Re: BUG: unable to handle kernel NULL pointer dereference at 00000000000000a8 in nilfs_segctor_do_co

[Ryusuke Konishi]

Hi,

This bug turned out to be caused by set_page_writeback() call for
segment summary buffers and super root buffers at
nilfs_segctor_prepare_write().

set_page_writeback() can call inc_wb_stat(inode_to_wb(inode),
WB_WRIEBACK) where inode_to_wb(inode) is NULL if inode_attach_wb() is
not called in advance.  To ensure inode_attach_wb() is called,
mark_buffer_dirty() should be called for those buffers.

The following patch fixes this issue, but I got another oops at
nilfs_segctor_complete_write() during a stress test.  So, I'm still
investigating.

Regards,
Ryusuke Konishi

===
diff --git a/fs/nilfs2/segment.c b/fs/nilfs2/segment.c
index 445eef4..f6b5ca8 100644
--- a/fs/nilfs2/segment.c
+++ b/fs/nilfs2/segment.c
@@ -1650,6 +1650,8 @@ static void nilfs_segctor_prepare_write(struct nilfs_sc_info *sci)
 
 		list_for_each_entry(bh, &segbuf->sb_segsum_buffers,
 				    b_assoc_buffers) {
+			set_buffer_uptodate(bh);
+			mark_buffer_dirty(bh);
 			if (bh->b_page != bd_page) {
 				if (bd_page) {
 					lock_page(bd_page);
@@ -1665,6 +1667,8 @@ static void nilfs_segctor_prepare_write(struct nilfs_sc_info *sci)
 				    b_assoc_buffers) {
 			set_buffer_async_write(bh);
 			if (bh == segbuf->sb_super_root) {
+				set_buffer_uptodate(bh);
+				mark_buffer_dirty(bh);
 				if (bh->b_page != bd_page) {
 					lock_page(bd_page);
 					clear_page_dirty_for_io(bd_page);
===


On Thu, 30 Apr 2020 08:27:47 -0700, Tom <tommytoad0@xxxxxxxxx> wrote:
> Thank you!  This is very helpful information, and does seem to be a
> workaround.
> 
> Like you, I have my home directory on a separate NILFS2 filesystem. As
> a temporary solution, I removed the line from /etc/fstab for that
> filesystem and added your dd suggestion along with a manual mount of
> the home filesystem to /etc/rc.local.  /home is now mounted properly
> at boot with any of the newer kernels I tried.
> 
> Thanks,
> Tom
> 
> On 4/30/20 5:38 AM, Hideki EIRAKU wrote:
>>> In Msg <874kuapb2s.fsf@xxxxxxxxxx>;
>>>     Subject "Re: BUG: unable to handle kernel NULL pointer dereference at
>>>     00000000000000a8 in nilfs_segctor_do_construct":
>>>
>>>> Tomas Hlavaty <tom@xxxxxxxxxx> writes:
>>>>>>> 2) Can you mount the corrupted(?) partition from a recent version of
>>>>>>> kernel ?
>>>>
>>>> I tried the following Linux kernel versions:
>>>>
>>>> - v4.19
>>>> - v5.4
>>>> - v5.5.11
>>>>
>>>> and still get the crash
>> I found conditions to reproduce this issue with Linux 5.7-rc3:
>> - CONFIG_MEMCG=y *and* CONFIG_BLK_CGROUP=y
>> - When the NILFS2 file system writes to a device, the device file has
>>    never written by other programs since boot
>> The following is an example with CONFIG_MEMCG=y and
>> CONFIG_BLK_CGROUP=y kernel.  If you do mkfs and mount it, it works
>> because the mkfs command has written data to the device file before
>> mounting:
>> # mkfs -t nilfs2 /dev/sda1
>> mkfs.nilfs2 (nilfs-utils 2.2.7)
>> Start writing file system initial data to the device
>>         Blocksize:4096  Device:/dev/sda1  Device Size:267386880
>> File system initialization succeeded !!
>> # mount /dev/sda1 /mnt
>> # touch /mnt
>> # sync
>> #
>> Loopback mount seems to be the same - if you do losetup, mkfs and
>> mount on a loopback device, it works:
>> # losetup /dev/loop0 foo
>> # mkfs -t nilfs2 /dev/loop0
>> mkfs.nilfs2 (nilfs-utils 2.2.7)
>> Start writing file system initial data to the device
>>         Blocksize:4096  Device:/dev/loop0  Device Size:267386880
>> File system initialization succeeded !!
>> # mount /dev/sda1 /mnt
>> # touch /mnt
>> # sync
>> #
>> But if you do mkfs on a file and use mount -o loop, it may fail,
>> depending on whether the loopback device assigned by the mount command
>> was used or not before mounting:
>> # /sbin/mkfs.nilfs2 ./foo
>> mkfs.nilfs2 (nilfs-utils 2.2.7)
>> Start writing file system initial data to the device
>>         Blocksize:4096  Device:./foo  Device Size:268435456
>> File system initialization succeeded !!
>> # mount -o loop ./foo /mnt
>> [ 36.371331] NILFS (loop0): segctord starting. Construction interval =
>> 5 seconds, CP frequency < 30 seconds
>> # touch /mnt
>> # sync
>> [ 40.252869] BUG: kernel NULL pointer dereference, address:
>> 00000000000000a8
>> (snip)
>> After reboot, it fails:
>> # mount /dev/sda1 /mnt
>> [ 14.021188] NILFS (sda1): segctord starting. Construction interval =
>> 5 seconds, CP frequency < 30 seconds
>> # touch /mnt
>> # sync
>> [ 20.576309] BUG: kernel NULL pointer dereference, address:
>> 00000000000000a8
>> (snip)
>> But if you do dummy write to the device file before mounting, it
>> works:
>> # dd if=/dev/sda1 of=/dev/sda1 count=1
>> 1+0 records in
>> 1+0 records out
>> 512 bytes copied, 0.0135982 s, 37.7 kB/s
>> # mount /dev/sda1 /mnt
>> [   52.604560] NILFS (sda1): mounting unchecked fs
>> [   52.613335] NILFS (sda1): recovery complete
>> [ 52.613877] NILFS (sda1): segctord starting. Construction interval =
>> 5 seconds, CP frequency < 30 seconds
>> # touch /mnt
>> # sync
>> #
>> # losetup /dev/loop0 foo
>> # dd if=/dev/loop0 of=/dev/loop0 count=1
>> 1+0 records in
>> 1+0 records out
>> 512 bytes copied, 0.0243797 s, 21.0 kB/s
>> # mount /dev/loop0 /mnt
>> [  271.915595] NILFS (loop0): mounting unchecked fs
>> [  272.049603] NILFS (loop0): recovery complete
>> [ 272.049724] NILFS (loop0): segctord starting. Construction interval
>> = 5 seconds, CP frequency < 30 seconds
>> # touch /mnt
>> # sync
>> #
>> I think the dummy write is a simple workaround for now, unless
>> mounting NILFS2 at boot time.  But I have been using NILFS2 /home for
>> years, I would like to know better workarounds.
>>