[v2] ext4: fix insufficient credits when writing back large folios

[PATCH v2 3/6] ext4: restart handle if credits are insufficient during allocating blocks

Posted by Zhang Yi 8 months ago

From: Zhang Yi <yi.zhang@huawei.com>

After large folios are supported on ext4, writing back a sufficiently
large and discontinuous folio may consume a significant number of
journal credits, placing considerable strain on the journal. For
example, in a 20GB filesystem with 1K block size and 1MB journal size,
writing back a 2MB folio could require thousands of credits in the
worst-case scenario (when each block is discontinuous and distributed
across different block groups), potentially exceeding the journal size.
This issue can also occur in ext4_write_begin() and ext4_page_mkwrite()
when delalloc is not enabled.

Fix this by ensuring that there are sufficient journal credits before
allocating an extent in mpage_map_one_extent() and _ext4_get_block(). If
there are not enough credits, return -EAGAIN, exit the current mapping
loop, restart a new handle and a new transaction, and allocating blocks
on this folio again in the next iteration.

Suggested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
---
 fs/ext4/inode.c | 45 +++++++++++++++++++++++++++++++++++++++------
 1 file changed, 39 insertions(+), 6 deletions(-)

diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index d0db6e3bf158..b51de58518b2 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -877,20 +877,44 @@ static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
 	} while (unlikely(!try_cmpxchg(&bh->b_state, &old_state, new_state)));
 }
 
+/*
+ * Make sure that the current journal transaction has enough credits to map
+ * one extent. Return -EAGAIN if it cannot extend the current running
+ * transaction.
+ */
+static inline int ext4_journal_ensure_extent_credits(handle_t *handle,
+						     struct inode *inode)
+{
+	int needed_credits;
+	int ret;
+
+	needed_credits = ext4_chunk_trans_blocks(inode, 1);
+	ret = __ext4_journal_ensure_credits(handle, needed_credits,
+					    needed_credits, 0);
+	return ret <= 0 ? ret : -EAGAIN;
+}
+
 static int _ext4_get_block(struct inode *inode, sector_t iblock,
 			   struct buffer_head *bh, int flags)
 {
 	struct ext4_map_blocks map;
+	handle_t *handle = ext4_journal_current_handle();
 	int ret = 0;
 
 	if (ext4_has_inline_data(inode))
 		return -ERANGE;
 
+	/* Make sure transaction has enough credits for this extent */
+	if (flags & EXT4_GET_BLOCKS_CREATE) {
+		ret = ext4_journal_ensure_extent_credits(handle, inode);
+		if (ret)
+			return ret;
+	}
+
 	map.m_lblk = iblock;
 	map.m_len = bh->b_size >> inode->i_blkbits;
 
-	ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
-			      flags);
+	ret = ext4_map_blocks(handle, inode, &map, flags);
 	if (ret > 0) {
 		map_bh(bh, inode->i_sb, map.m_pblk);
 		ext4_update_bh_state(bh, map.m_flags);
@@ -1374,8 +1398,9 @@ static int ext4_write_begin(struct file *file, struct address_space *mapping,
 				ext4_orphan_del(NULL, inode);
 		}
 
-		if (ret == -ENOSPC &&
-		    ext4_should_retry_alloc(inode->i_sb, &retries))
+		if (ret == -EAGAIN ||
+		    (ret == -ENOSPC &&
+		     ext4_should_retry_alloc(inode->i_sb, &retries)))
 			goto retry_journal;
 		folio_put(folio);
 		return ret;
@@ -2324,6 +2349,11 @@ static int mpage_map_one_extent(handle_t *handle, struct mpage_da_data *mpd)
 	int get_blocks_flags;
 	int err, dioread_nolock;
 
+	/* Make sure transaction has enough credits for this extent */
+	err = ext4_journal_ensure_extent_credits(handle, inode);
+	if (err < 0)
+		return err;
+
 	trace_ext4_da_write_pages_extent(inode, map);
 	/*
 	 * Call ext4_map_blocks() to allocate any delayed allocation blocks, or
@@ -2446,7 +2476,7 @@ static int mpage_map_and_submit_extent(handle_t *handle,
 			 * In the case of ENOSPC, if ext4_count_free_blocks()
 			 * is non-zero, a commit should free up blocks.
 			 */
-			if ((err == -ENOMEM) ||
+			if ((err == -ENOMEM) || (err == -EAGAIN) ||
 			    (err == -ENOSPC && ext4_count_free_clusters(sb))) {
 				/*
 				 * We may have already allocated extents for
@@ -2953,6 +2983,8 @@ static int ext4_do_writepages(struct mpage_da_data *mpd)
 			ret = 0;
 			continue;
 		}
+		if (ret == -EAGAIN)
+			ret = 0;
 		/* Fatal error - ENOMEM, EIO... */
 		if (ret)
 			break;
@@ -6722,7 +6754,8 @@ vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
 		}
 	}
 	ext4_journal_stop(handle);
-	if (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
+	if (err == -EAGAIN ||
+	    (err == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries)))
 		goto retry_alloc;
 out_ret:
 	ret = vmf_fs_error(err);
-- 
2.46.1

Re: [PATCH v2 3/6] ext4: restart handle if credits are insufficient during allocating blocks

Posted by Jan Kara 7 months, 3 weeks ago

On Wed 11-06-25 19:16:22, Zhang Yi wrote:
> From: Zhang Yi <yi.zhang@huawei.com>
> 
> After large folios are supported on ext4, writing back a sufficiently
> large and discontinuous folio may consume a significant number of
> journal credits, placing considerable strain on the journal. For
> example, in a 20GB filesystem with 1K block size and 1MB journal size,
> writing back a 2MB folio could require thousands of credits in the
> worst-case scenario (when each block is discontinuous and distributed
> across different block groups), potentially exceeding the journal size.
> This issue can also occur in ext4_write_begin() and ext4_page_mkwrite()
> when delalloc is not enabled.
> 
> Fix this by ensuring that there are sufficient journal credits before
> allocating an extent in mpage_map_one_extent() and _ext4_get_block(). If
> there are not enough credits, return -EAGAIN, exit the current mapping
> loop, restart a new handle and a new transaction, and allocating blocks
> on this folio again in the next iteration.
> 
> Suggested-by: Jan Kara <jack@suse.cz>
> Signed-off-by: Zhang Yi <yi.zhang@huawei.com>

...

>  static int _ext4_get_block(struct inode *inode, sector_t iblock,
>  			   struct buffer_head *bh, int flags)
>  {
>  	struct ext4_map_blocks map;
> +	handle_t *handle = ext4_journal_current_handle();
>  	int ret = 0;
>  
>  	if (ext4_has_inline_data(inode))
>  		return -ERANGE;
>  
> +	/* Make sure transaction has enough credits for this extent */
> +	if (flags & EXT4_GET_BLOCKS_CREATE) {
> +		ret = ext4_journal_ensure_extent_credits(handle, inode);
> +		if (ret)
> +			return ret;
> +	}
> +
>  	map.m_lblk = iblock;
>  	map.m_len = bh->b_size >> inode->i_blkbits;
>  
> -	ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
> -			      flags);
> +	ret = ext4_map_blocks(handle, inode, &map, flags);

Good spotting with ext4_page_mkwrite() and ext4_write_begin() also needing
this treatment! But rather then hiding the transaction extension in
_ext4_get_block() I'd do this in ext4_block_write_begin() where it is much
more obvious (and also it is much more obvious who needs to be prepared for
handling EAGAIN error). Otherwise the patch looks good!

								Honza

-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [PATCH v2 3/6] ext4: restart handle if credits are insufficient during allocating blocks

Posted by Zhang Yi 7 months, 3 weeks ago

On 2025/6/20 0:33, Jan Kara wrote:
> On Wed 11-06-25 19:16:22, Zhang Yi wrote:
>> From: Zhang Yi <yi.zhang@huawei.com>
>>
>> After large folios are supported on ext4, writing back a sufficiently
>> large and discontinuous folio may consume a significant number of
>> journal credits, placing considerable strain on the journal. For
>> example, in a 20GB filesystem with 1K block size and 1MB journal size,
>> writing back a 2MB folio could require thousands of credits in the
>> worst-case scenario (when each block is discontinuous and distributed
>> across different block groups), potentially exceeding the journal size.
>> This issue can also occur in ext4_write_begin() and ext4_page_mkwrite()
>> when delalloc is not enabled.
>>
>> Fix this by ensuring that there are sufficient journal credits before
>> allocating an extent in mpage_map_one_extent() and _ext4_get_block(). If
>> there are not enough credits, return -EAGAIN, exit the current mapping
>> loop, restart a new handle and a new transaction, and allocating blocks
>> on this folio again in the next iteration.
>>
>> Suggested-by: Jan Kara <jack@suse.cz>
>> Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
> 
> ...
> 
>>  static int _ext4_get_block(struct inode *inode, sector_t iblock,
>>  			   struct buffer_head *bh, int flags)
>>  {
>>  	struct ext4_map_blocks map;
>> +	handle_t *handle = ext4_journal_current_handle();
>>  	int ret = 0;
>>  
>>  	if (ext4_has_inline_data(inode))
>>  		return -ERANGE;
>>  
>> +	/* Make sure transaction has enough credits for this extent */
>> +	if (flags & EXT4_GET_BLOCKS_CREATE) {
>> +		ret = ext4_journal_ensure_extent_credits(handle, inode);
>> +		if (ret)
>> +			return ret;
>> +	}
>> +
>>  	map.m_lblk = iblock;
>>  	map.m_len = bh->b_size >> inode->i_blkbits;
>>  
>> -	ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
>> -			      flags);
>> +	ret = ext4_map_blocks(handle, inode, &map, flags);
> 
> Good spotting with ext4_page_mkwrite() and ext4_write_begin() also needing
> this treatment! But rather then hiding the transaction extension in
> _ext4_get_block() I'd do this in ext4_block_write_begin() where it is much
> more obvious (and also it is much more obvious who needs to be prepared for
> handling EAGAIN error). Otherwise the patch looks good!
> 

Yes, I completely agree with you. However, unfortunately, do this in
ext4_block_write_begin() only works for ext4_write_begin().
ext4_page_mkwrite() does not call ext4_block_write_begin() to allocate
blocks, it call the vfs helper __block_write_begin_int() instead.

vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
{
	...
	if (!ext4_should_journal_data(inode)) {
		err = block_page_mkwrite(vma, vmf, get_block);
	...
}


So...

Thanks,
Yi.

Re: [PATCH v2 3/6] ext4: restart handle if credits are insufficient during allocating blocks

Posted by Jan Kara 7 months, 3 weeks ago

On Fri 20-06-25 13:00:32, Zhang Yi wrote:
> On 2025/6/20 0:33, Jan Kara wrote:
> > On Wed 11-06-25 19:16:22, Zhang Yi wrote:
> >> From: Zhang Yi <yi.zhang@huawei.com>
> >>
> >> After large folios are supported on ext4, writing back a sufficiently
> >> large and discontinuous folio may consume a significant number of
> >> journal credits, placing considerable strain on the journal. For
> >> example, in a 20GB filesystem with 1K block size and 1MB journal size,
> >> writing back a 2MB folio could require thousands of credits in the
> >> worst-case scenario (when each block is discontinuous and distributed
> >> across different block groups), potentially exceeding the journal size.
> >> This issue can also occur in ext4_write_begin() and ext4_page_mkwrite()
> >> when delalloc is not enabled.
> >>
> >> Fix this by ensuring that there are sufficient journal credits before
> >> allocating an extent in mpage_map_one_extent() and _ext4_get_block(). If
> >> there are not enough credits, return -EAGAIN, exit the current mapping
> >> loop, restart a new handle and a new transaction, and allocating blocks
> >> on this folio again in the next iteration.
> >>
> >> Suggested-by: Jan Kara <jack@suse.cz>
> >> Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
> > 
> > ...
> > 
> >>  static int _ext4_get_block(struct inode *inode, sector_t iblock,
> >>  			   struct buffer_head *bh, int flags)
> >>  {
> >>  	struct ext4_map_blocks map;
> >> +	handle_t *handle = ext4_journal_current_handle();
> >>  	int ret = 0;
> >>  
> >>  	if (ext4_has_inline_data(inode))
> >>  		return -ERANGE;
> >>  
> >> +	/* Make sure transaction has enough credits for this extent */
> >> +	if (flags & EXT4_GET_BLOCKS_CREATE) {
> >> +		ret = ext4_journal_ensure_extent_credits(handle, inode);
> >> +		if (ret)
> >> +			return ret;
> >> +	}
> >> +
> >>  	map.m_lblk = iblock;
> >>  	map.m_len = bh->b_size >> inode->i_blkbits;
> >>  
> >> -	ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
> >> -			      flags);
> >> +	ret = ext4_map_blocks(handle, inode, &map, flags);
> > 
> > Good spotting with ext4_page_mkwrite() and ext4_write_begin() also needing
> > this treatment! But rather then hiding the transaction extension in
> > _ext4_get_block() I'd do this in ext4_block_write_begin() where it is much
> > more obvious (and also it is much more obvious who needs to be prepared for
> > handling EAGAIN error). Otherwise the patch looks good!
> > 
> 
> Yes, I completely agree with you. However, unfortunately, do this in
> ext4_block_write_begin() only works for ext4_write_begin().
> ext4_page_mkwrite() does not call ext4_block_write_begin() to allocate
> blocks, it call the vfs helper __block_write_begin_int() instead.
> 
> vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
> {
> 	...
> 	if (!ext4_should_journal_data(inode)) {
> 		err = block_page_mkwrite(vma, vmf, get_block);
> 	...
> }
> 
> 
> So...

Right, I forgot about the nodelalloc case. But since we do most of things
by hand for data=journal mode, perhaps we could lift some code from
data=journal mode and reuse it for nodelalloc as well like:

        folio_lock(folio);
        size = i_size_read(inode);
        /* Page got truncated from under us? */
        if (folio->mapping != mapping || folio_pos(folio) > size) {
                ret = VM_FAULT_NOPAGE;
                goto out_error;
        }

        len = folio_size(folio);
        if (folio_pos(folio) + len > size)
                len = size - folio_pos(folio);
                
        err = ext4_block_write_begin(handle, folio, 0, len,
                                     get_block);
	if (err)
		goto out_error;
	if (!ext4_should_journal_data(inode))
		block_commit_write(folio, 0, len);
		folio_mark_dirty(folio);
	} else {
	        if (ext4_journal_folio_buffers(handle, folio, len)) {
	        	ret = VM_FAULT_SIGBUS;
		        goto out_error;
		}
	}
	ext4_journal_stop(handle);
	folio_wait_stable(folio);

We get an additional bonus for not waiting for page writeback with
transaction handle held (which is a potential deadlock vector). What do you
think?

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [PATCH v2 3/6] ext4: restart handle if credits are insufficient during allocating blocks

Posted by Zhang Yi 7 months, 3 weeks ago

On 2025/6/20 22:18, Jan Kara wrote:
> On Fri 20-06-25 13:00:32, Zhang Yi wrote:
>> On 2025/6/20 0:33, Jan Kara wrote:
>>> On Wed 11-06-25 19:16:22, Zhang Yi wrote:
>>>> From: Zhang Yi <yi.zhang@huawei.com>
>>>>
>>>> After large folios are supported on ext4, writing back a sufficiently
>>>> large and discontinuous folio may consume a significant number of
>>>> journal credits, placing considerable strain on the journal. For
>>>> example, in a 20GB filesystem with 1K block size and 1MB journal size,
>>>> writing back a 2MB folio could require thousands of credits in the
>>>> worst-case scenario (when each block is discontinuous and distributed
>>>> across different block groups), potentially exceeding the journal size.
>>>> This issue can also occur in ext4_write_begin() and ext4_page_mkwrite()
>>>> when delalloc is not enabled.
>>>>
>>>> Fix this by ensuring that there are sufficient journal credits before
>>>> allocating an extent in mpage_map_one_extent() and _ext4_get_block(). If
>>>> there are not enough credits, return -EAGAIN, exit the current mapping
>>>> loop, restart a new handle and a new transaction, and allocating blocks
>>>> on this folio again in the next iteration.
>>>>
>>>> Suggested-by: Jan Kara <jack@suse.cz>
>>>> Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
>>>
>>> ...
>>>
>>>>  static int _ext4_get_block(struct inode *inode, sector_t iblock,
>>>>  			   struct buffer_head *bh, int flags)
>>>>  {
>>>>  	struct ext4_map_blocks map;
>>>> +	handle_t *handle = ext4_journal_current_handle();
>>>>  	int ret = 0;
>>>>  
>>>>  	if (ext4_has_inline_data(inode))
>>>>  		return -ERANGE;
>>>>  
>>>> +	/* Make sure transaction has enough credits for this extent */
>>>> +	if (flags & EXT4_GET_BLOCKS_CREATE) {
>>>> +		ret = ext4_journal_ensure_extent_credits(handle, inode);
>>>> +		if (ret)
>>>> +			return ret;
>>>> +	}
>>>> +
>>>>  	map.m_lblk = iblock;
>>>>  	map.m_len = bh->b_size >> inode->i_blkbits;
>>>>  
>>>> -	ret = ext4_map_blocks(ext4_journal_current_handle(), inode, &map,
>>>> -			      flags);
>>>> +	ret = ext4_map_blocks(handle, inode, &map, flags);
>>>
>>> Good spotting with ext4_page_mkwrite() and ext4_write_begin() also needing
>>> this treatment! But rather then hiding the transaction extension in
>>> _ext4_get_block() I'd do this in ext4_block_write_begin() where it is much
>>> more obvious (and also it is much more obvious who needs to be prepared for
>>> handling EAGAIN error). Otherwise the patch looks good!
>>>
>>
>> Yes, I completely agree with you. However, unfortunately, do this in
>> ext4_block_write_begin() only works for ext4_write_begin().
>> ext4_page_mkwrite() does not call ext4_block_write_begin() to allocate
>> blocks, it call the vfs helper __block_write_begin_int() instead.
>>
>> vm_fault_t ext4_page_mkwrite(struct vm_fault *vmf)
>> {
>> 	...
>> 	if (!ext4_should_journal_data(inode)) {
>> 		err = block_page_mkwrite(vma, vmf, get_block);
>> 	...
>> }
>>
>>
>> So...
> 
> Right, I forgot about the nodelalloc case. But since we do most of things
> by hand for data=journal mode, perhaps we could lift some code from
> data=journal mode and reuse it for nodelalloc as well like:
> 
>         folio_lock(folio);
>         size = i_size_read(inode);
>         /* Page got truncated from under us? */
>         if (folio->mapping != mapping || folio_pos(folio) > size) {
>                 ret = VM_FAULT_NOPAGE;
>                 goto out_error;
>         }
> 
>         len = folio_size(folio);
>         if (folio_pos(folio) + len > size)
>                 len = size - folio_pos(folio);
>                 
>         err = ext4_block_write_begin(handle, folio, 0, len,
>                                      get_block);
> 	if (err)
> 		goto out_error;
> 	if (!ext4_should_journal_data(inode))
> 		block_commit_write(folio, 0, len);
> 		folio_mark_dirty(folio);
> 	} else {
> 	        if (ext4_journal_folio_buffers(handle, folio, len)) {
> 	        	ret = VM_FAULT_SIGBUS;
> 		        goto out_error;
> 		}
> 	}
> 	ext4_journal_stop(handle);
> 	folio_wait_stable(folio);
> 
> We get an additional bonus for not waiting for page writeback with
> transaction handle held (which is a potential deadlock vector). What do you
> think?
> 

Yeah, this solution looks nice to me, it should works! Thank you for
the suggestion.

Best regards,
Yi.

[PATCH v2 1/6] ext4: move the calculation of wbc->nr_to_write to mpage_folio_done()
[PATCH v2 2/6] ext4: fix stale data if it bail out of the extents mapping loop
[PATCH v2 3/6] ext4: restart handle if credits are insufficient during allocating blocks
[PATCH v2 4/6] ext4: correct the reserved credits for extent conversion
[PATCH v2 5/6] ext4/jbd2: reintroduce jbd2_journal_blocks_per_page()
[PATCH v2 6/6] ext4: fix insufficient credits calculation in ext4_meta_trans_blocks()