when writing file with direct_IO on bcachefs, then performance is
much lower than other fs due to write back throttle in block layer:

        wbt_wait+1
        __rq_qos_throttle+32
        blk_mq_submit_bio+394
        submit_bio_noacct_nocheck+649
        bch2_submit_wbio_replicas+538
        __bch2_write+2539
        bch2_direct_write+1663
        bch2_write_iter+318
        aio_write+355
        io_submit_one+1224
        __x64_sys_io_submit+169
        do_syscall_64+134
        entry_SYSCALL_64_after_hwframe+110

add set REQ_SYNC and REQ_IDLE in bio->bi_opf as standard dirct-io

Signed-off-by: zhuxiaohui <zhuxiaohui.400@bytedance.com>
---
 fs/bcachefs/fs-io-direct.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/bcachefs/fs-io-direct.c b/fs/bcachefs/fs-io-direct.c
index 33cb6da3a5ad..f49e6c0f0f68 100644
--- a/fs/bcachefs/fs-io-direct.c
+++ b/fs/bcachefs/fs-io-direct.c
@@ -536,7 +536,7 @@ static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
 		if (likely(!dio->iter.count) || dio->op.error)
 			break;
 
-		bio_reset(bio, NULL, REQ_OP_WRITE);
+		bio_reset(bio, NULL, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
 	}
 out:
 	return bch2_dio_write_done(dio);
@@ -618,7 +618,7 @@ ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
 
 	bio = bio_alloc_bioset(NULL,
 			       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
-			       REQ_OP_WRITE,
+			       REQ_OP_WRITE | REQ_SYNC | REQ_IDLE,
 			       GFP_KERNEL,
 			       &c->dio_write_bioset);
 	dio = container_of(bio, struct dio_write, op.wbio.bio);
-- 
2.41.0

On Tue, Mar 26, 2024 at 08:03:45PM +0800, zhuxiaohui wrote:
> when writing file with direct_IO on bcachefs, then performance is
> much lower than other fs due to write back throttle in block layer:
> 
>         wbt_wait+1
>         __rq_qos_throttle+32
>         blk_mq_submit_bio+394
>         submit_bio_noacct_nocheck+649
>         bch2_submit_wbio_replicas+538
>         __bch2_write+2539
>         bch2_direct_write+1663
>         bch2_write_iter+318
>         aio_write+355
>         io_submit_one+1224
>         __x64_sys_io_submit+169
>         do_syscall_64+134
>         entry_SYSCALL_64_after_hwframe+110
> 
> add set REQ_SYNC and REQ_IDLE in bio->bi_opf as standard dirct-io
> 
> Signed-off-by: zhuxiaohui <zhuxiaohui.400@bytedance.com>

Can I get you interested in a bigger project?

The writeback throttling code is a problem; it really shouldn't be
specific to writeback, it really ought to be a general purpose
backpressure mechanism.

Unfortunately, I've found the code to be opaque, practically to the
point of obfuscation, so it's going to be a difficult nut to crack.

But the lack of higher level, more workable backpressure is a problem,
it leads to queueing delays and priority inversions in filesystem code
and below.

> ---
>  fs/bcachefs/fs-io-direct.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/bcachefs/fs-io-direct.c b/fs/bcachefs/fs-io-direct.c
> index 33cb6da3a5ad..f49e6c0f0f68 100644
> --- a/fs/bcachefs/fs-io-direct.c
> +++ b/fs/bcachefs/fs-io-direct.c
> @@ -536,7 +536,7 @@ static __always_inline long bch2_dio_write_loop(struct dio_write *dio)
>  		if (likely(!dio->iter.count) || dio->op.error)
>  			break;
>  
> -		bio_reset(bio, NULL, REQ_OP_WRITE);
> +		bio_reset(bio, NULL, REQ_OP_WRITE | REQ_SYNC | REQ_IDLE);
>  	}
>  out:
>  	return bch2_dio_write_done(dio);
> @@ -618,7 +618,7 @@ ssize_t bch2_direct_write(struct kiocb *req, struct iov_iter *iter)
>  
>  	bio = bio_alloc_bioset(NULL,
>  			       bio_iov_vecs_to_alloc(iter, BIO_MAX_VECS),
> -			       REQ_OP_WRITE,
> +			       REQ_OP_WRITE | REQ_SYNC | REQ_IDLE,
>  			       GFP_KERNEL,
>  			       &c->dio_write_bioset);
>  	dio = container_of(bio, struct dio_write, op.wbio.bio);
> -- 
> 2.41.0
>

Thanks for your patience and explanation, that seems a interesting issue