On Thu 25-07-19 18:44:09, Chandan Rajendra wrote:
> When executing generic/388 on a ppc64le machine, we notice the following
> call trace,
>
> VFS: brelse: Trying to free free buffer
> WARNING: CPU: 0 PID: 6637 at /root/repos/linux/fs/buffer.c:1195 __brelse+0x84/0xc0
>
> Call Trace:
> __brelse+0x80/0xc0 (unreliable)
> invalidate_bh_lru+0x78/0xc0
> on_each_cpu_mask+0xa8/0x130
> on_each_cpu_cond_mask+0x130/0x170
> invalidate_bh_lrus+0x44/0x60
> invalidate_bdev+0x38/0x70
> ext4_put_super+0x294/0x560
> generic_shutdown_super+0xb0/0x170
> kill_block_super+0x38/0xb0
> deactivate_locked_super+0xa4/0xf0
> cleanup_mnt+0x164/0x1d0
> task_work_run+0x110/0x160
> do_notify_resume+0x414/0x460
> ret_from_except_lite+0x70/0x74
>
> The above occurs due to the following sequence of events,
> 1. Get a buffer head for holding a descriptor buffer for a Revoke
> record. This causes buffer_head->b_count to have a value of 2,
> - The first ref count corresponds to JBD code holding a reference to the
> buffer head.
> - Another ref count correponds to adding the buffer head to the
> per-cpu LRU list.
> The buffer head is also added to the list of log_bufs tracking the
> buffer heads that need to be written to the on-disk journal.
> 2. When writing the revoke record to the disk, if journal is aborted,
> flush_descriptor() gives up one ref count by invoking put_bh().
> 3. jbd2_journal_commit_transaction() then loops across the buffer heads
> in the log_bufs list. While doing so, it decrements the buffer head's ref
> count to 0 by invoking __brelse().
> 4. On unmount, invalidate_bdev() invokes __brelse() via
> invalidate_bh_lru() to remove the buffer head from the per-cpu LRU
> list. Here __brelse() prints the call trace shown above since
> buffer_head->b_count already has a value of 0.
>
> Hence this commit removes the call to put_bh() inside
> flush_descriptor().
>
> Signed-off-by: Chandan Rajendra <chandan@xxxxxxxxxxxxx>
Thanks for the patch! It patch looks good, just the analysis above doesn't
really explain why it is correct to remove the put_bh() call in
flush_descriptor(). So I'd just replace the above points with:
The warning happens because flush_descriptor() drops bh reference it does
not own. The bh reference acquired by jbd2_journal_get_descriptor_buffer()
is owned by the log_bufs list and gets released when this list is
processed. The reference for doing IO is only acquired in
write_dirty_buffer() later in flush_descriptor().
With this updated feel free to add:
Reviewed-by: Jan Kara <jack@xxxxxxx>
Honza
> ---
> fs/jbd2/revoke.c | 4 +---
> 1 file changed, 1 insertion(+), 3 deletions(-)
>
> diff --git a/fs/jbd2/revoke.c b/fs/jbd2/revoke.c
> index 69b9bc329964..f08073d7bbf5 100644
> --- a/fs/jbd2/revoke.c
> +++ b/fs/jbd2/revoke.c
> @@ -638,10 +638,8 @@ static void flush_descriptor(journal_t *journal,
> {
> jbd2_journal_revoke_header_t *header;
>
> - if (is_journal_aborted(journal)) {
> - put_bh(descriptor);
> + if (is_journal_aborted(journal))
> return;
> - }
>
> header = (jbd2_journal_revoke_header_t *)descriptor->b_data;
> header->r_count = cpu_to_be32(offset);
> --
> 2.19.1
>
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
