[PATCH v4 2/5] mm: LARGE_ANON_FOLIO for improved performance
Ryan Roberts
ryan.roberts at arm.com
Wed Aug 2 01:02:18 PDT 2023
On 01/08/2023 07:36, Yu Zhao wrote:
> On Fri, Jul 28, 2023 at 4:13 AM Ryan Roberts <ryan.roberts at arm.com> wrote:
>>
>> On 27/07/2023 05:31, Yu Zhao wrote:
>>> On Wed, Jul 26, 2023 at 10:41 AM Yu Zhao <yuzhao at google.com> wrote:
>>>>
>>>> On Wed, Jul 26, 2023 at 3:52 AM Ryan Roberts <ryan.roberts at arm.com> wrote:
>>>>>
>>>>> Introduce LARGE_ANON_FOLIO feature, which allows anonymous memory to be
>>>>> allocated in large folios of a determined order. All pages of the large
>>>>> folio are pte-mapped during the same page fault, significantly reducing
>>>>> the number of page faults. The number of per-page operations (e.g. ref
>>>>> counting, rmap management lru list management) are also significantly
>>>>> reduced since those ops now become per-folio.
>>>>>
>>>>> The new behaviour is hidden behind the new LARGE_ANON_FOLIO Kconfig,
>>>>> which defaults to disabled for now; The long term aim is for this to
>>>>> defaut to enabled, but there are some risks around internal
>>>>> fragmentation that need to be better understood first.
>>>>>
>>>>> When enabled, the folio order is determined as such: For a vma, process
>>>>> or system that has explicitly disabled THP, we continue to allocate
>>>>> order-0. THP is most likely disabled to avoid any possible internal
>>>>> fragmentation so we honour that request.
>>>>>
>>>>> Otherwise, the return value of arch_wants_pte_order() is used. For vmas
>>>>> that have not explicitly opted-in to use transparent hugepages (e.g.
>>>>> where thp=madvise and the vma does not have MADV_HUGEPAGE), then
>>>>> arch_wants_pte_order() is limited to 64K (or PAGE_SIZE, whichever is
>>>>> bigger). This allows for a performance boost without requiring any
>>>>> explicit opt-in from the workload while limitting internal
>>>>> fragmentation.
>>>>>
>>>>> If the preferred order can't be used (e.g. because the folio would
>>>>> breach the bounds of the vma, or because ptes in the region are already
>>>>> mapped) then we fall back to a suitable lower order; first
>>>>> PAGE_ALLOC_COSTLY_ORDER, then order-0.
>>>>>
>>>>> arch_wants_pte_order() can be overridden by the architecture if desired.
>>>>> Some architectures (e.g. arm64) can coalsece TLB entries if a contiguous
>>>>> set of ptes map physically contigious, naturally aligned memory, so this
>>>>> mechanism allows the architecture to optimize as required.
>>>>>
>>>>> Here we add the default implementation of arch_wants_pte_order(), used
>>>>> when the architecture does not define it, which returns -1, implying
>>>>> that the HW has no preference. In this case, mm will choose it's own
>>>>> default order.
>>>>>
>>>>> Signed-off-by: Ryan Roberts <ryan.roberts at arm.com>
>>>>> ---
>>>>> include/linux/pgtable.h | 13 ++++
>>>>> mm/Kconfig | 10 +++
>>>>> mm/memory.c | 166 ++++++++++++++++++++++++++++++++++++----
>>>>> 3 files changed, 172 insertions(+), 17 deletions(-)
>>>>>
>>>>> diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
>>>>> index 5063b482e34f..2a1d83775837 100644
>>>>> --- a/include/linux/pgtable.h
>>>>> +++ b/include/linux/pgtable.h
>>>>> @@ -313,6 +313,19 @@ static inline bool arch_has_hw_pte_young(void)
>>>>> }
>>>>> #endif
>>>>>
>>>>> +#ifndef arch_wants_pte_order
>>>>> +/*
>>>>> + * Returns preferred folio order for pte-mapped memory. Must be in range [0,
>>>>> + * PMD_SHIFT-PAGE_SHIFT) and must not be order-1 since THP requires large folios
>>>>> + * to be at least order-2. Negative value implies that the HW has no preference
>>>>> + * and mm will choose it's own default order.
>>>>> + */
>>>>> +static inline int arch_wants_pte_order(void)
>>>>> +{
>>>>> + return -1;
>>>>> +}
>>>>> +#endif
>>>>> +
>>>>> #ifndef __HAVE_ARCH_PTEP_GET_AND_CLEAR
>>>>> static inline pte_t ptep_get_and_clear(struct mm_struct *mm,
>>>>> unsigned long address,
>>>>> diff --git a/mm/Kconfig b/mm/Kconfig
>>>>> index 09130434e30d..fa61ea160447 100644
>>>>> --- a/mm/Kconfig
>>>>> +++ b/mm/Kconfig
>>>>> @@ -1238,4 +1238,14 @@ config LOCK_MM_AND_FIND_VMA
>>>>>
>>>>> source "mm/damon/Kconfig"
>>>>>
>>>>> +config LARGE_ANON_FOLIO
>>>>> + bool "Allocate large folios for anonymous memory"
>>>>> + depends on TRANSPARENT_HUGEPAGE
>>>>> + default n
>>>>> + help
>>>>> + Use large (bigger than order-0) folios to back anonymous memory where
>>>>> + possible, even for pte-mapped memory. This reduces the number of page
>>>>> + faults, as well as other per-page overheads to improve performance for
>>>>> + many workloads.
>>>>> +
>>>>> endmenu
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 01f39e8144ef..64c3f242c49a 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -4050,6 +4050,127 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
>>>>> return ret;
>>>>> }
>>>>>
>>>>> +static bool vmf_pte_range_changed(struct vm_fault *vmf, int nr_pages)
>>>>> +{
>>>>> + int i;
>>>>> +
>>>>> + if (nr_pages == 1)
>>>>> + return vmf_pte_changed(vmf);
>>>>> +
>>>>> + for (i = 0; i < nr_pages; i++) {
>>>>> + if (!pte_none(ptep_get_lockless(vmf->pte + i)))
>>>>> + return true;
>>>>> + }
>>>>> +
>>>>> + return false;
>>>>> +}
>>>>> +
>>>>> +#ifdef CONFIG_LARGE_ANON_FOLIO
>>>>> +#define ANON_FOLIO_MAX_ORDER_UNHINTED \
>>>>> + (ilog2(max_t(unsigned long, SZ_64K, PAGE_SIZE)) - PAGE_SHIFT)
>>>>> +
>>>>> +static int anon_folio_order(struct vm_area_struct *vma)
>>>>> +{
>>>>> + int order;
>>>>> +
>>>>> + /*
>>>>> + * If THP is explicitly disabled for either the vma, the process or the
>>>>> + * system, then this is very likely intended to limit internal
>>>>> + * fragmentation; in this case, don't attempt to allocate a large
>>>>> + * anonymous folio.
>>>>> + *
>>>>> + * Else, if the vma is eligible for thp, allocate a large folio of the
>>>>> + * size preferred by the arch. Or if the arch requested a very small
>>>>> + * size or didn't request a size, then use PAGE_ALLOC_COSTLY_ORDER,
>>>>> + * which still meets the arch's requirements but means we still take
>>>>> + * advantage of SW optimizations (e.g. fewer page faults).
>>>>> + *
>>>>> + * Finally if thp is enabled but the vma isn't eligible, take the
>>>>> + * arch-preferred size and limit it to ANON_FOLIO_MAX_ORDER_UNHINTED.
>>>>> + * This ensures workloads that have not explicitly opted-in take benefit
>>>>> + * while capping the potential for internal fragmentation.
>>>>> + */
>>>>
>>>> What empirical evidence is SZ_64K based on?
>>>> What workloads would benefit from it?
>>>> How much would they benefit from it?
>>>> Would they benefit more or less from different values?
>>>> How much internal fragmentation would it cause?
>>>> What cost function was used to arrive at the conclusion that its
>>>> benefits outweigh its costs?
>>
>> Sorry this has taken a little while to reply to; I've been re-running my perf
>> tests with the modern patches to recomfirm old data.
>
> Thanks for the data!
>
>> In terms of empirical evidence, I've run the kernel compilation benchmark (yes I
>> know its a narrow use case, but I figure some data is better than no data), for
>> all values of ANON_FOLIO_MAX_ORDER_UNHINTED {4k, 16k, 32k, 64k, 128k, 256k}.
>>
>> I've run each test 15 times across 5 system reboots on Ampere Altra (arm64),
>
> What about x86 and ppc? Do we expect they might perform similarly wrt
> different page sizes?
It's my assumption that they should behave similarly, but I haven't actually
tested it. Thanks to Yin Fengwei for the kind offer to run these tests on x86!
Yin: I have a test tool that will automate running this and gather up the
results. Not sure if this is useful to you? I can share if you want? I also
slightly modified the code to add a boot param to set the
ANON_FOLIO_MAX_ORDER_UNHINTED value, and the test tool automatically injected
the values.
>
>> with the kernel configured for 4K base pages - I could rerun for other base page
>> sizes if we want to go further down this route.
>>
>> I've captured run time and peak memory usage, and taken the mean. The stdev for
>> the peak memory usage is big-ish, but I'm confident this still captures the
>> central tendancy well:
>>
>> | MAX_ORDER_UNHINTED | real-time | kern-time | user-time | peak memory |
>> |:-------------------|------------:|------------:|------------:|:------------|
>> | 4k | 0.0% | 0.0% | 0.0% | 0.0% |
>> | 16k | -3.6% | -26.5% | -0.5% | -0.1% |
>> | 32k | -4.8% | -37.4% | -0.6% | -0.1% |
>> | 64k | -5.7% | -42.0% | -0.6% | -1.1% |
>> | 128k | -5.6% | -42.1% | -0.7% | 1.4% |
>> | 256k | -4.9% | -41.9% | -0.4% | 1.9% |
>>
>> 64K looks like the clear sweet spot to me.
>
> Were the tests done under memory pressure?
No. I have the capability to run these tests in a memcg with limited max memory
though to force swap. I planned to do some sweeps increasing memory pressure,
all for ANON_FOLIO_MAX_ORDER_UNHINTED=64k. Doing this for all values above will
take too much time, but I could do them at a single value of max memory if that
helps? I'm not sure how I would choose that single value though? (probably do
the sweep for 64k then choose a sensible point on that graph?).
I agree 64KB might be a
> reasonable value, but I don't think we can or need to make a
> conclusion at this point: there are still pending questions from my
> list.
You mean there are still pending questions from your list above, or you have
others that you havent asked yet? If the former, I've answered all of the above
to the best of my ability. My view is that this value is always going to be
tuned empirically so its difficult to answer with absolutes. What can I do to
improve confidence? If the latter, then please let me know your other questions.
>
> Just to double check: we only need ANON_FOLIO_MAX_ORDER_UNHINTED
> because of hugepage_vma_check(), is it correct?
tldr; yes, correct.
My problem is that for arm64 16k and 64k base page configs, the contpte size is
2M. It's my view that this is too big to allocate when it has not been
explicitly asked for. And I think experience with THP (which is 2M for 4K
systems today) demonstrates that. But I would still like to benefit from reduced
SW overhead where possible (i.e. reduced page faults), and I would still like to
use the contpte 2M mappings when the user has signalled that they can tolerate
the potential internal fragmentation (MADV_HUGEPAGE).
In practical terms, ANON_FOLIO_MAX_ORDER_UNHINTED does not affect arm64/4K at
all (because the contpte size is 64K) and it does not impact the other 4K base
page arches, which don't currently implement arch_wants_pte_order() meaning they
get the default PAGE_ALLOC_COSTLY_ORDER=3=32k.
>
>> I know you have argued for using a page order in the past, rather than a size in
>> bytes. But my argument is that user space is mostly doing mmaps based on sizes
>> independent of the base page size (an assumption!) and a system's memory is
>> obviously a fixed quantity that doesn't it doesn't change with base page size.
>> So it feels more natural to limit internal fragmentation based on an absolute
>> size rather than a quantity of pages. Kyril have also suggested using absolute
>> sizes in the past [1].
>>
>> It's also worth mentioning that the file-backed memory "fault_around" mechanism
>> chooses 64K.
>
> This example actually is against your argument:
> 1. There have been multiple reports that fault around hurt
> performances and had to be disabled for some workloads over the years> -- ANON_FOLIO_MAX_ORDER_UNHINTED is likely to cause regressions too.
I don't see how ANON_FOLIO_MAX_ORDER_UNHINTED can cause regressions; it's adding
a limit making the behaviour of Large Anon Folios more similar to the old
behaviour than it otherwise would be. Without it, we will be allocating 2M
folios in some cases which would be much more likely to cause regression in
unprepared apps IMHO.
> 2. Not only can fault around be disabled, its default value can be
> changed too -- this series can't do either.
I had a mechanism for that in the previous version, but discussion concluded
that we should avoid adding the control for now and add it only if/when we have
identified a workload that will benefit.
> 3. Most importantly, fault around does not do high-order allocations
> -- this series does, and high-order allocations can be very difficult
> under memory pressure.
But ANON_FOLIO_MAX_ORDER_UNHINTED *reduces* the order from what it otherwise
would be. So I don't see how its making things worse?
>
>> If this approach really looks unacceptable, I have a couple of other ideas. But
>> I personally favour the approach that is already in the patch.
>
> I understand. If the answer to my question above is yes, then let's
> take a step back and figure out whether overloading existing ABIs is
> acceptable or not. Does this sound good to you?
Yes, good idea. Hopefully my explanation above (and all the previous threads)
gives you a good idea for the problem as I see it, and how I think hooking the
THP hints is helpful to the solution. If I've understood David's previuous
remarks correctly, then this also aligns with David's opinions (David you could
confirm/deny this please?). Could you explain your position that hooking these
ABIs is a bad approach?
>
>> 1) Add a large/small flag to arch_wants_pte_order(). arm64, at least, actually
>> has 2 mechanisms, HPA and contpte. Currently arm64 is always returning the
>> contpte order, but with a flag, it could return contpte order for large, and HPA
>> order for small. (I know we previously passed the vma and we didn't like that,
>> and this is pretty similar). I still think the SW (core-mm) needs a way to
>> sensibly limit internal fragmentation though, so personally I still think having
>> an upper limit in this case is useful.
>>
>> 2) More radical: move to a per-vma auto-tuning solution, which looks at the
>> fault pattern and maintains an allocation order in the VMA, which is modified
>> based on fault pattern. e.g. When we get faults that occur immediately adjacent
>> to the allocated range, we increase; when we get faults not connected to
>> previously allocated pages we decrease. I think it's an interesting thing to
>> look at, but certainly prefer that it's not part of an MVP implementation.
>>
>> [1]
>> https://lore.kernel.org/linux-mm/20230414140948.7pcaz6niyr2tpa7s@box.shutemov.name/
>>
>>
>>>>
>>>>> + if ((vma->vm_flags & VM_NOHUGEPAGE) ||
>>>>> + test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags) ||
>>>>> + !hugepage_flags_enabled())
>>>>> + order = 0;
>>>>> + else {
>>>>> + order = max(arch_wants_pte_order(), PAGE_ALLOC_COSTLY_ORDER);
>>>>> +
>>>>> + if (!hugepage_vma_check(vma, vma->vm_flags, false, true, true))
>>>>> + order = min(order, ANON_FOLIO_MAX_ORDER_UNHINTED);
>>>>> + }
>>>
>>> I'm a bit surprised to see the above: why can we overload existing
>>> ABIs? I don't think we can.
>>
>> I think this is all covered by the conversation with David against v2; see [2]
>> and proceeding replies. Argument is that VM_NOHUGEPAGE (and friends) is really a
>> request from user space to optimize for the least memory wastage possible and
>> avoid populating ptes that have not been expressly requested.
>>
>> [2]
>> https://lore.kernel.org/linux-mm/524bacd2-4a47-2b8b-6685-c46e31a01631@redhat.com/
>
> Thanks for the info.
>
> I think there might be a misunderstanding here.
>
> David, can you please clarify whether you suggested we overland
> (change the semantics) of existing ABIs?
>
> This sounds like a big red flag to me. If that's really what you
> suggest, can you shed some light on why this is acceptable to existing
> userspace at all?
>
> Thanks.
More information about the linux-arm-kernel
mailing list