As Greg pointed out, if we get a mangled mdsmap or fsmap, then something
has gone very wrong, and we should avoid doing any activity on the
filesystem.
When this occurs, shut down the mount the same way we would with a
forced umount by calling ceph_umount_begin when decoding fails on either
map. This causes most operations done against the filesystem to return
an error. Any dirty data or caps in the cache will be dropped as well.
The effect is not reversible, so the only remedy is to umount.
[ idryomov: print fsmap decoding error ]
URL: https://tracker.ceph.com/issues/52303
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Acked-by: Greg Farnum <gfarnum@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The "error_string" in the metadata of MClientSession is being
parsed by kclient to validate whether the session is blocklisted.
The "error_string" is for humans and shouldn't be relied on it.
Hence added the flag to MClientsession to indicate the session
is blocklisted.
[ jlayton: minor formatting cleanup ]
URL: https://tracker.ceph.com/issues/47450
Signed-off-by: Kotresh HR <khiremat@redhat.com>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Move remove_capsnaps to caps.c. Move the part of remove_session_caps_cb
under i_ceph_lock into a separate function that lives in caps.c. Have
remove_session_caps_cb call the new helper after taking the lock.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The existing logic relies on ci->i_auth_cap being NULL, but if we end up
removing the auth cap early, then we'll do a lot of useless work and
lock-taking on the remaining caps. Ensure that we only do the auth cap
removal when we're _actually_ removing the auth cap.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
This function does a lot of list-shuffling with cap flushes, all to
avoid possibly freeing a slab allocation under spinlock (which is
totally ok). Simplify the code by just detaching and freeing the cap
flushes in place.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Currently, we check the wb_err too early for directories, before all of
the unsafe child requests have been waited on. In order to fix that we
need to check the mapping->wb_err later nearer to the end of ceph_fsync.
We also have an overly-complex method for tracking errors after
blocklisting. The errors recorded in cleanup_session_requests go to a
completely separate field in the inode, but we end up reporting them the
same way we would for any other error (in fsync).
There's no real benefit to tracking these errors in two different
places, since the only reporting mechanism for them is in fsync, and
we'd need to advance them both every time.
Given that, we can just remove i_meta_err, and convert the places that
used it to instead just use mapping->wb_err instead. That also fixes
the original problem by ensuring that we do a check_and_advance of the
wb_err at the end of the fsync op.
Cc: stable@vger.kernel.org
URL: https://tracker.ceph.com/issues/52864
Reported-by: Patrick Donnelly <pdonnell@redhat.com>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
These are very chatty, racy, and not terribly useful. Just remove them.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
For example in the case of a forced umount, we'll remove all the session
caps even if they are dirty. Move the warning to a wrapper function and
make most of the callers use it. Call the core function when removing
caps due to a forced umount.
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Force umount will try to close the sessions by setting the session
state to _CLOSING. We don't want to WARN in this situation, since it's
expected.
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
capsnaps will take inode references via ihold when queueing to flush.
When force unmounting, the client will just close the sessions and
may never get a flush reply, causing a leak and inode ref leak.
Fix this by removing the capsnaps for an inode when removing the caps.
URL: https://tracker.ceph.com/issues/52295
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
In the case where the export MDS has crashed just after the EImportStart
journal is flushed, a standby MDS takes over for it and when replaying
the EImportStart journal the MDS will wait the client to reconnect. That
may never happen because the client may not have registered or opened
the sessions yet.
When receiving a new map, ensure we reconnect to valid export targets as
well if their sessions don't exist yet.
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The ceph_mds_client and ceph_mds_session structures are kzalloc'ed so
there's no need to explicitly initialize either of their fields to 0.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The first thing metric_delayed_work does is check mdsc->stopping,
and then return immediately if it's set. That's good since we would
have already torn down the metric structures at this point, otherwise,
but there is no locking around mdsc->stopping.
It's possible that the ceph_metric_destroy call could race with the
delayed_work, in which case we could end up with the delayed_work
accessing destroyed percpu variables.
At this point in the mdsc teardown, the "stopping" flag has already been
set, so there's no benefit to flushing the work. Move the work
cancellation in ceph_metric_destroy ahead of the percpu variable
destruction, and eliminate the flush_delayed_work call in
ceph_mdsc_destroy.
Fixes: 18f473b384 ("ceph: periodically send perf metrics to MDSes")
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The ceph_cap_flush structures are usually dynamically allocated, but
the ceph_cap_snap has an embedded one.
When force umounting, the client will try to remove all the session
caps. During this, it will free them, but that should not be done
with the ones embedded in a capsnap.
Fix this by adding a new boolean that indicates that the cap flush is
embedded in a capsnap, and skip freeing it if that's set.
At the same time, switch to using list_del_init() when detaching the
i_list and g_list heads. It's possible for a forced umount to remove
these objects but then handle_cap_flushsnap_ack() races in and does the
list_del_init() again, corrupting memory.
Cc: stable@vger.kernel.org
URL: https://tracker.ceph.com/issues/52283
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Function ceph_check_delayed_caps() is called from the mdsc->delayed_work
workqueue and it can be kept looping for quite some time if caps keep
being added back to the mdsc->cap_delay_list. This may result in the
watchdog tainting the kernel with the softlockup flag.
This patch breaks this loop if the caps have been recently (i.e. during
the loop execution). Any new caps added to the list will be handled in
the next run.
Also, allow schedule_delayed() callers to explicitly set the delay value
instead of defaulting to 5s, so we can ensure that it runs soon
afterward if it looks like there is more work.
Cc: stable@vger.kernel.org
URL: https://tracker.ceph.com/issues/46284
Signed-off-by: Luis Henriques <lhenriques@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
If MDSs aren't available while mounting a filesystem, the session state
will transition from SESSION_OPENING to SESSION_CLOSING. And in that
scenario check_session_state() will be called from delayed_work() and
trigger this WARN.
Avoid this by only WARNing after a session has already been established
(i.e., the s_ttl will be different from 0).
Fixes: 62575e270f ("ceph: check session state after bumping session->s_seq")
Signed-off-by: Luis Henriques <lhenriques@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Currently, we set the r_parent pointer but then don't take a reference
to it until we submit the request. If we end up freeing the req before
that point, then we'll do a iput when we shouldn't.
Instead, take the inode reference in the callers, so that it's always
safe to call ceph_mdsc_put_request on the req, even before submission.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Luis Henriques <lhenriques@suse.de>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Now that we don't need to hold session->s_mutex or the snap_rwsem when
calling ceph_check_caps, we can eliminate ceph_async_iput and just use
normal iput calls.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Turn s_cap_gen field into an atomic_t, and just rely on the fact that we
hold the s_mutex when changing the s_cap_ttl field.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Luis Henriques <lhenriques@suse.de>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The MDS reserves a set of inodes for its own usage, and these should
never be accessible to clients. Add a new helper to vet a proposed
inode number against that range, and complain loudly and refuse to
create or look it up if it's in it.
Also, ensure that the MDS doesn't try to delegate inodes that are in
that range or lower. Print a warning if it does, and don't save the
range in the xarray.
URL: https://tracker.ceph.com/issues/49922
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
We need to use i_size_read(), which properly handles the torn read
case on 32-bit arches.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Add support for grabbing the rsnaps value out of the inode info in
traces, and exposing that via ceph.dir.rsnaps xattr.
Signed-off-by: Yanhu Cao <gmayyyha@gmail.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Since a few years, kernel addresses are no longer included in oops
dumps, at least on x86. All we get is a symbol name with offset and
size.
This is a problem for ceph_connection_operations handlers, especially
con->ops->dispatch(). All three handlers have the same name and there
is little context to disambiguate between e.g. monitor and OSD clients
because almost everything is inlined. gdb sneakily stops at the first
matching symbol, so one has to resort to nm and addr2line.
Some of these are already prefixed with mon_, osd_ or mds_. Let's do
the same for all others.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Acked-by: Jeff Layton <jlayton@kernel.org>
On reconnect, cap and dentry releases are dropped and the fields
that follow must be reencoded into the freed space. Currently these
are timestamp and gid_list, but gid_list isn't reencoded. This
results in
failed to decode message of type 24 v4: End of buffer
errors on the MDS.
While at it, make a change to encode gid_list unconditionally,
without regard to what head/which version was used as a result
of checking whether CEPH_FEATURE_FS_BTIME is supported or not.
URL: https://tracker.ceph.com/issues/48618
Fixes: 4f1ddb1ea8 ("ceph: implement updated ceph_mds_request_head structure")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Implement msgr2.1 wire protocol, available since nautilus 14.2.11
and octopus 15.2.5. msgr2.0 wire protocol is not implemented -- it
has several security, integrity and robustness issues and therefore
considered deprecated.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
In preparation for msgr2, make the cluster send us maps with addrvecs
including both LEGACY and MSGR2 addrs instead of a single LEGACY addr.
This means advertising support for SERVER_NAUTILUS and also some older
features: SERVER_MIMIC, MONENC and MONNAMES.
MONNAMES and MONENC are actually pre-argonaut, we just never updated
ceph_monmap_decode() for them. Decoding is unconditional, see commit
23c625ce30 ("libceph: assume argonaut on the server side").
SERVER_MIMIC doesn't bear any meaning for the kernel client.
Since ceph_decode_entity_addrvec() is guarded by encoding version
checks (and in msgr2 case it is guarded implicitly by the fact that
server is speaking msgr2), we assume MSG_ADDR2 for it.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
- request service tickets together with auth ticket. Currently we get
auth ticket via CEPHX_GET_AUTH_SESSION_KEY op and then request service
tickets via CEPHX_GET_PRINCIPAL_SESSION_KEY op in a separate message.
Since nautilus, desired service tickets are shared togther with auth
ticket in CEPHX_GET_AUTH_SESSION_KEY reply.
- propagate session key and connection secret, if any. In preparation
for msgr2, update handle_reply() and verify_authorizer_reply() auth
ops to propagate session key and connection secret. Since nautilus,
if secure mode is negotiated, connection secret is shared either in
CEPHX_GET_AUTH_SESSION_KEY reply (for mons) or in a final authorizer
reply (for osds and mdses).
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
When we added the btime feature in mainline ceph, we had to extend
struct ceph_mds_request_args so that it could be set. Implement the same
in the kernel client.
Rename ceph_mds_request_head with a _old extension, and a union
ceph_mds_request_args_ext to allow for the extended size of the new
header format.
Add the appropriate code to handle both formats in struct
create_request_message and key the behavior on whether the peer supports
CEPH_FEATURE_FS_BTIME.
The gid_list field in the payload is now populated from the saved
credential. For now, we don't add any support for setting the btime via
setattr, but this does enable us to add that in the future.
[ idryomov: break unnecessarily long lines ]
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
We can always get the mdsc from the session, so there's no need to pass
it in as a separate argument. Pass the session to __prepare_send_request
as well, to prepare for later patches that will need to access it.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Replace req->r_uid/r_gid with an r_cred pointer and take a reference to
that at the point where we previously would sample the two. Use that to
populate the uid and gid in the header and release the reference when
the request is freed.
This should enable us to later add support for sending supplementary
group lists in MDS requests.
[ idryomov: break unnecessarily long lines ]
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
When setting the file/dir layout, it may need data pool info. So
in mds server, it needs to check the osdmap. At present, if mds
doesn't find the data pool specified, it will try to get the latest
osdmap. Now if pass the osd epoch for setxattr, the mds server can
only check this epoch of osdmap.
URL: https://tracker.ceph.com/issues/48504
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The variable i is being initialized with a value that is never read
and it is being updated later with a new value in a for-loop. The
initialization is redundant and can be removed.
Addresses-Coverity: ("Unused value")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Geng Jichao reported a rather complex deadlock involving several
moving parts:
1) readahead is issued against an inode and some of its pages are locked
while the read is in flight
2) the same inode is evicted from the cache, and this task gets stuck
waiting for the page lock because of the above readahead
3) another task is processing a reply trace, and looks up the inode
being evicted while holding the s_mutex. That ends up waiting for the
eviction to complete
4) a write reply for an unrelated inode is then processed in the
ceph_con_workfn job. It calls ceph_check_caps after putting wrbuffer
caps, and that gets stuck waiting on the s_mutex held by 3.
The reply to "1" is stuck behind the write reply in "4", so we deadlock
at that point.
This patch changes the trace processing to call ceph_get_inode outside
of the s_mutex and snap_rwsem, which should break the cycle above.
[ idryomov: break unnecessarily long lines ]
URL: https://tracker.ceph.com/issues/47998
Reported-by: Geng Jichao <gengjichao@jd.com>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Luis Henriques <lhenriques@suse.de>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Convert some decodes into unused variables into skips, and fix up some
non-kerneldoc comment headers to not start with "/**".
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Ilya noticed that the first access to a blacklisted mount would often
get back -EACCES, but then subsequent calls would be OK. The problem is
in __do_request. If the session is marked as REJECTED, a hard error is
returned instead of waiting for a new session to come into being.
When the session is REJECTED and the mount was done with
recover_session=clean, queue the request to the waiting_for_map queue,
which will be awoken after tearing down the old session. We can only
do this for sync requests though, so check for async ones first and
just let the callers redrive a sync request.
URL: https://tracker.ceph.com/issues/47385
Reported-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: "Yan, Zheng" <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
30 minutes is a long time to wait, and this makes it difficult to test
the feature by manually blocklisting clients. Remove the timeout
infrastructure and just allow the client to reconnect at will.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: "Yan, Zheng" <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
When recovering a session (a'la recover_session=clean), we want to do
all of the operations that we do on a forced umount, but changing the
mount state to SHUTDOWN is can cause queued MDS requests to fail when
the session comes back. Most of those can idle until the session is
recovered in this situation.
Reserve SHUTDOWN state for forced umount, and make a new RECOVER state
for the forced reconnect situation. Change several tests for equality with
SHUTDOWN to test for that or RECOVER.
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: "Yan, Zheng" <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Some messages sent by the MDS entail a session sequence number
increment, and the MDS will drop certain types of requests on the floor
when the sequence numbers don't match.
In particular, a REQUEST_CLOSE message can cross with one of the
sequence morphing messages from the MDS which can cause the client to
stall, waiting for a response that will never come.
Originally, this meant an up to 5s delay before the recurring workqueue
job kicked in and resent the request, but a recent change made it so
that the client would never resend, causing a 60s stall unmounting and
sometimes a blockisting event.
Add a new helper for incrementing the session sequence and then testing
to see whether a REQUEST_CLOSE needs to be resent, and move the handling
of CEPH_MDS_SESSION_CLOSING into that function. Change all of the
bare sequence counter increments to use the new helper.
Reorganize check_session_state with a switch statement. It should no
longer be called when the session is CLOSING, so throw a warning if it
ever is (but still handle that case sanely).
[ idryomov: whitespace, pr_err() call fixup ]
URL: https://tracker.ceph.com/issues/47563
Fixes: fa99677342 ("ceph: fix potential mdsc use-after-free crash")
Reported-by: Patrick Donnelly <pdonnell@redhat.com>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
error_string key in the metadata map of MClientSession message
is intended for humans, but unfortunately became part of the on-wire
format with the introduction of recover_session=clean mode in commit
131d7eb4fa ("ceph: auto reconnect after blacklisted").
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Since nautilus, MDS tracks dirfrags whose child inodes have caps in open
file table. When MDS recovers, it prefetches all of these dirfrags. This
avoids using backtrace to load inodes. But dirfrags prefetch may load
lots of useless inodes into cache, and make MDS run out of memory.
Recent MDS adds an option that disables dirfrags prefetch. When dirfrags
prefetch is disabled. Recovering MDS only prefetches corresponding dir
inodes. Including inodes' parent/d_name in cap reconnect message can
help MDS to load inodes into its cache.
Signed-off-by: "Yan, Zheng" <zyan@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Most session messages contain a feature mask, but the MDS will
routinely send a REJECT message with one that is zero-length.
Commit 0fa8263367 ("ceph: fix endianness bug when handling MDS
session feature bits") fixed the decoding of the feature mask,
but failed to account for the MDS sending a zero-length feature
mask. This causes REJECT message decoding to fail.
Skip trying to decode a feature mask if the word count is zero.
Cc: stable@vger.kernel.org
URL: https://tracker.ceph.com/issues/46823
Fixes: 0fa8263367 ("ceph: fix endianness bug when handling MDS session feature bits")
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Ilya Dryomov <idryomov@gmail.com>
Tested-by: Patrick Donnelly <pdonnell@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
When doing some tests with multiple mds, we were seeing many mds
forwarding requests between them, causing clients to resend.
If the request is a modification operation and the mode is set to
USE_AUTH_MDS, then the auth mds should be selected to handle the
request. If auth mds for frag is already set, then it should be returned
directly without further processing.
The current logic is wrong because it only returns directly if
mode is USE_AUTH_MDS, but we want to do that for all modes. If we don't,
then when the frag's mds is not equal to cap session's mds, the request
will get sent to the wrong MDS needlessly.
Drop the mode check in this condition.
Signed-off-by: Yanhu Cao <gmayyyha@gmail.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
If the ceph_mdsc_init() fails, it will free the mdsc already.
Reported-by: syzbot+b57f46d8d6ea51960b8c@syzkaller.appspotmail.com
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
