Re: [RFC PATCH v1] mm/filemap: Allow arch to request folio size for exec memory

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Sat, Jan 13, 2024 at 12:15 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>
> On 12/01/2024 10:13, Barry Song wrote:
> > On Fri, Jan 12, 2024 at 4:41 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> >>
> >> Change the readahead config so that if it is being requested for an
> >> executable mapping, do a synchronous read of an arch-specified size in a
> >> naturally aligned manner.
> >>
> >> On arm64 if memory is physically contiguous and naturally aligned to the
> >> "contpte" size, we can use contpte mappings, which improves utilization
> >> of the TLB. When paired with the "multi-size THP" changes, this works
> >> well to reduce dTLB pressure. However iTLB pressure is still high due to
> >> executable mappings having a low liklihood of being in the required
> >> folio size and mapping alignment, even when the filesystem supports
> >> readahead into large folios (e.g. XFS).
> >>
> >> The reason for the low liklihood is that the current readahead algorithm
> >> starts with an order-2 folio and increases the folio order by 2 every
> >> time the readahead mark is hit. But most executable memory is faulted in
> >> fairly randomly and so the readahead mark is rarely hit and most
> >> executable folios remain order-2. This is observed impirically and
> >> confirmed from discussion with a gnu linker expert; in general, the
> >> linker does nothing to group temporally accessed text together
> >> spacially. Additionally, with the current read-around approach there are
> >> no alignment guarrantees between the file and folio. This is
> >> insufficient for arm64's contpte mapping requirement (order-4 for 4K
> >> base pages).
> >>
> >> So it seems reasonable to special-case the read(ahead) logic for
> >> executable mappings. The trade-off is performance improvement (due to
> >> more efficient storage of the translations in iTLB) vs potential read
> >> amplification (due to reading too much data around the fault which won't
> >> be used), and the latter is independent of base page size. I've chosen
> >> 64K folio size for arm64 which benefits both the 4K and 16K base page
> >> size configs and shouldn't lead to any further read-amplification since
> >> the old read-around path was (usually) reading blocks of 128K (with the
> >> last 32K being async).
> >>
> >> Performance Benchmarking
> >> ------------------------
> >>
> >> The below shows kernel compilation and speedometer javascript benchmarks
> >> on Ampere Altra arm64 system. (The contpte patch series is applied in
> >> the baseline).
> >>
> >> First, confirmation that this patch causes more memory to be contained
> >> in 64K folios (this is for all file-backed memory so includes
> >> non-executable too):
> >>
> >> | File-backed folios      |   Speedometer   |  Kernel Compile |
> >> | by size as percentage   |-----------------|-----------------|
> >> | of all mapped file mem  | before |  after | before |  after |
> >> |=========================|========|========|========|========|
> >> |file-thp-aligned-16kB    |    45% |     9% |    46% |     7% |
> >> |file-thp-aligned-32kB    |     2% |     0% |     3% |     1% |
> >> |file-thp-aligned-64kB    |     3% |    63% |     5% |    80% |
> >> |file-thp-aligned-128kB   |    11% |    11% |     0% |     0% |
> >> |file-thp-unaligned-16kB  |     1% |     0% |     3% |     1% |
> >> |file-thp-unaligned-128kB |     1% |     0% |     0% |     0% |
> >> |file-thp-partial         |     0% |     0% |     0% |     0% |
> >> |-------------------------|--------|--------|--------|--------|
> >> |file-cont-aligned-64kB   |    16% |    75% |     5% |    80% |
> >>
> >> The above shows that for both use cases, the amount of file memory
> >> backed by 16K folios reduces and the amount backed by 64K folios
> >> increases significantly. And the amount of memory that is contpte-mapped
> >> significantly increases (last line).
> >>
> >> And this is reflected in performance improvement:
> >>
> >> Kernel Compilation (smaller is faster):
> >> | kernel   |   real-time |   kern-time |   user-time |   peak memory |
> >> |----------|-------------|-------------|-------------|---------------|
> >> | before   |        0.0% |        0.0% |        0.0% |          0.0% |
> >> | after    |       -1.6% |       -2.1% |       -1.7% |          0.0% |
> >>
> >> Speedometer (bigger is faster):
> >> | kernel   |   runs_per_min |   peak memory |
> >> |----------|----------------|---------------|
> >> | before   |           0.0% |          0.0% |
> >> | after    |           1.3% |          1.0% |
> >>
> >> Both benchmarks show a ~1.5% improvement once the patch is applied.
> >>
> >> Alternatives
> >> ------------
> >>
> >> I considered (and rejected for now - but I anticipate this patch will
> >> stimulate discussion around what the best approach is) alternative
> >> approaches:
> >>
> >>   - Expose a global user-controlled knob to set the preferred folio
> >>     size; this would move policy to user space and allow (e.g.) setting
> >>     it to PMD-size for even better iTLB utilizaiton. But this would add
> >>     ABI, and I prefer to start with the simplest approach first. It also
> >>     has the downside that a change wouldn't apply to memory already in
> >>     the page cache that is in active use (e.g. libc) so we don't get the
> >>     same level of utilization as for something that is fixed from boot.
> >>
> >>   - Add a per-vma attribute to allow user space to specify preferred
> >>     folio size for memory faulted from the range. (we've talked about
> >>     such a control in the context of mTHP). The dynamic loader would
> >>     then be responsible for adding the annotations. Again this feels
> >>     like something that could be added later if value was demonstrated.
> >>
> >>   - Enhance MADV_COLLAPSE to collapse to THP sizes less than PMD-size.
> >>     This would still require dynamic linker involvement, but would
> >>     additionally neccessitate a copy and all memory in the range would
> >>     be synchronously faulted in, adding to application load time. It
> >>     would work for filesystems that don't support large folios though.
> >>
> >> Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx>
> >> ---
> >>
> >> Hi all,
> >>
> >> I originally concocted something similar to this, with Matthew's help, as a
> >> quick proof of concept hack. Since then I've tried a few different approaches
> >> but always came back to this as the simplest solution. I expect this will raise
> >> a few eyebrows but given it is providing a real performance win, I hope we can
> >> converge to something that can be upstreamed.
> >>
> >> This depends on my contpte series to actually set the contiguous bit in the page
> >> table.
> >>
> >> Thanks,
> >> Ryan
> >>
> >>
> >>  arch/arm64/include/asm/pgtable.h | 12 ++++++++++++
> >>  include/linux/pgtable.h          | 12 ++++++++++++
> >>  mm/filemap.c                     | 19 +++++++++++++++++++
> >>  3 files changed, 43 insertions(+)
> >>
> >> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> >> index f5bf059291c3..8f8f3f7eb8d8 100644
> >> --- a/arch/arm64/include/asm/pgtable.h
> >> +++ b/arch/arm64/include/asm/pgtable.h
> >> @@ -1143,6 +1143,18 @@ static inline void update_mmu_cache_range(struct vm_fault *vmf,
> >>   */
> >>  #define arch_wants_old_prefaulted_pte  cpu_has_hw_af
> >>
> >> +/*
> >> + * Request exec memory is read into pagecache in at least 64K folios. The
> >> + * trade-off here is performance improvement due to storing translations more
> >> + * effciently in the iTLB vs the potential for read amplification due to reading
> >> + * data from disk that won't be used. The latter is independent of base page
> >> + * size, so we set a page-size independent block size of 64K. This size can be
> >> + * contpte-mapped when 4K base pages are in use (16 pages into 1 iTLB entry),
> >> + * and HPA can coalesce it (4 pages into 1 TLB entry) when 16K base pages are in
> >> + * use.
> >> + */
> >> +#define arch_wants_exec_folio_order(void) ilog2(SZ_64K >> PAGE_SHIFT)
> >> +
> >>  static inline bool pud_sect_supported(void)
> >>  {
> >>         return PAGE_SIZE == SZ_4K;
> >> diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
> >> index 170925379534..57090616d09c 100644
> >> --- a/include/linux/pgtable.h
> >> +++ b/include/linux/pgtable.h
> >> @@ -428,6 +428,18 @@ static inline bool arch_has_hw_pte_young(void)
> >>  }
> >>  #endif
> >>
> >> +#ifndef arch_wants_exec_folio_order
> >> +/*
> >> + * Returns preferred minimum folio order for executable file-backed memory. Must
> >> + * be in range [0, PMD_ORDER]. Negative value implies that the HW has no
> >> + * preference and mm will not special-case executable memory in the pagecache.
> >> + */
> >> +static inline int arch_wants_exec_folio_order(void)
> >> +{
> >> +       return -1;
> >> +}
> >> +#endif
> >> +
> >>  #ifndef arch_check_zapped_pte
> >>  static inline void arch_check_zapped_pte(struct vm_area_struct *vma,
> >>                                          pte_t pte)
> >> diff --git a/mm/filemap.c b/mm/filemap.c
> >> index 67ba56ecdd32..80a76d755534 100644
> >> --- a/mm/filemap.c
> >> +++ b/mm/filemap.c
> >> @@ -3115,6 +3115,25 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
> >>         }
> >>  #endif
> >>
> >> +       /*
> >> +        * Allow arch to request a preferred minimum folio order for executable
> >> +        * memory. This can often be beneficial to performance if (e.g.) arm64
> >> +        * can contpte-map the folio. Executable memory rarely benefits from
> >> +        * read-ahead anyway, due to its random access nature.
> >> +        */
> >> +       if (vm_flags & VM_EXEC) {
> >> +               int order = arch_wants_exec_folio_order();
> >> +
> >> +               if (order >= 0) {
> >> +                       fpin = maybe_unlock_mmap_for_io(vmf, fpin);
> >> +                       ra->size = 1UL << order;
> >> +                       ra->async_size = 0;
> >> +                       ractl._index &= ~((unsigned long)ra->size - 1);
> >> +                       page_cache_ra_order(&ractl, ra, order);
> >> +                       return fpin;
> >> +               }
> >> +       }
> >
> > I don't know, but most filesystems don't support large mapping,even iomap.
>
> True, but more are coming. For example ext4 is in the works:
> https://lore.kernel.org/all/20240102123918.799062-1-yi.zhang@xxxxxxxxxxxxxxx/

right, hopefully more filesystems will join.

>
> > This patch might negatively affect them. i feel we need to check
> > mapping_large_folio_support() at least.
>
> page_cache_ra_order() does this check and falls back to small folios if needed.
> So correctness-wise it all works out. I guess your concern is performance due to
> effectively removing the async readahead aspect? But if that is a problem, then
> it's not just a problem if we are reading small folios, so I don't think the
> proposed check is correct.

My point is that this patch is actually changing two things.
1. readahead index/size and async_size=0
2. try to use CONT-PTE

for filesystems which support large mapping, we are getting 2 to help
improve performance; for filesystems without large_mapping, 1 has
been changed from losing read-around,
        /*
         * mmap read-around
         */
        fpin = maybe_unlock_mmap_for_io(vmf, fpin);
        ra->start = max_t(long, 0, vmf->pgoff - ra->ra_pages / 2);
        ra->size = ra->ra_pages;
        ra->async_size = ra->ra_pages / 4;
        ractl._index = ra->start;
        page_cache_ra_order(&ractl, ra, 0);

We probably need data to prove this makes no regression. otherwise,
it is safer to let the code have no side effects on other file systems if
we haven't data.

>
> Perhaps an alternative would be to double ra->size and set ra->async_size to
> (ra->size / 2)? That would ensure we always have 64K aligned blocks but would
> give us an async portion so readahead can still happen.

this might be worth to try as PMD is exactly doing this because async
can decrease
the latency of subsequent page faults.

#ifdef CONFIG_TRANSPARENT_HUGEPAGE
        /* Use the readahead code, even if readahead is disabled */
        if (vm_flags & VM_HUGEPAGE) {
                fpin = maybe_unlock_mmap_for_io(vmf, fpin);
                ractl._index &= ~((unsigned long)HPAGE_PMD_NR - 1);
                ra->size = HPAGE_PMD_NR;
                /*
                 * Fetch two PMD folios, so we get the chance to actually
                 * readahead, unless we've been told not to.
                 */
                if (!(vm_flags & VM_RAND_READ))
                        ra->size *= 2;
                ra->async_size = HPAGE_PMD_NR;
                page_cache_ra_order(&ractl, ra, HPAGE_PMD_ORDER);
                return fpin;
        }
#endif

>
> I don't feel very expert with this area of the code so I might be talking
> rubbish - would be great to hear from others.
>
> >
> >> +
> >>         /* If we don't want any read-ahead, don't bother */
> >>         if (vm_flags & VM_RAND_READ)
> >>                 return fpin;
> >> --
> >> 2.25.1
> >

BTW, is it also possible that user space also wants to map some data
as cont-pte hugepage? just like we have a strong VM_HUGEPAGE
flag for PMD THP?

Thanks
Barry





[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [NTFS 3]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [NTFS 3]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux