Details:
- New store dump/write-meta didn't use actual metadata parameters from the header
- New store write-meta calculated small entry sizes incorrectly
- Old store write-journal imported entries with checksums incorrectly
- Old store write-meta didn't import block_csums at all (it was using a wrong json key)
- New store fixes:
- Fix repeated rollback logic
- Fix crash on rolled back object compaction
- Fix postpone_load possibly merging different object chains
- Fix block_csums import in vitastor-disk write-meta
- Fix header checksum after vitastor-disk write-meta
- Old store fixes:
- Fix batched fsync possibly skipped by some flush coroutines
- Improve ENOSPC test, fix possible crash on ENOSPC
- Add fsyncs to vitastor-disk prepare
- Fix possible crash on pg_lock check failure in sec_read_bmp
- Fix VitastorFS initialization when local_reads are enabled
Important fixes (except the new store):
- Fixed a possible use-after-free in the OSD during error handling of initial
commit/rollback of objects in EC pools.
- Fixed a possible free of an invalid pointer in the OSD during read errors
from snapshot/clone chains in EC pools.
- Fixed possibly incorrect handling of commit/rollback operations in EC pools
during pool PG count changes.
- Fixed the inverted fsync enable parameter in the ublk driver (fsync was not
enabled on pools without immediate_commit).
- Added the raw-ls command for debugging purposes to find object versions in
the cluster using listing operations.
New store fixes:
- Improved startup speed by using LSN-based sorting only for objects with a
large number of intermediate versions.
- Added skip_double_claim option as a temporary workaround to fix the rare OSD
startup error with the "double claimed block" message, observed by several
users. This option does not affect data integrity.
- Fixed incorrect rechecking of small writes during startup, which in theory
could lead to duplicate small write object entries on the OSD.
- Fixed fsync operation for disks with a writeback cache (without capacitors):
- Fixed incorrect semantics of consecutive fsyncs (next fsync was not blocked
by the previous one).
- Added fsync when copying small writes from the buffer to the data device
(somehow forgotten during initial development).
- Added fsync after the initial garbage collection during OSD startup.
- Fixed incorrect cast of LSN from uint64 to uint32, breaking fsync when
reaching LSN 2^32.
- Added missing verification of the metadata header checksum during startup.
- Fixed incorrect updating of object checksums in perfect_csum_update=true mode.
- Fixed a possible OSD crash with "assertion failed" when processing a malformed
EC STABILIZE operation.
- Fixed the accounting of active compactor coroutines.
- Removed broken and untested new->old store conversion support.
Minor issues fixed:
- Incorrect accounting of OSD local operation statistics in replicated pools.
- Missing non-zero exitcodes on vitastor-disk resize command errors.
- Missing reset of the list of inconsistent objects during PG restarts.
- Theoretically possible hangs of various OSD operations when working with
completely corrupted objects (without a single available copy), and possibly
in some other very rare situations.
- Incorrect fsyncs when deleting objects from pools without immediate_commit
(on disks with a writeback cache), which previously could leave garbage when
deleting misplaced objects.
- Possible crash/memory corruption of the NFS server during a targeted attack
on NFS-RDMA.
- Possibly incorrect handling of ENOSPC/EIO write errors in replicated pools,
leading to inability to retry the write later.
- Possible crash instead of a clean error exit when starting an OSD with the
old storage engine on a disk with corrupted journal data.
- Shallow copying of PG configuration in the monitor, however, not related to
actual bugs.
- Incorrect checking of allocated blocks in the QEMU driver in an unused code
branch (without the BDRV_WANT_ZERO flag).
- Possible memory leak on read errors of corrupted objects.
- Possible incorrect PG states when corrupted objects are detected.
- Possible failure to mark all "bad" copies of an object during scrubs without
checksums and with a large number of replicas (> 4).
- Incorrect checksum calculation in the old storage engine when
bitmap_granularity < 4096 (a practically unused configuration).
- Theoretically possible OSD crash in rare cases during a scrub and simultaneous
object recovery.
- Theoretically possible OSD crash when handling PING operation errors.
- Slightly suboptimal logic for reusing the RDMA send buffer.
- Possible memory leak when canceling an already running scrub via no_scrub.
- Possible memory corruption when a client (e.g., QEMU code) passes invalid
buffers and the writeback cache is enabled.
- Potentially incorrect search for corrupted parts of EC objects (inability
to find a "good" combination) during a scrub with checksums disabled.
- Possible additional memory usage on the OSD side when handling failed reads
from snapshots (not a leak however - the memory was freed upon client
disconnection).
- Potential sudden write slowdown at certain pg epoch values due to incorrect
epoch update logic in etcd.
Vitaliy Filippov