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 << " " << *b
.shared_blob
1712 void BlueStore::Blob::discard_unallocated(Collection
*coll
)
1714 if (get_blob().is_shared()) {
1717 if (get_blob().is_compressed()) {
1718 bool discard
= false;
1719 bool all_invalid
= true;
1720 for (auto e
: get_blob().get_extents()) {
1721 if (!e
.is_valid()) {
1724 all_invalid
= false;
1727 assert(discard
== all_invalid
); // in case of compressed blob all
1728 // or none pextents are invalid.
1730 shared_blob
->bc
.discard(shared_blob
->get_cache(), 0,
1731 get_blob().get_logical_length());
1735 for (auto e
: get_blob().get_extents()) {
1736 if (!e
.is_valid()) {
1737 ldout(coll
->store
->cct
, 20) << __func__
<< " 0x" << std::hex
<< pos
1739 << std::dec
<< dendl
;
1740 shared_blob
->bc
.discard(shared_blob
->get_cache(), pos
, e
.length
);
1744 if (get_blob().can_prune_tail()) {
1745 dirty_blob().prune_tail();
1746 used_in_blob
.prune_tail(get_blob().get_ondisk_length());
1747 auto cct
= coll
->store
->cct
; //used by dout
1748 dout(20) << __func__
<< " pruned tail, now " << get_blob() << dendl
;
1753 void BlueStore::Blob::get_ref(
1758 // Caller has to initialize Blob's logical length prior to increment
1759 // references. Otherwise one is neither unable to determine required
1760 // amount of counters in case of per-au tracking nor obtain min_release_size
1761 // for single counter mode.
1762 assert(get_blob().get_logical_length() != 0);
1763 auto cct
= coll
->store
->cct
;
1764 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1765 << std::dec
<< " " << *this << dendl
;
1767 if (used_in_blob
.is_empty()) {
1768 uint32_t min_release_size
=
1769 get_blob().get_release_size(coll
->store
->min_alloc_size
);
1770 uint64_t l
= get_blob().get_logical_length();
1771 dout(20) << __func__
<< " init 0x" << std::hex
<< l
<< ", "
1772 << min_release_size
<< std::dec
<< dendl
;
1773 used_in_blob
.init(l
, min_release_size
);
1780 bool BlueStore::Blob::put_ref(
1786 PExtentVector logical
;
1788 auto cct
= coll
->store
->cct
;
1789 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
1790 << std::dec
<< " " << *this << dendl
;
1792 bool empty
= used_in_blob
.put(
1797 // nothing to release
1798 if (!empty
&& logical
.empty()) {
1802 bluestore_blob_t
& b
= dirty_blob();
1803 return b
.release_extents(empty
, logical
, r
);
1806 bool BlueStore::Blob::can_reuse_blob(uint32_t min_alloc_size
,
1807 uint32_t target_blob_size
,
1809 uint32_t *length0
) {
1810 assert(min_alloc_size
);
1811 assert(target_blob_size
);
1812 if (!get_blob().is_mutable()) {
1816 uint32_t length
= *length0
;
1817 uint32_t end
= b_offset
+ length
;
1819 // Currently for the sake of simplicity we omit blob reuse if data is
1820 // unaligned with csum chunk. Later we can perform padding if needed.
1821 if (get_blob().has_csum() &&
1822 ((b_offset
% get_blob().get_csum_chunk_size()) != 0 ||
1823 (end
% get_blob().get_csum_chunk_size()) != 0)) {
1827 auto blen
= get_blob().get_logical_length();
1828 uint32_t new_blen
= blen
;
1830 // make sure target_blob_size isn't less than current blob len
1831 target_blob_size
= MAX(blen
, target_blob_size
);
1833 if (b_offset
>= blen
) {
1834 // new data totally stands out of the existing blob
1837 // new data overlaps with the existing blob
1838 new_blen
= MAX(blen
, end
);
1840 uint32_t overlap
= 0;
1841 if (new_blen
> blen
) {
1842 overlap
= blen
- b_offset
;
1847 if (!get_blob().is_unallocated(b_offset
, overlap
)) {
1848 // abort if any piece of the overlap has already been allocated
1853 if (new_blen
> blen
) {
1854 int64_t overflow
= int64_t(new_blen
) - target_blob_size
;
1855 // Unable to decrease the provided length to fit into max_blob_size
1856 if (overflow
>= length
) {
1860 // FIXME: in some cases we could reduce unused resolution
1861 if (get_blob().has_unused()) {
1866 new_blen
-= overflow
;
1871 if (new_blen
> blen
) {
1872 dirty_blob().add_tail(new_blen
);
1873 used_in_blob
.add_tail(new_blen
,
1874 get_blob().get_release_size(min_alloc_size
));
1880 void BlueStore::Blob::split(Collection
*coll
, uint32_t blob_offset
, Blob
*r
)
1882 auto cct
= coll
->store
->cct
; //used by dout
1883 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1884 << " start " << *this << dendl
;
1885 assert(blob
.can_split());
1886 assert(used_in_blob
.can_split());
1887 bluestore_blob_t
&lb
= dirty_blob();
1888 bluestore_blob_t
&rb
= r
->dirty_blob();
1892 &(r
->used_in_blob
));
1894 lb
.split(blob_offset
, rb
);
1895 shared_blob
->bc
.split(shared_blob
->get_cache(), blob_offset
, r
->shared_blob
->bc
);
1897 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1898 << " finish " << *this << dendl
;
1899 dout(10) << __func__
<< " 0x" << std::hex
<< blob_offset
<< std::dec
1900 << " and " << *r
<< dendl
;
1903 #ifndef CACHE_BLOB_BL
1904 void BlueStore::Blob::decode(
1906 bufferptr::iterator
& p
,
1909 bool include_ref_map
)
1911 denc(blob
, p
, struct_v
);
1912 if (blob
.is_shared()) {
1915 if (include_ref_map
) {
1917 used_in_blob
.decode(p
);
1919 used_in_blob
.clear();
1920 bluestore_extent_ref_map_t legacy_ref_map
;
1921 legacy_ref_map
.decode(p
);
1922 for (auto r
: legacy_ref_map
.ref_map
) {
1926 r
.second
.refs
* r
.second
.length
);
1935 ostream
& operator<<(ostream
& out
, const BlueStore::Extent
& e
)
1937 return out
<< std::hex
<< "0x" << e
.logical_offset
<< "~" << e
.length
1938 << ": 0x" << e
.blob_offset
<< "~" << e
.length
<< std::dec
1943 BlueStore::OldExtent
* BlueStore::OldExtent::create(CollectionRef c
,
1948 OldExtent
* oe
= new OldExtent(lo
, o
, l
, b
);
1949 b
->put_ref(c
.get(), o
, l
, &(oe
->r
));
1950 oe
->blob_empty
= b
->get_referenced_bytes() == 0;
1957 #define dout_prefix *_dout << "bluestore.extentmap(" << this << ") "
1959 BlueStore::ExtentMap::ExtentMap(Onode
*o
)
1962 o
->c
->store
->cct
->_conf
->bluestore_extent_map_inline_shard_prealloc_size
) {
1965 void BlueStore::ExtentMap::update(KeyValueDB::Transaction t
,
1968 auto cct
= onode
->c
->store
->cct
; //used by dout
1969 dout(20) << __func__
<< " " << onode
->oid
<< (force
? " force" : "") << dendl
;
1970 if (onode
->onode
.extent_map_shards
.empty()) {
1971 if (inline_bl
.length() == 0) {
1973 // we need to encode inline_bl to measure encoded length
1974 bool never_happen
= encode_some(0, OBJECT_MAX_SIZE
, inline_bl
, &n
);
1975 assert(!never_happen
);
1976 size_t len
= inline_bl
.length();
1977 dout(20) << __func__
<< " inline shard " << len
<< " bytes from " << n
1978 << " extents" << dendl
;
1979 if (!force
&& len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
1980 request_reshard(0, OBJECT_MAX_SIZE
);
1984 // will persist in the onode key.
1986 // pending shard update
1987 struct dirty_shard_t
{
1990 dirty_shard_t(Shard
*s
) : shard(s
) {}
1992 vector
<dirty_shard_t
> encoded_shards
;
1993 // allocate slots for all shards in a single call instead of
1994 // doing multiple allocations - one per each dirty shard
1995 encoded_shards
.reserve(shards
.size());
1997 auto p
= shards
.begin();
1999 while (p
!= shards
.end()) {
2000 assert(p
->shard_info
->offset
>= prev_p
->shard_info
->offset
);
2005 if (n
== shards
.end()) {
2006 endoff
= OBJECT_MAX_SIZE
;
2008 endoff
= n
->shard_info
->offset
;
2010 encoded_shards
.emplace_back(dirty_shard_t(&(*p
)));
2011 bufferlist
& bl
= encoded_shards
.back().bl
;
2012 if (encode_some(p
->shard_info
->offset
, endoff
- p
->shard_info
->offset
,
2015 derr
<< __func__
<< " encode_some needs reshard" << dendl
;
2019 size_t len
= bl
.length();
2021 dout(20) << __func__
<< " shard 0x" << std::hex
2022 << p
->shard_info
->offset
<< std::dec
<< " is " << len
2023 << " bytes (was " << p
->shard_info
->bytes
<< ") from "
2024 << p
->extents
<< " extents" << dendl
;
2027 if (len
> cct
->_conf
->bluestore_extent_map_shard_max_size
) {
2028 // we are big; reshard ourselves
2029 request_reshard(p
->shard_info
->offset
, endoff
);
2031 // avoid resharding the trailing shard, even if it is small
2032 else if (n
!= shards
.end() &&
2033 len
< g_conf
->bluestore_extent_map_shard_min_size
) {
2034 assert(endoff
!= OBJECT_MAX_SIZE
);
2035 if (p
== shards
.begin()) {
2036 // we are the first shard, combine with next shard
2037 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2039 // combine either with the previous shard or the next,
2040 // whichever is smaller
2041 if (prev_p
->shard_info
->bytes
> n
->shard_info
->bytes
) {
2042 request_reshard(p
->shard_info
->offset
, endoff
+ 1);
2044 request_reshard(prev_p
->shard_info
->offset
, endoff
);
2053 if (needs_reshard()) {
2057 // schedule DB update for dirty shards
2059 for (auto& it
: encoded_shards
) {
2060 it
.shard
->dirty
= false;
2061 it
.shard
->shard_info
->bytes
= it
.bl
.length();
2062 generate_extent_shard_key_and_apply(
2064 it
.shard
->shard_info
->offset
,
2066 [&](const string
& final_key
) {
2067 t
->set(PREFIX_OBJ
, final_key
, it
.bl
);
2074 bid_t
BlueStore::ExtentMap::allocate_spanning_blob_id()
2076 if (spanning_blob_map
.empty())
2078 bid_t bid
= spanning_blob_map
.rbegin()->first
+ 1;
2079 // bid is valid and available.
2082 // Find next unused bid;
2083 bid
= rand() % (numeric_limits
<bid_t
>::max() + 1);
2084 const auto begin_bid
= bid
;
2086 if (!spanning_blob_map
.count(bid
))
2090 if (bid
< 0) bid
= 0;
2092 } while (bid
!= begin_bid
);
2093 assert(0 == "no available blob id");
2096 void BlueStore::ExtentMap::reshard(
2098 KeyValueDB::Transaction t
)
2100 auto cct
= onode
->c
->store
->cct
; // used by dout
2102 dout(10) << __func__
<< " 0x[" << std::hex
<< needs_reshard_begin
<< ","
2103 << needs_reshard_end
<< ")" << std::dec
2104 << " of " << onode
->onode
.extent_map_shards
.size()
2105 << " shards on " << onode
->oid
<< dendl
;
2106 for (auto& p
: spanning_blob_map
) {
2107 dout(20) << __func__
<< " spanning blob " << p
.first
<< " " << *p
.second
2110 // determine shard index range
2111 unsigned si_begin
= 0, si_end
= 0;
2112 if (!shards
.empty()) {
2113 while (si_begin
+ 1 < shards
.size() &&
2114 shards
[si_begin
+ 1].shard_info
->offset
<= needs_reshard_begin
) {
2117 needs_reshard_begin
= shards
[si_begin
].shard_info
->offset
;
2118 for (si_end
= si_begin
; si_end
< shards
.size(); ++si_end
) {
2119 if (shards
[si_end
].shard_info
->offset
>= needs_reshard_end
) {
2120 needs_reshard_end
= shards
[si_end
].shard_info
->offset
;
2124 if (si_end
== shards
.size()) {
2125 needs_reshard_end
= OBJECT_MAX_SIZE
;
2127 dout(20) << __func__
<< " shards [" << si_begin
<< "," << si_end
<< ")"
2128 << " over 0x[" << std::hex
<< needs_reshard_begin
<< ","
2129 << needs_reshard_end
<< ")" << std::dec
<< dendl
;
2132 fault_range(db
, needs_reshard_begin
, needs_reshard_end
);
2134 // we may need to fault in a larger interval later must have all
2135 // referring extents for spanning blobs loaded in order to have
2136 // accurate use_tracker values.
2137 uint32_t spanning_scan_begin
= needs_reshard_begin
;
2138 uint32_t spanning_scan_end
= needs_reshard_end
;
2142 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2143 generate_extent_shard_key_and_apply(
2144 onode
->key
, shards
[i
].shard_info
->offset
, &key
,
2145 [&](const string
& final_key
) {
2146 t
->rmkey(PREFIX_OBJ
, final_key
);
2151 // calculate average extent size
2153 unsigned extents
= 0;
2154 if (onode
->onode
.extent_map_shards
.empty()) {
2155 bytes
= inline_bl
.length();
2156 extents
= extent_map
.size();
2158 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2159 bytes
+= shards
[i
].shard_info
->bytes
;
2160 extents
+= shards
[i
].extents
;
2163 unsigned target
= cct
->_conf
->bluestore_extent_map_shard_target_size
;
2164 unsigned slop
= target
*
2165 cct
->_conf
->bluestore_extent_map_shard_target_size_slop
;
2166 unsigned extent_avg
= bytes
/ MAX(1, extents
);
2167 dout(20) << __func__
<< " extent_avg " << extent_avg
<< ", target " << target
2168 << ", slop " << slop
<< dendl
;
2171 unsigned estimate
= 0;
2172 unsigned offset
= needs_reshard_begin
;
2173 vector
<bluestore_onode_t::shard_info
> new_shard_info
;
2174 unsigned max_blob_end
= 0;
2175 Extent
dummy(needs_reshard_begin
);
2176 for (auto e
= extent_map
.lower_bound(dummy
);
2177 e
!= extent_map
.end();
2179 if (e
->logical_offset
>= needs_reshard_end
) {
2182 dout(30) << " extent " << *e
<< dendl
;
2184 // disfavor shard boundaries that span a blob
2185 bool would_span
= (e
->logical_offset
< max_blob_end
) || e
->blob_offset
;
2187 estimate
+ extent_avg
> target
+ (would_span
? slop
: 0)) {
2189 if (offset
== needs_reshard_begin
) {
2190 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2191 new_shard_info
.back().offset
= offset
;
2192 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2193 << std::dec
<< dendl
;
2195 offset
= e
->logical_offset
;
2196 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2197 new_shard_info
.back().offset
= offset
;
2198 dout(20) << __func__
<< " new shard 0x" << std::hex
<< offset
2199 << std::dec
<< dendl
;
2202 estimate
+= extent_avg
;
2203 unsigned bs
= e
->blob_start();
2204 if (bs
< spanning_scan_begin
) {
2205 spanning_scan_begin
= bs
;
2207 uint32_t be
= e
->blob_end();
2208 if (be
> max_blob_end
) {
2211 if (be
> spanning_scan_end
) {
2212 spanning_scan_end
= be
;
2215 if (new_shard_info
.empty() && (si_begin
> 0 ||
2216 si_end
< shards
.size())) {
2217 // we resharded a partial range; we must produce at least one output
2219 new_shard_info
.emplace_back(bluestore_onode_t::shard_info());
2220 new_shard_info
.back().offset
= needs_reshard_begin
;
2221 dout(20) << __func__
<< " new shard 0x" << std::hex
<< needs_reshard_begin
2222 << std::dec
<< " (singleton degenerate case)" << dendl
;
2225 auto& sv
= onode
->onode
.extent_map_shards
;
2226 dout(20) << __func__
<< " new " << new_shard_info
<< dendl
;
2227 dout(20) << __func__
<< " old " << sv
<< dendl
;
2229 // no old shards to keep
2230 sv
.swap(new_shard_info
);
2231 init_shards(true, true);
2233 // splice in new shards
2234 sv
.erase(sv
.begin() + si_begin
, sv
.begin() + si_end
);
2235 shards
.erase(shards
.begin() + si_begin
, shards
.begin() + si_end
);
2237 sv
.begin() + si_begin
,
2238 new_shard_info
.begin(),
2239 new_shard_info
.end());
2240 shards
.insert(shards
.begin() + si_begin
, new_shard_info
.size(), Shard());
2241 si_end
= si_begin
+ new_shard_info
.size();
2243 assert(sv
.size() == shards
.size());
2245 // note that we need to update every shard_info of shards here,
2246 // as sv might have been totally re-allocated above
2247 for (unsigned i
= 0; i
< shards
.size(); i
++) {
2248 shards
[i
].shard_info
= &sv
[i
];
2251 // mark newly added shards as dirty
2252 for (unsigned i
= si_begin
; i
< si_end
; ++i
) {
2253 shards
[i
].loaded
= true;
2254 shards
[i
].dirty
= true;
2257 dout(20) << __func__
<< " fin " << sv
<< dendl
;
2261 // no more shards; unspan all previously spanning blobs
2262 auto p
= spanning_blob_map
.begin();
2263 while (p
!= spanning_blob_map
.end()) {
2265 dout(30) << __func__
<< " un-spanning " << *p
->second
<< dendl
;
2266 p
= spanning_blob_map
.erase(p
);
2269 // identify new spanning blobs
2270 dout(20) << __func__
<< " checking spanning blobs 0x[" << std::hex
2271 << spanning_scan_begin
<< "," << spanning_scan_end
<< ")" << dendl
;
2272 if (spanning_scan_begin
< needs_reshard_begin
) {
2273 fault_range(db
, spanning_scan_begin
,
2274 needs_reshard_begin
- spanning_scan_begin
);
2276 if (spanning_scan_end
> needs_reshard_end
) {
2277 fault_range(db
, needs_reshard_end
,
2278 spanning_scan_end
- needs_reshard_end
);
2280 auto sp
= sv
.begin() + si_begin
;
2281 auto esp
= sv
.end();
2282 unsigned shard_start
= sp
->offset
;
2286 shard_end
= OBJECT_MAX_SIZE
;
2288 shard_end
= sp
->offset
;
2290 Extent
dummy(needs_reshard_begin
);
2291 for (auto e
= extent_map
.lower_bound(dummy
); e
!= extent_map
.end(); ++e
) {
2292 if (e
->logical_offset
>= needs_reshard_end
) {
2295 dout(30) << " extent " << *e
<< dendl
;
2296 while (e
->logical_offset
>= shard_end
) {
2297 shard_start
= shard_end
;
2301 shard_end
= OBJECT_MAX_SIZE
;
2303 shard_end
= sp
->offset
;
2305 dout(30) << __func__
<< " shard 0x" << std::hex
<< shard_start
2306 << " to 0x" << shard_end
<< std::dec
<< dendl
;
2308 if (e
->blob_escapes_range(shard_start
, shard_end
- shard_start
)) {
2309 if (!e
->blob
->is_spanning()) {
2310 // We have two options: (1) split the blob into pieces at the
2311 // shard boundaries (and adjust extents accordingly), or (2)
2312 // mark it spanning. We prefer to cut the blob if we can. Note that
2313 // we may have to split it multiple times--potentially at every
2315 bool must_span
= false;
2316 BlobRef b
= e
->blob
;
2317 if (b
->can_split()) {
2318 uint32_t bstart
= e
->blob_start();
2319 uint32_t bend
= e
->blob_end();
2320 for (const auto& sh
: shards
) {
2321 if (bstart
< sh
.shard_info
->offset
&&
2322 bend
> sh
.shard_info
->offset
) {
2323 uint32_t blob_offset
= sh
.shard_info
->offset
- bstart
;
2324 if (b
->can_split_at(blob_offset
)) {
2325 dout(20) << __func__
<< " splitting blob, bstart 0x"
2326 << std::hex
<< bstart
<< " blob_offset 0x"
2327 << blob_offset
<< std::dec
<< " " << *b
<< dendl
;
2328 b
= split_blob(b
, blob_offset
, sh
.shard_info
->offset
);
2329 // switch b to the new right-hand side, in case it
2330 // *also* has to get split.
2331 bstart
+= blob_offset
;
2332 onode
->c
->store
->logger
->inc(l_bluestore_blob_split
);
2343 auto bid
= allocate_spanning_blob_id();
2345 spanning_blob_map
[b
->id
] = b
;
2346 dout(20) << __func__
<< " adding spanning " << *b
<< dendl
;
2350 if (e
->blob
->is_spanning()) {
2351 spanning_blob_map
.erase(e
->blob
->id
);
2353 dout(30) << __func__
<< " un-spanning " << *e
->blob
<< dendl
;
2359 clear_needs_reshard();
2362 bool BlueStore::ExtentMap::encode_some(
2368 auto cct
= onode
->c
->store
->cct
; //used by dout
2369 Extent
dummy(offset
);
2370 auto start
= extent_map
.lower_bound(dummy
);
2371 uint32_t end
= offset
+ length
;
2373 __u8 struct_v
= 2; // Version 2 differs from v1 in blob's ref_map
2374 // serialization only. Hence there is no specific
2375 // handling at ExtentMap level.
2379 bool must_reshard
= false;
2380 for (auto p
= start
;
2381 p
!= extent_map
.end() && p
->logical_offset
< end
;
2383 assert(p
->logical_offset
>= offset
);
2384 p
->blob
->last_encoded_id
= -1;
2385 if (!p
->blob
->is_spanning() && p
->blob_escapes_range(offset
, length
)) {
2386 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2387 << std::dec
<< " hit new spanning blob " << *p
<< dendl
;
2388 request_reshard(p
->blob_start(), p
->blob_end());
2389 must_reshard
= true;
2391 if (!must_reshard
) {
2392 denc_varint(0, bound
); // blobid
2393 denc_varint(0, bound
); // logical_offset
2394 denc_varint(0, bound
); // len
2395 denc_varint(0, bound
); // blob_offset
2397 p
->blob
->bound_encode(
2400 p
->blob
->shared_blob
->get_sbid(),
2408 denc(struct_v
, bound
);
2409 denc_varint(0, bound
); // number of extents
2412 auto app
= bl
.get_contiguous_appender(bound
);
2413 denc(struct_v
, app
);
2414 denc_varint(n
, app
);
2421 uint64_t prev_len
= 0;
2422 for (auto p
= start
;
2423 p
!= extent_map
.end() && p
->logical_offset
< end
;
2426 bool include_blob
= false;
2427 if (p
->blob
->is_spanning()) {
2428 blobid
= p
->blob
->id
<< BLOBID_SHIFT_BITS
;
2429 blobid
|= BLOBID_FLAG_SPANNING
;
2430 } else if (p
->blob
->last_encoded_id
< 0) {
2431 p
->blob
->last_encoded_id
= n
+ 1; // so it is always non-zero
2432 include_blob
= true;
2433 blobid
= 0; // the decoder will infer the id from n
2435 blobid
= p
->blob
->last_encoded_id
<< BLOBID_SHIFT_BITS
;
2437 if (p
->logical_offset
== pos
) {
2438 blobid
|= BLOBID_FLAG_CONTIGUOUS
;
2440 if (p
->blob_offset
== 0) {
2441 blobid
|= BLOBID_FLAG_ZEROOFFSET
;
2443 if (p
->length
== prev_len
) {
2444 blobid
|= BLOBID_FLAG_SAMELENGTH
;
2446 prev_len
= p
->length
;
2448 denc_varint(blobid
, app
);
2449 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2450 denc_varint_lowz(p
->logical_offset
- pos
, app
);
2452 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2453 denc_varint_lowz(p
->blob_offset
, app
);
2455 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2456 denc_varint_lowz(p
->length
, app
);
2458 pos
= p
->logical_end();
2460 p
->blob
->encode(app
, struct_v
, p
->blob
->shared_blob
->get_sbid(), false);
2464 /*derr << __func__ << bl << dendl;
2465 derr << __func__ << ":";
2472 unsigned BlueStore::ExtentMap::decode_some(bufferlist
& bl
)
2474 auto cct
= onode
->c
->store
->cct
; //used by dout
2476 derr << __func__ << ":";
2481 assert(bl
.get_num_buffers() <= 1);
2482 auto p
= bl
.front().begin_deep();
2485 // Version 2 differs from v1 in blob's ref_map
2486 // serialization only. Hence there is no specific
2487 // handling at ExtentMap level below.
2488 assert(struct_v
== 1 || struct_v
== 2);
2491 denc_varint(num
, p
);
2492 vector
<BlobRef
> blobs(num
);
2494 uint64_t prev_len
= 0;
2498 Extent
*le
= new Extent();
2500 denc_varint(blobid
, p
);
2501 if ((blobid
& BLOBID_FLAG_CONTIGUOUS
) == 0) {
2503 denc_varint_lowz(gap
, p
);
2506 le
->logical_offset
= pos
;
2507 if ((blobid
& BLOBID_FLAG_ZEROOFFSET
) == 0) {
2508 denc_varint_lowz(le
->blob_offset
, p
);
2510 le
->blob_offset
= 0;
2512 if ((blobid
& BLOBID_FLAG_SAMELENGTH
) == 0) {
2513 denc_varint_lowz(prev_len
, p
);
2515 le
->length
= prev_len
;
2517 if (blobid
& BLOBID_FLAG_SPANNING
) {
2518 dout(30) << __func__
<< " getting spanning blob "
2519 << (blobid
>> BLOBID_SHIFT_BITS
) << dendl
;
2520 le
->assign_blob(get_spanning_blob(blobid
>> BLOBID_SHIFT_BITS
));
2522 blobid
>>= BLOBID_SHIFT_BITS
;
2524 le
->assign_blob(blobs
[blobid
- 1]);
2527 Blob
*b
= new Blob();
2529 b
->decode(onode
->c
, p
, struct_v
, &sbid
, false);
2531 onode
->c
->open_shared_blob(sbid
, b
);
2534 // we build ref_map dynamically for non-spanning blobs
2542 extent_map
.insert(*le
);
2549 void BlueStore::ExtentMap::bound_encode_spanning_blobs(size_t& p
)
2551 // Version 2 differs from v1 in blob's ref_map
2552 // serialization only. Hence there is no specific
2553 // handling at ExtentMap level.
2557 denc_varint((uint32_t)0, p
);
2558 size_t key_size
= 0;
2559 denc_varint((uint32_t)0, key_size
);
2560 p
+= spanning_blob_map
.size() * key_size
;
2561 for (const auto& i
: spanning_blob_map
) {
2562 i
.second
->bound_encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2566 void BlueStore::ExtentMap::encode_spanning_blobs(
2567 bufferlist::contiguous_appender
& p
)
2569 // Version 2 differs from v1 in blob's ref_map
2570 // serialization only. Hence there is no specific
2571 // handling at ExtentMap level.
2575 denc_varint(spanning_blob_map
.size(), p
);
2576 for (auto& i
: spanning_blob_map
) {
2577 denc_varint(i
.second
->id
, p
);
2578 i
.second
->encode(p
, struct_v
, i
.second
->shared_blob
->get_sbid(), true);
2582 void BlueStore::ExtentMap::decode_spanning_blobs(
2583 bufferptr::iterator
& p
)
2587 // Version 2 differs from v1 in blob's ref_map
2588 // serialization only. Hence there is no specific
2589 // handling at ExtentMap level.
2590 assert(struct_v
== 1 || struct_v
== 2);
2595 BlobRef
b(new Blob());
2596 denc_varint(b
->id
, p
);
2597 spanning_blob_map
[b
->id
] = b
;
2599 b
->decode(onode
->c
, p
, struct_v
, &sbid
, true);
2600 onode
->c
->open_shared_blob(sbid
, b
);
2604 void BlueStore::ExtentMap::init_shards(bool loaded
, bool dirty
)
2606 shards
.resize(onode
->onode
.extent_map_shards
.size());
2608 for (auto &s
: onode
->onode
.extent_map_shards
) {
2609 shards
[i
].shard_info
= &s
;
2610 shards
[i
].loaded
= loaded
;
2611 shards
[i
].dirty
= dirty
;
2616 void BlueStore::ExtentMap::fault_range(
2621 auto cct
= onode
->c
->store
->cct
; //used by dout
2622 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2623 << std::dec
<< dendl
;
2624 auto start
= seek_shard(offset
);
2625 auto last
= seek_shard(offset
+ length
);
2630 assert(last
>= start
);
2632 while (start
<= last
) {
2633 assert((size_t)start
< shards
.size());
2634 auto p
= &shards
[start
];
2636 dout(30) << __func__
<< " opening shard 0x" << std::hex
2637 << p
->shard_info
->offset
<< std::dec
<< dendl
;
2639 generate_extent_shard_key_and_apply(
2640 onode
->key
, p
->shard_info
->offset
, &key
,
2641 [&](const string
& final_key
) {
2642 int r
= db
->get(PREFIX_OBJ
, final_key
, &v
);
2644 derr
<< __func__
<< " missing shard 0x" << std::hex
2645 << p
->shard_info
->offset
<< std::dec
<< " for " << onode
->oid
2651 p
->extents
= decode_some(v
);
2653 dout(20) << __func__
<< " open shard 0x" << std::hex
2654 << p
->shard_info
->offset
<< std::dec
2655 << " (" << v
.length() << " bytes)" << dendl
;
2656 assert(p
->dirty
== false);
2657 assert(v
.length() == p
->shard_info
->bytes
);
2658 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_misses
);
2660 onode
->c
->store
->logger
->inc(l_bluestore_onode_shard_hits
);
2666 void BlueStore::ExtentMap::dirty_range(
2670 auto cct
= onode
->c
->store
->cct
; //used by dout
2671 dout(30) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2672 << std::dec
<< dendl
;
2673 if (shards
.empty()) {
2674 dout(20) << __func__
<< " mark inline shard dirty" << dendl
;
2678 auto start
= seek_shard(offset
);
2679 auto last
= seek_shard(offset
+ length
);
2683 assert(last
>= start
);
2684 while (start
<= last
) {
2685 assert((size_t)start
< shards
.size());
2686 auto p
= &shards
[start
];
2688 dout(20) << __func__
<< " shard 0x" << std::hex
<< p
->shard_info
->offset
2689 << std::dec
<< " is not loaded, can't mark dirty" << dendl
;
2690 assert(0 == "can't mark unloaded shard dirty");
2693 dout(20) << __func__
<< " mark shard 0x" << std::hex
2694 << p
->shard_info
->offset
<< std::dec
<< " dirty" << dendl
;
2701 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::find(
2704 Extent
dummy(offset
);
2705 return extent_map
.find(dummy
);
2708 BlueStore::extent_map_t::iterator
BlueStore::ExtentMap::seek_lextent(
2711 Extent
dummy(offset
);
2712 auto fp
= extent_map
.lower_bound(dummy
);
2713 if (fp
!= extent_map
.begin()) {
2715 if (fp
->logical_end() <= offset
) {
2722 BlueStore::extent_map_t::const_iterator
BlueStore::ExtentMap::seek_lextent(
2723 uint64_t offset
) const
2725 Extent
dummy(offset
);
2726 auto fp
= extent_map
.lower_bound(dummy
);
2727 if (fp
!= extent_map
.begin()) {
2729 if (fp
->logical_end() <= offset
) {
2736 bool BlueStore::ExtentMap::has_any_lextents(uint64_t offset
, uint64_t length
)
2738 auto fp
= seek_lextent(offset
);
2739 if (fp
== extent_map
.end() || fp
->logical_offset
>= offset
+ length
) {
2745 int BlueStore::ExtentMap::compress_extent_map(
2749 auto cct
= onode
->c
->store
->cct
; //used by dout
2750 if (extent_map
.empty())
2753 auto p
= seek_lextent(offset
);
2754 if (p
!= extent_map
.begin()) {
2755 --p
; // start to the left of offset
2757 // the caller should have just written to this region
2758 assert(p
!= extent_map
.end());
2760 // identify the *next* shard
2761 auto pshard
= shards
.begin();
2762 while (pshard
!= shards
.end() &&
2763 p
->logical_offset
>= pshard
->shard_info
->offset
) {
2767 if (pshard
!= shards
.end()) {
2768 shard_end
= pshard
->shard_info
->offset
;
2770 shard_end
= OBJECT_MAX_SIZE
;
2774 for (++n
; n
!= extent_map
.end(); p
= n
++) {
2775 if (n
->logical_offset
> offset
+ length
) {
2776 break; // stop after end
2778 while (n
!= extent_map
.end() &&
2779 p
->logical_end() == n
->logical_offset
&&
2780 p
->blob
== n
->blob
&&
2781 p
->blob_offset
+ p
->length
== n
->blob_offset
&&
2782 n
->logical_offset
< shard_end
) {
2783 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
2784 << " next shard 0x" << shard_end
<< std::dec
2785 << " merging " << *p
<< " and " << *n
<< dendl
;
2786 p
->length
+= n
->length
;
2790 if (n
== extent_map
.end()) {
2793 if (n
->logical_offset
>= shard_end
) {
2794 assert(pshard
!= shards
.end());
2796 if (pshard
!= shards
.end()) {
2797 shard_end
= pshard
->shard_info
->offset
;
2799 shard_end
= OBJECT_MAX_SIZE
;
2803 if (removed
&& onode
) {
2804 onode
->c
->store
->logger
->inc(l_bluestore_extent_compress
, removed
);
2809 void BlueStore::ExtentMap::punch_hole(
2813 old_extent_map_t
*old_extents
)
2815 auto p
= seek_lextent(offset
);
2816 uint64_t end
= offset
+ length
;
2817 while (p
!= extent_map
.end()) {
2818 if (p
->logical_offset
>= end
) {
2821 if (p
->logical_offset
< offset
) {
2822 if (p
->logical_end() > end
) {
2823 // split and deref middle
2824 uint64_t front
= offset
- p
->logical_offset
;
2825 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ front
,
2827 old_extents
->push_back(*oe
);
2829 p
->blob_offset
+ front
+ length
,
2830 p
->length
- front
- length
,
2836 assert(p
->logical_end() > offset
); // else seek_lextent bug
2837 uint64_t keep
= offset
- p
->logical_offset
;
2838 OldExtent
* oe
= OldExtent::create(c
, offset
, p
->blob_offset
+ keep
,
2839 p
->length
- keep
, p
->blob
);
2840 old_extents
->push_back(*oe
);
2846 if (p
->logical_offset
+ p
->length
<= end
) {
2847 // deref whole lextent
2848 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2849 p
->length
, p
->blob
);
2850 old_extents
->push_back(*oe
);
2855 uint64_t keep
= p
->logical_end() - end
;
2856 OldExtent
* oe
= OldExtent::create(c
, p
->logical_offset
, p
->blob_offset
,
2857 p
->length
- keep
, p
->blob
);
2858 old_extents
->push_back(*oe
);
2860 add(end
, p
->blob_offset
+ p
->length
- keep
, keep
, p
->blob
);
2866 BlueStore::Extent
*BlueStore::ExtentMap::set_lextent(
2868 uint64_t logical_offset
,
2869 uint64_t blob_offset
, uint64_t length
, BlobRef b
,
2870 old_extent_map_t
*old_extents
)
2872 // We need to have completely initialized Blob to increment its ref counters.
2873 assert(b
->get_blob().get_logical_length() != 0);
2875 // Do get_ref prior to punch_hole to prevent from putting reused blob into
2876 // old_extents list if we overwre the blob totally
2877 // This might happen during WAL overwrite.
2878 b
->get_ref(onode
->c
, blob_offset
, length
);
2881 punch_hole(c
, logical_offset
, length
, old_extents
);
2884 Extent
*le
= new Extent(logical_offset
, blob_offset
, length
, b
);
2885 extent_map
.insert(*le
);
2886 if (spans_shard(logical_offset
, length
)) {
2887 request_reshard(logical_offset
, logical_offset
+ length
);
2892 BlueStore::BlobRef
BlueStore::ExtentMap::split_blob(
2894 uint32_t blob_offset
,
2897 auto cct
= onode
->c
->store
->cct
; //used by dout
2899 uint32_t end_pos
= pos
+ lb
->get_blob().get_logical_length() - blob_offset
;
2900 dout(20) << __func__
<< " 0x" << std::hex
<< pos
<< " end 0x" << end_pos
2901 << " blob_offset 0x" << blob_offset
<< std::dec
<< " " << *lb
2903 BlobRef rb
= onode
->c
->new_blob();
2904 lb
->split(onode
->c
, blob_offset
, rb
.get());
2906 for (auto ep
= seek_lextent(pos
);
2907 ep
!= extent_map
.end() && ep
->logical_offset
< end_pos
;
2909 if (ep
->blob
!= lb
) {
2912 if (ep
->logical_offset
< pos
) {
2914 size_t left
= pos
- ep
->logical_offset
;
2915 Extent
*ne
= new Extent(pos
, 0, ep
->length
- left
, rb
);
2916 extent_map
.insert(*ne
);
2918 dout(30) << __func__
<< " split " << *ep
<< dendl
;
2919 dout(30) << __func__
<< " to " << *ne
<< dendl
;
2922 assert(ep
->blob_offset
>= blob_offset
);
2925 ep
->blob_offset
-= blob_offset
;
2926 dout(30) << __func__
<< " adjusted " << *ep
<< dendl
;
2935 #define dout_prefix *_dout << "bluestore.onode(" << this << ")." << __func__ << " "
2937 void BlueStore::Onode::flush()
2939 if (flushing_count
.load()) {
2940 ldout(c
->store
->cct
, 20) << __func__
<< " cnt:" << flushing_count
<< dendl
;
2941 std::unique_lock
<std::mutex
> l(flush_lock
);
2942 while (flushing_count
.load()) {
2946 ldout(c
->store
->cct
, 20) << __func__
<< " done" << dendl
;
2949 // =======================================================
2952 /// Checks for writes to the same pextent within a blob
2953 bool BlueStore::WriteContext::has_conflict(
2957 uint64_t min_alloc_size
)
2959 assert((loffs
% min_alloc_size
) == 0);
2960 assert((loffs_end
% min_alloc_size
) == 0);
2961 for (auto w
: writes
) {
2963 auto loffs2
= P2ALIGN(w
.logical_offset
, min_alloc_size
);
2964 auto loffs2_end
= P2ROUNDUP(w
.logical_offset
+ w
.length0
, min_alloc_size
);
2965 if ((loffs
<= loffs2
&& loffs_end
> loffs2
) ||
2966 (loffs
>= loffs2
&& loffs
< loffs2_end
)) {
2974 // =======================================================
2978 #define dout_prefix *_dout << "bluestore.DeferredBatch(" << this << ") "
2980 void BlueStore::DeferredBatch::prepare_write(
2982 uint64_t seq
, uint64_t offset
, uint64_t length
,
2983 bufferlist::const_iterator
& blp
)
2985 _discard(cct
, offset
, length
);
2986 auto i
= iomap
.insert(make_pair(offset
, deferred_io()));
2987 assert(i
.second
); // this should be a new insertion
2988 i
.first
->second
.seq
= seq
;
2989 blp
.copy(length
, i
.first
->second
.bl
);
2990 i
.first
->second
.bl
.reassign_to_mempool(
2991 mempool::mempool_bluestore_writing_deferred
);
2992 dout(20) << __func__
<< " seq " << seq
2993 << " 0x" << std::hex
<< offset
<< "~" << length
2994 << " crc " << i
.first
->second
.bl
.crc32c(-1)
2995 << std::dec
<< dendl
;
2996 seq_bytes
[seq
] += length
;
2997 #ifdef DEBUG_DEFERRED
3002 void BlueStore::DeferredBatch::_discard(
3003 CephContext
*cct
, uint64_t offset
, uint64_t length
)
3005 generic_dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
3006 << std::dec
<< dendl
;
3007 auto p
= iomap
.lower_bound(offset
);
3008 if (p
!= iomap
.begin()) {
3010 auto end
= p
->first
+ p
->second
.bl
.length();
3013 head
.substr_of(p
->second
.bl
, 0, offset
- p
->first
);
3014 dout(20) << __func__
<< " keep head " << p
->second
.seq
3015 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3016 << " -> 0x" << head
.length() << std::dec
<< dendl
;
3017 auto i
= seq_bytes
.find(p
->second
.seq
);
3018 assert(i
!= seq_bytes
.end());
3019 if (end
> offset
+ length
) {
3021 tail
.substr_of(p
->second
.bl
, offset
+ length
- p
->first
,
3022 end
- (offset
+ length
));
3023 dout(20) << __func__
<< " keep tail " << p
->second
.seq
3024 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3025 << " -> 0x" << tail
.length() << std::dec
<< dendl
;
3026 auto &n
= iomap
[offset
+ length
];
3028 n
.seq
= p
->second
.seq
;
3029 i
->second
-= length
;
3031 i
->second
-= end
- offset
;
3033 assert(i
->second
>= 0);
3034 p
->second
.bl
.swap(head
);
3038 while (p
!= iomap
.end()) {
3039 if (p
->first
>= offset
+ length
) {
3042 auto i
= seq_bytes
.find(p
->second
.seq
);
3043 assert(i
!= seq_bytes
.end());
3044 auto end
= p
->first
+ p
->second
.bl
.length();
3045 if (end
> offset
+ length
) {
3046 unsigned drop_front
= offset
+ length
- p
->first
;
3047 unsigned keep_tail
= end
- (offset
+ length
);
3048 dout(20) << __func__
<< " truncate front " << p
->second
.seq
3049 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3050 << " drop_front 0x" << drop_front
<< " keep_tail 0x" << keep_tail
3051 << " to 0x" << (offset
+ length
) << "~" << keep_tail
3052 << std::dec
<< dendl
;
3053 auto &s
= iomap
[offset
+ length
];
3054 s
.seq
= p
->second
.seq
;
3055 s
.bl
.substr_of(p
->second
.bl
, drop_front
, keep_tail
);
3056 i
->second
-= drop_front
;
3058 dout(20) << __func__
<< " drop " << p
->second
.seq
3059 << " 0x" << std::hex
<< p
->first
<< "~" << p
->second
.bl
.length()
3060 << std::dec
<< dendl
;
3061 i
->second
-= p
->second
.bl
.length();
3063 assert(i
->second
>= 0);
3068 void BlueStore::DeferredBatch::_audit(CephContext
*cct
)
3070 map
<uint64_t,int> sb
;
3071 for (auto p
: seq_bytes
) {
3072 sb
[p
.first
] = 0; // make sure we have the same set of keys
3075 for (auto& p
: iomap
) {
3076 assert(p
.first
>= pos
);
3077 sb
[p
.second
.seq
] += p
.second
.bl
.length();
3078 pos
= p
.first
+ p
.second
.bl
.length();
3080 assert(sb
== seq_bytes
);
3087 #define dout_prefix *_dout << "bluestore(" << store->path << ").collection(" << cid << " " << this << ") "
3089 BlueStore::Collection::Collection(BlueStore
*ns
, Cache
*c
, coll_t cid
)
3093 lock("BlueStore::Collection::lock", true, false),
3099 void BlueStore::Collection::open_shared_blob(uint64_t sbid
, BlobRef b
)
3101 assert(!b
->shared_blob
);
3102 const bluestore_blob_t
& blob
= b
->get_blob();
3103 if (!blob
.is_shared()) {
3104 b
->shared_blob
= new SharedBlob(this);
3108 b
->shared_blob
= shared_blob_set
.lookup(sbid
);
3109 if (b
->shared_blob
) {
3110 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3111 << std::dec
<< " had " << *b
->shared_blob
<< dendl
;
3113 b
->shared_blob
= new SharedBlob(sbid
, this);
3114 shared_blob_set
.add(this, b
->shared_blob
.get());
3115 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3116 << std::dec
<< " opened " << *b
->shared_blob
3121 void BlueStore::Collection::load_shared_blob(SharedBlobRef sb
)
3123 if (!sb
->is_loaded()) {
3127 auto sbid
= sb
->get_sbid();
3128 get_shared_blob_key(sbid
, &key
);
3129 int r
= store
->db
->get(PREFIX_SHARED_BLOB
, key
, &v
);
3131 lderr(store
->cct
) << __func__
<< " sbid 0x" << std::hex
<< sbid
3132 << std::dec
<< " not found at key "
3133 << pretty_binary_string(key
) << dendl
;
3134 assert(0 == "uh oh, missing shared_blob");
3138 sb
->persistent
= new bluestore_shared_blob_t(sbid
);
3139 bufferlist::iterator p
= v
.begin();
3140 ::decode(*(sb
->persistent
), p
);
3141 ldout(store
->cct
, 10) << __func__
<< " sbid 0x" << std::hex
<< sbid
3142 << std::dec
<< " loaded shared_blob " << *sb
<< dendl
;
3146 void BlueStore::Collection::make_blob_shared(uint64_t sbid
, BlobRef b
)
3148 ldout(store
->cct
, 10) << __func__
<< " " << *b
<< dendl
;
3149 assert(!b
->shared_blob
->is_loaded());
3152 bluestore_blob_t
& blob
= b
->dirty_blob();
3153 blob
.set_flag(bluestore_blob_t::FLAG_SHARED
);
3155 // update shared blob
3156 b
->shared_blob
->loaded
= true;
3157 b
->shared_blob
->persistent
= new bluestore_shared_blob_t(sbid
);
3158 shared_blob_set
.add(this, b
->shared_blob
.get());
3159 for (auto p
: blob
.get_extents()) {
3161 b
->shared_blob
->get_ref(
3166 ldout(store
->cct
, 20) << __func__
<< " now " << *b
<< dendl
;
3169 uint64_t BlueStore::Collection::make_blob_unshared(SharedBlob
*sb
)
3171 ldout(store
->cct
, 10) << __func__
<< " " << *sb
<< dendl
;
3172 assert(sb
->is_loaded());
3174 uint64_t sbid
= sb
->get_sbid();
3175 shared_blob_set
.remove(sb
);
3177 delete sb
->persistent
;
3178 sb
->sbid_unloaded
= 0;
3179 ldout(store
->cct
, 20) << __func__
<< " now " << *sb
<< dendl
;
3183 BlueStore::OnodeRef
BlueStore::Collection::get_onode(
3184 const ghobject_t
& oid
,
3187 assert(create
? lock
.is_wlocked() : lock
.is_locked());
3190 if (cid
.is_pg(&pgid
)) {
3191 if (!oid
.match(cnode
.bits
, pgid
.ps())) {
3192 lderr(store
->cct
) << __func__
<< " oid " << oid
<< " not part of "
3193 << pgid
<< " bits " << cnode
.bits
<< dendl
;
3198 OnodeRef o
= onode_map
.lookup(oid
);
3202 mempool::bluestore_cache_other::string key
;
3203 get_object_key(store
->cct
, oid
, &key
);
3205 ldout(store
->cct
, 20) << __func__
<< " oid " << oid
<< " key "
3206 << pretty_binary_string(key
) << dendl
;
3209 int r
= store
->db
->get(PREFIX_OBJ
, key
.c_str(), key
.size(), &v
);
3210 ldout(store
->cct
, 20) << " r " << r
<< " v.len " << v
.length() << dendl
;
3212 if (v
.length() == 0) {
3213 assert(r
== -ENOENT
);
3214 if (!store
->cct
->_conf
->bluestore_debug_misc
&&
3218 // new object, new onode
3219 on
= new Onode(this, oid
, key
);
3223 on
= new Onode(this, oid
, key
);
3225 bufferptr::iterator p
= v
.front().begin_deep();
3226 on
->onode
.decode(p
);
3228 // initialize extent_map
3229 on
->extent_map
.decode_spanning_blobs(p
);
3230 if (on
->onode
.extent_map_shards
.empty()) {
3231 denc(on
->extent_map
.inline_bl
, p
);
3232 on
->extent_map
.decode_some(on
->extent_map
.inline_bl
);
3234 on
->extent_map
.init_shards(false, false);
3238 return onode_map
.add(oid
, o
);
3241 void BlueStore::Collection::split_cache(
3244 ldout(store
->cct
, 10) << __func__
<< " to " << dest
<< dendl
;
3246 // lock (one or both) cache shards
3247 std::lock(cache
->lock
, dest
->cache
->lock
);
3248 std::lock_guard
<std::recursive_mutex
> l(cache
->lock
, std::adopt_lock
);
3249 std::lock_guard
<std::recursive_mutex
> l2(dest
->cache
->lock
, std::adopt_lock
);
3251 int destbits
= dest
->cnode
.bits
;
3253 bool is_pg
= dest
->cid
.is_pg(&destpg
);
3256 auto p
= onode_map
.onode_map
.begin();
3257 while (p
!= onode_map
.onode_map
.end()) {
3258 if (!p
->second
->oid
.match(destbits
, destpg
.pgid
.ps())) {
3259 // onode does not belong to this child
3262 OnodeRef o
= p
->second
;
3263 ldout(store
->cct
, 20) << __func__
<< " moving " << o
<< " " << o
->oid
3266 cache
->_rm_onode(p
->second
);
3267 p
= onode_map
.onode_map
.erase(p
);
3270 dest
->cache
->_add_onode(o
, 1);
3271 dest
->onode_map
.onode_map
[o
->oid
] = o
;
3272 dest
->onode_map
.cache
= dest
->cache
;
3274 // move over shared blobs and buffers. cover shared blobs from
3275 // both extent map and spanning blob map (the full extent map
3276 // may not be faulted in)
3277 vector
<SharedBlob
*> sbvec
;
3278 for (auto& e
: o
->extent_map
.extent_map
) {
3279 sbvec
.push_back(e
.blob
->shared_blob
.get());
3281 for (auto& b
: o
->extent_map
.spanning_blob_map
) {
3282 sbvec
.push_back(b
.second
->shared_blob
.get());
3284 for (auto sb
: sbvec
) {
3285 if (sb
->coll
== dest
) {
3286 ldout(store
->cct
, 20) << __func__
<< " already moved " << *sb
3290 ldout(store
->cct
, 20) << __func__
<< " moving " << *sb
<< dendl
;
3292 if (sb
->get_sbid()) {
3293 ldout(store
->cct
, 20) << __func__
3294 << " moving registration " << *sb
<< dendl
;
3295 shared_blob_set
.remove(sb
);
3296 dest
->shared_blob_set
.add(dest
, sb
);
3298 if (dest
->cache
!= cache
) {
3299 for (auto& i
: sb
->bc
.buffer_map
) {
3300 if (!i
.second
->is_writing()) {
3301 ldout(store
->cct
, 20) << __func__
<< " moving " << *i
.second
3303 dest
->cache
->_move_buffer(cache
, i
.second
.get());
3312 // =======================================================
3314 void *BlueStore::MempoolThread::entry()
3316 Mutex::Locker
l(lock
);
3318 uint64_t meta_bytes
=
3319 mempool::bluestore_cache_other::allocated_bytes() +
3320 mempool::bluestore_cache_onode::allocated_bytes();
3321 uint64_t onode_num
=
3322 mempool::bluestore_cache_onode::allocated_items();
3324 if (onode_num
< 2) {
3328 float bytes_per_onode
= (float)meta_bytes
/ (float)onode_num
;
3329 size_t num_shards
= store
->cache_shards
.size();
3330 float target_ratio
= store
->cache_meta_ratio
+ store
->cache_data_ratio
;
3331 // A little sloppy but should be close enough
3332 uint64_t shard_target
= target_ratio
* (store
->cache_size
/ num_shards
);
3334 for (auto i
: store
->cache_shards
) {
3335 i
->trim(shard_target
,
3336 store
->cache_meta_ratio
,
3337 store
->cache_data_ratio
,
3341 store
->_update_cache_logger();
3344 wait
+= store
->cct
->_conf
->bluestore_cache_trim_interval
;
3345 cond
.WaitInterval(lock
, wait
);
3351 // =======================================================
3356 #define dout_prefix *_dout << "bluestore.OmapIteratorImpl(" << this << ") "
3358 BlueStore::OmapIteratorImpl::OmapIteratorImpl(
3359 CollectionRef c
, OnodeRef o
, KeyValueDB::Iterator it
)
3360 : c(c
), o(o
), it(it
)
3362 RWLock::RLocker
l(c
->lock
);
3363 if (o
->onode
.has_omap()) {
3364 get_omap_key(o
->onode
.nid
, string(), &head
);
3365 get_omap_tail(o
->onode
.nid
, &tail
);
3366 it
->lower_bound(head
);
3370 int BlueStore::OmapIteratorImpl::seek_to_first()
3372 RWLock::RLocker
l(c
->lock
);
3373 if (o
->onode
.has_omap()) {
3374 it
->lower_bound(head
);
3376 it
= KeyValueDB::Iterator();
3381 int BlueStore::OmapIteratorImpl::upper_bound(const string
& after
)
3383 RWLock::RLocker
l(c
->lock
);
3384 if (o
->onode
.has_omap()) {
3386 get_omap_key(o
->onode
.nid
, after
, &key
);
3387 ldout(c
->store
->cct
,20) << __func__
<< " after " << after
<< " key "
3388 << pretty_binary_string(key
) << dendl
;
3389 it
->upper_bound(key
);
3391 it
= KeyValueDB::Iterator();
3396 int BlueStore::OmapIteratorImpl::lower_bound(const string
& to
)
3398 RWLock::RLocker
l(c
->lock
);
3399 if (o
->onode
.has_omap()) {
3401 get_omap_key(o
->onode
.nid
, to
, &key
);
3402 ldout(c
->store
->cct
,20) << __func__
<< " to " << to
<< " key "
3403 << pretty_binary_string(key
) << dendl
;
3404 it
->lower_bound(key
);
3406 it
= KeyValueDB::Iterator();
3411 bool BlueStore::OmapIteratorImpl::valid()
3413 RWLock::RLocker
l(c
->lock
);
3414 bool r
= o
->onode
.has_omap() && it
&& it
->valid() &&
3415 it
->raw_key().second
<= tail
;
3416 if (it
&& it
->valid()) {
3417 ldout(c
->store
->cct
,20) << __func__
<< " is at "
3418 << pretty_binary_string(it
->raw_key().second
)
3424 int BlueStore::OmapIteratorImpl::next(bool validate
)
3426 RWLock::RLocker
l(c
->lock
);
3427 if (o
->onode
.has_omap()) {
3435 string
BlueStore::OmapIteratorImpl::key()
3437 RWLock::RLocker
l(c
->lock
);
3438 assert(it
->valid());
3439 string db_key
= it
->raw_key().second
;
3441 decode_omap_key(db_key
, &user_key
);
3445 bufferlist
BlueStore::OmapIteratorImpl::value()
3447 RWLock::RLocker
l(c
->lock
);
3448 assert(it
->valid());
3453 // =====================================
3456 #define dout_prefix *_dout << "bluestore(" << path << ") "
3459 static void aio_cb(void *priv
, void *priv2
)
3461 BlueStore
*store
= static_cast<BlueStore
*>(priv
);
3462 BlueStore::AioContext
*c
= static_cast<BlueStore::AioContext
*>(priv2
);
3463 c
->aio_finish(store
);
3466 BlueStore::BlueStore(CephContext
*cct
, const string
& path
)
3467 : ObjectStore(cct
, path
),
3468 throttle_bytes(cct
, "bluestore_throttle_bytes",
3469 cct
->_conf
->bluestore_throttle_bytes
),
3470 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3471 cct
->_conf
->bluestore_throttle_bytes
+
3472 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3473 kv_sync_thread(this),
3474 kv_finalize_thread(this),
3475 mempool_thread(this)
3478 cct
->_conf
->add_observer(this);
3479 set_cache_shards(1);
3482 BlueStore::BlueStore(CephContext
*cct
,
3484 uint64_t _min_alloc_size
)
3485 : ObjectStore(cct
, path
),
3486 throttle_bytes(cct
, "bluestore_throttle_bytes",
3487 cct
->_conf
->bluestore_throttle_bytes
),
3488 throttle_deferred_bytes(cct
, "bluestore_throttle_deferred_bytes",
3489 cct
->_conf
->bluestore_throttle_bytes
+
3490 cct
->_conf
->bluestore_throttle_deferred_bytes
),
3491 kv_sync_thread(this),
3492 kv_finalize_thread(this),
3493 min_alloc_size(_min_alloc_size
),
3494 min_alloc_size_order(ctz(_min_alloc_size
)),
3495 mempool_thread(this)
3498 cct
->_conf
->add_observer(this);
3499 set_cache_shards(1);
3502 BlueStore::~BlueStore()
3504 for (auto f
: finishers
) {
3509 cct
->_conf
->remove_observer(this);
3513 assert(bluefs
== NULL
);
3514 assert(fsid_fd
< 0);
3515 assert(path_fd
< 0);
3516 for (auto i
: cache_shards
) {
3519 cache_shards
.clear();
3522 const char **BlueStore::get_tracked_conf_keys() const
3524 static const char* KEYS
[] = {
3525 "bluestore_csum_type",
3526 "bluestore_compression_mode",
3527 "bluestore_compression_algorithm",
3528 "bluestore_compression_min_blob_size",
3529 "bluestore_compression_min_blob_size_ssd",
3530 "bluestore_compression_min_blob_size_hdd",
3531 "bluestore_compression_max_blob_size",
3532 "bluestore_compression_max_blob_size_ssd",
3533 "bluestore_compression_max_blob_size_hdd",
3534 "bluestore_compression_required_ratio",
3535 "bluestore_max_alloc_size",
3536 "bluestore_prefer_deferred_size",
3537 "bluestore_deferred_batch_ops",
3538 "bluestore_deferred_batch_ops_hdd",
3539 "bluestore_deferred_batch_ops_ssd",
3540 "bluestore_throttle_bytes",
3541 "bluestore_throttle_deferred_bytes",
3542 "bluestore_throttle_cost_per_io_hdd",
3543 "bluestore_throttle_cost_per_io_ssd",
3544 "bluestore_throttle_cost_per_io",
3545 "bluestore_max_blob_size",
3546 "bluestore_max_blob_size_ssd",
3547 "bluestore_max_blob_size_hdd",
3553 void BlueStore::handle_conf_change(const struct md_config_t
*conf
,
3554 const std::set
<std::string
> &changed
)
3556 if (changed
.count("bluestore_csum_type")) {
3559 if (changed
.count("bluestore_compression_mode") ||
3560 changed
.count("bluestore_compression_algorithm") ||
3561 changed
.count("bluestore_compression_min_blob_size") ||
3562 changed
.count("bluestore_compression_max_blob_size")) {
3567 if (changed
.count("bluestore_max_blob_size") ||
3568 changed
.count("bluestore_max_blob_size_ssd") ||
3569 changed
.count("bluestore_max_blob_size_hdd")) {
3571 // only after startup
3575 if (changed
.count("bluestore_prefer_deferred_size") ||
3576 changed
.count("bluestore_max_alloc_size") ||
3577 changed
.count("bluestore_deferred_batch_ops") ||
3578 changed
.count("bluestore_deferred_batch_ops_hdd") ||
3579 changed
.count("bluestore_deferred_batch_ops_ssd")) {
3581 // only after startup
3585 if (changed
.count("bluestore_throttle_cost_per_io") ||
3586 changed
.count("bluestore_throttle_cost_per_io_hdd") ||
3587 changed
.count("bluestore_throttle_cost_per_io_ssd")) {
3589 _set_throttle_params();
3592 if (changed
.count("bluestore_throttle_bytes")) {
3593 throttle_bytes
.reset_max(conf
->bluestore_throttle_bytes
);
3594 throttle_deferred_bytes
.reset_max(
3595 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3597 if (changed
.count("bluestore_throttle_deferred_bytes")) {
3598 throttle_deferred_bytes
.reset_max(
3599 conf
->bluestore_throttle_bytes
+ conf
->bluestore_throttle_deferred_bytes
);
3603 void BlueStore::_set_compression()
3605 auto m
= Compressor::get_comp_mode_type(cct
->_conf
->bluestore_compression_mode
);
3609 derr
<< __func__
<< " unrecognized value '"
3610 << cct
->_conf
->bluestore_compression_mode
3611 << "' for bluestore_compression_mode, reverting to 'none'"
3613 comp_mode
= Compressor::COMP_NONE
;
3616 compressor
= nullptr;
3618 if (comp_mode
== Compressor::COMP_NONE
) {
3619 dout(10) << __func__
<< " compression mode set to 'none', "
3620 << "ignore other compression setttings" << dendl
;
3624 if (cct
->_conf
->bluestore_compression_max_blob_size
) {
3625 comp_min_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size
;
3628 if (bdev
->is_rotational()) {
3629 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_hdd
;
3631 comp_min_blob_size
= cct
->_conf
->bluestore_compression_min_blob_size_ssd
;
3635 if (cct
->_conf
->bluestore_compression_max_blob_size
) {
3636 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size
;
3639 if (bdev
->is_rotational()) {
3640 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_hdd
;
3642 comp_max_blob_size
= cct
->_conf
->bluestore_compression_max_blob_size_ssd
;
3646 auto& alg_name
= cct
->_conf
->bluestore_compression_algorithm
;
3647 if (!alg_name
.empty()) {
3648 compressor
= Compressor::create(cct
, alg_name
);
3650 derr
<< __func__
<< " unable to initialize " << alg_name
.c_str() << " compressor"
3655 dout(10) << __func__
<< " mode " << Compressor::get_comp_mode_name(comp_mode
)
3656 << " alg " << (compressor
? compressor
->get_type_name() : "(none)")
3660 void BlueStore::_set_csum()
3662 csum_type
= Checksummer::CSUM_NONE
;
3663 int t
= Checksummer::get_csum_string_type(cct
->_conf
->bluestore_csum_type
);
3664 if (t
> Checksummer::CSUM_NONE
)
3667 dout(10) << __func__
<< " csum_type "
3668 << Checksummer::get_csum_type_string(csum_type
)
3672 void BlueStore::_set_throttle_params()
3674 if (cct
->_conf
->bluestore_throttle_cost_per_io
) {
3675 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io
;
3678 if (bdev
->is_rotational()) {
3679 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_hdd
;
3681 throttle_cost_per_io
= cct
->_conf
->bluestore_throttle_cost_per_io_ssd
;
3685 dout(10) << __func__
<< " throttle_cost_per_io " << throttle_cost_per_io
3688 void BlueStore::_set_blob_size()
3690 if (cct
->_conf
->bluestore_max_blob_size
) {
3691 max_blob_size
= cct
->_conf
->bluestore_max_blob_size
;
3694 if (bdev
->is_rotational()) {
3695 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_hdd
;
3697 max_blob_size
= cct
->_conf
->bluestore_max_blob_size_ssd
;
3700 dout(10) << __func__
<< " max_blob_size 0x" << std::hex
<< max_blob_size
3701 << std::dec
<< dendl
;
3704 int BlueStore::_set_cache_sizes()
3707 if (cct
->_conf
->bluestore_cache_size
) {
3708 cache_size
= cct
->_conf
->bluestore_cache_size
;
3710 // choose global cache size based on backend type
3711 if (bdev
->is_rotational()) {
3712 cache_size
= cct
->_conf
->bluestore_cache_size_hdd
;
3714 cache_size
= cct
->_conf
->bluestore_cache_size_ssd
;
3717 cache_meta_ratio
= cct
->_conf
->bluestore_cache_meta_ratio
;
3718 cache_kv_ratio
= cct
->_conf
->bluestore_cache_kv_ratio
;
3720 double cache_kv_max
= cct
->_conf
->bluestore_cache_kv_max
;
3721 double cache_kv_max_ratio
= 0;
3723 // if cache_kv_max is negative, disable it
3724 if (cache_size
> 0 && cache_kv_max
>= 0) {
3725 cache_kv_max_ratio
= (double) cache_kv_max
/ (double) cache_size
;
3726 if (cache_kv_max_ratio
< 1.0 && cache_kv_max_ratio
< cache_kv_ratio
) {
3727 dout(1) << __func__
<< " max " << cache_kv_max_ratio
3728 << " < ratio " << cache_kv_ratio
3730 cache_meta_ratio
= cache_meta_ratio
+ cache_kv_ratio
- cache_kv_max_ratio
;
3731 cache_kv_ratio
= cache_kv_max_ratio
;
3736 (double)1.0 - (double)cache_meta_ratio
- (double)cache_kv_ratio
;
3738 if (cache_meta_ratio
< 0 || cache_meta_ratio
> 1.0) {
3739 derr
<< __func__
<< "bluestore_cache_meta_ratio (" << cache_meta_ratio
3740 << ") must be in range [0,1.0]" << dendl
;
3743 if (cache_kv_ratio
< 0 || cache_kv_ratio
> 1.0) {
3744 derr
<< __func__
<< "bluestore_cache_kv_ratio (" << cache_kv_ratio
3745 << ") must be in range [0,1.0]" << dendl
;
3748 if (cache_meta_ratio
+ cache_kv_ratio
> 1.0) {
3749 derr
<< __func__
<< "bluestore_cache_meta_ratio (" << cache_meta_ratio
3750 << ") + bluestore_cache_kv_ratio (" << cache_kv_ratio
3751 << ") = " << cache_meta_ratio
+ cache_kv_ratio
<< "; must be <= 1.0"
3755 if (cache_data_ratio
< 0) {
3756 // deal with floating point imprecision
3757 cache_data_ratio
= 0;
3759 dout(1) << __func__
<< " cache_size " << cache_size
3760 << " meta " << cache_meta_ratio
3761 << " kv " << cache_kv_ratio
3762 << " data " << cache_data_ratio
3767 void BlueStore::_init_logger()
3769 PerfCountersBuilder
b(cct
, "bluestore",
3770 l_bluestore_first
, l_bluestore_last
);
3771 b
.add_time_avg(l_bluestore_kv_flush_lat
, "kv_flush_lat",
3772 "Average kv_thread flush latency",
3773 "fl_l", PerfCountersBuilder::PRIO_INTERESTING
);
3774 b
.add_time_avg(l_bluestore_kv_commit_lat
, "kv_commit_lat",
3775 "Average kv_thread commit latency");
3776 b
.add_time_avg(l_bluestore_kv_lat
, "kv_lat",
3777 "Average kv_thread sync latency",
3778 "k_l", PerfCountersBuilder::PRIO_INTERESTING
);
3779 b
.add_time_avg(l_bluestore_state_prepare_lat
, "state_prepare_lat",
3780 "Average prepare state latency");
3781 b
.add_time_avg(l_bluestore_state_aio_wait_lat
, "state_aio_wait_lat",
3782 "Average aio_wait state latency",
3783 "io_l", PerfCountersBuilder::PRIO_INTERESTING
);
3784 b
.add_time_avg(l_bluestore_state_io_done_lat
, "state_io_done_lat",
3785 "Average io_done state latency");
3786 b
.add_time_avg(l_bluestore_state_kv_queued_lat
, "state_kv_queued_lat",
3787 "Average kv_queued state latency");
3788 b
.add_time_avg(l_bluestore_state_kv_committing_lat
, "state_kv_commiting_lat",
3789 "Average kv_commiting state latency");
3790 b
.add_time_avg(l_bluestore_state_kv_done_lat
, "state_kv_done_lat",
3791 "Average kv_done state latency");
3792 b
.add_time_avg(l_bluestore_state_deferred_queued_lat
, "state_deferred_queued_lat",
3793 "Average deferred_queued state latency");
3794 b
.add_time_avg(l_bluestore_state_deferred_aio_wait_lat
, "state_deferred_aio_wait_lat",
3795 "Average aio_wait state latency");
3796 b
.add_time_avg(l_bluestore_state_deferred_cleanup_lat
, "state_deferred_cleanup_lat",
3797 "Average cleanup state latency");
3798 b
.add_time_avg(l_bluestore_state_finishing_lat
, "state_finishing_lat",
3799 "Average finishing state latency");
3800 b
.add_time_avg(l_bluestore_state_done_lat
, "state_done_lat",
3801 "Average done state latency");
3802 b
.add_time_avg(l_bluestore_throttle_lat
, "throttle_lat",
3803 "Average submit throttle latency",
3804 "th_l", PerfCountersBuilder::PRIO_CRITICAL
);
3805 b
.add_time_avg(l_bluestore_submit_lat
, "submit_lat",
3806 "Average submit latency",
3807 "s_l", PerfCountersBuilder::PRIO_CRITICAL
);
3808 b
.add_time_avg(l_bluestore_commit_lat
, "commit_lat",
3809 "Average commit latency",
3810 "c_l", PerfCountersBuilder::PRIO_CRITICAL
);
3811 b
.add_time_avg(l_bluestore_read_lat
, "read_lat",
3812 "Average read latency",
3813 "r_l", PerfCountersBuilder::PRIO_CRITICAL
);
3814 b
.add_time_avg(l_bluestore_read_onode_meta_lat
, "read_onode_meta_lat",
3815 "Average read onode metadata latency");
3816 b
.add_time_avg(l_bluestore_read_wait_aio_lat
, "read_wait_aio_lat",
3817 "Average read latency");
3818 b
.add_time_avg(l_bluestore_compress_lat
, "compress_lat",
3819 "Average compress latency");
3820 b
.add_time_avg(l_bluestore_decompress_lat
, "decompress_lat",
3821 "Average decompress latency");
3822 b
.add_time_avg(l_bluestore_csum_lat
, "csum_lat",
3823 "Average checksum latency");
3824 b
.add_u64_counter(l_bluestore_compress_success_count
, "compress_success_count",
3825 "Sum for beneficial compress ops");
3826 b
.add_u64_counter(l_bluestore_compress_rejected_count
, "compress_rejected_count",
3827 "Sum for compress ops rejected due to low net gain of space");
3828 b
.add_u64_counter(l_bluestore_write_pad_bytes
, "write_pad_bytes",
3829 "Sum for write-op padded bytes");
3830 b
.add_u64_counter(l_bluestore_deferred_write_ops
, "deferred_write_ops",
3831 "Sum for deferred write op");
3832 b
.add_u64_counter(l_bluestore_deferred_write_bytes
, "deferred_write_bytes",
3833 "Sum for deferred write bytes", "def");
3834 b
.add_u64_counter(l_bluestore_write_penalty_read_ops
, "write_penalty_read_ops",
3835 "Sum for write penalty read ops");
3836 b
.add_u64(l_bluestore_allocated
, "bluestore_allocated",
3837 "Sum for allocated bytes");
3838 b
.add_u64(l_bluestore_stored
, "bluestore_stored",
3839 "Sum for stored bytes");
3840 b
.add_u64(l_bluestore_compressed
, "bluestore_compressed",
3841 "Sum for stored compressed bytes");
3842 b
.add_u64(l_bluestore_compressed_allocated
, "bluestore_compressed_allocated",
3843 "Sum for bytes allocated for compressed data");
3844 b
.add_u64(l_bluestore_compressed_original
, "bluestore_compressed_original",
3845 "Sum for original bytes that were compressed");
3847 b
.add_u64(l_bluestore_onodes
, "bluestore_onodes",
3848 "Number of onodes in cache");
3849 b
.add_u64_counter(l_bluestore_onode_hits
, "bluestore_onode_hits",
3850 "Sum for onode-lookups hit in the cache");
3851 b
.add_u64_counter(l_bluestore_onode_misses
, "bluestore_onode_misses",
3852 "Sum for onode-lookups missed in the cache");
3853 b
.add_u64_counter(l_bluestore_onode_shard_hits
, "bluestore_onode_shard_hits",
3854 "Sum for onode-shard lookups hit in the cache");
3855 b
.add_u64_counter(l_bluestore_onode_shard_misses
,
3856 "bluestore_onode_shard_misses",
3857 "Sum for onode-shard lookups missed in the cache");
3858 b
.add_u64(l_bluestore_extents
, "bluestore_extents",
3859 "Number of extents in cache");
3860 b
.add_u64(l_bluestore_blobs
, "bluestore_blobs",
3861 "Number of blobs in cache");
3862 b
.add_u64(l_bluestore_buffers
, "bluestore_buffers",
3863 "Number of buffers in cache");
3864 b
.add_u64(l_bluestore_buffer_bytes
, "bluestore_buffer_bytes",
3865 "Number of buffer bytes in cache");
3866 b
.add_u64(l_bluestore_buffer_hit_bytes
, "bluestore_buffer_hit_bytes",
3867 "Sum for bytes of read hit in the cache");
3868 b
.add_u64(l_bluestore_buffer_miss_bytes
, "bluestore_buffer_miss_bytes",
3869 "Sum for bytes of read missed in the cache");
3871 b
.add_u64_counter(l_bluestore_write_big
, "bluestore_write_big",
3872 "Large aligned writes into fresh blobs");
3873 b
.add_u64_counter(l_bluestore_write_big_bytes
, "bluestore_write_big_bytes",
3874 "Large aligned writes into fresh blobs (bytes)");
3875 b
.add_u64_counter(l_bluestore_write_big_blobs
, "bluestore_write_big_blobs",
3876 "Large aligned writes into fresh blobs (blobs)");
3877 b
.add_u64_counter(l_bluestore_write_small
, "bluestore_write_small",
3878 "Small writes into existing or sparse small blobs");
3879 b
.add_u64_counter(l_bluestore_write_small_bytes
, "bluestore_write_small_bytes",
3880 "Small writes into existing or sparse small blobs (bytes)");
3881 b
.add_u64_counter(l_bluestore_write_small_unused
,
3882 "bluestore_write_small_unused",
3883 "Small writes into unused portion of existing blob");
3884 b
.add_u64_counter(l_bluestore_write_small_deferred
,
3885 "bluestore_write_small_deferred",
3886 "Small overwrites using deferred");
3887 b
.add_u64_counter(l_bluestore_write_small_pre_read
,
3888 "bluestore_write_small_pre_read",
3889 "Small writes that required we read some data (possibly "
3890 "cached) to fill out the block");
3891 b
.add_u64_counter(l_bluestore_write_small_new
, "bluestore_write_small_new",
3892 "Small write into new (sparse) blob");
3894 b
.add_u64_counter(l_bluestore_txc
, "bluestore_txc", "Transactions committed");
3895 b
.add_u64_counter(l_bluestore_onode_reshard
, "bluestore_onode_reshard",
3896 "Onode extent map reshard events");
3897 b
.add_u64_counter(l_bluestore_blob_split
, "bluestore_blob_split",
3898 "Sum for blob splitting due to resharding");
3899 b
.add_u64_counter(l_bluestore_extent_compress
, "bluestore_extent_compress",
3900 "Sum for extents that have been removed due to compression");
3901 b
.add_u64_counter(l_bluestore_gc_merged
, "bluestore_gc_merged",
3902 "Sum for extents that have been merged due to garbage "
3904 logger
= b
.create_perf_counters();
3905 cct
->get_perfcounters_collection()->add(logger
);
3908 int BlueStore::_reload_logger()
3910 struct store_statfs_t store_statfs
;
3912 int r
= statfs(&store_statfs
);
3914 logger
->set(l_bluestore_allocated
, store_statfs
.allocated
);
3915 logger
->set(l_bluestore_stored
, store_statfs
.stored
);
3916 logger
->set(l_bluestore_compressed
, store_statfs
.compressed
);
3917 logger
->set(l_bluestore_compressed_allocated
, store_statfs
.compressed_allocated
);
3918 logger
->set(l_bluestore_compressed_original
, store_statfs
.compressed_original
);
3923 void BlueStore::_shutdown_logger()
3925 cct
->get_perfcounters_collection()->remove(logger
);
3929 int BlueStore::get_block_device_fsid(CephContext
* cct
, const string
& path
,
3932 bluestore_bdev_label_t label
;
3933 int r
= _read_bdev_label(cct
, path
, &label
);
3936 *fsid
= label
.osd_uuid
;
3940 int BlueStore::_open_path()
3942 assert(path_fd
< 0);
3943 path_fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_DIRECTORY
));
3946 derr
<< __func__
<< " unable to open " << path
<< ": " << cpp_strerror(r
)
3953 void BlueStore::_close_path()
3955 VOID_TEMP_FAILURE_RETRY(::close(path_fd
));
3959 int BlueStore::_write_bdev_label(string path
, bluestore_bdev_label_t label
)
3961 dout(10) << __func__
<< " path " << path
<< " label " << label
<< dendl
;
3963 ::encode(label
, bl
);
3964 uint32_t crc
= bl
.crc32c(-1);
3966 assert(bl
.length() <= BDEV_LABEL_BLOCK_SIZE
);
3967 bufferptr
z(BDEV_LABEL_BLOCK_SIZE
- bl
.length());
3969 bl
.append(std::move(z
));
3971 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_WRONLY
));
3974 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
3978 int r
= bl
.write_fd(fd
);
3980 derr
<< __func__
<< " failed to write to " << path
3981 << ": " << cpp_strerror(r
) << dendl
;
3983 VOID_TEMP_FAILURE_RETRY(::close(fd
));
3987 int BlueStore::_read_bdev_label(CephContext
* cct
, string path
,
3988 bluestore_bdev_label_t
*label
)
3990 dout(10) << __func__
<< dendl
;
3991 int fd
= TEMP_FAILURE_RETRY(::open(path
.c_str(), O_RDONLY
));
3994 derr
<< __func__
<< " failed to open " << path
<< ": " << cpp_strerror(fd
)
3999 int r
= bl
.read_fd(fd
, BDEV_LABEL_BLOCK_SIZE
);
4000 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4002 derr
<< __func__
<< " failed to read from " << path
4003 << ": " << cpp_strerror(r
) << dendl
;
4007 uint32_t crc
, expected_crc
;
4008 bufferlist::iterator p
= bl
.begin();
4010 ::decode(*label
, p
);
4012 t
.substr_of(bl
, 0, p
.get_off());
4014 ::decode(expected_crc
, p
);
4016 catch (buffer::error
& e
) {
4017 derr
<< __func__
<< " unable to decode label at offset " << p
.get_off()
4022 if (crc
!= expected_crc
) {
4023 derr
<< __func__
<< " bad crc on label, expected " << expected_crc
4024 << " != actual " << crc
<< dendl
;
4027 dout(10) << __func__
<< " got " << *label
<< dendl
;
4031 int BlueStore::_check_or_set_bdev_label(
4032 string path
, uint64_t size
, string desc
, bool create
)
4034 bluestore_bdev_label_t label
;
4036 label
.osd_uuid
= fsid
;
4038 label
.btime
= ceph_clock_now();
4039 label
.description
= desc
;
4040 int r
= _write_bdev_label(path
, label
);
4044 int r
= _read_bdev_label(cct
, path
, &label
);
4047 if (cct
->_conf
->bluestore_debug_permit_any_bdev_label
) {
4048 dout(20) << __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4049 << " and fsid " << fsid
<< " check bypassed" << dendl
;
4051 else if (label
.osd_uuid
!= fsid
) {
4052 derr
<< __func__
<< " bdev " << path
<< " fsid " << label
.osd_uuid
4053 << " does not match our fsid " << fsid
<< dendl
;
4060 void BlueStore::_set_alloc_sizes(void)
4062 max_alloc_size
= cct
->_conf
->bluestore_max_alloc_size
;
4064 if (cct
->_conf
->bluestore_prefer_deferred_size
) {
4065 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size
;
4068 if (bdev
->is_rotational()) {
4069 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_hdd
;
4071 prefer_deferred_size
= cct
->_conf
->bluestore_prefer_deferred_size_ssd
;
4075 if (cct
->_conf
->bluestore_deferred_batch_ops
) {
4076 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops
;
4079 if (bdev
->is_rotational()) {
4080 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_hdd
;
4082 deferred_batch_ops
= cct
->_conf
->bluestore_deferred_batch_ops_ssd
;
4086 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
4087 << std::dec
<< " order " << min_alloc_size_order
4088 << " max_alloc_size 0x" << std::hex
<< max_alloc_size
4089 << " prefer_deferred_size 0x" << prefer_deferred_size
4091 << " deferred_batch_ops " << deferred_batch_ops
4095 int BlueStore::_open_bdev(bool create
)
4097 assert(bdev
== NULL
);
4098 string p
= path
+ "/block";
4099 bdev
= BlockDevice::create(cct
, p
, aio_cb
, static_cast<void*>(this));
4100 int r
= bdev
->open(p
);
4104 if (bdev
->supported_bdev_label()) {
4105 r
= _check_or_set_bdev_label(p
, bdev
->get_size(), "main", create
);
4110 // initialize global block parameters
4111 block_size
= bdev
->get_block_size();
4112 block_mask
= ~(block_size
- 1);
4113 block_size_order
= ctz(block_size
);
4114 assert(block_size
== 1u << block_size_order
);
4115 // and set cache_size based on device type
4116 r
= _set_cache_sizes();
4130 void BlueStore::_close_bdev()
4138 int BlueStore::_open_fm(bool create
)
4141 fm
= FreelistManager::create(cct
, freelist_type
, db
, PREFIX_ALLOC
);
4144 // initialize freespace
4145 dout(20) << __func__
<< " initializing freespace" << dendl
;
4146 KeyValueDB::Transaction t
= db
->get_transaction();
4149 bl
.append(freelist_type
);
4150 t
->set(PREFIX_SUPER
, "freelist_type", bl
);
4152 fm
->create(bdev
->get_size(), t
);
4154 // allocate superblock reserved space. note that we do not mark
4155 // bluefs space as allocated in the freelist; we instead rely on
4157 fm
->allocate(0, SUPER_RESERVED
, t
);
4159 uint64_t reserved
= 0;
4160 if (cct
->_conf
->bluestore_bluefs
) {
4161 assert(bluefs_extents
.num_intervals() == 1);
4162 interval_set
<uint64_t>::iterator p
= bluefs_extents
.begin();
4163 reserved
= p
.get_start() + p
.get_len();
4164 dout(20) << __func__
<< " reserved 0x" << std::hex
<< reserved
<< std::dec
4165 << " for bluefs" << dendl
;
4167 ::encode(bluefs_extents
, bl
);
4168 t
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
4169 dout(20) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
4170 << std::dec
<< dendl
;
4172 reserved
= SUPER_RESERVED
;
4175 if (cct
->_conf
->bluestore_debug_prefill
> 0) {
4176 uint64_t end
= bdev
->get_size() - reserved
;
4177 dout(1) << __func__
<< " pre-fragmenting freespace, using "
4178 << cct
->_conf
->bluestore_debug_prefill
<< " with max free extent "
4179 << cct
->_conf
->bluestore_debug_prefragment_max
<< dendl
;
4180 uint64_t start
= P2ROUNDUP(reserved
, min_alloc_size
);
4181 uint64_t max_b
= cct
->_conf
->bluestore_debug_prefragment_max
/ min_alloc_size
;
4182 float r
= cct
->_conf
->bluestore_debug_prefill
;
4186 while (!stop
&& start
< end
) {
4187 uint64_t l
= (rand() % max_b
+ 1) * min_alloc_size
;
4188 if (start
+ l
> end
) {
4190 l
= P2ALIGN(l
, min_alloc_size
);
4192 assert(start
+ l
<= end
);
4194 uint64_t u
= 1 + (uint64_t)(r
* (double)l
);
4195 u
= P2ROUNDUP(u
, min_alloc_size
);
4196 if (start
+ l
+ u
> end
) {
4197 u
= end
- (start
+ l
);
4198 // trim to align so we don't overflow again
4199 u
= P2ALIGN(u
, min_alloc_size
);
4202 assert(start
+ l
+ u
<= end
);
4204 dout(20) << " free 0x" << std::hex
<< start
<< "~" << l
4205 << " use 0x" << u
<< std::dec
<< dendl
;
4208 // break if u has been trimmed to nothing
4212 fm
->allocate(start
+ l
, u
, t
);
4216 db
->submit_transaction_sync(t
);
4221 derr
<< __func__
<< " freelist init failed: " << cpp_strerror(r
) << dendl
;
4229 void BlueStore::_close_fm()
4231 dout(10) << __func__
<< dendl
;
4238 int BlueStore::_open_alloc()
4240 assert(alloc
== NULL
);
4241 assert(bdev
->get_size());
4242 alloc
= Allocator::create(cct
, cct
->_conf
->bluestore_allocator
,
4246 lderr(cct
) << __func__
<< " Allocator::unknown alloc type "
4247 << cct
->_conf
->bluestore_allocator
4252 uint64_t num
= 0, bytes
= 0;
4254 dout(1) << __func__
<< " opening allocation metadata" << dendl
;
4255 // initialize from freelist
4256 fm
->enumerate_reset();
4257 uint64_t offset
, length
;
4258 while (fm
->enumerate_next(&offset
, &length
)) {
4259 alloc
->init_add_free(offset
, length
);
4263 fm
->enumerate_reset();
4264 dout(1) << __func__
<< " loaded " << pretty_si_t(bytes
)
4265 << " in " << num
<< " extents"
4268 // also mark bluefs space as allocated
4269 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
4270 alloc
->init_rm_free(e
.get_start(), e
.get_len());
4272 dout(10) << __func__
<< " marked bluefs_extents 0x" << std::hex
4273 << bluefs_extents
<< std::dec
<< " as allocated" << dendl
;
4278 void BlueStore::_close_alloc()
4286 int BlueStore::_open_fsid(bool create
)
4288 assert(fsid_fd
< 0);
4292 fsid_fd
= ::openat(path_fd
, "fsid", flags
, 0644);
4295 derr
<< __func__
<< " " << cpp_strerror(err
) << dendl
;
4301 int BlueStore::_read_fsid(uuid_d
*uuid
)
4304 memset(fsid_str
, 0, sizeof(fsid_str
));
4305 int ret
= safe_read(fsid_fd
, fsid_str
, sizeof(fsid_str
));
4307 derr
<< __func__
<< " failed: " << cpp_strerror(ret
) << dendl
;
4314 if (!uuid
->parse(fsid_str
)) {
4315 derr
<< __func__
<< " unparsable uuid " << fsid_str
<< dendl
;
4321 int BlueStore::_write_fsid()
4323 int r
= ::ftruncate(fsid_fd
, 0);
4326 derr
<< __func__
<< " fsid truncate failed: " << cpp_strerror(r
) << dendl
;
4329 string str
= stringify(fsid
) + "\n";
4330 r
= safe_write(fsid_fd
, str
.c_str(), str
.length());
4332 derr
<< __func__
<< " fsid write failed: " << cpp_strerror(r
) << dendl
;
4335 r
= ::fsync(fsid_fd
);
4338 derr
<< __func__
<< " fsid fsync failed: " << cpp_strerror(r
) << dendl
;
4344 void BlueStore::_close_fsid()
4346 VOID_TEMP_FAILURE_RETRY(::close(fsid_fd
));
4350 int BlueStore::_lock_fsid()
4353 memset(&l
, 0, sizeof(l
));
4355 l
.l_whence
= SEEK_SET
;
4356 int r
= ::fcntl(fsid_fd
, F_SETLK
, &l
);
4359 derr
<< __func__
<< " failed to lock " << path
<< "/fsid"
4360 << " (is another ceph-osd still running?)"
4361 << cpp_strerror(err
) << dendl
;
4367 bool BlueStore::is_rotational()
4370 return bdev
->is_rotational();
4373 bool rotational
= true;
4374 int r
= _open_path();
4377 r
= _open_fsid(false);
4380 r
= _read_fsid(&fsid
);
4386 r
= _open_bdev(false);
4389 rotational
= bdev
->is_rotational();
4399 bool BlueStore::test_mount_in_use()
4401 // most error conditions mean the mount is not in use (e.g., because
4402 // it doesn't exist). only if we fail to lock do we conclude it is
4405 int r
= _open_path();
4408 r
= _open_fsid(false);
4413 ret
= true; // if we can't lock, it is in use
4420 int BlueStore::_open_db(bool create
)
4424 string fn
= path
+ "/db";
4427 ceph::shared_ptr
<Int64ArrayMergeOperator
> merge_op(new Int64ArrayMergeOperator
);
4431 kv_backend
= cct
->_conf
->bluestore_kvbackend
;
4433 r
= read_meta("kv_backend", &kv_backend
);
4435 derr
<< __func__
<< " unable to read 'kv_backend' meta" << dendl
;
4439 dout(10) << __func__
<< " kv_backend = " << kv_backend
<< dendl
;
4443 do_bluefs
= cct
->_conf
->bluestore_bluefs
;
4446 r
= read_meta("bluefs", &s
);
4448 derr
<< __func__
<< " unable to read 'bluefs' meta" << dendl
;
4453 } else if (s
== "0") {
4456 derr
<< __func__
<< " bluefs = " << s
<< " : not 0 or 1, aborting"
4461 dout(10) << __func__
<< " do_bluefs = " << do_bluefs
<< dendl
;
4463 rocksdb::Env
*env
= NULL
;
4465 dout(10) << __func__
<< " initializing bluefs" << dendl
;
4466 if (kv_backend
!= "rocksdb") {
4467 derr
<< " backend must be rocksdb to use bluefs" << dendl
;
4470 bluefs
= new BlueFS(cct
);
4475 bfn
= path
+ "/block.db";
4476 if (::stat(bfn
.c_str(), &st
) == 0) {
4477 r
= bluefs
->add_block_device(BlueFS::BDEV_DB
, bfn
);
4479 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4480 << cpp_strerror(r
) << dendl
;
4484 if (bluefs
->bdev_support_label(BlueFS::BDEV_DB
)) {
4485 r
= _check_or_set_bdev_label(
4487 bluefs
->get_block_device_size(BlueFS::BDEV_DB
),
4488 "bluefs db", create
);
4491 << " check block device(" << bfn
<< ") label returned: "
4492 << cpp_strerror(r
) << dendl
;
4497 bluefs
->add_block_extent(
4500 bluefs
->get_block_device_size(BlueFS::BDEV_DB
) - SUPER_RESERVED
);
4502 bluefs_shared_bdev
= BlueFS::BDEV_SLOW
;
4503 bluefs_single_shared_device
= false;
4504 } else if (::lstat(bfn
.c_str(), &st
) == -1) {
4505 bluefs_shared_bdev
= BlueFS::BDEV_DB
;
4507 //symlink exist is bug
4508 derr
<< __func__
<< " " << bfn
<< " link target doesn't exist" << dendl
;
4514 bfn
= path
+ "/block";
4515 r
= bluefs
->add_block_device(bluefs_shared_bdev
, bfn
);
4517 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4518 << cpp_strerror(r
) << dendl
;
4522 // note: we always leave the first SUPER_RESERVED (8k) of the device unused
4524 bdev
->get_size() * (cct
->_conf
->bluestore_bluefs_min_ratio
+
4525 cct
->_conf
->bluestore_bluefs_gift_ratio
);
4526 initial
= MAX(initial
, cct
->_conf
->bluestore_bluefs_min
);
4527 // align to bluefs's alloc_size
4528 initial
= P2ROUNDUP(initial
, cct
->_conf
->bluefs_alloc_size
);
4529 // put bluefs in the middle of the device in case it is an HDD
4530 uint64_t start
= P2ALIGN((bdev
->get_size() - initial
) / 2,
4531 cct
->_conf
->bluefs_alloc_size
);
4532 bluefs
->add_block_extent(bluefs_shared_bdev
, start
, initial
);
4533 bluefs_extents
.insert(start
, initial
);
4536 bfn
= path
+ "/block.wal";
4537 if (::stat(bfn
.c_str(), &st
) == 0) {
4538 r
= bluefs
->add_block_device(BlueFS::BDEV_WAL
, bfn
);
4540 derr
<< __func__
<< " add block device(" << bfn
<< ") returned: "
4541 << cpp_strerror(r
) << dendl
;
4545 if (bluefs
->bdev_support_label(BlueFS::BDEV_WAL
)) {
4546 r
= _check_or_set_bdev_label(
4548 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
),
4549 "bluefs wal", create
);
4551 derr
<< __func__
<< " check block device(" << bfn
4552 << ") label returned: " << cpp_strerror(r
) << dendl
;
4558 bluefs
->add_block_extent(
4559 BlueFS::BDEV_WAL
, BDEV_LABEL_BLOCK_SIZE
,
4560 bluefs
->get_block_device_size(BlueFS::BDEV_WAL
) -
4561 BDEV_LABEL_BLOCK_SIZE
);
4563 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "true");
4564 bluefs_single_shared_device
= false;
4565 } else if (::lstat(bfn
.c_str(), &st
) == -1) {
4566 cct
->_conf
->set_val("rocksdb_separate_wal_dir", "false");
4568 //symlink exist is bug
4569 derr
<< __func__
<< " " << bfn
<< " link target doesn't exist" << dendl
;
4577 r
= bluefs
->mount();
4579 derr
<< __func__
<< " failed bluefs mount: " << cpp_strerror(r
) << dendl
;
4582 if (cct
->_conf
->bluestore_bluefs_env_mirror
) {
4583 rocksdb::Env
*a
= new BlueRocksEnv(bluefs
);
4584 rocksdb::Env
*b
= rocksdb::Env::Default();
4586 string cmd
= "rm -rf " + path
+ "/db " +
4587 path
+ "/db.slow " +
4589 int r
= system(cmd
.c_str());
4592 env
= new rocksdb::EnvMirror(b
, a
, false, true);
4594 env
= new BlueRocksEnv(bluefs
);
4596 // simplify the dir names, too, as "seen" by rocksdb
4600 if (bluefs_shared_bdev
== BlueFS::BDEV_SLOW
) {
4601 // we have both block.db and block; tell rocksdb!
4602 // note: the second (last) size value doesn't really matter
4603 ostringstream db_paths
;
4604 uint64_t db_size
= bluefs
->get_block_device_size(BlueFS::BDEV_DB
);
4605 uint64_t slow_size
= bluefs
->get_block_device_size(BlueFS::BDEV_SLOW
);
4606 db_paths
<< fn
<< ","
4607 << (uint64_t)(db_size
* 95 / 100) << " "
4608 << fn
+ ".slow" << ","
4609 << (uint64_t)(slow_size
* 95 / 100);
4610 cct
->_conf
->set_val("rocksdb_db_paths", db_paths
.str(), false);
4611 dout(10) << __func__
<< " set rocksdb_db_paths to "
4612 << cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths") << dendl
;
4617 if (cct
->_conf
->rocksdb_separate_wal_dir
)
4618 env
->CreateDir(fn
+ ".wal");
4619 if (cct
->_conf
->get_val
<std::string
>("rocksdb_db_paths").length())
4620 env
->CreateDir(fn
+ ".slow");
4622 } else if (create
) {
4623 int r
= ::mkdir(fn
.c_str(), 0755);
4626 if (r
< 0 && r
!= -EEXIST
) {
4627 derr
<< __func__
<< " failed to create " << fn
<< ": " << cpp_strerror(r
)
4633 if (cct
->_conf
->rocksdb_separate_wal_dir
) {
4634 string walfn
= path
+ "/db.wal";
4635 r
= ::mkdir(walfn
.c_str(), 0755);
4638 if (r
< 0 && r
!= -EEXIST
) {
4639 derr
<< __func__
<< " failed to create " << walfn
4640 << ": " << cpp_strerror(r
)
4647 db
= KeyValueDB::create(cct
,
4650 static_cast<void*>(env
));
4652 derr
<< __func__
<< " error creating db" << dendl
;
4658 // delete env manually here since we can't depend on db to do this
4665 FreelistManager::setup_merge_operators(db
);
4666 db
->set_merge_operator(PREFIX_STAT
, merge_op
);
4668 db
->set_cache_size(cache_size
* cache_kv_ratio
);
4670 if (kv_backend
== "rocksdb")
4671 options
= cct
->_conf
->bluestore_rocksdb_options
;
4674 r
= db
->create_and_open(err
);
4678 derr
<< __func__
<< " erroring opening db: " << err
.str() << dendl
;
4688 dout(1) << __func__
<< " opened " << kv_backend
4689 << " path " << fn
<< " options " << options
<< dendl
;
4699 void BlueStore::_close_db()
4711 int BlueStore::_reconcile_bluefs_freespace()
4713 dout(10) << __func__
<< dendl
;
4714 interval_set
<uint64_t> bset
;
4715 int r
= bluefs
->get_block_extents(bluefs_shared_bdev
, &bset
);
4717 if (bset
== bluefs_extents
) {
4718 dout(10) << __func__
<< " we agree bluefs has 0x" << std::hex
<< bset
4719 << std::dec
<< dendl
;
4722 dout(10) << __func__
<< " bluefs says 0x" << std::hex
<< bset
<< std::dec
4724 dout(10) << __func__
<< " super says 0x" << std::hex
<< bluefs_extents
4725 << std::dec
<< dendl
;
4727 interval_set
<uint64_t> overlap
;
4728 overlap
.intersection_of(bset
, bluefs_extents
);
4730 bset
.subtract(overlap
);
4731 if (!bset
.empty()) {
4732 derr
<< __func__
<< " bluefs extra 0x" << std::hex
<< bset
<< std::dec
4737 interval_set
<uint64_t> super_extra
;
4738 super_extra
= bluefs_extents
;
4739 super_extra
.subtract(overlap
);
4740 if (!super_extra
.empty()) {
4741 // This is normal: it can happen if we commit to give extents to
4742 // bluefs and we crash before bluefs commits that it owns them.
4743 dout(10) << __func__
<< " super extra " << super_extra
<< dendl
;
4744 for (interval_set
<uint64_t>::iterator p
= super_extra
.begin();
4745 p
!= super_extra
.end();
4747 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.get_start(), p
.get_len());
4754 int BlueStore::_balance_bluefs_freespace(PExtentVector
*extents
)
4759 vector
<pair
<uint64_t,uint64_t>> bluefs_usage
; // <free, total> ...
4760 bluefs
->get_usage(&bluefs_usage
);
4761 assert(bluefs_usage
.size() > bluefs_shared_bdev
);
4763 // fixme: look at primary bdev only for now
4764 uint64_t bluefs_free
= bluefs_usage
[bluefs_shared_bdev
].first
;
4765 uint64_t bluefs_total
= bluefs_usage
[bluefs_shared_bdev
].second
;
4766 float bluefs_free_ratio
= (float)bluefs_free
/ (float)bluefs_total
;
4768 uint64_t my_free
= alloc
->get_free();
4769 uint64_t total
= bdev
->get_size();
4770 float my_free_ratio
= (float)my_free
/ (float)total
;
4772 uint64_t total_free
= bluefs_free
+ my_free
;
4774 float bluefs_ratio
= (float)bluefs_free
/ (float)total_free
;
4776 dout(10) << __func__
4777 << " bluefs " << pretty_si_t(bluefs_free
)
4778 << " free (" << bluefs_free_ratio
4779 << ") bluestore " << pretty_si_t(my_free
)
4780 << " free (" << my_free_ratio
4781 << "), bluefs_ratio " << bluefs_ratio
4785 uint64_t reclaim
= 0;
4786 if (bluefs_ratio
< cct
->_conf
->bluestore_bluefs_min_ratio
) {
4787 gift
= cct
->_conf
->bluestore_bluefs_gift_ratio
* total_free
;
4788 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4789 << " < min_ratio " << cct
->_conf
->bluestore_bluefs_min_ratio
4790 << ", should gift " << pretty_si_t(gift
) << dendl
;
4791 } else if (bluefs_ratio
> cct
->_conf
->bluestore_bluefs_max_ratio
) {
4792 reclaim
= cct
->_conf
->bluestore_bluefs_reclaim_ratio
* total_free
;
4793 if (bluefs_total
- reclaim
< cct
->_conf
->bluestore_bluefs_min
)
4794 reclaim
= bluefs_total
- cct
->_conf
->bluestore_bluefs_min
;
4795 dout(10) << __func__
<< " bluefs_ratio " << bluefs_ratio
4796 << " > max_ratio " << cct
->_conf
->bluestore_bluefs_max_ratio
4797 << ", should reclaim " << pretty_si_t(reclaim
) << dendl
;
4799 if (bluefs_total
< cct
->_conf
->bluestore_bluefs_min
&&
4800 cct
->_conf
->bluestore_bluefs_min
<
4801 (uint64_t)(cct
->_conf
->bluestore_bluefs_max_ratio
* total_free
)) {
4802 uint64_t g
= cct
->_conf
->bluestore_bluefs_min
- bluefs_total
;
4803 dout(10) << __func__
<< " bluefs_total " << bluefs_total
4804 << " < min " << cct
->_conf
->bluestore_bluefs_min
4805 << ", should gift " << pretty_si_t(g
) << dendl
;
4812 // round up to alloc size
4813 gift
= P2ROUNDUP(gift
, cct
->_conf
->bluefs_alloc_size
);
4815 // hard cap to fit into 32 bits
4816 gift
= MIN(gift
, 1ull<<31);
4817 dout(10) << __func__
<< " gifting " << gift
4818 << " (" << pretty_si_t(gift
) << ")" << dendl
;
4820 // fixme: just do one allocation to start...
4821 int r
= alloc
->reserve(gift
);
4824 AllocExtentVector exts
;
4825 int64_t alloc_len
= alloc
->allocate(gift
, cct
->_conf
->bluefs_alloc_size
,
4828 if (alloc_len
< (int64_t)gift
) {
4829 derr
<< __func__
<< " allocate failed on 0x" << std::hex
<< gift
4830 << " min_alloc_size 0x" << min_alloc_size
<< std::dec
<< dendl
;
4832 assert(0 == "allocate failed, wtf");
4835 for (auto& p
: exts
) {
4836 bluestore_pextent_t e
= bluestore_pextent_t(p
);
4837 dout(1) << __func__
<< " gifting " << e
<< " to bluefs" << dendl
;
4838 extents
->push_back(e
);
4845 // reclaim from bluefs?
4847 // round up to alloc size
4848 reclaim
= P2ROUNDUP(reclaim
, cct
->_conf
->bluefs_alloc_size
);
4850 // hard cap to fit into 32 bits
4851 reclaim
= MIN(reclaim
, 1ull<<31);
4852 dout(10) << __func__
<< " reclaiming " << reclaim
4853 << " (" << pretty_si_t(reclaim
) << ")" << dendl
;
4855 while (reclaim
> 0) {
4856 // NOTE: this will block and do IO.
4857 AllocExtentVector extents
;
4858 int r
= bluefs
->reclaim_blocks(bluefs_shared_bdev
, reclaim
,
4861 derr
<< __func__
<< " failed to reclaim space from bluefs"
4865 for (auto e
: extents
) {
4866 bluefs_extents
.erase(e
.offset
, e
.length
);
4867 bluefs_extents_reclaiming
.insert(e
.offset
, e
.length
);
4868 reclaim
-= e
.length
;
4878 void BlueStore::_commit_bluefs_freespace(
4879 const PExtentVector
& bluefs_gift_extents
)
4881 dout(10) << __func__
<< dendl
;
4882 for (auto& p
: bluefs_gift_extents
) {
4883 bluefs
->add_block_extent(bluefs_shared_bdev
, p
.offset
, p
.length
);
4887 int BlueStore::_open_collections(int *errors
)
4889 assert(coll_map
.empty());
4890 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_COLL
);
4891 for (it
->upper_bound(string());
4895 if (cid
.parse(it
->key())) {
4899 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
4901 bufferlist bl
= it
->value();
4902 bufferlist::iterator p
= bl
.begin();
4904 ::decode(c
->cnode
, p
);
4905 } catch (buffer::error
& e
) {
4906 derr
<< __func__
<< " failed to decode cnode, key:"
4907 << pretty_binary_string(it
->key()) << dendl
;
4910 dout(20) << __func__
<< " opened " << cid
<< " " << c
<< dendl
;
4913 derr
<< __func__
<< " unrecognized collection " << it
->key() << dendl
;
4921 void BlueStore::_open_statfs()
4924 int r
= db
->get(PREFIX_STAT
, "bluestore_statfs", &bl
);
4926 if (size_t(bl
.length()) >= sizeof(vstatfs
.values
)) {
4927 auto it
= bl
.begin();
4930 dout(10) << __func__
<< " store_statfs is corrupt, using empty" << dendl
;
4934 dout(10) << __func__
<< " store_statfs missed, using empty" << dendl
;
4938 int BlueStore::_setup_block_symlink_or_file(
4944 dout(20) << __func__
<< " name " << name
<< " path " << epath
4945 << " size " << size
<< " create=" << (int)create
<< dendl
;
4950 if (epath
.length()) {
4951 r
= ::symlinkat(epath
.c_str(), path_fd
, name
.c_str());
4954 derr
<< __func__
<< " failed to create " << name
<< " symlink to "
4955 << epath
<< ": " << cpp_strerror(r
) << dendl
;
4959 if (!epath
.compare(0, strlen(SPDK_PREFIX
), SPDK_PREFIX
)) {
4960 int fd
= ::openat(path_fd
, epath
.c_str(), flags
, 0644);
4963 derr
<< __func__
<< " failed to open " << epath
<< " file: "
4964 << cpp_strerror(r
) << dendl
;
4967 string serial_number
= epath
.substr(strlen(SPDK_PREFIX
));
4968 r
= ::write(fd
, serial_number
.c_str(), serial_number
.size());
4969 assert(r
== (int)serial_number
.size());
4970 dout(1) << __func__
<< " created " << name
<< " symlink to "
4972 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4976 int fd
= ::openat(path_fd
, name
.c_str(), flags
, 0644);
4978 // block file is present
4980 int r
= ::fstat(fd
, &st
);
4982 S_ISREG(st
.st_mode
) && // if it is a regular file
4983 st
.st_size
== 0) { // and is 0 bytes
4984 r
= ::ftruncate(fd
, size
);
4987 derr
<< __func__
<< " failed to resize " << name
<< " file to "
4988 << size
<< ": " << cpp_strerror(r
) << dendl
;
4989 VOID_TEMP_FAILURE_RETRY(::close(fd
));
4993 if (cct
->_conf
->bluestore_block_preallocate_file
) {
4994 #ifdef HAVE_POSIX_FALLOCATE
4995 r
= ::posix_fallocate(fd
, 0, size
);
4997 derr
<< __func__
<< " failed to prefallocate " << name
<< " file to "
4998 << size
<< ": " << cpp_strerror(r
) << dendl
;
4999 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5003 char data
[1024*128];
5004 for (uint64_t off
= 0; off
< size
; off
+= sizeof(data
)) {
5005 if (off
+ sizeof(data
) > size
)
5006 r
= ::write(fd
, data
, size
- off
);
5008 r
= ::write(fd
, data
, sizeof(data
));
5011 derr
<< __func__
<< " failed to prefallocate w/ write " << name
<< " file to "
5012 << size
<< ": " << cpp_strerror(r
) << dendl
;
5013 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5019 dout(1) << __func__
<< " resized " << name
<< " file to "
5020 << pretty_si_t(size
) << "B" << dendl
;
5022 VOID_TEMP_FAILURE_RETRY(::close(fd
));
5026 derr
<< __func__
<< " failed to open " << name
<< " file: "
5027 << cpp_strerror(r
) << dendl
;
5035 int BlueStore::mkfs()
5037 dout(1) << __func__
<< " path " << path
<< dendl
;
5043 r
= read_meta("mkfs_done", &done
);
5045 dout(1) << __func__
<< " already created" << dendl
;
5046 if (cct
->_conf
->bluestore_fsck_on_mkfs
) {
5047 r
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5049 derr
<< __func__
<< " fsck found fatal error: " << cpp_strerror(r
)
5054 derr
<< __func__
<< " fsck found " << r
<< " errors" << dendl
;
5058 return r
; // idempotent
5064 r
= read_meta("type", &type
);
5066 if (type
!= "bluestore") {
5067 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5071 r
= write_meta("type", "bluestore");
5077 freelist_type
= "bitmap";
5083 r
= _open_fsid(true);
5089 goto out_close_fsid
;
5091 r
= _read_fsid(&old_fsid
);
5092 if (r
< 0 || old_fsid
.is_zero()) {
5093 if (fsid
.is_zero()) {
5094 fsid
.generate_random();
5095 dout(1) << __func__
<< " generated fsid " << fsid
<< dendl
;
5097 dout(1) << __func__
<< " using provided fsid " << fsid
<< dendl
;
5099 // we'll write it later.
5101 if (!fsid
.is_zero() && fsid
!= old_fsid
) {
5102 derr
<< __func__
<< " on-disk fsid " << old_fsid
5103 << " != provided " << fsid
<< dendl
;
5105 goto out_close_fsid
;
5110 r
= _setup_block_symlink_or_file("block", cct
->_conf
->bluestore_block_path
,
5111 cct
->_conf
->bluestore_block_size
,
5112 cct
->_conf
->bluestore_block_create
);
5114 goto out_close_fsid
;
5115 if (cct
->_conf
->bluestore_bluefs
) {
5116 r
= _setup_block_symlink_or_file("block.wal", cct
->_conf
->bluestore_block_wal_path
,
5117 cct
->_conf
->bluestore_block_wal_size
,
5118 cct
->_conf
->bluestore_block_wal_create
);
5120 goto out_close_fsid
;
5121 r
= _setup_block_symlink_or_file("block.db", cct
->_conf
->bluestore_block_db_path
,
5122 cct
->_conf
->bluestore_block_db_size
,
5123 cct
->_conf
->bluestore_block_db_create
);
5125 goto out_close_fsid
;
5128 r
= _open_bdev(true);
5130 goto out_close_fsid
;
5134 goto out_close_bdev
;
5141 KeyValueDB::Transaction t
= db
->get_transaction();
5144 ::encode((uint64_t)0, bl
);
5145 t
->set(PREFIX_SUPER
, "nid_max", bl
);
5146 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
5149 // choose min_alloc_size
5150 if (cct
->_conf
->bluestore_min_alloc_size
) {
5151 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size
;
5154 if (bdev
->is_rotational()) {
5155 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_hdd
;
5157 min_alloc_size
= cct
->_conf
->bluestore_min_alloc_size_ssd
;
5161 // make sure min_alloc_size is power of 2 aligned.
5162 if (!ISP2(min_alloc_size
)) {
5163 derr
<< __func__
<< " min_alloc_size 0x"
5164 << std::hex
<< min_alloc_size
<< std::dec
5165 << " is not power of 2 aligned!"
5173 ::encode((uint64_t)min_alloc_size
, bl
);
5174 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
5177 ondisk_format
= latest_ondisk_format
;
5178 _prepare_ondisk_format_super(t
);
5179 db
->submit_transaction_sync(t
);
5183 r
= write_meta("kv_backend", cct
->_conf
->bluestore_kvbackend
);
5187 r
= write_meta("bluefs", stringify((int)cct
->_conf
->bluestore_bluefs
));
5191 if (fsid
!= old_fsid
) {
5194 derr
<< __func__
<< " error writing fsid: " << cpp_strerror(r
) << dendl
;
5211 cct
->_conf
->bluestore_fsck_on_mkfs
) {
5212 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mkfs_deep
);
5216 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5222 // indicate success by writing the 'mkfs_done' file
5223 r
= write_meta("mkfs_done", "yes");
5227 derr
<< __func__
<< " failed, " << cpp_strerror(r
) << dendl
;
5229 dout(0) << __func__
<< " success" << dendl
;
5234 void BlueStore::set_cache_shards(unsigned num
)
5236 dout(10) << __func__
<< " " << num
<< dendl
;
5237 size_t old
= cache_shards
.size();
5239 cache_shards
.resize(num
);
5240 for (unsigned i
= old
; i
< num
; ++i
) {
5241 cache_shards
[i
] = Cache::create(cct
, cct
->_conf
->bluestore_cache_type
,
5246 int BlueStore::_mount(bool kv_only
)
5248 dout(1) << __func__
<< " path " << path
<< dendl
;
5252 int r
= read_meta("type", &type
);
5254 derr
<< __func__
<< " failed to load os-type: " << cpp_strerror(r
)
5259 if (type
!= "bluestore") {
5260 derr
<< __func__
<< " expected bluestore, but type is " << type
<< dendl
;
5265 if (cct
->_conf
->bluestore_fsck_on_mount
) {
5266 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_mount_deep
);
5270 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5275 int r
= _open_path();
5278 r
= _open_fsid(false);
5282 r
= _read_fsid(&fsid
);
5290 r
= _open_bdev(false);
5294 r
= _open_db(false);
5301 r
= _open_super_meta();
5305 r
= _open_fm(false);
5313 r
= _open_collections();
5317 r
= _reload_logger();
5322 r
= _reconcile_bluefs_freespace();
5329 r
= _deferred_replay();
5333 mempool_thread
.init();
5358 int BlueStore::umount()
5361 dout(1) << __func__
<< dendl
;
5364 _osr_unregister_all();
5366 mempool_thread
.shutdown();
5368 dout(20) << __func__
<< " stopping kv thread" << dendl
;
5370 _reap_collections();
5372 dout(20) << __func__
<< " closing" << dendl
;
5382 if (cct
->_conf
->bluestore_fsck_on_umount
) {
5383 int rc
= fsck(cct
->_conf
->bluestore_fsck_on_umount_deep
);
5387 derr
<< __func__
<< " fsck found " << rc
<< " errors" << dendl
;
5394 static void apply(uint64_t off
,
5396 uint64_t granularity
,
5397 BlueStore::mempool_dynamic_bitset
&bitset
,
5399 std::function
<void(uint64_t,
5400 BlueStore::mempool_dynamic_bitset
&)> f
) {
5401 auto end
= ROUND_UP_TO(off
+ len
, granularity
);
5403 uint64_t pos
= off
/ granularity
;
5409 int BlueStore::_fsck_check_extents(
5410 const ghobject_t
& oid
,
5411 const PExtentVector
& extents
,
5413 mempool_dynamic_bitset
&used_blocks
,
5414 store_statfs_t
& expected_statfs
)
5416 dout(30) << __func__
<< " oid " << oid
<< " extents " << extents
<< dendl
;
5418 for (auto e
: extents
) {
5421 expected_statfs
.allocated
+= e
.length
;
5423 expected_statfs
.compressed_allocated
+= e
.length
;
5425 bool already
= false;
5427 e
.offset
, e
.length
, block_size
, used_blocks
, __func__
,
5428 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5435 derr
<< " " << oid
<< " extent " << e
5436 << " or a subset is already allocated" << dendl
;
5439 if (e
.end() > bdev
->get_size()) {
5440 derr
<< " " << oid
<< " extent " << e
5441 << " past end of block device" << dendl
;
5448 int BlueStore::fsck(bool deep
)
5450 dout(1) << __func__
<< (deep
? " (deep)" : " (shallow)") << " start" << dendl
;
5453 typedef btree::btree_set
<
5454 uint64_t,std::less
<uint64_t>,
5455 mempool::bluestore_fsck::pool_allocator
<uint64_t>> uint64_t_btree_t
;
5456 uint64_t_btree_t used_nids
;
5457 uint64_t_btree_t used_omap_head
;
5458 uint64_t_btree_t used_sbids
;
5460 mempool_dynamic_bitset used_blocks
;
5461 KeyValueDB::Iterator it
;
5462 store_statfs_t expected_statfs
, actual_statfs
;
5464 list
<ghobject_t
> oids
;
5466 bluestore_extent_ref_map_t ref_map
;
5469 mempool::bluestore_fsck::map
<uint64_t,sb_info_t
> sb_info
;
5471 uint64_t num_objects
= 0;
5472 uint64_t num_extents
= 0;
5473 uint64_t num_blobs
= 0;
5474 uint64_t num_spanning_blobs
= 0;
5475 uint64_t num_shared_blobs
= 0;
5476 uint64_t num_sharded_objects
= 0;
5477 uint64_t num_object_shards
= 0;
5479 utime_t start
= ceph_clock_now();
5481 int r
= _open_path();
5484 r
= _open_fsid(false);
5488 r
= _read_fsid(&fsid
);
5496 r
= _open_bdev(false);
5500 r
= _open_db(false);
5504 r
= _open_super_meta();
5508 r
= _open_fm(false);
5516 r
= _open_collections(&errors
);
5520 mempool_thread
.init();
5522 // we need finishers and kv_{sync,finalize}_thread *just* for replay
5524 r
= _deferred_replay();
5529 used_blocks
.resize(bdev
->get_size() / block_size
);
5531 0, SUPER_RESERVED
, block_size
, used_blocks
, "0~SUPER_RESERVED",
5532 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5538 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
5540 e
.get_start(), e
.get_len(), block_size
, used_blocks
, "bluefs",
5541 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5554 // get expected statfs; fill unaffected fields to be able to compare
5556 statfs(&actual_statfs
);
5557 expected_statfs
.total
= actual_statfs
.total
;
5558 expected_statfs
.available
= actual_statfs
.available
;
5561 dout(1) << __func__
<< " walking object keyspace" << dendl
;
5562 it
= db
->get_iterator(PREFIX_OBJ
);
5566 mempool::bluestore_fsck::list
<string
> expecting_shards
;
5567 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5568 if (g_conf
->bluestore_debug_fsck_abort
) {
5571 dout(30) << " key " << pretty_binary_string(it
->key()) << dendl
;
5572 if (is_extent_shard_key(it
->key())) {
5573 while (!expecting_shards
.empty() &&
5574 expecting_shards
.front() < it
->key()) {
5575 derr
<< __func__
<< " error: missing shard key "
5576 << pretty_binary_string(expecting_shards
.front())
5579 expecting_shards
.pop_front();
5581 if (!expecting_shards
.empty() &&
5582 expecting_shards
.front() == it
->key()) {
5584 expecting_shards
.pop_front();
5590 get_key_extent_shard(it
->key(), &okey
, &offset
);
5591 derr
<< __func__
<< " error: stray shard 0x" << std::hex
<< offset
5592 << std::dec
<< dendl
;
5593 if (expecting_shards
.empty()) {
5594 derr
<< __func__
<< " error: " << pretty_binary_string(it
->key())
5595 << " is unexpected" << dendl
;
5599 while (expecting_shards
.front() > it
->key()) {
5600 derr
<< __func__
<< " error: saw " << pretty_binary_string(it
->key())
5602 derr
<< __func__
<< " error: exp "
5603 << pretty_binary_string(expecting_shards
.front()) << dendl
;
5605 expecting_shards
.pop_front();
5606 if (expecting_shards
.empty()) {
5614 int r
= get_key_object(it
->key(), &oid
);
5616 derr
<< __func__
<< " error: bad object key "
5617 << pretty_binary_string(it
->key()) << dendl
;
5622 oid
.shard_id
!= pgid
.shard
||
5623 oid
.hobj
.pool
!= (int64_t)pgid
.pool() ||
5624 !c
->contains(oid
)) {
5626 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
=
5628 p
!= coll_map
.end();
5630 if (p
->second
->contains(oid
)) {
5636 derr
<< __func__
<< " error: stray object " << oid
5637 << " not owned by any collection" << dendl
;
5641 c
->cid
.is_pg(&pgid
);
5642 dout(20) << __func__
<< " collection " << c
->cid
<< dendl
;
5645 if (!expecting_shards
.empty()) {
5646 for (auto &k
: expecting_shards
) {
5647 derr
<< __func__
<< " error: missing shard key "
5648 << pretty_binary_string(k
) << dendl
;
5651 expecting_shards
.clear();
5654 dout(10) << __func__
<< " " << oid
<< dendl
;
5655 RWLock::RLocker
l(c
->lock
);
5656 OnodeRef o
= c
->get_onode(oid
, false);
5658 if (o
->onode
.nid
> nid_max
) {
5659 derr
<< __func__
<< " error: " << oid
<< " nid " << o
->onode
.nid
5660 << " > nid_max " << nid_max
<< dendl
;
5663 if (used_nids
.count(o
->onode
.nid
)) {
5664 derr
<< __func__
<< " error: " << oid
<< " nid " << o
->onode
.nid
5665 << " already in use" << dendl
;
5667 continue; // go for next object
5669 used_nids
.insert(o
->onode
.nid
);
5672 num_spanning_blobs
+= o
->extent_map
.spanning_blob_map
.size();
5673 o
->extent_map
.fault_range(db
, 0, OBJECT_MAX_SIZE
);
5676 if (!o
->extent_map
.shards
.empty()) {
5677 ++num_sharded_objects
;
5678 num_object_shards
+= o
->extent_map
.shards
.size();
5680 for (auto& s
: o
->extent_map
.shards
) {
5681 dout(20) << __func__
<< " shard " << *s
.shard_info
<< dendl
;
5682 expecting_shards
.push_back(string());
5683 get_extent_shard_key(o
->key
, s
.shard_info
->offset
,
5684 &expecting_shards
.back());
5685 if (s
.shard_info
->offset
>= o
->onode
.size
) {
5686 derr
<< __func__
<< " error: " << oid
<< " shard 0x" << std::hex
5687 << s
.shard_info
->offset
<< " past EOF at 0x" << o
->onode
.size
5688 << std::dec
<< dendl
;
5693 map
<BlobRef
,bluestore_blob_t::unused_t
> referenced
;
5695 mempool::bluestore_fsck::map
<BlobRef
,
5696 bluestore_blob_use_tracker_t
> ref_map
;
5697 for (auto& l
: o
->extent_map
.extent_map
) {
5698 dout(20) << __func__
<< " " << l
<< dendl
;
5699 if (l
.logical_offset
< pos
) {
5700 derr
<< __func__
<< " error: " << oid
<< " lextent at 0x"
5701 << std::hex
<< l
.logical_offset
5702 << " overlaps with the previous, which ends at 0x" << pos
5703 << std::dec
<< dendl
;
5706 if (o
->extent_map
.spans_shard(l
.logical_offset
, l
.length
)) {
5707 derr
<< __func__
<< " error: " << oid
<< " lextent at 0x"
5708 << std::hex
<< l
.logical_offset
<< "~" << l
.length
5709 << " spans a shard boundary"
5710 << std::dec
<< dendl
;
5713 pos
= l
.logical_offset
+ l
.length
;
5714 expected_statfs
.stored
+= l
.length
;
5716 const bluestore_blob_t
& blob
= l
.blob
->get_blob();
5718 auto& ref
= ref_map
[l
.blob
];
5719 if (ref
.is_empty()) {
5720 uint32_t min_release_size
= blob
.get_release_size(min_alloc_size
);
5721 uint32_t l
= blob
.get_logical_length();
5722 ref
.init(l
, min_release_size
);
5728 if (blob
.has_unused()) {
5729 auto p
= referenced
.find(l
.blob
);
5730 bluestore_blob_t::unused_t
*pu
;
5731 if (p
== referenced
.end()) {
5732 pu
= &referenced
[l
.blob
];
5736 uint64_t blob_len
= blob
.get_logical_length();
5737 assert((blob_len
% (sizeof(*pu
)*8)) == 0);
5738 assert(l
.blob_offset
+ l
.length
<= blob_len
);
5739 uint64_t chunk_size
= blob_len
/ (sizeof(*pu
)*8);
5740 uint64_t start
= l
.blob_offset
/ chunk_size
;
5742 ROUND_UP_TO(l
.blob_offset
+ l
.length
, chunk_size
) / chunk_size
;
5743 for (auto i
= start
; i
< end
; ++i
) {
5748 for (auto &i
: referenced
) {
5749 dout(20) << __func__
<< " referenced 0x" << std::hex
<< i
.second
5750 << std::dec
<< " for " << *i
.first
<< dendl
;
5751 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5752 if (i
.second
& blob
.unused
) {
5753 derr
<< __func__
<< " error: " << oid
<< " blob claims unused 0x"
5754 << std::hex
<< blob
.unused
5755 << " but extents reference 0x" << i
.second
5756 << " on blob " << *i
.first
<< dendl
;
5759 if (blob
.has_csum()) {
5760 uint64_t blob_len
= blob
.get_logical_length();
5761 uint64_t unused_chunk_size
= blob_len
/ (sizeof(blob
.unused
)*8);
5762 unsigned csum_count
= blob
.get_csum_count();
5763 unsigned csum_chunk_size
= blob
.get_csum_chunk_size();
5764 for (unsigned p
= 0; p
< csum_count
; ++p
) {
5765 unsigned pos
= p
* csum_chunk_size
;
5766 unsigned firstbit
= pos
/ unused_chunk_size
; // [firstbit,lastbit]
5767 unsigned lastbit
= (pos
+ csum_chunk_size
- 1) / unused_chunk_size
;
5768 unsigned mask
= 1u << firstbit
;
5769 for (unsigned b
= firstbit
+ 1; b
<= lastbit
; ++b
) {
5772 if ((blob
.unused
& mask
) == mask
) {
5773 // this csum chunk region is marked unused
5774 if (blob
.get_csum_item(p
) != 0) {
5775 derr
<< __func__
<< " error: " << oid
5776 << " blob claims csum chunk 0x" << std::hex
<< pos
5777 << "~" << csum_chunk_size
5778 << " is unused (mask 0x" << mask
<< " of unused 0x"
5779 << blob
.unused
<< ") but csum is non-zero 0x"
5780 << blob
.get_csum_item(p
) << std::dec
<< " on blob "
5781 << *i
.first
<< dendl
;
5788 for (auto &i
: ref_map
) {
5790 const bluestore_blob_t
& blob
= i
.first
->get_blob();
5791 bool equal
= i
.first
->get_blob_use_tracker().equal(i
.second
);
5793 derr
<< __func__
<< " error: " << oid
<< " blob " << *i
.first
5794 << " doesn't match expected ref_map " << i
.second
<< dendl
;
5797 if (blob
.is_compressed()) {
5798 expected_statfs
.compressed
+= blob
.get_compressed_payload_length();
5799 expected_statfs
.compressed_original
+=
5800 i
.first
->get_referenced_bytes();
5802 if (blob
.is_shared()) {
5803 if (i
.first
->shared_blob
->get_sbid() > blobid_max
) {
5804 derr
<< __func__
<< " error: " << oid
<< " blob " << blob
5805 << " sbid " << i
.first
->shared_blob
->get_sbid() << " > blobid_max "
5806 << blobid_max
<< dendl
;
5808 } else if (i
.first
->shared_blob
->get_sbid() == 0) {
5809 derr
<< __func__
<< " error: " << oid
<< " blob " << blob
5810 << " marked as shared but has uninitialized sbid"
5814 sb_info_t
& sbi
= sb_info
[i
.first
->shared_blob
->get_sbid()];
5815 sbi
.sb
= i
.first
->shared_blob
;
5816 sbi
.oids
.push_back(oid
);
5817 sbi
.compressed
= blob
.is_compressed();
5818 for (auto e
: blob
.get_extents()) {
5820 sbi
.ref_map
.get(e
.offset
, e
.length
);
5824 errors
+= _fsck_check_extents(oid
, blob
.get_extents(),
5825 blob
.is_compressed(),
5832 int r
= _do_read(c
.get(), o
, 0, o
->onode
.size
, bl
, 0);
5835 derr
<< __func__
<< " error: " << oid
<< " error during read: "
5836 << cpp_strerror(r
) << dendl
;
5840 if (o
->onode
.has_omap()) {
5841 if (used_omap_head
.count(o
->onode
.nid
)) {
5842 derr
<< __func__
<< " error: " << oid
<< " omap_head " << o
->onode
.nid
5843 << " already in use" << dendl
;
5846 used_omap_head
.insert(o
->onode
.nid
);
5851 dout(1) << __func__
<< " checking shared_blobs" << dendl
;
5852 it
= db
->get_iterator(PREFIX_SHARED_BLOB
);
5854 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5855 string key
= it
->key();
5857 if (get_key_shared_blob(key
, &sbid
)) {
5858 derr
<< __func__
<< " error: bad key '" << key
5859 << "' in shared blob namespace" << dendl
;
5863 auto p
= sb_info
.find(sbid
);
5864 if (p
== sb_info
.end()) {
5865 derr
<< __func__
<< " error: found stray shared blob data for sbid 0x"
5866 << std::hex
<< sbid
<< std::dec
<< dendl
;
5870 sb_info_t
& sbi
= p
->second
;
5871 bluestore_shared_blob_t
shared_blob(sbid
);
5872 bufferlist bl
= it
->value();
5873 bufferlist::iterator blp
= bl
.begin();
5874 ::decode(shared_blob
, blp
);
5875 dout(20) << __func__
<< " " << *sbi
.sb
<< " " << shared_blob
<< dendl
;
5876 if (shared_blob
.ref_map
!= sbi
.ref_map
) {
5877 derr
<< __func__
<< " error: shared blob 0x" << std::hex
<< sbid
5878 << std::dec
<< " ref_map " << shared_blob
.ref_map
5879 << " != expected " << sbi
.ref_map
<< dendl
;
5882 PExtentVector extents
;
5883 for (auto &r
: shared_blob
.ref_map
.ref_map
) {
5884 extents
.emplace_back(bluestore_pextent_t(r
.first
, r
.second
.length
));
5886 errors
+= _fsck_check_extents(p
->second
.oids
.front(),
5888 p
->second
.compressed
,
5889 used_blocks
, expected_statfs
);
5894 for (auto &p
: sb_info
) {
5895 derr
<< __func__
<< " error: shared_blob 0x" << p
.first
5896 << " key is missing (" << *p
.second
.sb
<< ")" << dendl
;
5899 if (!(actual_statfs
== expected_statfs
)) {
5900 derr
<< __func__
<< " error: actual " << actual_statfs
5901 << " != expected " << expected_statfs
<< dendl
;
5905 dout(1) << __func__
<< " checking for stray omap data" << dendl
;
5906 it
= db
->get_iterator(PREFIX_OMAP
);
5908 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5910 _key_decode_u64(it
->key().c_str(), &omap_head
);
5911 if (used_omap_head
.count(omap_head
) == 0) {
5912 derr
<< __func__
<< " error: found stray omap data on omap_head "
5913 << omap_head
<< dendl
;
5919 dout(1) << __func__
<< " checking deferred events" << dendl
;
5920 it
= db
->get_iterator(PREFIX_DEFERRED
);
5922 for (it
->lower_bound(string()); it
->valid(); it
->next()) {
5923 bufferlist bl
= it
->value();
5924 bufferlist::iterator p
= bl
.begin();
5925 bluestore_deferred_transaction_t wt
;
5928 } catch (buffer::error
& e
) {
5929 derr
<< __func__
<< " error: failed to decode deferred txn "
5930 << pretty_binary_string(it
->key()) << dendl
;
5934 dout(20) << __func__
<< " deferred " << wt
.seq
5935 << " ops " << wt
.ops
.size()
5936 << " released 0x" << std::hex
<< wt
.released
<< std::dec
<< dendl
;
5937 for (auto e
= wt
.released
.begin(); e
!= wt
.released
.end(); ++e
) {
5939 e
.get_start(), e
.get_len(), block_size
, used_blocks
, "deferred",
5940 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5948 dout(1) << __func__
<< " checking freelist vs allocated" << dendl
;
5950 // remove bluefs_extents from used set since the freelist doesn't
5951 // know they are allocated.
5952 for (auto e
= bluefs_extents
.begin(); e
!= bluefs_extents
.end(); ++e
) {
5954 e
.get_start(), e
.get_len(), block_size
, used_blocks
, "bluefs_extents",
5955 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5960 fm
->enumerate_reset();
5961 uint64_t offset
, length
;
5962 while (fm
->enumerate_next(&offset
, &length
)) {
5963 bool intersects
= false;
5965 offset
, length
, block_size
, used_blocks
, "free",
5966 [&](uint64_t pos
, mempool_dynamic_bitset
&bs
) {
5975 derr
<< __func__
<< " error: free extent 0x" << std::hex
<< offset
5976 << "~" << length
<< std::dec
5977 << " intersects allocated blocks" << dendl
;
5981 fm
->enumerate_reset();
5982 size_t count
= used_blocks
.count();
5983 if (used_blocks
.size() != count
) {
5984 assert(used_blocks
.size() > count
);
5985 derr
<< __func__
<< " error: leaked some space;"
5986 << (used_blocks
.size() - count
) * min_alloc_size
5987 << " bytes leaked" << dendl
;
5993 mempool_thread
.shutdown();
6000 it
.reset(); // before db is closed
6009 // fatal errors take precedence
6013 dout(2) << __func__
<< " " << num_objects
<< " objects, "
6014 << num_sharded_objects
<< " of them sharded. "
6016 dout(2) << __func__
<< " " << num_extents
<< " extents to "
6017 << num_blobs
<< " blobs, "
6018 << num_spanning_blobs
<< " spanning, "
6019 << num_shared_blobs
<< " shared."
6022 utime_t duration
= ceph_clock_now() - start
;
6023 dout(1) << __func__
<< " finish with " << errors
<< " errors in "
6024 << duration
<< " seconds" << dendl
;
6028 void BlueStore::collect_metadata(map
<string
,string
> *pm
)
6030 dout(10) << __func__
<< dendl
;
6031 bdev
->collect_metadata("bluestore_bdev_", pm
);
6033 (*pm
)["bluefs"] = "1";
6034 (*pm
)["bluefs_single_shared_device"] = stringify((int)bluefs_single_shared_device
);
6035 bluefs
->collect_metadata(pm
);
6037 (*pm
)["bluefs"] = "0";
6041 int BlueStore::statfs(struct store_statfs_t
*buf
)
6044 buf
->total
= bdev
->get_size();
6045 buf
->available
= alloc
->get_free();
6048 // part of our shared device is "free" according to BlueFS
6049 // Don't include bluestore_bluefs_min because that space can't
6050 // be used for any other purpose.
6051 buf
->available
+= bluefs
->get_free(bluefs_shared_bdev
) - cct
->_conf
->bluestore_bluefs_min
;
6053 // include dedicated db, too, if that isn't the shared device.
6054 if (bluefs_shared_bdev
!= BlueFS::BDEV_DB
) {
6055 buf
->total
+= bluefs
->get_total(BlueFS::BDEV_DB
);
6060 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
6062 buf
->allocated
= vstatfs
.allocated();
6063 buf
->stored
= vstatfs
.stored();
6064 buf
->compressed
= vstatfs
.compressed();
6065 buf
->compressed_original
= vstatfs
.compressed_original();
6066 buf
->compressed_allocated
= vstatfs
.compressed_allocated();
6069 dout(20) << __func__
<< *buf
<< dendl
;
6076 BlueStore::CollectionRef
BlueStore::_get_collection(const coll_t
& cid
)
6078 RWLock::RLocker
l(coll_lock
);
6079 ceph::unordered_map
<coll_t
,CollectionRef
>::iterator cp
= coll_map
.find(cid
);
6080 if (cp
== coll_map
.end())
6081 return CollectionRef();
6085 void BlueStore::_queue_reap_collection(CollectionRef
& c
)
6087 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6088 std::lock_guard
<std::mutex
> l(reap_lock
);
6089 removed_collections
.push_back(c
);
6092 void BlueStore::_reap_collections()
6094 list
<CollectionRef
> removed_colls
;
6096 std::lock_guard
<std::mutex
> l(reap_lock
);
6097 removed_colls
.swap(removed_collections
);
6100 bool all_reaped
= true;
6102 for (list
<CollectionRef
>::iterator p
= removed_colls
.begin();
6103 p
!= removed_colls
.end();
6105 CollectionRef c
= *p
;
6106 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< dendl
;
6107 if (c
->onode_map
.map_any([&](OnodeRef o
) {
6109 if (o
->flushing_count
.load()) {
6110 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " " << o
->oid
6111 << " flush_txns " << o
->flushing_count
<< dendl
;
6119 c
->onode_map
.clear();
6120 dout(10) << __func__
<< " " << c
<< " " << c
->cid
<< " done" << dendl
;
6124 dout(10) << __func__
<< " all reaped" << dendl
;
6128 void BlueStore::_update_cache_logger()
6130 uint64_t num_onodes
= 0;
6131 uint64_t num_extents
= 0;
6132 uint64_t num_blobs
= 0;
6133 uint64_t num_buffers
= 0;
6134 uint64_t num_buffer_bytes
= 0;
6135 for (auto c
: cache_shards
) {
6136 c
->add_stats(&num_onodes
, &num_extents
, &num_blobs
,
6137 &num_buffers
, &num_buffer_bytes
);
6139 logger
->set(l_bluestore_onodes
, num_onodes
);
6140 logger
->set(l_bluestore_extents
, num_extents
);
6141 logger
->set(l_bluestore_blobs
, num_blobs
);
6142 logger
->set(l_bluestore_buffers
, num_buffers
);
6143 logger
->set(l_bluestore_buffer_bytes
, num_buffer_bytes
);
6149 ObjectStore::CollectionHandle
BlueStore::open_collection(const coll_t
& cid
)
6151 return _get_collection(cid
);
6154 bool BlueStore::exists(const coll_t
& cid
, const ghobject_t
& oid
)
6156 CollectionHandle c
= _get_collection(cid
);
6159 return exists(c
, oid
);
6162 bool BlueStore::exists(CollectionHandle
&c_
, const ghobject_t
& oid
)
6164 Collection
*c
= static_cast<Collection
*>(c_
.get());
6165 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
6172 RWLock::RLocker
l(c
->lock
);
6173 OnodeRef o
= c
->get_onode(oid
, false);
6174 if (!o
|| !o
->exists
)
6181 int BlueStore::stat(
6183 const ghobject_t
& oid
,
6187 CollectionHandle c
= _get_collection(cid
);
6190 return stat(c
, oid
, st
, allow_eio
);
6193 int BlueStore::stat(
6194 CollectionHandle
&c_
,
6195 const ghobject_t
& oid
,
6199 Collection
*c
= static_cast<Collection
*>(c_
.get());
6202 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
6205 RWLock::RLocker
l(c
->lock
);
6206 OnodeRef o
= c
->get_onode(oid
, false);
6207 if (!o
|| !o
->exists
)
6209 st
->st_size
= o
->onode
.size
;
6210 st
->st_blksize
= 4096;
6211 st
->st_blocks
= (st
->st_size
+ st
->st_blksize
- 1) / st
->st_blksize
;
6216 if (_debug_mdata_eio(oid
)) {
6218 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6222 int BlueStore::set_collection_opts(
6224 const pool_opts_t
& opts
)
6226 CollectionHandle ch
= _get_collection(cid
);
6229 Collection
*c
= static_cast<Collection
*>(ch
.get());
6230 dout(15) << __func__
<< " " << cid
<< " options " << opts
<< dendl
;
6233 RWLock::WLocker
l(c
->lock
);
6234 c
->pool_opts
= opts
;
6238 int BlueStore::read(
6240 const ghobject_t
& oid
,
6246 CollectionHandle c
= _get_collection(cid
);
6249 return read(c
, oid
, offset
, length
, bl
, op_flags
);
6252 int BlueStore::read(
6253 CollectionHandle
&c_
,
6254 const ghobject_t
& oid
,
6260 utime_t start
= ceph_clock_now();
6261 Collection
*c
= static_cast<Collection
*>(c_
.get());
6262 const coll_t
&cid
= c
->get_cid();
6263 dout(15) << __func__
<< " " << cid
<< " " << oid
6264 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6272 RWLock::RLocker
l(c
->lock
);
6273 utime_t start1
= ceph_clock_now();
6274 OnodeRef o
= c
->get_onode(oid
, false);
6275 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start1
);
6276 if (!o
|| !o
->exists
) {
6281 if (offset
== length
&& offset
== 0)
6282 length
= o
->onode
.size
;
6284 r
= _do_read(c
, o
, offset
, length
, bl
, op_flags
);
6288 if (r
== 0 && _debug_data_eio(oid
)) {
6290 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6291 } else if (cct
->_conf
->bluestore_debug_random_read_err
&&
6292 (rand() % (int)(cct
->_conf
->bluestore_debug_random_read_err
* 100.0)) == 0) {
6293 dout(0) << __func__
<< ": inject random EIO" << dendl
;
6296 dout(10) << __func__
<< " " << cid
<< " " << oid
6297 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
6298 << " = " << r
<< dendl
;
6299 logger
->tinc(l_bluestore_read_lat
, ceph_clock_now() - start
);
6303 // --------------------------------------------------------
6304 // intermediate data structures used while reading
6306 uint64_t logical_offset
;
6307 uint64_t blob_xoffset
; //region offset within the blob
6311 // used later in read process
6315 region_t(uint64_t offset
, uint64_t b_offs
, uint64_t len
)
6316 : logical_offset(offset
),
6317 blob_xoffset(b_offs
),
6319 region_t(const region_t
& from
)
6320 : logical_offset(from
.logical_offset
),
6321 blob_xoffset(from
.blob_xoffset
),
6322 length(from
.length
){}
6324 friend ostream
& operator<<(ostream
& out
, const region_t
& r
) {
6325 return out
<< "0x" << std::hex
<< r
.logical_offset
<< ":"
6326 << r
.blob_xoffset
<< "~" << r
.length
<< std::dec
;
6330 typedef list
<region_t
> regions2read_t
;
6331 typedef map
<BlueStore::BlobRef
, regions2read_t
> blobs2read_t
;
6333 int BlueStore::_do_read(
6344 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6345 << " size 0x" << o
->onode
.size
<< " (" << std::dec
6346 << o
->onode
.size
<< ")" << dendl
;
6349 if (offset
>= o
->onode
.size
) {
6353 // generally, don't buffer anything, unless the client explicitly requests
6355 bool buffered
= false;
6356 if (op_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
6357 dout(20) << __func__
<< " will do buffered read" << dendl
;
6359 } else if (cct
->_conf
->bluestore_default_buffered_read
&&
6360 (op_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
6361 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
6362 dout(20) << __func__
<< " defaulting to buffered read" << dendl
;
6366 if (offset
+ length
> o
->onode
.size
) {
6367 length
= o
->onode
.size
- offset
;
6370 utime_t start
= ceph_clock_now();
6371 o
->extent_map
.fault_range(db
, offset
, length
);
6372 logger
->tinc(l_bluestore_read_onode_meta_lat
, ceph_clock_now() - start
);
6375 ready_regions_t ready_regions
;
6377 // build blob-wise list to of stuff read (that isn't cached)
6378 blobs2read_t blobs2read
;
6379 unsigned left
= length
;
6380 uint64_t pos
= offset
;
6381 unsigned num_regions
= 0;
6382 auto lp
= o
->extent_map
.seek_lextent(offset
);
6383 while (left
> 0 && lp
!= o
->extent_map
.extent_map
.end()) {
6384 if (pos
< lp
->logical_offset
) {
6385 unsigned hole
= lp
->logical_offset
- pos
;
6389 dout(30) << __func__
<< " hole 0x" << std::hex
<< pos
<< "~" << hole
6390 << std::dec
<< dendl
;
6394 BlobRef bptr
= lp
->blob
;
6395 unsigned l_off
= pos
- lp
->logical_offset
;
6396 unsigned b_off
= l_off
+ lp
->blob_offset
;
6397 unsigned b_len
= std::min(left
, lp
->length
- l_off
);
6399 ready_regions_t cache_res
;
6400 interval_set
<uint32_t> cache_interval
;
6401 bptr
->shared_blob
->bc
.read(
6402 bptr
->shared_blob
->get_cache(), b_off
, b_len
, cache_res
, cache_interval
);
6403 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6404 << " need 0x" << b_off
<< "~" << b_len
6405 << " cache has 0x" << cache_interval
6406 << std::dec
<< dendl
;
6408 auto pc
= cache_res
.begin();
6411 if (pc
!= cache_res
.end() &&
6412 pc
->first
== b_off
) {
6413 l
= pc
->second
.length();
6414 ready_regions
[pos
].claim(pc
->second
);
6415 dout(30) << __func__
<< " use cache 0x" << std::hex
<< pos
<< ": 0x"
6416 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6420 if (pc
!= cache_res
.end()) {
6421 assert(pc
->first
> b_off
);
6422 l
= pc
->first
- b_off
;
6424 dout(30) << __func__
<< " will read 0x" << std::hex
<< pos
<< ": 0x"
6425 << b_off
<< "~" << l
<< std::dec
<< dendl
;
6426 blobs2read
[bptr
].emplace_back(region_t(pos
, b_off
, l
));
6437 // read raw blob data. use aio if we have >1 blobs to read.
6438 start
= ceph_clock_now(); // for the sake of simplicity
6439 // measure the whole block below.
6440 // The error isn't that much...
6441 vector
<bufferlist
> compressed_blob_bls
;
6442 IOContext
ioc(cct
, NULL
);
6443 for (auto& p
: blobs2read
) {
6444 BlobRef bptr
= p
.first
;
6445 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6446 << " need " << p
.second
<< std::dec
<< dendl
;
6447 if (bptr
->get_blob().is_compressed()) {
6448 // read the whole thing
6449 if (compressed_blob_bls
.empty()) {
6450 // ensure we avoid any reallocation on subsequent blobs
6451 compressed_blob_bls
.reserve(blobs2read
.size());
6453 compressed_blob_bls
.push_back(bufferlist());
6454 bufferlist
& bl
= compressed_blob_bls
.back();
6455 r
= bptr
->get_blob().map(
6456 0, bptr
->get_blob().get_ondisk_length(),
6457 [&](uint64_t offset
, uint64_t length
) {
6459 // use aio if there are more regions to read than those in this blob
6460 if (num_regions
> p
.second
.size()) {
6461 r
= bdev
->aio_read(offset
, length
, &bl
, &ioc
);
6463 r
= bdev
->read(offset
, length
, &bl
, &ioc
, false);
6472 for (auto& reg
: p
.second
) {
6473 // determine how much of the blob to read
6474 uint64_t chunk_size
= bptr
->get_blob().get_chunk_size(block_size
);
6475 reg
.r_off
= reg
.blob_xoffset
;
6476 uint64_t r_len
= reg
.length
;
6477 reg
.front
= reg
.r_off
% chunk_size
;
6479 reg
.r_off
-= reg
.front
;
6482 unsigned tail
= r_len
% chunk_size
;
6484 r_len
+= chunk_size
- tail
;
6486 dout(20) << __func__
<< " region 0x" << std::hex
6487 << reg
.logical_offset
6488 << ": 0x" << reg
.blob_xoffset
<< "~" << reg
.length
6489 << " reading 0x" << reg
.r_off
<< "~" << r_len
<< std::dec
6493 r
= bptr
->get_blob().map(
6495 [&](uint64_t offset
, uint64_t length
) {
6497 // use aio if there is more than one region to read
6498 if (num_regions
> 1) {
6499 r
= bdev
->aio_read(offset
, length
, ®
.bl
, &ioc
);
6501 r
= bdev
->read(offset
, length
, ®
.bl
, &ioc
, false);
6508 assert(reg
.bl
.length() == r_len
);
6512 if (ioc
.has_pending_aios()) {
6513 bdev
->aio_submit(&ioc
);
6514 dout(20) << __func__
<< " waiting for aio" << dendl
;
6517 logger
->tinc(l_bluestore_read_wait_aio_lat
, ceph_clock_now() - start
);
6519 // enumerate and decompress desired blobs
6520 auto p
= compressed_blob_bls
.begin();
6521 blobs2read_t::iterator b2r_it
= blobs2read
.begin();
6522 while (b2r_it
!= blobs2read
.end()) {
6523 BlobRef bptr
= b2r_it
->first
;
6524 dout(20) << __func__
<< " blob " << *bptr
<< std::hex
6525 << " need 0x" << b2r_it
->second
<< std::dec
<< dendl
;
6526 if (bptr
->get_blob().is_compressed()) {
6527 assert(p
!= compressed_blob_bls
.end());
6528 bufferlist
& compressed_bl
= *p
++;
6529 if (_verify_csum(o
, &bptr
->get_blob(), 0, compressed_bl
,
6530 b2r_it
->second
.front().logical_offset
) < 0) {
6534 r
= _decompress(compressed_bl
, &raw_bl
);
6538 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(), 0,
6541 for (auto& i
: b2r_it
->second
) {
6542 ready_regions
[i
.logical_offset
].substr_of(
6543 raw_bl
, i
.blob_xoffset
, i
.length
);
6546 for (auto& reg
: b2r_it
->second
) {
6547 if (_verify_csum(o
, &bptr
->get_blob(), reg
.r_off
, reg
.bl
,
6548 reg
.logical_offset
) < 0) {
6552 bptr
->shared_blob
->bc
.did_read(bptr
->shared_blob
->get_cache(),
6556 // prune and keep result
6557 ready_regions
[reg
.logical_offset
].substr_of(
6558 reg
.bl
, reg
.front
, reg
.length
);
6564 // generate a resulting buffer
6565 auto pr
= ready_regions
.begin();
6566 auto pr_end
= ready_regions
.end();
6568 while (pos
< length
) {
6569 if (pr
!= pr_end
&& pr
->first
== pos
+ offset
) {
6570 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6571 << ": data from 0x" << pr
->first
<< "~" << pr
->second
.length()
6572 << std::dec
<< dendl
;
6573 pos
+= pr
->second
.length();
6574 bl
.claim_append(pr
->second
);
6577 uint64_t l
= length
- pos
;
6579 assert(pr
->first
> pos
+ offset
);
6580 l
= pr
->first
- (pos
+ offset
);
6582 dout(30) << __func__
<< " assemble 0x" << std::hex
<< pos
6583 << ": zeros for 0x" << (pos
+ offset
) << "~" << l
6584 << std::dec
<< dendl
;
6589 assert(bl
.length() == length
);
6590 assert(pos
== length
);
6591 assert(pr
== pr_end
);
6596 int BlueStore::_verify_csum(OnodeRef
& o
,
6597 const bluestore_blob_t
* blob
, uint64_t blob_xoffset
,
6598 const bufferlist
& bl
,
6599 uint64_t logical_offset
) const
6603 utime_t start
= ceph_clock_now();
6604 int r
= blob
->verify_csum(blob_xoffset
, bl
, &bad
, &bad_csum
);
6610 blob
->get_csum_chunk_size(),
6611 [&](uint64_t offset
, uint64_t length
) {
6612 pex
.emplace_back(bluestore_pextent_t(offset
, length
));
6615 derr
<< __func__
<< " bad "
6616 << Checksummer::get_csum_type_string(blob
->csum_type
)
6617 << "/0x" << std::hex
<< blob
->get_csum_chunk_size()
6618 << " checksum at blob offset 0x" << bad
6619 << ", got 0x" << bad_csum
<< ", expected 0x"
6620 << blob
->get_csum_item(bad
/ blob
->get_csum_chunk_size()) << std::dec
6621 << ", device location " << pex
6622 << ", logical extent 0x" << std::hex
6623 << (logical_offset
+ bad
- blob_xoffset
) << "~"
6624 << blob
->get_csum_chunk_size() << std::dec
6625 << ", object " << o
->oid
6628 derr
<< __func__
<< " failed with exit code: " << cpp_strerror(r
) << dendl
;
6631 logger
->tinc(l_bluestore_csum_lat
, ceph_clock_now() - start
);
6635 int BlueStore::_decompress(bufferlist
& source
, bufferlist
* result
)
6638 utime_t start
= ceph_clock_now();
6639 bufferlist::iterator i
= source
.begin();
6640 bluestore_compression_header_t chdr
;
6642 int alg
= int(chdr
.type
);
6643 CompressorRef cp
= compressor
;
6644 if (!cp
|| (int)cp
->get_type() != alg
) {
6645 cp
= Compressor::create(cct
, alg
);
6649 // if compressor isn't available - error, because cannot return
6650 // decompressed data?
6651 derr
<< __func__
<< " can't load decompressor " << alg
<< dendl
;
6654 r
= cp
->decompress(i
, chdr
.length
, *result
);
6656 derr
<< __func__
<< " decompression failed with exit code " << r
<< dendl
;
6660 logger
->tinc(l_bluestore_decompress_lat
, ceph_clock_now() - start
);
6664 // this stores fiemap into interval_set, other variations
6665 // use it internally
6666 int BlueStore::_fiemap(
6667 CollectionHandle
&c_
,
6668 const ghobject_t
& oid
,
6671 interval_set
<uint64_t>& destset
)
6673 Collection
*c
= static_cast<Collection
*>(c_
.get());
6677 RWLock::RLocker
l(c
->lock
);
6679 OnodeRef o
= c
->get_onode(oid
, false);
6680 if (!o
|| !o
->exists
) {
6685 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6686 << " size 0x" << o
->onode
.size
<< std::dec
<< dendl
;
6688 boost::intrusive::set
<Extent
>::iterator ep
, eend
;
6689 if (offset
>= o
->onode
.size
)
6692 if (offset
+ length
> o
->onode
.size
) {
6693 length
= o
->onode
.size
- offset
;
6696 o
->extent_map
.fault_range(db
, offset
, length
);
6697 eend
= o
->extent_map
.extent_map
.end();
6698 ep
= o
->extent_map
.seek_lextent(offset
);
6699 while (length
> 0) {
6700 dout(20) << __func__
<< " offset " << offset
<< dendl
;
6701 if (ep
!= eend
&& ep
->logical_offset
+ ep
->length
<= offset
) {
6706 uint64_t x_len
= length
;
6707 if (ep
!= eend
&& ep
->logical_offset
<= offset
) {
6708 uint64_t x_off
= offset
- ep
->logical_offset
;
6709 x_len
= MIN(x_len
, ep
->length
- x_off
);
6710 dout(30) << __func__
<< " lextent 0x" << std::hex
<< offset
<< "~"
6711 << x_len
<< std::dec
<< " blob " << ep
->blob
<< dendl
;
6712 destset
.insert(offset
, x_len
);
6715 if (x_off
+ x_len
== ep
->length
)
6720 ep
->logical_offset
> offset
&&
6721 ep
->logical_offset
- offset
< x_len
) {
6722 x_len
= ep
->logical_offset
- offset
;
6730 dout(20) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
6731 << " size = 0x(" << destset
<< ")" << std::dec
<< dendl
;
6735 int BlueStore::fiemap(
6737 const ghobject_t
& oid
,
6742 CollectionHandle c
= _get_collection(cid
);
6745 return fiemap(c
, oid
, offset
, len
, bl
);
6748 int BlueStore::fiemap(
6749 CollectionHandle
&c_
,
6750 const ghobject_t
& oid
,
6755 interval_set
<uint64_t> m
;
6756 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6763 int BlueStore::fiemap(
6765 const ghobject_t
& oid
,
6768 map
<uint64_t, uint64_t>& destmap
)
6770 CollectionHandle c
= _get_collection(cid
);
6773 return fiemap(c
, oid
, offset
, len
, destmap
);
6776 int BlueStore::fiemap(
6777 CollectionHandle
&c_
,
6778 const ghobject_t
& oid
,
6781 map
<uint64_t, uint64_t>& destmap
)
6783 interval_set
<uint64_t> m
;
6784 int r
= _fiemap(c_
, oid
, offset
, length
, m
);
6786 m
.move_into(destmap
);
6791 int BlueStore::getattr(
6793 const ghobject_t
& oid
,
6797 CollectionHandle c
= _get_collection(cid
);
6800 return getattr(c
, oid
, name
, value
);
6803 int BlueStore::getattr(
6804 CollectionHandle
&c_
,
6805 const ghobject_t
& oid
,
6809 Collection
*c
= static_cast<Collection
*>(c_
.get());
6810 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
<< dendl
;
6816 RWLock::RLocker
l(c
->lock
);
6817 mempool::bluestore_cache_other::string
k(name
);
6819 OnodeRef o
= c
->get_onode(oid
, false);
6820 if (!o
|| !o
->exists
) {
6825 if (!o
->onode
.attrs
.count(k
)) {
6829 value
= o
->onode
.attrs
[k
];
6833 if (r
== 0 && _debug_mdata_eio(oid
)) {
6835 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6837 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
<< " " << name
6838 << " = " << r
<< dendl
;
6843 int BlueStore::getattrs(
6845 const ghobject_t
& oid
,
6846 map
<string
,bufferptr
>& aset
)
6848 CollectionHandle c
= _get_collection(cid
);
6851 return getattrs(c
, oid
, aset
);
6854 int BlueStore::getattrs(
6855 CollectionHandle
&c_
,
6856 const ghobject_t
& oid
,
6857 map
<string
,bufferptr
>& aset
)
6859 Collection
*c
= static_cast<Collection
*>(c_
.get());
6860 dout(15) << __func__
<< " " << c
->cid
<< " " << oid
<< dendl
;
6866 RWLock::RLocker
l(c
->lock
);
6868 OnodeRef o
= c
->get_onode(oid
, false);
6869 if (!o
|| !o
->exists
) {
6873 for (auto& i
: o
->onode
.attrs
) {
6874 aset
.emplace(i
.first
.c_str(), i
.second
);
6880 if (r
== 0 && _debug_mdata_eio(oid
)) {
6882 derr
<< __func__
<< " " << c
->cid
<< " " << oid
<< " INJECT EIO" << dendl
;
6884 dout(10) << __func__
<< " " << c
->cid
<< " " << oid
6885 << " = " << r
<< dendl
;
6889 int BlueStore::list_collections(vector
<coll_t
>& ls
)
6891 RWLock::RLocker
l(coll_lock
);
6892 for (ceph::unordered_map
<coll_t
, CollectionRef
>::iterator p
= coll_map
.begin();
6893 p
!= coll_map
.end();
6895 ls
.push_back(p
->first
);
6899 bool BlueStore::collection_exists(const coll_t
& c
)
6901 RWLock::RLocker
l(coll_lock
);
6902 return coll_map
.count(c
);
6905 int BlueStore::collection_empty(const coll_t
& cid
, bool *empty
)
6907 dout(15) << __func__
<< " " << cid
<< dendl
;
6908 vector
<ghobject_t
> ls
;
6910 int r
= collection_list(cid
, ghobject_t(), ghobject_t::get_max(), 1,
6913 derr
<< __func__
<< " collection_list returned: " << cpp_strerror(r
)
6917 *empty
= ls
.empty();
6918 dout(10) << __func__
<< " " << cid
<< " = " << (int)(*empty
) << dendl
;
6922 int BlueStore::collection_bits(const coll_t
& cid
)
6924 dout(15) << __func__
<< " " << cid
<< dendl
;
6925 CollectionRef c
= _get_collection(cid
);
6928 RWLock::RLocker
l(c
->lock
);
6929 dout(10) << __func__
<< " " << cid
<< " = " << c
->cnode
.bits
<< dendl
;
6930 return c
->cnode
.bits
;
6933 int BlueStore::collection_list(
6934 const coll_t
& cid
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
6935 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
6937 CollectionHandle c
= _get_collection(cid
);
6940 return collection_list(c
, start
, end
, max
, ls
, pnext
);
6943 int BlueStore::collection_list(
6944 CollectionHandle
&c_
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
6945 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
6947 Collection
*c
= static_cast<Collection
*>(c_
.get());
6948 dout(15) << __func__
<< " " << c
->cid
6949 << " start " << start
<< " end " << end
<< " max " << max
<< dendl
;
6952 RWLock::RLocker
l(c
->lock
);
6953 r
= _collection_list(c
, start
, end
, max
, ls
, pnext
);
6956 dout(10) << __func__
<< " " << c
->cid
6957 << " start " << start
<< " end " << end
<< " max " << max
6958 << " = " << r
<< ", ls.size() = " << ls
->size()
6959 << ", next = " << (pnext
? *pnext
: ghobject_t()) << dendl
;
6963 int BlueStore::_collection_list(
6964 Collection
*c
, const ghobject_t
& start
, const ghobject_t
& end
, int max
,
6965 vector
<ghobject_t
> *ls
, ghobject_t
*pnext
)
6972 ghobject_t static_next
;
6973 KeyValueDB::Iterator it
;
6974 string temp_start_key
, temp_end_key
;
6975 string start_key
, end_key
;
6976 bool set_next
= false;
6981 pnext
= &static_next
;
6983 if (start
== ghobject_t::get_max() ||
6984 start
.hobj
.is_max()) {
6987 get_coll_key_range(c
->cid
, c
->cnode
.bits
, &temp_start_key
, &temp_end_key
,
6988 &start_key
, &end_key
);
6989 dout(20) << __func__
6990 << " range " << pretty_binary_string(temp_start_key
)
6991 << " to " << pretty_binary_string(temp_end_key
)
6992 << " and " << pretty_binary_string(start_key
)
6993 << " to " << pretty_binary_string(end_key
)
6994 << " start " << start
<< dendl
;
6995 it
= db
->get_iterator(PREFIX_OBJ
);
6996 if (start
== ghobject_t() ||
6997 start
.hobj
== hobject_t() ||
6998 start
== c
->cid
.get_min_hobj()) {
6999 it
->upper_bound(temp_start_key
);
7003 get_object_key(cct
, start
, &k
);
7004 if (start
.hobj
.is_temp()) {
7006 assert(k
>= temp_start_key
&& k
< temp_end_key
);
7009 assert(k
>= start_key
&& k
< end_key
);
7011 dout(20) << " start from " << pretty_binary_string(k
)
7012 << " temp=" << (int)temp
<< dendl
;
7015 if (end
.hobj
.is_max()) {
7016 pend
= temp
? temp_end_key
: end_key
;
7018 get_object_key(cct
, end
, &end_key
);
7019 if (end
.hobj
.is_temp()) {
7025 pend
= temp
? temp_end_key
: end_key
;
7028 dout(20) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7030 if (!it
->valid() || it
->key() >= pend
) {
7032 dout(20) << __func__
<< " iterator not valid (end of db?)" << dendl
;
7034 dout(20) << __func__
<< " key " << pretty_binary_string(it
->key())
7035 << " >= " << end
<< dendl
;
7037 if (end
.hobj
.is_temp()) {
7040 dout(30) << __func__
<< " switch to non-temp namespace" << dendl
;
7042 it
->upper_bound(start_key
);
7044 dout(30) << __func__
<< " pend " << pretty_binary_string(pend
) << dendl
;
7049 dout(30) << __func__
<< " key " << pretty_binary_string(it
->key()) << dendl
;
7050 if (is_extent_shard_key(it
->key())) {
7055 int r
= get_key_object(it
->key(), &oid
);
7057 dout(20) << __func__
<< " oid " << oid
<< " end " << end
<< dendl
;
7058 if (ls
->size() >= (unsigned)max
) {
7059 dout(20) << __func__
<< " reached max " << max
<< dendl
;
7069 *pnext
= ghobject_t::get_max();
7075 int BlueStore::omap_get(
7076 const coll_t
& cid
, ///< [in] Collection containing oid
7077 const ghobject_t
&oid
, ///< [in] Object containing omap
7078 bufferlist
*header
, ///< [out] omap header
7079 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7082 CollectionHandle c
= _get_collection(cid
);
7085 return omap_get(c
, oid
, header
, out
);
7088 int BlueStore::omap_get(
7089 CollectionHandle
&c_
, ///< [in] Collection containing oid
7090 const ghobject_t
&oid
, ///< [in] Object containing omap
7091 bufferlist
*header
, ///< [out] omap header
7092 map
<string
, bufferlist
> *out
/// < [out] Key to value map
7095 Collection
*c
= static_cast<Collection
*>(c_
.get());
7096 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7099 RWLock::RLocker
l(c
->lock
);
7101 OnodeRef o
= c
->get_onode(oid
, false);
7102 if (!o
|| !o
->exists
) {
7106 if (!o
->onode
.has_omap())
7110 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7112 get_omap_header(o
->onode
.nid
, &head
);
7113 get_omap_tail(o
->onode
.nid
, &tail
);
7114 it
->lower_bound(head
);
7115 while (it
->valid()) {
7116 if (it
->key() == head
) {
7117 dout(30) << __func__
<< " got header" << dendl
;
7118 *header
= it
->value();
7119 } else if (it
->key() >= tail
) {
7120 dout(30) << __func__
<< " reached tail" << dendl
;
7124 decode_omap_key(it
->key(), &user_key
);
7125 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7126 << " -> " << user_key
<< dendl
;
7127 (*out
)[user_key
] = it
->value();
7133 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7138 int BlueStore::omap_get_header(
7139 const coll_t
& cid
, ///< [in] Collection containing oid
7140 const ghobject_t
&oid
, ///< [in] Object containing omap
7141 bufferlist
*header
, ///< [out] omap header
7142 bool allow_eio
///< [in] don't assert on eio
7145 CollectionHandle c
= _get_collection(cid
);
7148 return omap_get_header(c
, oid
, header
, allow_eio
);
7151 int BlueStore::omap_get_header(
7152 CollectionHandle
&c_
, ///< [in] Collection containing oid
7153 const ghobject_t
&oid
, ///< [in] Object containing omap
7154 bufferlist
*header
, ///< [out] omap header
7155 bool allow_eio
///< [in] don't assert on eio
7158 Collection
*c
= static_cast<Collection
*>(c_
.get());
7159 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7162 RWLock::RLocker
l(c
->lock
);
7164 OnodeRef o
= c
->get_onode(oid
, false);
7165 if (!o
|| !o
->exists
) {
7169 if (!o
->onode
.has_omap())
7174 get_omap_header(o
->onode
.nid
, &head
);
7175 if (db
->get(PREFIX_OMAP
, head
, header
) >= 0) {
7176 dout(30) << __func__
<< " got header" << dendl
;
7178 dout(30) << __func__
<< " no header" << dendl
;
7182 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7187 int BlueStore::omap_get_keys(
7188 const coll_t
& cid
, ///< [in] Collection containing oid
7189 const ghobject_t
&oid
, ///< [in] Object containing omap
7190 set
<string
> *keys
///< [out] Keys defined on oid
7193 CollectionHandle c
= _get_collection(cid
);
7196 return omap_get_keys(c
, oid
, keys
);
7199 int BlueStore::omap_get_keys(
7200 CollectionHandle
&c_
, ///< [in] Collection containing oid
7201 const ghobject_t
&oid
, ///< [in] Object containing omap
7202 set
<string
> *keys
///< [out] Keys defined on oid
7205 Collection
*c
= static_cast<Collection
*>(c_
.get());
7206 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7209 RWLock::RLocker
l(c
->lock
);
7211 OnodeRef o
= c
->get_onode(oid
, false);
7212 if (!o
|| !o
->exists
) {
7216 if (!o
->onode
.has_omap())
7220 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7222 get_omap_key(o
->onode
.nid
, string(), &head
);
7223 get_omap_tail(o
->onode
.nid
, &tail
);
7224 it
->lower_bound(head
);
7225 while (it
->valid()) {
7226 if (it
->key() >= tail
) {
7227 dout(30) << __func__
<< " reached tail" << dendl
;
7231 decode_omap_key(it
->key(), &user_key
);
7232 dout(30) << __func__
<< " got " << pretty_binary_string(it
->key())
7233 << " -> " << user_key
<< dendl
;
7234 keys
->insert(user_key
);
7239 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7244 int BlueStore::omap_get_values(
7245 const coll_t
& cid
, ///< [in] Collection containing oid
7246 const ghobject_t
&oid
, ///< [in] Object containing omap
7247 const set
<string
> &keys
, ///< [in] Keys to get
7248 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7251 CollectionHandle c
= _get_collection(cid
);
7254 return omap_get_values(c
, oid
, keys
, out
);
7257 int BlueStore::omap_get_values(
7258 CollectionHandle
&c_
, ///< [in] Collection containing oid
7259 const ghobject_t
&oid
, ///< [in] Object containing omap
7260 const set
<string
> &keys
, ///< [in] Keys to get
7261 map
<string
, bufferlist
> *out
///< [out] Returned keys and values
7264 Collection
*c
= static_cast<Collection
*>(c_
.get());
7265 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7268 RWLock::RLocker
l(c
->lock
);
7271 OnodeRef o
= c
->get_onode(oid
, false);
7272 if (!o
|| !o
->exists
) {
7276 if (!o
->onode
.has_omap())
7279 _key_encode_u64(o
->onode
.nid
, &final_key
);
7280 final_key
.push_back('.');
7281 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7282 final_key
.resize(9); // keep prefix
7285 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7286 dout(30) << __func__
<< " got " << pretty_binary_string(final_key
)
7287 << " -> " << *p
<< dendl
;
7288 out
->insert(make_pair(*p
, val
));
7292 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7297 int BlueStore::omap_check_keys(
7298 const coll_t
& cid
, ///< [in] Collection containing oid
7299 const ghobject_t
&oid
, ///< [in] Object containing omap
7300 const set
<string
> &keys
, ///< [in] Keys to check
7301 set
<string
> *out
///< [out] Subset of keys defined on oid
7304 CollectionHandle c
= _get_collection(cid
);
7307 return omap_check_keys(c
, oid
, keys
, out
);
7310 int BlueStore::omap_check_keys(
7311 CollectionHandle
&c_
, ///< [in] Collection containing oid
7312 const ghobject_t
&oid
, ///< [in] Object containing omap
7313 const set
<string
> &keys
, ///< [in] Keys to check
7314 set
<string
> *out
///< [out] Subset of keys defined on oid
7317 Collection
*c
= static_cast<Collection
*>(c_
.get());
7318 dout(15) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< dendl
;
7321 RWLock::RLocker
l(c
->lock
);
7324 OnodeRef o
= c
->get_onode(oid
, false);
7325 if (!o
|| !o
->exists
) {
7329 if (!o
->onode
.has_omap())
7332 _key_encode_u64(o
->onode
.nid
, &final_key
);
7333 final_key
.push_back('.');
7334 for (set
<string
>::const_iterator p
= keys
.begin(); p
!= keys
.end(); ++p
) {
7335 final_key
.resize(9); // keep prefix
7338 if (db
->get(PREFIX_OMAP
, final_key
, &val
) >= 0) {
7339 dout(30) << __func__
<< " have " << pretty_binary_string(final_key
)
7340 << " -> " << *p
<< dendl
;
7343 dout(30) << __func__
<< " miss " << pretty_binary_string(final_key
)
7344 << " -> " << *p
<< dendl
;
7348 dout(10) << __func__
<< " " << c
->get_cid() << " oid " << oid
<< " = " << r
7353 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7354 const coll_t
& cid
, ///< [in] collection
7355 const ghobject_t
&oid
///< [in] object
7358 CollectionHandle c
= _get_collection(cid
);
7360 dout(10) << __func__
<< " " << cid
<< "doesn't exist" <<dendl
;
7361 return ObjectMap::ObjectMapIterator();
7363 return get_omap_iterator(c
, oid
);
7366 ObjectMap::ObjectMapIterator
BlueStore::get_omap_iterator(
7367 CollectionHandle
&c_
, ///< [in] collection
7368 const ghobject_t
&oid
///< [in] object
7371 Collection
*c
= static_cast<Collection
*>(c_
.get());
7372 dout(10) << __func__
<< " " << c
->get_cid() << " " << oid
<< dendl
;
7374 return ObjectMap::ObjectMapIterator();
7376 RWLock::RLocker
l(c
->lock
);
7377 OnodeRef o
= c
->get_onode(oid
, false);
7378 if (!o
|| !o
->exists
) {
7379 dout(10) << __func__
<< " " << oid
<< "doesn't exist" <<dendl
;
7380 return ObjectMap::ObjectMapIterator();
7383 dout(10) << __func__
<< " has_omap = " << (int)o
->onode
.has_omap() <<dendl
;
7384 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
7385 return ObjectMap::ObjectMapIterator(new OmapIteratorImpl(c
, o
, it
));
7388 // -----------------
7391 void BlueStore::_prepare_ondisk_format_super(KeyValueDB::Transaction
& t
)
7393 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7394 << " min_compat_ondisk_format " << min_compat_ondisk_format
7396 assert(ondisk_format
== latest_ondisk_format
);
7399 ::encode(ondisk_format
, bl
);
7400 t
->set(PREFIX_SUPER
, "ondisk_format", bl
);
7404 ::encode(min_compat_ondisk_format
, bl
);
7405 t
->set(PREFIX_SUPER
, "min_compat_ondisk_format", bl
);
7409 int BlueStore::_open_super_meta()
7415 db
->get(PREFIX_SUPER
, "nid_max", &bl
);
7416 bufferlist::iterator p
= bl
.begin();
7421 } catch (buffer::error
& e
) {
7422 derr
<< __func__
<< " unable to read nid_max" << dendl
;
7425 dout(10) << __func__
<< " old nid_max " << nid_max
<< dendl
;
7426 nid_last
= nid_max
.load();
7433 db
->get(PREFIX_SUPER
, "blobid_max", &bl
);
7434 bufferlist::iterator p
= bl
.begin();
7439 } catch (buffer::error
& e
) {
7440 derr
<< __func__
<< " unable to read blobid_max" << dendl
;
7443 dout(10) << __func__
<< " old blobid_max " << blobid_max
<< dendl
;
7444 blobid_last
= blobid_max
.load();
7450 db
->get(PREFIX_SUPER
, "freelist_type", &bl
);
7452 freelist_type
= std::string(bl
.c_str(), bl
.length());
7453 dout(10) << __func__
<< " freelist_type " << freelist_type
<< dendl
;
7455 assert("Not Support extent freelist manager" == 0);
7460 if (cct
->_conf
->bluestore_bluefs
) {
7461 bluefs_extents
.clear();
7463 db
->get(PREFIX_SUPER
, "bluefs_extents", &bl
);
7464 bufferlist::iterator p
= bl
.begin();
7466 ::decode(bluefs_extents
, p
);
7468 catch (buffer::error
& e
) {
7469 derr
<< __func__
<< " unable to read bluefs_extents" << dendl
;
7472 dout(10) << __func__
<< " bluefs_extents 0x" << std::hex
<< bluefs_extents
7473 << std::dec
<< dendl
;
7477 int32_t compat_ondisk_format
= 0;
7480 int r
= db
->get(PREFIX_SUPER
, "ondisk_format", &bl
);
7482 // base case: kraken bluestore is v1 and readable by v1
7483 dout(20) << __func__
<< " missing ondisk_format; assuming kraken"
7486 compat_ondisk_format
= 1;
7488 auto p
= bl
.begin();
7490 ::decode(ondisk_format
, p
);
7491 } catch (buffer::error
& e
) {
7492 derr
<< __func__
<< " unable to read ondisk_format" << dendl
;
7497 r
= db
->get(PREFIX_SUPER
, "min_compat_ondisk_format", &bl
);
7499 auto p
= bl
.begin();
7501 ::decode(compat_ondisk_format
, p
);
7502 } catch (buffer::error
& e
) {
7503 derr
<< __func__
<< " unable to read compat_ondisk_format" << dendl
;
7508 dout(10) << __func__
<< " ondisk_format " << ondisk_format
7509 << " compat_ondisk_format " << compat_ondisk_format
7513 if (latest_ondisk_format
< compat_ondisk_format
) {
7514 derr
<< __func__
<< " compat_ondisk_format is "
7515 << compat_ondisk_format
<< " but we only understand version "
7516 << latest_ondisk_format
<< dendl
;
7519 if (ondisk_format
< latest_ondisk_format
) {
7520 int r
= _upgrade_super();
7528 db
->get(PREFIX_SUPER
, "min_alloc_size", &bl
);
7529 auto p
= bl
.begin();
7533 min_alloc_size
= val
;
7534 min_alloc_size_order
= ctz(val
);
7535 assert(min_alloc_size
== 1u << min_alloc_size_order
);
7536 } catch (buffer::error
& e
) {
7537 derr
<< __func__
<< " unable to read min_alloc_size" << dendl
;
7540 dout(10) << __func__
<< " min_alloc_size 0x" << std::hex
<< min_alloc_size
7541 << std::dec
<< dendl
;
7545 _set_throttle_params();
7554 int BlueStore::_upgrade_super()
7556 dout(1) << __func__
<< " from " << ondisk_format
<< ", latest "
7557 << latest_ondisk_format
<< dendl
;
7558 assert(ondisk_format
> 0);
7559 assert(ondisk_format
< latest_ondisk_format
);
7561 if (ondisk_format
== 1) {
7563 // - super: added ondisk_format
7564 // - super: added min_readable_ondisk_format
7565 // - super: added min_compat_ondisk_format
7566 // - super: added min_alloc_size
7567 // - super: removed min_min_alloc_size
7568 KeyValueDB::Transaction t
= db
->get_transaction();
7571 db
->get(PREFIX_SUPER
, "min_min_alloc_size", &bl
);
7572 auto p
= bl
.begin();
7576 min_alloc_size
= val
;
7577 } catch (buffer::error
& e
) {
7578 derr
<< __func__
<< " failed to read min_min_alloc_size" << dendl
;
7581 t
->set(PREFIX_SUPER
, "min_alloc_size", bl
);
7582 t
->rmkey(PREFIX_SUPER
, "min_min_alloc_size");
7585 _prepare_ondisk_format_super(t
);
7586 int r
= db
->submit_transaction_sync(t
);
7591 dout(1) << __func__
<< " done" << dendl
;
7595 void BlueStore::_assign_nid(TransContext
*txc
, OnodeRef o
)
7601 uint64_t nid
= ++nid_last
;
7602 dout(20) << __func__
<< " " << nid
<< dendl
;
7604 txc
->last_nid
= nid
;
7608 uint64_t BlueStore::_assign_blobid(TransContext
*txc
)
7610 uint64_t bid
= ++blobid_last
;
7611 dout(20) << __func__
<< " " << bid
<< dendl
;
7612 txc
->last_blobid
= bid
;
7616 void BlueStore::get_db_statistics(Formatter
*f
)
7618 db
->get_statistics(f
);
7621 BlueStore::TransContext
*BlueStore::_txc_create(OpSequencer
*osr
)
7623 TransContext
*txc
= new TransContext(cct
, osr
);
7624 txc
->t
= db
->get_transaction();
7625 osr
->queue_new(txc
);
7626 dout(20) << __func__
<< " osr " << osr
<< " = " << txc
7627 << " seq " << txc
->seq
<< dendl
;
7631 void BlueStore::_txc_calc_cost(TransContext
*txc
)
7633 // this is about the simplest model for transaction cost you can
7634 // imagine. there is some fixed overhead cost by saying there is a
7635 // minimum of one "io". and then we have some cost per "io" that is
7636 // a configurable (with different hdd and ssd defaults), and add
7637 // that to the bytes value.
7638 int ios
= 1; // one "io" for the kv commit
7639 for (auto& p
: txc
->ioc
.pending_aios
) {
7640 ios
+= p
.iov
.size();
7642 auto cost
= throttle_cost_per_io
.load();
7643 txc
->cost
= ios
* cost
+ txc
->bytes
;
7644 dout(10) << __func__
<< " " << txc
<< " cost " << txc
->cost
<< " ("
7645 << ios
<< " ios * " << cost
<< " + " << txc
->bytes
7646 << " bytes)" << dendl
;
7649 void BlueStore::_txc_update_store_statfs(TransContext
*txc
)
7651 if (txc
->statfs_delta
.is_empty())
7654 logger
->inc(l_bluestore_allocated
, txc
->statfs_delta
.allocated());
7655 logger
->inc(l_bluestore_stored
, txc
->statfs_delta
.stored());
7656 logger
->inc(l_bluestore_compressed
, txc
->statfs_delta
.compressed());
7657 logger
->inc(l_bluestore_compressed_allocated
, txc
->statfs_delta
.compressed_allocated());
7658 logger
->inc(l_bluestore_compressed_original
, txc
->statfs_delta
.compressed_original());
7661 std::lock_guard
<std::mutex
> l(vstatfs_lock
);
7662 vstatfs
+= txc
->statfs_delta
;
7666 txc
->statfs_delta
.encode(bl
);
7668 txc
->t
->merge(PREFIX_STAT
, "bluestore_statfs", bl
);
7669 txc
->statfs_delta
.reset();
7672 void BlueStore::_txc_state_proc(TransContext
*txc
)
7675 dout(10) << __func__
<< " txc " << txc
7676 << " " << txc
->get_state_name() << dendl
;
7677 switch (txc
->state
) {
7678 case TransContext::STATE_PREPARE
:
7679 txc
->log_state_latency(logger
, l_bluestore_state_prepare_lat
);
7680 if (txc
->ioc
.has_pending_aios()) {
7681 txc
->state
= TransContext::STATE_AIO_WAIT
;
7682 txc
->had_ios
= true;
7683 _txc_aio_submit(txc
);
7688 case TransContext::STATE_AIO_WAIT
:
7689 txc
->log_state_latency(logger
, l_bluestore_state_aio_wait_lat
);
7690 _txc_finish_io(txc
); // may trigger blocked txc's too
7693 case TransContext::STATE_IO_DONE
:
7694 //assert(txc->osr->qlock.is_locked()); // see _txc_finish_io
7696 ++txc
->osr
->txc_with_unstable_io
;
7698 txc
->log_state_latency(logger
, l_bluestore_state_io_done_lat
);
7699 txc
->state
= TransContext::STATE_KV_QUEUED
;
7700 if (cct
->_conf
->bluestore_sync_submit_transaction
) {
7701 if (txc
->last_nid
>= nid_max
||
7702 txc
->last_blobid
>= blobid_max
) {
7703 dout(20) << __func__
7704 << " last_{nid,blobid} exceeds max, submit via kv thread"
7706 } else if (txc
->osr
->kv_committing_serially
) {
7707 dout(20) << __func__
<< " prior txc submitted via kv thread, us too"
7709 // note: this is starvation-prone. once we have a txc in a busy
7710 // sequencer that is committing serially it is possible to keep
7711 // submitting new transactions fast enough that we get stuck doing
7712 // so. the alternative is to block here... fixme?
7713 } else if (txc
->osr
->txc_with_unstable_io
) {
7714 dout(20) << __func__
<< " prior txc(s) with unstable ios "
7715 << txc
->osr
->txc_with_unstable_io
.load() << dendl
;
7716 } else if (cct
->_conf
->bluestore_debug_randomize_serial_transaction
&&
7717 rand() % cct
->_conf
->bluestore_debug_randomize_serial_transaction
7719 dout(20) << __func__
<< " DEBUG randomly forcing submit via kv thread"
7722 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
7723 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
7725 _txc_applied_kv(txc
);
7729 std::lock_guard
<std::mutex
> l(kv_lock
);
7730 kv_queue
.push_back(txc
);
7731 kv_cond
.notify_one();
7732 if (txc
->state
!= TransContext::STATE_KV_SUBMITTED
) {
7733 kv_queue_unsubmitted
.push_back(txc
);
7734 ++txc
->osr
->kv_committing_serially
;
7738 kv_throttle_costs
+= txc
->cost
;
7741 case TransContext::STATE_KV_SUBMITTED
:
7742 txc
->log_state_latency(logger
, l_bluestore_state_kv_committing_lat
);
7743 txc
->state
= TransContext::STATE_KV_DONE
;
7744 _txc_committed_kv(txc
);
7747 case TransContext::STATE_KV_DONE
:
7748 txc
->log_state_latency(logger
, l_bluestore_state_kv_done_lat
);
7749 if (txc
->deferred_txn
) {
7750 txc
->state
= TransContext::STATE_DEFERRED_QUEUED
;
7751 _deferred_queue(txc
);
7754 txc
->state
= TransContext::STATE_FINISHING
;
7757 case TransContext::STATE_DEFERRED_CLEANUP
:
7758 txc
->log_state_latency(logger
, l_bluestore_state_deferred_cleanup_lat
);
7759 txc
->state
= TransContext::STATE_FINISHING
;
7762 case TransContext::STATE_FINISHING
:
7763 txc
->log_state_latency(logger
, l_bluestore_state_finishing_lat
);
7768 derr
<< __func__
<< " unexpected txc " << txc
7769 << " state " << txc
->get_state_name() << dendl
;
7770 assert(0 == "unexpected txc state");
7776 void BlueStore::_txc_finish_io(TransContext
*txc
)
7778 dout(20) << __func__
<< " " << txc
<< dendl
;
7781 * we need to preserve the order of kv transactions,
7782 * even though aio will complete in any order.
7785 OpSequencer
*osr
= txc
->osr
.get();
7786 std::lock_guard
<std::mutex
> l(osr
->qlock
);
7787 txc
->state
= TransContext::STATE_IO_DONE
;
7789 // release aio contexts (including pinned buffers).
7790 txc
->ioc
.running_aios
.clear();
7792 OpSequencer::q_list_t::iterator p
= osr
->q
.iterator_to(*txc
);
7793 while (p
!= osr
->q
.begin()) {
7795 if (p
->state
< TransContext::STATE_IO_DONE
) {
7796 dout(20) << __func__
<< " " << txc
<< " blocked by " << &*p
<< " "
7797 << p
->get_state_name() << dendl
;
7800 if (p
->state
> TransContext::STATE_IO_DONE
) {
7806 _txc_state_proc(&*p
++);
7807 } while (p
!= osr
->q
.end() &&
7808 p
->state
== TransContext::STATE_IO_DONE
);
7810 if (osr
->kv_submitted_waiters
&&
7811 osr
->_is_all_kv_submitted()) {
7812 osr
->qcond
.notify_all();
7816 void BlueStore::_txc_write_nodes(TransContext
*txc
, KeyValueDB::Transaction t
)
7818 dout(20) << __func__
<< " txc " << txc
7819 << " onodes " << txc
->onodes
7820 << " shared_blobs " << txc
->shared_blobs
7824 for (auto o
: txc
->onodes
) {
7825 // finalize extent_map shards
7826 o
->extent_map
.update(t
, false);
7827 if (o
->extent_map
.needs_reshard()) {
7828 o
->extent_map
.reshard(db
, t
);
7829 o
->extent_map
.update(t
, true);
7830 if (o
->extent_map
.needs_reshard()) {
7831 dout(20) << __func__
<< " warning: still wants reshard, check options?"
7833 o
->extent_map
.clear_needs_reshard();
7835 logger
->inc(l_bluestore_onode_reshard
);
7840 denc(o
->onode
, bound
);
7841 o
->extent_map
.bound_encode_spanning_blobs(bound
);
7842 if (o
->onode
.extent_map_shards
.empty()) {
7843 denc(o
->extent_map
.inline_bl
, bound
);
7848 unsigned onode_part
, blob_part
, extent_part
;
7850 auto p
= bl
.get_contiguous_appender(bound
, true);
7852 onode_part
= p
.get_logical_offset();
7853 o
->extent_map
.encode_spanning_blobs(p
);
7854 blob_part
= p
.get_logical_offset() - onode_part
;
7855 if (o
->onode
.extent_map_shards
.empty()) {
7856 denc(o
->extent_map
.inline_bl
, p
);
7858 extent_part
= p
.get_logical_offset() - onode_part
- blob_part
;
7861 dout(20) << " onode " << o
->oid
<< " is " << bl
.length()
7862 << " (" << onode_part
<< " bytes onode + "
7863 << blob_part
<< " bytes spanning blobs + "
7864 << extent_part
<< " bytes inline extents)"
7866 t
->set(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size(), bl
);
7867 o
->flushing_count
++;
7870 // objects we modified but didn't affect the onode
7871 auto p
= txc
->modified_objects
.begin();
7872 while (p
!= txc
->modified_objects
.end()) {
7873 if (txc
->onodes
.count(*p
) == 0) {
7874 (*p
)->flushing_count
++;
7877 // remove dups with onodes list to avoid problems in _txc_finish
7878 p
= txc
->modified_objects
.erase(p
);
7882 // finalize shared_blobs
7883 for (auto sb
: txc
->shared_blobs
) {
7885 auto sbid
= sb
->get_sbid();
7886 get_shared_blob_key(sbid
, &key
);
7887 if (sb
->persistent
->empty()) {
7888 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
7889 << " is empty" << dendl
;
7890 t
->rmkey(PREFIX_SHARED_BLOB
, key
);
7893 ::encode(*(sb
->persistent
), bl
);
7894 dout(20) << " shared_blob 0x" << std::hex
<< sbid
<< std::dec
7895 << " is " << bl
.length() << " " << *sb
<< dendl
;
7896 t
->set(PREFIX_SHARED_BLOB
, key
, bl
);
7901 void BlueStore::BSPerfTracker::update_from_perfcounters(
7902 PerfCounters
&logger
)
7904 os_commit_latency
.consume_next(
7906 l_bluestore_commit_lat
));
7907 os_apply_latency
.consume_next(
7909 l_bluestore_commit_lat
));
7912 void BlueStore::_txc_finalize_kv(TransContext
*txc
, KeyValueDB::Transaction t
)
7914 dout(20) << __func__
<< " txc " << txc
<< std::hex
7915 << " allocated 0x" << txc
->allocated
7916 << " released 0x" << txc
->released
7917 << std::dec
<< dendl
;
7919 // We have to handle the case where we allocate *and* deallocate the
7920 // same region in this transaction. The freelist doesn't like that.
7921 // (Actually, the only thing that cares is the BitmapFreelistManager
7922 // debug check. But that's important.)
7923 interval_set
<uint64_t> tmp_allocated
, tmp_released
;
7924 interval_set
<uint64_t> *pallocated
= &txc
->allocated
;
7925 interval_set
<uint64_t> *preleased
= &txc
->released
;
7926 if (!txc
->allocated
.empty() && !txc
->released
.empty()) {
7927 interval_set
<uint64_t> overlap
;
7928 overlap
.intersection_of(txc
->allocated
, txc
->released
);
7929 if (!overlap
.empty()) {
7930 tmp_allocated
= txc
->allocated
;
7931 tmp_allocated
.subtract(overlap
);
7932 tmp_released
= txc
->released
;
7933 tmp_released
.subtract(overlap
);
7934 dout(20) << __func__
<< " overlap 0x" << std::hex
<< overlap
7935 << ", new allocated 0x" << tmp_allocated
7936 << " released 0x" << tmp_released
<< std::dec
7938 pallocated
= &tmp_allocated
;
7939 preleased
= &tmp_released
;
7943 // update freelist with non-overlap sets
7944 for (interval_set
<uint64_t>::iterator p
= pallocated
->begin();
7945 p
!= pallocated
->end();
7947 fm
->allocate(p
.get_start(), p
.get_len(), t
);
7949 for (interval_set
<uint64_t>::iterator p
= preleased
->begin();
7950 p
!= preleased
->end();
7952 dout(20) << __func__
<< " release 0x" << std::hex
<< p
.get_start()
7953 << "~" << p
.get_len() << std::dec
<< dendl
;
7954 fm
->release(p
.get_start(), p
.get_len(), t
);
7957 _txc_update_store_statfs(txc
);
7960 void BlueStore::_txc_applied_kv(TransContext
*txc
)
7962 for (auto ls
: { &txc
->onodes
, &txc
->modified_objects
}) {
7963 for (auto& o
: *ls
) {
7964 dout(20) << __func__
<< " onode " << o
<< " had " << o
->flushing_count
7966 if (--o
->flushing_count
== 0) {
7967 std::lock_guard
<std::mutex
> l(o
->flush_lock
);
7968 o
->flush_cond
.notify_all();
7974 void BlueStore::_txc_committed_kv(TransContext
*txc
)
7976 dout(20) << __func__
<< " txc " << txc
<< dendl
;
7978 // warning: we're calling onreadable_sync inside the sequencer lock
7979 if (txc
->onreadable_sync
) {
7980 txc
->onreadable_sync
->complete(0);
7981 txc
->onreadable_sync
= NULL
;
7983 unsigned n
= txc
->osr
->parent
->shard_hint
.hash_to_shard(m_finisher_num
);
7984 if (txc
->oncommit
) {
7985 logger
->tinc(l_bluestore_commit_lat
, ceph_clock_now() - txc
->start
);
7986 finishers
[n
]->queue(txc
->oncommit
);
7987 txc
->oncommit
= NULL
;
7989 if (txc
->onreadable
) {
7990 finishers
[n
]->queue(txc
->onreadable
);
7991 txc
->onreadable
= NULL
;
7994 if (!txc
->oncommits
.empty()) {
7995 finishers
[n
]->queue(txc
->oncommits
);
7999 void BlueStore::_txc_finish(TransContext
*txc
)
8001 dout(20) << __func__
<< " " << txc
<< " onodes " << txc
->onodes
<< dendl
;
8002 assert(txc
->state
== TransContext::STATE_FINISHING
);
8004 for (auto& sb
: txc
->shared_blobs_written
) {
8005 sb
->bc
.finish_write(sb
->get_cache(), txc
->seq
);
8007 txc
->shared_blobs_written
.clear();
8009 while (!txc
->removed_collections
.empty()) {
8010 _queue_reap_collection(txc
->removed_collections
.front());
8011 txc
->removed_collections
.pop_front();
8014 OpSequencerRef osr
= txc
->osr
;
8016 bool submit_deferred
= false;
8017 OpSequencer::q_list_t releasing_txc
;
8019 std::lock_guard
<std::mutex
> l(osr
->qlock
);
8020 txc
->state
= TransContext::STATE_DONE
;
8021 bool notify
= false;
8022 while (!osr
->q
.empty()) {
8023 TransContext
*txc
= &osr
->q
.front();
8024 dout(20) << __func__
<< " txc " << txc
<< " " << txc
->get_state_name()
8026 if (txc
->state
!= TransContext::STATE_DONE
) {
8027 if (txc
->state
== TransContext::STATE_PREPARE
&&
8028 deferred_aggressive
) {
8029 // for _osr_drain_preceding()
8032 if (txc
->state
== TransContext::STATE_DEFERRED_QUEUED
&&
8033 osr
->q
.size() > g_conf
->bluestore_max_deferred_txc
) {
8034 submit_deferred
= true;
8040 releasing_txc
.push_back(*txc
);
8044 osr
->qcond
.notify_all();
8046 if (osr
->q
.empty()) {
8047 dout(20) << __func__
<< " osr " << osr
<< " q now empty" << dendl
;
8051 while (!releasing_txc
.empty()) {
8052 // release to allocator only after all preceding txc's have also
8053 // finished any deferred writes that potentially land in these
8055 auto txc
= &releasing_txc
.front();
8056 _txc_release_alloc(txc
);
8057 releasing_txc
.pop_front();
8058 txc
->log_state_latency(logger
, l_bluestore_state_done_lat
);
8062 if (submit_deferred
) {
8063 // we're pinning memory; flush! we could be more fine-grained here but
8064 // i'm not sure it's worth the bother.
8065 deferred_try_submit();
8068 if (empty
&& osr
->zombie
) {
8069 dout(10) << __func__
<< " reaping empty zombie osr " << osr
<< dendl
;
8074 void BlueStore::_txc_release_alloc(TransContext
*txc
)
8076 // update allocator with full released set
8077 if (!cct
->_conf
->bluestore_debug_no_reuse_blocks
) {
8078 dout(10) << __func__
<< " " << txc
<< " " << txc
->released
<< dendl
;
8079 for (interval_set
<uint64_t>::iterator p
= txc
->released
.begin();
8080 p
!= txc
->released
.end();
8082 alloc
->release(p
.get_start(), p
.get_len());
8086 txc
->allocated
.clear();
8087 txc
->released
.clear();
8090 void BlueStore::_osr_drain_preceding(TransContext
*txc
)
8092 OpSequencer
*osr
= txc
->osr
.get();
8093 dout(10) << __func__
<< " " << txc
<< " osr " << osr
<< dendl
;
8094 ++deferred_aggressive
; // FIXME: maybe osr-local aggressive flag?
8096 // submit anything pending
8097 deferred_lock
.lock();
8098 if (osr
->deferred_pending
) {
8099 _deferred_submit_unlock(osr
);
8101 deferred_lock
.unlock();
8105 // wake up any previously finished deferred events
8106 std::lock_guard
<std::mutex
> l(kv_lock
);
8107 kv_cond
.notify_one();
8109 osr
->drain_preceding(txc
);
8110 --deferred_aggressive
;
8111 dout(10) << __func__
<< " " << osr
<< " done" << dendl
;
8114 void BlueStore::_osr_drain_all()
8116 dout(10) << __func__
<< dendl
;
8118 set
<OpSequencerRef
> s
;
8120 std::lock_guard
<std::mutex
> l(osr_lock
);
8123 dout(20) << __func__
<< " osr_set " << s
<< dendl
;
8125 ++deferred_aggressive
;
8127 // submit anything pending
8128 deferred_try_submit();
8131 // wake up any previously finished deferred events
8132 std::lock_guard
<std::mutex
> l(kv_lock
);
8133 kv_cond
.notify_one();
8136 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8137 kv_finalize_cond
.notify_one();
8139 for (auto osr
: s
) {
8140 dout(20) << __func__
<< " drain " << osr
<< dendl
;
8143 --deferred_aggressive
;
8145 dout(10) << __func__
<< " done" << dendl
;
8148 void BlueStore::_osr_unregister_all()
8150 set
<OpSequencerRef
> s
;
8152 std::lock_guard
<std::mutex
> l(osr_lock
);
8155 dout(10) << __func__
<< " " << s
<< dendl
;
8156 for (auto osr
: s
) {
8160 // break link from Sequencer to us so that this OpSequencer
8161 // instance can die with this mount/umount cycle. note that
8162 // we assume umount() will not race against ~Sequencer.
8163 assert(osr
->parent
);
8164 osr
->parent
->p
.reset();
8167 // nobody should be creating sequencers during umount either.
8169 std::lock_guard
<std::mutex
> l(osr_lock
);
8170 assert(osr_set
.empty());
8174 void BlueStore::_kv_start()
8176 dout(10) << __func__
<< dendl
;
8178 if (cct
->_conf
->bluestore_shard_finishers
) {
8179 if (cct
->_conf
->osd_op_num_shards
) {
8180 m_finisher_num
= cct
->_conf
->osd_op_num_shards
;
8183 if (bdev
->is_rotational()) {
8184 m_finisher_num
= cct
->_conf
->osd_op_num_shards_hdd
;
8186 m_finisher_num
= cct
->_conf
->osd_op_num_shards_ssd
;
8191 assert(m_finisher_num
!= 0);
8193 for (int i
= 0; i
< m_finisher_num
; ++i
) {
8195 oss
<< "finisher-" << i
;
8196 Finisher
*f
= new Finisher(cct
, oss
.str(), "finisher");
8197 finishers
.push_back(f
);
8200 for (auto f
: finishers
) {
8203 kv_sync_thread
.create("bstore_kv_sync");
8204 kv_finalize_thread
.create("bstore_kv_final");
8207 void BlueStore::_kv_stop()
8209 dout(10) << __func__
<< dendl
;
8211 std::unique_lock
<std::mutex
> l(kv_lock
);
8212 while (!kv_sync_started
) {
8216 kv_cond
.notify_all();
8219 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8220 while (!kv_finalize_started
) {
8221 kv_finalize_cond
.wait(l
);
8223 kv_finalize_stop
= true;
8224 kv_finalize_cond
.notify_all();
8226 kv_sync_thread
.join();
8227 kv_finalize_thread
.join();
8229 std::lock_guard
<std::mutex
> l(kv_lock
);
8233 std::lock_guard
<std::mutex
> l(kv_finalize_lock
);
8234 kv_finalize_stop
= false;
8236 dout(10) << __func__
<< " stopping finishers" << dendl
;
8237 for (auto f
: finishers
) {
8238 f
->wait_for_empty();
8241 dout(10) << __func__
<< " stopped" << dendl
;
8244 void BlueStore::_kv_sync_thread()
8246 dout(10) << __func__
<< " start" << dendl
;
8247 std::unique_lock
<std::mutex
> l(kv_lock
);
8248 assert(!kv_sync_started
);
8249 kv_sync_started
= true;
8250 kv_cond
.notify_all();
8252 assert(kv_committing
.empty());
8253 if (kv_queue
.empty() &&
8254 ((deferred_done_queue
.empty() && deferred_stable_queue
.empty()) ||
8255 !deferred_aggressive
)) {
8258 dout(20) << __func__
<< " sleep" << dendl
;
8260 dout(20) << __func__
<< " wake" << dendl
;
8262 deque
<TransContext
*> kv_submitting
;
8263 deque
<DeferredBatch
*> deferred_done
, deferred_stable
;
8264 uint64_t aios
= 0, costs
= 0;
8266 dout(20) << __func__
<< " committing " << kv_queue
.size()
8267 << " submitting " << kv_queue_unsubmitted
.size()
8268 << " deferred done " << deferred_done_queue
.size()
8269 << " stable " << deferred_stable_queue
.size()
8271 kv_committing
.swap(kv_queue
);
8272 kv_submitting
.swap(kv_queue_unsubmitted
);
8273 deferred_done
.swap(deferred_done_queue
);
8274 deferred_stable
.swap(deferred_stable_queue
);
8276 costs
= kv_throttle_costs
;
8278 kv_throttle_costs
= 0;
8279 utime_t start
= ceph_clock_now();
8282 dout(30) << __func__
<< " committing " << kv_committing
<< dendl
;
8283 dout(30) << __func__
<< " submitting " << kv_submitting
<< dendl
;
8284 dout(30) << __func__
<< " deferred_done " << deferred_done
<< dendl
;
8285 dout(30) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8287 bool force_flush
= false;
8288 // if bluefs is sharing the same device as data (only), then we
8289 // can rely on the bluefs commit to flush the device and make
8290 // deferred aios stable. that means that if we do have done deferred
8291 // txcs AND we are not on a single device, we need to force a flush.
8292 if (bluefs_single_shared_device
&& bluefs
) {
8295 } else if (kv_committing
.empty() && kv_submitting
.empty() &&
8296 deferred_stable
.empty()) {
8297 force_flush
= true; // there's nothing else to commit!
8298 } else if (deferred_aggressive
) {
8305 dout(20) << __func__
<< " num_aios=" << aios
8306 << " force_flush=" << (int)force_flush
8307 << ", flushing, deferred done->stable" << dendl
;
8308 // flush/barrier on block device
8311 // if we flush then deferred done are now deferred stable
8312 deferred_stable
.insert(deferred_stable
.end(), deferred_done
.begin(),
8313 deferred_done
.end());
8314 deferred_done
.clear();
8316 utime_t after_flush
= ceph_clock_now();
8318 // we will use one final transaction to force a sync
8319 KeyValueDB::Transaction synct
= db
->get_transaction();
8321 // increase {nid,blobid}_max? note that this covers both the
8322 // case where we are approaching the max and the case we passed
8323 // it. in either case, we increase the max in the earlier txn
8325 uint64_t new_nid_max
= 0, new_blobid_max
= 0;
8326 if (nid_last
+ cct
->_conf
->bluestore_nid_prealloc
/2 > nid_max
) {
8327 KeyValueDB::Transaction t
=
8328 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8329 new_nid_max
= nid_last
+ cct
->_conf
->bluestore_nid_prealloc
;
8331 ::encode(new_nid_max
, bl
);
8332 t
->set(PREFIX_SUPER
, "nid_max", bl
);
8333 dout(10) << __func__
<< " new_nid_max " << new_nid_max
<< dendl
;
8335 if (blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
/2 > blobid_max
) {
8336 KeyValueDB::Transaction t
=
8337 kv_submitting
.empty() ? synct
: kv_submitting
.front()->t
;
8338 new_blobid_max
= blobid_last
+ cct
->_conf
->bluestore_blobid_prealloc
;
8340 ::encode(new_blobid_max
, bl
);
8341 t
->set(PREFIX_SUPER
, "blobid_max", bl
);
8342 dout(10) << __func__
<< " new_blobid_max " << new_blobid_max
<< dendl
;
8345 for (auto txc
: kv_committing
) {
8346 if (txc
->state
== TransContext::STATE_KV_QUEUED
) {
8347 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8348 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction(txc
->t
);
8350 _txc_applied_kv(txc
);
8351 --txc
->osr
->kv_committing_serially
;
8352 txc
->state
= TransContext::STATE_KV_SUBMITTED
;
8353 if (txc
->osr
->kv_submitted_waiters
) {
8354 std::lock_guard
<std::mutex
> l(txc
->osr
->qlock
);
8355 if (txc
->osr
->_is_all_kv_submitted()) {
8356 txc
->osr
->qcond
.notify_all();
8361 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8362 txc
->log_state_latency(logger
, l_bluestore_state_kv_queued_lat
);
8365 --txc
->osr
->txc_with_unstable_io
;
8369 // release throttle *before* we commit. this allows new ops
8370 // to be prepared and enter pipeline while we are waiting on
8371 // the kv commit sync/flush. then hopefully on the next
8372 // iteration there will already be ops awake. otherwise, we
8373 // end up going to sleep, and then wake up when the very first
8374 // transaction is ready for commit.
8375 throttle_bytes
.put(costs
);
8377 PExtentVector bluefs_gift_extents
;
8379 after_flush
- bluefs_last_balance
>
8380 cct
->_conf
->bluestore_bluefs_balance_interval
) {
8381 bluefs_last_balance
= after_flush
;
8382 int r
= _balance_bluefs_freespace(&bluefs_gift_extents
);
8385 for (auto& p
: bluefs_gift_extents
) {
8386 bluefs_extents
.insert(p
.offset
, p
.length
);
8389 ::encode(bluefs_extents
, bl
);
8390 dout(10) << __func__
<< " bluefs_extents now 0x" << std::hex
8391 << bluefs_extents
<< std::dec
<< dendl
;
8392 synct
->set(PREFIX_SUPER
, "bluefs_extents", bl
);
8396 // cleanup sync deferred keys
8397 for (auto b
: deferred_stable
) {
8398 for (auto& txc
: b
->txcs
) {
8399 bluestore_deferred_transaction_t
& wt
= *txc
.deferred_txn
;
8400 if (!wt
.released
.empty()) {
8401 // kraken replay compat only
8402 txc
.released
= wt
.released
;
8403 dout(10) << __func__
<< " deferred txn has released "
8405 << " (we just upgraded from kraken) on " << &txc
<< dendl
;
8406 _txc_finalize_kv(&txc
, synct
);
8408 // cleanup the deferred
8410 get_deferred_key(wt
.seq
, &key
);
8411 synct
->rm_single_key(PREFIX_DEFERRED
, key
);
8415 // submit synct synchronously (block and wait for it to commit)
8416 int r
= cct
->_conf
->bluestore_debug_omit_kv_commit
? 0 : db
->submit_transaction_sync(synct
);
8420 nid_max
= new_nid_max
;
8421 dout(10) << __func__
<< " nid_max now " << nid_max
<< dendl
;
8423 if (new_blobid_max
) {
8424 blobid_max
= new_blobid_max
;
8425 dout(10) << __func__
<< " blobid_max now " << blobid_max
<< dendl
;
8429 utime_t finish
= ceph_clock_now();
8430 utime_t dur_flush
= after_flush
- start
;
8431 utime_t dur_kv
= finish
- after_flush
;
8432 utime_t dur
= finish
- start
;
8433 dout(20) << __func__
<< " committed " << kv_committing
.size()
8434 << " cleaned " << deferred_stable
.size()
8436 << " (" << dur_flush
<< " flush + " << dur_kv
<< " kv commit)"
8438 logger
->tinc(l_bluestore_kv_flush_lat
, dur_flush
);
8439 logger
->tinc(l_bluestore_kv_commit_lat
, dur_kv
);
8440 logger
->tinc(l_bluestore_kv_lat
, dur
);
8444 if (!bluefs_gift_extents
.empty()) {
8445 _commit_bluefs_freespace(bluefs_gift_extents
);
8447 for (auto p
= bluefs_extents_reclaiming
.begin();
8448 p
!= bluefs_extents_reclaiming
.end();
8450 dout(20) << __func__
<< " releasing old bluefs 0x" << std::hex
8451 << p
.get_start() << "~" << p
.get_len() << std::dec
8453 alloc
->release(p
.get_start(), p
.get_len());
8455 bluefs_extents_reclaiming
.clear();
8459 std::unique_lock
<std::mutex
> m(kv_finalize_lock
);
8460 if (kv_committing_to_finalize
.empty()) {
8461 kv_committing_to_finalize
.swap(kv_committing
);
8463 kv_committing_to_finalize
.insert(
8464 kv_committing_to_finalize
.end(),
8465 kv_committing
.begin(),
8466 kv_committing
.end());
8467 kv_committing
.clear();
8469 if (deferred_stable_to_finalize
.empty()) {
8470 deferred_stable_to_finalize
.swap(deferred_stable
);
8472 deferred_stable_to_finalize
.insert(
8473 deferred_stable_to_finalize
.end(),
8474 deferred_stable
.begin(),
8475 deferred_stable
.end());
8476 deferred_stable
.clear();
8478 kv_finalize_cond
.notify_one();
8482 // previously deferred "done" are now "stable" by virtue of this
8484 deferred_stable_queue
.swap(deferred_done
);
8487 dout(10) << __func__
<< " finish" << dendl
;
8488 kv_sync_started
= false;
8491 void BlueStore::_kv_finalize_thread()
8493 deque
<TransContext
*> kv_committed
;
8494 deque
<DeferredBatch
*> deferred_stable
;
8495 dout(10) << __func__
<< " start" << dendl
;
8496 std::unique_lock
<std::mutex
> l(kv_finalize_lock
);
8497 assert(!kv_finalize_started
);
8498 kv_finalize_started
= true;
8499 kv_finalize_cond
.notify_all();
8501 assert(kv_committed
.empty());
8502 assert(deferred_stable
.empty());
8503 if (kv_committing_to_finalize
.empty() &&
8504 deferred_stable_to_finalize
.empty()) {
8505 if (kv_finalize_stop
)
8507 dout(20) << __func__
<< " sleep" << dendl
;
8508 kv_finalize_cond
.wait(l
);
8509 dout(20) << __func__
<< " wake" << dendl
;
8511 kv_committed
.swap(kv_committing_to_finalize
);
8512 deferred_stable
.swap(deferred_stable_to_finalize
);
8514 dout(20) << __func__
<< " kv_committed " << kv_committed
<< dendl
;
8515 dout(20) << __func__
<< " deferred_stable " << deferred_stable
<< dendl
;
8517 while (!kv_committed
.empty()) {
8518 TransContext
*txc
= kv_committed
.front();
8519 assert(txc
->state
== TransContext::STATE_KV_SUBMITTED
);
8520 _txc_state_proc(txc
);
8521 kv_committed
.pop_front();
8524 for (auto b
: deferred_stable
) {
8525 auto p
= b
->txcs
.begin();
8526 while (p
!= b
->txcs
.end()) {
8527 TransContext
*txc
= &*p
;
8528 p
= b
->txcs
.erase(p
); // unlink here because
8529 _txc_state_proc(txc
); // this may destroy txc
8533 deferred_stable
.clear();
8535 if (!deferred_aggressive
) {
8536 if (deferred_queue_size
>= deferred_batch_ops
.load() ||
8537 throttle_deferred_bytes
.past_midpoint()) {
8538 deferred_try_submit();
8542 // this is as good a place as any ...
8543 _reap_collections();
8548 dout(10) << __func__
<< " finish" << dendl
;
8549 kv_finalize_started
= false;
8552 bluestore_deferred_op_t
*BlueStore::_get_deferred_op(
8553 TransContext
*txc
, OnodeRef o
)
8555 if (!txc
->deferred_txn
) {
8556 txc
->deferred_txn
= new bluestore_deferred_transaction_t
;
8558 txc
->deferred_txn
->ops
.push_back(bluestore_deferred_op_t());
8559 return &txc
->deferred_txn
->ops
.back();
8562 void BlueStore::_deferred_queue(TransContext
*txc
)
8564 dout(20) << __func__
<< " txc " << txc
<< " osr " << txc
->osr
<< dendl
;
8565 deferred_lock
.lock();
8566 if (!txc
->osr
->deferred_pending
&&
8567 !txc
->osr
->deferred_running
) {
8568 deferred_queue
.push_back(*txc
->osr
);
8570 if (!txc
->osr
->deferred_pending
) {
8571 txc
->osr
->deferred_pending
= new DeferredBatch(cct
, txc
->osr
.get());
8573 ++deferred_queue_size
;
8574 txc
->osr
->deferred_pending
->txcs
.push_back(*txc
);
8575 bluestore_deferred_transaction_t
& wt
= *txc
->deferred_txn
;
8576 for (auto opi
= wt
.ops
.begin(); opi
!= wt
.ops
.end(); ++opi
) {
8577 const auto& op
= *opi
;
8578 assert(op
.op
== bluestore_deferred_op_t::OP_WRITE
);
8579 bufferlist::const_iterator p
= op
.data
.begin();
8580 for (auto e
: op
.extents
) {
8581 txc
->osr
->deferred_pending
->prepare_write(
8582 cct
, wt
.seq
, e
.offset
, e
.length
, p
);
8585 if (deferred_aggressive
&&
8586 !txc
->osr
->deferred_running
) {
8587 _deferred_submit_unlock(txc
->osr
.get());
8589 deferred_lock
.unlock();
8593 void BlueStore::deferred_try_submit()
8595 dout(20) << __func__
<< " " << deferred_queue
.size() << " osrs, "
8596 << deferred_queue_size
<< " txcs" << dendl
;
8597 std::lock_guard
<std::mutex
> l(deferred_lock
);
8598 vector
<OpSequencerRef
> osrs
;
8599 osrs
.reserve(deferred_queue
.size());
8600 for (auto& osr
: deferred_queue
) {
8601 osrs
.push_back(&osr
);
8603 for (auto& osr
: osrs
) {
8604 if (osr
->deferred_pending
&& !osr
->deferred_running
) {
8605 _deferred_submit_unlock(osr
.get());
8606 deferred_lock
.lock();
8611 void BlueStore::_deferred_submit_unlock(OpSequencer
*osr
)
8613 dout(10) << __func__
<< " osr " << osr
8614 << " " << osr
->deferred_pending
->iomap
.size() << " ios pending "
8616 assert(osr
->deferred_pending
);
8617 assert(!osr
->deferred_running
);
8619 auto b
= osr
->deferred_pending
;
8620 deferred_queue_size
-= b
->seq_bytes
.size();
8621 assert(deferred_queue_size
>= 0);
8623 osr
->deferred_running
= osr
->deferred_pending
;
8624 osr
->deferred_pending
= nullptr;
8626 uint64_t start
= 0, pos
= 0;
8628 auto i
= b
->iomap
.begin();
8630 if (i
== b
->iomap
.end() || i
->first
!= pos
) {
8632 dout(20) << __func__
<< " write 0x" << std::hex
8633 << start
<< "~" << bl
.length()
8634 << " crc " << bl
.crc32c(-1) << std::dec
<< dendl
;
8635 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
8636 logger
->inc(l_bluestore_deferred_write_ops
);
8637 logger
->inc(l_bluestore_deferred_write_bytes
, bl
.length());
8638 int r
= bdev
->aio_write(start
, bl
, &b
->ioc
, false);
8642 if (i
== b
->iomap
.end()) {
8649 dout(20) << __func__
<< " seq " << i
->second
.seq
<< " 0x"
8650 << std::hex
<< pos
<< "~" << i
->second
.bl
.length() << std::dec
8655 pos
+= i
->second
.bl
.length();
8656 bl
.claim_append(i
->second
.bl
);
8660 // demote to deferred_submit_lock, then drop that too
8661 std::lock_guard
<std::mutex
> l(deferred_submit_lock
);
8662 deferred_lock
.unlock();
8663 bdev
->aio_submit(&b
->ioc
);
8666 void BlueStore::_deferred_aio_finish(OpSequencer
*osr
)
8668 dout(10) << __func__
<< " osr " << osr
<< dendl
;
8669 assert(osr
->deferred_running
);
8670 DeferredBatch
*b
= osr
->deferred_running
;
8673 std::lock_guard
<std::mutex
> l(deferred_lock
);
8674 assert(osr
->deferred_running
== b
);
8675 osr
->deferred_running
= nullptr;
8676 if (!osr
->deferred_pending
) {
8677 auto q
= deferred_queue
.iterator_to(*osr
);
8678 deferred_queue
.erase(q
);
8679 } else if (deferred_aggressive
) {
8680 dout(20) << __func__
<< " queuing async deferred_try_submit" << dendl
;
8681 finishers
[0]->queue(new FunctionContext([&](int) {
8682 deferred_try_submit();
8689 std::lock_guard
<std::mutex
> l2(osr
->qlock
);
8690 for (auto& i
: b
->txcs
) {
8691 TransContext
*txc
= &i
;
8692 txc
->state
= TransContext::STATE_DEFERRED_CLEANUP
;
8695 osr
->qcond
.notify_all();
8696 throttle_deferred_bytes
.put(costs
);
8697 std::lock_guard
<std::mutex
> l(kv_lock
);
8698 deferred_done_queue
.emplace_back(b
);
8701 // in the normal case, do not bother waking up the kv thread; it will
8702 // catch us on the next commit anyway.
8703 if (deferred_aggressive
) {
8704 std::lock_guard
<std::mutex
> l(kv_lock
);
8705 kv_cond
.notify_one();
8709 int BlueStore::_deferred_replay()
8711 dout(10) << __func__
<< " start" << dendl
;
8712 OpSequencerRef osr
= new OpSequencer(cct
, this);
8715 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_DEFERRED
);
8716 for (it
->lower_bound(string()); it
->valid(); it
->next(), ++count
) {
8717 dout(20) << __func__
<< " replay " << pretty_binary_string(it
->key())
8719 bluestore_deferred_transaction_t
*deferred_txn
=
8720 new bluestore_deferred_transaction_t
;
8721 bufferlist bl
= it
->value();
8722 bufferlist::iterator p
= bl
.begin();
8724 ::decode(*deferred_txn
, p
);
8725 } catch (buffer::error
& e
) {
8726 derr
<< __func__
<< " failed to decode deferred txn "
8727 << pretty_binary_string(it
->key()) << dendl
;
8728 delete deferred_txn
;
8732 TransContext
*txc
= _txc_create(osr
.get());
8733 txc
->deferred_txn
= deferred_txn
;
8734 txc
->state
= TransContext::STATE_KV_DONE
;
8735 _txc_state_proc(txc
);
8738 dout(20) << __func__
<< " draining osr" << dendl
;
8741 dout(10) << __func__
<< " completed " << count
<< " events" << dendl
;
8745 // ---------------------------
8748 int BlueStore::queue_transactions(
8750 vector
<Transaction
>& tls
,
8752 ThreadPool::TPHandle
*handle
)
8755 Context
*onreadable
;
8757 Context
*onreadable_sync
;
8758 ObjectStore::Transaction::collect_contexts(
8759 tls
, &onreadable
, &ondisk
, &onreadable_sync
);
8761 if (cct
->_conf
->objectstore_blackhole
) {
8762 dout(0) << __func__
<< " objectstore_blackhole = TRUE, dropping transaction"
8766 delete onreadable_sync
;
8769 utime_t start
= ceph_clock_now();
8770 // set up the sequencer
8774 osr
= static_cast<OpSequencer
*>(posr
->p
.get());
8775 dout(10) << __func__
<< " existing " << osr
<< " " << *osr
<< dendl
;
8777 osr
= new OpSequencer(cct
, this);
8780 dout(10) << __func__
<< " new " << osr
<< " " << *osr
<< dendl
;
8784 TransContext
*txc
= _txc_create(osr
);
8785 txc
->onreadable
= onreadable
;
8786 txc
->onreadable_sync
= onreadable_sync
;
8787 txc
->oncommit
= ondisk
;
8789 for (vector
<Transaction
>::iterator p
= tls
.begin(); p
!= tls
.end(); ++p
) {
8791 txc
->bytes
+= (*p
).get_num_bytes();
8792 _txc_add_transaction(txc
, &(*p
));
8794 _txc_calc_cost(txc
);
8796 _txc_write_nodes(txc
, txc
->t
);
8798 // journal deferred items
8799 if (txc
->deferred_txn
) {
8800 txc
->deferred_txn
->seq
= ++deferred_seq
;
8802 ::encode(*txc
->deferred_txn
, bl
);
8804 get_deferred_key(txc
->deferred_txn
->seq
, &key
);
8805 txc
->t
->set(PREFIX_DEFERRED
, key
, bl
);
8808 _txc_finalize_kv(txc
, txc
->t
);
8810 handle
->suspend_tp_timeout();
8812 utime_t tstart
= ceph_clock_now();
8813 throttle_bytes
.get(txc
->cost
);
8814 if (txc
->deferred_txn
) {
8815 // ensure we do not block here because of deferred writes
8816 if (!throttle_deferred_bytes
.get_or_fail(txc
->cost
)) {
8817 deferred_try_submit();
8818 throttle_deferred_bytes
.get(txc
->cost
);
8821 utime_t tend
= ceph_clock_now();
8824 handle
->reset_tp_timeout();
8826 logger
->inc(l_bluestore_txc
);
8829 _txc_state_proc(txc
);
8831 logger
->tinc(l_bluestore_submit_lat
, ceph_clock_now() - start
);
8832 logger
->tinc(l_bluestore_throttle_lat
, tend
- tstart
);
8836 void BlueStore::_txc_aio_submit(TransContext
*txc
)
8838 dout(10) << __func__
<< " txc " << txc
<< dendl
;
8839 bdev
->aio_submit(&txc
->ioc
);
8842 void BlueStore::_txc_add_transaction(TransContext
*txc
, Transaction
*t
)
8844 Transaction::iterator i
= t
->begin();
8846 _dump_transaction(t
);
8848 vector
<CollectionRef
> cvec(i
.colls
.size());
8850 for (vector
<coll_t
>::iterator p
= i
.colls
.begin(); p
!= i
.colls
.end();
8852 cvec
[j
] = _get_collection(*p
);
8854 vector
<OnodeRef
> ovec(i
.objects
.size());
8856 for (int pos
= 0; i
.have_op(); ++pos
) {
8857 Transaction::Op
*op
= i
.decode_op();
8861 if (op
->op
== Transaction::OP_NOP
)
8864 // collection operations
8865 CollectionRef
&c
= cvec
[op
->cid
];
8867 case Transaction::OP_RMCOLL
:
8869 const coll_t
&cid
= i
.get_cid(op
->cid
);
8870 r
= _remove_collection(txc
, cid
, &c
);
8876 case Transaction::OP_MKCOLL
:
8879 const coll_t
&cid
= i
.get_cid(op
->cid
);
8880 r
= _create_collection(txc
, cid
, op
->split_bits
, &c
);
8886 case Transaction::OP_SPLIT_COLLECTION
:
8887 assert(0 == "deprecated");
8890 case Transaction::OP_SPLIT_COLLECTION2
:
8892 uint32_t bits
= op
->split_bits
;
8893 uint32_t rem
= op
->split_rem
;
8894 r
= _split_collection(txc
, c
, cvec
[op
->dest_cid
], bits
, rem
);
8900 case Transaction::OP_COLL_HINT
:
8902 uint32_t type
= op
->hint_type
;
8905 bufferlist::iterator hiter
= hint
.begin();
8906 if (type
== Transaction::COLL_HINT_EXPECTED_NUM_OBJECTS
) {
8909 ::decode(pg_num
, hiter
);
8910 ::decode(num_objs
, hiter
);
8911 dout(10) << __func__
<< " collection hint objects is a no-op, "
8912 << " pg_num " << pg_num
<< " num_objects " << num_objs
8916 dout(10) << __func__
<< " unknown collection hint " << type
<< dendl
;
8922 case Transaction::OP_COLL_SETATTR
:
8926 case Transaction::OP_COLL_RMATTR
:
8930 case Transaction::OP_COLL_RENAME
:
8931 assert(0 == "not implemented");
8935 derr
<< __func__
<< " error " << cpp_strerror(r
)
8936 << " not handled on operation " << op
->op
8937 << " (op " << pos
<< ", counting from 0)" << dendl
;
8938 _dump_transaction(t
, 0);
8939 assert(0 == "unexpected error");
8942 // these operations implicity create the object
8943 bool create
= false;
8944 if (op
->op
== Transaction::OP_TOUCH
||
8945 op
->op
== Transaction::OP_WRITE
||
8946 op
->op
== Transaction::OP_ZERO
) {
8950 // object operations
8951 RWLock::WLocker
l(c
->lock
);
8952 OnodeRef
&o
= ovec
[op
->oid
];
8954 ghobject_t oid
= i
.get_oid(op
->oid
);
8955 o
= c
->get_onode(oid
, create
);
8957 if (!create
&& (!o
|| !o
->exists
)) {
8958 dout(10) << __func__
<< " op " << op
->op
<< " got ENOENT on "
8959 << i
.get_oid(op
->oid
) << dendl
;
8965 case Transaction::OP_TOUCH
:
8966 r
= _touch(txc
, c
, o
);
8969 case Transaction::OP_WRITE
:
8971 uint64_t off
= op
->off
;
8972 uint64_t len
= op
->len
;
8973 uint32_t fadvise_flags
= i
.get_fadvise_flags();
8976 r
= _write(txc
, c
, o
, off
, len
, bl
, fadvise_flags
);
8980 case Transaction::OP_ZERO
:
8982 uint64_t off
= op
->off
;
8983 uint64_t len
= op
->len
;
8984 r
= _zero(txc
, c
, o
, off
, len
);
8988 case Transaction::OP_TRIMCACHE
:
8990 // deprecated, no-op
8994 case Transaction::OP_TRUNCATE
:
8996 uint64_t off
= op
->off
;
8997 _truncate(txc
, c
, o
, off
);
9001 case Transaction::OP_REMOVE
:
9003 r
= _remove(txc
, c
, o
);
9007 case Transaction::OP_SETATTR
:
9009 string name
= i
.decode_string();
9012 r
= _setattr(txc
, c
, o
, name
, bp
);
9016 case Transaction::OP_SETATTRS
:
9018 map
<string
, bufferptr
> aset
;
9019 i
.decode_attrset(aset
);
9020 r
= _setattrs(txc
, c
, o
, aset
);
9024 case Transaction::OP_RMATTR
:
9026 string name
= i
.decode_string();
9027 r
= _rmattr(txc
, c
, o
, name
);
9031 case Transaction::OP_RMATTRS
:
9033 r
= _rmattrs(txc
, c
, o
);
9037 case Transaction::OP_CLONE
:
9039 OnodeRef
& no
= ovec
[op
->dest_oid
];
9041 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9042 no
= c
->get_onode(noid
, true);
9044 r
= _clone(txc
, c
, o
, no
);
9048 case Transaction::OP_CLONERANGE
:
9049 assert(0 == "deprecated");
9052 case Transaction::OP_CLONERANGE2
:
9054 OnodeRef
& no
= ovec
[op
->dest_oid
];
9056 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9057 no
= c
->get_onode(noid
, true);
9059 uint64_t srcoff
= op
->off
;
9060 uint64_t len
= op
->len
;
9061 uint64_t dstoff
= op
->dest_off
;
9062 r
= _clone_range(txc
, c
, o
, no
, srcoff
, len
, dstoff
);
9066 case Transaction::OP_COLL_ADD
:
9067 assert(0 == "not implemented");
9070 case Transaction::OP_COLL_REMOVE
:
9071 assert(0 == "not implemented");
9074 case Transaction::OP_COLL_MOVE
:
9075 assert(0 == "deprecated");
9078 case Transaction::OP_COLL_MOVE_RENAME
:
9079 case Transaction::OP_TRY_RENAME
:
9081 assert(op
->cid
== op
->dest_cid
);
9082 const ghobject_t
& noid
= i
.get_oid(op
->dest_oid
);
9083 OnodeRef
& no
= ovec
[op
->dest_oid
];
9085 no
= c
->get_onode(noid
, false);
9087 r
= _rename(txc
, c
, o
, no
, noid
);
9091 case Transaction::OP_OMAP_CLEAR
:
9093 r
= _omap_clear(txc
, c
, o
);
9096 case Transaction::OP_OMAP_SETKEYS
:
9099 i
.decode_attrset_bl(&aset_bl
);
9100 r
= _omap_setkeys(txc
, c
, o
, aset_bl
);
9103 case Transaction::OP_OMAP_RMKEYS
:
9106 i
.decode_keyset_bl(&keys_bl
);
9107 r
= _omap_rmkeys(txc
, c
, o
, keys_bl
);
9110 case Transaction::OP_OMAP_RMKEYRANGE
:
9113 first
= i
.decode_string();
9114 last
= i
.decode_string();
9115 r
= _omap_rmkey_range(txc
, c
, o
, first
, last
);
9118 case Transaction::OP_OMAP_SETHEADER
:
9122 r
= _omap_setheader(txc
, c
, o
, bl
);
9126 case Transaction::OP_SETALLOCHINT
:
9128 r
= _set_alloc_hint(txc
, c
, o
,
9129 op
->expected_object_size
,
9130 op
->expected_write_size
,
9131 op
->alloc_hint_flags
);
9136 derr
<< __func__
<< "bad op " << op
->op
<< dendl
;
9144 if (r
== -ENOENT
&& !(op
->op
== Transaction::OP_CLONERANGE
||
9145 op
->op
== Transaction::OP_CLONE
||
9146 op
->op
== Transaction::OP_CLONERANGE2
||
9147 op
->op
== Transaction::OP_COLL_ADD
||
9148 op
->op
== Transaction::OP_SETATTR
||
9149 op
->op
== Transaction::OP_SETATTRS
||
9150 op
->op
== Transaction::OP_RMATTR
||
9151 op
->op
== Transaction::OP_OMAP_SETKEYS
||
9152 op
->op
== Transaction::OP_OMAP_RMKEYS
||
9153 op
->op
== Transaction::OP_OMAP_RMKEYRANGE
||
9154 op
->op
== Transaction::OP_OMAP_SETHEADER
))
9155 // -ENOENT is usually okay
9161 const char *msg
= "unexpected error code";
9163 if (r
== -ENOENT
&& (op
->op
== Transaction::OP_CLONERANGE
||
9164 op
->op
== Transaction::OP_CLONE
||
9165 op
->op
== Transaction::OP_CLONERANGE2
))
9166 msg
= "ENOENT on clone suggests osd bug";
9169 // For now, if we hit _any_ ENOSPC, crash, before we do any damage
9170 // by partially applying transactions.
9171 msg
= "ENOSPC from bluestore, misconfigured cluster";
9173 if (r
== -ENOTEMPTY
) {
9174 msg
= "ENOTEMPTY suggests garbage data in osd data dir";
9177 derr
<< __func__
<< " error " << cpp_strerror(r
)
9178 << " not handled on operation " << op
->op
9179 << " (op " << pos
<< ", counting from 0)"
9181 derr
<< msg
<< dendl
;
9182 _dump_transaction(t
, 0);
9183 assert(0 == "unexpected error");
9191 // -----------------
9194 int BlueStore::_touch(TransContext
*txc
,
9198 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
9200 _assign_nid(txc
, o
);
9201 txc
->write_onode(o
);
9202 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
9206 void BlueStore::_dump_onode(OnodeRef o
, int log_level
)
9208 if (!cct
->_conf
->subsys
.should_gather(ceph_subsys_bluestore
, log_level
))
9210 dout(log_level
) << __func__
<< " " << o
<< " " << o
->oid
9211 << " nid " << o
->onode
.nid
9212 << " size 0x" << std::hex
<< o
->onode
.size
9213 << " (" << std::dec
<< o
->onode
.size
<< ")"
9214 << " expected_object_size " << o
->onode
.expected_object_size
9215 << " expected_write_size " << o
->onode
.expected_write_size
9216 << " in " << o
->onode
.extent_map_shards
.size() << " shards"
9217 << ", " << o
->extent_map
.spanning_blob_map
.size()
9218 << " spanning blobs"
9220 for (auto p
= o
->onode
.attrs
.begin();
9221 p
!= o
->onode
.attrs
.end();
9223 dout(log_level
) << __func__
<< " attr " << p
->first
9224 << " len " << p
->second
.length() << dendl
;
9226 _dump_extent_map(o
->extent_map
, log_level
);
9229 void BlueStore::_dump_extent_map(ExtentMap
&em
, int log_level
)
9232 for (auto& s
: em
.shards
) {
9233 dout(log_level
) << __func__
<< " shard " << *s
.shard_info
9234 << (s
.loaded
? " (loaded)" : "")
9235 << (s
.dirty
? " (dirty)" : "")
9238 for (auto& e
: em
.extent_map
) {
9239 dout(log_level
) << __func__
<< " " << e
<< dendl
;
9240 assert(e
.logical_offset
>= pos
);
9241 pos
= e
.logical_offset
+ e
.length
;
9242 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
9243 if (blob
.has_csum()) {
9245 unsigned n
= blob
.get_csum_count();
9246 for (unsigned i
= 0; i
< n
; ++i
)
9247 v
.push_back(blob
.get_csum_item(i
));
9248 dout(log_level
) << __func__
<< " csum: " << std::hex
<< v
<< std::dec
9251 std::lock_guard
<std::recursive_mutex
> l(e
.blob
->shared_blob
->get_cache()->lock
);
9252 for (auto& i
: e
.blob
->shared_blob
->bc
.buffer_map
) {
9253 dout(log_level
) << __func__
<< " 0x" << std::hex
<< i
.first
9254 << "~" << i
.second
->length
<< std::dec
9255 << " " << *i
.second
<< dendl
;
9260 void BlueStore::_dump_transaction(Transaction
*t
, int log_level
)
9262 dout(log_level
) << " transaction dump:\n";
9263 JSONFormatter
f(true);
9264 f
.open_object_section("transaction");
9271 void BlueStore::_pad_zeros(
9272 bufferlist
*bl
, uint64_t *offset
,
9273 uint64_t chunk_size
)
9275 auto length
= bl
->length();
9276 dout(30) << __func__
<< " 0x" << std::hex
<< *offset
<< "~" << length
9277 << " chunk_size 0x" << chunk_size
<< std::dec
<< dendl
;
9278 dout(40) << "before:\n";
9279 bl
->hexdump(*_dout
);
9282 size_t front_pad
= *offset
% chunk_size
;
9283 size_t back_pad
= 0;
9284 size_t pad_count
= 0;
9286 size_t front_copy
= MIN(chunk_size
- front_pad
, length
);
9287 bufferptr z
= buffer::create_page_aligned(chunk_size
);
9288 z
.zero(0, front_pad
, false);
9289 pad_count
+= front_pad
;
9290 bl
->copy(0, front_copy
, z
.c_str() + front_pad
);
9291 if (front_copy
+ front_pad
< chunk_size
) {
9292 back_pad
= chunk_size
- (length
+ front_pad
);
9293 z
.zero(front_pad
+ length
, back_pad
, false);
9294 pad_count
+= back_pad
;
9298 t
.substr_of(old
, front_copy
, length
- front_copy
);
9300 bl
->claim_append(t
);
9301 *offset
-= front_pad
;
9302 length
+= pad_count
;
9306 uint64_t end
= *offset
+ length
;
9307 unsigned back_copy
= end
% chunk_size
;
9309 assert(back_pad
== 0);
9310 back_pad
= chunk_size
- back_copy
;
9311 assert(back_copy
<= length
);
9312 bufferptr
tail(chunk_size
);
9313 bl
->copy(length
- back_copy
, back_copy
, tail
.c_str());
9314 tail
.zero(back_copy
, back_pad
, false);
9317 bl
->substr_of(old
, 0, length
- back_copy
);
9320 pad_count
+= back_pad
;
9322 dout(20) << __func__
<< " pad 0x" << std::hex
<< front_pad
<< " + 0x"
9323 << back_pad
<< " on front/back, now 0x" << *offset
<< "~"
9324 << length
<< std::dec
<< dendl
;
9325 dout(40) << "after:\n";
9326 bl
->hexdump(*_dout
);
9329 logger
->inc(l_bluestore_write_pad_bytes
, pad_count
);
9330 assert(bl
->length() == length
);
9333 void BlueStore::_do_write_small(
9337 uint64_t offset
, uint64_t length
,
9338 bufferlist::iterator
& blp
,
9341 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9342 << std::dec
<< dendl
;
9343 assert(length
< min_alloc_size
);
9344 uint64_t end_offs
= offset
+ length
;
9346 logger
->inc(l_bluestore_write_small
);
9347 logger
->inc(l_bluestore_write_small_bytes
, length
);
9350 blp
.copy(length
, bl
);
9352 // Look for an existing mutable blob we can use.
9353 auto begin
= o
->extent_map
.extent_map
.begin();
9354 auto end
= o
->extent_map
.extent_map
.end();
9355 auto ep
= o
->extent_map
.seek_lextent(offset
);
9358 if (ep
->blob_end() <= offset
) {
9363 if (prev_ep
!= begin
) {
9366 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9369 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9370 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9371 uint32_t alloc_len
= min_alloc_size
;
9372 auto offset0
= P2ALIGN(offset
, alloc_len
);
9376 // search suitable extent in both forward and reverse direction in
9377 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9378 // then check if blob can be reused via can_reuse_blob func or apply
9379 // direct/deferred write (the latter for extents including or higher
9380 // than 'offset' only).
9384 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9385 BlobRef b
= ep
->blob
;
9386 auto bstart
= ep
->blob_start();
9387 dout(20) << __func__
<< " considering " << *b
9388 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9389 if (bstart
>= end_offs
) {
9390 dout(20) << __func__
<< " ignoring distant " << *b
<< dendl
;
9391 } else if (!b
->get_blob().is_mutable()) {
9392 dout(20) << __func__
<< " ignoring immutable " << *b
<< dendl
;
9393 } else if (ep
->logical_offset
% min_alloc_size
!=
9394 ep
->blob_offset
% min_alloc_size
) {
9395 dout(20) << __func__
<< " ignoring offset-skewed " << *b
<< dendl
;
9397 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9398 // can we pad our head/tail out with zeros?
9399 uint64_t head_pad
, tail_pad
;
9400 head_pad
= P2PHASE(offset
, chunk_size
);
9401 tail_pad
= P2NPHASE(end_offs
, chunk_size
);
9402 if (head_pad
|| tail_pad
) {
9403 o
->extent_map
.fault_range(db
, offset
- head_pad
,
9404 end_offs
- offset
+ head_pad
+ tail_pad
);
9407 o
->extent_map
.has_any_lextents(offset
- head_pad
, chunk_size
)) {
9410 if (tail_pad
&& o
->extent_map
.has_any_lextents(end_offs
, tail_pad
)) {
9414 uint64_t b_off
= offset
- head_pad
- bstart
;
9415 uint64_t b_len
= length
+ head_pad
+ tail_pad
;
9417 // direct write into unused blocks of an existing mutable blob?
9418 if ((b_off
% chunk_size
== 0 && b_len
% chunk_size
== 0) &&
9419 b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9420 b
->get_blob().is_unused(b_off
, b_len
) &&
9421 b
->get_blob().is_allocated(b_off
, b_len
)) {
9422 _apply_padding(head_pad
, tail_pad
, bl
);
9424 dout(20) << __func__
<< " write to unused 0x" << std::hex
9425 << b_off
<< "~" << b_len
9426 << " pad 0x" << head_pad
<< " + 0x" << tail_pad
9427 << std::dec
<< " of mutable " << *b
<< dendl
;
9428 _buffer_cache_write(txc
, b
, b_off
, bl
,
9429 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9431 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
9432 if (b_len
<= prefer_deferred_size
) {
9433 dout(20) << __func__
<< " deferring small 0x" << std::hex
9434 << b_len
<< std::dec
<< " unused write via deferred" << dendl
;
9435 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9436 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9439 [&](uint64_t offset
, uint64_t length
) {
9440 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9445 b
->get_blob().map_bl(
9447 [&](uint64_t offset
, bufferlist
& t
) {
9448 bdev
->aio_write(offset
, t
,
9449 &txc
->ioc
, wctx
->buffered
);
9453 b
->dirty_blob().calc_csum(b_off
, bl
);
9454 dout(20) << __func__
<< " lex old " << *ep
<< dendl
;
9455 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, b_off
+ head_pad
, length
,
9457 &wctx
->old_extents
);
9458 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9459 txc
->statfs_delta
.stored() += le
->length
;
9460 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9461 logger
->inc(l_bluestore_write_small_unused
);
9464 // read some data to fill out the chunk?
9465 uint64_t head_read
= P2PHASE(b_off
, chunk_size
);
9466 uint64_t tail_read
= P2NPHASE(b_off
+ b_len
, chunk_size
);
9467 if ((head_read
|| tail_read
) &&
9468 (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
+ tail_read
) &&
9469 head_read
+ tail_read
< min_alloc_size
) {
9471 b_len
+= head_read
+ tail_read
;
9474 head_read
= tail_read
= 0;
9477 // chunk-aligned deferred overwrite?
9478 if (b
->get_blob().get_ondisk_length() >= b_off
+ b_len
&&
9479 b_off
% chunk_size
== 0 &&
9480 b_len
% chunk_size
== 0 &&
9481 b
->get_blob().is_allocated(b_off
, b_len
)) {
9483 _apply_padding(head_pad
, tail_pad
, bl
);
9485 dout(20) << __func__
<< " reading head 0x" << std::hex
<< head_read
9486 << " and tail 0x" << tail_read
<< std::dec
<< dendl
;
9489 int r
= _do_read(c
.get(), o
, offset
- head_pad
- head_read
, head_read
,
9491 assert(r
>= 0 && r
<= (int)head_read
);
9492 size_t zlen
= head_read
- r
;
9494 head_bl
.append_zero(zlen
);
9495 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9497 bl
.claim_prepend(head_bl
);
9498 logger
->inc(l_bluestore_write_penalty_read_ops
);
9502 int r
= _do_read(c
.get(), o
, offset
+ length
+ tail_pad
, tail_read
,
9504 assert(r
>= 0 && r
<= (int)tail_read
);
9505 size_t zlen
= tail_read
- r
;
9507 tail_bl
.append_zero(zlen
);
9508 logger
->inc(l_bluestore_write_pad_bytes
, zlen
);
9510 bl
.claim_append(tail_bl
);
9511 logger
->inc(l_bluestore_write_penalty_read_ops
);
9513 logger
->inc(l_bluestore_write_small_pre_read
);
9515 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9516 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9517 _buffer_cache_write(txc
, b
, b_off
, bl
,
9518 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9520 int r
= b
->get_blob().map(
9522 [&](uint64_t offset
, uint64_t length
) {
9523 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9527 if (b
->get_blob().csum_type
) {
9528 b
->dirty_blob().calc_csum(b_off
, bl
);
9531 dout(20) << __func__
<< " deferred write 0x" << std::hex
<< b_off
<< "~"
9532 << b_len
<< std::dec
<< " of mutable " << *b
9533 << " at " << op
->extents
<< dendl
;
9534 Extent
*le
= o
->extent_map
.set_lextent(c
, offset
, offset
- bstart
, length
,
9535 b
, &wctx
->old_extents
);
9536 b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9537 txc
->statfs_delta
.stored() += le
->length
;
9538 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9539 logger
->inc(l_bluestore_write_small_deferred
);
9542 // try to reuse blob if we can
9543 if (b
->can_reuse_blob(min_alloc_size
,
9547 assert(alloc_len
== min_alloc_size
); // expecting data always
9548 // fit into reused blob
9549 // Need to check for pending writes desiring to
9550 // reuse the same pextent. The rationale is that during GC two chunks
9551 // from garbage blobs(compressed?) can share logical space within the same
9552 // AU. That's in turn might be caused by unaligned len in clone_range2.
9553 // Hence the second write will fail in an attempt to reuse blob at
9554 // do_alloc_write().
9555 if (!wctx
->has_conflict(b
,
9557 offset0
+ alloc_len
,
9560 // we can't reuse pad_head/pad_tail since they might be truncated
9561 // due to existent extents
9562 uint64_t b_off
= offset
- bstart
;
9563 uint64_t b_off0
= b_off
;
9564 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9566 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9567 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9568 << " (0x" << b_off
<< "~" << length
<< ")"
9569 << std::dec
<< dendl
;
9571 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9572 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9574 logger
->inc(l_bluestore_write_small_unused
);
9581 } // if (ep != end && ep->logical_offset < offset + max_bsize)
9583 // check extent for reuse in reverse order
9584 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9585 BlobRef b
= prev_ep
->blob
;
9586 auto bstart
= prev_ep
->blob_start();
9587 dout(20) << __func__
<< " considering " << *b
9588 << " bstart 0x" << std::hex
<< bstart
<< std::dec
<< dendl
;
9589 if (b
->can_reuse_blob(min_alloc_size
,
9593 assert(alloc_len
== min_alloc_size
); // expecting data always
9594 // fit into reused blob
9595 // Need to check for pending writes desiring to
9596 // reuse the same pextent. The rationale is that during GC two chunks
9597 // from garbage blobs(compressed?) can share logical space within the same
9598 // AU. That's in turn might be caused by unaligned len in clone_range2.
9599 // Hence the second write will fail in an attempt to reuse blob at
9600 // do_alloc_write().
9601 if (!wctx
->has_conflict(b
,
9603 offset0
+ alloc_len
,
9606 uint64_t chunk_size
= b
->get_blob().get_chunk_size(block_size
);
9607 uint64_t b_off
= offset
- bstart
;
9608 uint64_t b_off0
= b_off
;
9609 _pad_zeros(&bl
, &b_off0
, chunk_size
);
9611 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9612 << " (0x" << b_off0
<< "~" << bl
.length() << ")"
9613 << " (0x" << b_off
<< "~" << length
<< ")"
9614 << std::dec
<< dendl
;
9616 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9617 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
,
9619 logger
->inc(l_bluestore_write_small_unused
);
9623 if (prev_ep
!= begin
) {
9627 prev_ep
= end
; // to avoid useless first extent re-check
9629 } // if (prev_ep != end && prev_ep->logical_offset >= min_off)
9630 } while (any_change
);
9634 BlobRef b
= c
->new_blob();
9635 uint64_t b_off
= P2PHASE(offset
, alloc_len
);
9636 uint64_t b_off0
= b_off
;
9637 _pad_zeros(&bl
, &b_off0
, block_size
);
9638 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9639 wctx
->write(offset
, b
, alloc_len
, b_off0
, bl
, b_off
, length
, true, true);
9640 logger
->inc(l_bluestore_write_small_new
);
9645 void BlueStore::_do_write_big(
9649 uint64_t offset
, uint64_t length
,
9650 bufferlist::iterator
& blp
,
9653 dout(10) << __func__
<< " 0x" << std::hex
<< offset
<< "~" << length
9654 << " target_blob_size 0x" << wctx
->target_blob_size
<< std::dec
9655 << " compress " << (int)wctx
->compress
9657 logger
->inc(l_bluestore_write_big
);
9658 logger
->inc(l_bluestore_write_big_bytes
, length
);
9659 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
->old_extents
);
9660 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9661 while (length
> 0) {
9662 bool new_blob
= false;
9663 uint32_t l
= MIN(max_bsize
, length
);
9667 //attempting to reuse existing blob
9668 if (!wctx
->compress
) {
9669 // look for an existing mutable blob we can reuse
9670 auto begin
= o
->extent_map
.extent_map
.begin();
9671 auto end
= o
->extent_map
.extent_map
.end();
9672 auto ep
= o
->extent_map
.seek_lextent(offset
);
9674 if (prev_ep
!= begin
) {
9677 prev_ep
= end
; // to avoid this extent check as it's a duplicate
9679 auto min_off
= offset
>= max_bsize
? offset
- max_bsize
: 0;
9680 // search suitable extent in both forward and reverse direction in
9681 // [offset - target_max_blob_size, offset + target_max_blob_size] range
9682 // then check if blob can be reused via can_reuse_blob func.
9686 if (ep
!= end
&& ep
->logical_offset
< offset
+ max_bsize
) {
9687 if (offset
>= ep
->blob_start() &&
9688 ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9689 offset
- ep
->blob_start(),
9692 b_off
= offset
- ep
->blob_start();
9693 prev_ep
= end
; // to avoid check below
9694 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9695 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9702 if (prev_ep
!= end
&& prev_ep
->logical_offset
>= min_off
) {
9703 if (prev_ep
->blob
->can_reuse_blob(min_alloc_size
, max_bsize
,
9704 offset
- prev_ep
->blob_start(),
9707 b_off
= offset
- prev_ep
->blob_start();
9708 dout(20) << __func__
<< " reuse blob " << *b
<< std::hex
9709 << " (0x" << b_off
<< "~" << l
<< ")" << std::dec
<< dendl
;
9710 } else if (prev_ep
!= begin
) {
9714 prev_ep
= end
; // to avoid useless first extent re-check
9717 } while (b
== nullptr && any_change
);
9727 wctx
->write(offset
, b
, l
, b_off
, t
, b_off
, l
, false, new_blob
);
9730 logger
->inc(l_bluestore_write_big_blobs
);
9734 int BlueStore::_do_alloc_write(
9740 dout(20) << __func__
<< " txc " << txc
9741 << " " << wctx
->writes
.size() << " blobs"
9745 auto max_bsize
= MAX(wctx
->target_blob_size
, min_alloc_size
);
9746 for (auto &wi
: wctx
->writes
) {
9747 need
+= wi
.blob_length
;
9749 int r
= alloc
->reserve(need
);
9751 derr
<< __func__
<< " failed to reserve 0x" << std::hex
<< need
<< std::dec
9759 if (wctx
->compress
) {
9761 "compression_algorithm",
9765 if (coll
->pool_opts
.get(pool_opts_t::COMPRESSION_ALGORITHM
, &val
)) {
9766 CompressorRef cp
= compressor
;
9767 if (!cp
|| cp
->get_type_name() != val
) {
9768 cp
= Compressor::create(cct
, val
);
9770 return boost::optional
<CompressorRef
>(cp
);
9772 return boost::optional
<CompressorRef
>();
9776 crr
= select_option(
9777 "compression_required_ratio",
9778 cct
->_conf
->bluestore_compression_required_ratio
,
9781 if(coll
->pool_opts
.get(pool_opts_t::COMPRESSION_REQUIRED_RATIO
, &val
)) {
9782 return boost::optional
<double>(val
);
9784 return boost::optional
<double>();
9790 int csum
= csum_type
.load();
9791 csum
= select_option(
9796 if(coll
->pool_opts
.get(pool_opts_t::CSUM_TYPE
, &val
)) {
9797 return boost::optional
<int>(val
);
9799 return boost::optional
<int>();
9803 for (auto& wi
: wctx
->writes
) {
9805 bluestore_blob_t
& dblob
= b
->dirty_blob();
9806 uint64_t b_off
= wi
.b_off
;
9807 bufferlist
*l
= &wi
.bl
;
9808 uint64_t final_length
= wi
.blob_length
;
9809 uint64_t csum_length
= wi
.blob_length
;
9810 unsigned csum_order
= block_size_order
;
9811 bufferlist compressed_bl
;
9812 bool compressed
= false;
9813 if(c
&& wi
.blob_length
> min_alloc_size
) {
9815 utime_t start
= ceph_clock_now();
9819 assert(wi
.blob_length
== l
->length());
9820 bluestore_compression_header_t chdr
;
9821 chdr
.type
= c
->get_type();
9822 // FIXME: memory alignment here is bad
9825 r
= c
->compress(*l
, t
);
9828 chdr
.length
= t
.length();
9829 ::encode(chdr
, compressed_bl
);
9830 compressed_bl
.claim_append(t
);
9831 uint64_t rawlen
= compressed_bl
.length();
9832 uint64_t newlen
= P2ROUNDUP(rawlen
, min_alloc_size
);
9833 uint64_t want_len_raw
= final_length
* crr
;
9834 uint64_t want_len
= P2ROUNDUP(want_len_raw
, min_alloc_size
);
9835 if (newlen
<= want_len
&& newlen
< final_length
) {
9836 // Cool. We compressed at least as much as we were hoping to.
9837 // pad out to min_alloc_size
9838 compressed_bl
.append_zero(newlen
- rawlen
);
9839 logger
->inc(l_bluestore_write_pad_bytes
, newlen
- rawlen
);
9840 dout(20) << __func__
<< std::hex
<< " compressed 0x" << wi
.blob_length
9841 << " -> 0x" << rawlen
<< " => 0x" << newlen
9842 << " with " << c
->get_type()
9843 << std::dec
<< dendl
;
9844 txc
->statfs_delta
.compressed() += rawlen
;
9845 txc
->statfs_delta
.compressed_original() += l
->length();
9846 txc
->statfs_delta
.compressed_allocated() += newlen
;
9848 final_length
= newlen
;
9849 csum_length
= newlen
;
9850 csum_order
= ctz(newlen
);
9851 dblob
.set_compressed(wi
.blob_length
, rawlen
);
9853 logger
->inc(l_bluestore_compress_success_count
);
9855 dout(20) << __func__
<< std::hex
<< " 0x" << l
->length()
9856 << " compressed to 0x" << rawlen
<< " -> 0x" << newlen
9857 << " with " << c
->get_type()
9858 << ", which is more than required 0x" << want_len_raw
9859 << " -> 0x" << want_len
9860 << ", leaving uncompressed"
9861 << std::dec
<< dendl
;
9862 logger
->inc(l_bluestore_compress_rejected_count
);
9864 logger
->tinc(l_bluestore_compress_lat
,
9865 ceph_clock_now() - start
);
9867 if (!compressed
&& wi
.new_blob
) {
9868 // initialize newly created blob only
9869 assert(dblob
.is_mutable());
9870 if (l
->length() != wi
.blob_length
) {
9871 // hrm, maybe we could do better here, but let's not bother.
9872 dout(20) << __func__
<< " forcing csum_order to block_size_order "
9873 << block_size_order
<< dendl
;
9874 csum_order
= block_size_order
;
9876 csum_order
= std::min(wctx
->csum_order
, ctz(l
->length()));
9878 // try to align blob with max_blob_size to improve
9879 // its reuse ratio, e.g. in case of reverse write
9880 uint32_t suggested_boff
=
9881 (wi
.logical_offset
- (wi
.b_off0
- wi
.b_off
)) % max_bsize
;
9882 if ((suggested_boff
% (1 << csum_order
)) == 0 &&
9883 suggested_boff
+ final_length
<= max_bsize
&&
9884 suggested_boff
> b_off
) {
9885 dout(20) << __func__
<< " forcing blob_offset to "
9886 << std::hex
<< suggested_boff
<< std::dec
<< dendl
;
9887 assert(suggested_boff
>= b_off
);
9888 csum_length
+= suggested_boff
- b_off
;
9889 b_off
= suggested_boff
;
9893 AllocExtentVector extents
;
9894 extents
.reserve(4); // 4 should be (more than) enough for most allocations
9895 int64_t got
= alloc
->allocate(final_length
, min_alloc_size
,
9896 max_alloc_size
.load(),
9898 assert(got
== (int64_t)final_length
);
9900 txc
->statfs_delta
.allocated() += got
;
9901 for (auto& p
: extents
) {
9902 bluestore_pextent_t e
= bluestore_pextent_t(p
);
9903 txc
->allocated
.insert(e
.offset
, e
.length
);
9906 dblob
.allocated(P2ALIGN(b_off
, min_alloc_size
), final_length
, extents
);
9908 dout(20) << __func__
<< " blob " << *b
9909 << " csum_type " << Checksummer::get_csum_type_string(csum
)
9910 << " csum_order " << csum_order
9911 << " csum_length 0x" << std::hex
<< csum_length
<< std::dec
9914 if (csum
!= Checksummer::CSUM_NONE
) {
9915 if (!dblob
.has_csum()) {
9916 dblob
.init_csum(csum
, csum_order
, csum_length
);
9918 dblob
.calc_csum(b_off
, *l
);
9920 if (wi
.mark_unused
) {
9921 auto b_end
= b_off
+ wi
.bl
.length();
9923 dblob
.add_unused(0, b_off
);
9925 if (b_end
< wi
.blob_length
) {
9926 dblob
.add_unused(b_end
, wi
.blob_length
- b_end
);
9930 Extent
*le
= o
->extent_map
.set_lextent(coll
, wi
.logical_offset
,
9931 b_off
+ (wi
.b_off0
- wi
.b_off
),
9935 wi
.b
->dirty_blob().mark_used(le
->blob_offset
, le
->length
);
9936 txc
->statfs_delta
.stored() += le
->length
;
9937 dout(20) << __func__
<< " lex " << *le
<< dendl
;
9938 _buffer_cache_write(txc
, wi
.b
, b_off
, wi
.bl
,
9939 wctx
->buffered
? 0 : Buffer::FLAG_NOCACHE
);
9942 if (!g_conf
->bluestore_debug_omit_block_device_write
) {
9943 if (l
->length() <= prefer_deferred_size
.load()) {
9944 dout(20) << __func__
<< " deferring small 0x" << std::hex
9945 << l
->length() << std::dec
<< " write via deferred" << dendl
;
9946 bluestore_deferred_op_t
*op
= _get_deferred_op(txc
, o
);
9947 op
->op
= bluestore_deferred_op_t::OP_WRITE
;
9948 int r
= b
->get_blob().map(
9950 [&](uint64_t offset
, uint64_t length
) {
9951 op
->extents
.emplace_back(bluestore_pextent_t(offset
, length
));
9957 b
->get_blob().map_bl(
9959 [&](uint64_t offset
, bufferlist
& t
) {
9960 bdev
->aio_write(offset
, t
, &txc
->ioc
, false);
9966 alloc
->unreserve(need
);
9971 void BlueStore::_wctx_finish(
9976 set
<SharedBlob
*> *maybe_unshared_blobs
)
9978 auto oep
= wctx
->old_extents
.begin();
9979 while (oep
!= wctx
->old_extents
.end()) {
9981 oep
= wctx
->old_extents
.erase(oep
);
9982 dout(20) << __func__
<< " lex_old " << lo
.e
<< dendl
;
9983 BlobRef b
= lo
.e
.blob
;
9984 const bluestore_blob_t
& blob
= b
->get_blob();
9985 if (blob
.is_compressed()) {
9986 if (lo
.blob_empty
) {
9987 txc
->statfs_delta
.compressed() -= blob
.get_compressed_payload_length();
9989 txc
->statfs_delta
.compressed_original() -= lo
.e
.length
;
9992 txc
->statfs_delta
.stored() -= lo
.e
.length
;
9994 dout(20) << __func__
<< " blob release " << r
<< dendl
;
9995 if (blob
.is_shared()) {
9996 PExtentVector final
;
9997 c
->load_shared_blob(b
->shared_blob
);
9999 b
->shared_blob
->put_ref(
10000 e
.offset
, e
.length
, &final
,
10001 b
->is_referenced() ? nullptr : maybe_unshared_blobs
);
10003 dout(20) << __func__
<< " shared_blob release " << final
10004 << " from " << *b
->shared_blob
<< dendl
;
10005 txc
->write_shared_blob(b
->shared_blob
);
10010 // we can't invalidate our logical extents as we drop them because
10011 // other lextents (either in our onode or others) may still
10012 // reference them. but we can throw out anything that is no
10013 // longer allocated. Note that this will leave behind edge bits
10014 // that are no longer referenced but not deallocated (until they
10015 // age out of the cache naturally).
10016 b
->discard_unallocated(c
.get());
10018 dout(20) << __func__
<< " release " << e
<< dendl
;
10019 txc
->released
.insert(e
.offset
, e
.length
);
10020 txc
->statfs_delta
.allocated() -= e
.length
;
10021 if (blob
.is_compressed()) {
10022 txc
->statfs_delta
.compressed_allocated() -= e
.length
;
10026 if (b
->is_spanning() && !b
->is_referenced()) {
10027 dout(20) << __func__
<< " spanning_blob_map removing empty " << *b
10029 o
->extent_map
.spanning_blob_map
.erase(b
->id
);
10034 void BlueStore::_do_write_data(
10041 WriteContext
*wctx
)
10043 uint64_t end
= offset
+ length
;
10044 bufferlist::iterator p
= bl
.begin();
10046 if (offset
/ min_alloc_size
== (end
- 1) / min_alloc_size
&&
10047 (length
!= min_alloc_size
)) {
10048 // we fall within the same block
10049 _do_write_small(txc
, c
, o
, offset
, length
, p
, wctx
);
10051 uint64_t head_offset
, head_length
;
10052 uint64_t middle_offset
, middle_length
;
10053 uint64_t tail_offset
, tail_length
;
10055 head_offset
= offset
;
10056 head_length
= P2NPHASE(offset
, min_alloc_size
);
10058 tail_offset
= P2ALIGN(end
, min_alloc_size
);
10059 tail_length
= P2PHASE(end
, min_alloc_size
);
10061 middle_offset
= head_offset
+ head_length
;
10062 middle_length
= length
- head_length
- tail_length
;
10065 _do_write_small(txc
, c
, o
, head_offset
, head_length
, p
, wctx
);
10068 if (middle_length
) {
10069 _do_write_big(txc
, c
, o
, middle_offset
, middle_length
, p
, wctx
);
10073 _do_write_small(txc
, c
, o
, tail_offset
, tail_length
, p
, wctx
);
10078 void BlueStore::_choose_write_options(
10081 uint32_t fadvise_flags
,
10082 WriteContext
*wctx
)
10084 if (fadvise_flags
& CEPH_OSD_OP_FLAG_FADVISE_WILLNEED
) {
10085 dout(20) << __func__
<< " will do buffered write" << dendl
;
10086 wctx
->buffered
= true;
10087 } else if (cct
->_conf
->bluestore_default_buffered_write
&&
10088 (fadvise_flags
& (CEPH_OSD_OP_FLAG_FADVISE_DONTNEED
|
10089 CEPH_OSD_OP_FLAG_FADVISE_NOCACHE
)) == 0) {
10090 dout(20) << __func__
<< " defaulting to buffered write" << dendl
;
10091 wctx
->buffered
= true;
10094 // apply basic csum block size
10095 wctx
->csum_order
= block_size_order
;
10097 // compression parameters
10098 unsigned alloc_hints
= o
->onode
.alloc_hint_flags
;
10099 auto cm
= select_option(
10100 "compression_mode",
10104 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MODE
, &val
)) {
10105 return boost::optional
<Compressor::CompressionMode
>(
10106 Compressor::get_comp_mode_type(val
));
10108 return boost::optional
<Compressor::CompressionMode
>();
10112 wctx
->compress
= (cm
!= Compressor::COMP_NONE
) &&
10113 ((cm
== Compressor::COMP_FORCE
) ||
10114 (cm
== Compressor::COMP_AGGRESSIVE
&&
10115 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_INCOMPRESSIBLE
) == 0) ||
10116 (cm
== Compressor::COMP_PASSIVE
&&
10117 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_COMPRESSIBLE
)));
10119 if ((alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_SEQUENTIAL_READ
) &&
10120 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_READ
) == 0 &&
10121 (alloc_hints
& (CEPH_OSD_ALLOC_HINT_FLAG_IMMUTABLE
|
10122 CEPH_OSD_ALLOC_HINT_FLAG_APPEND_ONLY
)) &&
10123 (alloc_hints
& CEPH_OSD_ALLOC_HINT_FLAG_RANDOM_WRITE
) == 0) {
10125 dout(20) << __func__
<< " will prefer large blob and csum sizes" << dendl
;
10127 if (o
->onode
.expected_write_size
) {
10128 wctx
->csum_order
= std::max(min_alloc_size_order
,
10129 (uint8_t)ctz(o
->onode
.expected_write_size
));
10131 wctx
->csum_order
= min_alloc_size_order
;
10134 if (wctx
->compress
) {
10135 wctx
->target_blob_size
= select_option(
10136 "compression_max_blob_size",
10137 comp_max_blob_size
.load(),
10140 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MAX_BLOB_SIZE
, &val
)) {
10141 return boost::optional
<uint64_t>((uint64_t)val
);
10143 return boost::optional
<uint64_t>();
10148 if (wctx
->compress
) {
10149 wctx
->target_blob_size
= select_option(
10150 "compression_min_blob_size",
10151 comp_min_blob_size
.load(),
10154 if(c
->pool_opts
.get(pool_opts_t::COMPRESSION_MIN_BLOB_SIZE
, &val
)) {
10155 return boost::optional
<uint64_t>((uint64_t)val
);
10157 return boost::optional
<uint64_t>();
10163 uint64_t max_bsize
= max_blob_size
.load();
10164 if (wctx
->target_blob_size
== 0 || wctx
->target_blob_size
> max_bsize
) {
10165 wctx
->target_blob_size
= max_bsize
;
10168 // set the min blob size floor at 2x the min_alloc_size, or else we
10169 // won't be able to allocate a smaller extent for the compressed
10171 if (wctx
->compress
&&
10172 wctx
->target_blob_size
< min_alloc_size
* 2) {
10173 wctx
->target_blob_size
= min_alloc_size
* 2;
10176 dout(20) << __func__
<< " prefer csum_order " << wctx
->csum_order
10177 << " target_blob_size 0x" << std::hex
<< wctx
->target_blob_size
10178 << std::dec
<< dendl
;
10181 int BlueStore::_do_gc(
10185 const GarbageCollector
& gc
,
10186 const WriteContext
& wctx
,
10187 uint64_t *dirty_start
,
10188 uint64_t *dirty_end
)
10190 auto& extents_to_collect
= gc
.get_extents_to_collect();
10192 WriteContext wctx_gc
;
10193 wctx_gc
.fork(wctx
); // make a clone for garbage collection
10195 for (auto it
= extents_to_collect
.begin();
10196 it
!= extents_to_collect
.end();
10199 int r
= _do_read(c
.get(), o
, it
->offset
, it
->length
, bl
, 0);
10200 assert(r
== (int)it
->length
);
10202 o
->extent_map
.fault_range(db
, it
->offset
, it
->length
);
10203 _do_write_data(txc
, c
, o
, it
->offset
, it
->length
, bl
, &wctx_gc
);
10204 logger
->inc(l_bluestore_gc_merged
, it
->length
);
10206 if (*dirty_start
> it
->offset
) {
10207 *dirty_start
= it
->offset
;
10210 if (*dirty_end
< it
->offset
+ it
->length
) {
10211 *dirty_end
= it
->offset
+ it
->length
;
10215 dout(30) << __func__
<< " alloc write" << dendl
;
10216 int r
= _do_alloc_write(txc
, c
, o
, &wctx_gc
);
10218 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10223 _wctx_finish(txc
, c
, o
, &wctx_gc
);
10227 int BlueStore::_do_write(
10234 uint32_t fadvise_flags
)
10238 dout(20) << __func__
10240 << " 0x" << std::hex
<< offset
<< "~" << length
10241 << " - have 0x" << o
->onode
.size
10242 << " (" << std::dec
<< o
->onode
.size
<< ")"
10244 << " fadvise_flags 0x" << std::hex
<< fadvise_flags
<< std::dec
10252 uint64_t end
= offset
+ length
;
10254 GarbageCollector
gc(c
->store
->cct
);
10256 auto dirty_start
= offset
;
10257 auto dirty_end
= end
;
10260 _choose_write_options(c
, o
, fadvise_flags
, &wctx
);
10261 o
->extent_map
.fault_range(db
, offset
, length
);
10262 _do_write_data(txc
, c
, o
, offset
, length
, bl
, &wctx
);
10263 r
= _do_alloc_write(txc
, c
, o
, &wctx
);
10265 derr
<< __func__
<< " _do_alloc_write failed with " << cpp_strerror(r
)
10270 // NB: _wctx_finish() will empty old_extents
10271 // so we must do gc estimation before that
10272 benefit
= gc
.estimate(offset
,
10278 _wctx_finish(txc
, c
, o
, &wctx
);
10279 if (end
> o
->onode
.size
) {
10280 dout(20) << __func__
<< " extending size to 0x" << std::hex
<< end
10281 << std::dec
<< dendl
;
10282 o
->onode
.size
= end
;
10285 if (benefit
>= g_conf
->bluestore_gc_enable_total_threshold
) {
10286 if (!gc
.get_extents_to_collect().empty()) {
10287 dout(20) << __func__
<< " perform garbage collection, "
10288 << "expected benefit = " << benefit
<< " AUs" << dendl
;
10289 r
= _do_gc(txc
, c
, o
, gc
, wctx
, &dirty_start
, &dirty_end
);
10291 derr
<< __func__
<< " _do_gc failed with " << cpp_strerror(r
)
10298 o
->extent_map
.compress_extent_map(dirty_start
, dirty_end
- dirty_start
);
10299 o
->extent_map
.dirty_range(dirty_start
, dirty_end
- dirty_start
);
10307 int BlueStore::_write(TransContext
*txc
,
10310 uint64_t offset
, size_t length
,
10312 uint32_t fadvise_flags
)
10314 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10315 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10317 _assign_nid(txc
, o
);
10318 int r
= _do_write(txc
, c
, o
, offset
, length
, bl
, fadvise_flags
);
10319 txc
->write_onode(o
);
10321 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10322 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10323 << " = " << r
<< dendl
;
10327 int BlueStore::_zero(TransContext
*txc
,
10330 uint64_t offset
, size_t length
)
10332 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10333 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10335 _assign_nid(txc
, o
);
10336 int r
= _do_zero(txc
, c
, o
, offset
, length
);
10337 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10338 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10339 << " = " << r
<< dendl
;
10343 int BlueStore::_do_zero(TransContext
*txc
,
10346 uint64_t offset
, size_t length
)
10348 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10349 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10356 o
->extent_map
.fault_range(db
, offset
, length
);
10357 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10358 o
->extent_map
.dirty_range(offset
, length
);
10359 _wctx_finish(txc
, c
, o
, &wctx
);
10361 if (offset
+ length
> o
->onode
.size
) {
10362 o
->onode
.size
= offset
+ length
;
10363 dout(20) << __func__
<< " extending size to " << offset
+ length
10366 txc
->write_onode(o
);
10368 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10369 << " 0x" << std::hex
<< offset
<< "~" << length
<< std::dec
10370 << " = " << r
<< dendl
;
10374 void BlueStore::_do_truncate(
10375 TransContext
*txc
, CollectionRef
& c
, OnodeRef o
, uint64_t offset
,
10376 set
<SharedBlob
*> *maybe_unshared_blobs
)
10378 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10379 << " 0x" << std::hex
<< offset
<< std::dec
<< dendl
;
10381 _dump_onode(o
, 30);
10383 if (offset
== o
->onode
.size
)
10386 if (offset
< o
->onode
.size
) {
10388 uint64_t length
= o
->onode
.size
- offset
;
10389 o
->extent_map
.fault_range(db
, offset
, length
);
10390 o
->extent_map
.punch_hole(c
, offset
, length
, &wctx
.old_extents
);
10391 o
->extent_map
.dirty_range(offset
, length
);
10392 _wctx_finish(txc
, c
, o
, &wctx
, maybe_unshared_blobs
);
10394 // if we have shards past EOF, ask for a reshard
10395 if (!o
->onode
.extent_map_shards
.empty() &&
10396 o
->onode
.extent_map_shards
.back().offset
>= offset
) {
10397 dout(10) << __func__
<< " request reshard past EOF" << dendl
;
10399 o
->extent_map
.request_reshard(offset
- 1, offset
+ length
);
10401 o
->extent_map
.request_reshard(0, length
);
10406 o
->onode
.size
= offset
;
10408 txc
->write_onode(o
);
10411 void BlueStore::_truncate(TransContext
*txc
,
10416 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10417 << " 0x" << std::hex
<< offset
<< std::dec
10419 _do_truncate(txc
, c
, o
, offset
);
10422 int BlueStore::_do_remove(
10427 set
<SharedBlob
*> maybe_unshared_blobs
;
10428 bool is_gen
= !o
->oid
.is_no_gen();
10429 _do_truncate(txc
, c
, o
, 0, is_gen
? &maybe_unshared_blobs
: nullptr);
10430 if (o
->onode
.has_omap()) {
10432 _do_omap_clear(txc
, o
->onode
.nid
);
10436 for (auto &s
: o
->extent_map
.shards
) {
10437 dout(20) << __func__
<< " removing shard 0x" << std::hex
10438 << s
.shard_info
->offset
<< std::dec
<< dendl
;
10439 generate_extent_shard_key_and_apply(o
->key
, s
.shard_info
->offset
, &key
,
10440 [&](const string
& final_key
) {
10441 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
10445 txc
->t
->rmkey(PREFIX_OBJ
, o
->key
.c_str(), o
->key
.size());
10447 o
->extent_map
.clear();
10448 o
->onode
= bluestore_onode_t();
10449 _debug_obj_on_delete(o
->oid
);
10451 if (!is_gen
|| maybe_unshared_blobs
.empty()) {
10455 // see if we can unshare blobs still referenced by the head
10456 dout(10) << __func__
<< " gen and maybe_unshared_blobs "
10457 << maybe_unshared_blobs
<< dendl
;
10458 ghobject_t nogen
= o
->oid
;
10459 nogen
.generation
= ghobject_t::NO_GEN
;
10460 OnodeRef h
= c
->onode_map
.lookup(nogen
);
10462 if (!h
|| !h
->exists
) {
10466 dout(20) << __func__
<< " checking for unshareable blobs on " << h
10467 << " " << h
->oid
<< dendl
;
10468 map
<SharedBlob
*,bluestore_extent_ref_map_t
> expect
;
10469 for (auto& e
: h
->extent_map
.extent_map
) {
10470 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10471 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10472 if (b
.is_shared() &&
10474 maybe_unshared_blobs
.count(sb
)) {
10475 b
.map(e
.blob_offset
, e
.length
, [&](uint64_t off
, uint64_t len
) {
10476 expect
[sb
].get(off
, len
);
10482 vector
<SharedBlob
*> unshared_blobs
;
10483 unshared_blobs
.reserve(maybe_unshared_blobs
.size());
10484 for (auto& p
: expect
) {
10485 dout(20) << " ? " << *p
.first
<< " vs " << p
.second
<< dendl
;
10486 if (p
.first
->persistent
->ref_map
== p
.second
) {
10487 SharedBlob
*sb
= p
.first
;
10488 dout(20) << __func__
<< " unsharing " << *sb
<< dendl
;
10489 unshared_blobs
.push_back(sb
);
10490 txc
->unshare_blob(sb
);
10491 uint64_t sbid
= c
->make_blob_unshared(sb
);
10493 get_shared_blob_key(sbid
, &key
);
10494 txc
->t
->rmkey(PREFIX_SHARED_BLOB
, key
);
10498 if (unshared_blobs
.empty()) {
10502 uint32_t b_start
= OBJECT_MAX_SIZE
;
10503 uint32_t b_end
= 0;
10504 for (auto& e
: h
->extent_map
.extent_map
) {
10505 const bluestore_blob_t
& b
= e
.blob
->get_blob();
10506 SharedBlob
*sb
= e
.blob
->shared_blob
.get();
10507 if (b
.is_shared() &&
10508 std::find(unshared_blobs
.begin(), unshared_blobs
.end(),
10509 sb
) != unshared_blobs
.end()) {
10510 dout(20) << __func__
<< " unsharing " << e
<< dendl
;
10511 bluestore_blob_t
& blob
= e
.blob
->dirty_blob();
10512 blob
.clear_flag(bluestore_blob_t::FLAG_SHARED
);
10513 if (e
.logical_offset
< b_start
) {
10514 b_start
= e
.logical_offset
;
10516 if (e
.logical_end() > b_end
) {
10517 b_end
= e
.logical_end();
10522 assert(b_end
> b_start
);
10523 h
->extent_map
.dirty_range(b_start
, b_end
- b_start
);
10524 txc
->write_onode(h
);
10529 int BlueStore::_remove(TransContext
*txc
,
10533 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10534 int r
= _do_remove(txc
, c
, o
);
10535 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10539 int BlueStore::_setattr(TransContext
*txc
,
10542 const string
& name
,
10545 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10546 << " " << name
<< " (" << val
.length() << " bytes)"
10549 if (val
.is_partial())
10550 o
->onode
.attrs
[name
.c_str()] = bufferptr(val
.c_str(), val
.length());
10552 o
->onode
.attrs
[name
.c_str()] = val
;
10553 txc
->write_onode(o
);
10554 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10555 << " " << name
<< " (" << val
.length() << " bytes)"
10556 << " = " << r
<< dendl
;
10560 int BlueStore::_setattrs(TransContext
*txc
,
10563 const map
<string
,bufferptr
>& aset
)
10565 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10566 << " " << aset
.size() << " keys"
10569 for (map
<string
,bufferptr
>::const_iterator p
= aset
.begin();
10570 p
!= aset
.end(); ++p
) {
10571 if (p
->second
.is_partial())
10572 o
->onode
.attrs
[p
->first
.c_str()] =
10573 bufferptr(p
->second
.c_str(), p
->second
.length());
10575 o
->onode
.attrs
[p
->first
.c_str()] = p
->second
;
10577 txc
->write_onode(o
);
10578 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10579 << " " << aset
.size() << " keys"
10580 << " = " << r
<< dendl
;
10585 int BlueStore::_rmattr(TransContext
*txc
,
10588 const string
& name
)
10590 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10591 << " " << name
<< dendl
;
10593 auto it
= o
->onode
.attrs
.find(name
.c_str());
10594 if (it
== o
->onode
.attrs
.end())
10597 o
->onode
.attrs
.erase(it
);
10598 txc
->write_onode(o
);
10601 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10602 << " " << name
<< " = " << r
<< dendl
;
10606 int BlueStore::_rmattrs(TransContext
*txc
,
10610 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10613 if (o
->onode
.attrs
.empty())
10616 o
->onode
.attrs
.clear();
10617 txc
->write_onode(o
);
10620 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10624 void BlueStore::_do_omap_clear(TransContext
*txc
, uint64_t id
)
10626 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
10627 string prefix
, tail
;
10628 get_omap_header(id
, &prefix
);
10629 get_omap_tail(id
, &tail
);
10630 it
->lower_bound(prefix
);
10631 while (it
->valid()) {
10632 if (it
->key() >= tail
) {
10633 dout(30) << __func__
<< " stop at " << pretty_binary_string(tail
)
10637 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
10638 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
10643 int BlueStore::_omap_clear(TransContext
*txc
,
10647 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10649 if (o
->onode
.has_omap()) {
10651 _do_omap_clear(txc
, o
->onode
.nid
);
10652 o
->onode
.clear_omap_flag();
10653 txc
->write_onode(o
);
10655 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10659 int BlueStore::_omap_setkeys(TransContext
*txc
,
10664 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10666 bufferlist::iterator p
= bl
.begin();
10668 if (!o
->onode
.has_omap()) {
10669 o
->onode
.set_omap_flag();
10670 txc
->write_onode(o
);
10672 txc
->note_modified_object(o
);
10675 _key_encode_u64(o
->onode
.nid
, &final_key
);
10676 final_key
.push_back('.');
10682 ::decode(value
, p
);
10683 final_key
.resize(9); // keep prefix
10685 dout(30) << __func__
<< " " << pretty_binary_string(final_key
)
10686 << " <- " << key
<< dendl
;
10687 txc
->t
->set(PREFIX_OMAP
, final_key
, value
);
10690 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10694 int BlueStore::_omap_setheader(TransContext
*txc
,
10699 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10702 if (!o
->onode
.has_omap()) {
10703 o
->onode
.set_omap_flag();
10704 txc
->write_onode(o
);
10706 txc
->note_modified_object(o
);
10708 get_omap_header(o
->onode
.nid
, &key
);
10709 txc
->t
->set(PREFIX_OMAP
, key
, bl
);
10711 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10715 int BlueStore::_omap_rmkeys(TransContext
*txc
,
10720 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10722 bufferlist::iterator p
= bl
.begin();
10726 if (!o
->onode
.has_omap()) {
10729 _key_encode_u64(o
->onode
.nid
, &final_key
);
10730 final_key
.push_back('.');
10735 final_key
.resize(9); // keep prefix
10737 dout(30) << __func__
<< " rm " << pretty_binary_string(final_key
)
10738 << " <- " << key
<< dendl
;
10739 txc
->t
->rmkey(PREFIX_OMAP
, final_key
);
10741 txc
->note_modified_object(o
);
10744 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10748 int BlueStore::_omap_rmkey_range(TransContext
*txc
,
10751 const string
& first
, const string
& last
)
10753 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< dendl
;
10754 KeyValueDB::Iterator it
;
10755 string key_first
, key_last
;
10757 if (!o
->onode
.has_omap()) {
10761 it
= db
->get_iterator(PREFIX_OMAP
);
10762 get_omap_key(o
->onode
.nid
, first
, &key_first
);
10763 get_omap_key(o
->onode
.nid
, last
, &key_last
);
10764 it
->lower_bound(key_first
);
10765 while (it
->valid()) {
10766 if (it
->key() >= key_last
) {
10767 dout(30) << __func__
<< " stop at " << pretty_binary_string(key_last
)
10771 txc
->t
->rmkey(PREFIX_OMAP
, it
->key());
10772 dout(30) << __func__
<< " rm " << pretty_binary_string(it
->key()) << dendl
;
10775 txc
->note_modified_object(o
);
10778 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
<< " = " << r
<< dendl
;
10782 int BlueStore::_set_alloc_hint(
10786 uint64_t expected_object_size
,
10787 uint64_t expected_write_size
,
10790 dout(15) << __func__
<< " " << c
->cid
<< " " << o
->oid
10791 << " object_size " << expected_object_size
10792 << " write_size " << expected_write_size
10793 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
10796 o
->onode
.expected_object_size
= expected_object_size
;
10797 o
->onode
.expected_write_size
= expected_write_size
;
10798 o
->onode
.alloc_hint_flags
= flags
;
10799 txc
->write_onode(o
);
10800 dout(10) << __func__
<< " " << c
->cid
<< " " << o
->oid
10801 << " object_size " << expected_object_size
10802 << " write_size " << expected_write_size
10803 << " flags " << ceph_osd_alloc_hint_flag_string(flags
)
10804 << " = " << r
<< dendl
;
10808 int BlueStore::_clone(TransContext
*txc
,
10813 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
10814 << newo
->oid
<< dendl
;
10816 if (oldo
->oid
.hobj
.get_hash() != newo
->oid
.hobj
.get_hash()) {
10817 derr
<< __func__
<< " mismatched hash on " << oldo
->oid
10818 << " and " << newo
->oid
<< dendl
;
10822 _assign_nid(txc
, newo
);
10826 _do_truncate(txc
, c
, newo
, 0);
10827 if (cct
->_conf
->bluestore_clone_cow
) {
10828 _do_clone_range(txc
, c
, oldo
, newo
, 0, oldo
->onode
.size
, 0);
10831 r
= _do_read(c
.get(), oldo
, 0, oldo
->onode
.size
, bl
, 0);
10834 r
= _do_write(txc
, c
, newo
, 0, oldo
->onode
.size
, bl
, 0);
10840 newo
->onode
.attrs
= oldo
->onode
.attrs
;
10843 if (newo
->onode
.has_omap()) {
10844 dout(20) << __func__
<< " clearing old omap data" << dendl
;
10846 _do_omap_clear(txc
, newo
->onode
.nid
);
10848 if (oldo
->onode
.has_omap()) {
10849 dout(20) << __func__
<< " copying omap data" << dendl
;
10850 if (!newo
->onode
.has_omap()) {
10851 newo
->onode
.set_omap_flag();
10853 KeyValueDB::Iterator it
= db
->get_iterator(PREFIX_OMAP
);
10855 get_omap_header(oldo
->onode
.nid
, &head
);
10856 get_omap_tail(oldo
->onode
.nid
, &tail
);
10857 it
->lower_bound(head
);
10858 while (it
->valid()) {
10859 if (it
->key() >= tail
) {
10860 dout(30) << __func__
<< " reached tail" << dendl
;
10863 dout(30) << __func__
<< " got header/data "
10864 << pretty_binary_string(it
->key()) << dendl
;
10866 rewrite_omap_key(newo
->onode
.nid
, it
->key(), &key
);
10867 txc
->t
->set(PREFIX_OMAP
, key
, it
->value());
10872 newo
->onode
.clear_omap_flag();
10875 txc
->write_onode(newo
);
10879 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
10880 << newo
->oid
<< " = " << r
<< dendl
;
10884 int BlueStore::_do_clone_range(
10893 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
10895 << " 0x" << std::hex
<< srcoff
<< "~" << length
<< " -> "
10896 << " 0x" << dstoff
<< "~" << length
<< std::dec
<< dendl
;
10897 oldo
->extent_map
.fault_range(db
, srcoff
, length
);
10898 newo
->extent_map
.fault_range(db
, dstoff
, length
);
10902 // hmm, this could go into an ExtentMap::dup() method.
10903 vector
<BlobRef
> id_to_blob(oldo
->extent_map
.extent_map
.size());
10904 for (auto &e
: oldo
->extent_map
.extent_map
) {
10905 e
.blob
->last_encoded_id
= -1;
10908 uint64_t end
= srcoff
+ length
;
10909 uint32_t dirty_range_begin
= 0;
10910 uint32_t dirty_range_end
= 0;
10911 for (auto ep
= oldo
->extent_map
.seek_lextent(srcoff
);
10912 ep
!= oldo
->extent_map
.extent_map
.end();
10915 if (e
.logical_offset
>= end
) {
10918 dout(20) << __func__
<< " src " << e
<< dendl
;
10920 bool blob_duped
= true;
10921 if (e
.blob
->last_encoded_id
>= 0) {
10922 // blob is already duped
10923 cb
= id_to_blob
[e
.blob
->last_encoded_id
];
10924 blob_duped
= false;
10927 const bluestore_blob_t
& blob
= e
.blob
->get_blob();
10928 // make sure it is shared
10929 if (!blob
.is_shared()) {
10930 c
->make_blob_shared(_assign_blobid(txc
), e
.blob
);
10931 if (dirty_range_begin
== 0) {
10932 dirty_range_begin
= e
.logical_offset
;
10934 assert(e
.logical_end() > 0);
10935 // -1 to exclude next potential shard
10936 dirty_range_end
= e
.logical_end() - 1;
10938 c
->load_shared_blob(e
.blob
->shared_blob
);
10941 e
.blob
->last_encoded_id
= n
;
10942 id_to_blob
[n
] = cb
;
10944 // bump the extent refs on the copied blob's extents
10945 for (auto p
: blob
.get_extents()) {
10946 if (p
.is_valid()) {
10947 e
.blob
->shared_blob
->get_ref(p
.offset
, p
.length
);
10950 txc
->write_shared_blob(e
.blob
->shared_blob
);
10951 dout(20) << __func__
<< " new " << *cb
<< dendl
;
10954 int skip_front
, skip_back
;
10955 if (e
.logical_offset
< srcoff
) {
10956 skip_front
= srcoff
- e
.logical_offset
;
10960 if (e
.logical_end() > end
) {
10961 skip_back
= e
.logical_end() - end
;
10965 Extent
*ne
= new Extent(e
.logical_offset
+ skip_front
+ dstoff
- srcoff
,
10966 e
.blob_offset
+ skip_front
,
10967 e
.length
- skip_front
- skip_back
, cb
);
10968 newo
->extent_map
.extent_map
.insert(*ne
);
10969 ne
->blob
->get_ref(c
.get(), ne
->blob_offset
, ne
->length
);
10970 // fixme: we may leave parts of new blob unreferenced that could
10971 // be freed (relative to the shared_blob).
10972 txc
->statfs_delta
.stored() += ne
->length
;
10973 if (e
.blob
->get_blob().is_compressed()) {
10974 txc
->statfs_delta
.compressed_original() += ne
->length
;
10976 txc
->statfs_delta
.compressed() +=
10977 cb
->get_blob().get_compressed_payload_length();
10980 dout(20) << __func__
<< " dst " << *ne
<< dendl
;
10983 if (dirty_range_end
> dirty_range_begin
) {
10984 oldo
->extent_map
.dirty_range(dirty_range_begin
,
10985 dirty_range_end
- dirty_range_begin
);
10986 txc
->write_onode(oldo
);
10988 txc
->write_onode(newo
);
10990 if (dstoff
+ length
> newo
->onode
.size
) {
10991 newo
->onode
.size
= dstoff
+ length
;
10993 newo
->extent_map
.dirty_range(dstoff
, length
);
10999 int BlueStore::_clone_range(TransContext
*txc
,
11003 uint64_t srcoff
, uint64_t length
, uint64_t dstoff
)
11005 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11006 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11007 << " to offset 0x" << dstoff
<< std::dec
<< dendl
;
11010 if (srcoff
+ length
> oldo
->onode
.size
) {
11015 _assign_nid(txc
, newo
);
11018 if (cct
->_conf
->bluestore_clone_cow
) {
11019 _do_zero(txc
, c
, newo
, dstoff
, length
);
11020 _do_clone_range(txc
, c
, oldo
, newo
, srcoff
, length
, dstoff
);
11023 r
= _do_read(c
.get(), oldo
, srcoff
, length
, bl
, 0);
11026 r
= _do_write(txc
, c
, newo
, dstoff
, bl
.length(), bl
, 0);
11032 txc
->write_onode(newo
);
11036 dout(10) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11037 << newo
->oid
<< " from 0x" << std::hex
<< srcoff
<< "~" << length
11038 << " to offset 0x" << dstoff
<< std::dec
11039 << " = " << r
<< dendl
;
11043 int BlueStore::_rename(TransContext
*txc
,
11047 const ghobject_t
& new_oid
)
11049 dout(15) << __func__
<< " " << c
->cid
<< " " << oldo
->oid
<< " -> "
11050 << new_oid
<< dendl
;
11052 ghobject_t old_oid
= oldo
->oid
;
11053 mempool::bluestore_cache_other::string new_okey
;
11056 if (newo
->exists
) {
11060 assert(txc
->onodes
.count(newo
) == 0);
11063 txc
->t
->rmkey(PREFIX_OBJ
, oldo
->key
.c_str(), oldo
->key
.size());
11067 oldo
->extent_map
.fault_range(db
, 0, oldo
->onode
.size
);
11068 get_object_key(cct
, new_oid
, &new_okey
);
11070 for (auto &s
: oldo
->extent_map
.shards
) {
11071 generate_extent_shard_key_and_apply(oldo
->key
, s
.shard_info
->offset
, &key
,
11072 [&](const string
& final_key
) {
11073 txc
->t
->rmkey(PREFIX_OBJ
, final_key
);
11081 txc
->write_onode(newo
);
11083 // this adjusts oldo->{oid,key}, and reset oldo to a fresh empty
11084 // Onode in the old slot
11085 c
->onode_map
.rename(oldo
, old_oid
, new_oid
, new_okey
);
11089 dout(10) << __func__
<< " " << c
->cid
<< " " << old_oid
<< " -> "
11090 << new_oid
<< " = " << r
<< dendl
;
11096 int BlueStore::_create_collection(
11102 dout(15) << __func__
<< " " << cid
<< " bits " << bits
<< dendl
;
11107 RWLock::WLocker
l(coll_lock
);
11115 cache_shards
[cid
.hash_to_shard(cache_shards
.size())],
11117 (*c
)->cnode
.bits
= bits
;
11118 coll_map
[cid
] = *c
;
11120 ::encode((*c
)->cnode
, bl
);
11121 txc
->t
->set(PREFIX_COLL
, stringify(cid
), bl
);
11125 dout(10) << __func__
<< " " << cid
<< " bits " << bits
<< " = " << r
<< dendl
;
11129 int BlueStore::_remove_collection(TransContext
*txc
, const coll_t
&cid
,
11132 dout(15) << __func__
<< " " << cid
<< dendl
;
11136 RWLock::WLocker
l(coll_lock
);
11141 size_t nonexistent_count
= 0;
11142 assert((*c
)->exists
);
11143 if ((*c
)->onode_map
.map_any([&](OnodeRef o
) {
11145 dout(10) << __func__
<< " " << o
->oid
<< " " << o
11146 << " exists in onode_map" << dendl
;
11149 ++nonexistent_count
;
11156 vector
<ghobject_t
> ls
;
11158 // Enumerate onodes in db, up to nonexistent_count + 1
11159 // then check if all of them are marked as non-existent.
11160 // Bypass the check if returned number is greater than nonexistent_count
11161 r
= _collection_list(c
->get(), ghobject_t(), ghobject_t::get_max(),
11162 nonexistent_count
+ 1, &ls
, &next
);
11164 bool exists
= false; //ls.size() > nonexistent_count;
11165 for (auto it
= ls
.begin(); !exists
&& it
< ls
.end(); ++it
) {
11166 dout(10) << __func__
<< " oid " << *it
<< dendl
;
11167 auto onode
= (*c
)->onode_map
.lookup(*it
);
11168 exists
= !onode
|| onode
->exists
;
11170 dout(10) << __func__
<< " " << *it
11171 << " exists in db" << dendl
;
11175 coll_map
.erase(cid
);
11176 txc
->removed_collections
.push_back(*c
);
11177 (*c
)->exists
= false;
11179 txc
->t
->rmkey(PREFIX_COLL
, stringify(cid
));
11182 dout(10) << __func__
<< " " << cid
11183 << " is non-empty" << dendl
;
11190 dout(10) << __func__
<< " " << cid
<< " = " << r
<< dendl
;
11194 int BlueStore::_split_collection(TransContext
*txc
,
11197 unsigned bits
, int rem
)
11199 dout(15) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11200 << " bits " << bits
<< dendl
;
11201 RWLock::WLocker
l(c
->lock
);
11202 RWLock::WLocker
l2(d
->lock
);
11205 // flush all previous deferred writes on this sequencer. this is a bit
11206 // heavyweight, but we need to make sure all deferred writes complete
11207 // before we split as the new collection's sequencer may need to order
11208 // this after those writes, and we don't bother with the complexity of
11209 // moving those TransContexts over to the new osr.
11210 _osr_drain_preceding(txc
);
11212 // move any cached items (onodes and referenced shared blobs) that will
11213 // belong to the child collection post-split. leave everything else behind.
11214 // this may include things that don't strictly belong to the now-smaller
11215 // parent split, but the OSD will always send us a split for every new
11218 spg_t pgid
, dest_pgid
;
11219 bool is_pg
= c
->cid
.is_pg(&pgid
);
11221 is_pg
= d
->cid
.is_pg(&dest_pgid
);
11224 // the destination should initially be empty.
11225 assert(d
->onode_map
.empty());
11226 assert(d
->shared_blob_set
.empty());
11227 assert(d
->cnode
.bits
== bits
);
11229 c
->split_cache(d
.get());
11231 // adjust bits. note that this will be redundant for all but the first
11232 // split call for this parent (first child).
11233 c
->cnode
.bits
= bits
;
11234 assert(d
->cnode
.bits
== bits
);
11238 ::encode(c
->cnode
, bl
);
11239 txc
->t
->set(PREFIX_COLL
, stringify(c
->cid
), bl
);
11241 dout(10) << __func__
<< " " << c
->cid
<< " to " << d
->cid
<< " "
11242 << " bits " << bits
<< " = " << r
<< dendl
;
11246 // DB key value Histogram
11247 #define KEY_SLAB 32
11248 #define VALUE_SLAB 64
11250 const string prefix_onode
= "o";
11251 const string prefix_onode_shard
= "x";
11252 const string prefix_other
= "Z";
11254 int BlueStore::DBHistogram::get_key_slab(size_t sz
)
11256 return (sz
/KEY_SLAB
);
11259 string
BlueStore::DBHistogram::get_key_slab_to_range(int slab
)
11261 int lower_bound
= slab
* KEY_SLAB
;
11262 int upper_bound
= (slab
+ 1) * KEY_SLAB
;
11263 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11267 int BlueStore::DBHistogram::get_value_slab(size_t sz
)
11269 return (sz
/VALUE_SLAB
);
11272 string
BlueStore::DBHistogram::get_value_slab_to_range(int slab
)
11274 int lower_bound
= slab
* VALUE_SLAB
;
11275 int upper_bound
= (slab
+ 1) * VALUE_SLAB
;
11276 string ret
= "[" + stringify(lower_bound
) + "," + stringify(upper_bound
) + ")";
11280 void BlueStore::DBHistogram::update_hist_entry(map
<string
, map
<int, struct key_dist
> > &key_hist
,
11281 const string
&prefix
, size_t key_size
, size_t value_size
)
11283 uint32_t key_slab
= get_key_slab(key_size
);
11284 uint32_t value_slab
= get_value_slab(value_size
);
11285 key_hist
[prefix
][key_slab
].count
++;
11286 key_hist
[prefix
][key_slab
].max_len
= MAX(key_size
, key_hist
[prefix
][key_slab
].max_len
);
11287 key_hist
[prefix
][key_slab
].val_map
[value_slab
].count
++;
11288 key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
=
11289 MAX(value_size
, key_hist
[prefix
][key_slab
].val_map
[value_slab
].max_len
);
11292 void BlueStore::DBHistogram::dump(Formatter
*f
)
11294 f
->open_object_section("rocksdb_value_distribution");
11295 for (auto i
: value_hist
) {
11296 f
->dump_unsigned(get_value_slab_to_range(i
.first
).data(), i
.second
);
11298 f
->close_section();
11300 f
->open_object_section("rocksdb_key_value_histogram");
11301 for (auto i
: key_hist
) {
11302 f
->dump_string("prefix", i
.first
);
11303 f
->open_object_section("key_hist");
11304 for ( auto k
: i
.second
) {
11305 f
->dump_unsigned(get_key_slab_to_range(k
.first
).data(), k
.second
.count
);
11306 f
->dump_unsigned("max_len", k
.second
.max_len
);
11307 f
->open_object_section("value_hist");
11308 for ( auto j
: k
.second
.val_map
) {
11309 f
->dump_unsigned(get_value_slab_to_range(j
.first
).data(), j
.second
.count
);
11310 f
->dump_unsigned("max_len", j
.second
.max_len
);
11312 f
->close_section();
11314 f
->close_section();
11316 f
->close_section();
11319 //Itrerates through the db and collects the stats
11320 void BlueStore::generate_db_histogram(Formatter
*f
)
11323 uint64_t num_onodes
= 0;
11324 uint64_t num_shards
= 0;
11325 uint64_t num_super
= 0;
11326 uint64_t num_coll
= 0;
11327 uint64_t num_omap
= 0;
11328 uint64_t num_deferred
= 0;
11329 uint64_t num_alloc
= 0;
11330 uint64_t num_stat
= 0;
11331 uint64_t num_others
= 0;
11332 uint64_t num_shared_shards
= 0;
11333 size_t max_key_size
=0, max_value_size
= 0;
11334 uint64_t total_key_size
= 0, total_value_size
= 0;
11335 size_t key_size
= 0, value_size
= 0;
11338 utime_t start
= ceph_clock_now();
11340 KeyValueDB::WholeSpaceIterator iter
= db
->get_iterator();
11341 iter
->seek_to_first();
11342 while (iter
->valid()) {
11343 dout(30) << __func__
<< " Key: " << iter
->key() << dendl
;
11344 key_size
= iter
->key_size();
11345 value_size
= iter
->value_size();
11346 hist
.value_hist
[hist
.get_value_slab(value_size
)]++;
11347 max_key_size
= MAX(max_key_size
, key_size
);
11348 max_value_size
= MAX(max_value_size
, value_size
);
11349 total_key_size
+= key_size
;
11350 total_value_size
+= value_size
;
11352 pair
<string
,string
> key(iter
->raw_key());
11354 if (key
.first
== PREFIX_SUPER
) {
11355 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SUPER
, key_size
, value_size
);
11357 } else if (key
.first
== PREFIX_STAT
) {
11358 hist
.update_hist_entry(hist
.key_hist
, PREFIX_STAT
, key_size
, value_size
);
11360 } else if (key
.first
== PREFIX_COLL
) {
11361 hist
.update_hist_entry(hist
.key_hist
, PREFIX_COLL
, key_size
, value_size
);
11363 } else if (key
.first
== PREFIX_OBJ
) {
11364 if (key
.second
.back() == ONODE_KEY_SUFFIX
) {
11365 hist
.update_hist_entry(hist
.key_hist
, prefix_onode
, key_size
, value_size
);
11368 hist
.update_hist_entry(hist
.key_hist
, prefix_onode_shard
, key_size
, value_size
);
11371 } else if (key
.first
== PREFIX_OMAP
) {
11372 hist
.update_hist_entry(hist
.key_hist
, PREFIX_OMAP
, key_size
, value_size
);
11374 } else if (key
.first
== PREFIX_DEFERRED
) {
11375 hist
.update_hist_entry(hist
.key_hist
, PREFIX_DEFERRED
, key_size
, value_size
);
11377 } else if (key
.first
== PREFIX_ALLOC
|| key
.first
== "b" ) {
11378 hist
.update_hist_entry(hist
.key_hist
, PREFIX_ALLOC
, key_size
, value_size
);
11380 } else if (key
.first
== PREFIX_SHARED_BLOB
) {
11381 hist
.update_hist_entry(hist
.key_hist
, PREFIX_SHARED_BLOB
, key_size
, value_size
);
11382 num_shared_shards
++;
11384 hist
.update_hist_entry(hist
.key_hist
, prefix_other
, key_size
, value_size
);
11390 utime_t duration
= ceph_clock_now() - start
;
11391 f
->open_object_section("rocksdb_key_value_stats");
11392 f
->dump_unsigned("num_onodes", num_onodes
);
11393 f
->dump_unsigned("num_shards", num_shards
);
11394 f
->dump_unsigned("num_super", num_super
);
11395 f
->dump_unsigned("num_coll", num_coll
);
11396 f
->dump_unsigned("num_omap", num_omap
);
11397 f
->dump_unsigned("num_deferred", num_deferred
);
11398 f
->dump_unsigned("num_alloc", num_alloc
);
11399 f
->dump_unsigned("num_stat", num_stat
);
11400 f
->dump_unsigned("num_shared_shards", num_shared_shards
);
11401 f
->dump_unsigned("num_others", num_others
);
11402 f
->dump_unsigned("max_key_size", max_key_size
);
11403 f
->dump_unsigned("max_value_size", max_value_size
);
11404 f
->dump_unsigned("total_key_size", total_key_size
);
11405 f
->dump_unsigned("total_value_size", total_value_size
);
11406 f
->close_section();
11410 dout(20) << __func__
<< " finished in " << duration
<< " seconds" << dendl
;
11414 void BlueStore::_flush_cache()
11416 dout(10) << __func__
<< dendl
;
11417 for (auto i
: cache_shards
) {
11419 assert(i
->empty());
11421 for (auto& p
: coll_map
) {
11422 assert(p
.second
->onode_map
.empty());
11423 assert(p
.second
->shared_blob_set
.empty());
11428 // For external caller.
11429 // We use a best-effort policy instead, e.g.,
11430 // we don't care if there are still some pinned onodes/data in the cache
11431 // after this command is completed.
11432 void BlueStore::flush_cache()
11434 dout(10) << __func__
<< dendl
;
11435 for (auto i
: cache_shards
) {
11440 void BlueStore::_apply_padding(uint64_t head_pad
,
11442 bufferlist
& padded
)
11445 padded
.prepend_zero(head_pad
);
11448 padded
.append_zero(tail_pad
);
11450 if (head_pad
|| tail_pad
) {
11451 dout(20) << __func__
<< " can pad head 0x" << std::hex
<< head_pad
11452 << " tail 0x" << tail_pad
<< std::dec
<< dendl
;
11453 logger
->inc(l_bluestore_write_pad_bytes
, head_pad
+ tail_pad
);
11457 // ===========================================