md/bitmap: wait for bitmap writes to complete during the tear down sequence

NULL pointer dereference was observed in super_written() when it tries to access the mddev structure.

[The below stack trace is from an older kernel, but the problem described in this patch applies to the mainline kernel.]

... ...

bio in the above stack is a bitmap write whose completion is invoked after the tear down sequence sets the mddev structure to NULL in rdev.

During tear down, there is an attempt to flush the bitmap writes, but it doesn't fully wait for all the bitmap writes to complete. For instance, md_bitmap_flush() is called to flush the bitmap writes, but the last call to md_bitmap_daemon_work() in md_bitmap_flush() could generate new bitmap writes for which there is no explicit wait to complete those writes. This results in a kernel panic when the completion routine, super_written() is called which tries to reference mddev in the rdev that has been set to
NULL(in unbind_rdev_from_array() by tear down sequence).

The solution is to call md_bitmap_wait_writes() after the last call to md_bitmap_daemon_work() in md_bitmap_flush() to ensure there are no pending bitmap writes before proceeding with the tear down.

Signed-off-by Sudhakar Panneerselvam <[email protected]>

---
  drivers/md/md-bitmap.c | 1 +
  1 file changed, 1 insertion(+)

drivers/md/md-bitmap.c +0 -0

index 200c5d0f08bf..e0fdc3a090c5 100644

Hello Sudhakar,

First, let's discuss with master branch kernel.

What command or action stands for "tear down" ?
From your description, it very like ioctl STOP_ARRAY.
Your crash was related with super_written, which is the callback for updating array sb, not bitmap sb. in md_update_sb() there is a sync point md_super_wait(), which will guarantee all sb bios finished successfully.

for your patch, do you check md_bitmap_free, which already done the your patch's job.

the call flow:

```
do_md_stop //by STOP_ARRAY
  + __md_stop_writes()
  |  md_bitmap_flush
  |  md_update_sb
  |   + md_super_write
  |   |  bio->bi_end_io = super_written
  |   + md_super_wait(mddev) //wait for all bios done
  + __md_stop(mddev)
  |  md_bitmap_destroy(mddev);
  |   + md_bitmap_free //see below
  + ...

md_bitmap_free
{
    ...
     //do your patch job.
     /* Shouldn't be needed - but just in case.... */
     wait_event(bitmap->write_wait,
            atomic_read(&bitmap->pending_writes) == 0);
    ...
}
```

Would you share more analysis or test results for your patch?

Thanks,
Heming

Quoted 37 lines

Hello Sudhakar,

First, let's discuss with master branch kernel.

What command or action stands for "tear down" ?
From your description, it very like ioctl STOP_ARRAY.
Your crash was related with super_written, which is the callback for updating array sb, not bitmap sb. in md_update_sb() there is a sync point md_super_wait(), which will guarantee all sb bios finished successfully.

for your patch, do you check md_bitmap_free, which already done the your patch's job.

the call flow:

```
do_md_stop //by STOP_ARRAY
   + __md_stop_writes()
   |  md_bitmap_flush
   |  md_update_sb
   |   + md_super_write
   |   |  bio->bi_end_io = super_written
   |   + md_super_wait(mddev) //wait for all bios done
   + __md_stop(mddev)
   |  md_bitmap_destroy(mddev);
   |   + md_bitmap_free //see below
   + ...

md_bitmap_free
{
     ...
      //do your patch job.
      /* Shouldn't be needed - but just in case.... */
      wait_event(bitmap->write_wait,
             atomic_read(&bitmap->pending_writes) == 0);
     ...
}
```

Quoted 1 line

Would you share more analysis or test results for your patch?

Hello Heming,

Please find more details about our system configuration and the detailed sequence that led to the crash.

We have a RAID1 configured on an external imsm container. /proc/mdstat output looks like this:

***
md24 : active raid1 sdb[1] sda[0]

       141557760 blocks super external:/md127/0 [2/2] [UU]
       bitmap: 1/2 pages [4KB], 65536KB chunk

md127 : inactive sdb[1](S) sda[0](S)
10402 blocks super external:imsm
***

All our system partition is laid out on top of the above RAID1 device as shown below.

***
/dev/md24p5 on / type xfs (rw,relatime,attr2,inode64,noquota) /dev/md24p1 on /boot type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p15 on /home type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p12 on /var type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p16 on /tmp type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p11 on /var/log type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p14 on /var/log/audit type xfs (rw,nodev,relatime,attr2,inode64,noquota)
***

In such a configuration, the kernel panic described in this patch occurs during the "shutdown" of the server. Since the system partition is laid out on the RAID1, there could be write I/Os going to the RAID device as part of system log and filesystem activity that typically occur during the shutdown.

From the core dump, I see that, after filesystems are unmounted and killing all
other userspace processes, "mdmon" thread initiates the "stop" on the md device.

***
COMMAND: "mdmon"

    TASK: ffff8ff2b1663c80  [THREAD_INFO: ffff8ff2b1663c80]
     CPU: 30

STATE: TASK_UNINTERRUPTIBLE
crash> bt
PID: 7390 TASK: ffff8ff2b1663c80 CPU: 30 COMMAND: "mdmon"
#0 [ffffb99d4eacba30] __schedule at ffffffff91884acc
#1 [ffffb99d4eacbad0] schedule at ffffffff918850e6
#2 [ffffb99d4eacbae8] schedule_timeout at ffffffff91889616 #3 [ffffb99d4eacbb78] wait_for_completion at ffffffff91885d1b #4 [ffffb99d4eacbbe0] __wait_rcu_gp at ffffffff9110c123
#5 [ffffb99d4eacbc28] synchronize_sched at ffffffff9111008e #6 [ffffb99d4eacbc78] unbind_rdev_from_array at ffffffff9169640f #7 [ffffb99d4eacbcc0] do_md_stop at ffffffff9169f163
#8 [ffffb99d4eacbd58] array_state_store at ffffffff9169f644 #9 [ffffb99d4eacbd90] md_attr_store at ffffffff9169b71e
#10 [ffffb99d4eacbdc8] sysfs_kf_write at ffffffff91320edf
#11 [ffffb99d4eacbdd8] kernfs_fop_write at ffffffff913203c4 #12 [ffffb99d4eacbe18] __vfs_write at ffffffff9128fcfa
#13 [ffffb99d4eacbea0] vfs_write at ffffffff9128ffd2
#14 [ffffb99d4eacbee0] sys_write at ffffffff9129025c
#15 [ffffb99d4eacbf28] do_syscall_64 at ffffffff91003a39
#16 [ffffb99d4eacbf50] entry_SYSCALL_64_after_hwframe at ffffffff91a001b1

     RIP: 00007ff3bfc846fd  RSP: 00007ff3bf8a6cf0  RFLAGS: 00000293
     RAX: ffffffffffffffda  RBX: 000055e257ef0941  RCX: 00007ff3bfc846fd
     RDX: 0000000000000005  RSI: 000055e257ef0941  RDI: 0000000000000010
     RBP: 0000000000000010   R8: 0000000000000001   R9: 0000000000000000
     R10: 00000000ffffff01  R11: 0000000000000293  R12: 0000000000000001
     R13: 0000000000000000  R14: 0000000000000001  R15: 0000000000000000
     ORIG_RAX: 0000000000000001  CS: 0033  SS: 002b

crash>
***

While handling the "md" stop in the kernel, the code sequence based on the above md configuration is

do_md_stop
+ __md_stop_writes

  | + md_bitmap_flush
  | | + md_bitmap_daemon_work
  | | | + md_bitmap_wait_writes()
  | | | | (This wait is for bitmap writes that happened up until now and to avoid
  | | | |  overlapping with the new bitmap writes.)
  | | | | (wait flag is set to zero for async writes to bitmap.)
  | | | + write_page()
  | | | | | (For external bitmap, bitmap->storage.file is NULL, hence ends up
  | | | | |  calling write_sb_page())
  | | | | + write_sb_page()
  | | | | | + md_super_write()
  | | | | | | (Since the wait flag is false, the bitmap write is submitted
  | | | | | |  without waiting for it to complete.)
  | | | | | | + mddev->pending_writes is incremented, then the IO is submitted
  | | | | | |   and the call returns without waiting
  | | + md_bitmap_update_sb() - (This call simply returns because
  | | | | "bitmap->mddev->bitmap_info.external" is true for external bitmaps)
  | + md_update_sb() - (This won't be called as the call is conditional and the
  | | | | | | condition evaluates to false in my setup(see below for crash info)
  + __md_stop
  | + md_bitmap_destroy
  | | + md_bitmap_free
  | | | + wait_event(bitmap->write_wait,
  | | | |            atomic_read(&bitmap->pending_writes) == 0);
  | | | | bitmap->pending_writes is zero, but the mddev->pending_writes is 1,
  | | | | so this call returns immediately.
  | md detachment continues while there is pending mddev I/O cauing NULL pointer
  | derefence when the I/O callback, super_written is called.

***

         crash> struct mddev.bitmap_info.external 0xffff8ffb62f13800
                 bitmap_info.external = 1,
         crash> struct mddev.in_sync 0xffff8ffb62f13800
                 in_sync = 1
         crash> struct mddev.sb_flags 0xffff8ffb62f13800
                 sb_flags = 0
         crash> struct mddev.ro 0xffff8ffb62f13800
                 ro = 0
         crash> struct mddev.cluster_info 0xffff8ffb62f13800
                 cluster_info = 0x0
         crash>

***

Please review and let me know your thoughts.

Quoted 157 lines · Sudhakar Panneerselvam

Quoted 37 lines

Hello Sudhakar,

First, let's discuss with master branch kernel.

What command or action stands for "tear down" ?
From your description, it very like ioctl STOP_ARRAY.
Your crash was related with super_written, which is the callback for updating array sb, not bitmap sb. in md_update_sb() there is a sync point md_super_wait(), which will guarantee all sb bios finished successfully.

for your patch, do you check md_bitmap_free, which already done the your patch's job.

the call flow:

```
do_md_stop //by STOP_ARRAY
   + __md_stop_writes()
   |  md_bitmap_flush
   |  md_update_sb
   |   + md_super_write
   |   |  bio->bi_end_io = super_written
   |   + md_super_wait(mddev) //wait for all bios done
   + __md_stop(mddev)
   |  md_bitmap_destroy(mddev);
   |   + md_bitmap_free //see below
   + ...

md_bitmap_free
{
     ...
      //do your patch job.
      /* Shouldn't be needed - but just in case.... */
      wait_event(bitmap->write_wait,
             atomic_read(&bitmap->pending_writes) == 0);
     ...
}
```

Quoted 1 line

Would you share more analysis or test results for your patch?

Hello Heming,

Please find more details about our system configuration and the detailed sequence that led to the crash.

We have a RAID1 configured on an external imsm container. /proc/mdstat output looks like this:

***
md24 : active raid1 sdb[1] sda[0]

       141557760 blocks super external:/md127/0 [2/2] [UU]
       bitmap: 1/2 pages [4KB], 65536KB chunk

md127 : inactive sdb[1](S) sda[0](S)
10402 blocks super external:imsm
***

All our system partition is laid out on top of the above RAID1 device as shown below.

***
/dev/md24p5 on / type xfs (rw,relatime,attr2,inode64,noquota) /dev/md24p1 on /boot type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p15 on /home type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p12 on /var type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p16 on /tmp type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p11 on /var/log type xfs (rw,nodev,relatime,attr2,inode64,noquota) /dev/md24p14 on /var/log/audit type xfs (rw,nodev,relatime,attr2,inode64,noquota)
***

In such a configuration, the kernel panic described in this patch occurs during the "shutdown" of the server. Since the system partition is laid out on the RAID1, there could be write I/Os going to the RAID device as part of system log and filesystem activity that typically occur during the shutdown.

From the core dump, I see that, after filesystems are unmounted and killing all
other userspace processes, "mdmon" thread initiates the "stop" on the md device.

***
COMMAND: "mdmon"

    TASK: ffff8ff2b1663c80  [THREAD_INFO: ffff8ff2b1663c80]
     CPU: 30

STATE: TASK_UNINTERRUPTIBLE
crash> bt
PID: 7390 TASK: ffff8ff2b1663c80 CPU: 30 COMMAND: "mdmon"
#0 [ffffb99d4eacba30] __schedule at ffffffff91884acc
#1 [ffffb99d4eacbad0] schedule at ffffffff918850e6
#2 [ffffb99d4eacbae8] schedule_timeout at ffffffff91889616 #3 [ffffb99d4eacbb78] wait_for_completion at ffffffff91885d1b #4 [ffffb99d4eacbbe0] __wait_rcu_gp at ffffffff9110c123
#5 [ffffb99d4eacbc28] synchronize_sched at ffffffff9111008e #6 [ffffb99d4eacbc78] unbind_rdev_from_array at ffffffff9169640f #7 [ffffb99d4eacbcc0] do_md_stop at ffffffff9169f163
#8 [ffffb99d4eacbd58] array_state_store at ffffffff9169f644 #9 [ffffb99d4eacbd90] md_attr_store at ffffffff9169b71e
#10 [ffffb99d4eacbdc8] sysfs_kf_write at ffffffff91320edf
#11 [ffffb99d4eacbdd8] kernfs_fop_write at ffffffff913203c4 #12 [ffffb99d4eacbe18] __vfs_write at ffffffff9128fcfa
#13 [ffffb99d4eacbea0] vfs_write at ffffffff9128ffd2
#14 [ffffb99d4eacbee0] sys_write at ffffffff9129025c
#15 [ffffb99d4eacbf28] do_syscall_64 at ffffffff91003a39
#16 [ffffb99d4eacbf50] entry_SYSCALL_64_after_hwframe at ffffffff91a001b1

     RIP: 00007ff3bfc846fd  RSP: 00007ff3bf8a6cf0  RFLAGS: 00000293
     RAX: ffffffffffffffda  RBX: 000055e257ef0941  RCX: 00007ff3bfc846fd
     RDX: 0000000000000005  RSI: 000055e257ef0941  RDI: 0000000000000010
     RBP: 0000000000000010   R8: 0000000000000001   R9: 0000000000000000
     R10: 00000000ffffff01  R11: 0000000000000293  R12: 0000000000000001
     R13: 0000000000000000  R14: 0000000000000001  R15: 0000000000000000
     ORIG_RAX: 0000000000000001  CS: 0033  SS: 002b

crash>
***

While handling the "md" stop in the kernel, the code sequence based on the above md configuration is

do_md_stop
+ __md_stop_writes

  | + md_bitmap_flush
  | | + md_bitmap_daemon_work
  | | | + md_bitmap_wait_writes()
  | | | | (This wait is for bitmap writes that happened up until now and to avoid
  | | | |  overlapping with the new bitmap writes.)
  | | | | (wait flag is set to zero for async writes to bitmap.)
  | | | + write_page()
  | | | | | (For external bitmap, bitmap->storage.file is NULL, hence ends up
  | | | | |  calling write_sb_page())
  | | | | + write_sb_page()
  | | | | | + md_super_write()
  | | | | | | (Since the wait flag is false, the bitmap write is submitted
  | | | | | |  without waiting for it to complete.)
  | | | | | | + mddev->pending_writes is incremented, then the IO is submitted
  | | | | | |   and the call returns without waiting
  | | + md_bitmap_update_sb() - (This call simply returns because
  | | | | "bitmap->mddev->bitmap_info.external" is true for external bitmaps)
  | + md_update_sb() - (This won't be called as the call is conditional and the
  | | | | | | condition evaluates to false in my setup(see below for crash info)
  + __md_stop
  | + md_bitmap_destroy
  | | + md_bitmap_free
  | | | + wait_event(bitmap->write_wait,
  | | | |            atomic_read(&bitmap->pending_writes) == 0);
  | | | | bitmap->pending_writes is zero, but the mddev->pending_writes is 1,
  | | | | so this call returns immediately.
  | md detachment continues while there is pending mddev I/O cauing NULL pointer
  | derefence when the I/O callback, super_written is called.

***

         crash> struct mddev.bitmap_info.external 0xffff8ffb62f13800
                 bitmap_info.external = 1,
         crash> struct mddev.in_sync 0xffff8ffb62f13800
                 in_sync = 1
         crash> struct mddev.sb_flags 0xffff8ffb62f13800
                 sb_flags = 0
         crash> struct mddev.ro 0xffff8ffb62f13800
                 ro = 0
         crash> struct mddev.cluster_info 0xffff8ffb62f13800
                 cluster_info = 0x0
         crash>

Quoted 4 lines

***

Please review and let me know your thoughts.

Hello

On previous mail, I was wrong to assume the array mddev->bitmap_info.external is 0. So my analysis was not suite for your case.

Now we understand your env, which is related with IMSM & external bitmap, and I have a little knowledge about this field. May need to add more people in the cc list.

Just from code logic, your analysis is reasonable.
The key of the crash is why md layer didn't do the waiting job for bitmap bios. There have wait points in md_bitmap_daemon_work, write_sb_page, md_bitmap_update_sb & md_update_sb. But in your scenario there all didn't take effect. At last md_bitmap_free do the additional wait bitmap->pending_writes but not mddev->pending_writes. the crash is triggered.

In my opinion, using a special wait is more clear than calling general md_bitmap_wait_writes(). the md_bitmap_wait_writes makes people feel bitmap module does repetitive clean job.

My idea like:
```

drivers/md/md-bitmap.c +3 -0

index 200c5d0f08bf..ea6fa5a2cb6b 100644
@@ -1723,6 +1723,8 @@ void md_bitmap_flush(struct mddev *mddev)
         bitmap->daemon_lastrun -= sleep;
         md_bitmap_daemon_work(mddev);
         md_bitmap_update_sb(bitmap);
+       if (mddev->bitmap_info.external)
+               md_super_wait(mddev);
  }
 
  /*

@@ -1746,6 +1748,7 @@ void md_bitmap_free(struct bitmap *bitmap)
         /* Shouldn't be needed - but just in case.... */
         wait_event(bitmap->write_wait,
                    atomic_read(&bitmap->pending_writes) == 0);
+       wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);
 
         /* release the bitmap file  */
         md_bitmap_file_unmap(&bitmap->storage);

```

Quoted 1 line

In my opinion, using a special wait is more clear than calling general md_bitmap_wait_writes(). the md_bitmap_wait_writes makes people

feel bitmap module does repetitive clean job.

drivers/md/md-bitmap.c void md_bitmap_flush(struct mddev *mddev) · line 1727

Quoted 13 lines

My idea like:
```

drivers/md/md-bitmap.c +2 -0

index 200c5d0f08bf..ea6fa5a2cb6b 100644
@@ -1723,6 +1723,8 @@ void md_bitmap_flush(struct mddev *mddev)
         bitmap->daemon_lastrun -= sleep;
         md_bitmap_daemon_work(mddev);
         md_bitmap_update_sb(bitmap);
+       if (mddev->bitmap_info.external)
+               md_super_wait(mddev);

drivers/md/md-bitmap.c void md_bitmap_free(struct bitmap *bitmap) · line 1751

Quoted 8 lines

drivers/md/md-bitmap.c +1 -0

  }
 
  /*

@@ -1746,6 +1748,7 @@ void md_bitmap_free(struct bitmap *bitmap)
         /* Shouldn't be needed - but just in case.... */
         wait_event(bitmap->write_wait,
                    atomic_read(&bitmap->pending_writes) == 0);
+       wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);

Quoted 4 lines

drivers/md/md-bitmap.c +0 -0

         /* release the bitmap file  */
         md_bitmap_file_unmap(&bitmap->storage);

```

I like Heming's version better.

Hi Sudhakar,

Are you able to reproduce the issue? If so, could you please verify that Heming's
change fixes the issue?

Thanks,
Song

Quoted 15 lines

Quoted 1 line

In my opinion, using a special wait is more clear than calling general md_bitmap_wait_writes(). the md_bitmap_wait_writes makes people

feel bitmap module does repetitive clean job.

drivers/md/md-bitmap.c void md_bitmap_flush(struct mddev *mddev) · line 1727

Quoted 13 lines

My idea like:
```

drivers/md/md-bitmap.c +2 -0

index 200c5d0f08bf..ea6fa5a2cb6b 100644
@@ -1723,6 +1723,8 @@ void md_bitmap_flush(struct mddev *mddev)
          bitmap->daemon_lastrun -= sleep;
          md_bitmap_daemon_work(mddev);
          md_bitmap_update_sb(bitmap);
+       if (mddev->bitmap_info.external)
+               md_super_wait(mddev);

Agreed. This looks much cleaner.

Quoted 2 lines

Quoted 2 lines

drivers/md/md-bitmap.c +0 -0

   }
 

Quoted 6 lines

drivers/md/md-bitmap.c void md_bitmap_free(struct bitmap *bitmap) · line 1751

Quoted 6 lines

drivers/md/md-bitmap.c +1 -0

   /*

@@ -1746,6 +1748,7 @@ void md_bitmap_free(struct bitmap *bitmap)
          /* Shouldn't be needed - but just in case.... */
          wait_event(bitmap->write_wait,
                     atomic_read(&bitmap->pending_writes) == 0);
+       wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);

I think this call looks redundant as this wait is already covered by md_update_sb() for non-external bitmaps and the proposed change in md_bitmap_flush() for external bitmaps. So, can we omit this change?

-----Original Message-----
From: [email protected] [mailto:[email protected]]
Sent: Monday, April 12, 2021 3:59 AM
To: Sudhakar Panneerselvam <[email protected]>; Song Liu <[email protected]>
Cc: [email protected]; [email protected]; [email protected]; [email protected]; Martin Petersen <[email protected]>
Subject: Re: [PATCH] md/bitmap: wait for bitmap writes to complete during the tear down sequence

Quoted 1 line · Sudhakar Panneerselvam

Quoted 1 line

Quoted 1 line

In my opinion, using a special wait is more clear than calling general md_bitmap_wait_writes(). the md_bitmap_wait_writes makes

people

Quoted 19 lines

Quoted 14 lines

feel bitmap module does repetitive clean job.

drivers/md/md-bitmap.c void md_bitmap_flush(struct mddev *mddev) · line 1727

Quoted 13 lines

My idea like:
```

drivers/md/md-bitmap.c +2 -0

index 200c5d0f08bf..ea6fa5a2cb6b 100644
@@ -1723,6 +1723,8 @@ void md_bitmap_flush(struct mddev *mddev)
          bitmap->daemon_lastrun -= sleep;
          md_bitmap_daemon_work(mddev);
          md_bitmap_update_sb(bitmap);
+       if (mddev->bitmap_info.external)
+               md_super_wait(mddev);

Agreed. This looks much cleaner.

Quoted 2 lines

Quoted 2 lines

drivers/md/md-bitmap.c +0 -0

   }
 

Quoted 8 lines

Quoted 6 lines

drivers/md/md-bitmap.c void md_bitmap_free(struct bitmap *bitmap) · line 1751

Quoted 6 lines

drivers/md/md-bitmap.c +1 -0

   /*

@@ -1746,6 +1748,7 @@ void md_bitmap_free(struct bitmap *bitmap)
          /* Shouldn't be needed - but just in case.... */
          wait_event(bitmap->write_wait,
                     atomic_read(&bitmap->pending_writes) == 0);
+       wait_event(mddev->sb_wait, atomic_read(&mddev->pending_writes)==0);

I think this call looks redundant as this wait is already covered by md_update_sb() for non-external bitmaps and the proposed change in

md_bitmap_flush() for external bitmaps. So, can we omit this change?

Yes, it's absolute redundant step, to add or remove this line is up to you.
I added this line for following the style of bitmap->pending_writes. The comment in this area also give a explanation.