Sage Weil [Tue, 6 Apr 2021 14:41:09 +0000 (09:41 -0500)]
qa/suites/rados/objectstore: separate store_test tests
This takes 5 hours currently.
- Separate out filestore and memstore into separate task (~1 hr)
- Split bluestore into -a and -b (a tests exclude SynethicMatrixC,
b tests include it)
Sage Weil [Mon, 12 Apr 2021 15:45:50 +0000 (11:45 -0400)]
Merge PR #40736 into master
* refs/pull/40736/head:
mgr/cephadm: rewrite/simplify describe_service
mgr/orchestrator: report osds as osd.unmanaged as appropriate
mgr/orchestrator: remove IMAGE ID from 'orch ls'
all the scripts except for test_cls_cas.sh under qa/workunits/cls
are executable. to be more consistent, add the executable bit to
test_cls_cas.sh as well.
also, these scripts are launched by src/script/gen-corpus.sh directly,
so it's convenient just call them.
if we happen to run this script on a host where /etc/ceph/ceph.conf is
available, ceph CLI would use it instead. so, point it to $PWD/ceph.conf
instead.
Changes to the socket code now result in returning EINVAL
In the past ENOENT was returned which is the FreeBSD error code
if DNS lookup does not work.
And that change is probably because somewhere in the code that
errorcode is not passed verbatim from the systemcall, but is
rewritten in extra evaluation.
Signed-off-by: Willem Jan Withagen <wjw@digiware.nl>
FreeBSD ceph-dencoder crashes in the exit() calls, due to
invalid pointer references during the release process of
the loaded libraries.
Often this is signaled by libc reporting:
__cxa_thread_call_dtors: dtr 0x47efc0 from unloaded dso, skipping
The cause for this is different behaviour between FreeBSD and Linux:
https://groups.google.com/g/bsdmailinglist/c/22ncTZAbDp4/m/Dii_pII5AwAJ
_The FreeBSD implementation here looks racy. If one thread dlcloses an
object while another thread is exiting, we can end up calling a
function at an invalid memory address. It also looks as if it may
be possible to unload one library, load another at the same address,
and end up executing entirely the wrong code, which would have some
serious security implications.
The GNU/Linux equivalent of this function locks the DSO in memory
until all references to it have gone away. A call to dlclose() on
GNU/Linux will not actually unload the library until all threads
with destructors in that library have been unloaded. I believe
that this reuses the same reference counting mechanism that
allows the same library to be dlopened and dlclosed multiple times.
Signed-off-by: Willem Jan Withagen <wjw@digiware.nl>
Sage Weil [Fri, 9 Apr 2021 20:26:00 +0000 (16:26 -0400)]
mgr/cephadm: rewrite/simplify describe_service
The prior implementation first tried to fabricate services based on the
running daemons, and then filled in defined services on top. This led
to duplication and a range of small errors.
Instead, flip this around: start with the services that are defined,
and only fill in 'unmanaged' services where we need to.
Drop the osd kludges and instead rely on DaemonDescription.service_id to
return the right thing.
rgw: test `radosgw-admin radoslist` and incomplete multiparts better
Make sure there are more than 1000 incomplete multiparts and also make
sure one of the incomplete multiparts has at least 1000 parts. This
test is done indirectly through rgw-orphan-list, which invokes
`radosgw-admin radoslist`.
Also, clean up shell flags, so script output is less verbose.
Signed-off-by: J. Eric Ivancich <ivancich@redhat.com>
Sage Weil [Sat, 3 Apr 2021 13:14:00 +0000 (09:14 -0400)]
cephadm: normalize unqualified repo digests to docker.io
A RepoDigests returned by docker|podman image inspect can either include
the docker.io/ prefix or not. For reasons that aren't entirely clear,
this may vary between hosts in a cluster. However, ceph/ceph@sha256:abc...
is the same thing as docker.io/ceph/ceph@sha256:abc..., and should be
treated as such. Otherwise, upgrade can get into a loop where it pulls
the image on a new host, finds the other variant of the repodigests,
sees no overlap, updates target_digests, and restarts. (It will then
find the first variant again on the first host and loop.)
Avoid this by normalizing any docker.io digests by always including the
docker.io/ prefix.
Note that it is technically possible that this assumption is wrong: it
may be that the image that already exists on the local host is from a
different registry in registries.conf's unqualified-search-registries.
However, we don't know which, since this is a search list. In practice,
it should be exceeding rare that an image that *we* are installing using
a fully-qualified image name will end up having an unqualified repodigest
in the local registry. Hopefully!
Fixes: https://tracker.ceph.com/issues/50114 Signed-off-by: Sage Weil <sage@newdream.net>
If we get an unqualified target image, assume it's docker.io. This
ensures that we're passing a fully-qualified target to docker|podman on
the various hosts and don't end up with something different based on the
per-host search path for unqualified image names.
This is my second attempt to rewrite the
second half of the mclock docs. The first attempt
is enshrined in https://github.com/ceph/ceph/pull/40571,
in which I got cute with git and got burned.
Sage Weil [Fri, 9 Apr 2021 19:35:17 +0000 (15:35 -0400)]
mgr/orchestrator: remove IMAGE ID from 'orch ls'
This is not very useful at this level:
- we see it from 'orch ps'
- it can be a mix of ids during upgrade
- some services may have multiple images at steady state (e.g., ingress)
When an incomplete multipart upload has in excess of 1000 parts,
looping over those parts was not handled property causing an infinite
loop. The paging/marker is now handled correctly.
Signed-off-by: J. Eric Ivancich <ivancich@redhat.com>
as FreeBSD also has msghdr but it does not have systemd, or
flags like MFD_ALLOW_SEALING, O_TMPFILE or F_SEAL_GROW. so
use WITH_SYSTEMD for enabling journald backend of logging system.
also move the option of "WITH_SYSTEMD" up so that the src/CMakeLists.txt
is able to see the variable of WITH_SYSTEMD defined by it.
myoungwon oh [Wed, 31 Mar 2021 06:22:01 +0000 (15:22 +0900)]
osd: remove unnecessary ref handling in _delete_oid
Let's consider the following case when handling a delete op.
1. Delete --> whiteouted
2. Make clone
In this case, current code clears chunk_map and calls dec_all_manifest_refcount()
in _delete_oid() even if the clone still has the references.
To fix this, This commit remove unnecessary ref handling in _delete_oid, and
makes finish_ctx() to handle ref handling, which can aware of whether the
clone is created or not.
Also, remove oi.size == 0 condition in finish_ctx() to handle ref. counting
upon a delete op with whitedouted clone.
YuanXin [Sun, 21 Feb 2021 07:08:07 +0000 (15:08 +0800)]
osd: avoid two copy with same src cancel each other
For cache tier, if some head object has two snaps, the two snaps share the same clone object,
and the clone object was flush/evicted from cache pool, when a rollback requests and a read
snap request to these two snaps at the same time will generate two promote requests to the
same clone object, these two promote requests will generate two copy ops with same src, than
the second copy op will cancel the first copy op by calling cancel_copy and kick_object_context_blocked,
but after calling kick_object_context_blocked, a new promote request corresponding to first
copy op will be restarted and generate a new copy op, the new copy op will cancel the second
copy op again, so two promote requests will cancel their copy op each other and run into dead
loop.