Alex Ainscow [Wed, 23 Apr 2025 14:41:11 +0000 (15:41 +0100)]
osd: Fix parity updates in truncates.
Previously in optimised EC, when truncating to a partial
stripe, the parity was not being updated. This fix reads
the non-truncated data from the final stripe and calculates
parity updates, which are written to the parity shards.
Alex Ainscow [Tue, 22 Apr 2025 12:41:19 +0000 (13:41 +0100)]
osd: Fix EC cache invalidation bug
With optimised EC, there were two bugs with cache invalidation:
1. If two invalidates were in the queue, its possible the second
invalidate might be cleared by the first.
2. Reads were being requested if size was being reduced.
Also, added a few debug improvements and some new asserts.
Alex Ainscow [Wed, 9 Apr 2025 12:49:49 +0000 (13:49 +0100)]
osd: Make EC alignment independent of page size.
Code which manipulates full pages is often faster. To exploit this
optimised EC was written to deal with 4k alignment wherever possible.
When inputs are not aligned, they are quickly aligned to 4k.
Not all architectures use 4k page sizes. Some power architectures for
example have a 64k page size. In such situations, it is unlikely that
using 64k page alignment will provide any performance boost, indeed it
is likely to hurt performance significantly. As such, EC has been
moved to maintain its internal alignment (4k), whcih can be configured.
This has the added advantage, that we can can potentially tweak this
value in the future.
Alex Ainscow [Wed, 16 Apr 2025 09:41:48 +0000 (10:41 +0100)]
osd: Fix Truncates in Optimised EC
The previous truncate code attempted to perform a non-aligned truncate by
creating a zero buffer at the end of the object, which was written.
The new code initially truncates to the exact size of the user object before
growing the object to the required 4k alignment. This simpler arrangement
also simplifies the rollback.
Bill Scales [Fri, 6 Jun 2025 12:28:14 +0000 (13:28 +0100)]
osd: EC optimizations fix bug when recovering only partial write objects
PGLog::reset_complete_to is not handling the scenario where all the
missing objects have a partial write that excludes updating the shard being
recovered as their most recent update. In this scenario the oldest need
is newer than newest log entry. Setting last_compelte to the head of the
log confuses code and makes it think that recovery has completed.
The fix is to hold last_complete one entry behind the head of the log
until all missing objects have been recovered.
PGLog::recover_got already does this when an object is recovered and the
remaining objects to recover match this scenario, so this fix just makes
reset_complete_to behave the same way as recover_got.
Bill Scales [Thu, 5 Jun 2025 10:17:06 +0000 (11:17 +0100)]
osd: EC optimizations correct pwlc after PG split
When a PG splits the log entries are divided between the two PGs,
this can result in PWLC refering to log entries in the other PG.
Rollback PWLC after the split so it is not further advanced that
the most recently completed log entry.
Non-primary shards can be missing log entries and may rollback
PWLC too far because of this, however this does not matter
because a split occurs at the start of a peering cycle and these
shards will be updated with the correct PWLC from the primary
shard later in the peering cycle when they are activated.
Bill Scales [Mon, 26 May 2025 13:33:12 +0000 (14:33 +0100)]
osd: EC optimizations overaggresive check for missing objects
Relax an assert in read_log_and_missing for optimized EC
pools. Because the log may not have entries for partial
writes but the missing list is calculated from the full
log the need version for a missing item may be newer than
the lastest log entry for that object.
ceph_objectstore_tool needs care because we don't want to add
extra dependencies. To minimise the dependencies, we always
relax the asserts when using this tool.
Signed-off-by: Bill Scales <bill_scales@uk.ibm.com> Signed-off-by: Alex Ainscow <aainscow@uk.ibm.com>
(cherry picked from commit 20e883fedaf19293e939c4cac44de196bd6c9c19)
Bill Scales [Wed, 21 May 2025 17:16:50 +0000 (18:16 +0100)]
osd: EC Optimizations fix proc_master_log handling of splits
For optimized EC pools proc_master_log needs to deal with
the other log being merged being behind the local log because
it is missing partial writes. This is done by finding the
point where the logs diverge and then checking whether local
log entries have been committed on all the shards.
A bug in this code meant that after a PG split (where there
may be gaps in the log due to entries moving to the other PG)
that the divergence point was not found and committed
partial writes ended up being discarded which creates
unfound objects.
Bill Scales [Tue, 20 May 2025 10:37:05 +0000 (11:37 +0100)]
osd: EC Optimizations fix missing call to partial_write
When a shard is backfilling and it receives a log entry where the
transaction is not applied it can skip the roll forward by
immediately advancing crt. However it is still necessary to
call partial_write in this scenario to keep the pwlc information
up to date.
Bill Scales [Wed, 14 May 2025 07:39:40 +0000 (08:39 +0100)]
osd: EC Optimizations bug fix for flip/flop acting set
EC optimizations pools have a set of non-primary shards which
cannot become the primary because they do not have all the
metadata updates. If one of these shards is chosen as the
primary it will set the acting set to force another shard to
be chosen.
It is important that the selected acting set is the same
acting set that will be chosen by the next primary (assuming
nothing else changes) otherwise a PG can get into a state where
the acting set flip/flops between two different states causing
the PG to get stuck in peering and hanging I/O.
A bug in update_peer_info meant that non-primary shards did not
present the same info to choose_acting_set as primary shards
because they were not updating their pg_info_t based on pwlc
information from other shards.
Alex Ainscow [Tue, 13 May 2025 11:55:14 +0000 (12:55 +0100)]
osd: Refuse to commit/rollforward beyond end of log.
In optimised EC, if transaction is applied to all shards, followed by a
partial transaction AND these two transactions overlap, then it is
possible for the non-primary shards to commit a version which is after
then end of the log.
This commit changes the apply_log such that the commit version will be
changed to the head of the log in such situations.
Alex Ainscow [Tue, 29 Apr 2025 11:02:07 +0000 (12:02 +0100)]
osd: Refactor partial_write to address multiple issues.
We fix a number of issues with partial_write here.
Fix an issue where it is unclear whether the empty PWLC state is
newer or older than a populated PWLC on another shard by always
updating the pwlc with an empty range, rather than blank.
This is an unfortunate small increase in metadata, so we should
come back to this in a later commit (or possibly later PR).
Normally a PG log consists of a set of log entries with each
log entry have a version number one greater than the previous
entry. When a PG splits the PG log is split so that each of the
new PGs only has log entries for objects in that PG, which
means there can be gaps between version numbers.
PGBackend::partial_write is trying to keep track of adjacent
log updates than do not update a particular shard storing
these as a range in partial_writes_last_complete. To do this
it must compare with the version number of the previous log
entry rather than testing for a version number increment of one.
Also simplify partial_writes to make it more readable.
Signed-off-by: Bill Scales <bill_scales@uk.ibm.com> Signed-off-by: Alex Ainscow <aainscow@uk.ibm.com>
(cherry picked from commit 2467406f3e22c0746ce20cd04b838dccedadf055)
Alex Ainscow [Tue, 29 Apr 2025 10:59:24 +0000 (11:59 +0100)]
osd: nonprimary shards are permitted to have a crt newer than head
Non-primary shards do not get updates for some transactions. It is possible
however for other transactions to increase the can_rollback_to to a later
version. This causes an assert for some operations.
Bill Scales [Fri, 25 Apr 2025 14:03:02 +0000 (15:03 +0100)]
osd: overaggressive assert in read_log_and_missing with optimized EC pool
read_log_and_missing is called during OSD initializaiton to sanity check
the PG log. One of its checks is too agressive for an optimized EC pool
where because of a partial write there can be a log entry but no update
to the object on this shard (other shards will have been updated). The
fix is to skip the checks when the log entry indicates this shard was
not updated.
Only affects pool with allow_ec_optimizations flag on.
Bill Scales [Thu, 29 May 2025 11:53:27 +0000 (12:53 +0100)]
osd: EC optimizations rework for pg_temp
Bug fixes for how pg_temp is used with optimized EC pools. For these
pools pg_temp is re-ordered with non-primary shards last. The acting
set was undoing this re-ordering in PeeringState, but this is too
late and results code getting the shard id wrong. One consequence
iof this was an OSD refusing to create a PG because of an incorrect
shard id.
This commit moves the re-ordering earlier into OSDMap::_get_temp_osds,
some changes are then required to OSDMap::clean_temps.
Bill Scales [Fri, 23 May 2025 09:45:46 +0000 (10:45 +0100)]
osd: EC Optimizations OSDMap::clean_temps preventing change of primary
clean_temps is clearing pg_temp if the acting set will be the same
as the up set. For optimized EC pools this is overaggressive because
there are scenarios where it is setting acting set to be the same as
up set to force an alternative shard to be chosen as primary - this
happens because the acting set is transformed to place non-primary
shards at the end of the pg_temp vector.
Detect this scenario and stop clean_temps from undoing the acting
set which is being set by PeeringState::choose_acting.
Bill Scales [Thu, 22 May 2025 12:12:57 +0000 (13:12 +0100)]
osd: EC optimizations bug in OSDMap::clean_temps
OSDMap clean_temps clears pg_temp for a PG when the up set
matches the acting_set. For optimized EC pools the pg_temp
is reordered to place primary shards first, this function
was not calling pgtemp_undo_primaryfirst to revert the
reordering.
This meant that a 2+1 EC PG with up set [1,2,3] and
a desired acting set [1,3,2] re-ordered the acting
set to produce pg_temp as [1,2,3] and then deleted this
because it equals the up set.
Calling pgtemp_undo_primaryfirst makes this code work
as intended.
Bill Scales [Mon, 19 May 2025 16:01:39 +0000 (17:01 +0100)]
osd: EC Optimizations fix routing of requests to non-zero shard id
Pools with EC optimizations can use pg_temp to modify the selection
of the primary. When this happens the clients route request to the
correct OSD, but wrong shard which causes hung I/Os or misdirected
I/Os.
Fix Objecter to select the correct shard when sending requests to
EC optimized pools. Fix OSD to modify the shard when receving
requests from legacy clients.
Add new unittests to test new functions for remapping the shard id.
Alex Ainscow [Thu, 24 Apr 2025 14:13:31 +0000 (15:13 +0100)]
osd: Cosmetic code cleanup and improve debug
All these changes are one of:
* Whitespace changes
* Addition/removal of debug
* Typos
* Make CPU-intensive debug statements deb ug level 30
* Remove unnecessary counter which did not really help with debug
* Extra comments
This should fix the chmod 777 /var/log/ceph failures.
We were missing the install task which resulted in no /var/log/ceph:
```
2025-07-07T08:55:44.586 INFO:teuthology.run_tasks:Running task ceph...
2025-07-07T08:55:44.679 INFO:tasks.ceph:Making ceph log dir writeable by
non-root...
2025-07-07T08:55:44.679 DEBUG:teuthology.orchestra.run.smithi144:> sudo
chmod 777 /var/log/ceph
2025-07-07T08:55:44.711
INFO:teuthology.orchestra.run.smithi144.stderr:chmod: cannot access
'/var/log/ceph': No such file or directory
```
Add comprehensive documentation for defining configuration options in
ceph-mgr modules, including all supported properties and their usage.
Previously, the documentation did not explain how to define ceph-mgr
module configuration options, despite subtle differences from other Ceph
components. This change documents all supported Option properties, their
types, and provides clear examples to help module developers properly
configure their options.
projects / ceph-ci.git / log