In the Linux kernel, the following vulnerability has been resolved:
io_uring/rw: fix missing NOWAIT check for O_DIRECT start write
When io_uring starts a write, it'll call kiocb_start_write() to bump the
super block rwsem, preventing any freezes from happening while that
write is in-flight. The freeze side will grab that rwsem for writing,
excluding any new writers from happening and waiting for existing writes
to finish. But io_uring unconditionally uses kiocb_start_write(), which
will block if someone is currently attempting to freeze the mount point.
This causes a deadlock where freeze is waiting for previous writes to
complete, but the previous writes cannot complete, as the task that is
supposed to complete them is blocked waiting on starting a new write.
This results in the following stuck trace showing that dependency with
the write blocked starting a new write:
task:fio state:D stack:0 pid:886 tgid:886 ppid:876
Call trace:
__switch_to+0x1d8/0x348
__schedule+0x8e8/0x2248
schedule+0x110/0x3f0
percpu_rwsem_wait+0x1e8/0x3f8
__percpu_down_read+0xe8/0x500
io_write+0xbb8/0xff8
io_issue_sqe+0x10c/0x1020
io_submit_sqes+0x614/0x2110
__arm64_sys_io_uring_enter+0x524/0x1038
invoke_syscall+0x74/0x268
el0_svc_common.constprop.0+0x160/0x238
do_el0_svc+0x44/0x60
el0_svc+0x44/0xb0
el0t_64_sync_handler+0x118/0x128
el0t_64_sync+0x168/0x170
INFO: task fsfreeze:7364 blocked for more than 15 seconds.
Not tainted 6.12.0-rc5-00063-g76aaf945701c #7963
with the attempting freezer stuck trying to grab the rwsem:
task:fsfreeze state:D stack:0 pid:7364 tgid:7364 ppid:995
Call trace:
__switch_to+0x1d8/0x348
__schedule+0x8e8/0x2248
schedule+0x110/0x3f0
percpu_down_write+0x2b0/0x680
freeze_super+0x248/0x8a8
do_vfs_ioctl+0x149c/0x1b18
__arm64_sys_ioctl+0xd0/0x1a0
invoke_syscall+0x74/0x268
el0_svc_common.constprop.0+0x160/0x238
do_el0_svc+0x44/0x60
el0_svc+0x44/0xb0
el0t_64_sync_handler+0x118/0x128
el0t_64_sync+0x168/0x170
Fix this by having the io_uring side honor IOCB_NOWAIT, and only attempt a
blocking grab of the super block rwsem if it isn't set. For normal issue
where IOCB_NOWAIT would always be set, this returns -EAGAIN which will
have io_uring core issue a blocking attempt of the write. That will in
turn also get completions run, ensuring forward progress.
Since freezing requires CAP_SYS_ADMIN in the first place, this isn't
something that can be triggered by a regular user.
References
In the Linux kernel, the following vulnerability has been resolved:
io_uring/rw: fix missing NOWAIT check for O_DIRECT start write
When io_uring starts a write, it'll call kiocb_start_write() to bump the
super block rwsem, preventing any freezes from happening while that
write is in-flight. The freeze side will grab that rwsem for writing,
excluding any new writers from happening and waiting for existing writes
to finish. But io_uring unconditionally uses kiocb_start_write(), which
will block if someone is currently attempting to freeze the mount point.
This causes a deadlock where freeze is waiting for previous writes to
complete, but the previous writes cannot complete, as the task that is
supposed to complete them is blocked waiting on starting a new write.
This results in the following stuck trace showing that dependency with
the write blocked starting a new write:
task:fio state:D stack:0 pid:886 tgid:886 ppid:876
Call trace:
__switch_to+0x1d8/0x348
__schedule+0x8e8/0x2248
schedule+0x110/0x3f0
percpu_rwsem_wait+0x1e8/0x3f8
__percpu_down_read+0xe8/0x500
io_write+0xbb8/0xff8
io_issue_sqe+0x10c/0x1020
io_submit_sqes+0x614/0x2110
__arm64_sys_io_uring_enter+0x524/0x1038
invoke_syscall+0x74/0x268
el0_svc_common.constprop.0+0x160/0x238
do_el0_svc+0x44/0x60
el0_svc+0x44/0xb0
el0t_64_sync_handler+0x118/0x128
el0t_64_sync+0x168/0x170
INFO: task fsfreeze:7364 blocked for more than 15 seconds.
Not tainted 6.12.0-rc5-00063-g76aaf945701c #7963
with the attempting freezer stuck trying to grab the rwsem:
task:fsfreeze state:D stack:0 pid:7364 tgid:7364 ppid:995
Call trace:
__switch_to+0x1d8/0x348
__schedule+0x8e8/0x2248
schedule+0x110/0x3f0
percpu_down_write+0x2b0/0x680
freeze_super+0x248/0x8a8
do_vfs_ioctl+0x149c/0x1b18
__arm64_sys_ioctl+0xd0/0x1a0
invoke_syscall+0x74/0x268
el0_svc_common.constprop.0+0x160/0x238
do_el0_svc+0x44/0x60
el0_svc+0x44/0xb0
el0t_64_sync_handler+0x118/0x128
el0t_64_sync+0x168/0x170
Fix this by having the io_uring side honor IOCB_NOWAIT, and only attempt a
blocking grab of the super block rwsem if it isn't set. For normal issue
where IOCB_NOWAIT would always be set, this returns -EAGAIN which will
have io_uring core issue a blocking attempt of the write. That will in
turn also get completions run, ensuring forward progress.
Since freezing requires CAP_SYS_ADMIN in the first place, this isn't
something that can be triggered by a regular user.
References