Let raid5d handle stripe in batch way to reduce conf->device_lock locking.
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: NeilBrown <neilb@suse.de>

make_request() does stripe release for every stripe and the stripe usually has
count 1, which makes previous release_stripe() optimization not work. In my
test, this release_stripe() becomes the heaviest pleace to take
conf->device_lock after previous patches applied.

Below patch makes stripe release batch. All the stripes will be released in
unplug. The STRIPE_ON_UNPLUG_LIST bit is to protect concurrent access stripe
lru.
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: NeilBrown <neilb@suse.de>

This will allow md/raid to know why the unplug was called,
and will be able to act according - if !from_schedule it
is safe to perform tasks which could themselves schedule.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

MD raid1 prepares to dispatch request in unplug callback. If make_request in
low level queue also uses unplug callback to dispatch request, the low level
queue's unplug callback will not be called. Recheck the callback list helps
this case.
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Both md and umem has similar code for getting notified on an
blk_finish_plug event.
Centralize this code in block/ and allow each driver to
provide its distinctive difference.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

This seemed like a good idea at the time, but after further thought I
cannot see it making a difference other than very occasionally and
testing to try to exercise the case it is most likely to help did not
show any performance difference by removing it.

So remove the counting of active plugs and allow 'pending writes' to
be activated at any time, not just when no plugs are active.

This is only relevant when there is a write-intent bitmap, and the
updating of the bitmap will likely introduce enough delay that
the single-threading of bitmap updates will be enough to collect large
numbers of updates together.

Removing this will make it easier to centralise the unplug code, and
will clear the other for other unplug enhancements which have a
measurable effect.
Signed-off-by: NeilBrown <neilb@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Add in-flight cmds to the tail. That way while searching
(during request completion),we will always get a hit on the
first element.
Signed-off-by: Chetan Loke <loke.chetan@gmail.com>
Acked-by: Paul.Clements@steeleye.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Unconditionally announce FLUSH/FUA to upper layers.
If the lower layers on either node do not actually support this,
generic_make_request() will deal with it.

If this causes performance regressions on your setup,
make sure there are no volatile caches involved,
and mount -o nobarrier or equivalent.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

We capped our max_bio_size respectively max_hw_sectors with
min_t(int, lower level limit, our limit);
unfortunately, some drivers, e.g. the kvm virtio block driver, initialize their
limits to "-1U", and that is of course a smaller "int" value than our limit.

Impact: we started to request 16 MB resync requests,
which lead to protocol error and a reconnect loop.

Fix all relevant constants and parameters to be unsigned int.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

If you do back to back wait-sync/invalidate on a Primary in a tight loop,
during application IO load, you could trigger a race:
  kernel: block drbd6: FIXME going to queue 'set_n_write from StartingSync'
	but 'write from resync_finished' still pending?

Fix this by changing the order of the drbd_queue_work() and
the wake_up() in dec_ap_pending(), and adding the additional
drbd_flush_workqueue() before requesting the full sync.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

Occasionally, if we disconnect, we triggered this assert:
  block drbd7: ASSERT FAILED tl_hash[27] == c30b0f04, expected NULL

hlist_del() happens only on master bio completion.

We used to wait for pending IO to complete before freeing tl_hash
on disconnect. We no longer do so, since we learned to "freeze"
IO on disconnect.

If the local disk is too slow, we may reach C_STANDALONE early,
and there are still some requests pending locally when we call
drbd_free_tl_hash().

If we now free the tl_hash, and later the local IO completion completes
the master bio, which then does hlist_del() and clobbers freed memory.

Do hlist_del_init() and hlist_add_fake() before kfree(tl_hash),
so the hlist_del() on master bio completion is harmless.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

In case we want to hard-reset from the local-io-error handler,
we need to call it before notifying the peer or aborting local IO.
Otherwise the peer will advance its data generation UUIDs even
if secondary.

This way, local io error looks like a "regular" node crash,
which reduces the number of different failure cases.
This may be useful in a bigger picture where crashed or otherwise
"misbehaving" nodes are automatically re-deployed.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

Fix asserts like
  block drbd0: in got_BlockAck:4634: rs_pending_cnt = -35 < 0 !

We reset the resync lru cache and related information (rs_pending_cnt),
once we successfully finished a resync or online verify, or if the
replication connection is lost.

We also need to reset it if a resync or online verify is aborted
because a lower level disk failed.

In that case the replication link is still established,
and we may still have packets queued in the network buffers
which want to touch rs_pending_cnt.

We do not have any synchronization mechanism to know for sure when all
such pending resync related packets have been drained.

To avoid this counter to go negative (and violate the ASSERT that it
will always be >= 0), just do not reset it when we lose a disk.

It is good enough to make sure it is re-initialized before the next
resync can start: reset it when we re-attach a disk.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

We cache the congestion status in mdev->congestion_reason whenever
drbd_congested() was called.
Reset this cached info before reporting it when reading /proc/drbd.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

If the drbd worker thread is synchronously waiting for some userland
callback, we don't want some casual pageout to block on us.
Have drbd_congested() report congestion in that case.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

Aborting local requests (not waiting for completion from the lower level
disk) is dangerous: if the master bio has been completed to upper
layers, data pages may be re-used for other things already.
If local IO is still pending and later completes,
this may cause crashes or corrupt unrelated data.

Only abort local IO if explicitly requested.
Intended use case is a lower level device that turned into a tarpit,
not completing io requests, not even doing error completion.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

Merge branch 'upstream' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/floppy into for-3.6/drivers

The SYSTEM_SUSPEND_DISK system state is never used, so drop it.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Merge emailed kgdb dmesg fixups patches from Anton Vorontsov:
 "The dmesg command appears to be broken after the printk rework.  The
  old logic in the kdb code makes no sense in terms of current
  printk/logging storage format, and KDB simply hangs forever upon
  entering 'dmesg' command.

  The first patch revives the command by switching to kmsg_dumper
  iterator.  As a side-effect, the code is now much more simpler.

  A few changes were needed in the printk.c: we needed unlocked variant
  of the kmsg_dumper iterator, but these can surely wait for 3.6.

  It's probably too late even for the first patch to go to 3.5, but I'll
  try to convince otherwise.  :-) Here we go:

   - The current code is broken for sure, and has no hope to work at
     all.  It is a regression
   - The new code works for me, and probably works for everyone else;
   - If it compiles (and I urge everyone to compile-test it on your
     setup), it hardly can make things worse."

* Merge emailed patches from Anton Vorontsov: (4 commits)
  kdb: Switch to nolock variants of kmsg_dump functions
  printk: Implement some unlocked kmsg_dump functions
  printk: Remove kdb_syslog_data
  kdb: Revive dmesg command

The locked variants are prone to deadlocks (suppose we got to the
debugger w/ the logbuf lock held), so let's switch to nolock variants.
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

If used from KDB, the locked variants are prone to deadlocks (suppose we
got to the debugger w/ the logbuf lock held).

So, we have to implement a few routines that grab no logbuf lock.

Yet we don't need these functions in modules, so we don't export them.
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

The function is no longer needed, so remove it.
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

The kgdb dmesg command is broken after the printk rework.  The old logic
in kdb code makes no sense in terms of current printk/logging storage
format, and KDB simply hangs forever.

This patch revives the command by switching to kmsg_dumper iterator.

The code is now much more simpler and shorter.
Signed-off-by: Anton Vorontsov <anton.vorontsov@linaro.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Pull late MIPS fixes from Ralf Baechle:
 "This fixes a number of lose ends in the MIPS code and various bug
  fixes.

  Aside of dropping some patch that should not be in this pull request
  everything has sat in -next for quite a while and there are no known
  issues.

  The biggest patch in this patch set moves the allocation of an array
  that is aliased to a function (for runtime generated code) to
  assembler code.  This avoids an issue with certain toolchains when
  building for microMIPS."

* 'upstream' of git://git.linux-mips.org/pub/scm/ralf/upstream-linus: (35 commits)
  MIPS: PCI: Move fixups from __init to __devinit.
  MIPS: Fix bug.h MIPS build regression
  MIPS: sync-r4k: remove redundant irq operation
  MIPS: smp: Warn on too early irq enable
  MIPS: call set_cpu_online() on cpu being brought up with irq disabled
  MIPS: call ->smp_finish() a little late
  MIPS: Yosemite: delay irq enable to ->smp_finish()
  MIPS: SMTC: delay irq enable to ->smp_finish()
  MIPS: BMIPS: delay irq enable to ->smp_finish()
  MIPS: Octeon: delay enable irq to ->smp_finish()
  MIPS: Oprofile: Fix build as a module.
  MIPS: BCM63XX: Fix BCM6368 IPSec clock bit
  MIPS: perf: Fix build error caused by unused counters_per_cpu_to_total()
  MIPS: Fix Magic SysRq L kernel crash.
  MIPS: BMIPS: Fix duplicate header inclusion.
  mips: mark const init data with __initconst instead of __initdata
  MIPS: cmpxchg.h: Add missing include
  MIPS: Malta may also be equipped with MIPS64 R2 processors.
  MIPS: Fix typo multipy -> multiply
  MIPS: Cavium: Fix duplicate ARCH_SPARSEMEM_ENABLE in kconfig.
  ...

Pull device-mapper discard fixes from Alasdair G Kergon:
  - avoid a crash in dm-raid1 when discards coincide with mirror
    recovery;
  - avoid discarding shared data that's still needed in dm-thin;
  - don't guarantee that discarded blocks will be wiped in dm-raid1.

* tag 'dm-3.5-fixes-2' of git://git.kernel.org/pub/scm/linux/kernel/git/agk/linux-dm:
  dm raid1: set discard_zeroes_data_unsupported
  dm thin: do not send discards to shared blocks
  dm raid1: fix crash with mirror recovery and discard

Pull pnfs/ore fixes from Boaz Harrosh:
 "These are catastrophic fixes to the pnfs objects-layout that were just
  discovered.  They are also destined for @stable.

  I have found these and worked on them at around RC1 time but
  unfortunately went to the hospital for kidney stones and had a very
  slow recovery.  I refrained from sending them as is, before proper
  testing, and surly I have found a bug just yesterday.

  So now they are all well tested, and have my sign-off.  Other then
  fixing the problem at hand, and assuming there are no bugs at the new
  code, there is low risk to any surrounding code.  And in anyway they
  affect only these paths that are now broken.  That is RAID5 in pnfs
  objects-layout code.  It does also affect exofs (which was not broken)
  but I have tested exofs and it is lower priority then objects-layout
  because no one is using exofs, but objects-layout has lots of users."

* 'for-linus' of git://git.open-osd.org/linux-open-osd:
  pnfs-obj: Fix __r4w_get_page when offset is beyond i_size
  pnfs-obj: don't leak objio_state if ore_write/read fails
  ore: Unlock r4w pages in exact reverse order of locking
  ore: Remove support of partial IO request (NFS crash)
  ore: Fix NFS crash by supporting any unaligned RAID IO

Pull UBIFS free space fix-up bugfix from Artem Bityutskiy:
 "It's been reported already twice recently:

    http://lists.infradead.org/pipermail/linux-mtd/2012-May/041408.html
    http://lists.infradead.org/pipermail/linux-mtd/2012-June/042422.html

  and we finally have the fix.  I am quite confident the fix is correct
  because I could reproduce the problem with nandsim and verify the fix.
  It was also verified by Iwo (the reporter).

  I am also confident that this is OK to merge the fix so late because
  this patch affects only the fixup functionality, which is not used by
  most users."

* tag 'upstream-3.5-rc8' of git://git.infradead.org/linux-ubifs:
  UBIFS: fix a bug in empty space fix-up

We can't guarantee that REQ_DISCARD on dm-mirror zeroes the data even if
the underlying disks support zero on discard.  So this patch sets
ti->discard_zeroes_data_unsupported.

For example, if the mirror is in the process of resynchronizing, it may
happen that kcopyd reads a piece of data, then discard is sent on the
same area and then kcopyd writes the piece of data to another leg.
Consequently, the data is not zeroed.

The flag was made available by commit 983c7db3
(dm crypt: always disable discard_zeroes_data).
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@kernel.org
Signed-off-by: Alasdair G Kergon <agk@redhat.com>

When process_discard receives a partial discard that doesn't cover a
full block, it sends this discard down to that block. Unfortunately, the
block can be shared and the discard would corrupt the other snapshots
sharing this block.

This patch detects block sharing and ends the discard with success when
sending it to the shared block.

The above change means that if the device supports discard it can't be
guaranteed that a discard request zeroes data. Therefore, we set
ti->discard_zeroes_data_unsupported.

Thin target discard support with this bug arrived in commit
104655fd (dm thin: support discards).
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@kernel.org
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>

This patch fixes a crash when a discard request is sent during mirror
recovery.

Firstly, some background.  Generally, the following sequence happens during
mirror synchronization:
- function do_recovery is called
- do_recovery calls dm_rh_recovery_prepare
- dm_rh_recovery_prepare uses a semaphore to limit the number
  simultaneously recovered regions (by default the semaphore value is 1,
  so only one region at a time is recovered)
- dm_rh_recovery_prepare calls __rh_recovery_prepare,
  __rh_recovery_prepare asks the log driver for the next region to
  recover. Then, it sets the region state to DM_RH_RECOVERING. If there
  are no pending I/Os on this region, the region is added to
  quiesced_regions list. If there are pending I/Os, the region is not
  added to any list. It is added to the quiesced_regions list later (by
  dm_rh_dec function) when all I/Os finish.
- when the region is on quiesced_regions list, there are no I/Os in
  flight on this region. The region is popped from the list in
  dm_rh_recovery_start function. Then, a kcopyd job is started in the
  recover function.
- when the kcopyd job finishes, recovery_complete is called. It calls
  dm_rh_recovery_end. dm_rh_recovery_end adds the region to
  recovered_regions or failed_recovered_regions list (depending on
  whether the copy operation was successful or not).

The above mechanism assumes that if the region is in DM_RH_RECOVERING
state, no new I/Os are started on this region. When I/O is started,
dm_rh_inc_pending is called, which increases reg->pending count. When
I/O is finished, dm_rh_dec is called. It decreases reg->pending count.
If the count is zero and the region was in DM_RH_RECOVERING state,
dm_rh_dec adds it to the quiesced_regions list.

Consequently, if we call dm_rh_inc_pending/dm_rh_dec while the region is
in DM_RH_RECOVERING state, it could be added to quiesced_regions list
multiple times or it could be added to this list when kcopyd is copying
data (it is assumed that the region is not on any list while kcopyd does
its jobs). This results in memory corruption and crash.

There already exist bypasses for REQ_FLUSH requests: REQ_FLUSH requests
do not belong to any region, so they are always added to the sync list
in do_writes. dm_rh_inc_pending does not increase count for REQ_FLUSH
requests. In mirror_end_io, dm_rh_dec is never called for REQ_FLUSH
requests. These bypasses avoid the crash possibility described above.

These bypasses were improperly implemented for REQ_DISCARD when
the mirror target gained discard support in commit
5fc2ffea (dm raid1: support discard).

In do_writes, REQ_DISCARD requests is always added to the sync queue and
immediately dispatched (even if the region is in DM_RH_RECOVERING).  However,
dm_rh_inc and dm_rh_dec is called for REQ_DISCARD resusts.  So it violates the
rule that no I/Os are started on DM_RH_RECOVERING regions, and causes the list
corruption described above.

This patch changes it so that REQ_DISCARD requests follow the same path
as REQ_FLUSH. This avoids the crash.

Reference: https://bugzilla.redhat.com/837607Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@kernel.org
Signed-off-by: Alasdair G Kergon <agk@redhat.com>

It is very common for the end of the file to be unaligned on
stripe size. But since we know it's beyond file's end then
the XOR should be preformed with all zeros.

Old code used to just read zeros out of the OSD devices, which is a great
waist. But what scares me more about this situation is that, we now have
pages attached to the file's mapping that are beyond i_size. I don't
like the kind of bugs this calls for.

Fix both birds, by returning a global zero_page, if offset is beyond
i_size.

TODO:
	Change the API to ->__r4w_get_page() so a NULL can be
	returned without being considered as error, since XOR API
	treats NULL entries as zero_pages.

[Bug since 3.2. Should apply the same way to all Kernels since]
CC: Stable Tree <stable@kernel.org>
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>

[Bug since 3.2 Kernel]
CC: Stable Tree <stable@kernel.org>
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>

The read-4-write pages are locked in address ascending order.
But where unlocked in a way easiest for coding. Fix that,
locks should be released in opposite order of locking, .i.e
descending address order.

I have not hit this dead-lock. It was found by inspecting the
dbug print-outs. I suspect there is an higher lock at caller that
protects us, but fix it regardless.
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>

Do to OOM situations the ore might fail to allocate all resources
needed for IO of the full request. If some progress was possible
it would proceed with a partial/short request, for the sake of
forward progress.

Since this crashes NFS-core and exofs is just fine without it just
remove this contraption, and fail.

TODO:
	Support real forward progress with some reserved allocations
	of resources, such as mem pools and/or bio_sets

[Bug since 3.2 Kernel]
CC: Stable Tree <stable@kernel.org>
CC: Benny Halevy <bhalevy@tonian.com>
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>

In RAID_5/6 We used to not permit an IO that it's end
byte is not stripe_size aligned and spans more than one stripe.
.i.e the caller must check if after submission the actual
transferred bytes is shorter, and would need to resubmit
a new IO with the remainder.

Exofs supports this, and NFS was supposed to support this
as well with it's short write mechanism. But late testing has
exposed a CRASH when this is used with none-RPC layout-drivers.

The change at NFS is deep and risky, in it's place the fix
at ORE to lift the limitation is actually clean and simple.
So here it is below.

The principal here is that in the case of unaligned IO on
both ends, beginning and end, we will send two read requests
one like old code, before the calculation of the first stripe,
and also a new site, before the calculation of the last stripe.
If any "boundary" is aligned or the complete IO is within a single
stripe. we do a single read like before.

The code is clean and simple by splitting the old _read_4_write
into 3 even parts:
1._read_4_write_first_stripe
2. _read_4_write_last_stripe
3. _read_4_write_execute

And calling 1+3 at the same place as before. 2+3 before last
stripe, and in the case of all in a single stripe then 1+2+3
is preformed additively.

Why did I not think of it before. Well I had a strike of
genius because I have stared at this code for 2 years, and did
not find this simple solution, til today. Not that I did not try.

This solution is much better for NFS than the previous supposedly
solution because the short write was dealt  with out-of-band after
IO_done, which would cause for a seeky IO pattern where as in here
we execute in order. At both solutions we do 2 separate reads, only
here we do it within a single IO request. (And actually combine two
writes into a single submission)

NFS/exofs code need not change since the ORE API communicates the new
shorter length on return, what will happen is that this case would not
occur anymore.

hurray!!

[Stable this is an NFS bug since 3.2 Kernel should apply cleanly]
CC: Stable Tree <stable@kernel.org>
Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>

UBIFS has a feature called "empty space fix-up" which is a quirk to work-around
limitations of dumb flasher programs. Namely, of those flashers that are unable
to skip NAND pages full of 0xFFs while flashing, resulting in empty space at
the end of half-filled eraseblocks to be unusable for UBIFS. This feature is
relatively new (introduced in v3.0).

The fix-up routine (fixup_free_space()) is executed only once at the very first
mount if the superblock has the 'space_fixup' flag set (can be done with -F
option of mkfs.ubifs). It basically reads all the UBIFS data and metadata and
writes it back to the same LEB. The routine assumes the image is pristine and
does not have anything in the journal.

There was a bug in 'fixup_free_space()' where it fixed up the log incorrectly.
All but one LEB of the log of a pristine file-system are empty. And one
contains just a commit start node. And 'fixup_free_space()' just unmapped this
LEB, which resulted in wiping the commit start node. As a result, some users
were unable to mount the file-system next time with the following symptom:

UBIFS error (pid 1): replay_log_leb: first log node at LEB 3:0 is not CS node
UBIFS error (pid 1): replay_log_leb: log error detected while replaying the log at LEB 3:0

The root-cause of this bug was that 'fixup_free_space()' wrongly assumed
that the beginning of empty space in the log head (c->lhead_offs) was known
on mount. However, it is not the case - it was always 0. UBIFS does not store
in it the master node and finds out by scanning the log on every mount.

The fix is simple - just pass commit start node size instead of 0 to
'fixup_leb()'.
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@linux.intel.com>
Cc: stable@vger.kernel.org [v3.0+]
Reported-by: Iwo Mergler <Iwo.Mergler@netcommwireless.com>
Tested-by: Iwo Mergler <Iwo.Mergler@netcommwireless.com>
Reported-by: James Nute <newten82@gmail.com>

Pull last minute Ceph fixes from Sage Weil:
 "The important one fixes a bug in the socket failure handling behavior
  that was turned up in some recent failure injection testing.  The
  other two are minor bug fixes."

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client:
  rbd: endian bug in rbd_req_cb()
  rbd: Fix ceph_snap_context size calculation
  libceph: fix messenger retry