On 8/12/21 12:51 AM, Christoph Hellwig wrote:
> On Wed, Aug 11, 2021 at 01:35:28PM -0600, Jens Axboe wrote:
>> The memset() used is measurably slower in targeted benchmarks. Get rid
>> of it and fill in the bio manually, in a separate helper.
>
> If you have some numbers if would be great to throw them in here.
It's about 1% of the overhead of the alloc after the cache, which
comes later in the series.
Percent│ return __underlying_memset(p, c, size);
│ lea 0x8(%r8),%rdi
│ bio_alloc_kiocb():
2.18 │ cmove %rax,%r9
│ memset():
│ mov %r8,%rcx
│ and $0xfffffffffffffff8,%rdi
│ movq $0x0,(%r8)
│ sub %rdi,%rcx
│ add $0x60,%ecx
│ shr $0x3,%ecx
55.02 │ rep stos %rax,%es:(%rdi)
This is on AMD, might look different on Intel, the manual clear seems
like a nice win on both. As a minor detail, avoids things like
re-setting bio->bi_pool for cached entries, as it never changes.
>> +static inline void __bio_init(struct bio *bio)
>
> Why is this split from bio_init and what are the criteria where an
> initialization goes?
Got rid of the helper.
>> + bio->bi_flags = bio->bi_ioprio = bio->bi_write_hint = 0;
>
> Please keep each initialization on a separate line.
Done
--
Jens Axboe
