[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <5ef6df9b-f4a7-43d1-9de7-19ad2d82bbff@arm.com>
Date: Mon, 1 Sep 2025 14:20:07 +0530
From: Dev Jain <dev.jain@....com>
To: David Hildenbrand <david@...hat.com>, akpm@...ux-foundation.org,
kas@...nel.org, willy@...radead.org, hughd@...gle.com
Cc: ziy@...dia.com, baolin.wang@...ux.alibaba.com,
lorenzo.stoakes@...cle.com, Liam.Howlett@...cle.com, npache@...hat.com,
ryan.roberts@....com, baohua@...nel.org, linux-mm@...ck.org,
linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH] mm: Enable khugepaged to operate on non-writable VMAs
On 01/09/25 2:02 pm, David Hildenbrand wrote:
> On 01.09.25 09:48, Dev Jain wrote:
>> Currently khugepaged does not collapse a region which does not have a
>> single writable page. This is wasteful since, apart from any
>> non-writable
>> memory mapped by the application, there are a lot of non-writable VMAs
>> which will benefit from collapsing - the VMAs of the executable, those
>> of the glibc, vvar and vdso, which won't be unmapped during the lifetime
>> of the process, as opposed to other VMAs which maybe unmapped.
>
> Are these anonymous folios? ("VMAs of the executable"), or is you
> description
> misleading?
Oops. I dropped writable everywhere and failed to notice that the
callsites are
all for anon collapse. So I'll only mention vvar, vdso and other
non-writable
VMAs in the description.
>
>> Therefore,
>> remove this restriction and allow khugepaged to collapse a VMA with
>> arbitrary protections.
>>
>> Along with this, currently MADV_COLLAPSE does not perform a collapse
>> on a
>> non-writable VMA, and this restriction is nowhere to be found on the
>> manpage - the restriction itself sounds wrong to me since the user knows
>> the protection of the memory it has mapped, so collapsing read-only
>> memory via madvise() should be a choice of the user which shouldn't
>> be overriden by the kernel.
>>
>> I dug into the history of this and couldn't find any concrete reason of
>> the current behaviour - [1] is the v1 of the original khugepaged patch
>> which required all ptes to be writable. [2] is the v1 of the patch which
>> changed this behaviour to require at least one pte to be writable. The
>> closest thing I could find was: in response to [2], Kirill says in [3] -
>> "As a side effect it will effectively allow collapse in PROT_READ vmas,
>> right? I'm not convinced it's a good idea." (Although Kirill realizes in
>> [4] that this was not the intention of the patch).
>>
>> I can see performance improvements on mmtests run on an arm64 machine
>> comparing with 6.17-rc2. (I) denotes statistically significant
>> improvement,
>> (R) denotes statistically significant regression (Please ignore the
>> numbers in the middle column):
>
> I once dug into that myself as well as part of
>
> commit 1bafe96e89f056cb6e25d47451fb16aee2c7c4d0
> Author: David Hildenbrand <david@...hat.com>
> Date: Wed Apr 24 14:26:30 2024 +0200
>
> mm/khugepaged: replace page_mapcount() check by
> folio_likely_mapped_shared()
>
> where I noted:
>
> Interestingly, khugepaged will only collapse an anonymous THP if
> at least
> one PTE is writable. After fork(), that means that something
> (usually a
> page fault) populated at least a single exclusive anonymous THP in
> that
> PMD range.
> The problem I was concerned with (also documented in that patch)
> should no
> longer apply ever since we changed how folio_maybe_mapped_shared()
> operates.
>
> So yes, I don't see a good reason to fail on R/O PTEs
>
>>
>> +------------------------------------+----------------------------------------------------------+-----------------------+--------------------------+
>>
>> | mmtests/hackbench | process-pipes-1
>> (seconds) | 0.145
>> | -0.06% |
>> | | process-pipes-4
>> (seconds) | 0.4335
>> | -0.27% |
>> | | process-pipes-7
>> (seconds) | 0.823
>> | (I) -12.13% |
>> | | process-pipes-12
>> (seconds) | 1.3538333333333334
>> | (I) -5.32% |
>> | | process-pipes-21
>> (seconds) | 1.8971666666666664
>> | (I) -2.87% |
>> | | process-pipes-30
>> (seconds) | 2.5023333333333335
>> | (I) -3.39% |
>> | | process-pipes-48
>> (seconds) | 3.4305
>> | (I) -5.65% |
>> | | process-pipes-79
>> (seconds) | 4.245833333333334
>> | (I) -6.74% |
>> | | process-pipes-110
>> (seconds) | 5.114833333333333
>> | (I) -6.26% |
>> | | process-pipes-141
>> (seconds) | 6.1885
>> | (I) -4.99% |
>> | | process-pipes-172
>> (seconds) | 7.231833333333334
>> | (I) -4.45% |
>> | | process-pipes-203
>> (seconds) | 8.393166666666668
>> | (I) -3.65% |
>> | | process-pipes-234
>> (seconds) | 9.487499999999999
>> | (I) -3.45% |
>> | | process-pipes-256
>> (seconds) | 10.316166666666666
>> | (I) -3.47% |
>> | | process-sockets-1
>> (seconds) | 0.289
>> | 2.13% |
>> | | process-sockets-4
>> (seconds) | 0.7596666666666666
>> | 1.02% |
>> | | process-sockets-7
>> (seconds) | 1.1663333333333334
>> | -0.26% |
>> | | process-sockets-12
>> (seconds) | 1.8641666666666665
>> | -1.24% |
>> | | process-sockets-21
>> (seconds) | 3.0773333333333333
>> | 0.01% |
>> | | process-sockets-30
>> (seconds) | 4.2405
>> | -0.15% |
>> | | process-sockets-48
>> (seconds) | 6.459666666666666
>> | 0.15% |
>> | | process-sockets-79
>> (seconds) | 10.156833333333333
>> | 1.45% |
>> | | process-sockets-110
>> (seconds) | 14.317833333333333
>> | -1.64% |
>> | | process-sockets-141
>> (seconds) | 20.8735
>> | (I) -4.27% |
>> | | process-sockets-172
>> (seconds) | 26.205333333333332
>> | 0.30% |
>> | | process-sockets-203
>> (seconds) | 31.298000000000002
>> | -1.71% |
>> | | process-sockets-234
>> (seconds) | 36.104000000000006
>> | -1.94% |
>> | | process-sockets-256
>> (seconds) | 39.44016666666667
>> | -0.71% |
>> | | thread-pipes-1
>> (seconds) | 0.17550000000000002
>> | 0.66% |
>> | | thread-pipes-4
>> (seconds) | 0.44716666666666666
>> | 1.66% |
>> | | thread-pipes-7
>> (seconds) | 0.7345
>> | -0.17% |
>> | | thread-pipes-12
>> (seconds) | 1.405833333333333
>> | (I) -4.12% |
>> | | thread-pipes-21
>> (seconds) | 2.0113333333333334
>> | (I) -2.13% |
>> | | thread-pipes-30
>> (seconds) | 2.6648333333333336
>> | (I) -3.78% |
>> | | thread-pipes-48
>> (seconds) | 3.6341666666666668
>> | (I) -5.77% |
>> | | thread-pipes-79
>> (seconds) | 4.4085
>> | (I) -5.31% |
>> | | thread-pipes-110
>> (seconds) | 5.374666666666666
>> | (I) -6.12% |
>> | | thread-pipes-141
>> (seconds) | 6.385666666666666
>> | (I) -4.00% |
>> | | thread-pipes-172
>> (seconds) | 7.403000000000001
>> | (I) -3.01% |
>> | | thread-pipes-203
>> (seconds) | 8.570333333333332
>> | (I) -2.62% |
>> | | thread-pipes-234
>> (seconds) | 9.719166666666666
>> | (I) -2.00% |
>> | | thread-pipes-256
>> (seconds) | 10.552833333333334
>> | (I) -2.30% |
>> | | thread-sockets-1
>> (seconds) | 0.3065
>> | (R) 2.39% |
>> +------------------------------------+----------------------------------------------------------+-----------------------+--------------------------+
>>
>>
>> +------------------------------------+----------------------------------------------------------+-----------------------+--------------------------+
>>
>> | mmtests/sysbench-mutex | sysbenchmutex-1
>> (usec) | 194.38333333333333
>> | -0.02% |
>> | | sysbenchmutex-4
>> (usec) | 200.875
>> | -0.02% |
>> | | sysbenchmutex-7
>> (usec) | 201.23000000000002
>> | 0.00% |
>> | | sysbenchmutex-12
>> (usec) | 201.77666666666664
>> | 0.12% |
>> | | sysbenchmutex-21
>> (usec) | 203.03
>> | -0.40% |
>> | | sysbenchmutex-30
>> (usec) | 203.285
>> | 0.08% |
>> | | sysbenchmutex-48
>> (usec) | 231.30000000000004
>> | 2.59% |
>> | | sysbenchmutex-79
>> (usec) | 362.075
>> | -0.80% |
>> | | sysbenchmutex-110
>> (usec) | 516.8233333333334
>> | -3.87% |
>> | | sysbenchmutex-128
>> (usec) | 593.3533333333334
>> | (I) -4.46% |
>> +------------------------------------+----------------------------------------------------------+-----------------------+--------------------------+
>>
>>
>> No regressions were observed with mm-selftests.
>>
>> [1]
>> https://lore.kernel.org/all/679861e2e81b32a0ae08.1264054854@v2.random/
>> [2]
>> https://lore.kernel.org/all/1421999256-3881-1-git-send-email-ebru.akagunduz@gmail.com/
>> [3] https://lore.kernel.org/all/20150123113701.GB5975@node.dhcp.inet.fi/
>> [4] https://lore.kernel.org/all/20150123155802.GA7011@node.dhcp.inet.fi/
>>
>> Signed-off-by: Dev Jain <dev.jain@....com>
>> ---
>> Based on mm-new.
>>
>> Not very sure of the tracing parts which this patch changes. I have kept
>> the writable portion for the tracing to maintain backward compat, just
>> dropped it as a collapse condition.
>>
>> include/trace/events/huge_memory.h | 2 +-
>> mm/khugepaged.c | 11 +++--------
>> 2 files changed, 4 insertions(+), 9 deletions(-)
>>
>> diff --git a/include/trace/events/huge_memory.h
>> b/include/trace/events/huge_memory.h
>> index 2305df6cb485..f2472c1c132a 100644
>> --- a/include/trace/events/huge_memory.h
>> +++ b/include/trace/events/huge_memory.h
>> @@ -19,7 +19,7 @@
>> EM( SCAN_PTE_NON_PRESENT, "pte_non_present") \
>> EM( SCAN_PTE_UFFD_WP, "pte_uffd_wp") \
>> EM( SCAN_PTE_MAPPED_HUGEPAGE, "pte_mapped_hugepage") \
>> - EM( SCAN_PAGE_RO, "no_writable_page") \
>> + EM( SCAN_PAGE_RO, "no_writable_page") /* deprecated */ \
>> EM( SCAN_LACK_REFERENCED_PAGE, "lack_referenced_page") \
>> EM( SCAN_PAGE_NULL, "page_null") \
>> EM( SCAN_SCAN_ABORT, "scan_aborted") \
>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>> index 4ec324a4c1fe..5ef8482597a9 100644
>> --- a/mm/khugepaged.c
>> +++ b/mm/khugepaged.c
>> @@ -39,7 +39,7 @@ enum scan_result {
>> SCAN_PTE_NON_PRESENT,
>> SCAN_PTE_UFFD_WP,
>> SCAN_PTE_MAPPED_HUGEPAGE,
>> - SCAN_PAGE_RO,
>> + SCAN_PAGE_RO, /* deprecated */
>
> Why can't we remove that completely.
>
>
Powered by blists - more mailing lists