Re: [PATCH v4] mm: use per_vma lock for MADV_DONTNEED

[David Hildenbrand <david@redhat.com>]

On 18.06.25 15:05, Lance Yang wrote:
> 
> 
> On 2025/6/18 18:18, David Hildenbrand wrote:
>> On 18.06.25 11:52, Barry Song wrote:
>>> On Wed, Jun 18, 2025 at 10:25 AM Lance Yang <lance.yang@linux.dev> wrote:
>>>>
>>>> Hi all,
>>>>
>>>> Crazy, the per-VMA lock for madvise is an absolute game-changer ;)
>>>>
>>>> On 2025/6/17 21:38, Lorenzo Stoakes wrote:
>>>> [...]
>>>>>
>>>>> On Sun, Jun 08, 2025 at 10:01:50AM +1200, Barry Song wrote:
>>>>>> From: Barry Song <v-songbaohua@oppo.com>
>>>>>>
>>>>>> Certain madvise operations, especially MADV_DONTNEED, occur far more
>>>>>> frequently than other madvise options, particularly in native and Java
>>>>>> heaps for dynamic memory management.
>>>>>>
>>>>>> Currently, the mmap_lock is always held during these operations,
>>>>>> even when
>>>>>> unnecessary. This causes lock contention and can lead to severe
>>>>>> priority
>>>>>> inversion, where low-priority threads—such as Android's
>>>>>> HeapTaskDaemon—
>>>>>> hold the lock and block higher-priority threads.
>>>>>>
>>>>>> This patch enables the use of per-VMA locks when the advised range
>>>>>> lies
>>>>>> entirely within a single VMA, avoiding the need for full VMA
>>>>>> traversal. In
>>>>>> practice, userspace heaps rarely issue MADV_DONTNEED across
>>>>>> multiple VMAs.
>>>>>>
>>>>>> Tangquan’s testing shows that over 99.5% of memory reclaimed by
>>>>>> Android
>>>>>> benefits from this per-VMA lock optimization. After extended runtime,
>>>>>> 217,735 madvise calls from HeapTaskDaemon used the per-VMA path, while
>>>>>> only 1,231 fell back to mmap_lock.
>>>>>>
>>>>>> To simplify handling, the implementation falls back to the standard
>>>>>> mmap_lock if userfaultfd is enabled on the VMA, avoiding the
>>>>>> complexity of
>>>>>> userfaultfd_remove().
>>>>>>
>>>>>> Many thanks to Lorenzo's work[1] on:
>>>>>> "Refactor the madvise() code to retain state about the locking mode
>>>>>> utilised for traversing VMAs.
>>>>>>
>>>>>> Then use this mechanism to permit VMA locking to be done later in the
>>>>>> madvise() logic and also to allow altering of the locking mode to
>>>>>> permit
>>>>>> falling back to an mmap read lock if required."
>>>>>>
>>>>>> One important point, as pointed out by Jann[2], is that
>>>>>> untagged_addr_remote() requires holding mmap_lock. This is because
>>>>>> address tagging on x86 and RISC-V is quite complex.
>>>>>>
>>>>>> Until untagged_addr_remote() becomes atomic—which seems unlikely in
>>>>>> the near future—we cannot support per-VMA locks for remote processes.
>>>>>> So for now, only local processes are supported.
>>>>
>>>> Just to put some numbers on it, I ran a micro-benchmark with 100
>>>> parallel threads, where each thread calls madvise() on its own 1GiB
> 
> Correction: it uses 256MiB chunks per thread, not 1GiB ...
> 
>>>> chunk of 64KiB mTHP-backed memory. The performance gain is huge:
>>>>
>>>> 1) MADV_DONTNEED saw its average time drop from 0.0508s to 0.0270s (~47%
>>>> faster)
>>>> 2) MADV_FREE     saw its average time drop from 0.3078s to 0.1095s (~64%
>>>> faster)
>>>
>>> Thanks for the report, Lance. I assume your micro-benchmark includes some
>>> explicit or implicit operations that may require mmap_write_lock().
>>> As  mmap_read_lock() only waits for writers and does not block other
>>> mmap_read_lock() calls.
>>
>> The number rather indicate that one test was run with (m)THPs enabled
>> and the other not? Just a thought. The locking overhead from my
>> experience is not that significant.
>>
> 
> Both tests were run with 64KiB mTHP enabled on an Intel(R) Xeon(R)
> Silver 4314 CPU. The micro-benchmark code is following:

Ah, I missed the "100 threads" above. Yeah, there should be plenty of 
locking contention with all the mprotect() in there.

-- 
Cheers,

David / dhildenb