If the tcp connection drops and we reconnect to reestablish a stateful
session (with the mds), we need to resend previously sent (and possibly
received) messages with the _same_ seq # so that they can be dropped on
the other end if needed. Only assign a new seq once after the message is
queued.
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>
We shouldn't leak any prior memory contents to other parties. And random
data, particularly in the 'version' field, can cause problems down the
line.
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>
OSD requests need to be resubmitted on any pg mapping change, not just when
the pg primary changes. Resending only when the primary changes results in
occasional 'hung' requests during osd cluster recovery or rebalancing.
Signed-off-by: Sage Weil <sage@newdream.net>
It's possible the MDS will not issue caps on a snapped inode, in which case
an open request may not __ceph_get_fmode(), botching the open file
counting. (This is actually a server bug, but the client shouldn't BUG out
in this case.)
Signed-off-by: Sage Weil <sage@newdream.net>
The osd request wasn't being unregistered when the osd returned a failure
code, even though the result was returned to the caller. This would cause
it to eventually time out, and then crash the kernel when it tried to
resend the request using a stale page vector.
Signed-off-by: Sage Weil <sage@newdream.net>
The ->writepages writeback_control is not still valid in the writepages
completion. We were touching it solely to adjust pages_skipped when there
was a writeback error (EIO, ENOSPC, EPERM due to bad osd credentials),
causing an oops in the writeback code shortly thereafter. Updating
pages_skipped on error isn't correct anyway, so let's just rip out this
(clearly broken) code to pass the wbc to the completion.
Signed-off-by: Sage Weil <sage@newdream.net>
Unregister and destroy the bdi in put_super, after mount is r/o, but before
put_anon_super releases the device name.
For symmetry, bdi_destroy in destroy_client (we bdi_init in create_client).
Only set s_bdi if bdi_register succeeds, since we use it to decide whether
to bdi_unregister.
Signed-off-by: Sage Weil <sage@newdream.net>
It's useless, since our allocations are already a power of 2. And it was
allocated per-instance (not globally), which caused a name collision when
we tried to mount a second file system with auth_x enabled.
Signed-off-by: Sage Weil <sage@newdream.net>
If a rename operation is resent to the MDS following an MDS restart, the
client does not get a full reply (containing the resulting metadata) back.
In that case, a ceph_rename() needs to compensate by doing anything useful
that fill_inode() would have, like d_move().
It also needs to invalidate the dentry (to workaround the vfs_rename_dir()
bug) and clear the dir complete flag, just like fill_trace().
Signed-off-by: Sage Weil <sage@newdream.net>
We can get old message seq #'s after a tcp reconnect for stateful sessions
(i.e., the MDS). If we get a higher seq #, that is an error, and we
shouldn't see any bad seq #'s for stateless (mon, osd) connections.
Signed-off-by: Sage Weil <sage@newdream.net>
The snap realm split was checking i_snap_realm, not the list_head, to
determine if an inode belonged in the new realm. The check always failed,
which meant we always moved the inode, corrupting the old realm's list and
causing various crashes.
Also wait to release old realm reference to avoid possibility of use after
free.
Signed-off-by: Sage Weil <sage@newdream.net>
d_move() reorders the d_subdirs list, breaking the readdir result caching.
Unless/until d_move preserves that ordering, clear CEPH_I_COMPLETE on
rename.
Signed-off-by: Sage Weil <sage@newdream.net>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client:
ceph: use separate class for ceph sockets' sk_lock
ceph: reserve one more caps space when doing readdir
ceph: queue_cap_snap should always queue dirty context
ceph: fix dentry reference leak in dcache readdir
ceph: decode v5 of osdmap (pool names) [protocol change]
ceph: fix ack counter reset on connection reset
ceph: fix leaked inode ref due to snap metadata writeback race
ceph: fix snap context reference leaks
ceph: allow writeback of snapped pages older than 'oldest' snapc
ceph: fix dentry rehashing on virtual .snap dir
Use a separate class for ceph sockets to prevent lockdep confusion.
Because ceph sockets only get passed kernel pointers, there is no
dependency from sk_lock -> mmap_sem. If we share the same class as other
sockets, lockdep detects a circular dependency from
mmap_sem (page fault) -> fs mutex -> sk_lock -> mmap_sem
because dependencies are noted from both ceph and user contexts. Using
a separate class prevents the sk_lock(ceph) -> mmap_sem dependency and
makes lockdep happy.
Signed-off-by: Sage Weil <sage@newdream.net>
We were missing space for the directory cap. The result was a BUG at
fs/ceph/caps.c:2178.
Signed-off-by: Yehuda Sadeh <yehuda@hq.newdream.net>
Signed-off-by: Sage Weil <sage@newdream.net>
This simplifies the calling convention, and fixes a bug where we queue a
capsnap with a context other than i_head_snapc (the one that matches the
dirty pages). The result was a BUG at fs/ceph/caps.c:2178 on writeback
completion when a capsnap matching the writeback snapc could not be found.
Signed-off-by: Sage Weil <sage@newdream.net>
When filldir returned an error (e.g. buffer full for a large directory),
we would leak a dentry reference, causing an oops on umount.
Signed-off-by: Sage Weil <sage@newdream.net>
Teach the client to decode an updated format for the osdmap. The new
format includes pool names, which will be useful shortly. Get this change
in earlier rather than later.
Signed-off-by: Sage Weil <sage@newdream.net>
If in_seq_acked isn't reset along with in_seq, we don't ack received
messages until we reach the old count, consuming gobs memory on the other
end of the connection and introducing a large delay when those messages
are eventually deleted.
Signed-off-by: Sage Weil <sage@newdream.net>
We create a ceph_cap_snap if there is dirty cap metadata (for writeback to
mds) OR dirty pages (for writeback to osd). It is thus possible that the
metadata has been written back to the MDS but the OSD data has not when
the cap_snap is created. This results in a cap_snap with dirty(caps) == 0.
The problem is that cap writeback to the MDS isn't necessary, and a
FLUSHSNAP cap op gets no ack from the MDS. This leaves the cap_snap
attached to the inode along with its inode reference.
Fix the problem by dropping the cap_snap if it becomes 'complete' (all
pages written out) and dirty(caps) == 0 in ceph_put_wrbuffer_cap_refs().
Also, BUG() in __ceph_flush_snaps() if we encounter a cap_snap with
dirty(caps) == 0.
Signed-off-by: Sage Weil <sage@newdream.net>
The get_oldest_context() helper takes a reference to the returned snap
context, but most callers weren't dropping that reference. Fix them.
Also drop the unused locked __get_oldest_context() variant.
Signed-off-by: Sage Weil <sage@newdream.net>
On snap deletion, we don't regenerate ceph_cap_snaps for inodes with dirty
pages because deletion does not affect metadata writeback. However, we
did run into problems when we went to write back the pages because the
'oldest' snapc is determined by the oldest cap_snap, and that may be the
newer snapc that reflects the deletion. This caused confusion and an
infinite loop in ceph_update_writeable_page().
Change the snapc checks to allow writeback of any snapc that is equal to
OR older than the 'oldest' snapc.
When there are no cap_snaps, we were also using the realm's latest snapc
for writeback, which complicates ceph_put_wrbufffer_cap_refs(). Instead,
use i_head_snapc, the most snapc used for the most recent ('head') data.
This makes the writeback snapc (ceph_osd_request.r_snapc) _always_ match a
capsnap or i_head_snapc.
Also, in writepags_finish(), drop the snapc referenced by the _page_
and do not assume it matches the request snapc (it may not anymore).
Signed-off-by: Sage Weil <sage@newdream.net>
If a lookup fails on the magic .snap directory, we bind it to a magic
snap directory inode in ceph_lookup_finish(). That code assumes the dentry
is unhashed, but a recent server-side change started returning NULL leases
on lookup failure, causing the .snap dentry to be hashed and NULL by
ceph_fill_trace().
This causes dentry hash chain corruption, or a dies when d_rehash()
includes
BUG_ON(!d_unhashed(entry));
So, avoid processing the NULL dentry lease if it the dentry matches the
snapdir name in ceph_fill_trace(). That allows the lookup completion to
properly bind it to the snapdir inode. BUG there if dentry is hashed to
be sure.
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>
Clear pointer to mds request after dropping the reference to
ensure we don't drop it again, as there is at least one error
path through this function that does not reset fi->last_readdir
to a new value.
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>
Currently, if the wait_event_interruptible is interrupted, we
return EAGAIN unconditionally and loop, such that we aren't, in
fact, interruptible. So, propagate ERESTARTSYS if we get it.
Signed-off-by: Sage Weil <sage@newdream.net>
We were rebuilding the snap context when it was not necessary
(i.e. when the realm seq hadn't changed _and_ the parent seq
was still older), which caused page snapc pointers to not match
the realm's snapc pointer (even though the snap context itself
was identical). This confused begin_write and put it into an
endless loop.
The correct logic is: rebuild snapc if _my_ realm seq changed, or
if my parent realm's seq is newer than mine (and thus mine needs
to be rebuilt too).
Signed-off-by: Sage Weil <sage@newdream.net>
We get a fault callback on _every_ tcp connection fault. Normally, we
want to reopen the connection when that happens. If the address we have
is bad, however, and connection attempts always result in a connection
refused or similar error, explicitly closing and reopening the msgr
connection just prevents the messenger's backoff logic from kicking in.
The result can be a console full of
[ 3974.417106] ceph: osd11 10.3.14.138:6800 connection failed
[ 3974.423295] ceph: osd11 10.3.14.138:6800 connection failed
[ 3974.429709] ceph: osd11 10.3.14.138:6800 connection failed
Instead, if we get a fault, and have outstanding requests, but the osd
address hasn't changed and the connection never successfully connected in
the first place, do nothing to the osd connection. The messenger layer
will back off and retry periodically, because we never connected and thus
the lossy bit is not set.
Instead, touch each request's r_stamp so that handle_timeout can tell the
request is still alive and kicking.
Signed-off-by: Sage Weil <sage@newdream.net>
Make variable name slightly more generic, since it will (soon)
reflect either the time the request was sent OR the time it was
last determined to be still retrying.
Signed-off-by: Sage Weil <sage@newdream.net>
The messenger fault was clearing the BUSY bit, for reasons unclear. This
made it possible for the con->ops->fault function to reopen the connection,
and requeue work in the workqueue--even though the current thread was
already in con_work.
This avoids a problem where the client busy loops with connection failures
on an unreachable OSD, but doesn't address the root cause of that problem.
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 incremental map decoding of pg pool updates wasn't skipping
the snaps and removed_snaps vectors. This caused osd requests
to stall when pool snapshots were created or fs snapshots were
deleted. Use a common helper for full and incremental map
decoders that decodes pools properly.
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>
We were releasing used caps (e.g. FILE_CACHE) from encode_inode_release
with MDS requests (e.g. setattr). We don't carry refs on most caps, so
this code worked most of the time, but for setattr (utimes) we try to
drop Fscr.
This causes cap state to get slightly out of sync with reality, and may
result in subsequent mds revoke messages getting ignored.
Fix by only releasing unused caps.
Signed-off-by: Sage Weil <sage@newdream.net>
Drop session mutex unconditionally in handle_cap_grant, and do the
check_caps from the handle_cap_grant helper. This avoids using a magic
return value.
Also avoid using a flag variable in the IMPORT case and call
check_caps at the appropriate point.
Signed-off-by: Sage Weil <sage@newdream.net>
Passing a session pointer to ceph_check_caps() used to mean it would leave
the session mutex locked. That wasn't always possible if it wasn't passed
CHECK_CAPS_AUTHONLY. If could unlock the passed session and lock a
differet session mutex, which was clearly wrong, and also emitted a
warning when it a racing CPU retook it and we did an unlock from the wrong
context.
This was only a problem when there was more than one MDS.
First, make ceph_check_caps unconditionally drop the session mutex, so that
it is free to lock other sessions as needed. Then adjust the one caller
that passes in a session (handle_cap_grant) accordingly.
Signed-off-by: Sage Weil <sage@newdream.net>
