Sorry, we didn’t notice this patch. It does fix the issue!

Thanks,
Sishuai

On Nov 6, 2020, at 6:15 PM, Song Liu [off-list ref] wrote:

On Fri, Nov 6, 2020 at 10:58 AM Gong, Sishuai [off-list ref] wrote:

quoted

Hi,

We found a concurrency bug in linux 5.3.11 that we were able to reproduce in x86 under specific interleavings. This bug causes a warning message “WARNING: linux-5.3.11/drivers/md/md.c:7279 md_ioctl+0x9cd/0x1b02”.

This bug is triggered when two kernel threads run the md_ioctl() function on the same resource interleave with each other. The code sets the mddev->flags to indicate that the resource is being modified and resets it after the modification. However, the current code allows another thread to execute after the mddev->flags is set but before it is reset, resulting in the warning message.

------------------------------------------
Kernel console output
[  140.524331] WARNING: CPU: 1 PID: 1815 at /tmp/tmp.B7zb7od2zE-5.3.11/extract/linux-5.3.11/drivers/md/md.c:7279 md_ioctl+0x9cd/0x1b02
[  145.438749] Modules linked in:
[  147.691130] CPU: 1 PID: 1815 Comm: ski-executor Not tainted 5.3.11 #1
[  150.333839] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007
[  153.712887] EIP: md_ioctl+0x9cd/0x1b02
[  157.464368] Code: ff ff ff e8 0f ed 91 ff c6 45 84 01 e9 10 ff ff ff 8d 83 74 01 00 00 e8 75 33 24 00 c6 45 84 00 be f0 ff ff ff e9 3e f7 ff ff <0f> 0b eb bf b0 00 eb 02 b0 01 84 c0 0f 84 2c f7 ff ff 89 7c 24 0c
[  168.813781] EAX: 00000002 EBX: f3df4800 ECX: f3df497c EDX: 00000002
[  171.890615] ESI: 00000000 EDI: 00000932 EBP: e527be2c ESP: e527bd98
[  175.465728] DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068 EFLAGS: 00000202
[  179.394439] CR0: 80050033 CR2: 08572568 CR3: 25242000 CR4: 00000690
[  183.140588] DR0: 00000000 DR1: 00000000 DR2: 00000000 DR3: 00000000
[  186.578976] DR6: 00000000 DR7: 00000000

------------------------------------------
Test input

The bug is triggered when the same kernel test program is executed concurrently by two different threads. In particular, it is triggered when the system call md_ioctl() interleaves with itself.

The test program is in Syzkaller’s format as follows:
r0 = openat$md(0xffffffffffffff9c, &(0x7f0000000000)='/dev/md0\x00', 0x0, 0x0)
ioctl$BLKTRACETEARDOWN(r0, 0x932, 0x0)



------------------------------------------
Interleaving

Our analysis revealed that the following interleaving can trigger this bug:

Thread 1                                                                                        Thread 2
                                                                                               md_open()
                                                                                               -if (test_bit(MD_CLOSING, &mddev->flags)) {
                                                                                                       mutex_unlock(&mddev->open_mutex);
                                                                                                       err = -ENODEV;
                                                                                                       goto out;
                                                                                                       }
                                                                                               (condition is false)
                                                                                               -…
                                                                                               -mutex_unlock(&mddev->open_mutex);
                                                                                               -…
                                                                                               -return err;
                                                                                               (md_open finishes correctly)
md_open()
-if (test_bit(MD_CLOSING, &mddev->flags)) {
       mutex_unlock(&mddev->open_mutex);
       err = -ENODEV;
       goto out;
}
(condition is false)
-...
-return err;
(md_open finishes correctly)

md_ioctl()
(drivers/md/md.c:7279)
-WARN_ON_ONCE(test_bit(MD_CLOSING, &mddev->flags));
-set_bit(MD_CLOSING, &mddev->flags);
-...
-mutex_unlock(&mddev->open_mutex);
                                                                                               md_ioctl()
                                                                                               (drivers/md/md.c:7279)
                                                                                               -WARN_ON_ONCE(test_bit(MD_CLOSING, &mddev->flags));
                                                                                               (warning message shows)
(drivers/md/md.c:7347)
-case STOP_ARRAY:
      err = do_md_stop(mddev, 0, bdev);
      goto unlock;
(mddev->flags will be cleared inside do_md_stop())

Thanks for the report. Could you please verify whether this commit
address this issue:

https://git.kernel.org/pub/scm/linux/kernel/git/song/md.git/commit/?h=md-next&id=e7f1456b5ee4e97934ae724e7015d95f88984df0

Thanks,
Song