Message-ID: <0f7f3c4e-530a-5cd2-2719-259e509366e4@oracle.com>
Date:   Fri, 26 Mar 2021 14:32:01 -0700
From:   Mike Kravetz <mike.kravetz@...cle.com>
To:     Roman Gushchin <guro@...com>, Minchan Kim <minchan@...nel.org>
Cc:     David Hildenbrand <david@...hat.com>,
        Michal Hocko <mhocko@...e.com>, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org, Shakeel Butt <shakeelb@...gle.com>,
        Oscar Salvador <osalvador@...e.de>,
        Muchun Song <songmuchun@...edance.com>,
        David Rientjes <rientjes@...gle.com>,
        Miaohe Lin <linmiaohe@...wei.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Matthew Wilcox <willy@...radead.org>,
        HORIGUCHI NAOYA <naoya.horiguchi@....com>,
        "Aneesh Kumar K . V" <aneesh.kumar@...ux.ibm.com>,
        Waiman Long <longman@...hat.com>, Peter Xu <peterx@...hat.com>,
        Mina Almasry <almasrymina@...gle.com>,
        Hillf Danton <hdanton@...a.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Joonsoo Kim <iamjoonsoo.kim@....com>
Subject: Re: [PATCH 1/8] mm: cma: introduce cma_release_nowait()

On 3/25/21 4:49 PM, Mike Kravetz wrote:
> On 3/25/21 4:19 PM, Roman Gushchin wrote:
>> On Thu, Mar 25, 2021 at 01:12:51PM -0700, Minchan Kim wrote:
>>> On Thu, Mar 25, 2021 at 06:15:11PM +0100, David Hildenbrand wrote:
>>>> On 25.03.21 17:56, Mike Kravetz wrote:
>>>>> On 3/25/21 3:22 AM, Michal Hocko wrote:
>>>>>> On Thu 25-03-21 10:56:38, David Hildenbrand wrote:
>>>>>>> On 25.03.21 01:28, Mike Kravetz wrote:
>>>>>>>> From: Roman Gushchin <guro@...com>
>>>>>>>>
>>>>>>>> cma_release() has to lock the cma_lock mutex to clear the cma bitmap.
>>>>>>>> It makes it a blocking function, which complicates its usage from
>>>>>>>> non-blocking contexts. For instance, hugetlbfs code is temporarily
>>>>>>>> dropping the hugetlb_lock spinlock to call cma_release().
>>>>>>>>
>>>>>>>> This patch introduces a non-blocking cma_release_nowait(), which
>>>>>>>> postpones the cma bitmap clearance. It's done later from a work
>>>>>>>> context. The first page in the cma allocation is used to store
>>>>>>>> the work struct. Because CMA allocations and de-allocations are
>>>>>>>> usually not that frequent, a single global workqueue is used.
>>>>>>>>
>>>>>>>> To make sure that subsequent cma_alloc() call will pass, cma_alloc()
>>>>>>>> flushes the cma_release_wq workqueue. To avoid a performance
>>>>>>>> regression in the case when only cma_release() is used, gate it
>>>>>>>> by a per-cma area flag, which is set by the first call
>>>>>>>> of cma_release_nowait().
>>>>>>>>
>>>>>>>> Signed-off-by: Roman Gushchin <guro@...com>
>>>>>>>> [mike.kravetz@...cle.com: rebased to v5.12-rc3-mmotm-2021-03-17-22-24]
>>>>>>>> Signed-off-by: Mike Kravetz <mike.kravetz@...cle.com>
>>>>>>>> ---
>>>>>>>
>>>>>>>
>>>>>>> 1. Is there a real reason this is a mutex and not a spin lock? It seems to
>>>>>>> only protect the bitmap. Are bitmaps that huge that we spend a significant
>>>>>>> amount of time in there?
>>>>>>
>>>>>> Good question. Looking at the code it doesn't seem that there is any
>>>>>> blockable operation or any heavy lifting done under the lock.
>>>>>> 7ee793a62fa8 ("cma: Remove potential deadlock situation") has introduced
>>>>>> the lock and there was a simple bitmat protection back then. I suspect
>>>>>> the patch just followed the cma_mutex lead and used the same type of the
>>>>>> lock. cma_mutex used to protect alloc_contig_range which is sleepable.
>>>>>>
>>>>>> This all suggests that there is no real reason to use a sleepable lock
>>>>>> at all and we do not need all this heavy lifting.
>>>>>>
>>>>>
>>>>> When Roman first proposed these patches, I brought up the same issue:
>>>>>
>>>>> https://lore.kernel.org/linux-mm/20201022023352.GC300658@carbon.dhcp.thefacebook.com/
>>>>>
>>>>> Previously, Roman proposed replacing the mutex with a spinlock but
>>>>> Joonsoo was opposed.
>>>>>
>>>>> Adding Joonsoo on Cc:
>>>>>
>>>>
>>>> There has to be a good reason not to. And if there is a good reason,
>>>> lockless clearing might be one feasible alternative.
>>>
>>> I also don't think nowait variant is good idea. If the scanning of
>>> bitmap is *really* significant, it might be signal that we need to
>>> introduce different technique or data structure not bitmap rather
>>> than a new API variant.
>>
>> I'd also prefer to just replace the mutex with a spinlock rather than fiddling
>> with a delayed release.
>>
> 
> I hope Joonsoo or someone else brings up specific concerns.  I do not
> know enough about all CMA use cases.  Certainly, in this specific use
> case converting to a spinlock would not be an issue.  Do note that we
> would want to convert to an irq safe spinlock and disable irqs if that
> makes any difference in the discussion.
> 

Suggestions on how to move forward would be appreciated.  I can think of
the following options.

- Use the the cma_release_nowait() routine as defined in this patch.

- Just change the mutex to an irq safe spinlock.  AFAICT, the potential
  downsides could be:
  - Interrupts disabled during long bitmap scans
  - Wasted cpu cycles (spinning) if there is much contention on lock
  Both of these would be more of an issue on small/embedded systems. I
  took a quick look at the callers of cma_alloc/cma_release and nothing
  stood out that could lead to high degrees of contention.  However, I
  could have missed something.

- Another idea I had was to allow the user to specify the locking type
  when creating a cma area.  In this way, cma areas which may have
  release calls from atomic context would be set up with an irq safe
  spinlock.  Others, would use the mutex.  I admit this is a hackish
  way to address the issue, but perhaps not much worse than the separate
  cma_release_nowait interface?

- Change the CMA bitmap to some other data structure and algorithm.
  This would obviously take more work.

Thanks,
-- 
Mike Kravetz

Powered by blists - more mailing lists