Even the 'disable_send_metrics' is true so when the session is
being opened it will always trigger to send the metric for the
first time.
Cc: [email protected]
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Venky Shankar <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If a client sends out a cap update dropping caps with the prior 'seq'
just before an incoming cap revoke request, then the client may drop
the revoke because it believes it's already released the requested
capabilities.
This causes the MDS to wait indefinitely for the client to respond
to the revoke. It's therefore always a good idea to ack the cap
revoke request with the bumped up 'seq'.
Cc: [email protected]
Link: https://tracker.ceph.com/issues/61782
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Reviewed-by: Patrick Donnelly <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
In case:
mds client
- Releases cap and put Inode
- Increase cap->seq and sends
revokes req to the client
- Receives release req and - Receives & drops the revoke req
skip removing the cap and
then eval the CInode and
issue or revoke caps again.
- Receives & drops the caps update
or revoke req
- Health warning for client
isn't responding to
mclientcaps(revoke)
All the IMPORT/REVOKE/GRANT cap ops will increase the session seq
in MDS side and then the client need to issue a cap release to
unblock MDS to remove the corresponding cap to unblock possible
waiters.
Link: https://tracker.ceph.com/issues/61332
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The 'i_wr_ref' is used to track the 'Fb' caps, while whenever the 'Fb'
caps is took the kclient will always take the 'Fw' caps at the same
time. That means it will always be a false check in __ceph_finish_cap_snap().
When writing to buffer the kclient will take both 'Fb|Fw' caps and then
write the contents to the buffer pages by increasing the 'i_wrbuffer_ref'
and then just release both 'Fb|Fw'. This is different with the user
space libcephfs, which will keep the 'Fb' being took and use 'i_wr_ref'
instead of 'i_wrbuffer_ref' to track this until the buffer is flushed
to Rados.
We need to defer flushing the capsnap until the corresponding buffer
pages are all flushed to Rados, and at the same time just trigger to
flush the buffer pages immediately.
Link: https://tracker.ceph.com/issues/48640
Link: https://tracker.ceph.com/issues/59343
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Blindly expanding the readahead windows will cause unneccessary
pagecache thrashing and also will introduce the network workload.
We should disable expanding the windows if the readahead is disabled
and also shouldn't expand the windows too much.
Expanding forward firstly instead of expanding backward for possible
sequential reads.
Bound `rreq->len` to the actual file size to restore the previous page
cache usage.
The posix_fadvise may change the maximum size of a file readahead.
Cc: [email protected]
Fixes: 4987005600 ("ceph: convert ceph_readpages to ceph_readahead")
Link: https://lore.kernel.org/ceph-devel/[email protected]
Link: https://www.spinics.net/lists/ceph-users/msg76183.html
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-and-tested-by: Hu Weiwen <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
For write requests the parent's mtime will be updated correspondingly.
And if the 'Xx' caps is issued and when releasing other caps together
with the write requests the MDS Locker will try to eval the xattr lock,
which need to change the locker state excl --> sync and need to do Xx
caps revocation.
Just voluntarily dropping CEPH_CAP_XATTR_EXCL caps to avoid a cap
revoke message, which could cause the mtime will be overwrote by stale
one.
[ idryomov: break unnecessarily long lines ]
Link: https://tracker.ceph.com/issues/61584
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When the msgs are corrupted we need to dump them and then it will
be easier to dig what has happened and where the issue is.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When the MDS rank is in clientreplay state, the metrics requests
will be discarded directly. Also, when there are a lot of known
client requests to recover from, the metrics requests will slow
down the MDS rank from getting to the active state sooner.
With this patch, we will send the metrics requests only when the
MDS rank is in active state.
Link: https://tracker.ceph.com/issues/61524
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
There is a race between capsnaps flush and removing the inode from
'mdsc->snap_flush_list' list:
== Thread A == == Thread B ==
ceph_queue_cap_snap()
-> allocate 'capsnapA'
->ihold('&ci->vfs_inode')
->add 'capsnapA' to 'ci->i_cap_snaps'
->add 'ci' to 'mdsc->snap_flush_list'
...
== Thread C ==
ceph_flush_snaps()
->__ceph_flush_snaps()
->__send_flush_snap()
handle_cap_flushsnap_ack()
->iput('&ci->vfs_inode')
this also will release 'ci'
...
== Thread D ==
ceph_handle_snap()
->flush_snaps()
->iterate 'mdsc->snap_flush_list'
->get the stale 'ci'
->remove 'ci' from ->ihold(&ci->vfs_inode) this
'mdsc->snap_flush_list' will WARNING
To fix this we will increase the inode's i_count ref when adding 'ci'
to the 'mdsc->snap_flush_list' list.
[ idryomov: need_put int -> bool ]
Cc: [email protected]
Link: https://bugzilla.redhat.com/show_bug.cgi?id=2209299
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Milind Changire <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Smatch static checker warning:
fs/ceph/mds_client.c:3968 reconnect_caps_cb()
warn: missing error code here? '__get_cap_for_mds()' failed. 'err' = '0'
[ idryomov: Dan says that Smatch considers it intentional only if the
"ret = 0;" assignment is within 4 or 5 lines of the goto. ]
Reported-by: Dan Carpenter <[email protected]>
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When exporting the kceph to NFS it may pass a DCACHE_DISCONNECTED
dentry for the link operation. Then it will parse this dentry as a
snapdir, and the mds will fail the link request as -EROFS.
MDS allow clients to pass a ino# instead of a path.
Link: https://tracker.ceph.com/issues/59515
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When trimming the caps and just after the 'session->s_cap_lock' is
released in ceph_iterate_session_caps() the cap maybe removed by
another thread, and when using the stale cap memory in the callbacks
it will trigger use-after-free crash.
We need to check the existence of the cap just after the 'ci->i_ceph_lock'
being acquired. And do nothing if it's already removed.
Cc: [email protected]
Link: https://tracker.ceph.com/issues/43272
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Luís Henriques <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
While the mapped IOs continue if we try to flush a file's buffer
we can see that the fsync() won't complete until the IOs finish.
This is analogous to Jan Kara's commit (f446daaea9 mm: implement
writeback livelock avoidance using page tagging), we will try to
avoid livelocks of writeback when some steadily creates dirty pages
in a mapping we are writing out.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If the xattr's value size is long enough the kernel will warn and
then will fail the xfstests test case.
Just print part of the value string if it's too long.
At the same time fix the function name issue in the debug logs.
Link: https://tracker.ceph.com/issues/58404
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The fallocate will try to clear the suid/sgid if a unprevileged user
changed the file.
There is no POSIX item requires that we should clear the suid/sgid
in fallocate code path but this is the default behaviour for most of
the filesystems and the VFS layer. And also the same for the write
code path, which have already support it.
And also we need to update the time stamps since the fallocate will
change the file contents.
Cc: [email protected]
Link: https://tracker.ceph.com/issues/58054
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When received corrupted snap trace we don't know what exactly has
happened in MDS side. And we shouldn't continue IOs and metadatas
access to MDS, which may corrupt or get incorrect contents.
This patch will just block all the further IO/MDS requests
immediately and then evict the kclient itself.
The reason why we still need to evict the kclient just after
blocking all the further IOs is that the MDS could revoke the caps
faster.
Link: https://tracker.ceph.com/issues/57686
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Venky Shankar <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
These flags are only used in ceph filesystem in fs/ceph, so just
move it to the place it should be.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Venky Shankar <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When ceph releasing the file_lock it will try to get the inode pointer
from the fl->fl_file, which the memory could already be released by
another thread in filp_close(). Because in VFS layer the fl->fl_file
doesn't increase the file's reference counter.
Will switch to use ceph dedicate lock info to track the inode.
And in ceph_fl_release_lock() we should skip all the operations if the
fl->fl_u.ceph.inode is not set, which should come from the request
file_lock. And we will set fl->fl_u.ceph.inode when inserting it to the
inode lock list, which is when copying the lock.
Link: https://tracker.ceph.com/issues/57986
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
For the POSIX locks they are using the same owner, which is the
thread id. And multiple POSIX locks could be merged into single one,
so when checking whether the 'file' has locks may fail.
For a file where some openers use locking and others don't is a
really odd usage pattern though. Locks are like stoplights -- they
only work if everyone pays attention to them.
Just switch ceph_get_caps() to check whether any locks are set on
the inode. If there are POSIX/OFD/FLOCK locks on the file at the
time, we should set CHECK_FILELOCK, regardless of what fd was used
to set the lock.
Fixes: ff5d913dfc ("ceph: return -EIO if read/write against filp that lost file locks")
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When unlinking a file the kclient will send a unlink request to MDS
by holding the dentry reference, and then the MDS will return 2 replies,
which are unsafe reply and a deferred safe reply.
After the unsafe reply received the kernel will return and succeed
the unlink request to user space apps.
Only when the safe reply received the dentry's reference will be
released. Or the dentry will only be unhashed from dcache. But when
the open_by_handle_at() begins to open the unlinked files it will
succeed.
The inode->i_count couldn't be used to check whether the inode is
opened or not.
Link: https://tracker.ceph.com/issues/56524
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Reviewed-by: Luís Henriques <[email protected]>
Tested-by: Luís Henriques <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
For write when trying to get the Fwb caps we need to keep waiting
on transition from WRBUFFER|WR -> WR to avoid a new WR sync write
from going before a prior buffered writeback happens.
While for read there is no need to wait on transition from
RDCACHE|RD -> RD, and we can just exclude the revoking caps and
force to sync read.
Signed-off-by: Xiubo Li <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The netfs_write_begin() won't set the folio if the return value
is non-zero.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The f_frsize maybe changed in the quota size is less than the defualt
4MB.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When the quota is approaching we need to notify it to the MDS as
soon as possible, or the client could write to the directory more
than expected.
This will flush the dirty caps without delaying after each write,
though this couldn't prevent the real size of a directory exceed
the quota but could prevent it as soon as possible.
Link: https://tracker.ceph.com/issues/56180
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Luís Henriques <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If the 'i_inline_version' is 1, that means the file is just new
created and there shouldn't have any inline data in it, we should
skip retrieving the inline data from MDS.
This also could help reduce possiblity of dead lock issue introduce
by the inline data and Fcr caps.
Gradually we will remove the inline feature from kclient after ceph's
scrub too have support to unline the inline data, currently this
could help reduce the teuthology test failures.
This is possiblly could also fix a bug that for some old clients if
they couldn't explictly uninline the inline data when writing, the
inline version will keep as 1 always. We may always reading non-exist
data from inline data.
Signed-off-by: Xiubo Li <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
For async create we will always try to choose the auth MDS of frag
the dentry belonged to of the parent directory to send the request
and ususally this works fine, but if the MDS migrated the directory
to another MDS before it could be handled the request will be
forwarded. And then the auth cap will be changed.
We need to update the auth cap in this case before the request is
forwarded.
Link: https://tracker.ceph.com/issues/55857
Signed-off-by: Xiubo Li <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The 'old_size' is a __le64 type since birth, not sure why the
kclient incorrectly switched it to __le32. This change is okay
won't break anything because union will always allocate more memory
than the 'open' member needed.
Rename 'file_replication' to 'pool' as ceph did. Though this 'open'
struct may never be used in kclient in future, it's confusing when
going through the ceph code.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The incorrect comment is misleading. Acutally the last members
in ceph_mds_caps strcut is a union for none export and export
bodies.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
And for the 'Xs' caps for getxattr we will also choose the auth MDS,
because the MDS side code is buggy due to setxattr won't notify the
replica MDSes when the values changed and the replica MDS will return
the old values. Though we will fix it in MDS code, but this still
makes sense for old ceph.
Link: https://tracker.ceph.com/issues/55331
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Ilya Dryomov <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If the connection was accidently closed due to the socket issue or
something else the clients will try to open the opened sessions, the
MDSes will send the session open reply one more time if the clients
support the notify feature.
When the clients retry to open the sessions the s_seq will be 0 as
default, we need to update it anyway.
Link: https://tracker.ceph.com/issues/53911
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
In async unlink case the kclient won't wait for the first reply
from MDS and just drop all the links and unhash the dentry and then
succeeds immediately.
For any new create/link/rename,etc requests followed by using the
same file names we must wait for the first reply of the inflight
unlink request, or the MDS possibly will fail these following
requests with -EEXIST if the inflight async unlink request was
delayed for some reasons.
And the worst case is that for the none async openc request it will
successfully open the file if the CDentry hasn't been unlinked yet,
but later the previous delayed async unlink request will remove the
CDenty. That means the just created file is possiblly deleted later
by accident.
We need to wait for the inflight async unlink requests to finish
when creating new files/directories by using the same file names.
Link: https://tracker.ceph.com/issues/55332
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
This macro was added but never be used. And check the ceph code
there has another CEPHFS_FEATURES_MDS_REQUIRED but always be empty.
We should clean up all this related code, which make no sense but
introducing confusion.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Reviewed-by: Luís Henriques <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
check_write_begin() will unlock and put the folio when return
non-zero. So we should avoid unlocking and putting it twice in
netfs layer.
Change the way ->check_write_begin() works in the following two ways:
(1) Pass it a pointer to the folio pointer, allowing it to unlock and put
the folio prior to doing the stuff it wants to do, provided it clears
the folio pointer.
(2) Change the return values such that 0 with folio pointer set means
continue, 0 with folio pointer cleared means re-get and all error
codes indicating an error (no special treatment for -EAGAIN).
[ bagasdotme: use Sphinx code text syntax for *foliop pointer ]
Cc: [email protected]
Link: https://tracker.ceph.com/issues/56423
Link: https://lore.kernel.org/r/[email protected]
Co-developed-by: David Howells <[email protected]>
Signed-off-by: Xiubo Li <[email protected]>
Signed-off-by: David Howells <[email protected]>
Signed-off-by: Bagas Sanjaya <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If the task is placed in the TASK_INTERRUPTIBLE state it will sleep
until either something explicitly wakes it up, or a non-masked signal
is received. Switch to TASK_KILLABLE to avoid the noises.
Cc: Matthew Wilcox <[email protected]>
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If the pagecaches writeback just finished and the i_wrbuffer_ref
reaches zero it will try to trigger ceph_check_caps(). But if just
before ceph_check_caps() the i_wrbuffer_ref could be increased
again by mmap/cache write, then the Fwb revoke will fail.
We need to try to queue a writeback in this case instead of
triggering the writeback by BDI's delayed work per 5 seconds.
URL: https://tracker.ceph.com/issues/46904
URL: https://tracker.ceph.com/issues/55377
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
1, mount with wsync.
2, create a file with O_RDWR, and the request was sent to mds.0:
ceph_atomic_open()-->
ceph_mdsc_do_request(openc)
finish_open(file, dentry, ceph_open)-->
ceph_open()-->
ceph_init_file()-->
ceph_init_file_info()-->
ceph_uninline_data()-->
{
...
if (inline_version == 1 || /* initial version, no data */
inline_version == CEPH_INLINE_NONE)
goto out_unlock;
...
}
The inline_version will be 1, which is the initial version for the
new create file. And here the ci->i_inline_version will keep with 1,
it's buggy.
3, buffer write to the file immediately:
ceph_write_iter()-->
ceph_get_caps(file, need=Fw, want=Fb, ...);
generic_perform_write()-->
a_ops->write_begin()-->
ceph_write_begin()-->
netfs_write_begin()-->
netfs_begin_read()-->
netfs_rreq_submit_slice()-->
netfs_read_from_server()-->
rreq->netfs_ops->issue_read()-->
ceph_netfs_issue_read()-->
{
...
if (ci->i_inline_version != CEPH_INLINE_NONE &&
ceph_netfs_issue_op_inline(subreq))
return;
...
}
ceph_put_cap_refs(ci, Fwb);
The ceph_netfs_issue_op_inline() will send a getattr(Fsr) request to
mds.1.
4, then the mds.1 will request the rd lock for CInode::filelock from
the auth mds.0, the mds.0 will do the CInode::filelock state transation
from excl --> sync, but it need to revoke the Fxwb caps back from the
clients.
While the kernel client has aleady held the Fwb caps and waiting for
the getattr(Fsr).
It's deadlock!
URL: https://tracker.ceph.com/issues/55377
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If any 'x' caps is issued we can just choose the auth MDS instead
of the random replica MDSes. Because only when the Locker is in
LOCK_EXEC state will the loner client could get the 'x' caps. And
if we send the getattr requests to any replica MDS it must auth pin
and tries to rdlock from the auth MDS, and then the auth MDS need
to do the Locker state transition to LOCK_SYNC. And after that the
lock state will change back.
This cost much when doing the Locker state transition and usually
will need to revoke caps from clients.
URL: https://tracker.ceph.com/issues/55240
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Before waiting for a request's safe reply, we will send the mdlog flush
request to the relevant MDS. And this will also flush the mdlog for all
the other unsafe requests in the same session, so we can record the last
session and no need to flush mdlog again in the next loop. But there
still have cases that it may send the mdlog flush requst twice or more,
but that should be not often.
Rename wait_unsafe_requests() to
flush_mdlog_and_wait_mdsc_unsafe_requests() to make it more
descriptive.
[xiubli: fold in MDS request refcount leak fix from Jeff]
URL: https://tracker.ceph.com/issues/55284
URL: https://tracker.ceph.com/issues/55411
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Rename it to flush_mdlog_and_wait_inode_unsafe_requests() to make
it more descriptive.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
From the posix and the initial statx supporting commit comments,
the AT_STATX_DONT_SYNC is a lightweight stat and the
AT_STATX_FORCE_SYNC is a heaverweight one. And also checked all
the other current usage about these two flags they are all doing
the same, that is only when the AT_STATX_FORCE_SYNC is not set
and the AT_STATX_DONT_SYNC is set will they skip sync retriving
the attributes from storage.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: David Howells <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The MDS will always refresh the dentry lease when removing the files
or directories. And if the dentry is still hashed, we can update
the dentry lease and no need to do the lookup from the MDS later.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The type of 'r_attempts' in kernel 'ceph_mds_request' is 'int',
while in 'ceph_mds_request_head' the type of 'num_retry' is '__u8'.
So in case the request retries exceeding 256 times, the MDS will
receive a incorrect retry seq.
In this case it's ususally a bug in MDS and continue retrying the
request makes no sense. For now let's limit it to 256. In future
this could be fixed in ceph code, so avoid using the hardcode here.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The type of 'num_fwd' in ceph 'MClientRequestForward' is 'int32_t',
while in 'ceph_mds_request_head' the type is '__u8'. So in case
the request bounces between MDSes exceeding 256 times, the client
will get stuck.
In this case it's ususally a bug in MDS and continue bouncing the
request makes no sense.
URL: https://tracker.ceph.com/issues/55130
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Reviewed-by: Luís Henriques <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The ceph_mdsc_lease_release() has been removed by commit 8aa152c778
(ceph: remove ceph_mdsc_lease_release). ceph_mdsc_lease_send_msg will
never be called with CEPH_MDS_LEASE_RELEASE.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The pages in the file mapping maybe reclaimed and reused by other
subsystems and the page->private maybe used as flags field or
something else, if later that pages are used by page caches again
the page->private maybe not cleared as expected.
Here will check the PG_private bit instead of the folio->private.
Cc: [email protected]
URL: https://tracker.ceph.com/issues/55421
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Luis Henriques <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Once the session is opened the s->s_ttl will be set, and when receiving
a new mdsmap and the MDS map is changed, it will be possibly will close
some sessions and open new ones. And then some sessions will be in
CLOSING state evening without unmounting.
URL: https://tracker.ceph.com/issues/54979
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Reset the last_readdir at the same time, and add a comment explaining
why we don't free last_readdir when dir_emit returns false.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The ceph_get_inode() will search for or insert a new inode into the
hash for the given vino, and return a reference to it. If new is
non-NULL, its reference is consumed.
We should release the reference when in error handing cases.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
To make the logs more readable such as for log likes:
ceph: will move 00000000a42b796b to split realm 100000003ed 000000007146df45
With this it will always show the inode numbers instead the inode
addresses.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Ceph have removed this macro and the 0x3 will be use for global dummy
snaprealm.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
We will only track the uppest parent snapshot realm from which we
need to rebuild the snapshot contexts _downward_ in hierarchy. For
all the others having no new snapshot we will do nothing.
This fix will avoid calling ceph_queue_cap_snap() on some inodes
inappropriately. For example, with the code in mainline, suppose there
are 2 directory hierarchies (with 6 directories total), like this:
/dir_X1/dir_X2/dir_X3/
/dir_Y1/dir_Y2/dir_Y3/
Firstly, make a snapshot under /dir_X1/dir_X2/.snap/snap_X2, then make a
root snapshot under /.snap/root_snap. Every time we make snapshots under
/dir_Y1/..., the kclient will always try to rebuild the snap context for
snap_X2 realm and finally will always try to queue cap snaps for dir_Y2
and dir_Y3, which makes no sense.
That's because the snap_X2's seq is 2 and root_snap's seq is 3. So when
creating a new snapshot under /dir_Y1/... the new seq will be 4, and
the mds will send the kclient a snapshot backtrace in _downward_
order: seqs 4, 3.
When ceph_update_snap_trace() is called, it will always rebuild the from
the last realm, that's the root_snap. So later when rebuilding the snap
context, the current logic will always cause it to rebuild the snap_X2
realm and then try to queue cap snaps for all the inodes related in that
realm, even though it's not necessary.
This is accompanied by a lot of these sorts of dout messages:
"ceph: queue_cap_snap 00000000a42b796b nothing dirty|writing"
Fix the logic to avoid this situation.
Also, the 'invalidate' word is not precise here. In actuality, it will
cause a rebuild of the existing snapshot contexts or just build
non-existent ones. Rename it to 'rebuild_snapcs'.
URL: https://tracker.ceph.com/issues/44100
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
This potentially will cause a bug in future if using an old ceph
version that sends a smaller inode struct, which can cause some members
to be skipped in handle_reply.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
There could be huge number of capsnaps around at any given time. On
x86_64 the structure is 248 bytes, which will be rounded up to 256 bytes
by kzalloc. Move this to a dedicated slabcache to save 8 bytes for each.
[ jlayton: use kmem_cache_zalloc ]
Signed-off-by: Xiubo Li <[email protected]>
Signed-off-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If MDS return ESTALE, that means the MDS has already iterated all the
possible active MDSes including the auth MDS or the inode is under
purging. No need to retry in auth MDS and will just return ESTALE
directly. Retrying in this situation will cause an infinite loop.
Also, retrying like this would prevent the kernel VFS layer ESTALE
handling from working properly. An ESTALE error is usually an indication
that the dcache is wrong, so we want to allow the VFS to redo the lookup
and revalidate it properly.
URL: https://tracker.ceph.com/issues/53504
Signed-off-by: Xiubo Li <[email protected]>
Acked-by: Greg Farnum <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
When failing to allocate the sessions memory we should make sure
the req1 and req2 and the sessions get put. And also in case the
max_sessions decreased so when kreallocate the new memory some
sessions maybe missed being put.
And if the max_sessions is 0 krealloc will return ZERO_SIZE_PTR,
which will lead to a distinct access fault.
URL: https://tracker.ceph.com/issues/53819
Fixes: e1a4541ec0 ("ceph: flush the mdlog before waiting on unsafe reqs")
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Venky Shankar <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The newcaps has already included the Ls, no need to check it again.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The smatch static checker warned about an uninitialized symbol usage in
this function, in the case where ceph_mdsc_build_path returns an error.
It turns out that that case is harmless, but it just looks sketchy.
Initialize the variable at declaration time, and remove the unneeded
setting of it later.
Fixes: a33f6432b3 ("ceph: encode inodes' parent/d_name in cap reconnect message")
Reported-by: Dan Carpenter <[email protected]>
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Currently, if the sync read handler ends up reading more from the last
object in the file than the i_size indicates, then it'll end up
returning the wrong length. Ensure that we cap the returned length and
pos at the EOF.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If the max_mds is decreased in a cephfs cluster, there is a window
of time before the MDSs are removed. If a map goes out during this
period, the mdsmap may show the decreased max_mds but still shows
those MDSes as in or in the export target list.
Ensure that we don't fail the map decode in that case.
Cc: [email protected]
URL: https://tracker.ceph.com/issues/52436
Fixes: d517b3983d ("ceph: reconnect to the export targets on new mdsmaps")
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
If the new size is the same as the current size, the MDS will do nothing
but change the mtime/atime. POSIX doesn't mandate that the filesystems
must update them in this case, so just ignore it instead.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
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 <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
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 <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
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 <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
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 <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
For the client requests who will have unsafe and safe replies from
MDS daemons, in the MDS side the MDS daemons won't flush the mdlog
(journal log) immediatelly, because they think it's unnecessary.
That's true for most cases but not all, likes the fsync request.
The fsync will wait until all the unsafe replied requests to be
safely replied.
Normally if there have multiple threads or clients are running, the
whole mdlog in MDS daemons could be flushed in time if any request
will trigger the mdlog submit thread. So usually we won't experience
the normal operations will stuck for a long time. But in case there
has only one client with only thread is running, the stuck phenomenon
maybe obvious and the worst case it must wait at most 5 seconds to
wait the mdlog to be flushed by the MDS's tick thread periodically.
This patch will trigger to flush the mdlog in the relevant and auth
MDSes to which the in-flight requests are sent just before waiting
the unsafe requests to finish.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
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: [email protected]
URL: https://tracker.ceph.com/issues/52283
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
This will collect IO's total size and then calculate the average
size, and also will collect the min/max IO sizes.
The debugfs will show the size metrics in bytes and will let the
userspace applications to switch to what they need.
URL: https://tracker.ceph.com/issues/49913
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
The new __update_stdev() helper will only compute the standard
deviation.
Signed-off-by: Xiubo Li <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
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 <[email protected]>
Reviewed-by: Jeff Layton <[email protected]>
Signed-off-by: Ilya Dryomov <[email protected]>
Xiubo Li