Re: [RFC PATCH v4 33/34] ext4: don't mark IOMAP_F_DIRTY for buffer write

[Zhang Yi <yi.zhang@huaweicloud.com>]

On 2024/5/1 17:27, Dave Chinner wrote:
> On Wed, Apr 10, 2024 at 11:03:12PM +0800, Zhang Yi wrote:
>> From: Zhang Yi <yi.zhang@huawei.com>
>>
>> The data sync dirty check in ext4_inode_datasync_dirty() is expansive
>> since jbd2_transaction_committed() holds journal->j_state lock when
>> journal is enabled, it costs a lot in high-concurrency iomap buffered
>> read/write paths, but we never check IOMAP_F_DIRTY in these cases, so
>> let's check it only in swap file, dax and direct IO cases. Tested by
>> Unixbench on 100GB ramdisk:
>>
>> ./Run -c 128 -i 10 fstime fsbuffer fsdisk
>>
>>   == without this patch ==
>>   128 CPUs in system; running 128 parallel copies of tests
>>
>>   File Copy 1024 bufsize 2000 maxblocks       6332521.0 KBps
>>   File Copy 256 bufsize 500 maxblocks         1639726.0 KBps
>>   File Copy 4096 bufsize 8000 maxblocks      24018572.0 KBps
>>
>>   == with this patch ==
>>   128 CPUs in system; running 128 parallel copies of tests
>>
>>   File Copy 1024 bufsize 2000 maxblocks      49229257.0 KBps
>>   File Copy 256 bufsize 500 maxblocks        24057510.0 KBps
>>   File Copy 4096 bufsize 8000 maxblocks      75704437.0 KBps
>>
>> Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
>> ---
>>  fs/ext4/inode.c | 10 +++++++---
>>  1 file changed, 7 insertions(+), 3 deletions(-)
>>
>> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
>> index 1cb219d347af..269503749ef5 100644
>> --- a/fs/ext4/inode.c
>> +++ b/fs/ext4/inode.c
>> @@ -3281,9 +3281,13 @@ static void ext4_set_iomap(struct inode *inode, struct iomap *iomap,
>>  	 * there is no other metadata changes being made or are pending.
>>  	 */
>>  	iomap->flags = 0;
>> -	if (ext4_inode_datasync_dirty(inode) ||
>> -	    offset + length > i_size_read(inode))
>> -		iomap->flags |= IOMAP_F_DIRTY;
>> +	if ((flags & (IOMAP_DAX | IOMAP_REPORT)) ||
>> +	    ((flags & (IOMAP_WRITE | IOMAP_DIRECT)) ==
>> +	     (IOMAP_WRITE | IOMAP_DIRECT))) {
>> +		if (offset + length > i_size_read(inode) ||
>> +		    ext4_inode_datasync_dirty(inode))
>> +			iomap->flags |= IOMAP_F_DIRTY;
>> +	}
> 
> NACK. This just adds a nasty landmine that anyone working on the
> iomap infrastructure can step on. i.e. any time we add a new check
> for IOMAP_F_DIRTY in the generic infrastructure, ext4 is going to
> break because it won't set the IOMAP_F_DIRTY flag correctly.
> 
> If checking an inode is dirty is expensive on ext4, then make it
> less expensive and everyone will benefit.
> 
> /me goes and looks at jbd2_transaction_committed()
> 
> Oh, it it's just a sequence number comparison, and it needs a lock
> because it has to dereference the running/committed transactions
> structures to get the current sequence numbers. Why not just store
> the commiting/running transaction tids in the journal_t, and then
> you can sample them without needing any locking and the whole
> ext4_inode_datasync_dirty() scalability problem goes away...
> 

Indeed, it could be useful, and it seems could also simplify many
other jbd2 processes.

Thanks,
Yi.