On 26/11/2019 17:11, Jens Axboe wrote:
On 11/26/19 9:54 AM, John Garry wrote:
On 26/11/2019 15:14, Jens Axboe wrote:
On 11/26/19 2:14 AM, Hannes Reinecke wrote:
Instead of allocating the tag bitmap in place we should be using a
pointer. This is in preparation for shared host-wide bitmaps.
Not a huge fan of this, it's an extra indirection in the hot path
of both submission and completion.
Hi Jens,
Thanks for having a look.
I checked the disassembly for blk_mq_get_tag() as a sample - which I
assume is one hot path function which you care about - and the cost of
the indirection is a load instruction instead of an add, denoted by ***,
below:
Hi Jens,
I'm not that worried about an extra instruction, my worry is the extra
load is from different memory. When it's embedded in the struct, we're
on the same cache line or adjacent.
Right, so the earlier iteration of this series kept the embedded struct
and we simply pointed at that, so I wouldn't expect a caching issue of
different memory in that case.
Cheers,
John
