The use of a completion when waiting for session shutdown during umount is
inappropriate, given the complexity of the condition. For multiple MDS's,
this resulted in the umount thread spinning, often preventing the session
close message from being processed in some cases.
Switch to a waitqueue and defined a condition helper. This cleans things
up nicely.
Signed-off-by: Sage Weil <sage@newdream.net>
Implement flock inode operation to support advisory file locking. All
lock/unlock operations are synchronous with the MDS. Lock state is
sent when reconnecting to a recovering MDS to restore the shared lock
state.
Signed-off-by: Greg Farnum <gregf@hq.newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
Encode either old or v2 encoding of client_reconnect message, depending on
whether the peer has the FLOCK feature bit.
Signed-off-by: Sage Weil <sage@newdream.net>
When we get a cap EXPORT message, make sure we are connected to all export
targets to ensure we can handle the matching IMPORT.
Signed-off-by: Sage Weil <sage@newdream.net>
If an MDS we are talking to may have failed, we need to open sessions to
its potential export targets to ensure that any in-progress migration that
may have involved some of our caps is properly handled.
Signed-off-by: Sage Weil <sage@newdream.net>
Caps related accounting is now being done per mds client instead
of just being global. This prepares ground work for a later revision
of the caps preallocated reservation list.
Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
This fixes an issue triggered by running concurrent syncs. One of the syncs
would go through while the other would just hang indefinitely. In any case, we
never actually want to wake a single waiter, so the *_all functions should
be used.
Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
Strip the cap and dentry releases from replayed messages. They can
cause the shared state to get out of sync because they were generated
(with the request message) earlier, and no longer reflect the current
client state.
Signed-off-by: Sage Weil <sage@newdream.net>
Replayed rename operations (after an mds failure/recovery) were broken
because the request paths were regenerated from the dentry names, which
get mangled when d_move() is called.
Instead, resend the previous request message when replaying completed
operations. Just make sure the REPLAY flag is set and the target ino is
filled in.
This fixes problems with workloads doing renames when the MDS restarts,
where the rename operation appears to succeed, but on mds restart then
fails (leading to client confusion, app breakage, etc.).
Signed-off-by: Sage Weil <sage@newdream.net>
This fixes a race between handle_reply finishing an mds request, signalling
completion, and then dropping the request structing and its dentry+inode
refs, and pre_umount function waiting for requests to finish before
letting the vfs tear down the dcache. If umount was delayed waiting for
mds requests, we could race and BUG in shrink_dcache_for_umount_subtree
because of a slow dput.
This delays umount until the msgr queue flushes, which means handle_reply
will exit and will have dropped the ceph_mds_request struct. I'm assuming
the VFS has already ensured that its calls have all completed and those
request refs have thus been dropped as well (I haven't seen that race, at
least).
Signed-off-by: Sage Weil <sage@newdream.net>
If we have enough memory to allocate a new cap release message, do so, so
that we can send a partial release message immediately. This keeps us from
making the MDS wait when the cap release it needs is in a partially full
release message.
If we fail because of ENOMEM, oh well, they'll just have to wait a bit
longer.
Signed-off-by: Sage Weil <sage@newdream.net>
If we get an IMPORT that give us a cap, but we don't have the inode, queue
a release (and try to send it immediately) so that the MDS doesn't get
stuck waiting for us.
Signed-off-by: Sage Weil <sage@newdream.net>
If the client revokes a lease with a higher seq than what we have, keep
the mds's seq, so that it honors our release. Otherwise, we can hang
indefinitely.
Signed-off-by: Sage Weil <sage@newdream.net>
If an mds request is aborted (timeout, SIGKILL), it is left registered to
keep our state in sync with the mds. If we get a forward notification,
though, we know the request didn't succeed and we can unregister it
safely. We were trying to resend it, but then bailing out (and not
unregistering) in __do_request.
Signed-off-by: Sage Weil <sage@newdream.net>
The lease code includes a mask in the CEPH_LOCK_* namespace, but that
namespace is changing, and only one mask (formerly _DN == 1) is used, so
hard code for that value for now.
If we ever extend this code to handle leases over different data types we
can extend it accordingly.
Signed-off-by: Sage Weil <sage@newdream.net>
The underlying problem is that many mds requests can't be restarted. For
example, a restarted create() would return -EEXIST if the original request
succeeds. However, we do not want a hung MDS to hang the client too. So,
use the _killable wait_for_completion variants to abort on SIGKILL but
nothing else.
Signed-off-by: Sage Weil <sage@newdream.net>
The C99 specification states in section 6.11.5:
The placement of a storage-class specifier other than at the beginning
of the declaration specifiers in a declaration is an obsolescent
feature.
Signed-off-by: Tobias Klauser <tklauser@distanz.ch>
Signed-off-by: Sage Weil <sage@newdream.net>
This is essential, as for the rados block device we'll need
to run in different contexts that would need flags that
are other than GFP_NOFS.
Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
We invalidate I_COMPLETE and dentry leases in two places: on aborted mds
request and on request replay. Use common helper to avoid duplicate code.
Signed-off-by: Sage Weil <sage@newdream.net>
The remove_session_caps() helper is called when an MDS closes out our
session (either normally, or as a result of a failed reconnect), and when
we tear down state for umount. If we remove the last cap, and there are
no cap migrations in progress, then there is little hope of us flushing
out that data to the mds (without heroic efforts to reconnect and flush).
So, to avoid leaving inodes pinned (due to dirty state) and crashing after
umount, throw out dirty caps state and unpin the inodes. Print a warning
to the console so we know something was lost.
NOTE: Although we drop wrbuffer refs, we don't actually mark pages clean;
maybe a truncate should be queued?
Signed-off-by: Sage Weil <sage@newdream.net>
Currently, if our session is closed (due to a timeout, or explicit close,
or whatever), we just sit there doing nothing unless/until the MDS
restarts, at which point we try to reconnect.
Change client to attempt an immediate reconnect if our session is closed.
Note that currently the MDS doesn't support this, and our attempt will
fail. We'll get a session CLOSE, our caps and dirty cap state will be
dropped, and the client will be free to attempt to reconnect. That's
clearly not as nice as a successful reconnect, but it at least allows us
to try to carry on, and in the future the MDS will support a reconnect
and we will fare better.
Signed-off-by: Sage Weil <sage@newdream.net>
Pass a ceph_mds_session, since the caller has it.
Remove the dead code for sending empty reconnects. It used to be used
when the MDS contacted _us_ to solicit a reconnect, and we could reply
saying "go away, I have no session." Now we only send reconnects based
on the mds map, and only when we do in fact have an open session.
Signed-off-by: Sage Weil <sage@newdream.net>
We used to infer reconnect success by watching the MDS state, essentially
assuming that hearing nothing meant things were ok. That wasn't
particularly reliable. Instead, the MDS replies with an explicit OPEN
message to indicate success.
Strictly speaking, this is a protocol change, but it is a backwards
compatible one that does not break new clients + old servers or old
clients + new servers. At least not yet.
Drop unused @all argument from kick_requests while we're at it.
Signed-off-by: Sage Weil <sage@newdream.net>
On OPENING we shouldn't have any caps (or releases).
On CLOSING, we should wait until we succeed (and throw it all out), or
don't (and are OPEN again).
On RECONNECTING we can wait until we are OPEN.
Signed-off-by: Sage Weil <sage@newdream.net>
If the MDS restarts, the expire caps state is no longer shared, and can be
thrown out. Caps state will be rebuilt on the MDS during the reconnect
process that follows. Zero out any release messages and adjust the
release counter accordingly.
Signed-off-by: Sage Weil <sage@newdream.net>
We only need to pass in front_len. Callers can attach any other payload
pieces (middle, data) as they see fit.
Signed-off-by: Sage Weil <sage@newdream.net>
Returning ERR_PTR(-ENOMEM) is useless extra work. Return NULL on failure
instead, and fix up the callers (about half of which were wrong anyway).
Signed-off-by: Sage Weil <sage@newdream.net>
If we abort a request, we return to caller, but the request may still
complete. And if we hold the dir FILE_EXCL bit, we may not release a
lease when sending a request. A simple un-tar, control-c, un-tar again
will reproduce the bug (manifested as a 'Cannot open: File exists').
Ensure we invalidate affected dentry leases (as well dir I_COMPLETE) so
we don't have valid (but incorrect) leases. Do the same, consistently, at
other sites where I_COMPLETE is similarly cleared.
Signed-off-by: Sage Weil <sage@newdream.net>
When we abort requests we need to prevent fill_trace et al from doing
anything that relies on locks held by the VFS caller. This fixes a race
between the reply handler and the abort code, ensuring that continue
holding the dir mutex until the reply handler completes.
Signed-off-by: Sage Weil <sage@newdream.net>
We would occasionally BUG out in the reply handler because r_reply was
nonzero, due to a race with ceph_mdsc_do_request temporarily setting
r_reply to an ERR_PTR value. This is unnecessary, messy, and also wrong
in the EIO case.
Clean up by consistently using r_err for errors and r_reply for messages.
Also fix the abort logic to trigger consistently for all errors that return
to the caller early (e.g., EIO from timeout case). If an abort races with
a reply, use the result from the reply.
Also fix locking for r_err, r_reply update in the reply handler.
Signed-off-by: Sage Weil <sage@newdream.net>
The iterate_session_caps helper traverses the session caps list and tries
to grab an inode reference. However, the __ceph_remove_cap was clearing
the inode backpointer _before_ removing itself from the session list,
causing a null pointer dereference.
Clear cap->ci under protection of s_cap_lock to avoid the race, and to
tightly couple the list and backpointer state. Use a local flag to
indicate whether we are releasing the cap, as cap->session may be modified
by a racing thread in iterate_session_caps.
Signed-off-by: Sage Weil <sage@newdream.net>
The session->s_waiting list is protected by mdsc->mutex, not s_mutex. This
was causing (rare) s_waiting list corruption.
Fix errors paths too, while we're here. A more thorough cleanup of this
function is coming soon.
Signed-off-by: Sage Weil <sage@newdream.net>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
There was a use after free in __unregister_request that would trigger
whenever the request map held the last reference. This appears to have
triggered an oops during 'umount -f' when requests are being torn down.
Signed-off-by: Sage Weil <sage@newdream.net>
Fix a broken check that a reply came back from the same MDS we sent the
request to. I don't think a case that actually triggers this would ever
come up in practice, but it's clearly wrong and easy to fix.
Reported-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>
Return ERR_PTR(-ENOMEM) if kmalloc() fails. We handle allocation
failures the same way later in the function.
Signed-off-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Sage Weil <sage@newdream.net>
Prevent duplicate 'mds0 caps stale' message from spamming the console every
few seconds while the MDS restarts. Set s_renew_requested earlier, so that
we only print the message once, even if we don't send an actual request.
Signed-off-by: Sage Weil <sage@newdream.net>
The wait_unsafe_requests() helper dropped the mdsc mutex to wait
for each request to complete, and then examined r_node to get the
next request after retaking the lock. But the request completion
removes the request from the tree, so r_node was always undefined
at this point. Since it's a small race, it usually led to a
valid request, but not always. The result was an occasional
crash in rb_next() while dereferencing node->rb_left.
Fix this by clearing the rb_node when removing the request from
the request tree, and not walking off into the weeds when we
are done waiting for a request. Since the request we waited on
will _always_ be out of the request tree, take a ref on the next
request, in the hopes that it won't be. But if it is, it's ok:
we can start over from the beginning (and traverse over older read
requests again).
Signed-off-by: Sage Weil <sage@newdream.net>