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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 11/14/19 10:41 AM, John Garry wrote:
> On 13/11/2019 18:38, Hannes Reinecke wrote:
>>> 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.
>>>
>> Actually, I _do_ prefer keeping both in sync.
>> We might want to check if the 'normal' tag is set (typically it would
>> not, but then, who knows ...)
>> The beauty here is that both 'shared' and 'normal' tag are in sync, so
>> if a driver would be wanting to use the tag as index into a command
>> array it can do so without any surprises.
>>
>> Why do you think it's not ideal?
> 
> A few points:
> - Getting a bit from one tagset and then setting it in another tagset is
> a bit clunky.
Yes, that's true.
But painstakingly trying to find a free bit in a bitmask when we already
know which to pick is also a bit daft.

> - There may be an atomicity of the getting the shared tag bit and
> setting the hctx tag bit - I don't think that there is.

That was precisely what I've alluded to in 'We might want to check if
the normal tag is set'.
Typically the 'normal' tag would be free (as the shared tag set out of
necessity needs to be the combination of all hctx tag sets).
Any difference here _is_ a programming error, and should be flagged as
such (sbitmap_test_and_set() anyone?)
We might have ordering issues on release, as we really should drop the
hctx tag before the shared tag; but when we observe that we should be fine.

> - Consider that sometimes we may want to check if there is space on a hw
> queue - checking the hctx tags is not really proper any longer, as
> typically there would always be space on hctx, but not always the shared
> tags. We did delete blk_mq_can_queue() yesterday, which would be an
> example of that. Need to check if there are others.
> 
Clearly, this needs an audit of all functions accessing the hctx tag
space; maybe it's worth having a pre-requisite patchset differentiating
between hctx tags and global, shared tags. Hmm.

> Having said all that, the obvious advantage is performance gain, can
> still use request.tag and so maybe less intrusive changes.
> 
> I'll have a look at the implementation. The devil is mostly in the
> detail...
> 
True.
And, incidentally, if we run with shared tage we can skip the scheduling
section in blk_mq_get_tag(); if we're out of tags, we're out of tags,
and no rescheduling will help as we don't _have_ other tagsets to look at.

But overall I like this approach.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      Teamlead Storage & Networking
hare@xxxxxxx			                  +49 911 74053 688
SUSE Software Solutions Germany GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 247165 (AG München), GF: Felix Imendörffer



[Index of Archives]     [Linux RAID]     [Linux SCSI]     [Linux ATA RAID]     [IDE]     [Linux Wireless]     [Linux Kernel]     [ATH6KL]     [Linux Bluetooth]     [Linux Netdev]     [Kernel Newbies]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Device Mapper]

  Powered by Linux