Add complete infrastructure for per-subvolume I/O metrics collection
and reporting to the MDS. This enables administrators to monitor I/O
patterns at the subvolume granularity, which is useful for multi-tenant
CephFS deployments.
This patch adds:
- CEPHFS_FEATURE_SUBVOLUME_METRICS feature flag for MDS negotiation
- CEPH_SUBVOLUME_ID_NONE constant (0) for unknown/unset state
- Red-black tree based metrics tracker for efficient per-subvolume
aggregation with kmem_cache for entry allocations
- Wire format encoding matching the MDS C++ AggregatedIOMetrics struct
- Integration with the existing CLIENT_METRICS message
- Recording of I/O operations from file read/write and writeback paths
- Debugfs interfaces for monitoring (metrics/subvolumes, metrics/metric_features)
Metrics tracked per subvolume include:
- Read/write operation counts
- Read/write byte counts
- Read/write latency sums (for average calculation)
The metrics are periodically sent to the MDS as part of the existing
metrics reporting infrastructure when the MDS advertises support for
the SUBVOLUME_METRICS feature.
CEPH_SUBVOLUME_ID_NONE enforces subvolume_id immutability. Following
the FUSE client convention, 0 means unknown/unset. Once an inode has
a valid (non-zero) subvolume_id, it should not change during the
inode's lifetime.
Signed-off-by: Alex Markuze <amarkuze@redhat.com>
Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
ceph_mdsc_build_path() must be called with a zero-initialized
ceph_path_info parameter, or else the following
ceph_mdsc_free_path_info() may crash.
Example crash (on Linux 6.18.12):
virt_to_cache: Object is not a Slab page!
WARNING: CPU: 184 PID: 2871736 at mm/slub.c:6732 kmem_cache_free+0x316/0x400
[...]
Call Trace:
[...]
ceph_open+0x13d/0x3e0
do_dentry_open+0x134/0x480
vfs_open+0x2a/0xe0
path_openat+0x9a3/0x1160
[...]
cache_from_obj: Wrong slab cache. names_cache but object is from ceph_inode_info
WARNING: CPU: 184 PID: 2871736 at mm/slub.c:6746 kmem_cache_free+0x2dd/0x400
[...]
kernel BUG at mm/slub.c:634!
Oops: invalid opcode: 0000 [#1] SMP NOPTI
RIP: 0010:__slab_free+0x1a4/0x350
Some of the ceph_mdsc_build_path() callers had initializers, but
others had not, even though they were all added by commit 15f519e9f8
("ceph: fix race condition validating r_parent before applying state").
The ones without initializer are suspectible to random crashes. (I can
imagine it could even be possible to exploit this bug to elevate
privileges.)
Unfortunately, these Ceph functions are undocumented and its semantics
can only be derived from the code. I see that ceph_mdsc_build_path()
initializes the structure only on success, but not on error.
Calling ceph_mdsc_free_path_info() after a failed
ceph_mdsc_build_path() call does not even make sense, but that's what
all callers do, and for it to be safe, the structure must be
zero-initialized. The least intrusive approach to fix this is
therefore to add initializers everywhere.
Cc: stable@vger.kernel.org
Fixes: 15f519e9f8 ("ceph: fix race condition validating r_parent before applying state")
Signed-off-by: Max Kellermann <max.kellermann@ionos.com>
Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Add validation to ensure the cached parent directory inode matches the
directory info in MDS replies. This prevents client-side race conditions
where concurrent operations (e.g. rename) cause r_parent to become stale
between request initiation and reply processing, which could lead to
applying state changes to incorrect directory inodes.
[ idryomov: folded a kerneldoc fixup and a follow-up fix from Alex to
move CEPH_CAP_PIN reference when r_parent is updated:
When the parent directory lock is not held, req->r_parent can become
stale and is updated to point to the correct inode. However, the
associated CEPH_CAP_PIN reference was not being adjusted. The
CEPH_CAP_PIN is a reference on an inode that is tracked for
accounting purposes. Moving this pin is important to keep the
accounting balanced. When the pin was not moved from the old parent
to the new one, it created two problems: The reference on the old,
stale parent was never released, causing a reference leak.
A reference for the new parent was never acquired, creating the risk
of a reference underflow later in ceph_mdsc_release_request(). This
patch corrects the logic by releasing the pin from the old parent and
acquiring it for the new parent when r_parent is switched. This
ensures reference accounting stays balanced. ]
Cc: stable@vger.kernel.org
Signed-off-by: Alex Markuze <amarkuze@redhat.com>
Reviewed-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Initially, ceph_fs_debugfs_init() had temporary
name buffer with hardcoded length of 80 symbols.
Then, it was hardcoded again for 100 symbols.
Finally, it makes sense to exchange hardcoded
value on properly defined constant and 255 symbols
should be enough for any name case.
Signed-off-by: Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>
Reviewed-by: Patrick Donnelly <pdonnell@ibm.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Multiple CephFS mounts on a host is increasingly common so
disambiguating messages like this is necessary and will make it easier
to debug issues.
At the same this will improve the debug logs to make them easier to
troubleshooting issues, such as print the ino# instead only printing
the memory addresses of the corresponding inodes and print the dentry
names instead of the corresponding memory addresses for the dentry,etc.
Link: https://tracker.ceph.com/issues/61590
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Patrick Donnelly <pdonnell@redhat.com>
Reviewed-by: Milind Changire <mchangir@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
When trimming the caps and just after the 'session->s_cap_lock' is
released in ceph_iterate_session_caps() the cap maybe removed by
another thread, and when using the stale cap memory in the callbacks
it will trigger use-after-free crash.
We need to check the existence of the cap just after the 'ci->i_ceph_lock'
being acquired. And do nothing if it's already removed.
Cc: stable@vger.kernel.org
Link: https://tracker.ceph.com/issues/43272
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Luís Henriques <lhenriques@suse.de>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
This patch adds latency and size metrics for remote object copies
operations ("copyfrom"). For now, these metrics will be available on the
client only, they won't be sent to the MDS.
Signed-off-by: Luís Henriques <lhenriques@suse.de>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
This patch modifies struct ceph_client_metric so that each metric block
(read, write and metadata) becomes an element in a array. This allows to
also re-write the helper functions that handle these blocks, making them
simpler and, above all, reduce the amount of copy&paste every time a new
metric is added.
Thus, for each of these metrics there will be a new struct ceph_metric
entry that'll will contain all the sizes and latencies fields (and a lock).
Note however that the metadata metric doesn't really use the size_fields,
and thus this metric won't be shown in the debugfs '../metrics/size' file.
Signed-off-by: Luís Henriques <lhenriques@suse.de>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Currently, all the metrics are grouped together in a single file, making
it difficult to process this file from scripts. Furthermore, as new
metrics are added, processing this file will become even more challenging.
This patch turns the 'metric' file into a directory that will contain
several files, one for each metric.
Signed-off-by: Luís Henriques <lhenriques@suse.de>
Reviewed-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
This will collect IO's total size and then calculate the average
size, and also will collect the min/max IO sizes.
The debugfs will show the size metrics in bytes and will let the
userspace applications to switch to what they need.
URL: https://tracker.ceph.com/issues/49913
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
In client for each inode, it may have many opened files and may
have been pinned in more than one MDS servers. And some inodes
are idle, which have no any opened files.
This patch will show these metrics in the debugfs, likes:
item total
-----------------------------------------
opened files / total inodes 14 / 5
pinned i_caps / total inodes 7 / 5
opened inodes / total inodes 3 / 5
Will send these metrics to ceph, which will be used by the `fs top`,
later.
[ jlayton: drop unrelated hunk, count hashed inodes instead of
allocated ones ]
URL: https://tracker.ceph.com/issues/47005
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
In multi-mds, the 'caps' debugfs file will have duplicate ino,
add the 'mds' column to indicate which mds session the cap belongs to.
Signed-off-by: Yanhu Cao <gmayyyha@gmail.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Tuan and Ulrich mentioned that they were hitting a problem on s390x,
which has a 32-bit ino_t value, even though it's a 64-bit arch (for
historical reasons).
I think the current handling of inode numbers in the ceph driver is
wrong. It tries to use 32-bit inode numbers on 32-bit arches, but that's
actually not a problem. 32-bit arches can deal with 64-bit inode numbers
just fine when userland code is compiled with LFS support (the common
case these days).
What we really want to do is just use 64-bit numbers everywhere, unless
someone has mounted with the ino32 mount option. In that case, we want
to ensure that we hash the inode number down to something that will fit
in 32 bits before presenting the value to userland.
Add new helper functions that do this, and only do the conversion before
presenting these values to userland in getattr and readdir.
The inode table hashvalue is changed to just cast the inode number to
unsigned long, as low-order bits are the most likely to vary anyway.
While it's not strictly required, we do want to put something in
inode->i_ino. Instead of basing it on BITS_PER_LONG, however, base it on
the size of the ino_t type.
NOTE: This is a user-visible change on 32-bit arches:
1/ inode numbers will be seen to have changed between kernel versions.
32-bit arches will see large inode numbers now instead of the hashed
ones they saw before.
2/ any really old software not built with LFS support may start failing
stat() calls with -EOVERFLOW on inode numbers >2^32. Nothing much we
can do about these, but hopefully the intersection of people running
such code on ceph will be very small.
The workaround for both problems is to mount with "-o ino32".
[ idryomov: changelog tweak ]
URL: https://tracker.ceph.com/issues/46828
Reported-by: Ulrich Weigand <Ulrich.Weigand@de.ibm.com>
Reported-and-Tested-by: Tuan Hoang1 <Tuan.Hoang1@ibm.com>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: "Yan, Zheng" <zyan@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The variable mds is being initialized with a value that is never read
and it is being updated later with a new value. The initialization is
redundant and can be removed.
Addresses-Coverity: ("Unused value")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
This will help to reduce using the global mdsc->mutex lock in many
places.
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Calculate the latency for OSD read requests. Add a new r_end_stamp
field to struct ceph_osd_request that will hold the time of that
the reply was received. Use that to calculate the RTT for each call,
and divide the sum of those by number of calls to get averate RTT.
Keep a tally of RTT for OSD writes and number of calls to track average
latency of OSD writes.
URL: https://tracker.ceph.com/issues/43215
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
For dentry leases, only count the hit/miss info triggered from the vfs
calls. For the cases like request reply handling and ceph_trim_dentries,
ignore them.
For now, these are only viewable using debugfs. Future patches will
allow the client to send the stats to the MDS.
The output looks like:
item total miss hit
-------------------------------------------------
d_lease 11 7 141
URL: https://tracker.ceph.com/issues/43215
Signed-off-by: Xiubo Li <xiubli@redhat.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Alex MarkuzeandIlya Dryomov
Max KellermannandIlya Dryomov
Viacheslav DubeykoandIlya Dryomov
Thorsten BlumandIlya Dryomov
Xiubo LiandIlya Dryomov
Venky ShankarandIlya Dryomov
Luís HenriquesandIlya Dryomov
Yanhu CaoandIlya Dryomov
Jeff LaytonandIlya Dryomov
Colin Ian KingandIlya Dryomov
Yufen YuandJens Axboe