Re: [PATCH] mm: Respect mmap hint address when aligning for THP

[Vlastimil Babka <vbabka@suse.cz>]

On 11/15/24 23:15, Yang Shi wrote:
> On Fri, Nov 15, 2024 at 1:52 PM Kalesh Singh <kaleshsingh@google.com> wrote:
>>
>> Commit efa7df3e3bb5 ("mm: align larger anonymous mappings on THP
>> boundaries") updated __get_unmapped_area() to align the start address
>> for the VMA to a PMD boundary if CONFIG_TRANSPARENT_HUGEPAGE=y.
>>
>> It does this by effectively looking up a region that is of size,
>> request_size + PMD_SIZE, and aligning up the start to a PMD boundary.
>>
>> Commit 4ef9ad19e176 ("mm: huge_memory: don't force huge page alignment
>> on 32 bit") opted out of this for 32bit due to regressions in mmap base
>> randomization.
>>
>> Commit d4148aeab412 ("mm, mmap: limit THP alignment of anonymous
>> mappings to PMD-aligned sizes") restricted this to only mmap sizes that
>> are multiples of the PMD_SIZE due to reported regressions in some
>> performance benchmarks -- which seemed mostly due to the reduced spatial
>> locality of related mappings due to the forced PMD-alignment.
>>
>> Another unintended side effect has emerged: When a user specifies an mmap
>> hint address, the THP alignment logic modifies the behavior, potentially
>> ignoring the hint even if a sufficiently large gap exists at the requested
>> hint location.
>>
>> Example Scenario:
>>
>> Consider the following simplified virtual address (VA) space:
>>
>>     ...
>>
>>     0x200000-0x400000 --- VMA A
>>     0x400000-0x600000 --- Hole
>>     0x600000-0x800000 --- VMA B
>>
>>     ...
>>
>> A call to mmap() with hint=0x400000 and len=0x200000 behaves differently:
>>
>>   - Before THP alignment: The requested region (size 0x200000) fits into
>>     the gap at 0x400000, so the hint is respected.
>>
>>   - After alignment: The logic searches for a region of size
>>     0x400000 (len + PMD_SIZE) starting at 0x400000.
>>     This search fails due to the mapping at 0x600000 (VMA B), and the hint
>>     is ignored, falling back to arch_get_unmapped_area[_topdown]().

Hmm looks like the search is not done in the optimal way regardless of
whether or not it ignores a hint - it should be able to find the hole, no?

>> In general the hint is effectively ignored, if there is any
>> existing mapping in the below range:
>>
>>      [mmap_hint + mmap_size, mmap_hint + mmap_size + PMD_SIZE)
>>
>> This changes the semantics of mmap hint; from ""Respect the hint if a
>> sufficiently large gap exists at the requested location" to "Respect the
>> hint only if an additional PMD-sized gap exists beyond the requested size".
>>
>> This has performance implications for allocators that allocate their heap
>> using mmap but try to keep it "as contiguous as possible" by using the
>> end of the exisiting heap as the address hint. With the new behavior
>> it's more likely to get a much less contiguous heap, adding extra
>> fragmentation and performance overhead.
>>
>> To restore the expected behavior; don't use thp_get_unmapped_area_vmflags()
>> when the user provided a hint address.

Agreed, the hint should take precendence.

> Thanks for fixing it. I agree we should respect the hint address. But
> this patch actually just fixed anonymous mapping and the file mappings
> which don't support thp_get_unmapped_area(). So I think you should
> move the hint check to __thp_get_unmapped_area().
> 
> And Vlastimil's fix d4148aeab412 ("mm, mmap: limit THP alignment of
> anonymous mappings to PMD-aligned sizes") should be moved to there too
> IMHO.

This was brought up, but I didn't want to do it as part of the stable fix as
that would change even situations that Rik's change didn't.
If the mmap hint change is another stable hotfix, I wouldn't conflate it
either. But we can try it for further development. But careful about just
moving the code as-is, the file-based mappings are different than anonymous
memory and I believe file offsets matter:

https://lore.kernel.org/all/9d7c73f6-1e1a-458b-93c6-3b44959022e0@suse.cz/

https://lore.kernel.org/all/5f7a49e8-0416-4648-a704-a7a67e8cd894@suse.cz/

>> Cc: Andrew Morton <akpm@linux-foundation.org>
>> Cc: Vlastimil Babka <vbabka@suse.cz>
>> Cc: Yang Shi <yang@os.amperecomputing.com>
>> Cc: Rik van Riel <riel@surriel.com>
>> Cc: Ryan Roberts <ryan.roberts@arm.com>
>> Cc: Suren Baghdasaryan <surenb@google.com>
>> Cc: Minchan Kim <minchan@kernel.org>
>> Cc: Hans Boehm <hboehm@google.com>
>> Cc: Lokesh Gidra <lokeshgidra@google.com>
>> Cc: <stable@vger.kernel.org>
>> Fixes: efa7df3e3bb5 ("mm: align larger anonymous mappings on THP boundaries")
>> Signed-off-by: Kalesh Singh <kaleshsingh@google.com>

Reviewed-by: Vlastimil Babka <vbabka@suse.cz>

>> ---
>>  mm/mmap.c | 1 +
>>  1 file changed, 1 insertion(+)
>>
>> diff --git a/mm/mmap.c b/mm/mmap.c
>> index 79d541f1502b..2f01f1a8e304 100644
>> --- a/mm/mmap.c
>> +++ b/mm/mmap.c
>> @@ -901,6 +901,7 @@ __get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
>>         if (get_area) {
>>                 addr = get_area(file, addr, len, pgoff, flags);
>>         } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)
>> +                  && !addr /* no hint */
>>                    && IS_ALIGNED(len, PMD_SIZE)) {
>>                 /* Ensures that larger anonymous mappings are THP aligned. */
>>                 addr = thp_get_unmapped_area_vmflags(file, addr, len,
>>
>> base-commit: 2d5404caa8c7bb5c4e0435f94b28834ae5456623
>> --
>> 2.47.0.338.g60cca15819-goog
>>