On 10/31/24 7:54 AM, Jens Axboe wrote:
> On 10/31/24 5:20 AM, Peter Mann wrote:
>> Hello,
>>
>> it appears that there is a high probability of a deadlock occuring when performing fsfreeze on a filesystem which is currently performing multiple io_uring O_DIRECT writes.
>>
>> Steps to reproduce:
>> 1. Mount xfs or ext4 filesystem on /mnt
>>
>> 2. Start writing to the filesystem. Must use io_uring, direct io and iodepth>1 to reproduce:
>> fio --ioengine=io_uring --direct=1 --bs=4k --size=100M --rw=randwrite --loops=100000 --iodepth=32 --name=test --filename=/mnt/fio_test
>>
>> 3. Run this in another shell. For me it deadlocks almost immediately:
>> while true; do fsfreeze -f /mnt/; echo froze; fsfreeze -u /mnt/; echo unfroze; done
>>
>> 4. Fsfreeze and all tasks attempting to write /mnt get stuck:
>> At this point all stuck processes cannot be killed by SIGKILL and they are stuck in uninterruptible sleep.
>> If you try 'touch /mnt/a' for example, the new process gets stuck in the exact same way as well.
>>
>> This gets printed when running 6.11.4 with some debug options enabled:
>> [ 539.586122] Showing all locks held in the system:
>> [ 539.612972] 1 lock held by khungtaskd/35:
>> [ 539.626204] #0: ffffffffb3b1c100 (rcu_read_lock){....}-{1:2}, at: debug_show_all_locks+0x32/0x1e0
>> [ 539.640561] 1 lock held by dmesg/640:
>> [ 539.654282] #0: ffff9fd541a8e0e0 (&user->lock){+.+.}-{3:3}, at: devkmsg_read+0x74/0x2d0
>> [ 539.669220] 2 locks held by fio/647:
>> [ 539.684253] #0: ffff9fd54fe720b0 (&ctx->uring_lock){+.+.}-{3:3}, at: __do_sys_io_uring_enter+0x5c2/0x820
>> [ 539.699565] #1: ffff9fd541a8d450 (sb_writers#15){++++}-{0:0}, at: io_issue_sqe+0x9c/0x780
>> [ 539.715587] 2 locks held by fio/648:
>> [ 539.732293] #0: ffff9fd54fe710b0 (&ctx->uring_lock){+.+.}-{3:3}, at: __do_sys_io_uring_enter+0x5c2/0x820
>> [ 539.749121] #1: ffff9fd541a8d450 (sb_writers#15){++++}-{0:0}, at: io_issue_sqe+0x9c/0x780
>> [ 539.765484] 2 locks held by fio/649:
>> [ 539.781483] #0: ffff9fd541a8f0b0 (&ctx->uring_lock){+.+.}-{3:3}, at: __do_sys_io_uring_enter+0x5c2/0x820
>> [ 539.798785] #1: ffff9fd541a8d450 (sb_writers#15){++++}-{0:0}, at: io_issue_sqe+0x9c/0x780
>> [ 539.815466] 2 locks held by fio/650:
>> [ 539.831966] #0: ffff9fd54fe740b0 (&ctx->uring_lock){+.+.}-{3:3}, at: __do_sys_io_uring_enter+0x5c2/0x820
>> [ 539.849527] #1: ffff9fd541a8d450 (sb_writers#15){++++}-{0:0}, at: io_issue_sqe+0x9c/0x780
>> [ 539.867469] 1 lock held by fsfreeze/696:
>> [ 539.884565] #0: ffff9fd541a8d450 (sb_writers#15){++++}-{0:0}, at: freeze_super+0x20a/0x600
>>
>> I reproduced this bug on nvme, sata ssd, virtio disks and lvm logical volumes.
>> It deadlocks on all kernels that I tried (all on amd64):
>> 6.12-rc5 (compiled from kernel.org)
>> 6.11.4 (compiled from kernel.org)
>> 6.10.11-1~bpo12+1 (debian)
>> 6.1.0-23 (debian)
>> 5.14.0-427.40.1.el9_4.x86_64 (rocky linux)
>> 5.10.0-33-amd64 (debian)
>>
>> I tried to compile some older ones to check if it's a regression, but
>> those either didn't compile or didn't boot in my VM, sorry about that.
>> If you have anything specific for me to try, I'm happy to help.
>>
>> Found this issue as well, so it seems like it's not just me:
>> https://gitlab.com/qemu-project/qemu/-/issues/881
>> Note that mariadb 10.6 adds support for io_uring, and that proxmox backups perform fsfreeze in the guest VM.
>>
>> Originally I discovered this after a scheduled lvm snapshot of mariadb
>> got stuck. It appears that lvm calls dm_suspend, which then calls
>> freeze_super, so it looks like the same bug to me. I discovered the
>> simpler fsfreeze/fio reproduction method when I tried to find a
>> workaround.
>
> Thanks for the report! I'm pretty sure this is due to the freezing not
> allowing task_work to run, which prevents completions from being run.
> Hence you run into a situation where freezing isn't running the very IO
> completions that will free up the rwsem, with IO issue being stuck on
> the freeze having started.
>
> I'll take a look...
Can you try the below? Probably easiest on 6.12-rc5 as you already
tested that and should apply directly.
diff --git a/io_uring/rw.c b/io_uring/rw.c
index 30448f343c7f..ea057ec4365f 100644
--- a/io_uring/rw.c
+++ b/io_uring/rw.c
@@ -1013,6 +1013,18 @@ int io_read_mshot(struct io_kiocb *req, unsigned int issue_flags)
return IOU_OK;
}
+static bool io_kiocb_start_write(struct io_kiocb *req, struct kiocb *kiocb)
+{
+ if (!(req->flags & REQ_F_ISREG))
+ return true;
+ if (!(kiocb->ki_flags & IOCB_NOWAIT)) {
+ kiocb_start_write(kiocb);
+ return true;
+ }
+
+ return sb_start_write_trylock(file_inode(kiocb->ki_filp)->i_sb);
+}
+
int io_write(struct io_kiocb *req, unsigned int issue_flags)
{
bool force_nonblock = issue_flags & IO_URING_F_NONBLOCK;
@@ -1050,8 +1062,8 @@ int io_write(struct io_kiocb *req, unsigned int issue_flags)
if (unlikely(ret))
return ret;
- if (req->flags & REQ_F_ISREG)
- kiocb_start_write(kiocb);
+ if (unlikely(!io_kiocb_start_write(req, kiocb)))
+ return -EAGAIN;
kiocb->ki_flags |= IOCB_WRITE;
if (likely(req->file->f_op->write_iter))
--
Jens Axboe
