Hi Andrew,
在 2021/2/6 7:45, Andrew Morton 写道:
> On Wed, 3 Feb 2021 14:14:50 +0800 Shaokun Zhang <zhangshaokun@xxxxxxxxxxxxx> wrote:
>
>> From: Yang Guo <guoyang2@xxxxxxxxxx>
>>
>> clear_buffer_new() is used to clear buffer new stat. When PAGE_SIZE
>> is 64K, most buffer heads in the list are not needed to clear.
>> clear_buffer_new() has an enpensive atomic modification operation,
>> Let's add checking buffer head before clear it as __block_write_begin_int
>> does which is good for performance.
>
> Did this produce any measurable improvement?
It has been tested on Huwei Kunpeng 920 which is ARM64 platform and test commond is below:
numactl --cpunodebind=0 --membind=0 fio -name=randwrite -numjobs=16 -filename=/mnt/test1
-rw=randwrite -ioengine=libaio -direct=0 -iodepth=64 -sync=0 -norandommap -group_reporting
-runtime=60 -time_based -bs=4k -size=5G
The test result before patch:
WRITE: bw=930MiB/s (976MB/s), 930MiB/s-930MiB/s (976MB/s-976MB/s), io=54.5GiB (58.5GB),
run=60001-60001msec
The test result after patch:
WRITE: bw=958MiB/s (1005MB/s), 958MiB/s-958MiB/s (1005MB/s-1005MB/s), io=56.1GiB (60.3GB),
run=60001-60001msec
>
> Perhaps we should give clear_buffer_x() the same optimization as
> set_buffer_x()?
>
Good catch,
but we check it more about it, if we do it the same as set_buffer_x(),
many more codes will be fixed, such as ext4_wait_block_bitmap
it has done sanity check using buffer_new and clear_buffer_new
will check it again.
Thanks,
Shaokun
>
> static __always_inline void set_buffer_##name(struct buffer_head *bh) \
> { \
> if (!test_bit(BH_##bit, &(bh)->b_state)) \
> set_bit(BH_##bit, &(bh)->b_state); \
> } \
> static __always_inline void clear_buffer_##name(struct buffer_head *bh) \
> { \
> clear_bit(BH_##bit, &(bh)->b_state); \
> } \
>
>
> .
>
