This reverts commit ba9d114ec5.
.. which introduced a regression that prevented all lingering requests
requeued in kick_requests() from ever being sent to the OSDs, resulting
in a lot of missed notifies. In retrospect it's pretty obvious that
r_req_lru_item item in the case of lingering requests can be used not
only for notarget, but also for unsent linkage due to how tightly
actual map and enqueue operations are coupled in __map_request().
The assertion that was being silenced is taken care of in the previous
("libceph: request a new osdmap if lingering request maps to no osd")
commit: by always kicking homeless lingering requests we ensure that
none of them ends up on the notarget list outside of the critical
section guarded by request_mutex.
Cc: stable@vger.kernel.org # 3.18+, needs b049453221 "libceph: request a new osdmap if lingering request maps to no osd"
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
This commit does two things. First, if there are any homeless
lingering requests, we now request a new osdmap even if the osdmap that
is being processed brought no changes, i.e. if a given lingering
request turned homeless in one of the previous epochs and remained
homeless in the current epoch. Not doing so leaves us with a stale
osdmap and as a result we may miss our window for reestablishing the
watch and lose notifies.
MON=1 OSD=1:
# cat linger-needmap.sh
#!/bin/bash
rbd create --size 1 test
DEV=$(rbd map test)
ceph osd out 0
rbd map dne/dne # obtain a new osdmap as a side effect (!)
sleep 1
ceph osd in 0
rbd resize --size 2 test
# rbd info test | grep size -> 2M
# blockdev --getsize $DEV -> 1M
N.B.: Not obtaining a new osdmap in between "osd out" and "osd in"
above is enough to make it miss that resize notify, but that is a
bug^Wlimitation of ceph watch/notify v1.
Second, homeless lingering requests are now kicked just like those
lingering requests whose mapping has changed. This is mainly to
recognize that a homeless lingering request makes no sense and to
preserve the invariant that a registered lingering request is not
sitting on any of r_req_lru_item lists. This spares us a WARN_ON,
which commit ba9d114ec5 ("libceph: clear r_req_lru_item in
__unregister_linger_request()") tried to fix the _wrong_ way.
Cc: stable@vger.kernel.org # 3.10+
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
a255651d4c ("ceph: ensure auth ops are defined before use") made
kfree() in put_osd() conditional on the authorizer. A mechanical
mistake most likely - fix it.
Cc: Alex Elder <elder@linaro.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>
It turns out it's possible to get __remove_osd() called twice on the
same OSD. That doesn't sit well with rb_erase() - depending on the
shape of the tree we can get a NULL dereference, a soft lockup or
a random crash at some point in the future as we end up touching freed
memory. One scenario that I was able to reproduce is as follows:
<osd3 is idle, on the osd lru list>
<con reset - osd3>
con_fault_finish()
osd_reset()
<osdmap - osd3 down>
ceph_osdc_handle_map()
<takes map_sem>
kick_requests()
<takes request_mutex>
reset_changed_osds()
__reset_osd()
__remove_osd()
<releases request_mutex>
<releases map_sem>
<takes map_sem>
<takes request_mutex>
__kick_osd_requests()
__reset_osd()
__remove_osd() <-- !!!
A case can be made that osd refcounting is imperfect and reworking it
would be a proper resolution, but for now Sage and I decided to fix
this by adding a safe guard around __remove_osd().
Fixes: http://tracker.ceph.com/issues/8087
Cc: Sage Weil <sage@redhat.com>
Cc: stable@vger.kernel.org # 3.9+: 7c6e6fc53e: libceph: assert both regular and lingering lists in __remove_osd()
Cc: stable@vger.kernel.org # 3.9+: cc9f1f518c: libceph: change from BUG to WARN for __remove_osd() asserts
Cc: stable@vger.kernel.org # 3.9+
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>
allow specifying position of extent operation in multi-operations
osd request. This is required for cephfs to convert inline data to
normal data (compare xattr, then write object).
Signed-off-by: Yan, Zheng <zyan@redhat.com>
Reviewed-by: Ilya Dryomov <idryomov@redhat.com>
Add CEPH_OSD_OP_CREATE support. Also change libceph to not treat
CEPH_OSD_OP_DELETE as an extent op and add an assert to that end.
Signed-off-by: Yan, Zheng <zyan@redhat.com>
Reviewed-by: Ilya Dryomov <idryomov@redhat.com>
kick_requests() can put linger requests on the notarget list. This
means we need to clear the much-overloaded req->r_req_lru_item in
__unregister_linger_request() as well, or we get an assertion failure
in ceph_osdc_release_request() - !list_empty(&req->r_req_lru_item).
AFAICT the assumption was that registered linger requests cannot be on
any of req->r_req_lru_item lists, but that's clearly not the case.
Signed-off-by: Ilya Dryomov <idryomov@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Requests have to be unlinked from both osd->o_requests (normal
requests) and osd->o_linger_requests (linger requests) lists when
clearing req->r_osd. Otherwise __unregister_linger_request() gets
confused and we trip over a !list_empty(&osd->o_linger_requests)
assert in __remove_osd().
MON=1 OSD=1:
# cat remove-osd.sh
#!/bin/bash
rbd create --size 1 test
DEV=$(rbd map test)
ceph osd out 0
sleep 3
rbd map dne/dne # obtain a new osdmap as a side effect
rbd unmap $DEV & # will block
sleep 3
ceph osd in 0
Signed-off-by: Ilya Dryomov <idryomov@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Bring in missing osd ops and strings, use macros to eliminate multiple
points of maintenance.
Signed-off-by: Ilya Dryomov <idryomov@redhat.com>
Reviewed-by: Sage Weil <sage@redhat.com>
queue_work() doesn't "fail to queue", it returns false if work was
already on a queue, which can't happen here since we allocate
event_work right before we queue it. So don't bother at all.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Use the more common pr_warn.
Other miscellanea:
o Coalesce formats
o Realign arguments
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Both not yet registered (r_linger && list_empty(&r_linger_item)) and
registered linger requests should use the new tid on resend to avoid
the dup op detection logic on the OSDs, yet we were doing this only for
"registered" case. Factor out and simplify the "registered" logic and
use the new helper for "not registered" case as well.
Fixes: http://tracker.ceph.com/issues/8806
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Introduce ceph_osdc_cancel_request() intended for canceling requests
from the higher layers (rbd and cephfs). Because higher layers are in
charge and are supposed to know what and when they are canceling, the
request is not completed, only unref'ed and removed from the libceph
data structures.
__cancel_request() is no longer called before __unregister_request(),
because __unregister_request() unconditionally revokes r_request and
there is no point in trying to do it twice.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
We should check if request is on the linger request list of any of the
OSDs, not whether request is registered or not.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Linger requests that have not yet been registered should not be
unregistered by __unregister_linger_request(). This messes up ref
count and leads to use-after-free.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
It is important that both regular and lingering requests lists are
empty when the OSD is removed.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Add some WARN_ONs to alert us when we try to destroy requests that are
still registered.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Add dout()s to ceph_osdc_request_{get,put}(). Also move them to .c and
turn kref release callback into a static function.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Abstract out __move_osd_to_lru() logic from __unregister_request() and
__unregister_linger_request().
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
So that:
req->r_osd_item --> osd->o_requests list
req->r_linger_osd_item --> osd->o_linger_requests list
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Sparse complained about this bogus extern on definition of
a function.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull Ceph updates from Sage Weil:
"The biggest chunk is a series of patches from Ilya that add support
for new Ceph osd and crush map features, including some new tunables,
primary affinity, and the new encoding that is needed for erasure
coding support. This brings things into parity with the server side
and the looming firefly release. There is also support for allocation
hints in RBD that help limit fragmentation on the server side.
There is also a series of patches from Zheng fixing NFS reexport,
directory fragmentation support, flock vs fnctl behavior, and some
issues with clustered MDS.
Finally, there are some miscellaneous fixes from Yunchuan Wen for
fscache, Fabian Frederick for ACLs, and from me for fsync(dirfd)
behavior"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client: (79 commits)
ceph: skip invalid dentry during dcache readdir
libceph: dump pool {read,write}_tier to debugfs
libceph: output primary affinity values on osdmap updates
ceph: flush cap release queue when trimming session caps
ceph: don't grabs open file reference for aborted request
ceph: drop extra open file reference in ceph_atomic_open()
ceph: preallocate buffer for readdir reply
libceph: enable PRIMARY_AFFINITY feature bit
libceph: redo ceph_calc_pg_primary() in terms of ceph_calc_pg_acting()
libceph: add support for osd primary affinity
libceph: add support for primary_temp mappings
libceph: return primary from ceph_calc_pg_acting()
libceph: switch ceph_calc_pg_acting() to new helpers
libceph: introduce apply_temps() helper
libceph: introduce pg_to_raw_osds() and raw_to_up_osds() helpers
libceph: ceph_can_shift_osds(pool) and pool type defines
libceph: ceph_osd_{exists,is_up,is_down}(osd) definitions
libceph: enable OSDMAP_ENC feature bit
libceph: primary_affinity decode bits
libceph: primary_affinity infrastructure
...
In preparation for adding support for primary_temp, stop assuming
primaryness: add a primary out parameter to ceph_calc_pg_acting() and
change call sites accordingly. Primary is now specified separately
from the order of osds in the set.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Split osdmap allocation and initialization into a separate function,
ceph_osdmap_decode().
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
This is primarily for rbd's benefit and is supposed to combat
fragmentation:
"... knowing that rbd images have a 4m size, librbd can pass a hint
that will let the osd do the xfs allocation size ioctl on new files so
that they are allocated in 1m or 4m chunks. We've seen cases where
users with rbd workloads have very high levels of fragmentation in xfs
and this would mitigate that and probably have a pretty nice
performance benefit."
SETALLOCHINT is considered advisory, so our backwards compatibility
mechanism here is to set FAILOK flag for all SETALLOCHINT ops.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Encode ceph_osd_op::flags field so that it gets sent over the wire.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
One of my pet coding style peeves is the practice of
adding extra return; at the end of function.
Kill several instances of this in network code.
I suppose some coccinelle wizardy could do this automatically.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Handling redirect replies requires both map_sem and request_mutex.
Taking map_sem unconditionally near the top of handle_reply() avoids
possible race conditions that arise from releasing request_mutex to be
able to acquire map_sem in redirect reply case. (Lock ordering is:
map_sem, request_mutex, crush_mutex.)
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Factor out logic from ceph_osdc_start_request() into a new helper,
__ceph_osdc_start_request(). ceph_osdc_start_request() now amounts to
taking locks and calling __ceph_osdc_start_request().
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
msgpool_op_reply message pool isn't destroyed if workqueue construction
fails. Fix it.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Follow redirect replies from osds, for details see ceph.git commit
fbbe3ad1220799b7bb00ea30fce581c5eadaf034.
v1 (current) version of redirect reply consists of oloc and oid, which
expands to pool, key, nspace, hash and oid. However, server-side code
that would populate anything other than pool doesn't exist yet, and
hence this commit adds support for pool redirects only. To make sure
that future server-side updates don't break us, we decode all fields
and, if any of key, nspace, hash or oid have a non-default value, error
out with "corrupt osd_op_reply ..." message.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Rename ceph_osd_request::r_{oloc,oid} to r_base_{oloc,oid} before
introducing r_target_{oloc,oid} needed for redirects.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Overwrite ceph_osd_request::r_oloc.pool with read_tier for read ops and
write_tier for write and read+write ops (aka basic tiering support).
{read,write}_tier are part of pg_pool_t since v9. This commit bumps
our pg_pool_t decode compat version from v7 to v9, all new fields
except for {read,write}_tier are ignored.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Switch ceph_calc_ceph_pg() to new oloc and oid abstractions and rename
it to ceph_oloc_oid_to_pg() to make its purpose more clear.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
In preparation for tiering support, which would require having two
(base and target) object names for each osd request and also copying
those names around, introduce struct ceph_object_id (oid) and a couple
helpers to facilitate those copies and encapsulate the fact that object
name is not necessarily a NUL-terminated string.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
In preparation for adding oid abstraction, rename MAX_OBJ_NAME_SIZE to
CEPH_MAX_OID_NAME_LEN.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Instead of relying on pool fields in ceph_file_layout (for mapping) and
ceph_pg (for enconding), start using ceph_object_locator (oloc)
abstraction. Note that userspace oloc currently consists of pool, key,
nspace and hash fields, while this one contains only a pool. This is
OK, because at this point we only send (i.e. encode) olocs and never
have to receive (i.e. decode) them.
This makes keeping a copy of ceph_file_layout in every osd request
unnecessary, so ceph_osd_request::r_file_layout field is nuked.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The check that makes sure that we have enough memory allocated to read
in the entire header of the message in question is currently busted.
It compares front_len of the incoming message with iov_len field of
ceph_msg::front structure, which is used primarily to indicate the
amount of data already read in, and not the size of the allocated
buffer. Under certain conditions (e.g. a short read from a socket
followed by that socket's shutdown and owning ceph_connection reset)
this results in a warning similar to
[85688.975866] libceph: get_reply front 198 > preallocated 122 (4#0)
and, through another bug, leads to forever hung tasks and forced
reboots. Fix this by comparing front_len with front_alloc_len field of
struct ceph_msg, which stores the actual size of the buffer.
Fixes: http://tracker.ceph.com/issues/5425
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Rename front local variable to front_len in get_reply() to make its
purpose more clear.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
With the current full handling, there is a race between osds and
clients getting the first map marked full. If the osd wins, it will
return -ENOSPC to any writes, but the client may already have writes
in flight. This results in the client getting the error and
propagating it up the stack. For rbd, the block layer turns this into
EIO, which can cause corruption in filesystems above it.
To avoid this race, osds are being changed to drop writes that came
from clients with an osdmap older than the last osdmap marked full.
In order for this to work, clients must resend all writes after they
encounter a full -> not full transition in the osdmap. osds will wait
for an updated map instead of processing a request from a client with
a newer map, so resent writes will not be dropped by the osd unless
there is another not full -> full transition.
This approach requires both osds and clients to be fixed to avoid the
race. Old clients talking to osds with this fix may hang instead of
returning EIO and potentially corrupting an fs. New clients talking to
old osds have the same behavior as before if they encounter this race.
Fixes: http://tracker.ceph.com/issues/6938
Reviewed-by: Sage Weil <sage@inktank.com>
Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
The PAUSEWR and PAUSERD flags are meant to stop the cluster from
processing writes and reads, respectively. The FULL flag is set when
the cluster determines that it is out of space, and will no longer
process writes. PAUSEWR and PAUSERD are purely client-side settings
already implemented in userspace clients. The osd does nothing special
with these flags.
When the FULL flag is set, however, the osd responds to all writes
with -ENOSPC. For cephfs, this makes sense, but for rbd the block
layer translates this into EIO. If a cluster goes from full to
non-full quickly, a filesystem on top of rbd will not behave well,
since some writes succeed while others get EIO.
Fix this by blocking any writes when the FULL flag is set in the osd
client. This is the same strategy used by userspace, so apply it by
default. A follow-on patch makes this configurable.
__map_request() is called to re-target osd requests in case the
available osds changed. Add a paused field to a ceph_osd_request, and
set it whenever an appropriate osd map flag is set. Avoid queueing
paused requests in __map_request(), but force them to be resent if
they become unpaused.
Also subscribe to the next osd map from the monitor if any of these
flags are set, so paused requests can be unblocked as soon as
possible.
Fixes: http://tracker.ceph.com/issues/6079
Reviewed-by: Sage Weil <sage@inktank.com>
Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
Wake up possible waiters, invoke the call back if any, unregister the request
Signed-off-by: Li Wang <liwang@ubuntukylin.com>
Signed-off-by: Yunchuan Wen <yunchuanwen@ubuntukylin.com>
Signed-off-by: Sage Weil <sage@inktank.com>
Without a way to flush the osd client's notify workqueue, a watch
event that is unregistered could continue receiving callbacks
indefinitely.
Unregistering the event simply means no new notifies are added to the
queue, but there may still be events in the queue that will call the
watch callback for the event. If the queue is flushed after the event
is unregistered, the caller can be sure no more watch callbacks will
occur for the canceled watch.
Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
create_singlethread_workqueue() returns NULL on error, and it doesn't
return ERR_PTRs.
I tweaked the error handling a little to be consistent with earlier in
the function.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Sage Weil <sage@inktank.com>