- Fix one more bug in old store checksums O_o - checksum verification wasn't
working correctly on reads quickly after small initial offsetted writes
(when a BIG_WRITE was read from the journal).
- Fix clients possibly not destroyed after being switched to RDMA. Not 100%
verified, but in theory it could rarely lead to PGs hanging in peering/stopping
states with RDMA enabled.
- Use vitastor-cli in the Cinder driver for modifications instead of direct etcd
interaction.
- Fix mon_https_ca parsing.
- Fix Docker install.sh, add instructions for Podman.
Checksum verification wasn't working correctly on reads of small initial offsetted writes
from the journal (until the write was flushed to "clean" metadata).
It could in theory lead to hung operations on RDMA client disconnections.
Workflow (probable):
* OSD switches an outbound connection to PEER_RDMA and removes the FD from epoll
* But a receive request in io_uring is still active
* Receive response is never handled, refs remains > 0
* The connection is dropped, but osd_client_t is not destroyed because refs > 0
* Outbound ops are not canceled, suboperations hang, primary operations hang too
A single urgent fix for the new store: metadata was written incorrectly when
filled metadata area exceeded 4 GB due to a simple & stupid 32-bit integer overflow.
This was leading to OSDs corrupting metadata and being unable to start on the next
run.
The bug affected only the new store (meta_format=3) and OSDs with metadata area size
larger than 4 GB. You can check your OSD metadata area size by checking `data_offset`
in `vitastor-disk read-sb /dev/vitastor/osdXX-data` for single-disk OSDs or
`blockdev --getsize64 /dev/vitastor/osdXX-meta` for hybrid OSDs.
If you're affected, it's highly recommended to:
- Install updated packages without restarting OSDs
- Restart OSDs **one by one**
- If a restart of an OSD triggers "has_degraded" rebalance - wait for it to finish
before restarting another OSD
- If an OSD fails to start with "double-claimed" errors in the log - purge and recreate it,
and again wait for "has_degraded" rebalance to finish before proceeding to the next OSD
New store bug fixes:
- Fix very slow OSD startup (hours) possible in unlucky cases
- Fix incorrect garbage entry validation on start in some cases, preventing OSD startup
- Fix small_write_throttling not functioning and either hanging or crashing OSDs
- Fix "different fact_versions returned from subops" error possible with EC in some cases
- Add a recovery option for the new store: skip_corrupted_meta_entries
- Disable intent writes (even 4k) for hybrid OSDs by default for better write buffering
Other changes:
- Fix chmod/chown of a new empty VitastorFS root directory not working
- Fix vitastor-dd treating all < 32k reads as always empty
- Fix vitastor-disk always setting 32k csum_block_size for HDDs when checksums are enabled
- Fix OSD missing PG state updates and PGs hanging in "starting" in rare cases
- Fix a possible OSD crash with "map::at" message in rare cases when stopping a client
- Fix a possible OSD crash on start when RDMA is available but fails to initialize
- Fix removal and overwriting of totaled objects (objects with all corrupted copies)
- Do not allow data_block_size < 8*bitmap_granularity
- Remove RDMA ODP (On-Demand Paging) support
- Update antietcd to 1.2.4
- Update Docker build to Debian Trixie
- Add packages for AlmaLinux 10
- Fix csum_block_size > 0 unusable with atomic writes in the new store
(almost all atomic write requests generated invalid checksums with csum_block_size > 0)
- Fix monitor sometimes randomly failing to optimize PGs with the "problem is
infeasible or unbounded" message due to not waiting to read full stdout of lp_solve
- Remove reshard_abort optimisation to fix chunked resharding introduced in 3.0.2 possibly
corrupting in-memory OSD state when handling multiple rapid PG count change requests
- Fix removed inodes not disappearing from OSD statistics in the new store, leading to
bloating of the statistics with old inodes
- Increase test coverage for the new store and fix several minor bugs:
- Enabling/disabling of used_for_app was recalculating inode space statistics incorrectly
- Fix object metadata validation on OSD startup rejecting some correct sequences of entries,
possibly leading to OSD being unable to start
- Fix PG activation with EC failing in rare cases with EBUSY when requesting to commit an
already committed write
- Fix a theoretically possible metadata writeback issue on ENOSPC during commit
- Fix monitor failing to optimize PGs in presence of a host with name convertible to
js Number (like 04e278988710) :D
- Fix vitastor-cli dd sometimes (rarely) truncating the image when writing to stdout
- Fix a theoretically possible client connection object leak when io_uring is full
- Fix a leak of RDMA-CM connection objects
- Fix crashes with data_block_size < 32KB (useless setup, but anyway) (#113)
- Antietcd is now officially safe to use: the release includes a fixed version of antietcd
which passes Jepsen transaction serializability tests.
- Fix a huge checksum bug in the old store: incorrect checksums for small initial writes.
The bug affected writes of exactly csum_block_size (4k by default) into new (unallocated)
objects and generated invalid checksums in the store for the written 4k block. Moreover,
generation of these invalid checksums was very slow because it was calculating CRC32
for 4 GB of zeroes. If the block wasn't then overwritten as a part of a larger write
request it became unreadable even though the stored data was correct. The bug affected
all versions since 1.0.0, or since 2.3.0 because vitastor-disk didn't allow to enable
checksums prior to 2.3.0 because of another bug. O:-)
- Prevent OSD disconnections due to long blocking of event-loop caused by PG resharding
(moving objects between old and new PGs in memory) when changing pool PG count or
just on restart of an OSD with a large database (for example, with a filled 8 TB SSD).
The issue should now be fixed because OSD now performs resharding in chunks with pauses
between chunks.
- Prevent pools stuck in paused state on an aborted PG count change.
- Fix a possible OSD crash with "division by zero" when trying to handle an operation
before pool PG count is applied to the in-memory store.
Details:
- Write size should be exactly csum_block_size (4k by default)
- Write should be made into a new object (unallocated space)
- In this case, the checksum of the block was calculated as if the block was
padded with extra (2^32 - size) zero bytes due to a simple integer overflow
- As a cherry on the cake, such calculation was 'slightly' slow because it
was processing almost 4 GB of zeroes for a small write
Important fixes:
- Disable RWF_ATOMIC by default because Linux incorrectly requires all atomic
writes to be power-of-2-sized and length-aligned. Details: [use_atomic_flag](https://vitastor.io/en/docs/config/osd.html#use_atomic_flag)
- Fix cross-pool snapshots not working at all - always reading old data after taking the snapshot
- Fix level_placement (broken in 2.2.0)
- Fix CAS write return values in the client library (broken in 2.4.4), also breaking unaligned writes in VitastorFS
- Fix VitastorFS possibly losing some of intersecting parallel unaligned writes
- Prevent possible reads of the old data during unfinished intent writes in the new store
- Tests added for all of above problems to prevent future regressions
Other fixes:
- Allow to specify OSD tags and weights during prepare
- Only clear the first block instead of whole OSD metadata and journal areas during vitastor-disk
prepare - OSD anyway clears them on the first run
- Fix vitastor-cli dd non-seekable detection and error status
- Do not stop OSD on zero-copy tcp short send errors - these aren't a bug, but just caused
by TCP client disconnections
- Fix modify-osd not working after deleting the OSD configuration key from etcd
- Fix regular antietcd websocket disconnections in OSD due to lack of WS_PONG responses to WS_PING
- Reduce the number of allocated RDMA memory regions
- Allow routed RoCE
Vitaliy Filippov