linux-kernel - Re: [PATCH] arm64: mm: drop tlb flush operation when clearing the access bit

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <e818ef93-49f6-47c8-0d89-6330bf6687f8@linux.alibaba.com>
Date:   Wed, 25 Oct 2023 09:44:34 +0800
From:   Baolin Wang <baolin.wang@...ux.alibaba.com>
To:     Kefeng Wang <wangkefeng.wang@...wei.com>, catalin.marinas@....com,
        will@...nel.org
Cc:     akpm@...ux-foundation.org, v-songbaohua@...o.com,
        yuzhao@...gle.com, linux-mm@...ck.org,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] arm64: mm: drop tlb flush operation when clearing the
 access bit



On 10/24/2023 9:48 PM, Kefeng Wang wrote:
> 
> 
> On 2023/10/24 20:56, Baolin Wang wrote:
>> Now ptep_clear_flush_young() is only called by folio_referenced() to
>> check if the folio was referenced, and now it will call a tlb flush on
>> ARM64 architecture. However the tlb flush can be expensive on ARM64
>> servers, especially for the systems with a large CPU numbers.
>>
>> Similar to the x86 architecture, below comments also apply equally to
>> ARM64 architecture. So we can drop the tlb flush operation in
>> ptep_clear_flush_young() on ARM64 architecture to improve the 
>> performance.
>> "
>> /* Clearing the accessed bit without a TLB flush
>>   * doesn't cause data corruption. [ It could cause incorrect
>>   * page aging and the (mistaken) reclaim of hot pages, but the
>>   * chance of that should be relatively low. ]
>>   *
>>   * So as a performance optimization don't flush the TLB when
>>   * clearing the accessed bit, it will eventually be flushed by
>>   * a context switch or a VM operation anyway. [ In the rare
>>   * event of it not getting flushed for a long time the delay
>>   * shouldn't really matter because there's no real memory
>>   * pressure for swapout to react to. ]
>>   */
>> "
>> Running the thpscale to show some obvious improvements for compaction
>> latency with this patch:
>>                               base                   patched
>> Amean     fault-both-1      1093.19 (   0.00%)     1084.57 *   0.79%*
>> Amean     fault-both-3      2566.22 (   0.00%)     2228.45 *  13.16%*
>> Amean     fault-both-5      3591.22 (   0.00%)     3146.73 *  12.38%*
>> Amean     fault-both-7      4157.26 (   0.00%)     4113.67 *   1.05%*
>> Amean     fault-both-12     6184.79 (   0.00%)     5218.70 *  15.62%*
>> Amean     fault-both-18     9103.70 (   0.00%)     7739.71 *  14.98%*
>> Amean     fault-both-24    12341.73 (   0.00%)    10684.23 *  13.43%*
>> Amean     fault-both-30    15519.00 (   0.00%)    13695.14 *  11.75%*
>> Amean     fault-both-32    16189.15 (   0.00%)    14365.73 *  11.26%*
>>                         base       patched
>> Duration User         167.78      161.03
>> Duration System      1836.66     1673.01
>> Duration Elapsed     2074.58     2059.75
>>
>> Barry Song submitted a similar patch [1] before, that replaces the
>> ptep_clear_flush_young_notify() with ptep_clear_young_notify() in
>> folio_referenced_one(). However, I'm not sure if removing the tlb flush
>> operation is applicable to every architecture in kernel, so dropping
>> the tlb flush for ARM64 seems a sensible change.
> 
> At least x86/s390/riscv/powerpc already do it, also I think we could

Right.

> change pmdp_clear_flush_young_notify() too, since it is same with
> ptep_clear_flush_young_notify(),

Perhaps yes, but I'm still unsure if removing tlb flush for PMD entry is 
applicable to all architectures. Let's see the discussion in this 
thread. Thanks.