Skip to contentSkip to conversation

set kiobuf io_count once, instead of increment

PATCH In review Robert Read 3 participants 5 messages Feb 28, 2001 – Mar 2, 2001 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 set kiobuf io_count once, instead of increment

Currently in brw_kiovec, iobuf->io_count is being incremented as each bh is submitted, and decremented in the bh->b_end_io(). This means io_count can go to zero before all the bhs have been submitted, especially during a large request. This causes the end_kio_request() to be called before all of the io is complete.

This suggested patch against 2.4.2 sets io_count to the total amount before the bhs are submitted, although there is probably a better way to determine the io_count than this.

robert

fs/buffer.c +1 -2
diff -ru linux/fs/buffer.c linux-rm/fs/buffer.c
@@ -2085,6 +2085,7 @@
 		offset = iobuf->offset;
 		length = iobuf->length;
 		iobuf->errno = 0;
+		atomic_set(&iobuf->io_count, length/size);
 		
 		for (pageind = 0; pageind < iobuf->nr_pages; pageind++) {
 			map  = iobuf->maplist[pageind];
@@ -2119,8 +2120,6 @@
 				bh[bhind++] = tmp;
 				length -= size;
 				offset += size;
-
-				atomic_inc(&iobuf->io_count);
 
 				submit_bh(rw, tmp);
 				/* 
#1 lore raw
Quoted 24 lines · Robert Read

Currently in brw_kiovec, iobuf->io_count is being incremented as each bh is submitted, and decremented in the bh->b_end_io(). This means io_count can go to zero before all the bhs have been submitted, especially during a large request. This causes the end_kio_request() to be called before all of the io is complete.

This suggested patch against 2.4.2 sets io_count to the total amount before the bhs are submitted, although there is probably a better way to determine the io_count than this.

robert

diff -ru linux/fs/buffer.c linux-rm/fs/buffer.c

fs/buffer.c +1 -0
@@ -2085,6 +2085,7 @@
 		offset = iobuf->offset;
 		length = iobuf->length;
 		iobuf->errno = 0;
+		atomic_set(&iobuf->io_count, length/size);
 
 		for (pageind = 0; pageind < iobuf->nr_pages; pageind++) {
 			map  = iobuf->maplist[pageind];

fs/buffer.c

Quoted 9 lines
fs/buffer.c +0 -3
@@ -2119,8 +2120,6 @@
 				bh[bhind++] = tmp;
 				length -= size;
 				offset += size;
-
-				atomic_inc(&iobuf->io_count);
 
 				submit_bh(rw, tmp);
 				/* 
-

It seems your patch breaks bh allocation failure handling. If get_unused_buffer_head() fails, iobuf->io_count never reaches 0, so processes waiting on kiobuf_wait_for_io() will block forever.

#2 lore raw
In reply to Robert Read
Quoted 6 lines · Marcelo Tosatti

It seems your patch breaks bh allocation failure handling. If get_unused_buffer_head() fails, iobuf->io_count never reaches 0, so processes waiting on kiobuf_wait_for_io() will block forever.

This is true, but it looks like the brw_kiovec allocation failure handling is broken already; it's calling __put_unused_buffer_head on bhs without waiting for them to complete first. Also, the err won't be returned if the previous batch of bhs finished ok. It looks like brw_kiovec needs some work, but I'm going to need some coffee first...

robert

#3 lore raw
In reply to Robert Read
Quoted 5 lines · Robert Read

Currently in brw_kiovec, iobuf->io_count is being incremented as each bh is submitted, and decremented in the bh->b_end_io(). This means io_count can go to zero before all the bhs have been submitted, especially during a large request. This causes the end_kio_request() to be called before all of the io is complete.

brw_kiovec is currently entirely synchronous, so end_kio_request() calling is probably not a big deal right now. It would be much more important for an async version.

--Stephen

#4 lore raw
In reply to Robert Read

Hi,

Quoted 7 lines · Robert Read

On Tue, Feb 27, 2001 at 10:50:54PM -0300, Marcelo Tosatti wrote:

This is true, but it looks like the brw_kiovec allocation failure handling is broken already; it's calling __put_unused_buffer_head on bhs without waiting for them to complete first. Also, the err won't be returned if the previous batch of bhs finished ok. It looks like brw_kiovec needs some work, but I'm going to need some coffee first...

Right, looks like this happened when somebody was changing the bh submission mechanism to use submit_bh(). I'll fix it.

--Stephen