Re: [PATCH v8 6/7] mm, folio_zero_user: support clearing page ranges

[Ankur Arora <ankur.a.arora@oracle.com>]

David Hildenbrand (Red Hat) <david@kernel.org> writes:

> On 10.11.25 08:20, Ankur Arora wrote:
>> David Hildenbrand (Red Hat) <david@kernel.org> writes:
>>
>>> On 27.10.25 21:21, Ankur Arora wrote:
>>>> Clear contiguous page ranges in folio_zero_user() instead of clearing
>>>> a page-at-a-time. This enables CPU specific optimizations based on
>>>> the length of the region.
>>>> Operating on arbitrarily large regions can lead to high preemption
>>>> latency under cooperative preemption models. So, limit the worst
>>>> case preemption latency via architecture specified PAGE_CONTIG_NR
>>>> units.
>>>> The resultant performance depends on the kinds of optimizations
>>>> available to the CPU for the region being cleared. Two classes of
>>>> of optimizations:
>>>>     - clearing iteration costs can be amortized over a range larger
>>>>       than a single page.
>>>>     - cacheline allocation elision (seen on AMD Zen models).
>>>> Testing a demand fault workload shows an improved baseline from the
>>>> first optimization and a larger improvement when the region being
>>>> cleared is large enough for the second optimization.
>>>> AMD Milan (EPYC 7J13, boost=0, region=64GB on the local NUMA node):
>>>>    $ perf bench mem map -p $pg-sz -f demand -s 64GB -l 5
>>>>                       page-at-a-time     contiguous clearing      change
>>>>                     (GB/s  +- %stdev)     (GB/s  +- %stdev)
>>>>      pg-sz=2MB       12.92  +- 2.55%        17.03  +-  0.70%       + 31.8%
>>>> preempt=*
>>>>      pg-sz=1GB       17.14  +- 2.27%        18.04  +-  1.05% [#]   +  5.2%
>>>> preempt=none|voluntary
>>>>      pg-sz=1GB       17.26  +- 1.24%        42.17  +-  4.21%       +144.3%	preempt=full|lazy
>>>> [#] AMD Milan uses a threshold of LLC-size (~32MB) for eliding cacheline
>>>> allocation, which is larger than ARCH_PAGE_CONTIG_NR, so
>>>> preempt=none|voluntary see no improvement on the pg-sz=1GB.
>>>> Also as mentioned earlier, the baseline improvement is not specific to
>>>> AMD Zen platforms. Intel Icelakex (pg-sz=2MB|1GB) sees a similar
>>>> improvement as the Milan pg-sz=2MB workload above (~30%).
>>>> Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
>>>> Reviewed-by: Raghavendra K T <raghavendra.kt@amd.com>
>>>> Tested-by: Raghavendra K T <raghavendra.kt@amd.com>
>>>> ---
>>>>    include/linux/mm.h |  6 ++++++
>>>>    mm/memory.c        | 42 +++++++++++++++++++++---------------------
>>>>    2 files changed, 27 insertions(+), 21 deletions(-)
>>>> diff --git a/include/linux/mm.h b/include/linux/mm.h
>>>> index ecbcb76df9de..02db84667f97 100644
>>>> --- a/include/linux/mm.h
>>>> +++ b/include/linux/mm.h
>>>> @@ -3872,6 +3872,12 @@ static inline void clear_page_guard(struct zone *zone, struct page *page,
>>>>    				unsigned int order) {}
>>>>    #endif	/* CONFIG_DEBUG_PAGEALLOC */
>>>>    +#ifndef ARCH_PAGE_CONTIG_NR
>>>> +#define PAGE_CONTIG_NR	1
>>>> +#else
>>>> +#define PAGE_CONTIG_NR	ARCH_PAGE_CONTIG_NR
>>>> +#endif
>>>
>>> The name is a bit misleading. We need something that tells us that this is for
>>> patch-processing (clearing? maybe alter copying?) contig pages. Likely spelling
>>> out that this is for the non-preemptible case only.
>>>
>>> I assume we can drop the "CONTIG", just like clear_pages() doesn't contain it
>>> etc.
>>>
>>> CLEAR_PAGES_NON_PREEMPT_BATCH
>>>
>>> PROCESS_PAGES_NON_PREEMPT_BATCH
>> I think this version is clearer. And would be viable for copying as well.
>>
>>> Can you remind me again why this is arch specific, and why the default is 1
>>> instead of, say 2,4,8 ... ?
>> So, the only use for this value is to decide a reasonable frequency
>> for calling cond_resched() when operating on hugepages.
>> And the idea was the arch was best placed to have a reasonably safe
>> value based on the expected spread of bandwidths it might see across
>> uarchs. And the default choice of 1 was to keep it close to what we
>> have now.
>> Thinking about it now though, maybe it is better to instead do this
>> in common code. We could have two sets of defines,
>> PROCESS_PAGES_NON_PREEMPT_BATCH_{LARGE,SMALL}, the first for archs
>> that define __HAVE_ARCH_CLEAR_PAGES and the second, without.
>
> Right, avoiding this dependency on arch code would be nice.
>
> Also, it feels like something we can later optimize for archs without
> __HAVE_ARCH_CLEAR_PAGES in common code.

That makes sense. Will keep the default where it is (1) and just get
rid of the arch dependency.

--
ankur