lkml.org 
[lkml]   [2019]   [Nov]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH RFC 3/5] blk-mq: Facilitate a shared tags per tagset
From
Date
On 11/15/19 2:24 AM, John Garry wrote:
> Bart Van Assche wrote:
> > How about sharing tag sets across hardware
> > queues, e.g. like in the (totally untested) patch below?
>
> So this is similar in principle what Ming Lei came up with here:
> https://lore.kernel.org/linux-block/20190531022801.10003-1-ming.lei@redhat.com/
>
> However your implementation looks neater, which is good.
>
> My concern with this approach is that we can't differentiate which tags
> are allocated for which hctx, and sometimes we need to know that.
>
> An example here was blk_mq_queue_tag_busy_iter(), which iterates the
> bits for each hctx. This would just be broken by that change, unless we
> record which bits are associated with each hctx.

I disagree. In bt_iter() I added " && rq->mq_hctx == hctx" such that
blk_mq_queue_tag_busy_iter() only calls the callback function for
matching (hctx, rq) pairs.

> Another example was __blk_mq_tag_idle(), which looks problematic.

Please elaborate.

> For debugfs, when we examine
> /sys/kernel/debug/block/.../hctxX/tags_bitmap, wouldn't that be the tags
> for all hctx (hctx0)?
>
> For debugging reasons, I would say we want to know which tags are
> allocated for a specific hctx, as this is tightly related to the
> requests for that hctx.

That is an open issue in the patch I posted and something that needs to
be addressed. One way to address this is to change the
sbitmap_bitmap_show() calls into calls to a function that only shows
those bits for which rq->mq_hctx == hctx.

>> @@ -341,8 +341,11 @@ void blk_mq_tagset_busy_iter(struct
>> blk_mq_tag_set *tagset,
>>       int i;
>>
>>       for (i = 0; i < tagset->nr_hw_queues; i++) {
>> -        if (tagset->tags && tagset->tags[i])
>> +        if (tagset->tags && tagset->tags[i]) {
>>               blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
>
> As I mentioned earlier, wouldn't this iterate over all tags for all
> hctx's, when we just want the tags for hctx[i]?
>
> Thanks,
> John
>
> [Not trimming reply for future reference]
>
>> +            if (tagset->share_tags)
>> +                break;
>> +        }
>>       }
>>   }
>>   EXPORT_SYMBOL(blk_mq_tagset_busy_iter);

Since blk_mq_tagset_busy_iter() loops over all hardware queues all what
is changed is the order in which requests are examined. I am not aware
of any block driver that calls blk_mq_tagset_busy_iter() and that
depends on the order of the requests passed to the callback function.

Thanks,

Bart.

\
 
 \ /
  Last update: 2019-11-15 18:59    [W:0.124 / U:24.684 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site