Keith Busch <keith.bu...@intel.com> writes:
> On Tue, Mar 15, 2016 at 04:17:56PM +0100, Vitaly Kuznetsov wrote:
>> The reason of the slowdown is the fact that bios don't get merged and we
>> end up sending many short requests to the host. My investigation led me to
>> the following code (__bvec_gap_to_prev()):
>>
>> return offset ||
>> ((bprv->bv_offset + bprv->bv_len) & queue_virt_boundary(q));
>>
>> Here is an example: we have two bio_vec with the following content:
>> bprv.bv_offset = 512
>> bprv.bv_len = 512
>>
>> bnxt.bv_offset = 1024
>> bnxt.bv_len = 512
>>
>> bprv.bv_page == bnxt.bv_page
>> virt_boundary is set to PAGE_SIZE-1
>>
>> The above mentioned code will report that a gap will appear if we merge
>> these two (as offset = 1024) but this doesn't look sane. On top of that,
>> we have the following optimization in bio_add_pc_page():
>>
>> if (page == prev->bv_page &&
>> offset == prev->bv_offset + prev->bv_len) {
>> prev->bv_len += len;
>> bio->bi_iter.bi_size += len;
>> goto done;
>> }
>
> This part sounds odd. Why is a filesystem using bio_add_pc_page? Shouldn't
> these go through "bio_add_page" instead? That already has an optimization
> to combine bio's within the same page.
Not sure I know enough to comment here and it is most probably unrelated
to the issue I'm seeing (bio_add_pc_page() doesn't pop up when I do
'mkfs.ntfs') but in this particular place I see same page check before
we do bvec_gap_to_prev() but there is no such check in other places and
bios in the same page are always being split:
return offset || ((bprv->bv_offset + bprv->bv_len) & queue_virt_boundary(q));
will always return 'true' because offset is the offset of the second
bio. That's what I'm trying to address.
--
Vitaly
