syzbot
WARNING in smbdirect_connect_rdma_event_handler
Status:
moderation: reported on 2026/11/21 14:11
Subsystems:
cifs
Labels:
prio:normal
[Documentation on labels]
Reported-by: syzbot+2ab91c5ef00816058513@syzkaller.appspotmail.com
First crash: 8h30m, last: 8h30m
✨ AI-generated root cause summary:
smbdirect_connect_rdma_event_handler() accesses sc->first_error and sc->status without synchronization or READ_ONCE/barriers against concurrent teardown in smbdirect_connection_wait_for_connected(), __smbdirect_socket_schedule_cleanup(), and smbdirect_socket_destroy_sync(), which modify sc->first_error and sc->status (to SMBDIRECT_SOCKET_DISCONNECTED) before acquiring rdma_lock_handler(sc->rdma.cm_id). When the waiter is interrupted by a signal (-ERESTARTSYS), the cleanup/destroy path can advance sc->status to SMBDIRECT_SOCKET_DISCONNECTED and set sc->first_error = -ERESTARTSYS between the handler's sc->first_error check and its sc->status check in SMBDIRECT_CHECK_STATUS_DISCONNECT(), triggering the WARN_ONCE at fs/smb/smbdirect/connect.c:307. Moreover, because smbdirect_socket_destroy() disables work items before taking rdma_lock_handler(), the same race allows smbdirect_connect_rdma_connect() to re-initialize and schedule sc->idle.timer_work after it was disabled, risking a use-after-free after sc is freed.
▶
▼
✨ Full AI root cause analysis
Root Cause Analysis
The warning is caused by an unsynchronized concurrent state and lifecycle race on struct smbdirect_socket between the RDMA CM event callback (smbdirect_connect_rdma_event_handler()) and the connection waiter/cleanup/teardown path (smbdirect_connection_wait_for_connected(), __smbdirect_socket_schedule_cleanup(), smbdirect_socket_cleanup_work(), and smbdirect_socket_destroy_sync()).
Involved Functions, Structs, and Fields
Trigger Mechanism and Violated Invariant
-
Missing shared synchronization between RDMA CM callbacks and cleanup/teardown:
cma_work_handler() (drivers/infiniband/core/cma.c:3026) holds &id_priv->handler_mutex and invokes cma_cm_event_handler() -> smbdirect_connect_rdma_event_handler() for RDMA_CM_EVENT_ADDR_RESOLVED.
- Concurrently, a task in
smbdirect_connect_sync() -> smbdirect_connection_wait_for_connected() (fs/smb/smbdirect/connection.c:751) is sleeping in wait_event_interruptible_timeout() and is interrupted by a signal (ret == -ERESTARTSYS).
- That task calls
smbdirect_socket_schedule_cleanup(sc, -ERESTARTSYS) (connection.c:774) -> __smbdirect_socket_schedule_cleanup() (fs/smb/smbdirect/socket.c:262), which sets sc->first_error = -ERESTARTSYS (socket.c:286) and sc->status = SMBDIRECT_SOCKET_RESOLVE_ADDR_FAILED (socket.c:354), queues sc->disconnect_work, and returns -ERESTARTSYS to _smbd_get_connection() (fs/smb/client/smbdirect.c:324).
_smbd_get_connection() calls smbdirect_socket_release(sc) -> smbdirect_socket_destroy_sync(sc) (socket.c:714), which runs smbdirect_socket_cleanup_work(&sc->disconnect_work) (socket.c:756). In smbdirect_socket_cleanup_work(), for pre-established states (SMBDIRECT_SOCKET_RESOLVE_ADDR_FAILED), sc->status is transitioned to SMBDIRECT_SOCKET_DISCONNECTED (socket.c:493) without acquiring rdma_lock_handler(sc->rdma.cm_id).
smbdirect_socket_destroy() (socket.c:521) also runs disable_work_sync() / disable_delayed_work_sync(&sc->idle.timer_work) (socket.c:572-577) before finally calling rdma_lock_handler(sc->rdma.cm_id) (socket.c:580).
-
Unlocked, non-atomic checks of sc->first_error and sc->status (connect.c:302-307 and socket.h:683-731):
-
Related teardown ordering hazard in the same race:
- Because
smbdirect_socket_destroy() calls disable_delayed_work_sync(&sc->idle.timer_work) (socket.c:576) and disable_work_sync(&sc->connect.work) (socket.c:573) before acquiring rdma_lock_handler(sc->rdma.cm_id) (socket.c:580), and smbdirect_connect_rdma_connect() calls INIT_DELAYED_WORK(&sc->idle.timer_work, smbdirect_connection_idle_timer_work) and mod_delayed_work() (connect.c:235-237) without synchronizing against teardown, INIT_DELAYED_WORK() can clear the disabled flag and arm sc->idle.timer_work after smbdirect_socket_destroy() already disabled it, leaving a live delayed work item pointing to sc after kfree(sc) in smbdirect_socket_release_destroy().
Suggested Fix Direction
- Synchronize
sc->first_error and sc->status transitions between smbdirect_*_rdma_event_handler(), __smbdirect_socket_schedule_cleanup(), and smbdirect_socket_cleanup_work() (e.g., protecting socket state transitions with a dedicated spinlock, or holding rdma_lock_handler() around state/work teardown before smbdirect_socket_destroy() disables work items).
- Avoid calling
INIT_DELAYED_WORK() in smbdirect_connect_rdma_connect() (connect.c:235) when sc->idle.timer_work was already initialized and disabled in smbdirect_socket_init(), or ensure rdma_lock_handler() is acquired before disable_*_work_sync() in smbdirect_socket_destroy() (while ensuring smbdirect_socket_cleanup_work() / event handlers cannot deadlock on handler_mutex).
Sample crash report:
------------[ cut here ]------------
expected[RESOLVE_ADDR_RUNNING] != DISCONNECTED first_error=-ERESTARTSYS local=127.0.0.1:51036 remote=127.0.0.1:445
WARNING: fs/smb/smbdirect/connect.c:307 at smbdirect_connect_rdma_event_handler+0x1e02/0x2820 fs/smb/smbdirect/connect.c:307, CPU#1: kworker/u8:7/198
Modules linked in:
CPU: 1 UID: 0 PID: 198 Comm: kworker/u8:7 Not tainted syzkaller #0 PREEMPT_{RT,(full)}
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 11/06/2026
Workqueue: rdma_cm cma_work_handler
RIP: 0010:smbdirect_connect_rdma_event_handler+0x2086/0x2820 fs/smb/smbdirect/connect.c:307
Code: df 41 0f b6 04 06 84 c0 0f 85 67 06 00 00 49 63 8f 98 00 00 00 48 89 df 48 c7 c6 20 a8 be 8b 4c 89 ea 4d 89 e0 4c 8b 4c 24 08 <67> 48 0f b9 3a 4c 89 ff 48 c7 c6 60 97 be 8b 31 d2 48 c7 c1 6f 55
RSP: 0018:ffffc90004257a08 EFLAGS: 00010246
RAX: 0000000000000000 RBX: ffffffff90dcd1d0 RCX: fffffffffffffe00
RDX: ffffffff8bbeab40 RSI: ffffffff8bbea820 RDI: ffffffff90dcd1d0
RBP: ffff8880295a2098 R08: ffff888062fdc020 R09: ffff888062fdc0a0
R10: 0000000000000100 R11: 0000000000000012 R12: ffff888062fdc020
R13: ffffffff8bbeab40 R14: 1ffff110052b4413 R15: ffff8880295a2000
FS: 0000000000000000(0000) GS:ffff888124804000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00007f4a41662068 CR3: 00000000352ee000 CR4: 00000000003526f0
Call Trace:
<TASK>
cma_cm_event_handler+0xac/0x2c0 drivers/infiniband/core/cma.c:2181
cma_work_handler+0x265/0x520 drivers/infiniband/core/cma.c:3040
process_one_work kernel/workqueue.c:3495 [inline]
process_scheduled_works+0xc3d/0x1630 kernel/workqueue.c:3578
worker_thread+0xa47/0xfb0 kernel/workqueue.c:3659
kthread+0x38b/0x470 kernel/kthread.c:436
ret_from_fork+0x514/0xb70 arch/x86/kernel/process.c:158
ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
</TASK>
----------------
Code disassembly (best guess), 1 bytes skipped:
0: 41 0f b6 04 06 movzbl (%r14,%rax,1),%eax
5: 84 c0 test %al,%al
7: 0f 85 67 06 00 00 jne 0x674
d: 49 63 8f 98 00 00 00 movslq 0x98(%r15),%rcx
14: 48 89 df mov %rbx,%rdi
17: 48 c7 c6 20 a8 be 8b mov $0xffffffff8bbea820,%rsi
1e: 4c 89 ea mov %r13,%rdx
21: 4d 89 e0 mov %r12,%r8
24: 4c 8b 4c 24 08 mov 0x8(%rsp),%r9
* 29: 67 48 0f b9 3a ud1 (%edx),%rdi <-- trapping instruction
2e: 4c 89 ff mov %r15,%rdi
31: 48 c7 c6 60 97 be 8b mov $0xffffffff8bbe9760,%rsi
38: 31 d2 xor %edx,%edx
3a: 48 rex.W
3b: c7 .byte 0xc7
3c: c1 .byte 0xc1
3d: 6f outsl %ds:(%rsi),(%dx)
3e: 55 push %rbp
* Struck through repros no longer work on HEAD.