Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:2 as permitted sender) client-ip=2620:137:e000::3:2;
Message-ID: <ce48e5e3-8501-45ea-bc6a-290c7d93c6be@arm.com>
Date:   Thu, 7 Dec 2023 10:37:31 +0000
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [PATCH v8 04/10] mm: thp: Support allocation of anonymous
 multi-size THP
Content-Language: en-GB
From:   Ryan Roberts <ryan.roberts@arm.com>
To:     David Hildenbrand <david@redhat.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Matthew Wilcox <willy@infradead.org>,
        Yin Fengwei <fengwei.yin@intel.com>,
        Yu Zhao <yuzhao@google.com>,
        Catalin Marinas <catalin.marinas@arm.com>,
        Anshuman Khandual <anshuman.khandual@arm.com>,
        Yang Shi <shy828301@gmail.com>,
        "Huang, Ying" <ying.huang@intel.com>, Zi Yan <ziy@nvidia.com>,
        Luis Chamberlain <mcgrof@kernel.org>,
        Itaru Kitayama <itaru.kitayama@gmail.com>,
        "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>,
        John Hubbard <jhubbard@nvidia.com>,
        David Rientjes <rientjes@google.com>,
        Vlastimil Babka <vbabka@suse.cz>,
        Hugh Dickins <hughd@google.com>,
        Kefeng Wang <wangkefeng.wang@huawei.com>,
        Barry Song <21cnbao@gmail.com>,
        Alistair Popple <apopple@nvidia.com>
Cc:     linux-mm@kvack.org, linux-arm-kernel@lists.infradead.org,
        linux-kernel@vger.kernel.org
References: <20231204102027.57185-1-ryan.roberts@arm.com>
 <20231204102027.57185-5-ryan.roberts@arm.com>
 <71040a8c-4ea1-4f21-8ac8-65f7c25c217e@redhat.com>
 <f2896d7f-183b-48fb-a3aa-d21bf2257043@arm.com>
 <5c06e3fc-e8f0-4896-ab71-2317a71e812a@arm.com>
In-Reply-To: <5c06e3fc-e8f0-4896-ab71-2317a71e812a@arm.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk

On 06/12/2023 15:44, Ryan Roberts wrote:
> On 06/12/2023 14:19, Ryan Roberts wrote:
>> On 05/12/2023 16:32, David Hildenbrand wrote:
>>> On 04.12.23 11:20, Ryan Roberts wrote:
>>>> Introduce the logic to allow THP to be configured (through the new sysfs
>>>> interface we just added) to allocate large folios to back anonymous
>>>> memory, which are larger than the base page size but smaller than
>>>> PMD-size. We call this new THP extension "multi-size THP" (mTHP).
>>>>
>>>> mTHP continues to be PTE-mapped, but in many cases can still provide
>>>> similar benefits to traditional PMD-sized THP: Page faults are
>>>> significantly reduced (by a factor of e.g. 4, 8, 16, etc. depending on
>>>> the configured order), but latency spikes are much less prominent
>>>> because the size of each page isn't as huge as the PMD-sized variant and
>>>> there is less memory to clear in each page fault. 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.
>>>>
>>>> Some architectures also employ TLB compression mechanisms to squeeze
>>>> more entries in when a set of PTEs are virtually and physically
>>>> contiguous and approporiately aligned. In this case, TLB misses will
>>>> occur less often.
>>>>
>>>> The new behaviour is disabled by default, but can be enabled at runtime
>>>> by writing to /sys/kernel/mm/transparent_hugepage/hugepage-XXkb/enabled
>>>> (see documentation in previous commit). The long term aim is to change
>>>> the default to include suitable lower orders, but there are some risks
>>>> around internal fragmentation that need to be better understood first.
>>>>
>>>> Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
>>>
>>> In general, looks good to me, some comments/nits. And the usual "let's make sure
>>> we don't degrade order-0 and keep that as fast as possible" comment.
>>>
>>>> ---
>>>>   include/linux/huge_mm.h |   6 ++-
>>>>   mm/memory.c             | 106 ++++++++++++++++++++++++++++++++++++----
>>>>   2 files changed, 101 insertions(+), 11 deletions(-)
>>>>
>>>> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
>>>> index bd0eadd3befb..91a53b9835a4 100644
>>>> --- a/include/linux/huge_mm.h
>>>> +++ b/include/linux/huge_mm.h
>>>> @@ -68,9 +68,11 @@ extern struct kobj_attribute shmem_enabled_attr;
>>>>   #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
>>>>     /*
>>>> - * Mask of all large folio orders supported for anonymous THP.
>>>> + * Mask of all large folio orders supported for anonymous THP; all orders up to
>>>> + * and including PMD_ORDER, except order-0 (which is not "huge") and order-1
>>>> + * (which is a limitation of the THP implementation).
>>>>    */
>>>> -#define THP_ORDERS_ALL_ANON    BIT(PMD_ORDER)
>>>> +#define THP_ORDERS_ALL_ANON    ((BIT(PMD_ORDER + 1) - 1) & ~(BIT(0) | BIT(1)))
>>>>     /*
>>>>    * Mask of all large folio orders supported for file THP.
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 3ceeb0f45bf5..bf7e93813018 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>> @@ -4125,6 +4125,84 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
>>>>       return ret;
>>>>   }
>>>>   +static bool pte_range_none(pte_t *pte, int nr_pages)
>>>> +{
>>>> +    int i;
>>>> +
>>>> +    for (i = 0; i < nr_pages; i++) {
>>>> +        if (!pte_none(ptep_get_lockless(pte + i)))
>>>> +            return false;
>>>> +    }
>>>> +
>>>> +    return true;
>>>> +}
>>>> +
>>>> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
>>>> +static struct folio *alloc_anon_folio(struct vm_fault *vmf)
>>>> +{
>>>> +    gfp_t gfp;
>>>> +    pte_t *pte;
>>>> +    unsigned long addr;
>>>> +    struct folio *folio;
>>>> +    struct vm_area_struct *vma = vmf->vma;
>>>> +    unsigned long orders;
>>>> +    int order;
>>>
>>> Nit: reverse christmas tree encouraged ;)
>>
>> ACK will fix.
>>
>>>
>>>> +
>>>> +    /*
>>>> +     * If uffd is active for the vma we need per-page fault fidelity to
>>>> +     * maintain the uffd semantics.
>>>> +     */
>>>> +    if (userfaultfd_armed(vma))
>>>
>>> Nit: unlikely()
>>
>> ACK will fix.
>>
>>>
>>>> +        goto fallback;
>>>> +
>>>> +    /*
>>>> +     * Get a list of all the (large) orders below PMD_ORDER that are enabled
>>>> +     * for this vma. Then filter out the orders that can't be allocated over
>>>> +     * the faulting address and still be fully contained in the vma.
>>>> +     */
>>>> +    orders = thp_vma_allowable_orders(vma, vma->vm_flags, false, true, true,
>>>> +                      BIT(PMD_ORDER) - 1);
>>>> +    orders = thp_vma_suitable_orders(vma, vmf->address, orders);
>>>
>>> Comment: Both will eventually loop over all orders, correct? Could eventually be
>>> sped up in the future.
>>
>> No only thp_vma_suitable_orders() will loop. thp_vma_allowable_orders() only
>> loops if in_pf=false (it's true here).
>>
>>>
>>> Nit: the orders = ... order = ... looks like this might deserve a helper
>>> function that makes this easier to read.
>>
>> To be honest, the existing function that I've modified is a bit of a mess.
>> thp_vma_allowable_orders() calls thp_vma_suitable_orders() if we are not in a
>> page fault, because the page fault handlers already do that check themselves. It
>> would be nice to refactor the whole thing so that thp_vma_allowable_orders() is
>> a strict superset of thp_vma_suitable_orders(). Then this can just call
>> thp_vma_allowable_orders(). But that's going to start touching the PMD and PUD
>> handlers, so prefer if we leave that for a separate patch set.
>>
>>>
>>> Nit: Why call thp_vma_suitable_orders if the orders are already 0? Again, some
>>> helper might be reasonable where that is handled internally.
>>
>> Because thp_vma_suitable_orders() will handle it safely and is inline, so it
>> should just as efficient? This would go away with the refactoring described above.
>>
>>>
>>> Comment: For order-0 we'll always perform a function call to both
>>> thp_vma_allowable_orders() / thp_vma_suitable_orders(). We should perform some
>>> fast and efficient check if any <PMD THP are even enabled in the system / for
>>> this VMA, and in that case just fallback before doing more expensive checks.
>>
> 
> I just noticed I got these functions round the wrong way in my previous response:
> 
>> thp_vma_allowable_orders() is inline as you mentioned.
> 
> ^ Meant thp_vma_suitable_orders() here.
> 
>>
>> I was deliberately trying to keep all the decision logic in one place
>> (thp_vma_suitable_orders) because it's already pretty complicated. But if you
> 
> ^ Meant thp_vma_allowable_orders() here.
> 
> Sorry for the confusion.
> 
>> insist, how about this in the header:
>>
>> static inline
>> unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
>> 				       unsigned long vm_flags, bool smaps,
>> 				       bool in_pf, bool enforce_sysfs,
>> 				       unsigned long orders)
>> {
>> 	/* Optimization to check if required orders are enabled early. */
>> 	if (enforce_sysfs && vma_is_anonymous(vma)) {
>> 		unsigned long mask = READ_ONCE(huge_anon_orders_always);
>>
>> 		if (vm_flags & VM_HUGEPAGE)
>> 			mask |= READ_ONCE(huge_anon_orders_madvise);
>> 		if (hugepage_global_always() ||
>> 			((vm_flags & VM_HUGEPAGE) && hugepage_global_enabled()))
>> 			mask |= READ_ONCE(huge_anon_orders_inherit);
>>
>> 		orders &= mask;
>> 		if (!orders)
>> 			return 0;
>> 		
>> 		enforce_sysfs = false;
>> 	}
>>
>> 	return __thp_vma_allowable_orders(vma, vm_flags, smaps, in_pf,
>> 					  enforce_sysfs, orders);
>> }
>>
>> Then the above check can be removed from __thp_vma_allowable_orders() - it will
>> still retain the `if (enforce_sysfs && !vma_is_anonymous(vma))` part.
>>
>>
>>>
>>>> +
>>>> +    if (!orders)
>>>> +        goto fallback;
>>>> +
>>>> +    pte = pte_offset_map(vmf->pmd, vmf->address & PMD_MASK);
>>>> +    if (!pte)
>>>> +        return ERR_PTR(-EAGAIN);
>>>> +
>>>> +    order = first_order(orders);
>>>> +    while (orders) {
>>>> +        addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
>>>> +        vmf->pte = pte + pte_index(addr);
>>>> +        if (pte_range_none(vmf->pte, 1 << order))
>>>> +            break;
>>>
>>> Comment: Likely it would make sense to scan only once and determine the "largest
>>> none range" around that address, having the largest suitable order in mind.
>>
>> Yes, that's how I used to do it, but Yu Zhou requested simplifying to this,
>> IIRC. Perhaps this an optimization opportunity for later?
>>
>>>
>>>> +        order = next_order(&orders, order);
>>>> +    }
>>>> +
>>>> +    vmf->pte = NULL;
>>>
>>> Nit: Can you elaborate why you are messing with vmf->pte here? A simple helper
>>> variable will make this code look less magical. Unless I am missing something
>>> important :)
>>
>> Gahh, I used to pass the vmf to what pte_range_none() was refactored into (an
>> approach that was suggested by Yu Zhou IIRC). But since I did some refactoring
>> based on some comments from JohnH, I see I don't need that anymore. Agreed; it
>> will be much clearer just to use a local variable. Will fix.
>>
>>>
>>>> +    pte_unmap(pte);
>>>> +
>>>> +    gfp = vma_thp_gfp_mask(vma);
>>>> +
>>>> +    while (orders) {
>>>> +        addr = ALIGN_DOWN(vmf->address, PAGE_SIZE << order);
>>>> +        folio = vma_alloc_folio(gfp, order, vma, addr, true);
>>>> +        if (folio) {
>>>> +            clear_huge_page(&folio->page, addr, 1 << order);
>>>> +            return folio;
>>>> +        }
>>>> +        order = next_order(&orders, order);
>>>> +    }
>>>> +
>>>
>>> Queestion: would it make sense to combine both loops? I suspect memory
>>> allocations with pte_offset_map()/kmao are problematic.
>>
>> They are both operating on separate orders; next_order() is "consuming" an order
>> by removing the current one from the orders bitfield and returning the next one.
>>
>> So the first loop starts at the highest order and keeps checking lower orders
>> until one fully fits in the VMA. And the second loop starts at the first order
>> that was found to fully fits and loops to lower orders until an allocation is
>> successful.
>>
>> So I don't see a need to combine the loops.
>>
>>>
>>>> +fallback:
>>>> +    return vma_alloc_zeroed_movable_folio(vma, vmf->address);
>>>> +}
>>>> +#else
>>>> +#define alloc_anon_folio(vmf) \
>>>> +        vma_alloc_zeroed_movable_folio((vmf)->vma, (vmf)->address)
>>>> +#endif
>>>> +
>>>>   /*
>>>>    * We enter with non-exclusive mmap_lock (to exclude vma changes,
>>>>    * but allow concurrent faults), and pte mapped but not yet locked.
>>>> @@ -4132,6 +4210,9 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
>>>>    */
>>>>   static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
>>>>   {
>>>> +    int i;
>>>> +    int nr_pages = 1;
>>>> +    unsigned long addr = vmf->address;
>>>>       bool uffd_wp = vmf_orig_pte_uffd_wp(vmf);
>>>>       struct vm_area_struct *vma = vmf->vma;
>>>>       struct folio *folio;
>>>
>>> Nit: reverse christmas tree :)
>>
>> ACK
>>
>>>
>>>> @@ -4176,10 +4257,15 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
>>>>       /* Allocate our own private page. */
>>>>       if (unlikely(anon_vma_prepare(vma)))
>>>>           goto oom;
>>>> -    folio = vma_alloc_zeroed_movable_folio(vma, vmf->address);
>>>> +    folio = alloc_anon_folio(vmf);
>>>> +    if (IS_ERR(folio))
>>>> +        return 0;
>>>>       if (!folio)
>>>>           goto oom;
>>>>   +    nr_pages = folio_nr_pages(folio);
>>>> +    addr = ALIGN_DOWN(vmf->address, nr_pages * PAGE_SIZE);
>>>> +
>>>>       if (mem_cgroup_charge(folio, vma->vm_mm, GFP_KERNEL))
>>>>           goto oom_free_page;
>>>>       folio_throttle_swaprate(folio, GFP_KERNEL);
>>>> @@ -4196,12 +4282,13 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
>>>>       if (vma->vm_flags & VM_WRITE)
>>>>           entry = pte_mkwrite(pte_mkdirty(entry), vma);
>>>>   -    vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
>>>> -            &vmf->ptl);
>>>> +    vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, addr, &vmf->ptl);
>>>>       if (!vmf->pte)
>>>>           goto release;
>>>> -    if (vmf_pte_changed(vmf)) {
>>>> -        update_mmu_tlb(vma, vmf->address, vmf->pte);
>>>> +    if ((nr_pages == 1 && vmf_pte_changed(vmf)) ||
>>>> +        (nr_pages  > 1 && !pte_range_none(vmf->pte, nr_pages))) {
>>>> +        for (i = 0; i < nr_pages; i++)
>>>> +            update_mmu_tlb(vma, addr + PAGE_SIZE * i, vmf->pte + i);
>>>
>>> Comment: separating the order-0 case from the other case might make this easier
>>> to read.
>>
>> Yeah fair enough. Will fix.
>>
>>>
>>>>           goto release;
>>>>       }
>>>>   @@ -4216,16 +4303,17 @@ static vm_fault_t do_anonymous_page(struct vm_fault
>>>> *vmf)
>>>>           return handle_userfault(vmf, VM_UFFD_MISSING);
>>>>       }
>>>>   -    inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
>>>> -    folio_add_new_anon_rmap(folio, vma, vmf->address);
>>>> +    folio_ref_add(folio, nr_pages - 1);
>>>> +    add_mm_counter(vma->vm_mm, MM_ANONPAGES, nr_pages);
>>>> +    folio_add_new_anon_rmap(folio, vma, addr);
>>>>       folio_add_lru_vma(folio, vma);
>>>>   setpte:
>>>>       if (uffd_wp)
>>>>           entry = pte_mkuffd_wp(entry);
>>>> -    set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
>>>> +    set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr_pages);
>>>>         /* No need to invalidate - it was non-present before */
>>>> -    update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
>>>> +    update_mmu_cache_range(vmf, vma, addr, vmf->pte, nr_pages);
>>>>   unlock:
>>>>       if (vmf->pte)
>>>>           pte_unmap_unlock(vmf->pte, vmf->ptl);
>>>
>>> Benchmarking order-0 allocations might be interesting. There will be some added
>>> checks + multiple loops/conditionals for order-0 that could be avoided by having
>>> two separate code paths. If we can't measure a difference, all good.
>>
>> Yep will do - will post numbers once I have them. I've been assuming that the
>> major cost is clearing the page, but perhaps I'm wrong.
>>

I added a "write-fault-byte" benchmark to the microbenchmark tool you gave me.
This elides the normal memset page population routine, and instead writes the
first byte of every page while the timer is running.

I ran with 100 iterations per run, then ran the whole thing 16 times. I ran it
for a baseline kernel, as well as v8 (this series) and v9 (with changes from
your review). I repeated on Ampere Altra (bare metal) and Apple M2 (VM):

|              |        m2 vm        |        altra        |
|--------------|---------------------|--------------------:|
| kernel       |     mean |  std_rel |     mean |  std_rel |
|--------------|----------|----------|----------|---------:|
| baseline     |   0.000% |   0.341% |   0.000% |   3.581% |
| anonfolio-v8 |   0.005% |   0.272% |   5.068% |   1.128% |
| anonfolio-v9 |  -0.013% |   0.442% |   0.107% |   1.788% |

No measurable difference on M2, but altra has a slow down in v8 which is fixed
in v9; Looking at the changes, this is either down to the new unlikely() for the
uffd or due to moving the THP order check to be inline within
thp_vma_allowable_orders().

So I have all the changes done and perf numbers to show no regression for
order-0. I'm gonna do a final check and post v9 later today.

Thanks,
Ryan