jbd2: fix outstanding credits assert in jbd2_journal_commit_transaction()

We catch an assert problem in jbd2_journal_commit_transaction() when
doing fsstress and request falut injection tests. The problem is
happened in a race condition between jbd2_journal_commit_transaction()
and ext4_end_io_end(). Firstly, ext4_writepages() writeback dirty pages
and start reserved handle, and then the journal was aborted due to some
previous metadata IO error, jbd2_journal_abort() start to commit current
running transaction, the committing procedure could be raced by
ext4_end_io_end() and lead to subtract j_reserved_credits twice from
commit_transaction->t_outstanding_credits, finally the
t_outstanding_credits is mistakenly smaller than t_nr_buffers and
trigger assert.

kjournald2           kworker

jbd2_journal_commit_transaction()
 write_unlock(&journal->j_state_lock);
 atomic_sub(j_reserved_credits, t_outstanding_credits); //sub once

     	             jbd2_journal_start_reserved()
     	              start_this_handle()  //detect aborted journal
     	              jbd2_journal_free_reserved()  //get running transaction
                       read_lock(&journal->j_state_lock)
     	                __jbd2_journal_unreserve_handle()
     	               atomic_sub(j_reserved_credits, t_outstanding_credits);
                       //sub again
                       read_unlock(&journal->j_state_lock);

 journal->j_running_transaction = NULL;
 J_ASSERT(t_nr_buffers <= t_outstanding_credits) //bomb!!!

Fix this issue by using journal->j_state_lock to protect the subtraction
in jbd2_journal_commit_transaction().

Fixes: 96f1e0974575 ("jbd2: avoid long hold times of j_state_lock while committing a transaction")
Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
---
 fs/jbd2/commit.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

On Sat 11-06-22 21:04:26, Zhang Yi wrote:
> We catch an assert problem in jbd2_journal_commit_transaction() when
> doing fsstress and request falut injection tests. The problem is
> happened in a race condition between jbd2_journal_commit_transaction()
> and ext4_end_io_end(). Firstly, ext4_writepages() writeback dirty pages
> and start reserved handle, and then the journal was aborted due to some
> previous metadata IO error, jbd2_journal_abort() start to commit current
> running transaction, the committing procedure could be raced by
> ext4_end_io_end() and lead to subtract j_reserved_credits twice from
> commit_transaction->t_outstanding_credits, finally the
> t_outstanding_credits is mistakenly smaller than t_nr_buffers and
> trigger assert.
> 
> kjournald2           kworker
> 
> jbd2_journal_commit_transaction()
>  write_unlock(&journal->j_state_lock);
>  atomic_sub(j_reserved_credits, t_outstanding_credits); //sub once
> 
>      	             jbd2_journal_start_reserved()
>      	              start_this_handle()  //detect aborted journal
>      	              jbd2_journal_free_reserved()  //get running transaction
>                        read_lock(&journal->j_state_lock)
>      	                __jbd2_journal_unreserve_handle()
>      	               atomic_sub(j_reserved_credits, t_outstanding_credits);
>                        //sub again
>                        read_unlock(&journal->j_state_lock);
> 
>  journal->j_running_transaction = NULL;
>  J_ASSERT(t_nr_buffers <= t_outstanding_credits) //bomb!!!
> 
> Fix this issue by using journal->j_state_lock to protect the subtraction
> in jbd2_journal_commit_transaction().
> 
> Fixes: 96f1e0974575 ("jbd2: avoid long hold times of j_state_lock while committing a transaction")
> Signed-off-by: Zhang Yi <yi.zhang@huawei.com>

Thanks for the analysis and the fix! This is indeed subtle. This fix looks
good to me. Feel free to add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza

> ---
>  fs/jbd2/commit.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c
> index eb315e81f1a6..af1a9191368c 100644
> --- a/fs/jbd2/commit.c
> +++ b/fs/jbd2/commit.c
> @@ -553,13 +553,13 @@ void jbd2_journal_commit_transaction(journal_t *journal)
>  	 */
>  	jbd2_journal_switch_revoke_table(journal);
>  
> +	write_lock(&journal->j_state_lock);
>  	/*
>  	 * Reserved credits cannot be claimed anymore, free them
>  	 */
>  	atomic_sub(atomic_read(&journal->j_reserved_credits),
>  		   &commit_transaction->t_outstanding_credits);
>  
> -	write_lock(&journal->j_state_lock);
>  	trace_jbd2_commit_flushing(journal, commit_transaction);
>  	stats.run.rs_flushing = jiffies;
>  	stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
> -- 
> 2.31.1
>

On Sat, 11 Jun 2022 21:04:26 +0800, Zhang Yi wrote:
> We catch an assert problem in jbd2_journal_commit_transaction() when
> doing fsstress and request falut injection tests. The problem is
> happened in a race condition between jbd2_journal_commit_transaction()
> and ext4_end_io_end(). Firstly, ext4_writepages() writeback dirty pages
> and start reserved handle, and then the journal was aborted due to some
> previous metadata IO error, jbd2_journal_abort() start to commit current
> running transaction, the committing procedure could be raced by
> ext4_end_io_end() and lead to subtract j_reserved_credits twice from
> commit_transaction->t_outstanding_credits, finally the
> t_outstanding_credits is mistakenly smaller than t_nr_buffers and
> trigger assert.
> 
> [...]

Applied, thanks!

[1/1] jbd2: fix outstanding credits assert in jbd2_journal_commit_transaction()
      commit: f8dc286e4d942dab79d1814e0708ac91052a34fa

Best regards,