Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <4b47e6317aca3deeabf610a7f4839563ff2b25a1.camel@intel.com>
Subject: Re: [PATCH v2 2/9] mm/vmscan: remove unneeded can_split_huge_page
 check
From:   "ying.huang@intel.com" <ying.huang@intel.com>
To:     David Hildenbrand <david@redhat.com>,
        Miaohe Lin <linmiaohe@huawei.com>,
        Oscar Salvador <osalvador@suse.de>
Cc:     akpm@linux-foundation.org, songmuchun@bytedance.com,
        hch@infradead.org, willy@infradead.org, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org,
        Pavel Tatashin <pasha.tatashin@soleen.com>,
        John Hubbard <jhubbard@nvidia.com>,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Vlastimil Babka <vbabka@suse.cz>, Yu Zhao <yuzhao@google.com>
Date:   Fri, 15 Apr 2022 11:07:53 +0800
In-Reply-To: <a90dc108e54e19cf3aa3fd21f6321afa8f194adc.camel@intel.com>
References: <20220409093500.10329-1-linmiaohe@huawei.com>
         <20220409093500.10329-3-linmiaohe@huawei.com>
         <YlU/h0fdE1L846Bd@localhost.localdomain>
         <7455b680-3d89-5d3e-ba0e-6e4358b114a2@huawei.com>
         <b153b758-ce11-364a-2699-753b21250508@redhat.com>
         <a90dc108e54e19cf3aa3fd21f6321afa8f194adc.camel@intel.com>
Content-Type: text/plain; charset="UTF-8"
User-Agent: Evolution 3.38.3-1 
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Precedence: bulk

On Wed, 2022-04-13 at 09:26 +0800, ying.huang@intel.com wrote:
> On Tue, 2022-04-12 at 16:59 +0200, David Hildenbrand wrote:
> > On 12.04.22 15:42, Miaohe Lin wrote:
> > > On 2022/4/12 16:59, Oscar Salvador wrote:
> > > > On Sat, Apr 09, 2022 at 05:34:53PM +0800, Miaohe Lin wrote:
> > > > > We don't need to check can_split_folio() because folio_maybe_dma_pinned()
> > > > > is checked before. It will avoid the long term pinned pages to be swapped
> > > > > out. And we can live with short term pinned pages. Without can_split_folio
> > > > > checking we can simplify the code. Also activate_locked can be changed to
> > > > > keep_locked as it's just short term pinning.
> > > > 
> > > > What do you mean by "we can live with short term pinned pages"?
> > > > Does it mean that it was not pinned when we check
> > > > folio_maybe_dma_pinned() but now it is?
> > > > 
> > > > To me it looks like the pinning is fluctuating and we rely on
> > > > split_folio_to_list() to see whether we succeed or not, and if not
> > > > we give it another spin in the next round?
> > > 
> > > Yes. Short term pinned pages is relative to long term pinned pages and these pages won't be
> > > pinned for a noticeable time. So it's expected to split the folio successfully in the next
> > > round as the pinning is really fluctuating. Or am I miss something?
> > > 
> > 
> > Just so we're on the same page. folio_maybe_dma_pinned() only capture
> > FOLL_PIN, but not FOLL_GET. You can have long-term FOLL_GET right now
> > via vmsplice().
> 
> Per my original understanding, folio_maybe_dma_pinned() can be used to
> detect long-term pinned pages.  And it seems reasonable to skip the
> long-term pinned pages and try short-term pinned pages during page
> reclaiming.  But as you pointed out, vmsplice() doesn't use FOLL_PIN. 
> So if vmsplice() is expected to pin pages for long time, and we have no
> way to detect it, then we should keep can_split_folio() in the original
> code.
> 
> Copying more people who have worked on long-term pinning for comments.

Checked the discussion in the following thread,

https://lore.kernel.org/lkml/CA+CK2bBffHBxjmb9jmSKacm0fJMinyt3Nhk8Nx6iudcQSj80_w@mail.gmail.com/

It seems that from the practical point of view, folio_maybe_dma_pinned()
can identify most long-term pinned pages that may block memory hot-
remove or CMA allocation.  Although as David pointed out, some pages may
still be GUPed for long time (e.g. via vmsplice) even if
!folio_maybe_dma_pinned().

But from another point of view, can_split_huge_page() is cheap and THP
swapout is expensive (swap space, disk IO, and hard to be recovered), so
it may be better to keep can_split_huge_page() in shink_page_list().

Best Regards,
Huang, Ying

> 
> > can_split_folio() is more precise then folio_maybe_dma_pinned(), but
> > both are racy as long as the page is still mapped.
> > 
> > 
>