 if (cct
->_conf
->bluestore_compression_max_blob_size
) {
3629 comp_min_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size
;
3632 if (bdev
->is_rotational()) {
3633 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_hdd
;
3635 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_ssd
;
3639 if (cct
->_conf
->bluestore_compression_max_blob_size
) {
3640 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size
;
3643 if (bdev
->is_rotational()) {
3644 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_hdd
;
3646 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_ssd
;
3650 auto& alg_name
= cct
->_conf
->bluestore_compression_algorithm
;
3651 if (!alg_name
.empty()) {
3652 compressor
= Compressor::create(cct
, alg_name
);
3654 derr
<< __func__
<< " unable to initialize " << alg_name
.c_str() << " compressor"
3659 dout(10) << __func__
<< " mode " << Compressor::get_comp_mode_name(comp_mode
)
3660 << " alg " << (compressor
? compressor
->get_type_name() : "(none)")
3664 void BlueStore::_set_csum()
3666 csum_type
= Checksummer::CSUM_NONE
;
3667 int t
= Checksummer::get_csum_string_type(cct
->_conf
->bluestore_csum_type
);
3668 if (t
> Checksummer::CSUM_NONE
)
3671 dout(10) << __func__
<< " csum_type "
3672 << Checksummer::get_csum_type_string(csum_type
)
3676 void BlueStore::_set_throttle_params()
3678 if (cct
->_conf
->bluestore_throttle_cost_per_io
) {
3679 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io
;
3682 if (bdev
->is_rotational()) {
3683 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_hdd
;
3685 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_ssd
;
3689 dout(10) << __func__
<< " throttle_cost_per_io " << throttle_cost_per_io
3692 void BlueStore::_set_blob_size()
3694 if (cct
->_conf
->bluestore_max_blob_size
) {
3695 max_blob_size
= cct
->_conf
->bluestore_max_blob_size
;
3698 if (bdev
->is_rotational()) {
3699 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_hdd
;
3701 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_ssd
;
3704 dout(10) << __func__
<< " max_blob_size 0x" << std::hex
<< max_blob_size
3705 << std::dec
<< dendl
;
3708 int BlueStore::_set_cache_sizes()
3711 if (cct
->_conf
->bluestore_cache_size
) {
3712 cache_size
= cct
->_conf
->bluestore_cache_size
;
3714 // choose global cache size based on backend type
3715 if (bdev
->is_rotational()) {
3716 cache_size
= cct
->_conf
->bluestore_cache_size_hdd
;
3718 cache_size
= cct
->_conf
->bluestore_cache_size_ssd
;
3721 cache_meta_ratio
= cct
->_conf
->bluestore_cache_meta_ratio
;
3722 cache_kv_ratio
= cct
->_conf
->bluestore_cache_kv_ratio
;
3724 double cache_kv_max
= cct
->_conf
->bluestore_cache_kv_max
;
3725 double cache_kv_max_ratio
= 0;
3727 // if cache_kv_max is negative, disable it
3728 if (cache_size
> 0 && cache_kv_max
>= 0) {
3729 cache_kv_max_ratio
= (double) cache_kv_max
/ (double) cache_size
;
3730 if (cache_kv_max_ratio
< 1.0 && cache_kv_max_ratio
< cache_kv_ratio
) {
3731 dout(1) << __func__
<< " max " << cache_kv_max_ratio
3732 << " < ratio " << cache_kv_ratio
3734 cache_meta_ratio
= cache_meta_ratio
+ cache_kv_ratio
- cache_kv_max_ratio
;
3735 cache_kv_ratio
= cache_kv_max_ratio
;
3740 (double)1.0 - (double)cache_meta_ratio
- (double)cache_kv_ratio
;
3742 if (cache_meta_ratio
< 0 || cache_meta_ratio
> 1.0) {
3743 derr
<< __func__
<< " bluestore_cache_meta_ratio (" << cache_meta_ratio
3744 << ") must be in range [0,1.0]" << dendl
;
3747 if (cache_kv_ratio
< 0 || cache_kv_ratio
> 1.0) {
3748 derr
<< __func__
<< " bluestore_cache_kv_ratio (" << cache_kv_ratio
3749 << ") must be in range [0,1.0]" << dendl
;
3752 if (cache_meta_ratio
+ cache_kv_ratio
> 1.0) {
3753 derr
<< __func__
<< " bluestore_cache_meta_ratio (" << cache_meta_ratio
3754 << ") + bluestore_cache_kv_ratio (" << cache_kv_ratio
3755 << ") = " << cache_meta_ratio
+ cache_kv_ratio
<< "; must be <= 1.0"
3759 if (cache_data_ratio
< 0) {
3760 // deal with floating point imprecision
3761 cache_data_ratio
= 0;
3763 dout(1) << __func__
<< " cache_size " << cache_size
3764 << " meta " << cache_meta_ratio
3765 << " kv " << cache_kv_ratio
3766 << " data " << cache_data_ratio
3771 void BlueStore::_init_logger()
3773 PerfCountersBuilder
b(cct
, "bluestore",
3774 l_bluestore_first
, l_bluestore_last
);
3775 b
.add_time_avg(l_bluestore_kv_flush_lat
, "kv_flush_lat",
3776 "Average kv_thread flush latency",
3777 "fl_l", PerfCountersBuilder::PRIO_INTERESTING
);
3778 b
.add_time_avg(l_bluestore_kv_commit_lat
, "kv_commit_lat",
3779 "Average kv_thread commit latency");
3780 b
.add_time_avg(l_bluestore_kv_lat
, "kv_lat",
3781 "Average kv_thread sync latency",
3782 "k_l", PerfCountersBuilder::PRIO_INTERESTING
);
3783 b
.add_time_avg(l_bluestore_state_prepare_lat
, "state_prepare_lat",
3784 "Average prepare state latency");
3785 b
.add_time_avg(l_bluestore_state_aio_wait_lat
, "state_aio_wait_lat",
3786 "Average aio_wait state latency",
3787 "io_l", PerfCountersBuilder::PRIO_INTERESTING
);
3788 b
.add_time_avg(l_bluestore_state_io_done_lat
, "state_io_done_lat",
3789 "Average io_done state latency");
3790 b
.add_time_avg(l_bluestore_state_kv_queued_lat
, "state_kv_queued_lat",
3791 "Average kv_queued state latency");
3792 b
.add_time_avg(l_bluestore_state_kv_committing_lat
, "state_kv_commiting_lat",
3793 "Average kv_commiting state latency");
3794 b
.add_time_avg(l_bluestore_state_kv_done_lat
, "state_kv_done_lat",
3795 "Average kv_done state latency");
3796 b
.add_time_avg(l_bluestore_state_deferred_queued_lat
, "state_deferred_queued_lat",
3797 "Average deferred_queued state latency");
3798 b
.add_time_avg(l_bluestore_state_deferred_aio_wait_lat
, "state_deferred_aio_wait_lat",
3799 "Average aio_wait state latency");
3800 b
.add_time_avg(l_bluestore_state_deferred_cleanup_lat
, "state_deferred_cleanup_lat",
3801 "Average cleanup state latency");
3802 b
.add_time_avg(l_bluestore_state_finishing_lat
, "state_finishing_lat",
3803 "Average finishing state latency");
3804 b
.add_time_avg(l_bluestore_state_done_lat
, "state_done_lat",
3805 "Average done state latency");
3806 b
.add_time_avg(l_bluestore_throttle_lat
, "throttle_lat",
3807 "Average submit throttle latency",
3808 "th_l", PerfCountersBuilder::PRIO_CRITICAL
);
3809 b
.add_time_avg(l_bluestore_submit_lat
, "submit_lat",
3810 "Average submit latency",
3811 "s_l", PerfCountersBuilder::PRIO_CRITICAL
);
3812 b
.add_time_avg(l_bluestore_commit_lat
, "commit_lat",
3813 "Average commit latency",
3814 "c_l", PerfCountersBuilder::PRIO_CRITICAL
);
3815 b
.add_time_avg(l_bluestore_read_lat
, "read_lat",
3816 "Average read latency",
3817 "r_l", PerfCountersBuilder::PRIO_CRITICAL
);
3818 b
.add_time_avg(l_bluestore_read_onode_meta_lat
, "read_onode_meta_lat",
3819 "Average read onode metadata latency");
3820 b
.add_time_avg(l_bluestore_read_wait_aio_lat
, "read_wait_aio_lat",
3821 "Average read latency");
3822 b
.add_time_avg(l_bluestore_compress_lat
, "compress_lat",
3823 "Average compress latency");
3824 b
.add_time_avg(l_bluestore_decompress_lat
, "decompress_lat",
3825 "Average decompress latency");
3826 b
.add_time_avg(l_bluestore_csum_lat
, "csum_lat",
3827 "Average checksum latency");
3828 b
.add_u64_counter(l_bluestore_compress_success_count
, "compress_success_count",
3829 "Sum for beneficial compress ops");
3830 b
.add_u64_counter(l_bluestore_compress_rejected_count
, "compress_rejected_count",
3831 "Sum for compress ops rejected due to low net gain of space");
3832 b
.add_u64_counter(l_bluestore_write_pad_bytes
, "write_pad_bytes",
3833 "Sum for write-op padded bytes");
3834 b
.add_u64_counter(l_bluestore_deferred_write_ops
, "deferred_write_ops",
3835 "Sum for deferred write op");
3836 b
.add_u64_counter(l_bluestore_deferred_write_bytes
, "deferred_write_bytes",
3837 "Sum for deferred write bytes", "def");
3838 b
.add_u64_counter(l_bluestore_write_penalty_read_ops
, "write_penalty_read_ops",
3839 "Sum for write penalty read ops");
3840 b
.add_u64(l_bluestore_allocated
, "bluestore_allocated",
3841 "Sum for allocated bytes");
3842 b
.add_u64(l_bluestore_stored
, "bluestore_stored",
3843 "Sum for stored bytes");
3844 b
.add_u64(l_bluestore_compressed
, "bluestore_compressed",
3845 "Sum for stored compressed bytes");
3846 b
.add_u64(l_bluestore_compressed_allocated
, "bluestore_compressed_allocated",
3847 "Sum for bytes allocated for compressed data");
3848 b
.add_u64(l_bluestore_compressed_original
, "bluestore_compressed_original",
3849 "Sum for original bytes that were compressed");
3851 b
.add_u64(l_bluestore_onodes
, "bluestore_onodes",
3852 "Number of onodes in cache");
3853 b
.add_u64_counter(l_bluestore_onode_hits
, "bluestore_onode_hits",
3854 "Sum for onode-lookups hit in the cache");
3855 b
.add_u64_counter(l_bluestore_onode_misses
, "bluestore_onode_misses",
3856 "Sum for onode-lookups missed in the cache");
3857 b
.add_u64_counter(l_bluestore_onode_shard_hits
, "bluestore_onode_shard_hits",
3858 "Sum for onode-shard lookups hit in the cache");
3859 b
.add_u64_counter(l_bluestore_onode_shard_misses
,
3860 "bluestore_onode_shard_misses",
3861 "Sum for onode-shard lookups missed in the cache");
3862 b
.add_u64(l_bluestore_extents
, "bluestore_extents",
3863 "Number of extents in cache");
3864 b
.add_u64(l_bluestore_blobs
, "bluestore_blobs",
3865 "Number of blobs in cache");
3866 b
.add_u64(l_bluestore_buffers
, "bluestore_buffers",
3867 "Number of buffers in cache");
3868 b
.add_u64(l_bluestore_buffer_bytes
, "bluestore_buffer_bytes",
3869 "Number of buffer bytes in cache");
3870 b
.add_u64(l_bluestore_buffer_hit_bytes
, "bluestore_buffer_hit_bytes",
3871 "Sum for bytes of read hit in the cache");
3872 b
.add_u64(l_bluestore_buffer_miss_bytes
, "bluestore_buffer_miss_bytes",
3873 "Sum for bytes of read missed in the cache");
3875 b
.add_u64_counter(l_bluestore_write_big
, "bluestore_write_big",
3876 "Large aligned writes into fresh blobs");
3877 b
.add_u64_counter(l_bluestore_write_big_bytes
, "bluestore_write_big_bytes",
3878 "Large aligned writes into fresh blobs (bytes)");
3879 b
.add_u64_counter(l_bluestore_write_big_blobs
, "bluestore_write_big_blobs",
3880 "Large aligned writes into fresh blobs (blobs)");
3881 b
.add_u64_counter(l_bluestore_write_small
, "bluestore_write_small",
3882 "Small writes into existing or sparse small blobs");
3883 b
.add_u64_counter(l_bluestore_write_small_bytes
, "bluestore_write_small_bytes",
3884 "Small writes into existing or sparse small blobs (bytes)");
3885 b
.add_u64_counter(l_bluestore_write_small_unused
,
3886 "bluestore_write_small_unused",
3887 "Small writes into unused portion of existing blob");
3888 b
.add_u64_counter(l_bluestore_write_small_deferred
,
3889 "bluestore_write_small_deferred",
3890 "Small overwrites using deferred");
3891 b
.add_u64_counter(l_bluestore_write_small_pre_read
,
3892 "bluestore_write_small_pre_read",
3893 "Small writes that required we read some data (possibly "
3894 "cached) to fill out the block");
3895 b
.add_u64_counter(l_bluestore_write_small_new
, "bluestore_write_small_new",
3896 "Small write into new (sparse) blob");
3898 b
.add_u64_counter(l_bluestore_txc
, "bluestore_txc", "Transactions committed");
3899 b
.add_u64_counter(l_bluestore_onode_reshard
, "bluestore_onode_reshard",
3900 "Onode extent map reshard events");
3901 b
.add_u64_counter(l_bluestore_blob_split
, "bluestore_blob_split",
3902 "Sum for blob splitting due to resharding");
3903 b
.add_u64_counter(l_bluestore_extent_compress
, "bluestore_extent_compress",
3904 "Sum for extents that have been removed due to compression");
3905 b
.add_u64_counter(l_bluestore_gc_merged
, "bluestore_gc_merged",
3906 "Sum for extents that have been merged due to garbage "
3908 logger
= b
.create_perf_counters();
3909 cct
->get_perfcounters_collection()->add(logger
);
3912 int BlueStore::_reload_logger()
3914 struct store_statfs_t store_statfs
;
3916 int r
= statfs(&store_statfs
);
3918 logger
->set(l_bluestore_allocated
, store_statfs
.allocated
);
3919 logger
->set(l_bluestore_stored
, store_statfs
.stored
);
3920 logger
->set(l_bluestore_compressed
, store_statfs
.compressed
);
3921 logger
->set(l_bluestore_compressed_allocated
, store_statfs
.compressed_allocated
);
3922 logger
->set(l_bluestore_compressed_original
, store_statfs
.compressed_original
);
3927 void BlueStore::_shutdown_logger()
3929 cct
->get_perfcounters_collection()->remove(logger
);
3933 int BlueStore::get_block_device_fsid(CephContext
* cct
, const string
& path
,
3936 bluestore_bdev_label_t label
;
3937 int r
= _read_bdev_label(cct
, path
, &label
);
3940 *fsid
= label
.osd_uuid
;
3944 int BlueStore::_open_path()
3946 assert(path_fd
< 0);
3947 path_fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_DIRECTORY
));
3950 derr
<< __func__
<< " unable to open " << path
<< ": " << cpp_strerror(r
)
3957 void BlueStore::_close_path()
3959 VOID_TEMP_FAILURE_RETRY(::close(path_fd
));
3963 int BlueStore::_write_bdev_label(string path
, bluestore_bdev_label_t label
)
3965 dout(10) << __func__
<< " path " << path
<< " label " << label
<< dendl
;
3967 ::encode(label
, bl
);
3968 uint32_t crc
= bl
.crc32c(-1);
3970 assert(bl
.length() <= BDEV_LABEL_BLOCK_SIZE
);
3971 bufferptr
z(BDEV_LABEL_BLOCK_SIZE
- bl
.length());
3973 bl
.append(std::move(z
));
3975 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_WRONLY
));
3978 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
3982 int r
= bl
.write_fd(fd
);
3984 derr
<< __func__
<< " failed to write to " << path
3985 << ": " << cpp_strerror(r
) << dendl
;
3987 VOID_TEMP_FAILURE_RETRY(::close(fd
));
3991 int BlueStore::_read_bdev_label(CephContext
* cct
, string path
,
3992 bluestore_bdev_label_t
*label
)
3994 dout(10) << __func__
<< dendl
;
3995 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_RDONLY
));
3998 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
4003 int r
= bl
.read_fd(fd
, BDEV_LABEL_BLOCK_SIZE
);
4004 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4006 derr
<< __func__
<< " failed to read from " << path
4007 << ": " << cpp_strerror(r
) << dendl
;
4011 uint32_t crc
, expected_crc
;
4012 bufferlist::iterator p
= bl
.begin();
4014 ::decode(*label
, p
);
4016 t
.substr_of(bl
, 0, p
.get_off());
4018 ::decode(expected_crc
, p
);
4020 catch (buffer::error
& e
) {
4021 derr
<< __func__
<< " unable to decode label at offset " << p
.get_off()
4026 if (crc
!= expected_crc
) {
4027 derr
<< __func__
<< " bad crc on label, expected " << expected_crc
4028 << " != actual " << crc
<< dendl
;
4031 dout(10) << __func__
<< " got " << *label
<< dendl
;
4035 int BlueStore::_check_or_set_bdev_label(
4036 string path
, uint64_t size
, string desc
, bool create
)
4038 bluestore_bdev_label_t label
;
4040 label
.osd_uuid
= fsid
;
4042 label
.btime
= ceph_clock_now();
4043 label
.description
= desc
;
4044 int r
= _write_bdev_label(path
, label
);
4048 int r
= _read_bdev_label(cct
, path
, &label
);
4051 if (cct
->_conf
->bluestore_debug_permit_any_bdev_label
) {
4052 dout(20) << __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4053 << " and fsid " << fsid
<< " check bypassed" << dendl
;
4055 else if (label
.osd_uuid
!= fsid
) {
4056 derr
<< __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4057 << " does not match our fsid " << fsid
<< dendl
;
4064 void BlueStore::_set_alloc_sizes(void)
4066 max_alloc_size
= cct
->_conf
->bluestore_max_alloc_size
;
4068 if (cct
->_conf
->bluestore_prefer_deferred_size
) {
4069 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size
;
4072 if (bdev
->is_rotational()) {
4073 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_hdd
;
4075 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_ssd
;
4079 if (cct
->_conf
->bluestore_deferred_batch_ops
) {
4080 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops
;
4083 if (bdev
->is_rotational()) {
4084 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_hdd
;
4086 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_ssd
;
4090 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
4091 << std::dec
<< " order " << min_alloc_size_order
4092 << " max_alloc_size 0x" << std::hex
<< max_alloc_size
4093 << " prefer_deferred_size 0x" << prefer_deferred_size
4095 << " deferred_batch_ops " << deferred_batch_ops
4099 int BlueStore::_open_bdev(bool create
)
4101 assert(bdev
== NULL
);
4102 string p
= path
+ "/block";
4103 bdev
= BlockDevice::create(cct
, p
, aio_cb
, static_cast<void*>(this));
4104 int r
= bdev
->open(p
);
4108 if (bdev
->supported_bdev_label()) {
4109 r
= _check_or_set_bdev_label(p
, bdev
->get_size(), "main", create
);
4114 // initialize global block parameters
4115 block_size
= bdev
->get_block_size();
4116 block_mask
= ~(block_size
- 1);
4117 block_size_order
= ctz(block_size
);
4118 assert(block_size
== 1u << block_size_order
);
4119 // and set cache_size based on device type
4120 r
= _set_cache_sizes();
4134 void BlueStore::_close_bdev()
4142 int BlueStore::_open_fm(bool create
)
4145 fm
= FreelistManager::create(cct
, freelist_type
, db
, PREFIX_ALLOC
);
4148 // initialize freespace
4149 dout(20) << __func__
<< " initializing freespace" << dendl
;
4150 KeyValueDB::Transaction t
= db
->get_transaction();
4153 bl
.append(freelist_type
);
4154 t
->set(PREFIX_SUPER
, "freelist_type", bl
);
4156 fm
->create(bdev
->get_size(), t
);
4158 // allocate superblock reserved space. note that we do not mark
4159 // bluefs space as allocated in the freelist; we instead rely on
4161 fm
->allocate(0, SUPER_RESERVED
, t
);
4163 uint64_t reserved
= 0;
4164 if (cct
->_conf
->bluestore_bluefs
) {
4165 assert(bluefs_extents
.num_intervals() == 1);
4166 interval_set
<uint64_t>::iterator p
= bluefs_extents
.begin();
4167 reserved
= p
.get_start() + p
.get_len();
4168 dout(20) << __func__
<< " reserved 0x" << std::hex
<< reserved
<< std::dec
4169 << " for bluefs" << dendl
;
4171 ::encode(bluefs_extents
, bl
);
4172 t
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
4173 dout(20) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
4174 << std::dec
<< dendl
;
4176 reserved
= SUPER_RESERVED
;
4179 if (cct
->_conf
->bluestore_debug_prefill
> 0) {
4180 uint64_t end
= bdev
->get_size() - reserved
;
4181 dout(1) << __func__
<< " pre-fragmenting freespace, using "
4182 << cct
->_conf
->bluestore_debug_prefill
<< " with max free extent "
4183 << cct
->_conf
->bluestore_debug_prefragment_max
<< dendl
;
4184 uint64_t start
= P2ROUNDUP(reserved
, min_alloc_size
);
4185 uint64_t max_b
= cct
->_conf
->bluestore_debug_prefragment_max
/ min_alloc_size
;
4186 float r
= cct
->_conf
->bluestore_debug_prefill
;
4190 while (!stop
&& start
< end
) {
4191 uint64_t l
= (rand() % max_b
+ 1) * min_alloc_size
;
4192 if (start
+ l
> end
) {
4194 l
= P2ALIGN(l
, min_alloc_size
);
4196 assert(start
+ l
<= end
);
4198 uint64_t u
= 1 + (uint64_t)(r
* (double)l
);
4199 u
= P2ROUNDUP(u
, min_alloc_size
);
4200 if (start
+ l
+ u
> end
) {
4201 u
= end
- (start
+ l
);
4202 // trim to align so we don't overflow again
4203 u
= P2ALIGN(u
, min_alloc_size
);
4206 assert(start
+ l
+ u
<= end
);
4208 dout(20) << " free 0x" << std::hex
<< start
<< "~" << l
4209 << " use 0x" << u
<< std::dec
<< dendl
;
4212 // break if u has been trimmed to nothing
4216 fm
->allocate(start
+ l
, u
, t
);
4220 db
->submit_transaction_sync(t
);
4225 derr
<< __func__
<< " freelist init failed: " << cpp_strerror(r
) << dendl
;
4233 void BlueStore::_close_fm()
4235 dout(10) << __func__
<< dendl
;
4242 int BlueStore::_open_alloc()
4244 assert(alloc
== NULL
);
4245 assert(bdev
->get_size());
4246 alloc
= Allocator::create(cct
, cct
->_conf
->bluestore_allocator
,
4250 lderr(cct
) << __func__
<< " Allocator::unknown alloc type "
4251 << cct
->_conf
->bluestore_allocator
4256 uint64_t num
= 0, bytes
= 0;
4258 dout(1) << __func__
<< " opening allocation metadata" << dendl
;
4259 // initialize from freelist
4260 fm
->enumerate_reset();
4261 uint64_t offset
, length
;
4262 while (fm
->enumerate_next(&offset
, &length
)) {
4263 alloc
->init_add_free(offset
, length
);
4267 fm
->enumerate_reset();
4268 dout(1) << __func__
<< " loaded " << pretty_si_t(bytes
)
4269 << " in " << num
<< " extents"
4272 // also mark bluefs space as allocated
4273 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
4274 alloc
->init_rm_free(e
.get_start(), e
.get_len());
4276 dout(10) << __func__
<< " marked bluefs_extents 0x" << std::hex
4277 << bluefs_extents
<< std::dec
<< " as allocated" << dendl
;
4282 void BlueStore::_close_alloc()
4290 int BlueStore::_open_fsid(bool create
)
4292 assert(fsid_fd
< 0);
4296 fsid_fd
= ::openat(path_fd
, "fsid", flags
, 0644);
4299 derr
<< __func__
<< " " << cpp_strerror(err
) << dendl
;
4305 int BlueStore::_read_fsid(uuid_d
*uuid
)
4308 memset(fsid_str
, 0, sizeof(fsid_str
));
4309 int ret
= safe_read(fsid_fd
, fsid_str
, sizeof(fsid_str
));
4311 derr
<< __func__
<< " failed: " << cpp_strerror(ret
) << dendl
;
4318 if (!uuid
->parse(fsid_str
)) {
4319 derr
<< __func__
<< " unparsable uuid " << fsid_str
<< dendl
;
4325 int BlueStore::_write_fsid()
4327 int r
= ::ftruncate(fsid_fd
, 0);
4330 derr
<< __func__
<< " fsid truncate failed: " << cpp_strerror(r
) << dendl
;
4333 string str
= stringify(fsid
) + "\n";
4334 r
= safe_write(fsid_fd
, str
.c_str(), str
.length());
4336 derr
<< __func__
<< " fsid write failed: " << cpp_strerror(r
) << dendl
;
4339 r
= ::fsync(fsid_fd
);
4342 derr
<< __func__
<< " fsid fsync failed: " << cpp_strerror(r
) << dendl
;
4348 void BlueStore::_close_fsid()
4350 VOID_TEMP_FAILURE_RETRY(::close(fsid_fd
));
4354 int BlueStore::_lock_fsid()
4357 memset(&l
, 0, sizeof(l
));
4359 l
.l_whence
= SEEK_SET
;
4360 int r
= ::fcntl(fsid_fd
, F_SETLK
, &l
);
4363 derr
<< __func__
<< " failed to lock " << path
<< "/fsid"
4364 << " (is another ceph-osd still running?)"
4365 << cpp_strerror(err
) << dendl
;
4371 bool BlueStore::is_rotational()
4374 return bdev
->is_rotational();
4377 bool rotational
= true;
4378 int r
= _open_path();
4381 r
= _open_fsid(false);
4384 r
= _read_fsid(&fsid
);
4390 r
= _open_bdev(false);
4393 rotational
= bdev
->is_rotational();
4403 bool BlueStore::is_journal_rotational()
4406 dout(5) << __func__
<< " bluefs disabled, default to store media type"
4408 return is_rotational();
4410 dout(10) << __func__
<< " " << (int)bluefs
->wal_is_rotational() << dendl
;
4411 return bluefs
->wal_is_rotational();
4414 bool BlueStore::test_mount_in_use()
4416 // most error conditions mean the mount is not in use (e.g., because
4417 // it doesn't exist). only if we fail to lock do we conclude it is
4420 int r
= _open_path();
4423 r
= _open_fsid(false);
4428 ret
= true; // if we can't lock, it is in use
4435 int BlueStore::_open_db(bool create
)
4439 string fn
= path
+ "/db";
4442 ceph::shared_ptr
<Int64ArrayMergeOperator
> merge_op(new Int64ArrayMergeOperator
);
4446 kv_backend
= cct
->_conf
->bluestore_kvbackend
;
4448 r
= read_meta("kv_backend", &kv_backend
);
4450 derr
<< __func__
<< " unable to read 'kv_backend' meta" << dendl
;
4454 dout(10) << __func__
<< " kv_backend = " << kv_backend
<< dendl
;
4458 do_bluefs
= cct
->_conf
->bluestore_bluefs
;
4461 r
= read_meta("bluefs", &s
);
4463 derr
<< __func__
<< " unable to read 'bluefs' meta" << dendl
;
4468 } else if (s
== "0") {
4471 derr
<< __func__
<< " bluefs = " << s
<< " : not 0 or 1, aborting"
4476 dout(10) << __func__
<< " do_bluefs = " << do_bluefs
<< dendl
;
4478 rocksdb::Env
*env
= NULL
;
4480 dout(10) << __func__
<< " initializing bluefs" << dendl
;
4481 if (kv_backend
!= "rocksdb") {
4482 derr
<< " backend must be rocksdb to use bluefs" << dendl
;
4485 bluefs
= new BlueFS(cct
);
4490 bfn
= path
+ "/block.db";
4491 if (::stat(bfn
.c_str(), &st
) == 0) {
4492 r
= bluefs
->add_block_device(BlueFS::BDEV_DB
, bfn
);
4494 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4495 << cpp_strerror(r
) << dendl
;
4499 if (bluefs
->bdev_support_label(BlueFS::BDEV_DB
)) {
4500 r
= _check_or_set_bdev_label(
4502 bluefs
->get_block_device_size(BlueFS::BDEV_DB
),
4503 "bluefs db", create
);
4506 << " check block device(" << bfn
<< ") label returned: "
4507 << cpp_strerror(r
) << dendl
;
4512 bluefs
->add_block_extent(
4515 bluefs
->get_block_device_size(BlueFS::BDEV_DB
) - SUPER_RESERVED
);
4517 bluefs_shared_bdev
= BlueFS::BDEV_SLOW
;
4518 bluefs_single_shared_device
= false;
4519 } else if (::lstat(bfn
.c_str(), &st
) == -1) {
4520 bluefs_shared_bdev
= BlueFS::BDEV_DB
;
4522 //symlink exist is bug
4523 derr
<< __func__
<< " " << bfn
<< " link target doesn't exist" << dendl
;
4529 bfn
= path
+ "/block";
4530 r
= bluefs
->add_block_device(bluefs_shared_bdev
, bfn
);
4532 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4533 << cpp_strerror(r
) << dendl
;
4537 // note: we always leave the first SUPER_RESERVED (8k) of the device unused
4539 bdev
->get_size() * (cct
->_conf
->bluestore_bluefs_min_ratio
+
4540 cct
->_conf
->bluestore_bluefs_gift_ratio
);
4541 initial
= MAX(initial
, cct
->_conf
->bluestore_bluefs_min
);
4542 // align to bluefs's alloc_size
4543 initial
= P2ROUNDUP(initial
, cct
->_conf
->bluefs_alloc_size
);
4544 // put bluefs in the middle of the device in case it is an HDD
4545 uint64_t start
= P2ALIGN((bdev
->get_size() - initial
) / 2,
4546 cct
->_conf
->bluefs_alloc_size
);
4547 bluefs
->add_block_extent(bluefs_shared_bdev
, start
, initial
);
4548 bluefs_extents
.insert(start
, initial
);
4551 bfn
= path
+ "/block.wal";
4552 if (::stat(bfn
.c_str(), &st
) == 0) {
4553 r
= bluefs
->add_block_device(BlueFS::BDEV_WAL
, bfn
);
4555 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4556 << cpp_strerror(r
) << dendl
;
4560 if (bluefs
->bdev_support_label(BlueFS::BDEV_WAL
)) {
4561 r
= _check_or_set_bdev_label(
4563 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
),
4564 "bluefs wal", create
);
4566 derr
<< __func__
<< " check block device(" << bfn
4567 << ") label returned: " << cpp_strerror(r
) << dendl
;
4573 bluefs
->add_block_extent(
4574 BlueFS::BDEV_WAL
, BDEV_LABEL_BLOCK_SIZE
,
4575 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
) -
4576 BDEV_LABEL_BLOCK_SIZE
);
4578 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "true");
4579 bluefs_single_shared_device
= false;
4580 } else if (::lstat(bfn
.c_str(), &st
) == -1) {
4581 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "false");
4583 //symlink exist is bug
4584 derr
<< __func__
<< " " << bfn
<< " link target doesn't exist" << dendl
;
4592 r
= bluefs
->mount();
4594 derr
<< __func__
<< " failed bluefs mount: " << cpp_strerror(r
) << dendl
;
4597 if (cct
->_conf
->bluestore_bluefs_env_mirror
) {
4598 rocksdb::Env
*a
= new BlueRocksEnv(bluefs
);
4599 rocksdb::Env
*b
= rocksdb::Env::Default();
4601 string cmd
= "rm -rf " + path
+ "/db " +
4602 path
+ "/db.slow " +
4604 int r
= system(cmd
.c_str());
4607 env
= new rocksdb::EnvMirror(b
, a
, false, true);
4609 env
= new BlueRocksEnv(bluefs
);
4611 // simplify the dir names, too, as "seen" by rocksdb
4615 if (bluefs_shared_bdev
== BlueFS::BDEV_SLOW
) {
4616 // we have both block.db and block; tell rocksdb!
4617 // note: the second (last) size value doesn't really matter
4618 ostringstream db_paths
;
4619 uint64_t db_size
= bluefs
->get_block_device_size(BlueFS::BDEV_DB
);
4620 uint64_t slow_size
= bluefs
->get_block_device_size(BlueFS::BDEV_SLOW
);
4621 db_paths
<< fn
<< ","
4622 << (uint64_t)(db_size
* 95 / 100) << " "
4623 << fn
+ ".slow" << ","
4624 << (uint64_t)(slow_size
* 95 / 100);
4625 cct
->_conf
->set_val("rocksdb_db_paths", db_paths
.str(), false);
4626 dout(10) << __func__
<< " set rocksdb_db_paths to "
4627 << cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths") << dendl
;
4632 if (cct
->_conf
->rocksdb_separate_wal_dir
)
4633 env
->CreateDir(fn
+ ".wal");
4634 if (cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths").length())
4635 env
->CreateDir(fn
+ ".slow");
4637 } else if (create
) {
4638 int r
= ::mkdir(fn
.c_str(), 0755);
4641 if (r
< 0 && r
!= -EEXIST
) {
4642 derr
<< __func__
<< " failed to create " << fn
<< ": " << cpp_strerror(r
)
4648 if (cct
->_conf
->rocksdb_separate_wal_dir
) {
4649 string walfn
= path
+ "/db.wal";
4650 r
= ::mkdir(walfn
.c_str(), 0755);
4653 if (r
< 0 && r
!= -EEXIST
) {
4654 derr
<< __func__
<< " failed to create " << walfn
4655 << ": " << cpp_strerror(r
)
4662 db
= KeyValueDB::create(cct
,
4665 static_cast<void*>(env
));
4667 derr
<< __func__
<< " error creating db" << dendl
;
4673 // delete env manually here since we can't depend on db to do this
4680 FreelistManager::setup_merge_operators(db
);
4681 db
->set_merge_operator(PREFIX_STAT
, merge_op
);
4683 db
->set_cache_size(cache_size
* cache_kv_ratio
);
4685 if (kv_backend
== "rocksdb")
4686 options
= cct
->_conf
->bluestore_rocksdb_options
;
4689 r
= db
->create_and_open(err
);
4693 derr
<< __func__
<< " erroring opening db: " << err
.str() << dendl
;
4703 dout(1) << __func__
<< " opened " << kv_backend
4704 << " path " << fn
<< " options " << options
<< dendl
;
4714 void BlueStore::_close_db()
4726 int BlueStore::_reconcile_bluefs_freespace()
4728 dout(10) << __func__
<< dendl
;
4729 interval_set
<uint64_t> bset
;
4730 int r
= bluefs
->get_block_extents(bluefs_shared_bdev
, &bset
);
4732 if (bset
== bluefs_extents
) {
4733 dout(10) << __func__
<< " we agree bluefs has 0x" << std::hex
<< bset
4734 << std::dec
<< dendl
;
4737 dout(10) << __func__
<< " bluefs says 0x" << std::hex
<< bset
<< std::dec
4739 dout(10) << __func__
<< " super says 0x" << std::hex
<< bluefs_extents
4740 << std::dec
<< dendl
;
4742 interval_set
<uint64_t> overlap
;
4743 overlap
.intersection_of(bset
, bluefs_extents
);
4745 bset
.subtract(overlap
);
4746 if (!bset
.empty()) {
4747 derr
<< __func__
<< " bluefs extra 0x" << std::hex
<< bset
<< std::dec
4752 interval_set
<uint64_t> super_extra
;
4753 super_extra
= bluefs_extents
;
4754 super_extra
.subtract(overlap
);
4755 if (!super_extra
.empty()) {
4756 // This is normal: it can happen if we commit to give extents to
4757 // bluefs and we crash before bluefs commits that it owns them.
4758 dout(10) << __func__
<< " super extra " << super_extra
<< dendl
;
4759 for (interval_set
<uint64_t>::iterator p
= super_extra
.begin();
4760 p
!= super_extra
.end();
4762 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.get_start(), p
.get_len());
4769 int BlueStore::_balance_bluefs_freespace(PExtentVector
*extents
)
4774 vector
<pair
<uint64_t,uint64_t>> bluefs_usage
; // <free, total> ...
4775 bluefs
->get_usage(&bluefs_usage
);
4776 assert(bluefs_usage
.size() > bluefs_shared_bdev
);
4778 // fixme: look at primary bdev only for now
4779 uint64_t bluefs_free
= bluefs_usage
[bluefs_shared_bdev
].first
;
4780 uint64_t bluefs_total
= bluefs_usage
[bluefs_shared_bdev
].second
;
4781 float bluefs_free_ratio
= (float)bluefs_free
/ (float)bluefs_total
;
4783 uint64_t my_free
= alloc
->get_free();
4784 uint64_t total
= bdev
->get_size();
4785 float my_free_ratio
= (float)my_free
/ (float)total
;
4787 uint64_t total_free
= bluefs_free
+ my_free
;
4789 float bluefs_ratio
= (float)bluefs_free
/ (float)total_free
;
4791 dout(10) << __func__
4792 << " bluefs " << pretty_si_t(bluefs_free
)
4793 << " free (" << bluefs_free_ratio
4794 << ") bluestore " << pretty_si_t(my_free
)
4795 << " free (" << my_free_ratio
4796 << "), bluefs_ratio " << bluefs_ratio
4800 uint64_t reclaim
= 0;
4801 if (bluefs_ratio
< cct
->_conf
->bluestore_bluefs_min_ratio
) {
4802 gift
= cct
->_conf
->bluestore_bluefs_gift_ratio
* total_free
;
4803 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4804 << " < min_ratio " << cct
->_conf
->bluestore_bluefs_min_ratio
4805 << ", should gift " << pretty_si_t(gift
) << dendl
;
4806 } else if (bluefs_ratio
> cct
->_conf
->bluestore_bluefs_max_ratio
) {
4807 reclaim
= cct
->_conf
->bluestore_bluefs_reclaim_ratio
* total_free
;
4808 if (bluefs_total
- reclaim
< cct
->_conf
->bluestore_bluefs_min
)
4809 reclaim
= bluefs_total
- cct
->_conf
->bluestore_bluefs_min
;
4810 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4811 << " > max_ratio " << cct
->_conf
->bluestore_bluefs_max_ratio
4812 << ", should reclaim " << pretty_si_t(reclaim
) << dendl
;
4814 if (bluefs_total
< cct
->_conf
->bluestore_bluefs_min
&&
4815 cct
->_conf
->bluestore_bluefs_min
<
4816 (uint64_t)(cct
->_conf
->bluestore_bluefs_max_ratio
* total_free
)) {
4817 uint64_t g
= cct
->_conf
->bluestore_bluefs_min
- bluefs_total
;
4818 dout(10) << __func__
<< " bluefs_total " << bluefs_total
4819 << " < min " << cct
->_conf
->bluestore_bluefs_min
4820 << ", should gift " << pretty_si_t(g
) << dendl
;
4827 // round up to alloc size
4828 gift
= P2ROUNDUP(gift
, cct
->_conf
->bluefs_alloc_size
);
4830 // hard cap to fit into 32 bits
4831 gift
= MIN(gift
, 1ull<<31);
4832 dout(10) << __func__
<< " gifting " << gift
4833 << " (" << pretty_si_t(gift
) << ")" << dendl
;
4835 // fixme: just do one allocation to start...
4836 int r
= alloc
->reserve(gift
);
4839 AllocExtentVector exts
;
4840 int64_t alloc_len
= alloc
->allocate(gift
, cct
->_conf
->bluefs_alloc_size
,
4843 if (alloc_len
< (int64_t)gift
) {
4844 derr
<< __func__
<< " allocate failed on 0x" << std::hex
<< gift
4845 << " min_alloc_size 0x" << min_alloc_size
<< std::dec
<< dendl
;
4847 assert(0 == "allocate failed, wtf");
4850 for (auto& p
: exts
) {
4851 bluestore_pextent_t e
= bluestore_pextent_t(p
);
4852 dout(1) << __func__
<< " gifting " << e
<< " to bluefs" << dendl
;
4853 extents
->push_back(e
);
4860 // reclaim from bluefs?
4862 // round up to alloc size
4863 reclaim
= P2ROUNDUP(reclaim
, cct
->_conf
->bluefs_alloc_size
);
4865 // hard cap to fit into 32 bits
4866 reclaim
= MIN(reclaim
, 1ull<<31);
4867 dout(10) << __func__
<< " reclaiming " << reclaim
4868 << " (" << pretty_si_t(reclaim
) << ")" << dendl
;
4870 while (reclaim
> 0) {
4871 // NOTE: this will block and do IO.
4872 AllocExtentVector extents
;
4873 int r
= bluefs
->reclaim_blocks(bluefs_shared_bdev
, reclaim
,
4876 derr
<< __func__
<< " failed to reclaim space from bluefs"
4880 for (auto e
: extents
) {
4881 bluefs_extents
.erase(e
.offset
, e
.length
);
4882 bluefs_extents_reclaiming
.insert(e
.offset
, e
.length
);
4883 reclaim
-= e
.length
;
4893 void BlueStore::_commit_bluefs_freespace(
4894 const PExtentVector
& bluefs_gift_extents
)
4896 dout(10) << __func__
<< dendl
;
4897 for (auto& p
: bluefs_gift_extents
) {
4898 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.offset
, p
.length
);
4902 int BlueStore::_open_collections(int *errors
)
4904 assert(coll_map
.empty());
4905 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_COLL
);
4906 for (it
->upper_bound(string());
4910 if (cid
.parse(it
->key())) {
4914 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
4916 bufferlist bl
= it
->value();
4917 bufferlist::iterator p
= bl
.begin();
4919 ::decode(c
->cnode
, p
);
4920 } catch (buffer::error
& e
) {
4921 derr
<< __func__
<< " failed to decode cnode, key:"
4922 << pretty_binary_string(it
->key()) << dendl
;
4925 dout(20) << __func__
<< " opened " << cid
<< " " << c
<< dendl
;
4928 derr
<< __func__
<< " unrecognized collection " << it
->key() << dendl
;
4936 void BlueStore::_open_statfs()
4939 int r
= db
->get(PREFIX_STAT
, "bluestore_statfs", &bl
);
4941 if (size_t(bl
.length()) >= sizeof(vstatfs
.values
)) {
4942 auto it
= bl
.begin();
4945 dout(10) << __func__
<< " store_statfs is corrupt, using empty" << dendl
;
4949 dout(10) << __func__
<< " store_statfs missed, using empty" << dendl
;
4953 int BlueStore::_setup_block_symlink_or_file(
4959 dout(20) << __func__
<< " name " << name
<< " path " << epath
4960 << " size " << size
<< " create=" << (int)create
<< dendl
;
4965 if (epath
.length()) {
4966 r
= ::symlinkat(epath
.c_str(), path_fd
, name
.c_str());
4969 derr
<< __func__
<< " failed to create " << name
<< " symlink to "
4970 << epath
<< ": " << cpp_strerror(r
) << dendl
;
4974 if (!epath
.compare(0, strlen(SPDK_PREFIX
), SPDK_PREFIX
)) {
4975 int fd
= ::openat(path_fd
, epath
.c_str(), flags
, 0644);
4978 derr
<< __func__
<< " failed to open " << epath
<< " file: "
4979 << cpp_strerror(r
) << dendl
;
4982 string serial_number
= epath
.substr(strlen(SPDK_PREFIX
));
4983 r
= ::write(fd
, serial_number
.c_str(), serial_number
.size());
4984 assert(r
== (int)serial_number
.size());
4985 dout(1) << __func__
<< " created " << name
<< " symlink to "
4987 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4991 int fd
= ::openat(path_fd
, name
.c_str(), flags
, 0644);
4993 // block file is present
4995 int r
= ::fstat(fd
, &st
);
4997 S_ISREG(st
.st_mode
) && // if it is a regular file
4998 st
.st_size
== 0) { // and is 0 bytes
4999 r
= ::ftruncate(fd
, size
);
5002 derr
<< __func__
<< " failed to resize " << name
<< " file to "
5003 << size
<< ": " << cpp_strerror(r
) << dendl
;
5004 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5008 if (cct
->_conf
->bluestore_block_preallocate_file
) {
5009 #ifdef HAVE_POSIX_FALLOCATE
5010 r
= ::posix_fallocate(fd
, 0, size
);
5012 derr
<< __func__
<< " failed to prefallocate " << name
<< " file to "
5013 << size
<< ": " << cpp_strerror(r
) << dendl
;
5014 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5018 char data
[1024*128];
5019 for (uint64_t off
= 0; off
< size
; off
+= sizeof(data
)) {
5020 if (off
+ sizeof(data
) > size
)
5021 r
= ::write(fd
, data
, size
- off
);
5023 r
= ::write(fd
, data
, sizeof(data
));
5026 derr
<< __func__
<< " failed to prefallocate w/ write " << name
<< " file to "
5027 << size
<< ": " << cpp_strerror(r
) << dendl
;
5028 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5034 dout(1) << __func__
<< " resized " << name
<< " file to "
5035 << pretty_si_t(size
) << "B" << dendl
;
5037 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5041 derr
<< __func__
<< " failed to open " << name
<< " file: "
5042 << cpp_strerror(r
) << dendl
;
5050 int BlueStore::mkfs()
5052 dout(1) << __func__
<< " path " << path
<< dendl
;
5058 r
= read_meta("mkfs_done", &done
);
5060 dout(1) << __func__
<< " already created" << dendl
;
5061 if (cct
->_conf
->bluestore_fsck_on_mkfs
) {
5062 r
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5064 derr
<< __func__
<< " fsck found fatal error: " << cpp_strerror(r
)
5069 derr
<< __func__
<< " fsck found " << r
<< " errors" << dendl
;
5073 return r
; // idempotent
5079 r
= read_meta("type", &type
);
5081 if (type
!= "bluestore") {
5082 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5086 r
= write_meta("type", "bluestore");
5092 freelist_type
= "bitmap";
5098 r
= _open_fsid(true);
5104 goto out_close_fsid
;
5106 r
= _read_fsid(&old_fsid
);
5107 if (r
< 0 || old_fsid
.is_zero()) {
5108 if (fsid
.is_zero()) {
5109 fsid
.generate_random();
5110 dout(1) << __func__
<< " generated fsid " << fsid
<< dendl
;
5112 dout(1) << __func__
<< " using provided fsid " << fsid
<< dendl
;
5114 // we'll write it later.
5116 if (!fsid
.is_zero() && fsid
!= old_fsid
) {
5117 derr
<< __func__
<< " on-disk fsid " << old_fsid
5118 << " != provided " << fsid
<< dendl
;
5120 goto out_close_fsid
;
5125 r
= _setup_block_symlink_or_file("block", cct
->_conf
->bluestore_block_path
,
5126 cct
->_conf
->bluestore_block_size
,
5127 cct
->_conf
->bluestore_block_create
);
5129 goto out_close_fsid
;
5130 if (cct
->_conf
->bluestore_bluefs
) {
5131 r
= _setup_block_symlink_or_file("block.wal", cct
->_conf
->bluestore_block_wal_path
,
5132 cct
->_conf
->bluestore_block_wal_size
,
5133 cct
->_conf
->bluestore_block_wal_create
);
5135 goto out_close_fsid
;
5136 r
= _setup_block_symlink_or_file("block.db", cct
->_conf
->bluestore_block_db_path
,
5137 cct
->_conf
->bluestore_block_db_size
,
5138 cct
->_conf
->bluestore_block_db_create
);
5140 goto out_close_fsid
;
5143 r
= _open_bdev(true);
5145 goto out_close_fsid
;
5149 goto out_close_bdev
;
5156 KeyValueDB::Transaction t
= db
->get_transaction();
5159 ::encode((uint64_t)0, bl
);
5160 t
->set(PREFIX_SUPER
, "nid_max", bl
);
5161 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
5164 // choose min_alloc_size
5165 if (cct
->_conf
->bluestore_min_alloc_size
) {
5166 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size
;
5169 if (bdev
->is_rotational()) {
5170 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_hdd
;
5172 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_ssd
;
5176 // make sure min_alloc_size is power of 2 aligned.
5177 if (!ISP2(min_alloc_size
)) {
5178 derr
<< __func__
<< " min_alloc_size 0x"
5179 << std::hex
<< min_alloc_size
<< std::dec
5180 << " is not power of 2 aligned!"
5188 ::encode((uint64_t)min_alloc_size
, bl
);
5189 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
5192 ondisk_format
= latest_ondisk_format
;
5193 _prepare_ondisk_format_super(t
);
5194 db
->submit_transaction_sync(t
);
5198 r
= write_meta("kv_backend", cct
->_conf
->bluestore_kvbackend
);
5202 r
= write_meta("bluefs", stringify((int)cct
->_conf
->bluestore_bluefs
));
5206 if (fsid
!= old_fsid
) {
5209 derr
<< __func__
<< " error writing fsid: " << cpp_strerror(r
) << dendl
;
5226 cct
->_conf
->bluestore_fsck_on_mkfs
) {
5227 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5231 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5237 // indicate success by writing the 'mkfs_done' file
5238 r
= write_meta("mkfs_done", "yes");
5242 derr
<< __func__
<< " failed, " << cpp_strerror(r
) << dendl
;
5244 dout(0) << __func__
<< " success" << dendl
;
5249 void BlueStore::set_cache_shards(unsigned num
)
5251 dout(10) << __func__
<< " " << num
<< dendl
;
5252 size_t old
= cache_shards
.size();
5254 cache_shards
.resize(num
);
5255 for (unsigned i
= old
; i
< num
; ++i
) {
5256 cache_shards
[i
] = Cache::create(cct
, cct
->_conf
->bluestore_cache_type
,
5261 int BlueStore::_mount(bool kv_only
)
5263 dout(1) << __func__
<< " path " << path
<< dendl
;
5267 int r
= read_meta("type", &type
);
5269 derr
<< __func__
<< " failed to load os-type: " << cpp_strerror(r
)
5274 if (type
!= "bluestore") {
5275 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5280 if (cct
->_conf
->bluestore_fsck_on_mount
) {
5281 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mount_deep
);
5285 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5290 int r
= _open_path();
5293 r
= _open_fsid(false);
5297 r
= _read_fsid(&fsid
);
5305 r
= _open_bdev(false);
5309 r
= _open_db(false);
5316 r
= _open_super_meta();
5320 r
= _open_fm(false);
5328 r
= _open_collections();
5332 r
= _reload_logger();
5337 r
= _reconcile_bluefs_freespace();
5344 r
= _deferred_replay();
5348 mempool_thread
.init();
5373 int BlueStore::umount()
5376 dout(1) << __func__
<< dendl
;
5379 _osr_unregister_all();
5381 mempool_thread
.shutdown();
5383 dout(20) << __func__
<< " stopping kv thread" << dendl
;
5385 _reap_collections();
5387 dout(20) << __func__
<< " closing" << dendl
;
5397 if (cct
->_conf
->bluestore_fsck_on_umount
) {
5398 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_umount_deep
);
5402 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5409 static void apply(uint64_t off
,
5411 uint64_t granularity
,
5412 BlueStore::mempool_dynamic_bitset
&bitset
,
5414 std::function
<void(uint64_t,
5415 BlueStore::mempool_dynamic_bitset
&)> f
) {
5416 auto end
= ROUND_UP_TO(off
+ len
, granularity
);
5418 uint64_t pos
= off
/ granularity
;
5424 int BlueStore::_fsck_check_extents(
5425 const ghobject_t
& oid
,
5426 const PExtentVector
& extents
,
5428 mempool_dynamic_bitset
&used_blocks
,
5429 store_statfs_t
& expected_statfs
)
5431 dout(30) << __func__
<< " oid " << oid
<< " extents " << extents
<< dendl
;
5433 for (auto e
: extents
) {
5436 expected_statfs
.allocated
+= e
.length
;
5438 expected_statfs
.compressed_allocated
+= e
.length
;
5440 bool already
= false;
5442 e
.offset
, e
.length
, block_size
, used_blocks
, __func__
,
5443 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5450 derr
<< " " << oid
<< " extent " << e
5451 << " or a subset is already allocated" << dendl
;
5454 if (e
.end() > bdev
->get_size()) {
5455 derr
<< " " << oid
<< " extent " << e
5456 << " past end of block device" << dendl
;
5463 int BlueStore::fsck(bool deep
)
5465 dout(1) << __func__
<< (deep
? " (deep)" : " (shallow)") << " start" << dendl
;
5468 typedef btree::btree_set
<
5469 uint64_t,std::less
<uint64_t>,
5470 mempool::bluestore_fsck::pool_allocator
<uint64_t>> uint64_t_btree_t
;
5471 uint64_t_btree_t used_nids
;
5472 uint64_t_btree_t used_omap_head
;
5473 uint64_t_btree_t used_sbids
;
5475 mempool_dynamic_bitset used_blocks
;
5476 KeyValueDB::Iterator it
;
5477 store_statfs_t expected_statfs
, actual_statfs
;
5479 list
<ghobject_t
> oids
;
5481 bluestore_extent_ref_map_t ref_map
;
5484 mempool::bluestore_fsck::map
<uint64_t,sb_info_t
> sb_info
;
5486 uint64_t num_objects
= 0;
5487 uint64_t num_extents
= 0;
5488 uint64_t num_blobs
= 0;
5489 uint64_t num_spanning_blobs
= 0;
5490 uint64_t num_shared_blobs
= 0;
5491 uint64_t num_sharded_objects
= 0;
5492 uint64_t num_object_shards
= 0;
5494 utime_t start
= ceph_clock_now();
5496 int r
= _open_path();
5499 r
= _open_fsid(false);
5503 r
= _read_fsid(&fsid
);
5511 r
= _open_bdev(false);
5515 r
= _open_db(false);
5519 r
= _open_super_meta();
5523 r
= _open_fm(false);
5531 r
= _open_collections(&errors
);
5535 mempool_thread
.init();
5537 // we need finishers and kv_{sync,finalize}_thread *just* for replay
5539 r
= _deferred_replay();
5544 used_blocks
.resize(bdev
->get_size() / block_size
);
5546 0, SUPER_RESERVED
, block_size
, used_blocks
, "0~SUPER_RESERVED",
5547 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5553 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
5555 e
.get_start(), e
.get_len(), block_size
, used_blocks
, "bluefs",
5556 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5569 // get expected statfs; fill unaffected fields to be able to compare
5571 statfs(&actual_statfs
);
5572 expected_statfs
.total
= actual_statfs
.total
;
5573 expected_statfs
.available
= actual_statfs
.available
;
5576 dout(1) << __func__
<< " walking object keyspace" << dendl
;
5577 it
= db
->get_iterator(PREFIX_OBJ
);
5581 mempool::bluestore_fsck::list
<string
> expecting_shards
;
5582 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5583 if (g_conf
->bluestore_debug_fsck_abort
) {
5586 dout(30) << " key " << pretty_binary_string(it
->key()) << dendl
;
5587 if (is_extent_shard_key(it
->key())) {
5588 while (!expecting_shards
.empty() &&
5589 expecting_shards
.front() < it
->key()) {
5590 derr
<< __func__
<< " error: missing shard key "
5591 << pretty_binary_string(expecting_shards
.front())
5594 expecting_shards
.pop_front();
5596 if (!expecting_shards
.empty() &&
5597 expecting_shards
.front() == it
->key()) {
5599 expecting_shards
.pop_front();
5605 get_key_extent_shard(it
->key(), &okey
, &offset
);
5606 derr
<< __func__
<< " error: stray shard 0x" << std::hex
<< offset
5607 << std::dec
<< dendl
;
5608 if (expecting_shards
.empty()) {
5609 derr
<< __func__
<< " error: " << pretty_binary_string(it
->key())
5610 << " is unexpected" << dendl
;
5614 while (expecting_shards
.front() > it
->key()) {
5615 derr
<< __func__
<< " error: saw " << pretty_binary_string(it
->key())
5617 derr
<< __func__
<< " error: exp "
5618 << pretty_binary_string(expecting_shards
.front()) << dendl
;
5620 expecting_shards
.pop_front();
5621 if (expecting_shards
.empty()) {
5629 int r
= get_key_object(it
->key(), &oid
);
5631 derr
<< __func__
<< " error: bad object key "
5632 << pretty_binary_string(it
->key()) << dendl
;
5637 oid
.shard_id
!= pgid
.shard
||
5638 oid
.hobj
.pool
!= (int64_t)pgid
.pool() ||
5639 !c
->contains(oid
)) {
5641 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
=
5643 p
!= coll_map
.end();
5645 if (p
->second
->contains(oid
)) {
5651 derr
<< __func__
<< " error: stray object " << oid
5652 << " not owned by any collection" << dendl
;
5656 c
->cid
.is_pg(&pgid
);
5657 dout(20) << __func__
<< " collection " << c
->cid
<< dendl
;
5660 if (!expecting_shards
.empty()) {
5661 for (auto &k
: expecting_shards
) {
5662 derr
<< __func__
<< " error: missing shard key "
5663 << pretty_binary_string(k
) << dendl
;
5666 expecting_shards
.clear();
5669 dout(10) << __func__
<< " " << oid
<< dendl
;
5670 RWLock::RLocker
l(c
->lock
);
5671 OnodeRef o
= c
->get_onode(oid
, false);
5673 if (o
->onode
.nid
> nid_max
) {
5674 derr
<< __func__
<< " error: " << oid
<< " nid " << o
->onode
.nid
5675 << " > nid_max " << nid_max
<< dendl
;
5678 if (used_nids
.count(o
->onode
.nid
)) {
5679 derr
<< __func__
<< " error: " << oid
<< " nid " << o
->onode
.nid
5680 << " already in use" << dendl
;
5682 continue; // go for next object
5684 used_nids
.insert(o
->onode
.nid
);
5687 num_spanning_blobs
+= o
->extent_map
.spanning_blob_map
.size();
5688 o
->extent_map
.fault_range(db
, 0, OBJECT_MAX_SIZE
);
5691 if (!o
->extent_map
.shards
.empty()) {
5692 ++num_sharded_objects
;
5693 num_object_shards
+= o
->extent_map
.shards
.size();
5695 for (auto& s
: o
->extent_map
.shards
) {
5696 dout(20) << __func__
<< " shard " << *s
.shard_info
<< dendl
;
5697 expecting_shards
.push_back(string());
5698 get_extent_shard_key(o
->key
, s
.shard_info
->offset
,
5699 &expecting_shards
.back());
5700 if (s
.shard_info
->offset
>= o
->onode
.size
) {
5701 derr
<< __func__
<< " error: " << oid
<< " shard 0x" << std::hex
5702 << s
.shard_info
->offset
<< " past EOF at 0x" << o
->onode
.size
5703 << std::dec
<< dendl
;
5708 map
<BlobRef
,bluestore_blob_t::unused_t
> referenced
;
5710 mempool::bluestore_fsck::map
<BlobRef
,
5711 bluestore_blob_use_tracker_t
> ref_map
;
5712 for (auto& l
: o
->extent_map
.extent_map
) {
5713 dout(20) << __func__
<< " " << l
<< dendl
;
5714 if (l
.logical_offset
< pos
) {
5715 derr
<< __func__
<< " error: " << oid
<< " lextent at 0x"
5716 << std::hex
<< l
.logical_offset
5717 << " overlaps with the previous, which ends at 0x" << pos
5718 << std::dec
<< dendl
;
5721 if (o
->extent_map
.spans_shard(l
.logical_offset
, l
.length
)) {
5722 derr
<< __func__
<< " error: " << oid
<< " lextent at 0x"
5723 << std::hex
<< l
.logical_offset
<< "~" << l
.length
5724 << " spans a shard boundary"
5725 << std::dec
<< dendl
;
5728 pos
= l
.logical_offset
+ l
.length
;
5729 expected_statfs
.stored
+= l
.length
;
5731 const bluestore_blob_t
& blob
= l
.blob
->get_blob();
5733 auto& ref
= ref_map
[l
.blob
];
5734 if (ref
.is_empty()) {
5735 uint32_t min_release_size
= blob
.get_release_size(min_alloc_size
);
5736 uint32_t l
= blob
.get_logical_length();
5737 ref
.init(l
, min_release_size
);
5743 if (blob
.has_unused()) {
5744 auto p
= referenced
.find(l
.blob
);
5745 bluestore_blob_t::unused_t
*pu
;
5746 if (p
== referenced
.end()) {
5747 pu
= &referenced
[l
.blob
];
5751 uint64_t blob_len
= blob
.get_logical_length();
5752 assert((blob_len
% (sizeof(*pu
)*8)) == 0);
5753 assert(l
.blob_offset
+ l
.length
<= blob_len
);
5754 uint64_t chunk_size
= blob_len
/ (sizeof(*pu
)*8);
5755 uint64_t start
= l
.blob_offset
/ chunk_size
;
5757 ROUND_UP_TO(l
.blob_offset
+ l
.length
, chunk_size
) / chunk_size
;
5758 for (auto i
= start
; i
< end
; ++i
) {
5763 for (auto &i
: referenced
) {
5764 dout(20) << __func__
<< " referenced 0x" << std::hex
<< i
.second
5765 << std::dec
<< " for " << *i
.first
<< dendl
;
5766 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5767 if (i
.second
& blob
.unused
) {
5768 derr
<< __func__
<< " error: " << oid
<< " blob claims unused 0x"
5769 << std::hex
<< blob
.unused
5770 << " but extents reference 0x" << i
.second
5771 << " on blob " << *i
.first
<< dendl
;
5774 if (blob
.has_csum()) {
5775 uint64_t blob_len
= blob
.get_logical_length();
5776 uint64_t unused_chunk_size
= blob_len
/ (sizeof(blob
.unused
)*8);
5777 unsigned csum_count
= blob
.get_csum_count();
5778 unsigned csum_chunk_size
= blob
.get_csum_chunk_size();
5779 for (unsigned p
= 0; p
< csum_count
; ++p
) {
5780 unsigned pos
= p
* csum_chunk_size
;
5781 unsigned firstbit
= pos
/ unused_chunk_size
; // [firstbit,lastbit]
5782 unsigned lastbit
= (pos
+ csum_chunk_size
- 1) / unused_chunk_size
;
5783 unsigned mask
= 1u << firstbit
;
5784 for (unsigned b
= firstbit
+ 1; b
<= lastbit
; ++b
) {
5787 if ((blob
.unused
& mask
) == mask
) {
5788 // this csum chunk region is marked unused
5789 if (blob
.get_csum_item(p
) != 0) {
5790 derr
<< __func__
<< " error: " << oid
5791 << " blob claims csum chunk 0x" << std::hex
<< pos
5792 << "~" << csum_chunk_size
5793 << " is unused (mask 0x" << mask
<< " of unused 0x"
5794 << blob
.unused
<< ") but csum is non-zero 0x"
5795 << blob
.get_csum_item(p
) << std::dec
<< " on blob "
5796 << *i
.first
<< dendl
;
5803 for (auto &i
: ref_map
) {
5805 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5806 bool equal
= i
.first
->get_blob_use_tracker().equal(i
.second
);
5808 derr
<< __func__
<< " error: " << oid
<< " blob " << *i
.first
5809 << " doesn't match expected ref_map " << i
.second
<< dendl
;
5812 if (blob
.is_compressed()) {
5813 expected_statfs
.compressed
+= blob
.get_compressed_payload_length();
5814 expected_statfs
.compressed_original
+=
5815 i
.first
->get_referenced_bytes();
5817 if (blob
.is_shared()) {
5818 if (i
.first
->shared_blob
->get_sbid() > blobid_max
) {
5819 derr
<< __func__
<< " error: " << oid
<< " blob " << blob
5820 << " sbid " << i
.first
->shared_blob
->get_sbid() << " > blobid_max "
5821 << blobid_max
<< dendl
;
5823 } else if (i
.first
->shared_blob
->get_sbid() == 0) {
5824 derr
<< __func__
<< " error: " << oid
<< " blob " << blob
5825 << " marked as shared but has uninitialized sbid"
5829 sb_info_t
& sbi
= sb_info
[i
.first
->shared_blob
->get_sbid()];
5830 sbi
.sb
= i
.first
->shared_blob
;
5831 sbi
.oids
.push_back(oid
);
5832 sbi
.compressed
= blob
.is_compressed();
5833 for (auto e
: blob
.get_extents()) {
5835 sbi
.ref_map
.get(e
.offset
, e
.length
);
5839 errors
+= _fsck_check_extents(oid
, blob
.get_extents(),
5840 blob
.is_compressed(),
5847 int r
= _do_read(c
.get(), o
, 0, o
->onode
.size
, bl
, 0);
5850 derr
<< __func__
<< " error: " << oid
<< " error during read: "
5851 << cpp_strerror(r
) << dendl
;
5855 if (o
->onode
.has_omap()) {
5856 if (used_omap_head
.count(o
->onode
.nid
)) {
5857 derr
<< __func__
<< " error: " << oid
<< " omap_head " << o
->onode
.nid
5858 << " already in use" << dendl
;
5861 used_omap_head
.insert(o
->onode
.nid
);
5866 dout(1) << __func__
<< " checking shared_blobs" << dendl
;
5867 it
= db
->get_iterator(PREFIX_SHARED_BLOB
);
5869 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5870 string key
= it
->key();
5872 if (get_key_shared_blob(key
, &sbid
)) {
5873 derr
<< __func__
<< " error: bad key '" << key
5874 << "' in shared blob namespace" << dendl
;
5878 auto p
= sb_info
.find(sbid
);
5879 if (p
== sb_info
.end()) {
5880 derr
<< __func__
<< " error: found stray shared blob data for sbid 0x"
5881 << std::hex
<< sbid
<< std::dec
<< dendl
;
5885 sb_info_t
& sbi
= p
->second
;
5886 bluestore_shared_blob_t
shared_blob(sbid
);
5887 bufferlist bl
= it
->value();
5888 bufferlist::iterator blp
= bl
.begin();
5889 ::decode(shared_blob
, blp
);
5890 dout(20) << __func__
<< " " << *sbi
.sb
<< " " << shared_blob
<< dendl
;
5891 if (shared_blob
.ref_map
!= sbi
.ref_map
) {
5892 derr
<< __func__
<< " error: shared blob 0x" << std::hex
<< sbid
5893 << std::dec
<< " ref_map " << shared_blob
.ref_map
5894 << " != expected " << sbi
.ref_map
<< dendl
;
5897 PExtentVector extents
;
5898 for (auto &r
: shared_blob
.ref_map
.ref_map
) {
5899 extents
.emplace_back(bluestore_pextent_t(r
.first
, r
.second
.length
));
5901 errors
+= _fsck_check_extents(p
->second
.oids
.front(),
5903 p
->second
.compressed
,
5904 used_blocks
, expected_statfs
);
5909 for (auto &p
: sb_info
) {
5910 derr
<< __func__
<< " error: shared_blob 0x" << p
.first
5911 << " key is missing (" << *p
.second
.sb
<< ")" << dendl
;
5914 if (!(actual_statfs
== expected_statfs
)) {
5915 derr
<< __func__
<< " error: actual " << actual_statfs
5916 << " != expected " << expected_statfs
<< dendl
;
5920 dout(1) << __func__
<< " checking for stray omap data" << dendl
;
5921 it
= db
->get_iterator(PREFIX_OMAP
);
5923 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5925 _key_decode_u64(it
->key().c_str(), &omap_head
);
5926 if (used_omap_head
.count(omap_head
) == 0) {
5927 derr
<< __func__
<< " error: found stray omap data on omap_head "
5928 << omap_head
<< dendl
;
5934 dout(1) << __func__
<< " checking deferred events" << dendl
;
5935 it
= db
->get_iterator(PREFIX_DEFERRED
);
5937 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5938 bufferlist bl
= it
->value();
5939 bufferlist::iterator p
= bl
.begin();
5940 bluestore_deferred_transaction_t wt
;
5943 } catch (buffer::error
& e
) {
5944 derr
<< __func__
<< " error: failed to decode deferred txn "
5945 << pretty_binary_string(it
->key()) << dendl
;
5949 dout(20) << __func__
<< " deferred " << wt
.seq
5950 << " ops " << wt
.ops
.size()
5951 << " released 0x" << std::hex
<< wt
.released
<< std::dec
<< dendl
;
5952 for (auto e
= wt
.released
.begin(); e
!= wt
.released
.end(); ++e
) {
5954 e
.get_start(), e
.get_len(), block_size
, used_blocks
, "deferred",
5955 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5963 dout(1) << __func__
<< " checking freelist vs allocated" << dendl
;
5965 // remove bluefs_extents from used set since the freelist doesn't
5966 // know they are allocated.
5967 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
5969 e
.get_start(), e
.get_len(), block_size
, used_blocks
, "bluefs_extents",
5970 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5975 fm
->enumerate_reset();
5976 uint64_t offset
, length
;
5977 while (fm
->enumerate_next(&offset
, &length
)) {
5978 bool intersects
= false;
5980 offset
, length
, block_size
, used_blocks
, "free",
5981 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5990 derr
<< __func__
<< " error: free extent 0x" << std::hex
<< offset
5991 << "~" << length
<< std::dec
5992 << " intersects allocated blocks" << dendl
;
5996 fm
->enumerate_reset();
5997 size_t count
= used_blocks
.count();
5998 if (used_blocks
.size() != count
) {
5999 assert(used_blocks
.size() > count
);
6000 derr
<< __func__
<< " error: leaked some space;"
6001 << (used_blocks
.size() - count
) * min_alloc_size
6002 << " bytes leaked" << dendl
;
6008 mempool_thread
.shutdown();
6015 it
.reset(); // before db is closed
6024 // fatal errors take precedence
6028 dout(2) << __func__
<< " " << num_objects
<< " objects, "
6029 << num_sharded_objects
<< " of them sharded. "
6031 dout(2) << __func__
<< " " << num_extents
<< " extents to "
6032 << num_blobs
<< " blobs, "
6033 << num_spanning_blobs
<< " spanning, "
6034 << num_shared_blobs
<< " shared."
6037 utime_t duration
= ceph_clock_now() - start
;
6038 dout(1) << __func__
<< " finish with " << errors
<< " errors in "
6039 << duration
<< " seconds" << dendl
;
6043 void BlueStore::collect_metadata(map
<string
,string
> *pm
)
6045 dout(10) << __func__
<< dendl
;
6046 bdev
->collect_metadata("bluestore_bdev_", pm
);
6048 (*pm
)["bluefs"] = "1";
6049 (*pm
)["bluefs_single_shared_device"] = stringify((int)bluefs_single_shared_device
);
6050 bluefs
->collect_metadata(pm
);
6052 (*pm
)["bluefs"] = "0";
6056 int BlueStore::statfs(struct store_statfs_t
*buf
)
6059 buf
->total
= bdev
->get_size();
6060 buf
->available
= alloc
->get_free();
6063 // part of our shared device is "free" according to BlueFS
6064 // Don't include bluestore_bluefs_min because that space can't
6065 // be used for any other purpose.
6066 buf
->available
+= bluefs
->get_free(bluefs_shared_bdev
) - cct
->_conf
->bluestore_bluefs_min
;
6068 // include dedicated db, too, if that isn't the shared device.
6069 if (bluefs_shared_bdev
!= BlueFS::BDEV_DB
) {
6070 buf
->total
+= bluefs
->get_total(BlueFS::BDEV_DB
);
6075 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
6077 buf
->allocated
= vstatfs
.allocated();
6078 buf
->stored
= vstatfs
.stored();
6079 buf
->compressed
= vstatfs
.compressed();
6080 buf
->compressed_original
= vstatfs
.compressed_original();
6081 buf
->compressed_allocated
= vstatfs
.compressed_allocated();
6084 dout(20) << __func__
<< *buf
<< dendl
;
6091 BlueStore::CollectionRef
BlueStore::_get_collection(const coll_t
& cid
)
6093 RWLock::RLocker
l(coll_lock
);
6094 ceph::unordered_map
<coll_t
,CollectionRef
>::iterator cp
= coll_map
.find(cid
);
6095 if (cp
== coll_map
.end())
6096 return CollectionRef();
6100 void BlueStore::_queue_reap_collection(CollectionRef
& c
)
6102 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6103 std::lock_guard
<std::mutex
> l(reap_lock
);
6104 removed_collections
.push_back(c
);
6107 void BlueStore::_reap_collections()
6109 list
<CollectionRef
> removed_colls
;
6111 std::lock_guard
<std::mutex
> l(reap_lock
);
6112 removed_colls
.swap(removed_collections
);
6115 bool all_reaped
= true;
6117 for (list
<CollectionRef
>::iterator p
= removed_colls
.begin();
6118 p
!= removed_colls
.end();
6120 CollectionRef c
= *p
;
6121 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6122 if (c
->onode_map
.map_any([&](OnodeRef o
) {
6124 if (o
->flushing_count
.load()) {
6125 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " " << o
->oid
6126 << " flush_txns " << o
->flushing_count
<< dendl
;
6134 c
->onode_map
.clear();
6135 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " done" << dendl
;
6139 dout(10) << __func__
<< " all reaped" << dendl
;
6143 void BlueStore::_update_cache_logger()
6145 uint64_t num_onodes
= 0;
6146 uint64_t num_extents
= 0;
6147 uint64_t num_blobs
= 0;
6148 uint64_t num_buffers
= 0;
6149 uint64_t num_buffer_bytes
= 0;
6150 for (auto c
: cache_shards
) {
6151 c
->add_stats(&num_onodes
, &num_extents
, &num_blobs
,
6152 &num_buffers
, &num_buffer_bytes
);
6154 logger
->set(l_bluestore_onodes
, num_onodes
);
6155 logger
->set(l_bluestore_extents
, num_extents
);
6156 logger
->set(l_bluestore_blobs
, num_blobs
);
6157 logger
->set(l_bluestore_buffers
, num_buffers
);
6158 logger
->set(l_bluestore_buffer_bytes
, num_buffer_bytes
);
6164 ObjectStore::CollectionHandle
BlueStore::open_collection(const coll_t
& cid
)
6166 return _get_collection(cid
);
6169 bool BlueStore::exists(const coll_t
& cid
, const ghobject_t
& oid
)
6171 CollectionHandle c
= _get_collection(cid
);
6174 return exists(c
, oid
);
6177 bool BlueStore::exists(CollectionHandle
&c_
, const ghobject_t
& oid
)
6179 Collection
*c
= static_cast<Collection
*>(c_
.get());
6180 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
6187 RWLock::RLocker
l(c
->lock
);
6188 OnodeRef o
= c
->get_onode(oid
, false);
6189 if (!o
|| !o
->exists
)
6196 int BlueStore::stat(
6198 const ghobject_t
& oid
,
6202 CollectionHandle c
= _get_collection(cid
);
6205 return stat(c
, oid
, st
, allow_eio
);
6208 int BlueStore::stat(
6209 CollectionHandle
&c_
,
6210 const ghobject_t
& oid
,
6214 Collection
*c
= static_cast<Collection
*>(c_
.get());
6217 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
6220 RWLock::RLocker
l(c
->lock
);
6221 OnodeRef o
= c
->get_onode(oid
, false);
6222 if (!o
|| !o
->exists
)
6224 st
->st_size
= o
->onode
.size
;
6225 st
->st_blksize
= 4096;
6226 st
->st_blocks
= (st
->st_size
+ st
->st_blksize
- 1) / st
->st_blksize
;
6231 if (_debug_mdata_eio(oid
)) {
6233 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6237 int BlueStore::set_collection_opts(
6239 const pool_opts_t
& opts
)
6241 CollectionHandle ch
= _get_collection(cid
);
6244 Collection
*c
= static_cast<Collection
*>(ch
.get());
6245 dout(15) << __func__
<< " " << cid
<< " options " << opts
<< dendl
;
6248 RWLock::WLocker
l(c
->lock
);
6249 c
->pool_opts
= opts
;
6253 int BlueStore::read(
6255 const ghobject_t
& oid
,
6261 CollectionHandle c
= _get_collection(cid
);
6264 return read(c
, oid
, offset
, length
, bl
, op_flags
);
6267 int BlueStore::read(
6268 CollectionHandle
&c_
,
6269 const ghobject_t
& oid
,
6275 utime_t start
= ceph_clock_now();
6276 Collection
*c
= static_cast<Collection
*>(c_
.get());
6277 const coll_t
&cid
= c
->get_cid();
6278 dout(15) << __func__
<< " " << cid
<< " " << oid
6279 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6287 RWLock::RLocker
l(c
->lock
);
6288 utime_t start1
= ceph_clock_now();
6289 OnodeRef o
= c
->get_onode(oid
, false);
6290 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start1
);
6291 if (!o
|| !o
->exists
) {
6296 if (offset
== length
&& offset
== 0)
6297 length
= o
->onode
.size
;
6299 r
= _do_read(c
, o
, offset
, length
, bl
, op_flags
);
6303 if (r
== 0 && _debug_data_eio(oid
)) {
6305 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6306 } else if (cct
->_conf
->bluestore_debug_random_read_err
&&
6307 (rand() % (int)(cct
->_conf
->bluestore_debug_random_read_err
* 100.0)) == 0) {
6308 dout(0) << __func__
<< ": inject random EIO" << dendl
;
6311 dout(10) << __func__
<< " " << cid
<< " " << oid
6312 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6313 << " = " << r
<< dendl
;
6314 logger
->tinc(l_bluestore_read_lat
, ceph_clock_now() - start
);
6318 // --------------------------------------------------------
6319 // intermediate data structures used while reading
6321 uint64_t logical_offset
;
6322 uint64_t blob_xoffset
; //region offset within the blob
6326 // used later in read process
6330 region_t(uint64_t offset
, uint64_t b_offs
, uint64_t len
)
6331 : logical_offset(offset
),
6332 blob_xoffset(b_offs
),
6334 region_t(const region_t
& from
)
6335 : logical_offset(from
.logical_offset
),
6336 blob_xoffset(from
.blob_xoffset
),
6337 length(from
.length
){}
6339 friend ostream
& operator<<(ostream
& out
, const region_t
& r
) {
6340 return out
<< "0x" << std::hex
<< r
.logical_offset
<< ":"
6341 << r
.blob_xoffset
<< "~" << r
.length
<< std::dec
;
6345 typedef list
<region_t
> regions2read_t
;
6346 typedef map
<BlueStore::BlobRef
, regions2read_t
> blobs2read_t
;
6348 int BlueStore::_do_read(
6359 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6360 << " size 0x" << o
->onode
.size
<< " (" << std::dec
6361 << o
->onode
.size
<< ")" << dendl
;
6364 if (offset
>= o
->onode
.size
) {
6368 // generally, don't buffer anything, unless the client explicitly requests
6370 bool buffered
= false;
6371 if (op_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
6372 dout(20) << __func__
<< " will do buffered read" << dendl
;
6374 } else if (cct
->_conf
->bluestore_default_buffered_read
&&
6375 (op_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
6376 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
6377 dout(20) << __func__
<< " defaulting to buffered read" << dendl
;
6381 if (offset
+ length
> o
->onode
.size
) {
6382 length
= o
->onode
.size
- offset
;
6385 utime_t start
= ceph_clock_now();
6386 o
->extent_map
.fault_range(db
, offset
, length
);
6387 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start
);
6390 ready_regions_t ready_regions
;
6392 // build blob-wise list to of stuff read (that isn't cached)
6393 blobs2read_t blobs2read
;
6394 unsigned left
= length
;
6395 uint64_t pos
= offset
;
6396 unsigned num_regions
= 0;
6397 auto lp
= o
->extent_map
.seek_lextent(offset
);
6398 while (left
> 0 && lp
!= o
->extent_map
.extent_map
.end()) {
6399 if (pos
< lp
->logical_offset
) {
6400 unsigned hole
= lp
->logical_offset
- pos
;
6404 dout(30) << __func__
<< " hole 0x" << std::hex
<< pos
<< "~" << hole
6405 << std::dec
<< dendl
;
6409 BlobRef bptr
= lp
->blob
;
6410 unsigned l_off
= pos
- lp
->logical_offset
;
6411 unsigned b_off
= l_off
+ lp
->blob_offset
;
6412 unsigned b_len
= std::min(left
, lp
->length
- l_off
);
6414 ready_regions_t cache_res
;
6415 interval_set
<uint32_t> cache_interval
;
6416 bptr
->shared_blob
->bc
.read(
6417 bptr
->shared_blob
->get_cache(), b_off
, b_len
, cache_res
, cache_interval
);
6418 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6419 << " need 0x" << b_off
<< "~" << b_len
6420 << " cache has 0x" << cache_interval
6421 << std::dec
<< dendl
;
6423 auto pc
= cache_res
.begin();
6426 if (pc
!= cache_res
.end() &&
6427 pc
->first
== b_off
) {
6428 l
= pc
->second
.length();
6429 ready_regions
[pos
].claim(pc
->second
);
6430 dout(30) << __func__
<< " use cache 0x" << std::hex
<< pos
<< ": 0x"
6431 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6435 if (pc
!= cache_res
.end()) {
6436 assert(pc
->first
> b_off
);
6437 l
= pc
->first
- b_off
;
6439 dout(30) << __func__
<< " will read 0x" << std::hex
<< pos
<< ": 0x"
6440 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6441 blobs2read
[bptr
].emplace_back(region_t(pos
, b_off
, l
));
6452 // read raw blob data. use aio if we have >1 blobs to read.
6453 start
= ceph_clock_now(); // for the sake of simplicity
6454 // measure the whole block below.
6455 // The error isn't that much...
6456 vector
<bufferlist
> compressed_blob_bls
;
6457 IOContext
ioc(cct
, NULL
);
6458 for (auto& p
: blobs2read
) {
6459 BlobRef bptr
= p
.first
;
6460 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6461 << " need " << p
.second
<< std::dec
<< dendl
;
6462 if (bptr
->get_blob().is_compressed()) {
6463 // read the whole thing
6464 if (compressed_blob_bls
.empty()) {
6465 // ensure we avoid any reallocation on subsequent blobs
6466 compressed_blob_bls
.reserve(blobs2read
.size());
6468 compressed_blob_bls
.push_back(bufferlist());
6469 bufferlist
& bl
= compressed_blob_bls
.back();
6470 r
= bptr
->get_blob().map(
6471 0, bptr
->get_blob().get_ondisk_length(),
6472 [&](uint64_t offset
, uint64_t length
) {
6474 // use aio if there are more regions to read than those in this blob
6475 if (num_regions
> p
.second
.size()) {
6476 r
= bdev
->aio_read(offset
, length
, &bl
, &ioc
);
6478 r
= bdev
->read(offset
, length
, &bl
, &ioc
, false);
6487 for (auto& reg
: p
.second
) {
6488 // determine how much of the blob to read
6489 uint64_t chunk_size
= bptr
->get_blob().get_chunk_size(block_size
);
6490 reg
.r_off
= reg
.blob_xoffset
;
6491 uint64_t r_len
= reg
.length
;
6492 reg
.front
= reg
.r_off
% chunk_size
;
6494 reg
.r_off
-= reg
.front
;
6497 unsigned tail
= r_len
% chunk_size
;
6499 r_len
+= chunk_size
- tail
;
6501 dout(20) << __func__
<< " region 0x" << std::hex
6502 << reg
.logical_offset
6503 << ": 0x" << reg
.blob_xoffset
<< "~" << reg
.length
6504 << " reading 0x" << reg
.r_off
<< "~" << r_len
<< std::dec
6508 r
= bptr
->get_blob().map(
6510 [&](uint64_t offset
, uint64_t length
) {
6512 // use aio if there is more than one region to read
6513 if (num_regions
> 1) {
6514 r
= bdev
->aio_read(offset
, length
, ®
.bl
, &ioc
);
6516 r
= bdev
->read(offset
, length
, ®
.bl
, &ioc
, false);
6523 assert(reg
.bl
.length() == r_len
);
6527 if (ioc
.has_pending_aios()) {
6528 bdev
->aio_submit(&ioc
);
6529 dout(20) << __func__
<< " waiting for aio" << dendl
;
6532 logger
->tinc(l_bluestore_read_wait_aio_lat
, ceph_clock_now() - start
);
6534 // enumerate and decompress desired blobs
6535 auto p
= compressed_blob_bls
.begin();
6536 blobs2read_t::iterator b2r_it
= blobs2read
.begin();
6537 while (b2r_it
!= blobs2read
.end()) {
6538 BlobRef bptr
= b2r_it
->first
;
6539 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6540 << " need 0x" << b2r_it
->second
<< std::dec
<< dendl
;
6541 if (bptr
->get_blob().is_compressed()) {
6542 assert(p
!= compressed_blob_bls
.end());
6543 bufferlist
& compressed_bl
= *p
++;
6544 if (_verify_csum(o
, &bptr
->get_blob(), 0, compressed_bl
,
6545 b2r_it
->second
.front().logical_offset
) < 0) {
6549 r
= _decompress(compressed_bl
, &raw_bl
);
6553 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(), 0,
6556 for (auto& i
: b2r_it
->second
) {
6557 ready_regions
[i
.logical_offset
].substr_of(
6558 raw_bl
, i
.blob_xoffset
, i
.length
);
6561 for (auto& reg
: b2r_it
->second
) {
6562 if (_verify_csum(o
, &bptr
->get_blob(), reg
.r_off
, reg
.bl
,
6563 reg
.logical_offset
) < 0) {
6567 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(),
6571 // prune and keep result
6572 ready_regions
[reg
.logical_offset
].substr_of(
6573 reg
.bl
, reg
.front
, reg
.length
);
6579 // generate a resulting buffer
6580 auto pr
= ready_regions
.begin();
6581 auto pr_end
= ready_regions
.end();
6583 while (pos
< length
) {
6584 if (pr
!= pr_end
&& pr
->first
== pos
+ offset
) {
6585 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6586 << ": data from 0x" << pr
->first
<< "~" << pr
->second
.length()
6587 << std::dec
<< dendl
;
6588 pos
+= pr
->second
.length();
6589 bl
.claim_append(pr
->second
);
6592 uint64_t l
= length
- pos
;
6594 assert(pr
->first
> pos
+ offset
);
6595 l
= pr
->first
- (pos
+ offset
);
6597 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6598 << ": zeros for 0x" << (pos
+ offset
) << "~" << l
6599 << std::dec
<< dendl
;
6604 assert(bl
.length() == length
);
6605 assert(pos
== length
);
6606 assert(pr
== pr_end
);
6611 int BlueStore::_verify_csum(OnodeRef
& o
,
6612 const bluestore_blob_t
* blob
, uint64_t blob_xoffset
,
6613 const bufferlist
& bl
,
6614 uint64_t logical_offset
) const
6618 utime_t start
= ceph_clock_now();
6619 int r
= blob
->verify_csum(blob_xoffset
, bl
, &bad
, &bad_csum
);
6625 blob
->get_csum_chunk_size(),
6626 [&](uint64_t offset
, uint64_t length
) {
6627 pex
.emplace_back(bluestore_pextent_t(offset
, length
));
6630 derr
<< __func__
<< " bad "
6631 << Checksummer::get_csum_type_string(blob
->csum_type
)
6632 << "/0x" << std::hex
<< blob
->get_csum_chunk_size()
6633 << " checksum at blob offset 0x" << bad
6634 << ", got 0x" << bad_csum
<< ", expected 0x"
6635 << blob
->get_csum_item(bad
/ blob
->get_csum_chunk_size()) << std::dec
6636 << ", device location " << pex
6637 << ", logical extent 0x" << std::hex
6638 << (logical_offset
+ bad
- blob_xoffset
) << "~"
6639 << blob
->get_csum_chunk_size() << std::dec
6640 << ", object " << o
->oid
6643 derr
<< __func__
<< " failed with exit code: " << cpp_strerror(r
) << dendl
;
6646 logger
->tinc(l_bluestore_csum_lat
, ceph_clock_now() - start
);
6650 int BlueStore::_decompress(bufferlist
& source
, bufferlist
* result
)
6653 utime_t start
= ceph_clock_now();
6654 bufferlist::iterator i
= source
.begin();
6655 bluestore_compression_header_t chdr
;
6657 int alg
= int(chdr
.type
);
6658 CompressorRef cp
= compressor
;
6659 if (!cp
|| (int)cp
->get_type() != alg
) {
6660 cp
= Compressor::create(cct
, alg
);
6664 // if compressor isn't available - error, because cannot return
6665 // decompressed data?
6666 derr
<< __func__
<< " can't load decompressor " << alg
<< dendl
;
6669 r
= cp
->decompress(i
, chdr
.length
, *result
);
6671 derr
<< __func__
<< " decompression failed with exit code " << r
<< dendl
;
6675 logger
->tinc(l_bluestore_decompress_lat
, ceph_clock_now() - start
);
6679 // this stores fiemap into interval_set, other variations
6680 // use it internally
6681 int BlueStore::_fiemap(
6682 CollectionHandle
&c_
,
6683 const ghobject_t
& oid
,
6686 interval_set
<uint64_t>& destset
)
6688 Collection
*c
= static_cast<Collection
*>(c_
.get());
6692 RWLock::RLocker
l(c
->lock
);
6694 OnodeRef o
= c
->get_onode(oid
, false);
6695 if (!o
|| !o
->exists
) {
6700 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6701 << " size 0x" << o
->onode
.size
<< std::dec
<< dendl
;
6703 boost::intrusive::set
<Extent
>::iterator ep
, eend
;
6704 if (offset
>= o
->onode
.size
)
6707 if (offset
+ length
> o
->onode
.size
) {
6708 length
= o
->onode
.size
- offset
;
6711 o
->extent_map
.fault_range(db
, offset
, length
);
6712 eend
= o
->extent_map
.extent_map
.end();
6713 ep
= o
->extent_map
.seek_lextent(offset
);
6714 while (length
> 0) {
6715 dout(20) << __func__
<< " offset " << offset
<< dendl
;
6716 if (ep
!= eend
&& ep
->logical_offset
+ ep
->length
<= offset
) {
6721 uint64_t x_len
= length
;
6722 if (ep
!= eend
&& ep
->logical_offset
<= offset
) {
6723 uint64_t x_off
= offset
- ep
->logical_offset
;
6724 x_len
= MIN(x_len
, ep
->length
- x_off
);
6725 dout(30) << __func__
<< " lextent 0x" << std::hex
<< offset
<< "~"
6726 << x_len
<< std::dec
<< " blob " << ep
->blob
<< dendl
;
6727 destset
.insert(offset
, x_len
);
6730 if (x_off
+ x_len
== ep
->length
)
6735 ep
->logical_offset
> offset
&&
6736 ep
->logical_offset
- offset
< x_len
) {
6737 x_len
= ep
->logical_offset
- offset
;
6745 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6746 << " size = 0x(" << destset
<< ")" << std::dec
<< dendl
;
6750 int BlueStore::fiemap(
6752 const ghobject_t
& oid
,
6757 CollectionHandle c
= _get_collection(cid
);
6760 return fiemap(c
, oid
, offset
, len
, bl
);
6763 int BlueStore::fiemap(
6764 CollectionHandle
&c_
,
6765 const ghobject_t
& oid
,
6770 interval_set
<uint64_t> m
;
6771 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6778 int BlueStore::fiemap(
6780 const ghobject_t
& oid
,
6783 map
<uint64_t, uint64_t>& destmap
)
6785 CollectionHandle c
= _get_collection(cid
);
6788 return fiemap(c
, oid
, offset
, len
, destmap
);
6791 int BlueStore::fiemap(
6792 CollectionHandle
&c_
,
6793 const ghobject_t
& oid
,
6796 map
<uint64_t, uint64_t>& destmap
)
6798 interval_set
<uint64_t> m
;
6799 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6801 m
.move_into(destmap
);
6806 int BlueStore::getattr(
6808 const ghobject_t
& oid
,
6812 CollectionHandle c
= _get_collection(cid
);
6815 return getattr(c
, oid
, name
, value
);
6818 int BlueStore::getattr(
6819 CollectionHandle
&c_
,
6820 const ghobject_t
& oid
,
6824 Collection
*c
= static_cast<Collection
*>(c_
.get());
6825 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
<< dendl
;
6831 RWLock::RLocker
l(c
->lock
);
6832 mempool::bluestore_cache_other::string
k(name
);
6834 OnodeRef o
= c
->get_onode(oid
, false);
6835 if (!o
|| !o
->exists
) {
6840 if (!o
->onode
.attrs
.count(k
)) {
6844 value
= o
->onode
.attrs
[k
];
6848 if (r
== 0 && _debug_mdata_eio(oid
)) {
6850 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6852 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
6853 << " = " << r
<< dendl
;
6858 int BlueStore::getattrs(
6860 const ghobject_t
& oid
,
6861 map
<string
,bufferptr
>& aset
)
6863 CollectionHandle c
= _get_collection(cid
);
6866 return getattrs(c
, oid
, aset
);
6869 int BlueStore::getattrs(
6870 CollectionHandle
&c_
,
6871 const ghobject_t
& oid
,
6872 map
<string
,bufferptr
>& aset
)
6874 Collection
*c
= static_cast<Collection
*>(c_
.get());
6875 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
6881 RWLock::RLocker
l(c
->lock
);
6883 OnodeRef o
= c
->get_onode(oid
, false);
6884 if (!o
|| !o
->exists
) {
6888 for (auto& i
: o
->onode
.attrs
) {
6889 aset
.emplace(i
.first
.c_str(), i
.second
);
6895 if (r
== 0 && _debug_mdata_eio(oid
)) {
6897 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6899 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
6900 << " = " << r
<< dendl
;
6904 int BlueStore::list_collections(vector
<coll_t
>& ls
)
6906 RWLock::RLocker
l(coll_lock
);
6907 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
= coll_map
.begin();
6908 p
!= coll_map
.end();
6910 ls
.push_back(p
->first
);
6914 bool BlueStore::collection_exists(const coll_t
& c
)
6916 RWLock::RLocker
l(coll_lock
);
6917 return coll_map
.count(c
);
6920 int BlueStore::collection_empty(const coll_t
& cid
, bool *empty
)
6922 dout(15) << __func__
<< " " << cid
<< dendl
;
6923 vector
<ghobject_t
> ls
;
6925 int r
= collection_list(cid
, ghobject_t(), ghobject_t::get_max(), 1,
6928 derr
<< __func__
<< " collection_list returned: " << cpp_strerror(r
)
6932 *empty
= ls
.empty();
6933 dout(10) << __func__
<< " " << cid
<< " = " << (int)(*empty
) << dendl
;
6937 int BlueStore::collection_bits(const coll_t
& cid
)
6939 dout(15) << __func__
<< " " << cid
<< dendl
;
6940 CollectionRef c
= _get_collection(cid
);
6943 RWLock::RLocker
l(c
->lock
);
6944 dout(10) << __func__
<< " " << cid
<< " = " << c
->cnode
.bits
<< dendl
;
6945 return c
->cnode
.bits
;
6948 int BlueStore::collection_list(
6949 const coll_t
& cid
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
6950 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
6952 CollectionHandle c
= _get_collection(cid
);
6955 return collection_list(c
, start
, end
, max
, ls
, pnext
);
6958 int BlueStore::collection_list(
6959 CollectionHandle
&c_
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
6960 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
6962 Collection
*c
= static_cast<Collection
*>(c_
.get());
6963 dout(15) << __func__
<< " " << c
->cid
6964 << " start " << start
<< " end " << end
<< " max " << max
<< dendl
;
6967 RWLock::RLocker
l(c
->lock
);
6968 r
= _collection_list(c
, start
, end
, max
, ls
, pnext
);
6971 dout(10) << __func__
<< " " << c
->cid
6972 << " start " << start
<< " end " << end
<< " max " << max
6973 << " = " << r
<< ", ls.size() = " << ls
->size()
6974 << ", next = " << (pnext
? *pnext
: ghobject_t()) << dendl
;
6978 int BlueStore::_collection_list(
6979 Collection
*c
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
6980 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
6987 ghobject_t static_next
;
6988 KeyValueDB::Iterator it
;
6989 string temp_start_key
, temp_end_key
;
6990 string start_key
, end_key
;
6991 bool set_next
= false;
6996 pnext
= &static_next
;
6998 if (start
== ghobject_t::get_max() ||
6999 start
.hobj
.is_max()) {
7002 get_coll_key_range(c
->cid
, c
->cnode
.bits
, &temp_start_key
, &temp_end_key
,
7003 &start_key
, &end_key
);
7004 dout(20) << __func__
7005 << " range " << pretty_binary_string(temp_start_key
)
7006 << " to " << pretty_binary_string(temp_end_key
)
7007 << " and " << pretty_binary_string(start_key
)
7008 << " to " << pretty_binary_string(end_key
)
7009 << " start " << start
<< dendl
;
7010 it
= db
->get_iterator(PREFIX_OBJ
);
7011 if (start
== ghobject_t() ||
7012 start
.hobj
== hobject_t() ||
7013 start
== c
->cid
.get_min_hobj()) {
7014 it
->upper_bound(temp_start_key
);
7018 get_object_key(cct
, start
, &k
);
7019 if (start
.hobj
.is_temp()) {
7021 assert(k
>= temp_start_key
&& k
< temp_end_key
);
7024 assert(k
>= start_key
&& k
< end_key
);
7026 dout(20) << " start from " << pretty_binary_string(k
)
7027 << " temp=" << (int)temp
<< dendl
;
7030 if (end
.hobj
.is_max()) {
7031 pend
= temp
? temp_end_key
: end_key
;
7033 get_object_key(cct
, end
, &end_key
);
7034 if (end
.hobj
.is_temp()) {
7040 pend
= temp
? temp_end_key
: end_key
;
7043 dout(20) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7045 if (!it
->valid() || it
->key() >= pend
) {
7047 dout(20) << __func__
<< " iterator not valid (end of db?)" << dendl
;
7049 dout(20) << __func__
<< " key " << pretty_binary_string(it
->key())
7050 << " >= " << end
<< dendl
;
7052 if (end
.hobj
.is_temp()) {
7055 dout(30) << __func__
<< " switch to non-temp namespace" << dendl
;
7057 it
->upper_bound(start_key
);
7059 dout(30) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7064 dout(30) << __func__
<< " key " << pretty_binary_string(it
->key()) << dendl
;
7065 if (is_extent_shard_key(it
->key())) {
7070 int r
= get_key_object(it
->key(), &oid
);
7072 dout(20) << __func__
<< " oid " << oid
<< " end " << end
<< dendl
;
7073 if (ls
->size() >= (unsigned)max
) {
7074 dout(20) << __func__
<< " reached max " << max
<< dendl
;
7084 *pnext
= ghobject_t::get_max();
7090 int BlueStore::omap_get(
7091 const coll_t
& cid
, ///< [in] Collection containing oid
7092 const ghobject_t
&oid
, ///< [in] Object containing omap
7093 bufferlist
*header
, ///< [out] omap header
7094 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7097 CollectionHandle c
= _get_collection(cid
);
7100 return omap_get(c
, oid
, header
, out
);
7103 int BlueStore::omap_get(
7104 CollectionHandle
&c_
, ///< [in] Collection containing oid
7105 const ghobject_t
&oid
, ///< [in] Object containing omap
7106 bufferlist
*header
, ///< [out] omap header
7107 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7110 Collection
*c
= static_cast<Collection
*>(c_
.get());
7111 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7114 RWLock::RLocker
l(c
->lock
);
7116 OnodeRef o
= c
->get_onode(oid
, false);
7117 if (!o
|| !o
->exists
) {
7121 if (!o
->onode
.has_omap())
7125 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7127 get_omap_header(o
->onode
.nid
, &head
);
7128 get_omap_tail(o
->onode
.nid
, &tail
);
7129 it
->lower_bound(head
);
7130 while (it
->valid()) {
7131 if (it
->key() == head
) {
7132 dout(30) << __func__
<< " got header" << dendl
;
7133 *header
= it
->value();
7134 } else if (it
->key() >= tail
) {
7135 dout(30) << __func__
<< " reached tail" << dendl
;
7139 decode_omap_key(it
->key(), &user_key
);
7140 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7141 << " -> " << user_key
<< dendl
;
7142 (*out
)[user_key
] = it
->value();
7148 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7153 int BlueStore::omap_get_header(
7154 const coll_t
& cid
, ///< [in] Collection containing oid
7155 const ghobject_t
&oid
, ///< [in] Object containing omap
7156 bufferlist
*header
, ///< [out] omap header
7157 bool allow_eio
///< [in] don't assert on eio
7160 CollectionHandle c
= _get_collection(cid
);
7163 return omap_get_header(c
, oid
, header
, allow_eio
);
7166 int BlueStore::omap_get_header(
7167 CollectionHandle
&c_
, ///< [in] Collection containing oid
7168 const ghobject_t
&oid
, ///< [in] Object containing omap
7169 bufferlist
*header
, ///< [out] omap header
7170 bool allow_eio
///< [in] don't assert on eio
7173 Collection
*c
= static_cast<Collection
*>(c_
.get());
7174 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7177 RWLock::RLocker
l(c
->lock
);
7179 OnodeRef o
= c
->get_onode(oid
, false);
7180 if (!o
|| !o
->exists
) {
7184 if (!o
->onode
.has_omap())
7189 get_omap_header(o
->onode
.nid
, &head
);
7190 if (db
->get(PREFIX_OMAP
, head
, header
) >= 0) {
7191 dout(30) << __func__
<< " got header" << dendl
;
7193 dout(30) << __func__
<< " no header" << dendl
;
7197 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7202 int BlueStore::omap_get_keys(
7203 const coll_t
& cid
, ///< [in] Collection containing oid
7204 const ghobject_t
&oid
, ///< [in] Object containing omap
7205 set
<string
> *keys
///< [out] Keys defined on oid
7208 CollectionHandle c
= _get_collection(cid
);
7211 return omap_get_keys(c
, oid
, keys
);
7214 int BlueStore::omap_get_keys(
7215 CollectionHandle
&c_
, ///< [in] Collection containing oid
7216 const ghobject_t
&oid
, ///< [in] Object containing omap
7217 set
<string
> *keys
///< [out] Keys defined on oid
7220 Collection
*c
= static_cast<Collection
*>(c_
.get());
7221 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7224 RWLock::RLocker
l(c
->lock
);
7226 OnodeRef o
= c
->get_onode(oid
, false);
7227 if (!o
|| !o
->exists
) {
7231 if (!o
->onode
.has_omap())
7235 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7237 get_omap_key(o
->onode
.nid
, string(), &head
);
7238 get_omap_tail(o
->onode
.nid
, &tail
);
7239 it
->lower_bound(head
);
7240 while (it
->valid()) {
7241 if (it
->key() >= tail
) {
7242 dout(30) << __func__
<< " reached tail" << dendl
;
7246 decode_omap_key(it
->key(), &user_key
);
7247 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7248 << " -> " << user_key
<< dendl
;
7249 keys
->insert(user_key
);
7254 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7259 int BlueStore::omap_get_values(
7260 const coll_t
& cid
, ///< [in] Collection containing oid
7261 const ghobject_t
&oid
, ///< [in] Object containing omap
7262 const set
<string
> &keys
, ///< [in] Keys to get
7263 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7266 CollectionHandle c
= _get_collection(cid
);
7269 return omap_get_values(c
, oid
, keys
, out
);
7272 int BlueStore::omap_get_values(
7273 CollectionHandle
&c_
, ///< [in] Collection containing oid
7274 const ghobject_t
&oid
, ///< [in] Object containing omap
7275 const set
<string
> &keys
, ///< [in] Keys to get
7276 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7279 Collection
*c
= static_cast<Collection
*>(c_
.get());
7280 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7283 RWLock::RLocker
l(c
->lock
);
7286 OnodeRef o
= c
->get_onode(oid
, false);
7287 if (!o
|| !o
->exists
) {
7291 if (!o
->onode
.has_omap())
7294 _key_encode_u64(o
->onode
.nid
, &final_key
);
7295 final_key
.push_back('.');
7296 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7297 final_key
.resize(9); // keep prefix
7300 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7301 dout(30) << __func__
<< " got " << pretty_binary_string(final_key
)
7302 << " -> " << *p
<< dendl
;
7303 out
->insert(make_pair(*p
, val
));
7307 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7312 int BlueStore::omap_check_keys(
7313 const coll_t
& cid
, ///< [in] Collection containing oid
7314 const ghobject_t
&oid
, ///< [in] Object containing omap
7315 const set
<string
> &keys
, ///< [in] Keys to check
7316 set
<string
> *out
///< [out] Subset of keys defined on oid
7319 CollectionHandle c
= _get_collection(cid
);
7322 return omap_check_keys(c
, oid
, keys
, out
);
7325 int BlueStore::omap_check_keys(
7326 CollectionHandle
&c_
, ///< [in] Collection containing oid
7327 const ghobject_t
&oid
, ///< [in] Object containing omap
7328 const set
<string
> &keys
, ///< [in] Keys to check
7329 set
<string
> *out
///< [out] Subset of keys defined on oid
7332 Collection
*c
= static_cast<Collection
*>(c_
.get());
7333 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7336 RWLock::RLocker
l(c
->lock
);
7339 OnodeRef o
= c
->get_onode(oid
, false);
7340 if (!o
|| !o
->exists
) {
7344 if (!o
->onode
.has_omap())
7347 _key_encode_u64(o
->onode
.nid
, &final_key
);
7348 final_key
.push_back('.');
7349 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7350 final_key
.resize(9); // keep prefix
7353 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7354 dout(30) << __func__
<< " have " << pretty_binary_string(final_key
)
7355 << " -> " << *p
<< dendl
;
7358 dout(30) << __func__
<< " miss " << pretty_binary_string(final_key
)
7359 << " -> " << *p
<< dendl
;
7363 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7368 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7369 const coll_t
& cid
, ///< [in] collection
7370 const ghobject_t
&oid
///< [in] object
7373 CollectionHandle c
= _get_collection(cid
);
7375 dout(10) << __func__
<< " " << cid
<< "doesn't exist" <<dendl
;
7376 return ObjectMap::ObjectMapIterator();
7378 return get_omap_iterator(c
, oid
);
7381 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7382 CollectionHandle
&c_
, ///< [in] collection
7383 const ghobject_t
&oid
///< [in] object
7386 Collection
*c
= static_cast<Collection
*>(c_
.get());
7387 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
7389 return ObjectMap::ObjectMapIterator();
7391 RWLock::RLocker
l(c
->lock
);
7392 OnodeRef o
= c
->get_onode(oid
, false);
7393 if (!o
|| !o
->exists
) {
7394 dout(10) << __func__
<< " " << oid
<< "doesn't exist" <<dendl
;
7395 return ObjectMap::ObjectMapIterator();
7398 dout(10) << __func__
<< " has_omap = " << (int)o
->onode
.has_omap() <<dendl
;
7399 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7400 return ObjectMap::ObjectMapIterator(new OmapIteratorImpl(c
, o
, it
));
7403 // -----------------
7406 void BlueStore::_prepare_ondisk_format_super(KeyValueDB::Transaction
& t
)
7408 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7409 << " min_compat_ondisk_format " << min_compat_ondisk_format
7411 assert(ondisk_format
== latest_ondisk_format
);
7414 ::encode(ondisk_format
, bl
);
7415 t
->set(PREFIX_SUPER
, "ondisk_format", bl
);
7419 ::encode(min_compat_ondisk_format
, bl
);
7420 t
->set(PREFIX_SUPER
, "min_compat_ondisk_format", bl
);
7424 int BlueStore::_open_super_meta()
7430 db
->get(PREFIX_SUPER
, "nid_max", &bl
);
7431 bufferlist::iterator p
= bl
.begin();
7436 } catch (buffer::error
& e
) {
7437 derr
<< __func__
<< " unable to read nid_max" << dendl
;
7440 dout(10) << __func__
<< " old nid_max " << nid_max
<< dendl
;
7441 nid_last
= nid_max
.load();
7448 db
->get(PREFIX_SUPER
, "blobid_max", &bl
);
7449 bufferlist::iterator p
= bl
.begin();
7454 } catch (buffer::error
& e
) {
7455 derr
<< __func__
<< " unable to read blobid_max" << dendl
;
7458 dout(10) << __func__
<< " old blobid_max " << blobid_max
<< dendl
;
7459 blobid_last
= blobid_max
.load();
7465 db
->get(PREFIX_SUPER
, "freelist_type", &bl
);
7467 freelist_type
= std::string(bl
.c_str(), bl
.length());
7468 dout(10) << __func__
<< " freelist_type " << freelist_type
<< dendl
;
7470 assert("Not Support extent freelist manager" == 0);
7475 if (cct
->_conf
->bluestore_bluefs
) {
7476 bluefs_extents
.clear();
7478 db
->get(PREFIX_SUPER
, "bluefs_extents", &bl
);
7479 bufferlist::iterator p
= bl
.begin();
7481 ::decode(bluefs_extents
, p
);
7483 catch (buffer::error
& e
) {
7484 derr
<< __func__
<< " unable to read bluefs_extents" << dendl
;
7487 dout(10) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
7488 << std::dec
<< dendl
;
7492 int32_t compat_ondisk_format
= 0;
7495 int r
= db
->get(PREFIX_SUPER
, "ondisk_format", &bl
);
7497 // base case: kraken bluestore is v1 and readable by v1
7498 dout(20) << __func__
<< " missing ondisk_format; assuming kraken"
7501 compat_ondisk_format
= 1;
7503 auto p
= bl
.begin();
7505 ::decode(ondisk_format
, p
);
7506 } catch (buffer::error
& e
) {
7507 derr
<< __func__
<< " unable to read ondisk_format" << dendl
;
7512 r
= db
->get(PREFIX_SUPER
, "min_compat_ondisk_format", &bl
);
7514 auto p
= bl
.begin();
7516 ::decode(compat_ondisk_format
, p
);
7517 } catch (buffer::error
& e
) {
7518 derr
<< __func__
<< " unable to read compat_ondisk_format" << dendl
;
7523 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7524 << " compat_ondisk_format " << compat_ondisk_format
7528 if (latest_ondisk_format
< compat_ondisk_format
) {
7529 derr
<< __func__
<< " compat_ondisk_format is "
7530 << compat_ondisk_format
<< " but we only understand version "
7531 << latest_ondisk_format
<< dendl
;
7534 if (ondisk_format
< latest_ondisk_format
) {
7535 int r
= _upgrade_super();
7543 db
->get(PREFIX_SUPER
, "min_alloc_size", &bl
);
7544 auto p
= bl
.begin();
7548 min_alloc_size
= val
;
7549 min_alloc_size_order
= ctz(val
);
7550 assert(min_alloc_size
== 1u << min_alloc_size_order
);
7551 } catch (buffer::error
& e
) {
7552 derr
<< __func__
<< " unable to read min_alloc_size" << dendl
;
7555 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
7556 << std::dec
<< dendl
;
7560 _set_throttle_params();
7569 int BlueStore::_upgrade_super()
7571 dout(1) << __func__
<< " from " << ondisk_format
<< ", latest "
7572 << latest_ondisk_format
<< dendl
;
7573 assert(ondisk_format
> 0);
7574 assert(ondisk_format
< latest_ondisk_format
);
7576 if (ondisk_format
== 1) {
7578 // - super: added ondisk_format
7579 // - super: added min_readable_ondisk_format
7580 // - super: added min_compat_ondisk_format
7581 // - super: added min_alloc_size
7582 // - super: removed min_min_alloc_size
7583 KeyValueDB::Transaction t
= db
->get_transaction();
7586 db
->get(PREFIX_SUPER
, "min_min_alloc_size", &bl
);
7587 auto p
= bl
.begin();
7591 min_alloc_size
= val
;
7592 } catch (buffer::error
& e
) {
7593 derr
<< __func__
<< " failed to read min_min_alloc_size" << dendl
;
7596 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
7597 t
->rmkey(PREFIX_SUPER
, "min_min_alloc_size");
7600 _prepare_ondisk_format_super(t
);
7601 int r
= db
->submit_transaction_sync(t
);
7606 dout(1) << __func__
<< " done" << dendl
;
7610 void BlueStore::_assign_nid(TransContext
*txc
, OnodeRef o
)
7616 uint64_t nid
= ++nid_last
;
7617 dout(20) << __func__
<< " " << nid
<< dendl
;
7619 txc
->last_nid
= nid
;
7623 uint64_t BlueStore::_assign_blobid(TransContext
*txc
)
7625 uint64_t bid
= ++blobid_last
;
7626 dout(20) << __func__
<< " " << bid
<< dendl
;
7627 txc
->last_blobid
= bid
;
7631 void BlueStore::get_db_statistics(Formatter
*f
)
7633 db
->get_statistics(f
);
7636 BlueStore::TransContext
*BlueStore::_txc_create(OpSequencer
*osr
)
7638 TransContext
*txc
= new TransContext(cct
, osr
);
7639 txc
->t
= db
->get_transaction();
7640 osr
->queue_new(txc
);
7641 dout(20) << __func__
<< " osr " << osr
<< " = " << txc
7642 << " seq " << txc
->seq
<< dendl
;
7646 void BlueStore::_txc_calc_cost(TransContext
*txc
)
7648 // this is about the simplest model for transaction cost you can
7649 // imagine. there is some fixed overhead cost by saying there is a
7650 // minimum of one "io". and then we have some cost per "io" that is
7651 // a configurable (with different hdd and ssd defaults), and add
7652 // that to the bytes value.
7653 int ios
= 1; // one "io" for the kv commit
7654 for (auto& p
: txc
->ioc
.pending_aios
) {
7655 ios
+= p
.iov
.size();
7657 auto cost
= throttle_cost_per_io
.load();
7658 txc
->cost
= ios
* cost
+ txc
->bytes
;
7659 dout(10) << __func__
<< " " << txc
<< " cost " << txc
->cost
<< " ("
7660 << ios
<< " ios * " << cost
<< " + " << txc
->bytes
7661 << " bytes)" << dendl
;
7664 void BlueStore::_txc_update_store_statfs(TransContext
*txc
)
7666 if (txc
->statfs_delta
.is_empty())
7669 logger
->inc(l_bluestore_allocated
, txc
->statfs_delta
.allocated());
7670 logger
->inc(l_bluestore_stored
, txc
->statfs_delta
.stored());
7671 logger
->inc(l_bluestore_compressed
, txc
->statfs_delta
.compressed());
7672 logger
->inc(l_bluestore_compressed_allocated
, txc
->statfs_delta
.compressed_allocated());
7673 logger
->inc(l_bluestore_compressed_original
, txc
->statfs_delta
.compressed_original());
7676 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
7677 vstatfs
+= txc
->statfs_delta
;
7681 txc
->statfs_delta
.encode(bl
);
7683 txc
->t
->merge(PREFIX_STAT
, "bluestore_statfs", bl
);
7684 txc
->statfs_delta
.reset();
7687 void BlueStore::_txc_state_proc(TransContext
*txc
)
7690 dout(10) << __func__
<< " txc " << txc
7691 << " " << txc
->get_state_name() << dendl
;
7692 switch (txc
->state
) {
7693 case TransContext::STATE_PREPARE
:
7694 txc
->log_state_latency(logger
, l_bluestore_state_prepare_lat
);
7695 if (txc
->ioc
.has_pending_aios()) {
7696 txc
->state
= TransContext::STATE_AIO_WAIT
;
7697 txc
->had_ios
= true;
7698 _txc_aio_submit(txc
);
7703 case TransContext::STATE_AIO_WAIT
:
7704 txc
->log_state_latency(logger
, l_bluestore_state_aio_wait_lat
);
7705 _txc_finish_io(txc
); // may trigger blocked txc's too
7708 case TransContext::STATE_IO_DONE
:
7709 //assert(txc->osr->qlock.is_locked()); // see _txc_finish_io
7711 ++txc
->osr
->txc_with_unstable_io
;
7713 txc
->log_state_latency(logger
, l_bluestore_state_io_done_lat
);
7714 txc
->state
= TransContext::STATE_KV_QUEUED
;
7715 if (cct
->_conf
->bluestore_sync_submit_transaction
) {
7716 if (txc
->last_nid
>= nid_max
||
7717 txc
->last_blobid
>= blobid_max
) {
7718 dout(20) << __func__
7719 << " last_{nid,blobid} exceeds max, submit via kv thread"
7721 } else if (txc
->osr
->kv_committing_serially
) {
7722 dout(20) << __func__
<< " prior txc submitted via kv thread, us too"
7724 // note: this is starvation-prone. once we have a txc in a busy
7725 // sequencer that is committing serially it is possible to keep
7726 // submitting new transactions fast enough that we get stuck doing
7727 // so. the alternative is to block here... fixme?
7728 } else if (txc
->osr
->txc_with_unstable_io
) {
7729 dout(20) << __func__
<< " prior txc(s) with unstable ios "
7730 << txc
->osr
->txc_with_unstable_io
.load() << dendl
;
7731 } else if (cct
->_conf
->bluestore_debug_randomize_serial_transaction
&&
7732 rand() % cct
->_conf
->bluestore_debug_randomize_serial_transaction
7734 dout(20) << __func__
<< " DEBUG randomly forcing submit via kv thread"
7737 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
7738 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
7740 _txc_applied_kv(txc
);
7744 std::lock_guard
<std::mutex
> l(kv_lock
);
7745 kv_queue
.push_back(txc
);
7746 kv_cond
.notify_one();
7747 if (txc
->state
!= TransContext::STATE_KV_SUBMITTED
) {
7748 kv_queue_unsubmitted
.push_back(txc
);
7749 ++txc
->osr
->kv_committing_serially
;
7753 kv_throttle_costs
+= txc
->cost
;
7756 case TransContext::STATE_KV_SUBMITTED
:
7757 txc
->log_state_latency(logger
, l_bluestore_state_kv_committing_lat
);
7758 txc
->state
= TransContext::STATE_KV_DONE
;
7759 _txc_committed_kv(txc
);
7762 case TransContext::STATE_KV_DONE
:
7763 txc
->log_state_latency(logger
, l_bluestore_state_kv_done_lat
);
7764 if (txc
->deferred_txn
) {
7765 txc
->state
= TransContext::STATE_DEFERRED_QUEUED
;
7766 _deferred_queue(txc
);
7769 txc
->state
= TransContext::STATE_FINISHING
;
7772 case TransContext::STATE_DEFERRED_CLEANUP
:
7773 txc
->log_state_latency(logger
, l_bluestore_state_deferred_cleanup_lat
);
7774 txc
->state
= TransContext::STATE_FINISHING
;
7777 case TransContext::STATE_FINISHING
:
7778 txc
->log_state_latency(logger
, l_bluestore_state_finishing_lat
);
7783 derr
<< __func__
<< " unexpected txc " << txc
7784 << " state " << txc
->get_state_name() << dendl
;
7785 assert(0 == "unexpected txc state");
7791 void BlueStore::_txc_finish_io(TransContext
*txc
)
7793 dout(20) << __func__
<< " " << txc
<< dendl
;
7796 * we need to preserve the order of kv transactions,
7797 * even though aio will complete in any order.
7800 OpSequencer
*osr
= txc
->osr
.get();
7801 std::lock_guard
<std::mutex
> l(osr
->qlock
);
7802 txc
->state
= TransContext::STATE_IO_DONE
;
7804 // release aio contexts (including pinned buffers).
7805 txc
->ioc
.running_aios
.clear();
7807 OpSequencer::q_list_t::iterator p
= osr
->q
.iterator_to(*txc
);
7808 while (p
!= osr
->q
.begin()) {
7810 if (p
->state
< TransContext::STATE_IO_DONE
) {
7811 dout(20) << __func__
<< " " << txc
<< " blocked by " << &*p
<< " "
7812 << p
->get_state_name() << dendl
;
7815 if (p
->state
> TransContext::STATE_IO_DONE
) {
7821 _txc_state_proc(&*p
++);
7822 } while (p
!= osr
->q
.end() &&
7823 p
->state
== TransContext::STATE_IO_DONE
);
7825 if (osr
->kv_submitted_waiters
&&
7826 osr
->_is_all_kv_submitted()) {
7827 osr
->qcond
.notify_all();
7831 void BlueStore::_txc_write_nodes(TransContext
*txc
, KeyValueDB::Transaction t
)
7833 dout(20) << __func__
<< " txc " << txc
7834 << " onodes " << txc
->onodes
7835 << " shared_blobs " << txc
->shared_blobs
7839 for (auto o
: txc
->onodes
) {
7840 // finalize extent_map shards
7841 o
->extent_map
.update(t
, false);
7842 if (o
->extent_map
.needs_reshard()) {
7843 o
->extent_map
.reshard(db
, t
);
7844 o
->extent_map
.update(t
, true);
7845 if (o
->extent_map
.needs_reshard()) {
7846 dout(20) << __func__
<< " warning: still wants reshard, check options?"
7848 o
->extent_map
.clear_needs_reshard();
7850 logger
->inc(l_bluestore_onode_reshard
);
7855 denc(o
->onode
, bound
);
7856 o
->extent_map
.bound_encode_spanning_blobs(bound
);
7857 if (o
->onode
.extent_map_shards
.empty()) {
7858 denc(o
->extent_map
.inline_bl
, bound
);
7863 unsigned onode_part
, blob_part
, extent_part
;
7865 auto p
= bl
.get_contiguous_appender(bound
, true);
7867 onode_part
= p
.get_logical_offset();
7868 o
->extent_map
.encode_spanning_blobs(p
);
7869 blob_part
= p
.get_logical_offset() - onode_part
;
7870 if (o
->onode
.extent_map_shards
.empty()) {
7871 denc(o
->extent_map
.inline_bl
, p
);
7873 extent_part
= p
.get_logical_offset() - onode_part
- blob_part
;
7876 dout(20) << " onode " << o
->oid
<< " is " << bl
.length()
7877 << " (" << onode_part
<< " bytes onode + "
7878 << blob_part
<< " bytes spanning blobs + "
7879 << extent_part
<< " bytes inline extents)"
7881 t
->set(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size(), bl
);
7882 o
->flushing_count
++;
7885 // objects we modified but didn't affect the onode
7886 auto p
= txc
->modified_objects
.begin();
7887 while (p
!= txc
->modified_objects
.end()) {
7888 if (txc
->onodes
.count(*p
) == 0) {
7889 (*p
)->flushing_count
++;
7892 // remove dups with onodes list to avoid problems in _txc_finish
7893 p
= txc
->modified_objects
.erase(p
);
7897 // finalize shared_blobs
7898 for (auto sb
: txc
->shared_blobs
) {
7900 auto sbid
= sb
->get_sbid();
7901 get_shared_blob_key(sbid
, &key
);
7902 if (sb
->persistent
->empty()) {
7903 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
7904 << " is empty" << dendl
;
7905 t
->rmkey(PREFIX_SHARED_BLOB
, key
);
7908 ::encode(*(sb
->persistent
), bl
);
7909 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
7910 << " is " << bl
.length() << " " << *sb
<< dendl
;
7911 t
->set(PREFIX_SHARED_BLOB
, key
, bl
);
7916 void BlueStore::BSPerfTracker::update_from_perfcounters(
7917 PerfCounters
&logger
)
7919 os_commit_latency
.consume_next(
7921 l_bluestore_commit_lat
));
7922 os_apply_latency
.consume_next(
7924 l_bluestore_commit_lat
));
7927 void BlueStore::_txc_finalize_kv(TransContext
*txc
, KeyValueDB::Transaction t
)
7929 dout(20) << __func__
<< " txc " << txc
<< std::hex
7930 << " allocated 0x" << txc
->allocated
7931 << " released 0x" << txc
->released
7932 << std::dec
<< dendl
;
7934 // We have to handle the case where we allocate *and* deallocate the
7935 // same region in this transaction. The freelist doesn't like that.
7936 // (Actually, the only thing that cares is the BitmapFreelistManager
7937 // debug check. But that's important.)
7938 interval_set
<uint64_t> tmp_allocated
, tmp_released
;
7939 interval_set
<uint64_t> *pallocated
= &txc
->allocated
;
7940 interval_set
<uint64_t> *preleased
= &txc
->released
;
7941 if (!txc
->allocated
.empty() && !txc
->released
.empty()) {
7942 interval_set
<uint64_t> overlap
;
7943 overlap
.intersection_of(txc
->allocated
, txc
->released
);
7944 if (!overlap
.empty()) {
7945 tmp_allocated
= txc
->allocated
;
7946 tmp_allocated
.subtract(overlap
);
7947 tmp_released
= txc
->released
;
7948 tmp_released
.subtract(overlap
);
7949 dout(20) << __func__
<< " overlap 0x" << std::hex
<< overlap
7950 << ", new allocated 0x" << tmp_allocated
7951 << " released 0x" << tmp_released
<< std::dec
7953 pallocated
= &tmp_allocated
;
7954 preleased
= &tmp_released
;
7958 // update freelist with non-overlap sets
7959 for (interval_set
<uint64_t>::iterator p
= pallocated
->begin();
7960 p
!= pallocated
->end();
7962 fm
->allocate(p
.get_start(), p
.get_len(), t
);
7964 for (interval_set
<uint64_t>::iterator p
= preleased
->begin();
7965 p
!= preleased
->end();
7967 dout(20) << __func__
<< " release 0x" << std::hex
<< p
.get_start()
7968 << "~" << p
.get_len() << std::dec
<< dendl
;
7969 fm
->release(p
.get_start(), p
.get_len(), t
);
7972 _txc_update_store_statfs(txc
);
7975 void BlueStore::_txc_applied_kv(TransContext
*txc
)
7977 for (auto ls
: { &txc
->onodes
, &txc
->modified_objects
}) {
7978 for (auto& o
: *ls
) {
7979 dout(20) << __func__
<< " onode " << o
<< " had " << o
->flushing_count
7981 if (--o
->flushing_count
== 0) {
7982 std::lock_guard
<std::mutex
> l(o
->flush_lock
);
7983 o
->flush_cond
.notify_all();
7989 void BlueStore::_txc_committed_kv(TransContext
*txc
)
7991 dout(20) << __func__
<< " txc " << txc
<< dendl
;
7993 // warning: we're calling onreadable_sync inside the sequencer lock
7994 if (txc
->onreadable_sync
) {
7995 txc
->onreadable_sync
->complete(0);
7996 txc
->onreadable_sync
= NULL
;
7998 unsigned n
= txc
->osr
->parent
->shard_hint
.hash_to_shard(m_finisher_num
);
7999 if (txc
->oncommit
) {
8000 logger
->tinc(l_bluestore_commit_lat
, ceph_clock_now() - txc
->start
);
8001 finishers
[n
]->queue(txc
->oncommit
);
8002 txc
->oncommit
= NULL
;
8004 if (txc
->onreadable
) {
8005 finishers
[n
]->queue(txc
->onreadable
);
8006 txc
->onreadable
= NULL
;
8009 if (!txc
->oncommits
.empty()) {
8010 finishers
[n
]->queue(txc
->oncommits
);
8014 void BlueStore::_txc_finish(TransContext
*txc
)
8016 dout(20) << __func__
<< " " << txc
<< " onodes " << txc
->onodes
<< dendl
;
8017 assert(txc
->state
== TransContext::STATE_FINISHING
);
8019 for (auto& sb
: txc
->shared_blobs_written
) {
8020 sb
->bc
.finish_write(sb
->get_cache(), txc
->seq
);
8022 txc
->shared_blobs_written
.clear();
8024 while (!txc
->removed_collections
.empty()) {
8025 _queue_reap_collection(txc
->removed_collections
.front());
8026 txc
->removed_collections
.pop_front();
8029 OpSequencerRef osr
= txc
->osr
;
8031 bool submit_deferred
= false;
8032 OpSequencer::q_list_t releasing_txc
;
8034 std::lock_guard
<std::mutex
> l(osr
->qlock
);
8035 txc
->state
= TransContext::STATE_DONE
;
8036 bool notify
= false;
8037 while (!osr
->q
.empty()) {
8038 TransContext
*txc
= &osr
->q
.front();
8039 dout(20) << __func__
<< " txc " << txc
<< " " << txc
->get_state_name()
8041 if (txc
->state
!= TransContext::STATE_DONE
) {
8042 if (txc
->state
== TransContext::STATE_PREPARE
&&
8043 deferred_aggressive
) {
8044 // for _osr_drain_preceding()
8047 if (txc
->state
== TransContext::STATE_DEFERRED_QUEUED
&&
8048 osr
->q
.size() > g_conf
->bluestore_max_deferred_txc
) {
8049 submit_deferred
= true;
8055 releasing_txc
.push_back(*txc
);
8059 osr
->qcond
.notify_all();
8061 if (osr
->q
.empty()) {
8062 dout(20) << __func__
<< " osr " << osr
<< " q now empty" << dendl
;
8066 while (!releasing_txc
.empty()) {
8067 // release to allocator only after all preceding txc's have also
8068 // finished any deferred writes that potentially land in these
8070 auto txc
= &releasing_txc
.front();
8071 _txc_release_alloc(txc
);
8072 releasing_txc
.pop_front();
8073 txc
->log_state_latency(logger
, l_bluestore_state_done_lat
);
8077 if (submit_deferred
) {
8078 // we're pinning memory; flush! we could be more fine-grained here but
8079 // i'm not sure it's worth the bother.
8080 deferred_try_submit();
8083 if (empty
&& osr
->zombie
) {
8084 dout(10) << __func__
<< " reaping empty zombie osr " << osr
<< dendl
;
8089 void BlueStore::_txc_release_alloc(TransContext
*txc
)
8091 // update allocator with full released set
8092 if (!cct
->_conf
->bluestore_debug_no_reuse_blocks
) {
8093 dout(10) << __func__
<< " " << txc
<< " " << txc
->released
<< dendl
;
8094 for (interval_set
<uint64_t>::iterator p
= txc
->released
.begin();
8095 p
!= txc
->released
.end();
8097 alloc
->release(p
.get_start(), p
.get_len());
8101 txc
->allocated
.clear();
8102 txc
->released
.clear();
8105 void BlueStore::_osr_drain_preceding(TransContext
*txc
)
8107 OpSequencer
*osr
= txc
->osr
.get();
8108 dout(10) << __func__
<< " " << txc
<< " osr " << osr
<< dendl
;
8109 ++deferred_aggressive
; // FIXME: maybe osr-local aggressive flag?
8111 // submit anything pending
8112 deferred_lock
.lock();
8113 if (osr
->deferred_pending
) {
8114 _deferred_submit_unlock(osr
);
8116 deferred_lock
.unlock();
8120 // wake up any previously finished deferred events
8121 std::lock_guard
<std::mutex
> l(kv_lock
);
8122 kv_cond
.notify_one();
8124 osr
->drain_preceding(txc
);
8125 --deferred_aggressive
;
8126 dout(10) << __func__
<< " " << osr
<< " done" << dendl
;
8129 void BlueStore::_osr_drain_all()
8131 dout(10) << __func__
<< dendl
;
8133 set
<OpSequencerRef
> s
;
8135 std::lock_guard
<std::mutex
> l(osr_lock
);
8138 dout(20) << __func__
<< " osr_set " << s
<< dendl
;
8140 ++deferred_aggressive
;
8142 // submit anything pending
8143 deferred_try_submit();
8146 // wake up any previously finished deferred events
8147 std::lock_guard
<std::mutex
> l(kv_lock
);
8148 kv_cond
.notify_one();
8151 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8152 kv_finalize_cond
.notify_one();
8154 for (auto osr
: s
) {
8155 dout(20) << __func__
<< " drain " << osr
<< dendl
;
8158 --deferred_aggressive
;
8160 dout(10) << __func__
<< " done" << dendl
;
8163 void BlueStore::_osr_unregister_all()
8165 set
<OpSequencerRef
> s
;
8167 std::lock_guard
<std::mutex
> l(osr_lock
);
8170 dout(10) << __func__
<< " " << s
<< dendl
;
8171 for (auto osr
: s
) {
8175 // break link from Sequencer to us so that this OpSequencer
8176 // instance can die with this mount/umount cycle. note that
8177 // we assume umount() will not race against ~Sequencer.
8178 assert(osr
->parent
);
8179 osr
->parent
->p
.reset();
8182 // nobody should be creating sequencers during umount either.
8184 std::lock_guard
<std::mutex
> l(osr_lock
);
8185 assert(osr_set
.empty());
8189 void BlueStore::_kv_start()
8191 dout(10) << __func__
<< dendl
;
8193 if (cct
->_conf
->bluestore_shard_finishers
) {
8194 if (cct
->_conf
->osd_op_num_shards
) {
8195 m_finisher_num
= cct
->_conf
->osd_op_num_shards
;
8198 if (bdev
->is_rotational()) {
8199 m_finisher_num
= cct
->_conf
->osd_op_num_shards_hdd
;
8201 m_finisher_num
= cct
->_conf
->osd_op_num_shards_ssd
;
8206 assert(m_finisher_num
!= 0);
8208 for (int i
= 0; i
< m_finisher_num
; ++i
) {
8210 oss
<< "finisher-" << i
;
8211 Finisher
*f
= new Finisher(cct
, oss
.str(), "finisher");
8212 finishers
.push_back(f
);
8215 for (auto f
: finishers
) {
8218 kv_sync_thread
.create("bstore_kv_sync");
8219 kv_finalize_thread
.create("bstore_kv_final");
8222 void BlueStore::_kv_stop()
8224 dout(10) << __func__
<< dendl
;
8226 std::unique_lock
<std::mutex
> l(kv_lock
);
8227 while (!kv_sync_started
) {
8231 kv_cond
.notify_all();
8234 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8235 while (!kv_finalize_started
) {
8236 kv_finalize_cond
.wait(l
);
8238 kv_finalize_stop
= true;
8239 kv_finalize_cond
.notify_all();
8241 kv_sync_thread
.join();
8242 kv_finalize_thread
.join();
8244 std::lock_guard
<std::mutex
> l(kv_lock
);
8248 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8249 kv_finalize_stop
= false;
8251 dout(10) << __func__
<< " stopping finishers" << dendl
;
8252 for (auto f
: finishers
) {
8253 f
->wait_for_empty();
8256 dout(10) << __func__
<< " stopped" << dendl
;
8259 void BlueStore::_kv_sync_thread()
8261 dout(10) << __func__
<< " start" << dendl
;
8262 std::unique_lock
<std::mutex
> l(kv_lock
);
8263 assert(!kv_sync_started
);
8264 kv_sync_started
= true;
8265 kv_cond
.notify_all();
8267 assert(kv_committing
.empty());
8268 if (kv_queue
.empty() &&
8269 ((deferred_done_queue
.empty() && deferred_stable_queue
.empty()) ||
8270 !deferred_aggressive
)) {
8273 dout(20) << __func__
<< " sleep" << dendl
;
8275 dout(20) << __func__
<< " wake" << dendl
;
8277 deque
<TransContext
*> kv_submitting
;
8278 deque
<DeferredBatch
*> deferred_done
, deferred_stable
;
8279 uint64_t aios
= 0, costs
= 0;
8281 dout(20) << __func__
<< " committing " << kv_queue
.size()
8282 << " submitting " << kv_queue_unsubmitted
.size()
8283 << " deferred done " << deferred_done_queue
.size()
8284 << " stable " << deferred_stable_queue
.size()
8286 kv_committing
.swap(kv_queue
);
8287 kv_submitting
.swap(kv_queue_unsubmitted
);
8288 deferred_done
.swap(deferred_done_queue
);
8289 deferred_stable
.swap(deferred_stable_queue
);
8291 costs
= kv_throttle_costs
;
8293 kv_throttle_costs
= 0;
8294 utime_t start
= ceph_clock_now();
8297 dout(30) << __func__
<< " committing " << kv_committing
<< dendl
;
8298 dout(30) << __func__
<< " submitting " << kv_submitting
<< dendl
;
8299 dout(30) << __func__
<< " deferred_done " << deferred_done
<< dendl
;
8300 dout(30) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8302 bool force_flush
= false;
8303 // if bluefs is sharing the same device as data (only), then we
8304 // can rely on the bluefs commit to flush the device and make
8305 // deferred aios stable. that means that if we do have done deferred
8306 // txcs AND we are not on a single device, we need to force a flush.
8307 if (bluefs_single_shared_device
&& bluefs
) {
8310 } else if (kv_committing
.empty() && kv_submitting
.empty() &&
8311 deferred_stable
.empty()) {
8312 force_flush
= true; // there's nothing else to commit!
8313 } else if (deferred_aggressive
) {
8320 dout(20) << __func__
<< " num_aios=" << aios
8321 << " force_flush=" << (int)force_flush
8322 << ", flushing, deferred done->stable" << dendl
;
8323 // flush/barrier on block device
8326 // if we flush then deferred done are now deferred stable
8327 deferred_stable
.insert(deferred_stable
.end(), deferred_done
.begin(),
8328 deferred_done
.end());
8329 deferred_done
.clear();
8331 utime_t after_flush
= ceph_clock_now();
8333 // we will use one final transaction to force a sync
8334 KeyValueDB::Transaction synct
= db
->get_transaction();
8336 // increase {nid,blobid}_max? note that this covers both the
8337 // case where we are approaching the max and the case we passed
8338 // it. in either case, we increase the max in the earlier txn
8340 uint64_t new_nid_max
= 0, new_blobid_max
= 0;
8341 if (nid_last
+ cct
->_conf
->bluestore_nid_prealloc
/2 > nid_max
) {
8342 KeyValueDB::Transaction t
=
8343 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8344 new_nid_max
= nid_last
+ cct
->_conf
->bluestore_nid_prealloc
;
8346 ::encode(new_nid_max
, bl
);
8347 t
->set(PREFIX_SUPER
, "nid_max", bl
);
8348 dout(10) << __func__
<< " new_nid_max " << new_nid_max
<< dendl
;
8350 if (blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
/2 > blobid_max
) {
8351 KeyValueDB::Transaction t
=
8352 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8353 new_blobid_max
= blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
;
8355 ::encode(new_blobid_max
, bl
);
8356 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
8357 dout(10) << __func__
<< " new_blobid_max " << new_blobid_max
<< dendl
;
8360 for (auto txc
: kv_committing
) {
8361 if (txc
->state
== TransContext::STATE_KV_QUEUED
) {
8362 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8363 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
8365 _txc_applied_kv(txc
);
8366 --txc
->osr
->kv_committing_serially
;
8367 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
8368 if (txc
->osr
->kv_submitted_waiters
) {
8369 std::lock_guard
<std::mutex
> l(txc
->osr
->qlock
);
8370 if (txc
->osr
->_is_all_kv_submitted()) {
8371 txc
->osr
->qcond
.notify_all();
8376 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8377 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8380 --txc
->osr
->txc_with_unstable_io
;
8384 // release throttle *before* we commit. this allows new ops
8385 // to be prepared and enter pipeline while we are waiting on
8386 // the kv commit sync/flush. then hopefully on the next
8387 // iteration there will already be ops awake. otherwise, we
8388 // end up going to sleep, and then wake up when the very first
8389 // transaction is ready for commit.
8390 throttle_bytes
.put(costs
);
8392 PExtentVector bluefs_gift_extents
;
8394 after_flush
- bluefs_last_balance
>
8395 cct
->_conf
->bluestore_bluefs_balance_interval
) {
8396 bluefs_last_balance
= after_flush
;
8397 int r
= _balance_bluefs_freespace(&bluefs_gift_extents
);
8400 for (auto& p
: bluefs_gift_extents
) {
8401 bluefs_extents
.insert(p
.offset
, p
.length
);
8404 ::encode(bluefs_extents
, bl
);
8405 dout(10) << __func__
<< " bluefs_extents now 0x" << std::hex
8406 << bluefs_extents
<< std::dec
<< dendl
;
8407 synct
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
8411 // cleanup sync deferred keys
8412 for (auto b
: deferred_stable
) {
8413 for (auto& txc
: b
->txcs
) {
8414 bluestore_deferred_transaction_t
& wt
= *txc
.deferred_txn
;
8415 if (!wt
.released
.empty()) {
8416 // kraken replay compat only
8417 txc
.released
= wt
.released
;
8418 dout(10) << __func__
<< " deferred txn has released "
8420 << " (we just upgraded from kraken) on " << &txc
<< dendl
;
8421 _txc_finalize_kv(&txc
, synct
);
8423 // cleanup the deferred
8425 get_deferred_key(wt
.seq
, &key
);
8426 synct
->rm_single_key(PREFIX_DEFERRED
, key
);
8430 // submit synct synchronously (block and wait for it to commit)
8431 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction_sync(synct
);
8435 nid_max
= new_nid_max
;
8436 dout(10) << __func__
<< " nid_max now " << nid_max
<< dendl
;
8438 if (new_blobid_max
) {
8439 blobid_max
= new_blobid_max
;
8440 dout(10) << __func__
<< " blobid_max now " << blobid_max
<< dendl
;
8444 utime_t finish
= ceph_clock_now();
8445 utime_t dur_flush
= after_flush
- start
;
8446 utime_t dur_kv
= finish
- after_flush
;
8447 utime_t dur
= finish
- start
;
8448 dout(20) << __func__
<< " committed " << kv_committing
.size()
8449 << " cleaned " << deferred_stable
.size()
8451 << " (" << dur_flush
<< " flush + " << dur_kv
<< " kv commit)"
8453 logger
->tinc(l_bluestore_kv_flush_lat
, dur_flush
);
8454 logger
->tinc(l_bluestore_kv_commit_lat
, dur_kv
);
8455 logger
->tinc(l_bluestore_kv_lat
, dur
);
8459 if (!bluefs_gift_extents
.empty()) {
8460 _commit_bluefs_freespace(bluefs_gift_extents
);
8462 for (auto p
= bluefs_extents_reclaiming
.begin();
8463 p
!= bluefs_extents_reclaiming
.end();
8465 dout(20) << __func__
<< " releasing old bluefs 0x" << std::hex
8466 << p
.get_start() << "~" << p
.get_len() << std::dec
8468 alloc
->release(p
.get_start(), p
.get_len());
8470 bluefs_extents_reclaiming
.clear();
8474 std::unique_lock
<std::mutex
> m(kv_finalize_lock
);
8475 if (kv_committing_to_finalize
.empty()) {
8476 kv_committing_to_finalize
.swap(kv_committing
);
8478 kv_committing_to_finalize
.insert(
8479 kv_committing_to_finalize
.end(),
8480 kv_committing
.begin(),
8481 kv_committing
.end());
8482 kv_committing
.clear();
8484 if (deferred_stable_to_finalize
.empty()) {
8485 deferred_stable_to_finalize
.swap(deferred_stable
);
8487 deferred_stable_to_finalize
.insert(
8488 deferred_stable_to_finalize
.end(),
8489 deferred_stable
.begin(),
8490 deferred_stable
.end());
8491 deferred_stable
.clear();
8493 kv_finalize_cond
.notify_one();
8497 // previously deferred "done" are now "stable" by virtue of this
8499 deferred_stable_queue
.swap(deferred_done
);
8502 dout(10) << __func__
<< " finish" << dendl
;
8503 kv_sync_started
= false;
8506 void BlueStore::_kv_finalize_thread()
8508 deque
<TransContext
*> kv_committed
;
8509 deque
<DeferredBatch
*> deferred_stable
;
8510 dout(10) << __func__
<< " start" << dendl
;
8511 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8512 assert(!kv_finalize_started
);
8513 kv_finalize_started
= true;
8514 kv_finalize_cond
.notify_all();
8516 assert(kv_committed
.empty());
8517 assert(deferred_stable
.empty());
8518 if (kv_committing_to_finalize
.empty() &&
8519 deferred_stable_to_finalize
.empty()) {
8520 if (kv_finalize_stop
)
8522 dout(20) << __func__
<< " sleep" << dendl
;
8523 kv_finalize_cond
.wait(l
);
8524 dout(20) << __func__
<< " wake" << dendl
;
8526 kv_committed
.swap(kv_committing_to_finalize
);
8527 deferred_stable
.swap(deferred_stable_to_finalize
);
8529 dout(20) << __func__
<< " kv_committed " << kv_committed
<< dendl
;
8530 dout(20) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8532 while (!kv_committed
.empty()) {
8533 TransContext
*txc
= kv_committed
.front();
8534 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8535 _txc_state_proc(txc
);
8536 kv_committed
.pop_front();
8539 for (auto b
: deferred_stable
) {
8540 auto p
= b
->txcs
.begin();
8541 while (p
!= b
->txcs
.end()) {
8542 TransContext
*txc
= &*p
;
8543 p
= b
->txcs
.erase(p
); // unlink here because
8544 _txc_state_proc(txc
); // this may destroy txc
8548 deferred_stable
.clear();
8550 if (!deferred_aggressive
) {
8551 if (deferred_queue_size
>= deferred_batch_ops
.load() ||
8552 throttle_deferred_bytes
.past_midpoint()) {
8553 deferred_try_submit();
8557 // this is as good a place as any ...
8558 _reap_collections();
8563 dout(10) << __func__
<< " finish" << dendl
;
8564 kv_finalize_started
= false;
8567 bluestore_deferred_op_t
*BlueStore::_get_deferred_op(
8568 TransContext
*txc
, OnodeRef o
)
8570 if (!txc
->deferred_txn
) {
8571 txc
->deferred_txn
= new bluestore_deferred_transaction_t
;
8573 txc
->deferred_txn
->ops
.push_back(bluestore_deferred_op_t());
8574 return &txc
->deferred_txn
->ops
.back();
8577 void BlueStore::_deferred_queue(TransContext
*txc
)
8579 dout(20) << __func__
<< " txc " << txc
<< " osr " << txc
->osr
<< dendl
;
8580 deferred_lock
.lock();
8581 if (!txc
->osr
->deferred_pending
&&
8582 !txc
->osr
->deferred_running
) {
8583 deferred_queue
.push_back(*txc
->osr
);
8585 if (!txc
->osr
->deferred_pending
) {
8586 txc
->osr
->deferred_pending
= new DeferredBatch(cct
, txc
->osr
.get());
8588 ++deferred_queue_size
;
8589 txc
->osr
->deferred_pending
->txcs
.push_back(*txc
);
8590 bluestore_deferred_transaction_t
& wt
= *txc
->deferred_txn
;
8591 for (auto opi
= wt
.ops
.begin(); opi
!= wt
.ops
.end(); ++opi
) {
8592 const auto& op
= *opi
;
8593 assert(op
.op
== bluestore_deferred_op_t::OP_WRITE
);
8594 bufferlist::const_iterator p
= op
.data
.begin();
8595 for (auto e
: op
.extents
) {
8596 txc
->osr
->deferred_pending
->prepare_write(
8597 cct
, wt
.seq
, e
.offset
, e
.length
, p
);
8600 if (deferred_aggressive
&&
8601 !txc
->osr
->deferred_running
) {
8602 _deferred_submit_unlock(txc
->osr
.get());
8604 deferred_lock
.unlock();
8608 void BlueStore::deferred_try_submit()
8610 dout(20) << __func__
<< " " << deferred_queue
.size() << " osrs, "
8611 << deferred_queue_size
<< " txcs" << dendl
;
8612 std::lock_guard
<std::mutex
> l(deferred_lock
);
8613 vector
<OpSequencerRef
> osrs
;
8614 osrs
.reserve(deferred_queue
.size());
8615 for (auto& osr
: deferred_queue
) {
8616 osrs
.push_back(&osr
);
8618 for (auto& osr
: osrs
) {
8619 if (osr
->deferred_pending
&& !osr
->deferred_running
) {
8620 _deferred_submit_unlock(osr
.get());
8621 deferred_lock
.lock();
8626 void BlueStore::_deferred_submit_unlock(OpSequencer
*osr
)
8628 dout(10) << __func__
<< " osr " << osr
8629 << " " << osr
->deferred_pending
->iomap
.size() << " ios pending "
8631 assert(osr
->deferred_pending
);
8632 assert(!osr
->deferred_running
);
8634 auto b
= osr
->deferred_pending
;
8635 deferred_queue_size
-= b
->seq_bytes
.size();
8636 assert(deferred_queue_size
>= 0);
8638 osr
->deferred_running
= osr
->deferred_pending
;
8639 osr
->deferred_pending
= nullptr;
8641 uint64_t start
= 0, pos
= 0;
8643 auto i
= b
->iomap
.begin();
8645 if (i
== b
->iomap
.end() || i
->first
!= pos
) {
8647 dout(20) << __func__
<< " write 0x" << std::hex
8648 << start
<< "~" << bl
.length()
8649 << " crc " << bl
.crc32c(-1) << std::dec
<< dendl
;
8650 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
8651 logger
->inc(l_bluestore_deferred_write_ops
);
8652 logger
->inc(l_bluestore_deferred_write_bytes
, bl
.length());
8653 int r
= bdev
->aio_write(start
, bl
, &b
->ioc
, false);
8657 if (i
== b
->iomap
.end()) {
8664 dout(20) << __func__
<< " seq " << i
->second
.seq
<< " 0x"
8665 << std::hex
<< pos
<< "~" << i
->second
.bl
.length() << std::dec
8670 pos
+= i
->second
.bl
.length();
8671 bl
.claim_append(i
->second
.bl
);
8675 // demote to deferred_submit_lock, then drop that too
8676 std::lock_guard
<std::mutex
> l(deferred_submit_lock
);
8677 deferred_lock
.unlock();
8678 bdev
->aio_submit(&b
->ioc
);
8681 void BlueStore::_deferred_aio_finish(OpSequencer
*osr
)
8683 dout(10) << __func__
<< " osr " << osr
<< dendl
;
8684 assert(osr
->deferred_running
);
8685 DeferredBatch
*b
= osr
->deferred_running
;
8688 std::lock_guard
<std::mutex
> l(deferred_lock
);
8689 assert(osr
->deferred_running
== b
);
8690 osr
->deferred_running
= nullptr;
8691 if (!osr
->deferred_pending
) {
8692 auto q
= deferred_queue
.iterator_to(*osr
);
8693 deferred_queue
.erase(q
);
8694 } else if (deferred_aggressive
) {
8695 dout(20) << __func__
<< " queuing async deferred_try_submit" << dendl
;
8696 finishers
[0]->queue(new FunctionContext([&](int) {
8697 deferred_try_submit();
8704 std::lock_guard
<std::mutex
> l2(osr
->qlock
);
8705 for (auto& i
: b
->txcs
) {
8706 TransContext
*txc
= &i
;
8707 txc
->state
= TransContext::STATE_DEFERRED_CLEANUP
;
8710 osr
->qcond
.notify_all();
8711 throttle_deferred_bytes
.put(costs
);
8712 std::lock_guard
<std::mutex
> l(kv_lock
);
8713 deferred_done_queue
.emplace_back(b
);
8716 // in the normal case, do not bother waking up the kv thread; it will
8717 // catch us on the next commit anyway.
8718 if (deferred_aggressive
) {
8719 std::lock_guard
<std::mutex
> l(kv_lock
);
8720 kv_cond
.notify_one();
8724 int BlueStore::_deferred_replay()
8726 dout(10) << __func__
<< " start" << dendl
;
8727 OpSequencerRef osr
= new OpSequencer(cct
, this);
8730 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_DEFERRED
);
8731 for (it
->lower_bound(string()); it
->valid(); it
->next(), ++count
) {
8732 dout(20) << __func__
<< " replay " << pretty_binary_string(it
->key())
8734 bluestore_deferred_transaction_t
*deferred_txn
=
8735 new bluestore_deferred_transaction_t
;
8736 bufferlist bl
= it
->value();
8737 bufferlist::iterator p
= bl
.begin();
8739 ::decode(*deferred_txn
, p
);
8740 } catch (buffer::error
& e
) {
8741 derr
<< __func__
<< " failed to decode deferred txn "
8742 << pretty_binary_string(it
->key()) << dendl
;
8743 delete deferred_txn
;
8747 TransContext
*txc
= _txc_create(osr
.get());
8748 txc
->deferred_txn
= deferred_txn
;
8749 txc
->state
= TransContext::STATE_KV_DONE
;
8750 _txc_state_proc(txc
);
8753 dout(20) << __func__
<< " draining osr" << dendl
;
8756 dout(10) << __func__
<< " completed " << count
<< " events" << dendl
;
8760 // ---------------------------
8763 int BlueStore::queue_transactions(
8765 vector
<Transaction
>& tls
,
8767 ThreadPool::TPHandle
*handle
)
8770 Context
*onreadable
;
8772 Context
*onreadable_sync
;
8773 ObjectStore::Transaction::collect_contexts(
8774 tls
, &onreadable
, &ondisk
, &onreadable_sync
);
8776 if (cct
->_conf
->objectstore_blackhole
) {
8777 dout(0) << __func__
<< " objectstore_blackhole = TRUE, dropping transaction"
8781 delete onreadable_sync
;
8784 utime_t start
= ceph_clock_now();
8785 // set up the sequencer
8789 osr
= static_cast<OpSequencer
*>(posr
->p
.get());
8790 dout(10) << __func__
<< " existing " << osr
<< " " << *osr
<< dendl
;
8792 osr
= new OpSequencer(cct
, this);
8795 dout(10) << __func__
<< " new " << osr
<< " " << *osr
<< dendl
;
8799 TransContext
*txc
= _txc_create(osr
);
8800 txc
->onreadable
= onreadable
;
8801 txc
->onreadable_sync
= onreadable_sync
;
8802 txc
->oncommit
= ondisk
;
8804 for (vector
<Transaction
>::iterator p
= tls
.begin(); p
!= tls
.end(); ++p
) {
8806 txc
->bytes
+= (*p
).get_num_bytes();
8807 _txc_add_transaction(txc
, &(*p
));
8809 _txc_calc_cost(txc
);
8811 _txc_write_nodes(txc
, txc
->t
);
8813 // journal deferred items
8814 if (txc
->deferred_txn
) {
8815 txc
->deferred_txn
->seq
= ++deferred_seq
;
8817 ::encode(*txc
->deferred_txn
, bl
);
8819 get_deferred_key(txc
->deferred_txn
->seq
, &key
);
8820 txc
->t
->set(PREFIX_DEFERRED
, key
, bl
);
8823 _txc_finalize_kv(txc
, txc
->t
);
8825 handle
->suspend_tp_timeout();
8827 utime_t tstart
= ceph_clock_now();
8828 throttle_bytes
.get(txc
->cost
);
8829 if (txc
->deferred_txn
) {
8830 // ensure we do not block here because of deferred writes
8831 if (!throttle_deferred_bytes
.get_or_fail(txc
->cost
)) {
8832 dout(10) << __func__
<< " failed get throttle_deferred_bytes, aggressive"
8834 ++deferred_aggressive
;
8835 deferred_try_submit();
8836 throttle_deferred_bytes
.get(txc
->cost
);
8837 --deferred_aggressive
;
8840 utime_t tend
= ceph_clock_now();
8843 handle
->reset_tp_timeout();
8845 logger
->inc(l_bluestore_txc
);
8848 _txc_state_proc(txc
);
8850 logger
->tinc(l_bluestore_submit_lat
, ceph_clock_now() - start
);
8851 logger
->tinc(l_bluestore_throttle_lat
, tend
- tstart
);
8855 void BlueStore::_txc_aio_submit(TransContext
*txc
)
8857 dout(10) << __func__
<< " txc " << txc
<< dendl
;
8858 bdev
->aio_submit(&txc
->ioc
);
8861 void BlueStore::_txc_add_transaction(TransContext
*txc
, Transaction
*t
)
8863 Transaction::iterator i
= t
->begin();
8865 _dump_transaction(t
);
8867 vector
<CollectionRef
> cvec(i
.colls
.size());
8869 for (vector
<coll_t
>::iterator p
= i
.colls
.begin(); p
!= i
.colls
.end();
8871 cvec
[j
] = _get_collection(*p
);
8873 vector
<OnodeRef
> ovec(i
.objects
.size());
8875 for (int pos
= 0; i
.have_op(); ++pos
) {
8876 Transaction::Op
*op
= i
.decode_op();
8880 if (op
->op
== Transaction::OP_NOP
)
8883 // collection operations
8884 CollectionRef
&c
= cvec
[op
->cid
];
8886 case Transaction::OP_RMCOLL
:
8888 const coll_t
&cid
= i
.get_cid(op
->cid
);
8889 r
= _remove_collection(txc
, cid
, &c
);
8895 case Transaction::OP_MKCOLL
:
8898 const coll_t
&cid
= i
.get_cid(op
->cid
);
8899 r
= _create_collection(txc
, cid
, op
->split_bits
, &c
);
8905 case Transaction::OP_SPLIT_COLLECTION
:
8906 assert(0 == "deprecated");
8909 case Transaction::OP_SPLIT_COLLECTION2
:
8911 uint32_t bits
= op
->split_bits
;
8912 uint32_t rem
= op
->split_rem
;
8913 r
= _split_collection(txc
, c
, cvec
[op
->dest_cid
], bits
, rem
);
8919 case Transaction::OP_COLL_HINT
:
8921 uint32_t type
= op
->hint_type
;
8924 bufferlist::iterator hiter
= hint
.begin();
8925 if (type
== Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS
) {
8928 ::decode(pg_num
, hiter
);
8929 ::decode(num_objs
, hiter
);
8930 dout(10) << __func__
<< " collection hint objects is a no-op, "
8931 << " pg_num " << pg_num
<< " num_objects " << num_objs
8935 dout(10) << __func__
<< " unknown collection hint " << type
<< dendl
;
8941 case Transaction::OP_COLL_SETATTR
:
8945 case Transaction::OP_COLL_RMATTR
:
8949 case Transaction::OP_COLL_RENAME
:
8950 assert(0 == "not implemented");
8954 derr
<< __func__
<< " error " << cpp_strerror(r
)
8955 << " not handled on operation " << op
->op
8956 << " (op " << pos
<< ", counting from 0)" << dendl
;
8957 _dump_transaction(t
, 0);
8958 assert(0 == "unexpected error");
8961 // these operations implicity create the object
8962 bool create
= false;
8963 if (op
->op
== Transaction::OP_TOUCH
||
8964 op
->op
== Transaction::OP_WRITE
||
8965 op
->op
== Transaction::OP_ZERO
) {
8969 // object operations
8970 RWLock::WLocker
l(c
->lock
);
8971 OnodeRef
&o
= ovec
[op
->oid
];
8973 ghobject_t oid
= i
.get_oid(op
->oid
);
8974 o
= c
->get_onode(oid
, create
);
8976 if (!create
&& (!o
|| !o
->exists
)) {
8977 dout(10) << __func__
<< " op " << op
->op
<< " got ENOENT on "
8978 << i
.get_oid(op
->oid
) << dendl
;
8984 case Transaction::OP_TOUCH
:
8985 r
= _touch(txc
, c
, o
);
8988 case Transaction::OP_WRITE
:
8990 uint64_t off
= op
->off
;
8991 uint64_t len
= op
->len
;
8992 uint32_t fadvise_flags
= i
.get_fadvise_flags();
8995 r
= _write(txc
, c
, o
, off
, len
, bl
, fadvise_flags
);
8999 case Transaction::OP_ZERO
:
9001 uint64_t off
= op
->off
;
9002 uint64_t len
= op
->len
;
9003 r
= _zero(txc
, c
, o
, off
, len
);
9007 case Transaction::OP_TRIMCACHE
:
9009 // deprecated, no-op
9013 case Transaction::OP_TRUNCATE
:
9015 uint64_t off
= op
->off
;
9016 r
= _truncate(txc
, c
, o
, off
);
9020 case Transaction::OP_REMOVE
:
9022 r
= _remove(txc
, c
, o
);
9026 case Transaction::OP_SETATTR
:
9028 string name
= i
.decode_string();
9031 r
= _setattr(txc
, c
, o
, name
, bp
);
9035 case Transaction::OP_SETATTRS
:
9037 map
<string
, bufferptr
> aset
;
9038 i
.decode_attrset(aset
);
9039 r
= _setattrs(txc
, c
, o
, aset
);
9043 case Transaction::OP_RMATTR
:
9045 string name
= i
.decode_string();
9046 r
= _rmattr(txc
, c
, o
, name
);
9050 case Transaction::OP_RMATTRS
:
9052 r
= _rmattrs(txc
, c
, o
);
9056 case Transaction::OP_CLONE
:
9058 OnodeRef
& no
= ovec
[op
->dest_oid
];
9060 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9061 no
= c
->get_onode(noid
, true);
9063 r
= _clone(txc
, c
, o
, no
);
9067 case Transaction::OP_CLONERANGE
:
9068 assert(0 == "deprecated");
9071 case Transaction::OP_CLONERANGE2
:
9073 OnodeRef
& no
= ovec
[op
->dest_oid
];
9075 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9076 no
= c
->get_onode(noid
, true);
9078 uint64_t srcoff
= op
->off
;
9079 uint64_t len
= op
->len
;
9080 uint64_t dstoff
= op
->dest_off
;
9081 r
= _clone_range(txc
, c
, o
, no
, srcoff
, len
, dstoff
);
9085 case Transaction::OP_COLL_ADD
:
9086 assert(0 == "not implemented");
9089 case Transaction::OP_COLL_REMOVE
:
9090 assert(0 == "not implemented");
9093 case Transaction::OP_COLL_MOVE
:
9094 assert(0 == "deprecated");
9097 case Transaction::OP_COLL_MOVE_RENAME
:
9098 case Transaction::OP_TRY_RENAME
:
9100 assert(op
->cid
== op
->dest_cid
);
9101 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9102 OnodeRef
& no
= ovec
[op
->dest_oid
];
9104 no
= c
->get_onode(noid
, false);
9106 r
= _rename(txc
, c
, o
, no
, noid
);
9110 case Transaction::OP_OMAP_CLEAR
:
9112 r
= _omap_clear(txc
, c
, o
);
9115 case Transaction::OP_OMAP_SETKEYS
:
9118 i
.decode_attrset_bl(&aset_bl
);
9119 r
= _omap_setkeys(txc
, c
, o
, aset_bl
);
9122 case Transaction::OP_OMAP_RMKEYS
:
9125 i
.decode_keyset_bl(&keys_bl
);
9126 r
= _omap_rmkeys(txc
, c
, o
, keys_bl
);
9129 case Transaction::OP_OMAP_RMKEYRANGE
:
9132 first
= i
.decode_string();
9133 last
= i
.decode_string();
9134 r
= _omap_rmkey_range(txc
, c
, o
, first
, last
);
9137 case Transaction::OP_OMAP_SETHEADER
:
9141 r
= _omap_setheader(txc
, c
, o
, bl
);
9145 case Transaction::OP_SETALLOCHINT
:
9147 r
= _set_alloc_hint(txc
, c
, o
,
9148 op
->expected_object_size
,
9149 op
->expected_write_size
,
9150 op
->alloc_hint_flags
);
9155 derr
<< __func__
<< "bad op " << op
->op
<< dendl
;
9163 if (r
== -ENOENT
&& !(op
->op
== Transaction::OP_CLONERANGE
||
9164 op
->op
== Transaction::OP_CLONE
||
9165 op
->op
== Transaction::OP_CLONERANGE2
||
9166 op
->op
== Transaction::OP_COLL_ADD
||
9167 op
->op
== Transaction::OP_SETATTR
||
9168 op
->op
== Transaction::OP_SETATTRS
||
9169 op
->op
== Transaction::OP_RMATTR
||
9170 op
->op
== Transaction::OP_OMAP_SETKEYS
||
9171 op
->op
== Transaction::OP_OMAP_RMKEYS
||
9172 op
->op
== Transaction::OP_OMAP_RMKEYRANGE
||
9173 op
->op
== Transaction::OP_OMAP_SETHEADER
))
9174 // -ENOENT is usually okay
9180 const char *msg
= "unexpected error code";
9182 if (r
== -ENOENT
&& (op
->op
== Transaction::OP_CLONERANGE
||
9183 op
->op
== Transaction::OP_CLONE
||
9184 op
->op
== Transaction::OP_CLONERANGE2
))
9185 msg
= "ENOENT on clone suggests osd bug";
9188 // For now, if we hit _any_ ENOSPC, crash, before we do any damage
9189 // by partially applying transactions.
9190 msg
= "ENOSPC from bluestore, misconfigured cluster";
9192 if (r
== -ENOTEMPTY
) {
9193 msg
= "ENOTEMPTY suggests garbage data in osd data dir";
9196 derr
<< __func__
<< " error " << cpp_strerror(r
)
9197 << " not handled on operation " << op
->op
9198 << " (op " << pos
<< ", counting from 0)"
9200 derr
<< msg
<< dendl
;
9201 _dump_transaction(t
, 0);
9202 assert(0 == "unexpected error");
9210 // -----------------
9213 int BlueStore::_touch(TransContext
*txc
,
9217 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
9219 _assign_nid(txc
, o
);
9220 txc
->write_onode(o
);
9221 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
9225 void BlueStore::_dump_onode(OnodeRef o
, int log_level
)
9227 if (!cct
->_conf
->subsys
.should_gather(ceph_subsys_bluestore
, log_level
))
9229 dout(log_level
) << __func__
<< " " << o
<< " " << o
->oid
9230 << " nid " << o
->onode
.nid
9231 << " size 0x" << std::hex
<< o
->onode
.size
9232 << " (" << std::dec
<< o
->onode
.size
<< ")"
9233 << " expected_object_size " << o
->onode
.expected_object_size
9234 << " expected_write_size " << o
->onode
.expected_write_size
9235 << " in " << o
->onode
.extent_map_shards
.size() << " shards"
9236 << ", " << o
->extent_map
.spanning_blob_map
.size()
9237 << " spanning blobs"
9239 for (auto p
= o
->onode
.attrs
.begin();
9240 p
!= o
->onode
.attrs
.end();
9242 dout(log_level
) << __func__
<< " attr " << p
->first
9243 << " len " << p
->second
.length() << dendl
;
9245 _dump_extent_map(o
->extent_map
, log_level
);
9248 void BlueStore::_dump_extent_map(ExtentMap
&em
, int log_level
)
9251 for (auto& s
: em
.shards
) {
9252 dout(log_level
) << __func__
<< " shard " << *s
.shard_info
9253 << (s
.loaded
? " (loaded)" : "")
9254 << (s
.dirty
? " (dirty)" : "")
9257 for (auto& e
: em
.extent_map
) {
9258 dout(log_level
) << __func__
<< " " << e
<< dendl
;
9259 assert(e
.logical_offset
>= pos
);
9260 pos
= e
.logical_offset
+ e
.length
;
9261 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
9262 if (blob
.has_csum()) {
9264 unsigned n
= blob
.get_csum_count();
9265 for (unsigned i
= 0; i
< n
; ++i
)
9266 v
.push_back(blob
.get_csum_item(i
));
9267 dout(log_level
) << __func__
<< " csum: " << std::hex
<< v
<< std::dec
9270 std::lock_guard
<std::recursive_mutex
> l(e
.blob
->shared_blob
->get_cache()->lock
);
9271 for (auto& i
: e
.blob
->shared_blob
->bc
.buffer_map
) {
9272 dout(log_level
) << __func__
<< " 0x" << std::hex
<< i
.first
9273 << "~" << i
.second
->length
<< std::dec
9274 << " " << *i
.second
<< dendl
;
9279 void BlueStore::_dump_transaction(Transaction
*t
, int log_level
)
9281 dout(log_level
) << " transaction dump:\n";
9282 JSONFormatter
f(true);
9283 f
.open_object_section("transaction");
9290 void BlueStore::_pad_zeros(
9291 bufferlist
*bl
, uint64_t *offset
,
9292 uint64_t chunk_size
)
9294 auto length
= bl
->length();
9295 dout(30) << __func__
<< " 0x" << std::hex
<< *offset
<< "~" << length
9296 << " chunk_size 0x" << chunk_size
<< std::dec
<< dendl
;
9297 dout(40) << "before:\n";
9298 bl
->hexdump(*_dout
);
9301 size_t front_pad
= *offset
% chunk_size
;
9302 size_t back_pad
= 0;
9303 size_t pad_count
= 0;
9305 size_t front_copy
= MIN(chunk_size
- front_pad
, length
);
9306 bufferptr z
= buffer::create_page_aligned(chunk_size
);
9307 z
.zero(0, front_pad
, false);
9308 pad_count
+= front_pad
;
9309 bl
->copy(0, front_copy
, z
.c_str() + front_pad
);
9310 if (front_copy
+ front_pad
< chunk_size
) {
9311 back_pad
= chunk_size
- (length
+ front_pad
);
9312 z
.zero(front_pad
+ length
, back_pad
, false);
9313 pad_count
+= back_pad
;
9317 t
.substr_of(old
, front_copy
, length
- front_copy
);
9319 bl
->claim_append(t
);
9320 *offset
-= front_pad
;
9321 length
+= pad_count
;
9325 uint64_t end
= *offset
+ length
;
9326 unsigned back_copy
= end
% chunk_size
;
9328 assert(back_pad
== 0);
9329 back_pad
= chunk_size
- back_copy
;
9330 assert(back_copy
<= length
);
9331 bufferptr
tail(chunk_size
);
9332 bl
->copy(length
- back_copy
, back_copy
, tail
.c_str());
9333 tail
.zero(back_copy
, back_pad
, false);
9336 bl
->substr_of(old
, 0, length
- back_copy
);
9339 pad_count
+= back_pad
;
9341 dout(20) << __func__
<< " pad 0x" << std::hex
<< front_pad
<< " + 0x"
9342 << back_pad
<< " on front/back, now 0x" << *offset
<< "~"
9343 << length
<< std::dec
<< dendl
;
9344 dout(40) << "after:\n";
9345 bl
->hexdump(*_dout
);
9348 logger
->inc(l_bluestore_write_pad_bytes
, pad_count
);
9349 assert(bl
->length() == length
);
9352 void BlueStore::_do_write_small(
9356 uint64_t offset
, uint64_t length
,
9357 bufferlist::iterator
& blp
,
9360 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9361 << std::dec
<< dendl
;
9362 assert(length
< min_alloc_size
);
9363 uint64_t end_offs
= offset
+ length
;
9365 logger
->inc(l_bluestore_write_small
);
9366 logger
->inc(l_bluestore_write_small_bytes
, length
);
9369 blp
.copy(length
, bl
);
9371 // Look for an existing mutable blob we can use.
9372 auto begin
= o
->extent_map
.extent_map
.begin();
9373 auto end
= o
->extent_map
.extent_map
.end();
9374 auto ep
= o
->extent_map
.seek_lextent(offset
);
9377 if (ep
->blob_end() <= offset
) {
9382 if (prev_ep
!= begin
) {
9385 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9388 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9389 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9390 uint32_t alloc_len
= min_alloc_size
;
9391 auto offset0
= P2ALIGN(offset
, alloc_len
);
9395 // search suitable extent in both forward and reverse direction in
9396 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9397 // then check if blob can be reused via can_reuse_blob func or apply
9398 // direct/deferred write (the latter for extents including or higher
9399 // than 'offset' only).
9403 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9404 BlobRef b
= ep
->blob
;
9405 auto bstart
= ep
->blob_start();
9406 dout(20) << __func__
<< " considering " << *b
9407 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9408 if (bstart
>= end_offs
) {
9409 dout(20) << __func__
<< " ignoring distant " << *b
<< dendl
;
9410 } else if (!b
->get_blob().is_mutable()) {
9411 dout(20) << __func__
<< " ignoring immutable " << *b
<< dendl
;
9412 } else if (ep
->logical_offset
% min_alloc_size
!=
9413 ep
->blob_offset
% min_alloc_size
) {
9414 dout(20) << __func__
<< " ignoring offset-skewed " << *b
<< dendl
;
9416 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9417 // can we pad our head/tail out with zeros?
9418 uint64_t head_pad
, tail_pad
;
9419 head_pad
= P2PHASE(offset
, chunk_size
);
9420 tail_pad
= P2NPHASE(end_offs
, chunk_size
);
9421 if (head_pad
|| tail_pad
) {
9422 o
->extent_map
.fault_range(db
, offset
- head_pad
,
9423 end_offs
- offset
+ head_pad
+ tail_pad
);
9426 o
->extent_map
.has_any_lextents(offset
- head_pad
, chunk_size
)) {
9429 if (tail_pad
&& o
->extent_map
.has_any_lextents(end_offs
, tail_pad
)) {
9433 uint64_t b_off
= offset
- head_pad
- bstart
;
9434 uint64_t b_len
= length
+ head_pad
+ tail_pad
;
9436 // direct write into unused blocks of an existing mutable blob?
9437 if ((b_off
% chunk_size
== 0 && b_len
% chunk_size
== 0) &&
9438 b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9439 b
->get_blob().is_unused(b_off
, b_len
) &&
9440 b
->get_blob().is_allocated(b_off
, b_len
)) {
9441 _apply_padding(head_pad
, tail_pad
, bl
);
9443 dout(20) << __func__
<< " write to unused 0x" << std::hex
9444 << b_off
<< "~" << b_len
9445 << " pad 0x" << head_pad
<< " + 0x" << tail_pad
9446 << std::dec
<< " of mutable " << *b
<< dendl
;
9447 _buffer_cache_write(txc
, b
, b_off
, bl
,
9448 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9450 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
9451 if (b_len
<= prefer_deferred_size
) {
9452 dout(20) << __func__
<< " deferring small 0x" << std::hex
9453 << b_len
<< std::dec
<< " unused write via deferred" << dendl
;
9454 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9455 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9458 [&](uint64_t offset
, uint64_t length
) {
9459 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9464 b
->get_blob().map_bl(
9466 [&](uint64_t offset
, bufferlist
& t
) {
9467 bdev
->aio_write(offset
, t
,
9468 &txc
->ioc
, wctx
->buffered
);
9472 b
->dirty_blob().calc_csum(b_off
, bl
);
9473 dout(20) << __func__
<< " lex old " << *ep
<< dendl
;
9474 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, b_off
+ head_pad
, length
,
9476 &wctx
->old_extents
);
9477 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9478 txc
->statfs_delta
.stored() += le
->length
;
9479 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9480 logger
->inc(l_bluestore_write_small_unused
);
9483 // read some data to fill out the chunk?
9484 uint64_t head_read
= P2PHASE(b_off
, chunk_size
);
9485 uint64_t tail_read
= P2NPHASE(b_off
+ b_len
, chunk_size
);
9486 if ((head_read
|| tail_read
) &&
9487 (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
+ tail_read
) &&
9488 head_read
+ tail_read
< min_alloc_size
) {
9490 b_len
+= head_read
+ tail_read
;
9493 head_read
= tail_read
= 0;
9496 // chunk-aligned deferred overwrite?
9497 if (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9498 b_off
% chunk_size
== 0 &&
9499 b_len
% chunk_size
== 0 &&
9500 b
->get_blob().is_allocated(b_off
, b_len
)) {
9502 _apply_padding(head_pad
, tail_pad
, bl
);
9504 dout(20) << __func__
<< " reading head 0x" << std::hex
<< head_read
9505 << " and tail 0x" << tail_read
<< std::dec
<< dendl
;
9508 int r
= _do_read(c
.get(), o
, offset
- head_pad
- head_read
, head_read
,
9510 assert(r
>= 0 && r
<= (int)head_read
);
9511 size_t zlen
= head_read
- r
;
9513 head_bl
.append_zero(zlen
);
9514 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9516 bl
.claim_prepend(head_bl
);
9517 logger
->inc(l_bluestore_write_penalty_read_ops
);
9521 int r
= _do_read(c
.get(), o
, offset
+ length
+ tail_pad
, tail_read
,
9523 assert(r
>= 0 && r
<= (int)tail_read
);
9524 size_t zlen
= tail_read
- r
;
9526 tail_bl
.append_zero(zlen
);
9527 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9529 bl
.claim_append(tail_bl
);
9530 logger
->inc(l_bluestore_write_penalty_read_ops
);
9532 logger
->inc(l_bluestore_write_small_pre_read
);
9534 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9535 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9536 _buffer_cache_write(txc
, b
, b_off
, bl
,
9537 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9539 int r
= b
->get_blob().map(
9541 [&](uint64_t offset
, uint64_t length
) {
9542 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9546 if (b
->get_blob().csum_type
) {
9547 b
->dirty_blob().calc_csum(b_off
, bl
);
9550 dout(20) << __func__
<< " deferred write 0x" << std::hex
<< b_off
<< "~"
9551 << b_len
<< std::dec
<< " of mutable " << *b
9552 << " at " << op
->extents
<< dendl
;
9553 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, offset
- bstart
, length
,
9554 b
, &wctx
->old_extents
);
9555 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9556 txc
->statfs_delta
.stored() += le
->length
;
9557 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9558 logger
->inc(l_bluestore_write_small_deferred
);
9561 // try to reuse blob if we can
9562 if (b
->can_reuse_blob(min_alloc_size
,
9566 assert(alloc_len
== min_alloc_size
); // expecting data always
9567 // fit into reused blob
9568 // Need to check for pending writes desiring to
9569 // reuse the same pextent. The rationale is that during GC two chunks
9570 // from garbage blobs(compressed?) can share logical space within the same
9571 // AU. That's in turn might be caused by unaligned len in clone_range2.
9572 // Hence the second write will fail in an attempt to reuse blob at
9573 // do_alloc_write().
9574 if (!wctx
->has_conflict(b
,
9576 offset0
+ alloc_len
,
9579 // we can't reuse pad_head/pad_tail since they might be truncated
9580 // due to existent extents
9581 uint64_t b_off
= offset
- bstart
;
9582 uint64_t b_off0
= b_off
;
9583 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9585 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9586 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9587 << " (0x" << b_off
<< "~" << length
<< ")"
9588 << std::dec
<< dendl
;
9590 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9591 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9593 logger
->inc(l_bluestore_write_small_unused
);
9600 } // if (ep != end && ep->logical_offset < offset + max_bsize)
9602 // check extent for reuse in reverse order
9603 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9604 BlobRef b
= prev_ep
->blob
;
9605 auto bstart
= prev_ep
->blob_start();
9606 dout(20) << __func__
<< " considering " << *b
9607 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9608 if (b
->can_reuse_blob(min_alloc_size
,
9612 assert(alloc_len
== min_alloc_size
); // expecting data always
9613 // fit into reused blob
9614 // Need to check for pending writes desiring to
9615 // reuse the same pextent. The rationale is that during GC two chunks
9616 // from garbage blobs(compressed?) can share logical space within the same
9617 // AU. That's in turn might be caused by unaligned len in clone_range2.
9618 // Hence the second write will fail in an attempt to reuse blob at
9619 // do_alloc_write().
9620 if (!wctx
->has_conflict(b
,
9622 offset0
+ alloc_len
,
9625 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9626 uint64_t b_off
= offset
- bstart
;
9627 uint64_t b_off0
= b_off
;
9628 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9630 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9631 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9632 << " (0x" << b_off
<< "~" << length
<< ")"
9633 << std::dec
<< dendl
;
9635 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9636 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9638 logger
->inc(l_bluestore_write_small_unused
);
9642 if (prev_ep
!= begin
) {
9646 prev_ep
= end
; // to avoid useless first extent re-check
9648 } // if (prev_ep != end && prev_ep->logical_offset >= min_off)
9649 } while (any_change
);
9653 BlobRef b
= c
->new_blob();
9654 uint64_t b_off
= P2PHASE(offset
, alloc_len
);
9655 uint64_t b_off0
= b_off
;
9656 _pad_zeros(&bl
, &b_off0
, block_size
);
9657 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9658 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
, true, true);
9659 logger
->inc(l_bluestore_write_small_new
);
9664 void BlueStore::_do_write_big(
9668 uint64_t offset
, uint64_t length
,
9669 bufferlist::iterator
& blp
,
9672 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9673 << " target_blob_size 0x" << wctx
->target_blob_size
<< std::dec
9674 << " compress " << (int)wctx
->compress
9676 logger
->inc(l_bluestore_write_big
);
9677 logger
->inc(l_bluestore_write_big_bytes
, length
);
9678 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9679 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9680 while (length
> 0) {
9681 bool new_blob
= false;
9682 uint32_t l
= MIN(max_bsize
, length
);
9686 //attempting to reuse existing blob
9687 if (!wctx
->compress
) {
9688 // look for an existing mutable blob we can reuse
9689 auto begin
= o
->extent_map
.extent_map
.begin();
9690 auto end
= o
->extent_map
.extent_map
.end();
9691 auto ep
= o
->extent_map
.seek_lextent(offset
);
9693 if (prev_ep
!= begin
) {
9696 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9698 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9699 // search suitable extent in both forward and reverse direction in
9700 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9701 // then check if blob can be reused via can_reuse_blob func.
9705 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9706 if (offset
>= ep
->blob_start() &&
9707 ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9708 offset
- ep
->blob_start(),
9711 b_off
= offset
- ep
->blob_start();
9712 prev_ep
= end
; // to avoid check below
9713 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9714 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9721 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9722 if (prev_ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9723 offset
- prev_ep
->blob_start(),
9726 b_off
= offset
- prev_ep
->blob_start();
9727 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9728 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9729 } else if (prev_ep
!= begin
) {
9733 prev_ep
= end
; // to avoid useless first extent re-check
9736 } while (b
== nullptr && any_change
);
9746 wctx
->write(offset
, b
, l
, b_off
, t
, b_off
, l
, false, new_blob
);
9749 logger
->inc(l_bluestore_write_big_blobs
);
9753 int BlueStore::_do_alloc_write(
9759 dout(20) << __func__
<< " txc " << txc
9760 << " " << wctx
->writes
.size() << " blobs"
9764 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9765 for (auto &wi
: wctx
->writes
) {
9766 need
+= wi
.blob_length
;
9768 int r
= alloc
->reserve(need
);
9770 derr
<< __func__
<< " failed to reserve 0x" << std::hex
<< need
<< std::dec
9778 if (wctx
->compress
) {
9780 "compression_algorithm",
9784 if (coll
->pool_opts
.get(pool_opts_t::COMPRESSION_ALGORITHM
, &val
)) {
9785 CompressorRef cp
= compressor
;
9786 if (!cp
|| cp
->get_type_name() != val
) {
9787 cp
= Compressor::create(cct
, val
);
9789 return boost::optional
<CompressorRef
>(cp
);
9791 return boost::optional
<CompressorRef
>();
9795 crr
= select_option(
9796 "compression_required_ratio",
9797 cct
->_conf
->bluestore_compression_required_ratio
,
9800 if(coll
->pool_opts
.get(pool_opts_t::COMPRESSION_REQUIRED_RATIO
, &val
)) {
9801 return boost::optional
<double>(val
);
9803 return boost::optional
<double>();
9809 int csum
= csum_type
.load();
9810 csum
= select_option(
9815 if(coll
->pool_opts
.get(pool_opts_t::CSUM_TYPE
, &val
)) {
9816 return boost::optional
<int>(val
);
9818 return boost::optional
<int>();
9822 for (auto& wi
: wctx
->writes
) {
9824 bluestore_blob_t
& dblob
= b
->dirty_blob();
9825 uint64_t b_off
= wi
.b_off
;
9826 bufferlist
*l
= &wi
.bl
;
9827 uint64_t final_length
= wi
.blob_length
;
9828 uint64_t csum_length
= wi
.blob_length
;
9829 unsigned csum_order
= block_size_order
;
9830 bufferlist compressed_bl
;
9831 bool compressed
= false;
9832 if(c
&& wi
.blob_length
> min_alloc_size
) {
9834 utime_t start
= ceph_clock_now();
9838 assert(wi
.blob_length
== l
->length());
9839 bluestore_compression_header_t chdr
;
9840 chdr
.type
= c
->get_type();
9841 // FIXME: memory alignment here is bad
9844 r
= c
->compress(*l
, t
);
9847 chdr
.length
= t
.length();
9848 ::encode(chdr
, compressed_bl
);
9849 compressed_bl
.claim_append(t
);
9850 uint64_t rawlen
= compressed_bl
.length();
9851 uint64_t newlen
= P2ROUNDUP(rawlen
, min_alloc_size
);
9852 uint64_t want_len_raw
= final_length
* crr
;
9853 uint64_t want_len
= P2ROUNDUP(want_len_raw
, min_alloc_size
);
9854 if (newlen
<= want_len
&& newlen
< final_length
) {
9855 // Cool. We compressed at least as much as we were hoping to.
9856 // pad out to min_alloc_size
9857 compressed_bl
.append_zero(newlen
- rawlen
);
9858 logger
->inc(l_bluestore_write_pad_bytes
, newlen
- rawlen
);
9859 dout(20) << __func__
<< std::hex
<< " compressed 0x" << wi
.blob_length
9860 << " -> 0x" << rawlen
<< " => 0x" << newlen
9861 << " with " << c
->get_type()
9862 << std::dec
<< dendl
;
9863 txc
->statfs_delta
.compressed() += rawlen
;
9864 txc
->statfs_delta
.compressed_original() += l
->length();
9865 txc
->statfs_delta
.compressed_allocated() += newlen
;
9867 final_length
= newlen
;
9868 csum_length
= newlen
;
9869 csum_order
= ctz(newlen
);
9870 dblob
.set_compressed(wi
.blob_length
, rawlen
);
9872 logger
->inc(l_bluestore_compress_success_count
);
9874 dout(20) << __func__
<< std::hex
<< " 0x" << l
->length()
9875 << " compressed to 0x" << rawlen
<< " -> 0x" << newlen
9876 << " with " << c
->get_type()
9877 << ", which is more than required 0x" << want_len_raw
9878 << " -> 0x" << want_len
9879 << ", leaving uncompressed"
9880 << std::dec
<< dendl
;
9881 logger
->inc(l_bluestore_compress_rejected_count
);
9883 logger
->tinc(l_bluestore_compress_lat
,
9884 ceph_clock_now() - start
);
9886 if (!compressed
&& wi
.new_blob
) {
9887 // initialize newly created blob only
9888 assert(dblob
.is_mutable());
9889 if (l
->length() != wi
.blob_length
) {
9890 // hrm, maybe we could do better here, but let's not bother.
9891 dout(20) << __func__
<< " forcing csum_order to block_size_order "
9892 << block_size_order
<< dendl
;
9893 csum_order
= block_size_order
;
9895 csum_order
= std::min(wctx
->csum_order
, ctz(l
->length()));
9897 // try to align blob with max_blob_size to improve
9898 // its reuse ratio, e.g. in case of reverse write
9899 uint32_t suggested_boff
=
9900 (wi
.logical_offset
- (wi
.b_off0
- wi
.b_off
)) % max_bsize
;
9901 if ((suggested_boff
% (1 << csum_order
)) == 0 &&
9902 suggested_boff
+ final_length
<= max_bsize
&&
9903 suggested_boff
> b_off
) {
9904 dout(20) << __func__
<< " forcing blob_offset to "
9905 << std::hex
<< suggested_boff
<< std::dec
<< dendl
;
9906 assert(suggested_boff
>= b_off
);
9907 csum_length
+= suggested_boff
- b_off
;
9908 b_off
= suggested_boff
;
9912 AllocExtentVector extents
;
9913 extents
.reserve(4); // 4 should be (more than) enough for most allocations
9914 int64_t got
= alloc
->allocate(final_length
, min_alloc_size
,
9915 max_alloc_size
.load(),
9917 assert(got
== (int64_t)final_length
);
9919 txc
->statfs_delta
.allocated() += got
;
9920 for (auto& p
: extents
) {
9921 bluestore_pextent_t e
= bluestore_pextent_t(p
);
9922 txc
->allocated
.insert(e
.offset
, e
.length
);
9925 dblob
.allocated(P2ALIGN(b_off
, min_alloc_size
), final_length
, extents
);
9927 dout(20) << __func__
<< " blob " << *b
9928 << " csum_type " << Checksummer::get_csum_type_string(csum
)
9929 << " csum_order " << csum_order
9930 << " csum_length 0x" << std::hex
<< csum_length
<< std::dec
9933 if (csum
!= Checksummer::CSUM_NONE
) {
9934 if (!dblob
.has_csum()) {
9935 dblob
.init_csum(csum
, csum_order
, csum_length
);
9937 dblob
.calc_csum(b_off
, *l
);
9939 if (wi
.mark_unused
) {
9940 auto b_end
= b_off
+ wi
.bl
.length();
9942 dblob
.add_unused(0, b_off
);
9944 if (b_end
< wi
.blob_length
) {
9945 dblob
.add_unused(b_end
, wi
.blob_length
- b_end
);
9949 Extent
*le
= o
->extent_map
.set_lextent(coll
, wi
.logical_offset
,
9950 b_off
+ (wi
.b_off0
- wi
.b_off
),
9954 wi
.b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9955 txc
->statfs_delta
.stored() += le
->length
;
9956 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9957 _buffer_cache_write(txc
, wi
.b
, b_off
, wi
.bl
,
9958 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9961 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
9962 if (l
->length() <= prefer_deferred_size
.load()) {
9963 dout(20) << __func__
<< " deferring small 0x" << std::hex
9964 << l
->length() << std::dec
<< " write via deferred" << dendl
;
9965 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9966 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9967 int r
= b
->get_blob().map(
9969 [&](uint64_t offset
, uint64_t length
) {
9970 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9976 b
->get_blob().map_bl(
9978 [&](uint64_t offset
, bufferlist
& t
) {
9979 bdev
->aio_write(offset
, t
, &txc
->ioc
, false);
9985 alloc
->unreserve(need
);
9990 void BlueStore::_wctx_finish(
9995 set
<SharedBlob
*> *maybe_unshared_blobs
)
9997 auto oep
= wctx
->old_extents
.begin();
9998 while (oep
!= wctx
->old_extents
.end()) {
10000 oep
= wctx
->old_extents
.erase(oep
);
10001 dout(20) << __func__
<< " lex_old " << lo
.e
<< dendl
;
10002 BlobRef b
= lo
.e
.blob
;
10003 const bluestore_blob_t
& blob
= b
->get_blob();
10004 if (blob
.is_compressed()) {
10005 if (lo
.blob_empty
) {
10006 txc
->statfs_delta
.compressed() -= blob
.get_compressed_payload_length();
10008 txc
->statfs_delta
.compressed_original() -= lo
.e
.length
;
10011 txc
->statfs_delta
.stored() -= lo
.e
.length
;
10013 dout(20) << __func__
<< " blob release " << r
<< dendl
;
10014 if (blob
.is_shared()) {
10015 PExtentVector final
;
10016 c
->load_shared_blob(b
->shared_blob
);
10018 b
->shared_blob
->put_ref(
10019 e
.offset
, e
.length
, &final
,
10020 b
->is_referenced() ? nullptr : maybe_unshared_blobs
);
10022 dout(20) << __func__
<< " shared_blob release " << final
10023 << " from " << *b
->shared_blob
<< dendl
;
10024 txc
->write_shared_blob(b
->shared_blob
);
10029 // we can't invalidate our logical extents as we drop them because
10030 // other lextents (either in our onode or others) may still
10031 // reference them. but we can throw out anything that is no
10032 // longer allocated. Note that this will leave behind edge bits
10033 // that are no longer referenced but not deallocated (until they
10034 // age out of the cache naturally).
10035 b
->discard_unallocated(c
.get());
10037 dout(20) << __func__
<< " release " << e
<< dendl
;
10038 txc
->released
.insert(e
.offset
, e
.length
);
10039 txc
->statfs_delta
.allocated() -= e
.length
;
10040 if (blob
.is_compressed()) {
10041 txc
->statfs_delta
.compressed_allocated() -= e
.length
;
10045 if (b
->is_spanning() && !b
->is_referenced()) {
10046 dout(20) << __func__
<< " spanning_blob_map removing empty " << *b
10048 o
->extent_map
.spanning_blob_map
.erase(b
->id
);
10053 void BlueStore::_do_write_data(
10060 WriteContext
*wctx
)
10062 uint64_t end
= offset
+ length
;
10063 bufferlist::iterator p
= bl
.begin();
10065 if (offset
/ min_alloc_size
== (end
- 1) / min_alloc_size
&&
10066 (length
!= min_alloc_size
)) {
10067 // we fall within the same block
10068 _do_write_small(txc
, c
, o
, offset
, length
, p
, wctx
);
10070 uint64_t head_offset
, head_length
;
10071 uint64_t middle_offset
, middle_length
;
10072 uint64_t tail_offset
, tail_length
;
10074 head_offset
= offset
;
10075 head_length
= P2NPHASE(offset
, min_alloc_size
);
10077 tail_offset
= P2ALIGN(end
, min_alloc_size
);
10078 tail_length
= P2PHASE(end
, min_alloc_size
);
10080 middle_offset
= head_offset
+ head_length
;
10081 middle_length
= length
- head_length
- tail_length
;
10084 _do_write_small(txc
, c
, o
, head_offset
, head_length
, p
, wctx
);
10087 if (middle_length
) {
10088 _do_write_big(txc
, c
, o
, middle_offset
, middle_length
, p
, wctx
);
10092 _do_write_small(txc
, c
, o
, tail_offset
, tail_length
, p
, wctx
);
10097 void BlueStore::_choose_write_options(
10100 uint32_t fadvise_flags
,
10101 WriteContext
*wctx
)
10103 if (fadvise_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
10104 dout(20) << __func__
<< " will do buffered write" << dendl
;
10105 wctx
->buffered
= true;
10106 } else if (cct
->_conf
->bluestore_default_buffered_write
&&
10107 (fadvise_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
10108 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
10109 dout(20) << __func__
<< " defaulting to buffered write" << dendl
;
10110 wctx
->buffered
= true;
10113 // apply basic csum block size
10114 wctx
->csum_order
= block_size_order
;
10116 // compression parameters
10117 unsigned alloc_hints
= o
->onode
.alloc_hint_flags
;
10118 auto cm
= select_option(
10119 "compression_mode",
10123 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MODE
, &val
)) {
10124 return boost::optional
<Compressor::CompressionMode
>(
10125 Compressor::get_comp_mode_type(val
));
10127 return boost::optional
<Compressor::CompressionMode
>();
10131 wctx
->compress
= (cm
!= Compressor::COMP_NONE
) &&
10132 ((cm
== Compressor::COMP_FORCE
) ||
10133 (cm
== Compressor::COMP_AGGRESSIVE
&&
10134 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE
) == 0) ||
10135 (cm
== Compressor::COMP_PASSIVE
&&
10136 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE
)));
10138 if ((alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_SEQUENTIAL_READ
) &&
10139 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_READ
) == 0 &&
10140 (alloc_hints
& (CEPH_OSD_ALLOC_HINT_FLAG_IMMUTABLE
|
10141 CEPH_OSD_ALLOC_HINT_FLAG_APPEND_ONLY
)) &&
10142 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_WRITE
) == 0) {
10144 dout(20) << __func__
<< " will prefer large blob and csum sizes" << dendl
;
10146 if (o
->onode
.expected_write_size
) {
10147 wctx
->csum_order
= std::max(min_alloc_size_order
,
10148 (uint8_t)ctz(o
->onode
.expected_write_size
));
10150 wctx
->csum_order
= min_alloc_size_order
;
10153 if (wctx
->compress
) {
10154 wctx
->target_blob_size
= select_option(
10155 "compression_max_blob_size",
10156 comp_max_blob_size
.load(),
10159 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MAX_BLOB_SIZE
, &val
)) {
10160 return boost::optional
<uint64_t>((uint64_t)val
);
10162 return boost::optional
<uint64_t>();
10167 if (wctx
->compress
) {
10168 wctx
->target_blob_size
= select_option(
10169 "compression_min_blob_size",
10170 comp_min_blob_size
.load(),
10173 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MIN_BLOB_SIZE
, &val
)) {
10174 return boost::optional
<uint64_t>((uint64_t)val
);
10176 return boost::optional
<uint64_t>();
10182 uint64_t max_bsize
= max_blob_size
.load();
10183 if (wctx
->target_blob_size
== 0 || wctx
->target_blob_size
> max_bsize
) {
10184 wctx
->target_blob_size
= max_bsize
;
10187 // set the min blob size floor at 2x the min_alloc_size, or else we
10188 // won't be able to allocate a smaller extent for the compressed
10190 if (wctx
->compress
&&
10191 wctx
->target_blob_size
< min_alloc_size
* 2) {
10192 wctx
->target_blob_size
= min_alloc_size
* 2;
10195 dout(20) << __func__
<< " prefer csum_order " << wctx
->csum_order
10196 << " target_blob_size 0x" << std::hex
<< wctx
->target_blob_size
10197 << std::dec
<< dendl
;
10200 int BlueStore::_do_gc(
10204 const GarbageCollector
& gc
,
10205 const WriteContext
& wctx
,
10206 uint64_t *dirty_start
,
10207 uint64_t *dirty_end
)
10209 auto& extents_to_collect
= gc
.get_extents_to_collect();
10211 WriteContext wctx_gc
;
10212 wctx_gc
.fork(wctx
); // make a clone for garbage collection
10214 for (auto it
= extents_to_collect
.begin();
10215 it
!= extents_to_collect
.end();
10218 int r
= _do_read(c
.get(), o
, it
->offset
, it
->length
, bl
, 0);
10219 assert(r
== (int)it
->length
);
10221 o
->extent_map
.fault_range(db
, it
->offset
, it
->length
);
10222 _do_write_data(txc
, c
, o
, it
->offset
, it
->length
, bl
, &wctx_gc
);
10223 logger
->inc(l_bluestore_gc_merged
, it
->length
);
10225 if (*dirty_start
> it
->offset
) {
10226 *dirty_start
= it
->offset
;
10229 if (*dirty_end
< it
->offset
+ it
->length
) {
10230 *dirty_end
= it
->offset
+ it
->length
;
10234 dout(30) << __func__
<< " alloc write" << dendl
;
10235 int r
= _do_alloc_write(txc
, c
, o
, &wctx_gc
);
10237 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10242 _wctx_finish(txc
, c
, o
, &wctx_gc
);
10246 int BlueStore::_do_write(
10253 uint32_t fadvise_flags
)
10257 dout(20) << __func__
10259 << " 0x" << std::hex
<< offset
<< "~" << length
10260 << " - have 0x" << o
->onode
.size
10261 << " (" << std::dec
<< o
->onode
.size
<< ")"
10263 << " fadvise_flags 0x" << std::hex
<< fadvise_flags
<< std::dec
10271 uint64_t end
= offset
+ length
;
10273 GarbageCollector
gc(c
->store
->cct
);
10275 auto dirty_start
= offset
;
10276 auto dirty_end
= end
;
10279 _choose_write_options(c
, o
, fadvise_flags
, &wctx
);
10280 o
->extent_map
.fault_range(db
, offset
, length
);
10281 _do_write_data(txc
, c
, o
, offset
, length
, bl
, &wctx
);
10282 r
= _do_alloc_write(txc
, c
, o
, &wctx
);
10284 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10289 // NB: _wctx_finish() will empty old_extents
10290 // so we must do gc estimation before that
10291 benefit
= gc
.estimate(offset
,
10297 _wctx_finish(txc
, c
, o
, &wctx
);
10298 if (end
> o
->onode
.size
) {
10299 dout(20) << __func__
<< " extending size to 0x" << std::hex
<< end
10300 << std::dec
<< dendl
;
10301 o
->onode
.size
= end
;
10304 if (benefit
>= g_conf
->bluestore_gc_enable_total_threshold
) {
10305 if (!gc
.get_extents_to_collect().empty()) {
10306 dout(20) << __func__
<< " perform garbage collection, "
10307 << "expected benefit = " << benefit
<< " AUs" << dendl
;
10308 r
= _do_gc(txc
, c
, o
, gc
, wctx
, &dirty_start
, &dirty_end
);
10310 derr
<< __func__
<< " _do_gc failed with " << cpp_strerror(r
)
10317 o
->extent_map
.compress_extent_map(dirty_start
, dirty_end
- dirty_start
);
10318 o
->extent_map
.dirty_range(dirty_start
, dirty_end
- dirty_start
);
10326 int BlueStore::_write(TransContext
*txc
,
10329 uint64_t offset
, size_t length
,
10331 uint32_t fadvise_flags
)
10333 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10334 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10337 if (offset
+ length
>= OBJECT_MAX_SIZE
) {
10340 _assign_nid(txc
, o
);
10341 r
= _do_write(txc
, c
, o
, offset
, length
, bl
, fadvise_flags
);
10342 txc
->write_onode(o
);
10344 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10345 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10346 << " = " << r
<< dendl
;
10350 int BlueStore::_zero(TransContext
*txc
,
10353 uint64_t offset
, size_t length
)
10355 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10356 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10359 if (offset
+ length
>= OBJECT_MAX_SIZE
) {
10362 _assign_nid(txc
, o
);
10363 r
= _do_zero(txc
, c
, o
, offset
, length
);
10365 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10366 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10367 << " = " << r
<< dendl
;
10371 int BlueStore::_do_zero(TransContext
*txc
,
10374 uint64_t offset
, size_t length
)
10376 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10377 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10384 o
->extent_map
.fault_range(db
, offset
, length
);
10385 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10386 o
->extent_map
.dirty_range(offset
, length
);
10387 _wctx_finish(txc
, c
, o
, &wctx
);
10389 if (offset
+ length
> o
->onode
.size
) {
10390 o
->onode
.size
= offset
+ length
;
10391 dout(20) << __func__
<< " extending size to " << offset
+ length
10394 txc
->write_onode(o
);
10396 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10397 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10398 << " = " << r
<< dendl
;
10402 void BlueStore::_do_truncate(
10403 TransContext
*txc
, CollectionRef
& c
, OnodeRef o
, uint64_t offset
,
10404 set
<SharedBlob
*> *maybe_unshared_blobs
)
10406 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10407 << " 0x" << std::hex
<< offset
<< std::dec
<< dendl
;
10409 _dump_onode(o
, 30);
10411 if (offset
== o
->onode
.size
)
10414 if (offset
< o
->onode
.size
) {
10416 uint64_t length
= o
->onode
.size
- offset
;
10417 o
->extent_map
.fault_range(db
, offset
, length
);
10418 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10419 o
->extent_map
.dirty_range(offset
, length
);
10420 _wctx_finish(txc
, c
, o
, &wctx
, maybe_unshared_blobs
);
10422 // if we have shards past EOF, ask for a reshard
10423 if (!o
->onode
.extent_map_shards
.empty() &&
10424 o
->onode
.extent_map_shards
.back().offset
>= offset
) {
10425 dout(10) << __func__
<< " request reshard past EOF" << dendl
;
10427 o
->extent_map
.request_reshard(offset
- 1, offset
+ length
);
10429 o
->extent_map
.request_reshard(0, length
);
10434 o
->onode
.size
= offset
;
10436 txc
->write_onode(o
);
10439 int BlueStore::_truncate(TransContext
*txc
,
10444 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10445 << " 0x" << std::hex
<< offset
<< std::dec
10448 if (offset
>= OBJECT_MAX_SIZE
) {
10451 _do_truncate(txc
, c
, o
, offset
);
10453 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10454 << " 0x" << std::hex
<< offset
<< std::dec
10455 << " = " << r
<< dendl
;
10459 int BlueStore::_do_remove(
10464 set
<SharedBlob
*> maybe_unshared_blobs
;
10465 bool is_gen
= !o
->oid
.is_no_gen();
10466 _do_truncate(txc
, c
, o
, 0, is_gen
? &maybe_unshared_blobs
: nullptr);
10467 if (o
->onode
.has_omap()) {
10469 _do_omap_clear(txc
, o
->onode
.nid
);
10473 for (auto &s
: o
->extent_map
.shards
) {
10474 dout(20) << __func__
<< " removing shard 0x" << std::hex
10475 << s
.shard_info
->offset
<< std::dec
<< dendl
;
10476 generate_extent_shard_key_and_apply(o
->key
, s
.shard_info
->offset
, &key
,
10477 [&](const string
& final_key
) {
10478 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
10482 txc
->t
->rmkey(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size());
10484 o
->extent_map
.clear();
10485 o
->onode
= bluestore_onode_t();
10486 _debug_obj_on_delete(o
->oid
);
10488 if (!is_gen
|| maybe_unshared_blobs
.empty()) {
10492 // see if we can unshare blobs still referenced by the head
10493 dout(10) << __func__
<< " gen and maybe_unshared_blobs "
10494 << maybe_unshared_blobs
<< dendl
;
10495 ghobject_t nogen
= o
->oid
;
10496 nogen
.generation
= ghobject_t::NO_GEN
;
10497 OnodeRef h
= c
->onode_map
.lookup(nogen
);
10499 if (!h
|| !h
->exists
) {
10503 dout(20) << __func__
<< " checking for unshareable blobs on " << h
10504 << " " << h
->oid
<< dendl
;
10505 map
<SharedBlob
*,bluestore_extent_ref_map_t
> expect
;
10506 for (auto& e
: h
->extent_map
.extent_map
) {
10507 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10508 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10509 if (b
.is_shared() &&
10511 maybe_unshared_blobs
.count(sb
)) {
10512 b
.map(e
.blob_offset
, e
.length
, [&](uint64_t off
, uint64_t len
) {
10513 expect
[sb
].get(off
, len
);
10519 vector
<SharedBlob
*> unshared_blobs
;
10520 unshared_blobs
.reserve(maybe_unshared_blobs
.size());
10521 for (auto& p
: expect
) {
10522 dout(20) << " ? " << *p
.first
<< " vs " << p
.second
<< dendl
;
10523 if (p
.first
->persistent
->ref_map
== p
.second
) {
10524 SharedBlob
*sb
= p
.first
;
10525 dout(20) << __func__
<< " unsharing " << *sb
<< dendl
;
10526 unshared_blobs
.push_back(sb
);
10527 txc
->unshare_blob(sb
);
10528 uint64_t sbid
= c
->make_blob_unshared(sb
);
10530 get_shared_blob_key(sbid
, &key
);
10531 txc
->t
->rmkey(PREFIX_SHARED_BLOB
, key
);
10535 if (unshared_blobs
.empty()) {
10539 for (auto& e
: h
->extent_map
.extent_map
) {
10540 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10541 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10542 if (b
.is_shared() &&
10543 std::find(unshared_blobs
.begin(), unshared_blobs
.end(),
10544 sb
) != unshared_blobs
.end()) {
10545 dout(20) << __func__
<< " unsharing " << e
<< dendl
;
10546 bluestore_blob_t
& blob
= e
.blob
->dirty_blob();
10547 blob
.clear_flag(bluestore_blob_t::FLAG_SHARED
);
10548 h
->extent_map
.dirty_range(e
.logical_offset
, 1);
10551 txc
->write_onode(h
);
10556 int BlueStore::_remove(TransContext
*txc
,
10560 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10561 int r
= _do_remove(txc
, c
, o
);
10562 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10566 int BlueStore::_setattr(TransContext
*txc
,
10569 const string
& name
,
10572 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10573 << " " << name
<< " (" << val
.length() << " bytes)"
10576 if (val
.is_partial())
10577 o
->onode
.attrs
[name
.c_str()] = bufferptr(val
.c_str(), val
.length());
10579 o
->onode
.attrs
[name
.c_str()] = val
;
10580 txc
->write_onode(o
);
10581 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10582 << " " << name
<< " (" << val
.length() << " bytes)"
10583 << " = " << r
<< dendl
;
10587 int BlueStore::_setattrs(TransContext
*txc
,
10590 const map
<string
,bufferptr
>& aset
)
10592 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10593 << " " << aset
.size() << " keys"
10596 for (map
<string
,bufferptr
>::const_iterator p
= aset
.begin();
10597 p
!= aset
.end(); ++p
) {
10598 if (p
->second
.is_partial())
10599 o
->onode
.attrs
[p
->first
.c_str()] =
10600 bufferptr(p
->second
.c_str(), p
->second
.length());
10602 o
->onode
.attrs
[p
->first
.c_str()] = p
->second
;
10604 txc
->write_onode(o
);
10605 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10606 << " " << aset
.size() << " keys"
10607 << " = " << r
<< dendl
;
10612 int BlueStore::_rmattr(TransContext
*txc
,
10615 const string
& name
)
10617 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10618 << " " << name
<< dendl
;
10620 auto it
= o
->onode
.attrs
.find(name
.c_str());
10621 if (it
== o
->onode
.attrs
.end())
10624 o
->onode
.attrs
.erase(it
);
10625 txc
->write_onode(o
);
10628 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10629 << " " << name
<< " = " << r
<< dendl
;
10633 int BlueStore::_rmattrs(TransContext
*txc
,
10637 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10640 if (o
->onode
.attrs
.empty())
10643 o
->onode
.attrs
.clear();
10644 txc
->write_onode(o
);
10647 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10651 void BlueStore::_do_omap_clear(TransContext
*txc
, uint64_t id
)
10653 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
10654 string prefix
, tail
;
10655 get_omap_header(id
, &prefix
);
10656 get_omap_tail(id
, &tail
);
10657 it
->lower_bound(prefix
);
10658 while (it
->valid()) {
10659 if (it
->key() >= tail
) {
10660 dout(30) << __func__
<< " stop at " << pretty_binary_string(tail
)
10664 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
10665 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
10670 int BlueStore::_omap_clear(TransContext
*txc
,
10674 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10676 if (o
->onode
.has_omap()) {
10678 _do_omap_clear(txc
, o
->onode
.nid
);
10679 o
->onode
.clear_omap_flag();
10680 txc
->write_onode(o
);
10682 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10686 int BlueStore::_omap_setkeys(TransContext
*txc
,
10691 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10693 bufferlist::iterator p
= bl
.begin();
10695 if (!o
->onode
.has_omap()) {
10696 o
->onode
.set_omap_flag();
10697 txc
->write_onode(o
);
10699 txc
->note_modified_object(o
);
10702 _key_encode_u64(o
->onode
.nid
, &final_key
);
10703 final_key
.push_back('.');
10709 ::decode(value
, p
);
10710 final_key
.resize(9); // keep prefix
10712 dout(30) << __func__
<< " " << pretty_binary_string(final_key
)
10713 << " <- " << key
<< dendl
;
10714 txc
->t
->set(PREFIX_OMAP
, final_key
, value
);
10717 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10721 int BlueStore::_omap_setheader(TransContext
*txc
,
10726 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10729 if (!o
->onode
.has_omap()) {
10730 o
->onode
.set_omap_flag();
10731 txc
->write_onode(o
);
10733 txc
->note_modified_object(o
);
10735 get_omap_header(o
->onode
.nid
, &key
);
10736 txc
->t
->set(PREFIX_OMAP
, key
, bl
);
10738 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10742 int BlueStore::_omap_rmkeys(TransContext
*txc
,
10747 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10749 bufferlist::iterator p
= bl
.begin();
10753 if (!o
->onode
.has_omap()) {
10756 _key_encode_u64(o
->onode
.nid
, &final_key
);
10757 final_key
.push_back('.');
10762 final_key
.resize(9); // keep prefix
10764 dout(30) << __func__
<< " rm " << pretty_binary_string(final_key
)
10765 << " <- " << key
<< dendl
;
10766 txc
->t
->rmkey(PREFIX_OMAP
, final_key
);
10768 txc
->note_modified_object(o
);
10771 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10775 int BlueStore::_omap_rmkey_range(TransContext
*txc
,
10778 const string
& first
, const string
& last
)
10780 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10781 KeyValueDB::Iterator it
;
10782 string key_first
, key_last
;
10784 if (!o
->onode
.has_omap()) {
10788 it
= db
->get_iterator(PREFIX_OMAP
);
10789 get_omap_key(o
->onode
.nid
, first
, &key_first
);
10790 get_omap_key(o
->onode
.nid
, last
, &key_last
);
10791 it
->lower_bound(key_first
);
10792 while (it
->valid()) {
10793 if (it
->key() >= key_last
) {
10794 dout(30) << __func__
<< " stop at " << pretty_binary_string(key_last
)
10798 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
10799 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
10802 txc
->note_modified_object(o
);
10805 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10809 int BlueStore::_set_alloc_hint(
10813 uint64_t expected_object_size
,
10814 uint64_t expected_write_size
,
10817 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10818 << " object_size " << expected_object_size
10819 << " write_size " << expected_write_size
10820 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
10823 o
->onode
.expected_object_size
= expected_object_size
;
10824 o
->onode
.expected_write_size
= expected_write_size
;
10825 o
->onode
.alloc_hint_flags
= flags
;
10826 txc
->write_onode(o
);
10827 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10828 << " object_size " << expected_object_size
10829 << " write_size " << expected_write_size
10830 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
10831 << " = " << r
<< dendl
;
10835 int BlueStore::_clone(TransContext
*txc
,
10840 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
10841 << newo
->oid
<< dendl
;
10843 if (oldo
->oid
.hobj
.get_hash() != newo
->oid
.hobj
.get_hash()) {
10844 derr
<< __func__
<< " mismatched hash on " << oldo
->oid
10845 << " and " << newo
->oid
<< dendl
;
10849 _assign_nid(txc
, newo
);
10853 _do_truncate(txc
, c
, newo
, 0);
10854 if (cct
->_conf
->bluestore_clone_cow
) {
10855 _do_clone_range(txc
, c
, oldo
, newo
, 0, oldo
->onode
.size
, 0);
10858 r
= _do_read(c
.get(), oldo
, 0, oldo
->onode
.size
, bl
, 0);
10861 r
= _do_write(txc
, c
, newo
, 0, oldo
->onode
.size
, bl
, 0);
10867 newo
->onode
.attrs
= oldo
->onode
.attrs
;
10870 if (newo
->onode
.has_omap()) {
10871 dout(20) << __func__
<< " clearing old omap data" << dendl
;
10873 _do_omap_clear(txc
, newo
->onode
.nid
);
10875 if (oldo
->onode
.has_omap()) {
10876 dout(20) << __func__
<< " copying omap data" << dendl
;
10877 if (!newo
->onode
.has_omap()) {
10878 newo
->onode
.set_omap_flag();
10880 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
10882 get_omap_header(oldo
->onode
.nid
, &head
);
10883 get_omap_tail(oldo
->onode
.nid
, &tail
);
10884 it
->lower_bound(head
);
10885 while (it
->valid()) {
10886 if (it
->key() >= tail
) {
10887 dout(30) << __func__
<< " reached tail" << dendl
;
10890 dout(30) << __func__
<< " got header/data "
10891 << pretty_binary_string(it
->key()) << dendl
;
10893 rewrite_omap_key(newo
->onode
.nid
, it
->key(), &key
);
10894 txc
->t
->set(PREFIX_OMAP
, key
, it
->value());
10899 newo
->onode
.clear_omap_flag();
10902 txc
->write_onode(newo
);
10906 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
10907 << newo
->oid
<< " = " << r
<< dendl
;
10911 int BlueStore::_do_clone_range(
10920 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
10922 << " 0x" << std::hex
<< srcoff
<< "~" << length
<< " -> "
10923 << " 0x" << dstoff
<< "~" << length
<< std::dec
<< dendl
;
10924 oldo
->extent_map
.fault_range(db
, srcoff
, length
);
10925 newo
->extent_map
.fault_range(db
, dstoff
, length
);
10929 // hmm, this could go into an ExtentMap::dup() method.
10930 vector
<BlobRef
> id_to_blob(oldo
->extent_map
.extent_map
.size());
10931 for (auto &e
: oldo
->extent_map
.extent_map
) {
10932 e
.blob
->last_encoded_id
= -1;
10935 uint64_t end
= srcoff
+ length
;
10936 uint32_t dirty_range_begin
= 0;
10937 uint32_t dirty_range_end
= 0;
10938 bool src_dirty
= false;
10939 for (auto ep
= oldo
->extent_map
.seek_lextent(srcoff
);
10940 ep
!= oldo
->extent_map
.extent_map
.end();
10943 if (e
.logical_offset
>= end
) {
10946 dout(20) << __func__
<< " src " << e
<< dendl
;
10948 bool blob_duped
= true;
10949 if (e
.blob
->last_encoded_id
>= 0) {
10950 // blob is already duped
10951 cb
= id_to_blob
[e
.blob
->last_encoded_id
];
10952 blob_duped
= false;
10955 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
10956 // make sure it is shared
10957 if (!blob
.is_shared()) {
10958 c
->make_blob_shared(_assign_blobid(txc
), e
.blob
);
10961 dirty_range_begin
= e
.logical_offset
;
10963 assert(e
.logical_end() > 0);
10964 // -1 to exclude next potential shard
10965 dirty_range_end
= e
.logical_end() - 1;
10967 c
->load_shared_blob(e
.blob
->shared_blob
);
10970 e
.blob
->last_encoded_id
= n
;
10971 id_to_blob
[n
] = cb
;
10973 // bump the extent refs on the copied blob's extents
10974 for (auto p
: blob
.get_extents()) {
10975 if (p
.is_valid()) {
10976 e
.blob
->shared_blob
->get_ref(p
.offset
, p
.length
);
10979 txc
->write_shared_blob(e
.blob
->shared_blob
);
10980 dout(20) << __func__
<< " new " << *cb
<< dendl
;
10983 int skip_front
, skip_back
;
10984 if (e
.logical_offset
< srcoff
) {
10985 skip_front
= srcoff
- e
.logical_offset
;
10989 if (e
.logical_end() > end
) {
10990 skip_back
= e
.logical_end() - end
;
10994 Extent
*ne
= new Extent(e
.logical_offset
+ skip_front
+ dstoff
- srcoff
,
10995 e
.blob_offset
+ skip_front
,
10996 e
.length
- skip_front
- skip_back
, cb
);
10997 newo
->extent_map
.extent_map
.insert(*ne
);
10998 ne
->blob
->get_ref(c
.get(), ne
->blob_offset
, ne
->length
);
10999 // fixme: we may leave parts of new blob unreferenced that could
11000 // be freed (relative to the shared_blob).
11001 txc
->statfs_delta
.stored() += ne
->length
;
11002 if (e
.blob
->get_blob().is_compressed()) {
11003 txc
->statfs_delta
.compressed_original() += ne
->length
;
11005 txc
->statfs_delta
.compressed() +=
11006 cb
->get_blob().get_compressed_payload_length();
11009 dout(20) << __func__
<< " dst " << *ne
<< dendl
;
11013 oldo
->extent_map
.dirty_range(dirty_range_begin
,
11014 dirty_range_end
- dirty_range_begin
);
11015 txc
->write_onode(oldo
);
11017 txc
->write_onode(newo
);
11019 if (dstoff
+ length
> newo
->onode
.size
) {
11020 newo
->onode
.size
= dstoff
+ length
;
11022 newo
->extent_map
.dirty_range(dstoff
, length
);
11028 int BlueStore::_clone_range(TransContext
*txc
,
11032 uint64_t srcoff
, uint64_t length
, uint64_t dstoff
)
11034 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11035 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11036 << " to offset 0x" << dstoff
<< std::dec
<< dendl
;
11039 if (srcoff
+ length
>= OBJECT_MAX_SIZE
||
11040 dstoff
+ length
>= OBJECT_MAX_SIZE
) {
11044 if (srcoff
+ length
> oldo
->onode
.size
) {
11049 _assign_nid(txc
, newo
);
11052 if (cct
->_conf
->bluestore_clone_cow
) {
11053 _do_zero(txc
, c
, newo
, dstoff
, length
);
11054 _do_clone_range(txc
, c
, oldo
, newo
, srcoff
, length
, dstoff
);
11057 r
= _do_read(c
.get(), oldo
, srcoff
, length
, bl
, 0);
11060 r
= _do_write(txc
, c
, newo
, dstoff
, bl
.length(), bl
, 0);
11066 txc
->write_onode(newo
);
11070 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11071 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11072 << " to offset 0x" << dstoff
<< std::dec
11073 << " = " << r
<< dendl
;
11077 int BlueStore::_rename(TransContext
*txc
,
11081 const ghobject_t
& new_oid
)
11083 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11084 << new_oid
<< dendl
;
11086 ghobject_t old_oid
= oldo
->oid
;
11087 mempool::bluestore_cache_other::string new_okey
;
11090 if (newo
->exists
) {
11094 assert(txc
->onodes
.count(newo
) == 0);
11097 txc
->t
->rmkey(PREFIX_OBJ
, oldo
->key
.c_str(), oldo
->key
.size());
11101 oldo
->extent_map
.fault_range(db
, 0, oldo
->onode
.size
);
11102 get_object_key(cct
, new_oid
, &new_okey
);
11104 for (auto &s
: oldo
->extent_map
.shards
) {
11105 generate_extent_shard_key_and_apply(oldo
->key
, s
.shard_info
->offset
, &key
,
11106 [&](const string
& final_key
) {
11107 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
11115 txc
->write_onode(newo
);
11117 // this adjusts oldo->{oid,key}, and reset oldo to a fresh empty
11118 // Onode in the old slot
11119 c
->onode_map
.rename(oldo
, old_oid
, new_oid
, new_okey
);
11123 dout(10) << __func__
<< " " << c
->cid
<< " " << old_oid
<< " -> "
11124 << new_oid
<< " = " << r
<< dendl
;
11130 int BlueStore::_create_collection(
11136 dout(15) << __func__
<< " " << cid
<< " bits " << bits
<< dendl
;
11141 RWLock::WLocker
l(coll_lock
);
11149 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
11151 (*c
)->cnode
.bits
= bits
;
11152 coll_map
[cid
] = *c
;
11154 ::encode((*c
)->cnode
, bl
);
11155 txc
->t
->set(PREFIX_COLL
, stringify(cid
), bl
);
11159 dout(10) << __func__
<< " " << cid
<< " bits " << bits
<< " = " << r
<< dendl
;
11163 int BlueStore::_remove_collection(TransContext
*txc
, const coll_t
&cid
,
11166 dout(15) << __func__
<< " " << cid
<< dendl
;
11170 RWLock::WLocker
l(coll_lock
);
11175 size_t nonexistent_count
= 0;
11176 assert((*c
)->exists
);
11177 if ((*c
)->onode_map
.map_any([&](OnodeRef o
) {
11179 dout(10) << __func__
<< " " << o
->oid
<< " " << o
11180 << " exists in onode_map" << dendl
;
11183 ++nonexistent_count
;
11190 vector
<ghobject_t
> ls
;
11192 // Enumerate onodes in db, up to nonexistent_count + 1
11193 // then check if all of them are marked as non-existent.
11194 // Bypass the check if returned number is greater than nonexistent_count
11195 r
= _collection_list(c
->get(), ghobject_t(), ghobject_t::get_max(),
11196 nonexistent_count
+ 1, &ls
, &next
);
11198 bool exists
= false; //ls.size() > nonexistent_count;
11199 for (auto it
= ls
.begin(); !exists
&& it
< ls
.end(); ++it
) {
11200 dout(10) << __func__
<< " oid " << *it
<< dendl
;
11201 auto onode
= (*c
)->onode_map
.lookup(*it
);
11202 exists
= !onode
|| onode
->exists
;
11204 dout(10) << __func__
<< " " << *it
11205 << " exists in db" << dendl
;
11209 coll_map
.erase(cid
);
11210 txc
->removed_collections
.push_back(*c
);
11211 (*c
)->exists
= false;
11213 txc
->t
->rmkey(PREFIX_COLL
, stringify(cid
));
11216 dout(10) << __func__
<< " " << cid
11217 << " is non-empty" << dendl
;
11224 dout(10) << __func__
<< " " << cid
<< " = " << r
<< dendl
;
11228 int BlueStore::_split_collection(TransContext
*txc
,
11231 unsigned bits
, int rem
)
11233 dout(15) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11234 << " bits " << bits
<< dendl
;
11235 RWLock::WLocker
l(c
->lock
);
11236 RWLock::WLocker
l2(d
->lock
);
11239 // flush all previous deferred writes on this sequencer. this is a bit
11240 // heavyweight, but we need to make sure all deferred writes complete
11241 // before we split as the new collection's sequencer may need to order
11242 // this after those writes, and we don't bother with the complexity of
11243 // moving those TransContexts over to the new osr.
11244 _osr_drain_preceding(txc
);
11246 // move any cached items (onodes and referenced shared blobs) that will
11247 // belong to the child collection post-split. leave everything else behind.
11248 // this may include things that don't strictly belong to the now-smaller
11249 // parent split, but the OSD will always send us a split for every new
11252 spg_t pgid
, dest_pgid
;
11253 bool is_pg
= c
->cid
.is_pg(&pgid
);
11255 is_pg
= d
->cid
.is_pg(&dest_pgid
);
11258 // the destination should initially be empty.
11259 assert(d
->onode_map
.empty());
11260 assert(d
->shared_blob_set
.empty());
11261 assert(d
->cnode
.bits
== bits
);
11263 c
->split_cache(d
.get());
11265 // adjust bits. note that this will be redundant for all but the first
11266 // split call for this parent (first child).
11267 c
->cnode
.bits
= bits
;
11268 assert(d
->cnode
.bits
== bits
);
11272 ::encode(c
->cnode
, bl
);
11273 txc
->t
->set(PREFIX_COLL
, stringify(c
->cid
), bl
);
11275 dout(10) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11276 << " bits " << bits
<< " = " << r
<< dendl
;
11280 // DB key value Histogram
11281 #define KEY_SLAB 32
11282 #define VALUE_SLAB 64
11284 const string prefix_onode
= "o";
11285 const string prefix_onode_shard
= "x";
11286 const string prefix_other
= "Z";
11288 int BlueStore::DBHistogram::get_key_slab(size_t sz
)
11290 return (sz
/KEY_SLAB
);
11293 string
BlueStore::DBHistogram::get_key_slab_to_range(int slab
)
11295 int lower_bound
= slab
* KEY_SLAB
;
11296 int upper_bound
= (slab
+ 1) * KEY_SLAB
;
11297 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11301 int BlueStore::DBHistogram::get_value_slab(size_t sz
)
11303 return (sz
/VALUE_SLAB
);
11306 string
BlueStore::DBHistogram::get_value_slab_to_range(int slab
)
11308 int lower_bound
= slab
* VALUE_SLAB
;
11309 int upper_bound
= (slab
+ 1) * VALUE_SLAB
;
11310 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11314 void BlueStore::DBHistogram::update_hist_entry(map
<string
, map
<int, struct key_dist
> > &key_hist
,
11315 const string
&prefix
, size_t key_size
, size_t value_size
)
11317 uint32_t key_slab
= get_key_slab(key_size
);
11318 uint32_t value_slab
= get_value_slab(value_size
);
11319 key_hist
[prefix
][key_slab
].count
++;
11320 key_hist
[prefix
][key_slab
].max_len
= MAX(key_size
, key_hist
[prefix
][key_slab
].max_len
);
11321 key_hist
[prefix
][key_slab
].val_map
[value_slab
].count
++;
11322 key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
=
11323 MAX(value_size
, key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
);
11326 void BlueStore::DBHistogram::dump(Formatter
*f
)
11328 f
->open_object_section("rocksdb_value_distribution");
11329 for (auto i
: value_hist
) {
11330 f
->dump_unsigned(get_value_slab_to_range(i
.first
).data(), i
.second
);
11332 f
->close_section();
11334 f
->open_object_section("rocksdb_key_value_histogram");
11335 for (auto i
: key_hist
) {
11336 f
->dump_string("prefix", i
.first
);
11337 f
->open_object_section("key_hist");
11338 for ( auto k
: i
.second
) {
11339 f
->dump_unsigned(get_key_slab_to_range(k
.first
).data(), k
.second
.count
);
11340 f
->dump_unsigned("max_len", k
.second
.max_len
);
11341 f
->open_object_section("value_hist");
11342 for ( auto j
: k
.second
.val_map
) {
11343 f
->dump_unsigned(get_value_slab_to_range(j
.first
).data(), j
.second
.count
);
11344 f
->dump_unsigned("max_len", j
.second
.max_len
);
11346 f
->close_section();
11348 f
->close_section();
11350 f
->close_section();
11353 //Itrerates through the db and collects the stats
11354 void BlueStore::generate_db_histogram(Formatter
*f
)
11357 uint64_t num_onodes
= 0;
11358 uint64_t num_shards
= 0;
11359 uint64_t num_super
= 0;
11360 uint64_t num_coll
= 0;
11361 uint64_t num_omap
= 0;
11362 uint64_t num_deferred
= 0;
11363 uint64_t num_alloc
= 0;
11364 uint64_t num_stat
= 0;
11365 uint64_t num_others
= 0;
11366 uint64_t num_shared_shards
= 0;
11367 size_t max_key_size
=0, max_value_size
= 0;
11368 uint64_t total_key_size
= 0, total_value_size
= 0;
11369 size_t key_size
= 0, value_size
= 0;
11372 utime_t start
= ceph_clock_now();
11374 KeyValueDB::WholeSpaceIterator iter
= db
->get_iterator();
11375 iter
->seek_to_first();
11376 while (iter
->valid()) {
11377 dout(30) << __func__
<< " Key: " << iter
->key() << dendl
;
11378 key_size
= iter
->key_size();
11379 value_size
= iter
->value_size();
11380 hist
.value_hist
[hist
.get_value_slab(value_size
)]++;
11381 max_key_size
= MAX(max_key_size
, key_size
);
11382 max_value_size
= MAX(max_value_size
, value_size
);
11383 total_key_size
+= key_size
;
11384 total_value_size
+= value_size
;
11386 pair
<string
,string
> key(iter
->raw_key());
11388 if (key
.first
== PREFIX_SUPER
) {
11389 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SUPER
, key_size
, value_size
);
11391 } else if (key
.first
== PREFIX_STAT
) {
11392 hist
.update_hist_entry(hist
.key_hist
, PREFIX_STAT
, key_size
, value_size
);
11394 } else if (key
.first
== PREFIX_COLL
) {
11395 hist
.update_hist_entry(hist
.key_hist
, PREFIX_COLL
, key_size
, value_size
);
11397 } else if (key
.first
== PREFIX_OBJ
) {
11398 if (key
.second
.back() == ONODE_KEY_SUFFIX
) {
11399 hist
.update_hist_entry(hist
.key_hist
, prefix_onode
, key_size
, value_size
);
11402 hist
.update_hist_entry(hist
.key_hist
, prefix_onode_shard
, key_size
, value_size
);
11405 } else if (key
.first
== PREFIX_OMAP
) {
11406 hist
.update_hist_entry(hist
.key_hist
, PREFIX_OMAP
, key_size
, value_size
);
11408 } else if (key
.first
== PREFIX_DEFERRED
) {
11409 hist
.update_hist_entry(hist
.key_hist
, PREFIX_DEFERRED
, key_size
, value_size
);
11411 } else if (key
.first
== PREFIX_ALLOC
|| key
.first
== "b" ) {
11412 hist
.update_hist_entry(hist
.key_hist
, PREFIX_ALLOC
, key_size
, value_size
);
11414 } else if (key
.first
== PREFIX_SHARED_BLOB
) {
11415 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SHARED_BLOB
, key_size
, value_size
);
11416 num_shared_shards
++;
11418 hist
.update_hist_entry(hist
.key_hist
, prefix_other
, key_size
, value_size
);
11424 utime_t duration
= ceph_clock_now() - start
;
11425 f
->open_object_section("rocksdb_key_value_stats");
11426 f
->dump_unsigned("num_onodes", num_onodes
);
11427 f
->dump_unsigned("num_shards", num_shards
);
11428 f
->dump_unsigned("num_super", num_super
);
11429 f
->dump_unsigned("num_coll", num_coll
);
11430 f
->dump_unsigned("num_omap", num_omap
);
11431 f
->dump_unsigned("num_deferred", num_deferred
);
11432 f
->dump_unsigned("num_alloc", num_alloc
);
11433 f
->dump_unsigned("num_stat", num_stat
);
11434 f
->dump_unsigned("num_shared_shards", num_shared_shards
);
11435 f
->dump_unsigned("num_others", num_others
);
11436 f
->dump_unsigned("max_key_size", max_key_size
);
11437 f
->dump_unsigned("max_value_size", max_value_size
);
11438 f
->dump_unsigned("total_key_size", total_key_size
);
11439 f
->dump_unsigned("total_value_size", total_value_size
);
11440 f
->close_section();
11444 dout(20) << __func__
<< " finished in " << duration
<< " seconds" << dendl
;
11448 void BlueStore::_flush_cache()
11450 dout(10) << __func__
<< dendl
;
11451 for (auto i
: cache_shards
) {
11453 assert(i
->empty());
11455 for (auto& p
: coll_map
) {
11456 assert(p
.second
->onode_map
.empty());
11457 assert(p
.second
->shared_blob_set
.empty());
11462 // For external caller.
11463 // We use a best-effort policy instead, e.g.,
11464 // we don't care if there are still some pinned onodes/data in the cache
11465 // after this command is completed.
11466 void BlueStore::flush_cache()
11468 dout(10) << __func__
<< dendl
;
11469 for (auto i
: cache_shards
) {
11474 void BlueStore::_apply_padding(uint64_t head_pad
,
11476 bufferlist
& padded
)
11479 padded
.prepend_zero(head_pad
);
11482 padded
.append_zero(tail_pad
);
11484 if (head_pad
|| tail_pad
) {
11485 dout(20) << __func__
<< " can pad head 0x" << std::hex
<< head_pad
11486 << " tail 0x" << tail_pad
<< std::dec
<< dendl
;
11487 logger
->inc(l_bluestore_write_pad_bytes
, head_pad
+ tail_pad
);
11491 // ===========================================