Received-SPF: pass (google.com: domain of linux-kernel+bounces-205818-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Message-ID: <e4d167fe-cb1e-41d1-a144-00bfa14b7148@gmail.com>
Date: Fri, 7 Jun 2024 11:24:13 +0100
Precedence: bulk
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [PATCH v2 1/2] mm: clear pte for folios that are zero filled
To: David Hildenbrand <david@redhat.com>, akpm@linux-foundation.org,
 shakeel.butt@linux.dev, yosryahmed@google.com, willy@infradead.org
Cc: hannes@cmpxchg.org, nphamcs@gmail.com, chengming.zhou@linux.dev,
 linux-mm@kvack.org, linux-kernel@vger.kernel.org, kernel-team@meta.com
References: <20240604105950.1134192-1-usamaarif642@gmail.com>
 <20240604105950.1134192-2-usamaarif642@gmail.com>
 <6b1485b6-c2a1-45b8-8afe-7b211689070b@redhat.com>
 <f26a6ac2-48a9-4bae-89b9-a3f9b97ae9dc@redhat.com>
Content-Language: en-US
From: Usama Arif <usamaarif642@gmail.com>
In-Reply-To: <f26a6ac2-48a9-4bae-89b9-a3f9b97ae9dc@redhat.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit


On 04/06/2024 13:43, David Hildenbrand wrote:
> On 04.06.24 14:30, David Hildenbrand wrote:
>> On 04.06.24 12:58, Usama Arif wrote:
>>> Approximately 10-20% of pages to be swapped out are zero pages [1].
>>> Rather than reading/writing these pages to flash resulting
>>> in increased I/O and flash wear, the pte can be cleared for those
>>> addresses at unmap time while shrinking folio list. When this
>>> causes a page fault, do_pte_missing will take care of this page.
>>> With this patch, NVMe writes in Meta server fleet decreased
>>> by almost 10% with conventional swap setup (zswap disabled).
>>>
>>> [1] 
>>> https://lore.kernel.org/all/20171018104832epcms5p1b2232e2236258de3d03d1344dde9fce0@epcms5p1/
>>>
>>> Signed-off-by: Usama Arif <usamaarif642@gmail.com>
>>> ---
>>>    include/linux/rmap.h |   1 +
>>>    mm/rmap.c            | 163 
>>> ++++++++++++++++++++++---------------------
>>>    mm/vmscan.c          |  89 ++++++++++++++++-------
>>>    3 files changed, 150 insertions(+), 103 deletions(-)
>>>
>>> diff --git a/include/linux/rmap.h b/include/linux/rmap.h
>>> index bb53e5920b88..b36db1e886e4 100644
>>> --- a/include/linux/rmap.h
>>> +++ b/include/linux/rmap.h
>>> @@ -100,6 +100,7 @@ enum ttu_flags {
>>>                         * do a final flush if necessary */
>>>        TTU_RMAP_LOCKED        = 0x80,    /* do not grab rmap lock:
>>>                         * caller holds it */
>>> +    TTU_ZERO_FOLIO        = 0x100,/* zero folio */
>>>    };
>>>       #ifdef CONFIG_MMU
>>> diff --git a/mm/rmap.c b/mm/rmap.c
>>> index 52357d79917c..d98f70876327 100644
>>> --- a/mm/rmap.c
>>> +++ b/mm/rmap.c
>>> @@ -1819,96 +1819,101 @@ static bool try_to_unmap_one(struct folio 
>>> *folio, struct vm_area_struct *vma,
>>>                 */
>>>                dec_mm_counter(mm, mm_counter(folio));
>>>            } else if (folio_test_anon(folio)) {
>>> -            swp_entry_t entry = page_swap_entry(subpage);
>>> -            pte_t swp_pte;
>>> -            /*
>>> -             * Store the swap location in the pte.
>>> -             * See handle_pte_fault() ...
>>> -             */
>>> -            if (unlikely(folio_test_swapbacked(folio) !=
>>> -                    folio_test_swapcache(folio))) {
>>> +            if (flags & TTU_ZERO_FOLIO) {
>>> +                pte_clear(mm, address, pvmw.pte);
>>> +                dec_mm_counter(mm, MM_ANONPAGES);
>>
>> Is there an easy way to reduce the code churn and highlight the added 
>> code?
>>
>> Like
>>
>> } else if (folio_test_anon(folio) && (flags & TTU_ZERO_FOLIO)) {
>>
>> } else if (folio_test_anon(folio)) {
>>
>>
>>
>> Also to concerns that I want to spell out:
>>
>> (a) what stops the page from getting modified in the meantime? The CPU
>>       can write it until the TLB was flushed.
>>
Thanks for pointing this out David and Shakeel. This is a big issue in 
this v2, and as Shakeel pointed out in [1] we need to do a second rmap 
walk. Looking at how ksm deals with this in try_to_merge_one_page which 
calls write_protect_page for each vma (i.e. basically an rmap walk), 
this would be much more CPU expensive and complicated compared to v1 
[2], where the swap subsystem can handle all complexities. I will go 
back to my v1 solution for the next revision as its much more simpler 
and the memory usage is very low (0.003%) as pointed out by Johannes [3] 
which would likely go away with the memory savings of not having a 
zswap_entry for zero filled pages, and the solution being a lot simpler 
than what a valid v2 approach would look like.


[1] 
https://lore.kernel.org/all/nes73bwc5p6yhwt5tw3upxcqrn5kenn6lvqb6exrf4yppmz6jx@ywhuevpkxlvh/

[2] 
https://lore.kernel.org/all/20240530102126.357438-1-usamaarif642@gmail.com/

[3] https://lore.kernel.org/all/20240530122715.GB1222079@cmpxchg.org/

>> (b) do you properly handle if the page is pinned (or just got pinned)
>>       and we must not discard it?
>
> Oh, and I forgot, are you handling userfaultd as expected? IIRC there 
> are some really nasty side-effects with userfaultfd even when 
> userfaultfd is currently not registered for a VMA [1].
>
> [1] 
> https://lore.kernel.org/linux-mm/3a4b1027-df6e-31b8-b0de-ff202828228d@redhat.com/
>
> What should work is replacing all-zero anonymous pages by the shared 
> zeropage iff the anonymous page is not pinned and we synchronize 
> against GUP fast. Well, and we handle possible concurrent writes 
> accordingly.
>
> KSM does essentially that when told to de-duplicate the shared 
> zeropage, and I was thinking a while ago if we would want a 
> zeropage-only KSM version that doesn't need stable tress and all that, 
> but only deduplicates zero-filled pages into the shared zeropage in a 
> safe way.
>
Thanks for the pointer to KSM code.