Commit Graph

680 Commits (519f0810068f776614e83cbbbd332a30f4a60056)

Author SHA1 Message Date
Vitaliy Filippov 18a5fafa2a Fix rollback 2021-03-25 02:41:58 +03:00
Vitaliy Filippov 06f4978085 Fix fsync check in blockstore_flush (data fsyncs were disabled instead of journal fsyncs) 2021-03-25 02:41:58 +03:00
Vitaliy Filippov 7ebf1588c5 Check for immediate_commit==small in the OSD code 2021-03-25 02:41:58 +03:00
Vitaliy Filippov b0ad1e1e6d Remember writes as "unsynced" only after completing them
Previously BS_OP_SYNC could take unfinished writes and add them into the journal before
they were actually completed. This was leading to crashes with the message
"BUG: Unexpected dirty_entry 2000000000001:9f2a0000 v3 unstable state during flush: 338"
2021-03-25 02:41:58 +03:00
Vitaliy Filippov 0949f08407 Extract osd_primary write and sync code into separate files 2021-03-24 14:20:56 +03:00
Vitaliy Filippov 04a1f18fa5 Assign .req as a whole to always zero out the remaining part
Also clear .reply before processing the operation
2021-03-24 14:20:56 +03:00
Vitaliy Filippov cf9a641d66 Skip disconnected OSDs during sync 2021-03-24 14:20:56 +03:00
Vitaliy Filippov 05db1308aa Fix two potential read/write ordering problems (even though not yet seen in tests)
- Write operations could be 'stabilized' and previous versions could be
  purged from OSDs before the removal of version_override and following
  reads could potentially hit different version in EC pools
- Object was marked clean after completing the delete during recovery, so
  reads could in theory hit a deleted version and return nothing
2021-03-24 14:20:56 +03:00
Vitaliy Filippov 98b54ca948 Don't try to "recover" misplaced objects if it would make them degraded 2021-03-21 01:37:23 +03:00
Vitaliy Filippov 23225c5e62 Do not run ping on clients that are not yet connected 2021-03-21 01:37:23 +03:00
Vitaliy Filippov 7e6e1a5a82 Release 0.5.10
The version seems to be stable after this bunch of fixes :)

- Fix delete & write operation ordering during rebalance to not lose objects in the immediate_commit=off mode
- Fix a possible crash caused by very high iodepths
- Re-distribute PG primaries over OSDs that come up after a short downtime
- Allow to specify etcd URLs for OSDs with http://, do not die with a strange error if -etcd option is missing for fio
- Fix a journal flushing deadlock which sometimes occurred in the immediate_commit=off mode
- Fix a bug where OSDs could hang if the data device filled up
- Fix an allocator bug where it was unable to allocate up to last (n%64) data device blocks
- Fix monitor crash that occurred on removal of some etcd keys
- Fix a bug where PGs could remain incomplete due to incorrect PG history with just zeroes in osd_sets
2021-03-16 12:48:26 +03:00
Vitaliy Filippov 435045751d Delete objects only after a SYNC during rebalance in the non-immediate_commit mode
Previously OSDs could commit deletes before writes during recovery or rebalance
in the "lazy fsync" (immediate_commit=off) mode which could result in lost objects
2021-03-16 12:48:26 +03:00
Vitaliy Filippov c5fb1d5987 Do not duplicate blockstore operations when io_uring fills up
This bug was leading to OSDs dying with "Assertion `fulfilled == read_op->len' failed"
when testing fio -rw=randread -numjobs=8 -iodepth=128
2021-03-16 12:48:26 +03:00
Vitaliy Filippov 9f59381bea Re-distribute PG primaries over OSDs that come up after a short downtime 2021-03-16 12:48:26 +03:00
Vitaliy Filippov 9ac7e75178 Allow to specify etcd URLs for OSDs with http://, do not die with a strange error if -etcd option is missing for fio 2021-03-16 12:48:26 +03:00
Vitaliy Filippov 88671cf745 Fix a bug causing all flushers to wait for an fsync without actually trying to do it
This happened because flusher_count became dynamic and fsync_batch() was comparing the number
of flushers currently ready to do an fsync with the maximum number of flushers. Also the number
wasn't rechecked on every loop which was also incorrect.

Now the interrupted_rebalance test passes even without IMMEDIATE_COMMIT=1.
2021-03-13 17:27:29 +03:00
Vitaliy Filippov fe1749c427 Fix the multiple_interrupted_rebalance test 2021-03-13 17:19:45 +03:00
Vitaliy Filippov ceb9c28de7 Set default log_level before passing config to etcd_state_client 2021-03-13 17:19:45 +03:00
Vitaliy Filippov 299d7d7c95 Use common macro for get_sqe 2021-03-13 17:19:45 +03:00
Vitaliy Filippov d1526b415f Correctly resume writes when OSD is full to return an error 2021-03-13 17:19:45 +03:00
Vitaliy Filippov f49fd53d55 Fix a bug where allocator was unable to allocate up to last (n%64) blocks, add tests for it 2021-03-13 02:19:02 +03:00
Vitaliy Filippov dd76eda5e5 Test multiple interrupted rebalancings
Currently only passes with immediate_commit=all configuration
(env variable IMMEDIATE_COMMIT=1 for the bash script)
2021-03-12 12:55:44 +03:00
Vitaliy Filippov 87dbd8fa57 Use empty hash as the default value for some etcd keys in the monitor 2021-03-12 12:40:15 +03:00
Vitaliy Filippov b44f49aab2 Ignore zero OSDs in history osd_sets 2021-03-12 12:40:15 +03:00
Vitaliy Filippov 036555638e Release 0.5.9
- Fix two monitor bugs which led to objects being "logically lost" (physically
  present on some secondary OSDs while primary doesn't know about it) after multiple
  interrupted rebalancings
- Implement "no_recovery" and "no_rebalance" flags
2021-03-11 00:39:10 +03:00
Vitaliy Filippov af5155fcd9 Implement "no_recovery" and "no_rebalance" flags 2021-03-11 00:36:31 +03:00
Vitaliy Filippov 0d2efbecc9 Preserve previous PG history when changing PG distribution
Fixes incorrect PG history in case when a new rebalance is started
before the finish of the previous one which could make primary OSDs unable
to locate some objects on some secondaries.
2021-03-11 00:16:10 +03:00
Vitaliy Filippov e62e8b6bae Use real pg configuration instead of the "last clean" one for generating PG history
Basically fixes the bug introduced in 0.5.7 where an rebalance interrupted
by the monitor could result in forgetting objects moved to the new place
2021-03-10 02:01:44 +03:00
Vitaliy Filippov c4ba24c305 Do not print ping op latency 2021-03-10 02:01:44 +03:00
Vitaliy Filippov 19e47a0279 Release 0.5.8
- Add heartbeats (fixes failover in case of network issues or offline nodes)
- Fix a bug where a PG could incorrectly become listed as 'incomplete' if historical osd_sets
  included a set with the the PG's primary OSD as the only alive one
- Use osd_out_time = 10 minutes by default instead of 30 minutes
- Make monitors stick to a single selected etcd URL on start and not try to select random ones
  on every request - this was leading to etcd interaction errors when some etcds were unavailable
2020-03-09 02:38:17 +03:00
Vitaliy Filippov bd178ac20f Fix history osd_set check - local OSD is always available! 2021-03-09 02:18:18 +03:00
Vitaliy Filippov 7006875a24 Make monitor stick to one etcd until the restart 2021-03-09 02:15:38 +03:00
Vitaliy Filippov ad577c4aac Add PING operation and timeouts to detect OSD failures when a host goes down 2021-03-09 02:15:38 +03:00
Vitaliy Filippov 836635c518 Use osd_out_time = 10 minutes by default 2021-03-09 02:15:38 +03:00
Vitaliy Filippov 88a03f4e98 Release 0.5.7
- Fix multiple bugs leading to OSDs sometimes being unable to correctly activate PGs
  when a lot of PG peering events occurred in a small amount of time
- Fix a bug where OSDs could list incomplete object versions during peering. The bug
  manifested with "local rollback operation failed" messages in OSD logs
- Fix a bug where misplaced chunks for degraded and incomplete objects were not removed
  from extra OSDs during recovery
- Fix incorrect PG history configuration resulting in OSDs being unable to find some
  of the objects after a PG count change
- Simplify block layer write ordering logic
- Avoid extra data move when a lot of OSDs are first stopped for long time and then restarted
- Fix incorrect degraded & misplaced object statistics after a completed rebalance
- Fix incorrect usage of pg_minsize instead of the minimal possible object chunk count in EC pools
2021-03-08 23:37:02 +03:00
Vitaliy Filippov 2a5036669d Fix PG count change procedure
In previous versions PG histories were calculated incorrectly during
PG count change which led to objects being lost on OSDs not in PG's osd set.
2021-03-08 23:15:58 +03:00
Vitaliy Filippov 2e0c853180 Make test_change_pg_count check if any objects are lost during the test 2021-03-08 23:15:07 +03:00
Vitaliy Filippov e91ff2a9ec Only forget offline PGs if their state is not changed during reporting 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 086667f568 Do not check PG state key ownership if it doesn't exist yet
This fixes the bug where OSDs were sometimes trying to report updated PG states
infinitely without luck when PGs transitioned from 'starting' to 'peering' too fast
2021-03-08 17:04:10 +03:00
Vitaliy Filippov 73ce20e246 Add a test for the "reappear after move" case 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 1be94da437 Check & remove extra chunks for degraded / incomplete objects, too 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 80e12358a2 Use pg_data_size instead of pg_minsize for object state calculation 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 36c935ace6 Use std::vector for the blockstore submission queue 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 0d8b5e2ef9 Remove unused enqueue_op_first() 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 98f1e2c277 Rework write/sync ordering
Make syncs wait for all previous writes because it's the only way
to make sure that OSDs do not receive incomplete writes in LIST results
during peering when some writes are still in progress.

Also simplify blockstore submission queue logic.
2021-03-08 17:04:10 +03:00
Vitaliy Filippov 21e7686037 Fix possible "assertion failed: pg.inflight >= 0" error during PG stop 2021-03-08 17:04:10 +03:00
Vitaliy Filippov ab21a1908b Check for the dirty PG flag when trying to continue to stop it after sync 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 30d1ccd43e Fix an infinite loop when discarding list operations during stop_pg() 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 8bdd6d8d78 Reset PG state when stopping them 2021-03-08 17:04:10 +03:00
Vitaliy Filippov 09b3e4e789 Fix OSDs being unable to stop PGs that are 'peering', not 'active'
This was sometimes leading to incorrect misplaced and degraded object count statistics
2021-03-08 17:04:10 +03:00