On Mon, Jun 19, 2017 at 05:03:16PM +0100, Al Viro wrote:
> On Mon, Jun 19, 2017 at 09:01:36PM +0800, Sean Fu wrote:
> > Make alloc_page_buffers support circular buffer list and initialise
> > b_state field.
> > Optimize the performance by removing the buffer list traversal to create
> > circular buffer list.
> 
> > -		bh = head = alloc_page_buffers(page, blocksize, 1);
> > +		bh = head = alloc_page_buffers(page, blocksize, 1, 0, 0);
> 
> Frankly, I don't like that change of calling conventions; it's very easy to
> mess the order of arguments when using interfaces like that and it's hell
> to find when trying to debug the resulting mess.
> 
> Do you really get an observable change in performance?  What loads are
> triggering it?
Yes, I have got the performance change with ext3 file system which block
size is 1024 bytes. It has at least %5 performance improvement.

I have found the performance improvements when writting/reading a 800M
size of file on ext3 file system with 1024 block size.
In this case, Each page has four buffer. In other word, the buffer list
has 4 elements.

I have compared the time that the process spent in kernel mode.

Improvements via this patch
Before		After
Write:
0m5.604s	0m4.116s
0m4.408s	0m3.924s
0m4.184s	0m3.708s
0m4.352s	0m3.656s
0m4.380s	0m3.608s
0m4.240s	0m3.612s
0m4.460s	0m3.552s
0m4.072s	0m3.832s
0m4.300s	0m3.736s
0m4.400s	0m3.480s

Read:
0m3.128s	0m3.036s
0m2.976s	0m2.568s
0m3.384s	0m2.544s
0m3.112s	0m2.752s
0m2.924s	0m2.684s
0m3.084s	0m2.856s
0m3.348s	0m2.576s
0m3.000s	0m2.968s
0m3.012s	0m2.560s
0m2.768s	0m2.752s

Reproduce steps:
1 mkfs.ext3 -b 1024 /dev/sdb1
2 ./test_write.sh ./writetest 10

Test shell script:
#!/bin/bash
i=$2;
while test $((i)) -ge 1; do
	mount /dev/sdb1 /mnt/sdb1/
	time $1 -b 800 -o /mnt/sdb1/fileout
	rm /mnt/sdb1/fileout
	sync
	sleep 1
	umount /mnt/sdb1/
	echo "aaa"
	i=$i-1
done


The attachment are the code for writetest and test result.