The existing error handling scheme requires resetting err to -EINVAL
prior to calling any ceph_decode_* macro. This is ugly and fragile,
and there already are a few places where we would return 0 on error,
due to a missing reset. Fix this by adding a special e_inval label to
be used by all ceph_decode_* macros.
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>
Dump osdmap in hex on both full and incremental decode errors, to make
it easier to match the contents with error offset. dout() map epoch
and max_osd value on success.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
To save screen space in anticipation of more fields (e.g. primary
affinity).
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Add TUNABLES3 feature (chooseleaf_vary_r tunable) to a set of features
supported by default.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
This lets you adjust the vary_r tunable on a per-rule basis.
Reflects ceph.git commit f944ccc20aee60a7d8da7e405ec75ad1cd449fac.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
The current crush_choose_firstn code will re-use the same 'r' value for
the recursive call. That means that if we are hitting a collision or
rejection for some reason (say, an OSD that is marked out) and need to
retry, we will keep making the same (bad) choice in that recursive
selection.
Introduce a tunable that fixes that behavior by incorporating the parent
'r' value into the recursive starting point, so that a different path
will be taken in subsequent placement attempts.
Note that this was done from the get-go for the new crush_choose_indep
algorithm.
This was exposed by a user who was seeing PGs stuck in active+remapped
after reweight-by-utilization because the up set mapped to a single OSD.
Reflects ceph.git commit a8e6c9fbf88bad056dd05d3eb790e98a5e43451a.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
These two fields are misnomers; they are *retry* counts.
Reflects ceph.git commit f17caba8ae0cad7b6f8f35e53e5f73b444696835.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
Back in 27f4d1f6bc32c2ed7b2c5080cbd58b14df622607 we refactored the CRUSH
code to allow adjustment of the retry counts on a per-pool basis. That
commit had an off-by-one bug: the previous "tries" counter was a *retry*
count, not a *try* count, but the new code was passing in 1 meaning
there should be no retries.
Fix the ftotal vs tries comparison to use < instead of <= to fix the
problem. Note that the original code used <= here, which means the
global "choose_total_tries" tunable is actually counting retries.
Compensate for that by adding 1 in crush_do_rule when we pull the tunable
into the local variable.
This was noticed looking at output from a user provided osdmap.
Unfortunately the map doesn't illustrate the change in mapping behavior
and I haven't managed to construct one yet that does. Inspection of the
crush debug output now aligns with prior versions, though.
Reflects ceph.git commit 795704fd615f0b008dcc81aa088a859b2d075138.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
If buffer size is zero, return the size of layout vxattr. If buffer
size is not zero, check if it is large enough for layout vxattr.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
send_mds_reconnect() may call discard_cap_releases() after all
release messages have been dropped by cleanup_cap_releases()
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
When there is no more data, ceph_msg_data_{pages,pagelist}_advance()
should not move on to the next page.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
When adjusting caps client wants, MDS does not record caps that are
not allowed. For non-auth MDS, it does not record WR caps. So when
a MDS reply changes a non-auth cap to auth cap, client needs to set
cap's mds_wanted according to the reply.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
flock and posix lock should use fl->fl_file instead of process ID
as owner identifier. (posix lock uses fl->fl_owner. fl->fl_owner
is usually equal to fl->fl_file, but it also can be a customized
value). The process ID of who holds the lock is just for F_GETLK
fcntl(2).
The fix is rename the 'pid' fields of struct ceph_mds_request_args
and struct ceph_filelock to 'owner', rename 'pid_namespace' fields
to 'pid'. Assign fl->fl_file to the 'owner' field of lock messages.
We also set the most significant bit of the 'owner' field. MDS can
use that bit to distinguish between old and new clients.
The MDS counterpart of this patch modifies the flock code to not
take the 'pid_namespace' into consideration when checking conflict
locks.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
VFS does not directly pass flock's operation code to filesystem's
flock callback. It translates the operation code to the form how
posix lock's parameters are presented.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
handle following sequence of events:
- client releases a inode with i_max_size > 0. The release message
is queued. (is not sent to the auth MDS)
- a 'lookup' request reply from non-auth MDS returns the same inode.
- client opens the inode in write mode. The version of inode trace
in 'open' request reply is equal to the cached inode's version.
- client requests new max size. The MDS ignores the request because
it does not affect client's write range
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Only auth MDS can issue write caps to clients, so don't consider
write caps registered with non-auth MDS as valid.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Use the newly introduced LOOKUPNAME MDS request to connect child
inode to its parent directory.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
ceph_fh_to_parent() returns dentry that corresponds to the 'ino' field
of struct ceph_nfs_confh. This is wrong, it should return dentry that
corresponds to the 'parent_ino' field.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
MDS handles LOOKUPHASH and LOOKUPINO MDS requests in the same way.
So __cfh_to_dentry() is redundant.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Sage Weil <sage@inktank.com>
The object store limit needs to be updated after writing,
and this can be done provided the corresponding object has already
been initialized. Current object initialization is done asynchrously,
which introduce a race if a file is opened, then immediately followed
by a writing, the initialization may have not completed, the code will
reach the ASSERT in fscache_submit_exclusive_op() to cause kernel
bug.
Tested-by: Milosz Tanski <milosz@adfin.com>
Signed-off-by: Yunchuan Wen <yunchuanwen@ubuntukylin.com>
Signed-off-by: Min Chen <minchen@ubuntukylin.com>
Signed-off-by: Li Wang <liwang@ubuntukylin.com>
Synchronize object->store_limit[_l] with new inode->i_size after file writing.
Tested-by: Milosz Tanski <milosz@adfin.com>
Signed-off-by: Yunchuan Wen <yunchuanwen@ubuntukylin.com>
Signed-off-by: Min Chen <minchen@ubuntukylin.com>
Signed-off-by: Li Wang <liwang@ubuntukylin.com>
Add an interface to explicitly synchronize object->store_limit[_l]
with inode->i_size
Tested-by: Milosz Tanski <milosz@adfin.com>
Signed-off-by: Yunchuan Wen <yunchuanwen@ubuntukylin.com>
Signed-off-by: Min Chen <minchen@ubuntukylin.com>
Signed-off-by: Li Wang <liwang@ubuntukylin.com>
This is racy--we do not know whather d_parent has changed out from
underneath us because i_mutex is not held on the source inode's directory.
Also, taking this reference is useless.
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yan, Zheng <zheng.z.yan@intel.com>
Do not assume that r_old_dentry implies that r_old_dentry_dir is also
true. Separate out the ref cleanup and make the debugs dump behave when
it is NULL.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yan, Zheng <zheng.z.yan@intel.com>
The fsync(dirfd) only covers namespace operations, not inode updates.
We do not need to cover setattr variants or O_TRUNC.
Reported-by: Al Viro <viro@xeniv.linux.org.uk>
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yan, Zheng <zheng.z.yan@intel.com>
This is just old_dir; no reason to abuse the dcache pointers.
Reported-by: Al Viro <viro.zeniv.linux.org.uk>
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yan, Zheng <zheng.z.yan@intel.com>
If readdir 'frag' is adjusted, readdir 'offset' should be reset.
Otherwise some dentries may be lost when readdir and fragmenting
directory happen at the some.
Another way to fix this issue is let MDS adjust readdir 'frag'.
The code that handles MDS reply reset the readdir 'offset' if
the readdir reply is different than the requested one.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
When changing readdir postion, fi->next_offset should be set to 0
if the new postion is not in the first dirfrag.
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Comparing offset with inode->i_sb->s_maxbytes doesn't make sense for
directory. For a fragmented directory, offset (frag_t, off) can be
larger than inode->i_sb->s_maxbytes.
At the very beginning of ceph_dir_llseek(), local variable old_offset
is initialized to parameter offset. This doesn't make sense neither.
Old_offset should be ceph_make_fpos(fi->frag, fi->next_offset).
Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Reviewed-by: Alex Elder <elder@linaro.org>
In an effort to reduce fragmentation, prefix every rbd write with
a CEPH_OSD_OP_SETALLOCHINT osd op with an expected_write_size value set
to the object size (1 << order). Backwards compatibility is taken care
of on the libceph/osd side.
"The CEPH_OSD_OP_SETALLOCHINT hint is durable, in that it's enough to
do it once. The reason every rbd write is prefixed is that rbd doesn't
explicitly create objects and relies on writes creating them
implicitly, so there is no place to stick a single hint op into. To
get around that we decided to prefix every rbd write with a hint (just
like write and setattr ops, hint op will create an object implicitly if
it doesn't exist)."
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
In preparation for prefixing rbd writes with an allocation hint
introduce a num_ops parameter for rbd_osd_req_create(). The rationale
is that not every write request is a write op that needs to be prefixed
(e.g. watch op), so the num_ops logic needs to be in the callers.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Our longest osd request now contains 3 ops: copyup+hint+write.
Also, CEPH_OSD_MAX_OP value in a BUG_ON in rbd_osd_req_callback() was
hard-coded to 2. Fix it, and switch to rbd_assert while at it.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@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>
Doing rbd_obj_request_put() in rbd_img_request_fill() error paths is
not only insufficient, but also triggers an rbd_assert() in
rbd_obj_request_destroy():
Assertion failure in rbd_obj_request_destroy() at line 1867:
rbd_assert(obj_request->img_request == NULL);
rbd_img_obj_request_add() adds obj_requests to the img_request, the
opposite is rbd_img_obj_request_del(). Use it.
Fixes: http://tracker.ceph.com/issues/7327
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Commit 03507db631 ("rbd: fix buffer size for writes to images with
snapshots") moved the call to rbd_img_obj_request_add() up, making the
out_partial label bogus. Remove it.
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Alex Elder <elder@linaro.org>
With the addition of erasure coding support in the future, scratch
variable-length array in crush_do_rule_ary() is going to grow to at
least 200 bytes on average, on top of another 128 bytes consumed by
rawosd/osd arrays in the call chain. Replace it with a buffer inside
struct osdmap and a mutex. This shouldn't result in any contention,
because all osd requests were already serialized by request_mutex at
that point; the only unlocked caller was ceph_ioctl_get_dataloc().
Signed-off-by: Ilya Dryomov <ilya.dryomov@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Pull vfs fixes from Al Viro:
"Switch mnt_hash to hlist, turning the races between __lookup_mnt() and
hash modifications into false negatives from __lookup_mnt() (instead
of hangs)"
On the false negatives from __lookup_mnt():
"The *only* thing we care about is not getting stuck in __lookup_mnt().
If it misses an entry because something in front of it just got moved
around, etc, we are fine. We'll notice that mount_lock mismatch and
that'll be it"
* 'for-linus-2' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
switch mnt_hash to hlist
don't bother with propagate_mnt() unless the target is shared
keep shadowed vfsmounts together
resizable namespace.c hashes
I am the new kernel tree Documentation maintainer (except for parts that
are handled by other people, of course).
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Acked-by: Rob Landley <rob@landley.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull input updates from Dmitry Torokhov:
"Some more updates for the input subsystem.
You will get a fix for race in mousedev that has been causing quite a
few oopses lately and a small fixup for force feedback support in
evdev"
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input:
Input: mousedev - fix race when creating mixed device
Input: don't modify the id of ioctl-provided ff effect on upload failure
It its possible to configure your PAM stack to refuse login if audit
messages (about the login) were unable to be sent. This is common in
many distros and thus normal configuration of many containers. The PAM
modules determine if audit is enabled/disabled in the kernel based on
the return value from sending an audit message on the netlink socket.
If userspace gets back ECONNREFUSED it believes audit is disabled in the
kernel. If it gets any other error else it refuses to let the login
proceed.
Just about ever since the introduction of namespaces the kernel audit
subsystem has returned EPERM if the task sending a message was not in
the init user or pid namespace. So many forms of containers have never
worked if audit was enabled in the kernel.
BUT if the container was not in net_init then the kernel network code
would send ECONNREFUSED (instead of the audit code sending EPERM). Thus
by pure accident/dumb luck/bug if an admin configured the PAM stack to
reject all logins that didn't talk to audit, but then ran the login
untility in the non-init_net namespace, it would work!! Clearly this was
a bug, but it is a bug some people expected.
With the introduction of network namespace support in 3.14-rc1 the two
bugs stopped cancelling each other out. Now, containers in the
non-init_net namespace refused to let users log in (just like PAM was
configfured!) Obviously some people were not happy that what used to let
users log in, now didn't!
This fix is kinda hacky. We return ECONNREFUSED for all non-init
relevant namespaces. That means that not only will the old broken
non-init_net setups continue to work, now the broken non-init_pid or
non-init_user setups will 'work'. They don't really work, since audit
isn't logging things. But it's what most users want.
In 3.15 we should have patches to support not only the non-init_net
(3.14) namespace but also the non-init_pid and non-init_user namespace.
So all will be right in the world. This just opens the doors wide open
on 3.14 and hopefully makes users happy, if not the audit system...
Reported-by: Andre Tomt <andre@tomt.net>
Reported-by: Adam Richter <adam_richter2004@yahoo.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
