[RFC] Fix data missing when reusing bh which is ready to be checkpointed

Message ID 20221220150551.653925-1-chengzhihao1@huawei.com
State New
Headers
Series [RFC] Fix data missing when reusing bh which is ready to be checkpointed |

Commit Message

Zhihao Cheng Dec. 20, 2022, 3:05 p.m. UTC
  From: zhanchengbin <zhanchengbin1@huawei.com>

Following process will make data lost and could lead to a filesystem
corrupted problem:

1. jh(bh) is inserted into T1->t_checkpoint_list, bh is dirty, and
   jh->b_transaction = NULL
2. T1 is added into journal->j_checkpoint_transactions.
3. Get bh prepare to write while doing checkpoing:
           PA				    PB
   do_get_write_access             jbd2_log_do_checkpoint
    spin_lock(&jh->b_state_lock)
     if (buffer_dirty(bh))
      clear_buffer_dirty(bh)   // clear buffer dirty
       set_buffer_jbddirty(bh)
				    transaction =
				    journal->j_checkpoint_transactions
				    jh = transaction->t_checkpoint_list
				    if (!buffer_dirty(bh))
		                      __jbd2_journal_remove_checkpoint(jh)
				      // bh won't be flushed
		                    jbd2_cleanup_journal_tail
    __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved)
4. Aborting journal/Power-cut before writing latest bh on journal area.

In this way we get a corrupted filesystem with bh'data lost.

Fix it by wrapping clear_buffer_dirty(bh) and jh->b_transaction setting
into journal->j_list_lock, so that jbd2_log_do_checkpoint() will wait
until jh's new transaction fininshed even bh is currently not dirty.

Cc: <stable@kernel.org>
Signed-off-by: zhanchengbin <zhanchengbin1@huawei.com>
Signed-off-by: Baokun Li <libaokun1@huawei.com>
Signed-off-by: Zhihao Cheng <chengzhihao1@huawei.com>
---
 This is a quick fix, I need some suggestions about this patch, whether
 it will import new problems if this patch is applied.
 Yi suggests that the formal solution could be splitting
 journal->j_list_lock into two locks: one protects checkpoint list and
 the other one for other lists. Besides, jh->b_state_lock should be
 held while traversing transaction->t_checkpoint_list in
 jbd2_log_do_checkpoint()/journal_shrink_one_cp_list().

 fs/jbd2/transaction.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
  

Comments

Jan Kara Dec. 21, 2022, 10:13 a.m. UTC | #1
On Tue 20-12-22 23:05:51, Zhihao Cheng wrote:
> From: zhanchengbin <zhanchengbin1@huawei.com>
> 
> Following process will make data lost and could lead to a filesystem
> corrupted problem:
> 
> 1. jh(bh) is inserted into T1->t_checkpoint_list, bh is dirty, and
>    jh->b_transaction = NULL
> 2. T1 is added into journal->j_checkpoint_transactions.
> 3. Get bh prepare to write while doing checkpoing:
>            PA				    PB
>    do_get_write_access             jbd2_log_do_checkpoint
>     spin_lock(&jh->b_state_lock)
>      if (buffer_dirty(bh))
>       clear_buffer_dirty(bh)   // clear buffer dirty
>        set_buffer_jbddirty(bh)
> 				    transaction =
> 				    journal->j_checkpoint_transactions
> 				    jh = transaction->t_checkpoint_list
> 				    if (!buffer_dirty(bh))
> 		                      __jbd2_journal_remove_checkpoint(jh)
> 				      // bh won't be flushed
> 		                    jbd2_cleanup_journal_tail
>     __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved)
> 4. Aborting journal/Power-cut before writing latest bh on journal area.
> 
> In this way we get a corrupted filesystem with bh'data lost.
> 
> Fix it by wrapping clear_buffer_dirty(bh) and jh->b_transaction setting
> into journal->j_list_lock, so that jbd2_log_do_checkpoint() will wait
> until jh's new transaction fininshed even bh is currently not dirty.
> 
> Cc: <stable@kernel.org>
> Signed-off-by: zhanchengbin <zhanchengbin1@huawei.com>
> Signed-off-by: Baokun Li <libaokun1@huawei.com>
> Signed-off-by: Zhihao Cheng <chengzhihao1@huawei.com>
> ---
>  This is a quick fix, I need some suggestions about this patch, whether
>  it will import new problems if this patch is applied.
>  Yi suggests that the formal solution could be splitting
>  journal->j_list_lock into two locks: one protects checkpoint list and
>  the other one for other lists. Besides, jh->b_state_lock should be
>  held while traversing transaction->t_checkpoint_list in
>  jbd2_log_do_checkpoint()/journal_shrink_one_cp_list().
> 
>  fs/jbd2/transaction.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)

Good catch! Did you find it by code inspection or were you able to actually
trigger this problem?

I think there might be a simpler fix of the problem. Move the clearing
of buffer_dirty bit just before the call to __jbd2_journal_file_buffer().
We'll need to keep the buffer locked somewhat longer but that should not be
a huge deal.

								Honza


> diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
> index 6a404ac1c178..d22460001d6b 100644
> --- a/fs/jbd2/transaction.c
> +++ b/fs/jbd2/transaction.c
> @@ -990,6 +990,7 @@ do_get_write_access(handle_t *handle, struct journal_head *jh,
>   	start_lock = jiffies;
>  	lock_buffer(bh);
>  	spin_lock(&jh->b_state_lock);
> +	spin_lock(&journal->j_list_lock);
>  
>  	/* If it takes too long to lock the buffer, trace it */
>  	time_lock = jbd2_time_diff(start_lock, jiffies);
> @@ -1039,6 +1040,7 @@ do_get_write_access(handle_t *handle, struct journal_head *jh,
>  
>  	error = -EROFS;
>  	if (is_handle_aborted(handle)) {
> +		spin_unlock(&journal->j_list_lock);
>  		spin_unlock(&jh->b_state_lock);
>  		goto out;
>  	}
> @@ -1049,8 +1051,10 @@ do_get_write_access(handle_t *handle, struct journal_head *jh,
>  	 * b_next_transaction points to it
>  	 */
>  	if (jh->b_transaction == transaction ||
> -	    jh->b_next_transaction == transaction)
> +	    jh->b_next_transaction == transaction) {
> +		spin_unlock(&journal->j_list_lock);
>  		goto done;
> +	}
>  
>  	/*
>  	 * this is the first time this transaction is touching this buffer,
> @@ -1073,11 +1077,11 @@ do_get_write_access(handle_t *handle, struct journal_head *jh,
>  		 * Paired with barrier in jbd2_write_access_granted()
>  		 */
>  		smp_wmb();
> -		spin_lock(&journal->j_list_lock);
>  		__jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
>  		spin_unlock(&journal->j_list_lock);
>  		goto done;
>  	}
> +	spin_unlock(&journal->j_list_lock);
>  	/*
>  	 * If there is already a copy-out version of this buffer, then we don't
>  	 * need to make another one
> -- 
> 2.31.1
>
  
Zhihao Cheng Jan. 7, 2023, 9:28 a.m. UTC | #2
在 2022/12/21 18:13, Jan Kara 写道:
> On Tue 20-12-22 23:05:51, Zhihao Cheng wrote:
>> From: zhanchengbin <zhanchengbin1@huawei.com>
>>
>> Following process will make data lost and could lead to a filesystem
>> corrupted problem:
>>
>> 1. jh(bh) is inserted into T1->t_checkpoint_list, bh is dirty, and
>>     jh->b_transaction = NULL
>> 2. T1 is added into journal->j_checkpoint_transactions.
>> 3. Get bh prepare to write while doing checkpoing:
>>             PA				    PB
>>     do_get_write_access             jbd2_log_do_checkpoint
>>      spin_lock(&jh->b_state_lock)
>>       if (buffer_dirty(bh))
>>        clear_buffer_dirty(bh)   // clear buffer dirty
>>         set_buffer_jbddirty(bh)
>> 				    transaction =
>> 				    journal->j_checkpoint_transactions
>> 				    jh = transaction->t_checkpoint_list
>> 				    if (!buffer_dirty(bh))
>> 		                      __jbd2_journal_remove_checkpoint(jh)
>> 				      // bh won't be flushed
>> 		                    jbd2_cleanup_journal_tail
>>      __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved)
>> 4. Aborting journal/Power-cut before writing latest bh on journal area.

[...]
>>
>>   fs/jbd2/transaction.c | 8 ++++++--
>>   1 file changed, 6 insertions(+), 2 deletions(-)
> 
> Good catch! Did you find it by code inspection or were you able to actually
> trigger this problem?

By code inspection.
Reproducer: https://bugzilla.kernel.org/show_bug.cgi?id=216898
  

Patch

diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
index 6a404ac1c178..d22460001d6b 100644
--- a/fs/jbd2/transaction.c
+++ b/fs/jbd2/transaction.c
@@ -990,6 +990,7 @@  do_get_write_access(handle_t *handle, struct journal_head *jh,
  	start_lock = jiffies;
 	lock_buffer(bh);
 	spin_lock(&jh->b_state_lock);
+	spin_lock(&journal->j_list_lock);
 
 	/* If it takes too long to lock the buffer, trace it */
 	time_lock = jbd2_time_diff(start_lock, jiffies);
@@ -1039,6 +1040,7 @@  do_get_write_access(handle_t *handle, struct journal_head *jh,
 
 	error = -EROFS;
 	if (is_handle_aborted(handle)) {
+		spin_unlock(&journal->j_list_lock);
 		spin_unlock(&jh->b_state_lock);
 		goto out;
 	}
@@ -1049,8 +1051,10 @@  do_get_write_access(handle_t *handle, struct journal_head *jh,
 	 * b_next_transaction points to it
 	 */
 	if (jh->b_transaction == transaction ||
-	    jh->b_next_transaction == transaction)
+	    jh->b_next_transaction == transaction) {
+		spin_unlock(&journal->j_list_lock);
 		goto done;
+	}
 
 	/*
 	 * this is the first time this transaction is touching this buffer,
@@ -1073,11 +1077,11 @@  do_get_write_access(handle_t *handle, struct journal_head *jh,
 		 * Paired with barrier in jbd2_write_access_granted()
 		 */
 		smp_wmb();
-		spin_lock(&journal->j_list_lock);
 		__jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
 		spin_unlock(&journal->j_list_lock);
 		goto done;
 	}
+	spin_unlock(&journal->j_list_lock);
 	/*
 	 * If there is already a copy-out version of this buffer, then we don't
 	 * need to make another one