kernel/git/stable/linux.git/fs/ceph, branch master

Merge tag 'ceph-for-7.3-rc1' of https://github.com/ceph/ceph-client

2026-08-28T18:51:05+00:00

Pull ceph updates from Ilya Dryomov: "A wide variety of mostly CephFS fixes and cleanups, split between changes that address edge cases (Sam, Xiubo, Matthew), efficiency improvements (Max) and AI-assisted hardening (Michael, Jeremy). One thing that stands out is Alex's change to how CephFS behaves in NEARFULL scenarios: the long-standing "make all writes synchronous" behavior has become opt-in. It was always somewhat controversial and doesn't make much sense for modern deployments; the new default is to continue normal operation (i.e. buffer writes as MDS allows, etc). The behavior in case the cluster reaches any FULL state remains the same as before" * tag 'ceph-for-7.3-rc1' of https://github.com/ceph/ceph-client: (32 commits) ceph: force a cap message when a deferred revoke can't be acked immediately libceph: reject buckets with mismatched CRUSH ids ceph: reject export_targets ranks >= CEPH_MAX_MDS in mdsmap decode ceph: fix leaked inode reference on writeback abort at umount libceph: remove ceph_put_page_vector() libceph: validate banner payload length ceph: make nearfull sync writes opt-in ceph: do not repeat ceph_trim_dentries() if no progress possible ceph: drop mdsc->mutex before decoding the MDS reply ceph: fix UAF in check_new_map() on session freed during unlock ceph: fix UAF in __kick_flushing_caps() on cf entry freed during unlock ceph: pass inode pointer around instead of reloading it ceph: mark cap remove with RB_CLEAR_NODE() instead of setting ci=NULL ceph: add helper function ceph_cap_is_removed() ceph: make __ceph_remove_cap() static ceph: cap delegated inode count in ceph_parse_deleg_inos() ceph: bound num_export_targets array for mds info v2/v3 ceph: bound MDSCapAuth path and fs_name decode in handle_session() ceph: bound xattr value length in __build_xattrs() ceph: bound copied dentry name length in NFS export get_name ...

ceph: force a cap message when a deferred revoke can't be acked immediately

2026-08-26T17:57:29+00:00

When the MDS revokes capabilities, handle_cap_grant() normally guarantees a response by setting `CHECK_CAPS_FLUSH_FORCE` (see commit 31634d7597d8 ("ceph: force sending a cap update msg back to MDS for revoke op")), so ceph_check_caps() sends a cap message even if the client would otherwise decide it has nothing to do. That guarantee is skipped whenever the revoke has to be deferred (via revoke_wait): revoking Fb while dirty data is still buffered (writeback is queued first) or revoking Fc while pages are cached (async invalidation is queued first). In those cases, the ack is left to the deferred completion (ceph_put_wrbuffer_cap_refs() after writeback, or the invalidate worker after invalidation); both of which call ceph_check_caps(ci,0) i.e. without `CHECK_CAPS_FLUSH_FORCE`. Nothing gets sent under one of the following conditions: - the inode is retaining caps because the file was used recently (file_wanted != 0; retain |= CEPH_CAP_ANY) - the revoked cap is still used because the page was re-cached (e.g. a file being re-read) - the MDS has meanwhile re-granted, so `issued==implemented` and the client sees nothing being revoked The client then never emits the cap message which the MDS is waiting for. The MDS blocks on the revoke indefinitely and logs, for minutes or hours: client.NNN isn't responding to mclientcaps(revoke), ino 0x... pending pAsxLsXsxFsxcrwb issued pAsxLsXsxFsxcrwb, sent 964.899182 seconds ago The client-side state at that point shows the full cap set still issued, nothing in the revoking/flushing sets. Thus nothing gets sent. This patch fixes it by remembering that a forced response is expected. When a revoke is deferred, set `CEPH_I_FLUSH_FORCE` on the inode. ceph_check_caps() replays it as `CHECK_CAPS_FLUSH_FORCE`, so whichever path re-checks the inode next (the writeback/invalidate completion, the delayed worker, or any other caller) is guaranteed to send a cap message to the MDS. __prep_cap() clears the flag once a message is actually built. This is the deferred-path counterpart of the existing `CHECK_CAPS_FLUSH_FORCE` handling; a normal (non-deferred) revoke still forces the response inline as before. Cc: stable@vger.kernel.org Fixes: 31634d7597d8 ("ceph: force sending a cap update msg back to MDS for revoke op") Fixes: 257e6172ab36 ("ceph: don't let check_caps skip sending responses for revoke msgs") Signed-off-by: Max Kellermann Reviewed-by: Alex Markuze Signed-off-by: Ilya Dryomov

ceph: reject export_targets ranks >= CEPH_MAX_MDS in mdsmap decode

2026-08-26T17:57:29+00:00

MDSMap export_targets entries are monitor controlled. check_new_map() uses each entry as a bit number in a fixed stack bitmap, so a rank outside the protocol namespace can make set_bit() write past the end of the array. Reject ranks outside CEPH_MAX_MDS while decoding the map. Do not validate against possible_max_rank here because maps may legitimately reference ranks beyond a temporarily reduced max_mds. Cc: stable@vger.kernel.org Fixes: d517b3983dd3 ("ceph: reconnect to the export targets on new mdsmaps") Signed-off-by: Jérémy Jean Reviewed-by: Alex Markuze Signed-off-by: Alex Markuze Signed-off-by: Ilya Dryomov

ceph: fix leaked inode reference on writeback abort at umount

2026-08-26T17:57:28+00:00

ceph_dirty_folio() takes a wrbuffer claim on each newly dirtied folio: it bumps i_wrbuffer_ref (taking an ihold() on the 0->1 transition) and attaches the snap_context to folio->private. That claim is released only by ceph_put_wrbuffer_cap_refs(), which for a submitted write runs from writepages_finish(). In ceph_submit_write(), if ceph_inc_osd_stopping_blocker() fails -- which happens during umount -- the request is aborted before submission: the already-collected folios are only redirtied and unlocked, so writepages_finish() never runs and the claim is leaked. redirty_page_for_writepage() -> folio_redirty_for_writepage() -> filemap_dirty_folio() sets PG_dirty directly and does not go through ->dirty_folio, so ceph_dirty_folio() is not re-entered to rebalance it. Because every subsequent writeback also fails the osd_stopping_blocker, i_wrbuffer_ref never returns to 0, the ihold() is never dropped, and the inode cannot be evicted: VFS: Busy inodes after unmount of ceph kernel BUG at fs/super.c:650! Release the orphaned claim in the abort path before redirtying, via ceph_undo_wrbuffer_claim(): detach the snap_context, drop the wrbuffer reference (letting i_wrbuffer_ref reach 0 and iput() the inode), and drop the snap_context reference -- i.e. do what writepages_finish() would have done for these never-submitted folios. Only the locked_pages entries are undone; folios still in the fbatch were never dirty-cleared by this call (folio_clear_dirty_for_io() is the ownership-transfer point, and a successful move NULLs the fbatch slot), so they hold no claim this call owns. Cc: stable@vger.kernel.org Fixes: fd7449d937e7 ("ceph: fix generic/421 test failure") Signed-off-by: Matthew Brown Reviewed-by: Xiubo Li Signed-off-by: Ilya Dryomov

libceph: remove ceph_put_page_vector()

2026-08-26T17:57:28+00:00

ceph_put_page_vector() was paired with ceph_get_direct_page_vector(), which was removed in commit 97a385e55829 ("libceph: remove ceph_get_direct_page_vector()"). Its only remaining caller, finish_netfs_read(), uses it to put a page vector allocated with iov_iter_get_pages_alloc2(), which is confusing. Open-code the put_page() loop and kvfree() there instead. The caller passed dirty = false, so this also removes the dead dirty branch and with it a call to the deprecated set_page_dirty_lock(). Signed-off-by: Tal Zussman Reviewed-by: Ilya Dryomov Signed-off-by: Ilya Dryomov

ceph: make nearfull sync writes opt-in

2026-08-26T17:57:28+00:00

The kernel CephFS client has historically treated a cluster or pool NEARFULL condition as a request to force successful writes through generic_write_sync(). That effectively turns otherwise buffered writes into synchronous writes and can cause a severe throughput drop as soon as a single OSD or the file data pool crosses the nearfull threshold. On modern large clusters, NEARFULL is primarily an operator health signal rather than an immediate client-side capacity failure. Operators can still have substantial usable capacity while a cluster is rebalancing, splitting PGs, or expanding onto new devices. RBD, RGW and the userspace CephFS client do not impose this extra client-side sync-write throttle, so the kernel client behavior is surprising and operationally painful. Change the default behavior so NEARFULL no longer changes normal write-sync semantics. FULL and pool FULL still fail with -ENOSPC, and explicitly synchronous writes continue to be synced by generic_write_sync(). Add a nearfull_sync mount option for deployments that want the legacy backpressure behavior. When this option is set, successful writes are promoted to IOCB_DSYNC if the cluster or file data pool is marked NEARFULL, preserving the old behavior for conservative deployments. Link: https://tracker.ceph.com/issues/74849 Signed-off-by: Alex Markuze Reviewed-by: Xiubo Li Signed-off-by: Ilya Dryomov

ceph: do not repeat ceph_trim_dentries() if no progress possible

2026-08-26T17:57:28+00:00

ceph_cap_reclaim_work() re-queues itself for as long as ceph_trim_dentries() returns -EAGAIN, which happens whenever a lease walk exhausts its `nr_to_scan` budget. This creates a busy loop that consumes CPU without making any progress when there is nothing to reclaim: with no cap pressure (`count==0`) and every scanned lease still valid, each pass runs the full scan budget down to zero and returns `-EAGAIN`, only to be queued again immediately. The dir-lease walk made this worse. When `expire_dir_lease` is `false` (i.e. we have no intention of reclaiming dir leases), __dir_lease_check() returned `TOUCH` for every valid lease. `TOUCH` moves the dentry to the tail of the list and resets `di->time` via __dentry_dir_lease_touch(), so a walk over N valid leases pointlessly rewrote the list, refreshed the timestamps (preventing them from ever aging out) and always drained `nr_to_scan`, guaranteeing the `-EAGAIN` requeue. Fix this in three steps: - Return `KEEP` instead of `TOUCH` when `expire_dir_lease` is `false`. If we are not going to reclaim the lease, leave it in place instead of churning the list and resetting its timestamp; the walk then terminates naturally (or via `STOP` at the first fresh lease). - Only return `-EAGAIN` from the first (dentry-lease) walk when something was actually freed. A full batch that frees nothing means retrying the same list immediately is futile; fall through to the dir-lease walk instead. - After both walks, bail out with success (0) when nothing was freed and there is no cap pressure (`count==0`). There is no reason to keep retrying when we are not over the cap limit and made no progress. Under real cap pressure (`count>0`) the reclaim path is unchanged and still retries via `-EAGAIN`. Without this patch, I saw 500 ceph_trim_dentries() calls per second on our web servers. This is very visible in `/proc/lock_stat` (5 minute capture): class name con-bounces contentions waittime-min waittime-max waittime-total waittime-avg acq-bounces acquisitions holdtime-min holdtime-max holdtime-total holdtime-avg &mdsc->dentry_list_lock: 126180 128218 0.04 8063.44 15986965.20 124.69 1573354 5296812 0.04 8291.28 74164526.48 14.00 ----------------------- &mdsc->dentry_list_lock 111736 [<000000007b11e319>] __ceph_dentry_dir_lease_touch+0x7c/0xa8 &mdsc->dentry_list_lock 2631 [<0000000050597999>] __dentry_leases_walk+0x64/0x2c8 &mdsc->dentry_list_lock 3878 [<00000000c0022f62>] __ceph_dentry_lease_touch+0x5c/0xa8 &mdsc->dentry_list_lock 9973 [<000000002f27cb6f>] __dentry_lease_unlist+0x50/0xa0 ----------------------- &mdsc->dentry_list_lock 123621 [<0000000050597999>] __dentry_leases_walk+0x64/0x2c8 &mdsc->dentry_list_lock 1822 [<000000007b11e319>] __ceph_dentry_dir_lease_touch+0x7c/0xa8 &mdsc->dentry_list_lock 2720 [<000000002f27cb6f>] __dentry_lease_unlist+0x50/0xa0 &mdsc->dentry_list_lock 55 [<00000000c0022f62>] __ceph_dentry_lease_touch+0x5c/0xa8 With this patch: class name con-bounces contentions waittime-min waittime-max waittime-total waittime-avg acq-bounces acquisitions holdtime-min holdtime-max holdtime-total holdtime-avg &mdsc->dentry_list_lock: 1203 1215 0.16 408.88 33082.88 27.23 4320501 7357389 0.04 500.64 1961578.00 0.27 ----------------------- &mdsc->dentry_list_lock 1029 [<000000003c9aea8a>] __ceph_dentry_dir_lease_touch+0x7c/0xa8 &mdsc->dentry_list_lock 169 [<000000002038c577>] __dentry_lease_unlist+0x50/0xa0 &mdsc->dentry_list_lock 16 [<00000000c991106d>] __ceph_dentry_lease_touch+0x5c/0xa8 &mdsc->dentry_list_lock 1 [<00000000612fe15f>] __dentry_leases_walk+0x64/0x2c8 ----------------------- &mdsc->dentry_list_lock 158 [<000000002038c577>] __dentry_lease_unlist+0x50/0xa0 &mdsc->dentry_list_lock 858 [<000000003c9aea8a>] __ceph_dentry_dir_lease_touch+0x7c/0xa8 &mdsc->dentry_list_lock 182 [<00000000612fe15f>] __dentry_leases_walk+0x64/0x2c8 &mdsc->dentry_list_lock 17 [<00000000c991106d>] __ceph_dentry_lease_touch+0x5c/0xa8 __dentry_leases_walk() is almost gone. The total wait time is reduced by a factor of 483. That will give some latency gains to ceph_readdir(). Cc: stable@vger.kernel.org Fixes: 37c4efc1ddf9 ("ceph: periodically trim stale dentries") Signed-off-by: Max Kellermann Reviewed-by: Alex Markuze Signed-off-by: Ilya Dryomov

ceph: drop mdsc->mutex before decoding the MDS reply

2026-08-26T17:57:28+00:00

handle_reply() held `mdsc->mutex` across parse_reply_info(), i.e. across the full decode of the reply message. For large replies (a big readdir allocates and parses many dir_entries), this can take a while and blocks ceph_mdsc_submit_request() calls meanwhile. The decode does not need `mdsc->mutex`: parse_reply_info() mostly fills the request's `r_reply_info`. Create replies may also add delegated inode numbers to the session xarray, but that xarray is protected by its own lock and is not serialized by `mdsc->mutex` today. By the time we reach parse_reply_info(), all `mdsc->mutex`-protected state has already been updated under the lock (the request has either been unregistered (safe reply) or added to the session's unsafe list (unsafe reply)) and the request is pinned by the reference taken in lookup_get_request(). Drop `mdsc->mutex` before calling parse_reply_info() so reply decoding no longer blocks request submission. This only widens the existing unlocked window that already covers the heavier ceph_fill_trace() / ceph_readdir_prepopulate() processing, so no new races are introduced. Signed-off-by: Max Kellermann Reviewed-by: Xiubo Li Signed-off-by: Ilya Dryomov

ceph: fix UAF in check_new_map() on session freed during unlock

2026-08-26T17:57:28+00:00

check_new_map() iterates mdsc->sessions[] and for each active session drops mdsc->mutex to perform per-session operations. The forced-close path (rank removed from map) correctly takes a reference on s via ceph_get_mds_session() before releasing mdsc->mutex, but three other paths do not: Path A (address changed): mutex_unlock → mutex_lock(&s->s_mutex) Path B (reconnect): mutex_unlock → send_mds_reconnect(mdsc, s) Path C (active transition): mutex_unlock → mutex_lock(&s->s_mutex) Without the extra reference, another thread can acquire mdsc->mutex during the unlock window, call __unregister_session() which drops the last reference on s, and free it. The original thread then accesses freed memory via s->s_mutex. Fix by adding ceph_get_mds_session(s) before each mutex_unlock and ceph_put_mds_session(s) after the corresponding mutex_lock, matching the pattern already used in the forced-close path. Race timeline (Path A): Thread A (check_new_map) Thread B (another map update holds mdsc->mutex or session teardown) -------------------------- -------------------------- s = mdsc->sessions[i] (refcount == 1, held only by sessions[] array) mutex_unlock(&mdsc->mutex) ---> acquires mdsc->mutex __unregister_session(mdsc, s) sessions[i] = NULL ceph_put_mds_session(s) refcount: 1 -> 0 kfree(s) <--- freed! mutex_lock(&s->s_mutex) UAF on freed s->s_mutex Cc: stable@vger.kernel.org Signed-off-by: Xiubo Li Reviewed-by: Viacheslav Dubeyko Signed-off-by: Ilya Dryomov

ceph: fix UAF in __kick_flushing_caps() on cf entry freed during unlock

2026-08-26T17:57:28+00:00

list_for_each_entry() iterates ci->i_cap_flush_list but drops i_ceph_lock to send cap messages. During the unlock window, handle_cap_flush_ack() can acquire i_ceph_lock, detach cf entries with tid <= flush_tid from the list, release i_ceph_lock, and free them via ceph_free_cap_flush() outside any lock. When the original thread reacquires i_ceph_lock and the for-loop macro advances via cf = list_next_entry(cf, i_list), it dereferences cf->i_list.next on freed memory. The race timeline: __kick_flushing_caps() handle_cap_flush_ack() ----------------------- ----------------------- holds i_ceph_lock <--- iterates to cf (tid=10) prepares FLUSH message drops i_ceph_lock <--- __send_cap() ── FLUSH(tid=10) MDS sends FLUSH_ACK(tid=10) ---> acquires i_ceph_lock cf->tid(10) <= flush_tid(10), detaches cf from i_cap_flush_list drops i_ceph_lock ceph_free_cap_flush(cf) <- frees it! acquires i_ceph_lock <--- for-loop advances: cf = list_next_entry(cf, i_list) -- UAF on freed cf->i_list.next The cf was just sent by __kick_flushing_caps itself via __send_cap(). The MDS may respond with FLUSH_ACK quickly enough that handle_cap_flush_ack() frees cf before __kick_flushing_caps can finish the iteration. Fix by converting to a manual while loop: save the next pointer under i_ceph_lock before dropping it, then use the saved pointer after reacquiring, so the potentially-freed cf is never accessed again. Cc: stable@vger.kernel.org Signed-off-by: Xiubo Li Reviewed-by: Viacheslav Dubeyko Signed-off-by: Ilya Dryomov