Re: [PATCH RFC 3/5] blk-mq: Facilitate a shared tags per tagset

[John Garry <john.garry@huawei.com>]

On 13/11/2019 15:38, Hannes Reinecke wrote:
>>>> -        if (clear_ctx_on_error)
>>>> -            data->ctx = NULL;
>>>> -        blk_queue_exit(q);
>>>> -        return NULL;
>>>> +    if (data->hctx->shared_tags) {
>>>> +        shared_tag = blk_mq_get_shared_tag(data);
>>>> +        if (shared_tag == BLK_MQ_TAG_FAIL)
>>>> +            goto err_shared_tag;
>>>>        }
>>>>    -    rq = blk_mq_rq_ctx_init(data, tag, data->cmd_flags,
>>>> alloc_time_ns);
>>>> +    tag = blk_mq_get_tag(data);
>>>> +    if (tag == BLK_MQ_TAG_FAIL)
>>>> +        goto err_tag;
>>>> +
>>>> +    rq = blk_mq_rq_ctx_init(data, tag, shared_tag, data->cmd_flags,
>>>> alloc_time_ns);
>>>>        if (!op_is_flush(data->cmd_flags)) {
>>>>            rq->elv.icq = NULL;
>>>>            if (e && e->type->ops.prepare_request) {
>> Hi Hannes,
>>
>>> Why do you need to keep a parallel tag accounting between 'normal' and
>>> 'shared' tags here?
>>> Isn't is sufficient to get a shared tag only, and us that in lieo of the
>>> 'real' one?
>> In theory, yes. Just the 'shared' tag should be adequate.
>>
>> A problem I see with this approach is that we lose the identity of which
>> tags are allocated for each hctx. As an example for this, consider
>> blk_mq_queue_tag_busy_iter(), which iterates the bits for each hctx.
>> Now, if you're just using shared tags only, that wouldn't work.
>>
>> Consider blk_mq_can_queue() as another example - this tells us if a hctx
>> has any bits unset, while with only using shared tags it would tell if
>> any bits unset over all queues, and this change in semantics could break
>> things. At a glance, function __blk_mq_tag_idle() looks problematic also.
>>
>> And this is where it becomes messy to implement.
>>

Hi Hannes,

> Oh, my. Indeed, that's correct.

The tags could be kept in sync like this:

shared_tag = blk_mq_get_tag(shared_tagset);
if (shared_tag != -1)
	sbitmap_set(hctx->tags, shared_tag);

But that's obviously not ideal.

> 
> But then we don't really care _which_ shared tag is assigned; so
> wouldn't be we better off by just having an atomic counter here?
> Cache locality will be blown anyway ...
The atomic counter would solve the issuing more than Scsi_host.can_queue 
to the LLDD, but we still need a unique tag, which is what the shared 
tag is.

Thanks,
John