Let PGShardManager::invoke_on_each_shard_seq pass the local shard_services
instance instead of using an additional helper.
The downside of dropping the generic sharded_map_seq helper is that it is
able to support *any* (seastar::)sharded object. However, as shard_services
is the only user of it - directly using the local instance without the
helper seems easier to read.
Matan Breizman [Sun, 22 Jun 2025 10:10:10 +0000 (10:10 +0000)]
crimson/common/smp_helpers: fix reactor_map_seq
Copy f into reactor_map_seq which would be kept alive
due to this method being a coroutine. That way, we can ensure
the lambdas passed to each core that are capturing f by
reference would be safe.
Alternatively, we can also copy f by using it's copy ctor and
pass a copy to each shard:
co_await crimson::submit_to(core, F(f))
However, avoiding the copy is possible here due to the sequential
traversal. Note, seastar's invoke_on_all do copy each callback to
every shard and is running the invocation in parallel.
The above would have fixed f's captures to be invalid and result
in a segfaults on diffrent shards.
J. Eric Ivancich [Fri, 20 Jun 2025 20:00:54 +0000 (14:00 -0600)]
Merge pull request #63271 from rafaelweingartner/parameter_to_externalize_secret_key_ttl-upstream-2
rgw: Externalize Keystone secret key cache TTL
Reviewed-by: Matt Benjamin <mbenjamin@redhat.com> Reviewed-by: Adam C. Emerson <aemerson@redhat.com> Reviewed-by: Tobias Urdin <tobias.urdin@binero.com>
Ronen Friedman [Thu, 19 Jun 2025 15:27:38 +0000 (10:27 -0500)]
osd/scrub: clarify that osd_scrub_auto_repair_num_errors counts objects
'osd_scrub_auto_repair_num_errors' limits the number of damaged objects
that we will try to auto-repair during a scrub. Its documentation
referred to "number of errors", which did not fit the implementation.
Fixes: https://tracker.ceph.com/issues/71754 Fixes: Red Hat BZ2316244 Signed-off-by: Ronen Friedman <rfriedma@redhat.com>
Jon [Thu, 19 Jun 2025 14:16:36 +0000 (15:16 +0100)]
test/osd: Make error messages that occur based on configuration problems or user error more readable in ceph_test_rados_io_sequence
Move away from using ceph_abort for error messages that occur from configuration issues and user error to printing to std::error as it is not necissairy to produce a call stack and core dumps in these cases and much better to give an easily readable message to the user.
Signed-off-by: Jon Bailey <jonathan.bailey1@ibm.com>
Jon [Thu, 19 Jun 2025 14:14:04 +0000 (15:14 +0100)]
test/osd: Add erasure code plugins as a dependancy of ceph_test_rados_io_sequence
ceph_test_rados_io_sequence uses the plugins for determining erasure code values for testing, so we want these to be built as a dependancy of the application when built in isolation
Signed-off-by: Jon Bailey <jonathan.bailey1@ibm.com>
Kefu Chai [Thu, 19 Jun 2025 10:04:05 +0000 (18:04 +0800)]
rgw: avoid using std::aligned_storage_t
std::aligned_storage_t was deprecated in C++23, to be prepared for it,
let's use alignas for the same behavior. because the 3 * 8 (with LP64
data model) is not power-of-2, while `alignas()` requires an alignment
of power of 2. so we use `std::bit_ceil()` to calculate the minimum
alignment greater or equal to this number.
Kefu Chai [Thu, 19 Jun 2025 09:50:15 +0000 (17:50 +0800)]
common/static_ptr: avoid using std::aligned_storage_t
std::aligned_storage_t was deprecated in C++23, to be prepared for it,
let's use alignas for the same behavior. because the size of `Base`
class is not always power-of-2, while `alignas()` requires an alignment
of power of 2. so we use `std::bit_ceil()` to calculate the minimum
alignment greater or equal to its size.
Kefu Chai [Thu, 19 Jun 2025 08:52:59 +0000 (16:52 +0800)]
neorados: avoid using std::aligned_storage_t
std::aligned_storage_t was deprecated in C++23, to be prepared for
it, let's use alignas for the same behavior. because we do not always
pass a power-of-2 number to `std::aligned_storage_t`, while `alignas()`
requires an alignment of power of 2. so we use `std::bit_ceil()` to
calculate the minimum alignment greater or equal to the given number.
Venky Shankar [Tue, 20 May 2025 12:19:41 +0000 (12:19 +0000)]
client: do not check file size when inode does not have Fc caps
Since the client is holding Fr caps, the read request can be
directly sent to the OSD. The offset/in->size comparison check
is causing the read request to return with no data since in->size
isn't yet updated when another client does an extending write.
Kefu Chai [Thu, 19 Jun 2025 08:19:04 +0000 (16:19 +0800)]
common/io_exerciser: fix buffer overread in DataGenerator
Fix GCC-15 warning about reading uninitialized memory when copying
random data to fill remaining bytes in generated blocks.
The issue occurred when remainingBytes exceeded the 8-byte size of
the uint64_t rand1 variable, causing memcpy to read beyond the
variable's boundary. While this didn't cause crashes (reading from
stack) and the buffer was still properly filled with rand2, it
violated memory safety and generated compiler warnings.
Fixed by limiting the copy size to the actual size of the source
variable (sizeof(rand1)) to ensure we only read initialized memory.
Resolves GCC-15 warnings:
- DataGenerator.cc:76: memcpy reading 9-15 bytes from 8-byte region
- DataGenerator.cc:108: memcpy reading 9-15 bytes from 8-byte region
Currently, we use the "Check ceph config" CI check to remind users about
any configuration changes that were detected in the PR. There's no easy
way for the script to detect if the relevant docuemtations has been
updated for the config change that was detected.
Users might get confused to still see the CI check failing even after
updating the relevant docs. We update the text message to help diffuse
the confusion. If the users will still like to see the CI check go green
- they can comment `/config check ok` and re-run the failed test.
Yuval Lifshitz [Mon, 16 Jun 2025 11:05:25 +0000 (11:05 +0000)]
rgw/logging: make unique part of log file both random and incremental
new format will be: 10 char incremental count (so 32bit uint fit in it).
and 6 char alphanumeric random part.
this should fix possible race conditions in case of multisite
Kefu Chai [Wed, 18 Jun 2025 13:19:21 +0000 (21:19 +0800)]
ceph-object-corpus: update submodule
Update the ceph-object-corpus submodule to pick up the change to
mark cls_rbd_snap as forward incompatible since nautilus. This change
allows us to update ceph-dencoder to allocate fresh instances for each
decode operation instead of reusing existing ones. The uncoming
change in ceph-dencoder will allow us to identify the potential
compatibility break early.
Kefu Chai [Wed, 18 Jun 2025 09:22:36 +0000 (17:22 +0800)]
deb: use variable expansion to support systemd unit dir changes
Ubuntu changed the systemd unit directory location between releases:
- Jammy (22.04): /lib/systemd/system
- Noble (24.04): /usr/lib/systemd/system
To maintain compatibility across both versions, update .install files
to use brace expansion pattern {usr/,}lib/systemd/system/<service>.
This pattern works because dh_install uses bsd_glob() with GLOB_CSH
flags, which expands braces and matches files in both locations
depending on where CMakeLists.txt actually installed them.
Fixes installation issues when building packages on Noble while
maintaining backward compatibility with Jammy builds.
Venky Shankar [Tue, 27 May 2025 07:26:12 +0000 (07:26 +0000)]
qa/cephfs: ignore `OSD_DOWN/osds down` warning
Runs have started failing a lot with the human friendly variant
of the warning. OSD_DOWN is in the ignore list, however, the human
friendly warning (osds down) isn't.
Kefu Chai [Sat, 14 Jun 2025 13:44:05 +0000 (21:44 +0800)]
cls/rbd: use default values for non-decoded fields in test instances
Previously, test instances for cls_rbd_snap used non-default values
for the "parent" field, which is ignored during decoding. The
check-generated.sh test passed because they reused the same instance
for re-encoding, preserving undecoded fields.
An upcoming change will allocate new instances for each encode/decode
verification instead of reusing instances. This will expose
discrepancies between original test instances and re-encoded values
when fields contain non-default values but aren't decoded.
This change sets ignored fields to their default values in test
instances, ensuring consistency between encoding and decoding
operations regardless of the verification approach used.
Since the incompatibility of cls_rbd_snap's on-disk format was
introduced in 32b14ed1, which was introduced Ceph v14, we will
mark this version the first incompatible version in ceph-object-corpus
in the sense that the re-encoded cls_rbd_snap with v8 struct version
is different from the original copy if its parent field is set with
< v8 struct version.
projects / ceph.git / log