Re: [PATCH] mm/vmalloc: request large order pages from buddy allocator

[Dev Jain <dev.jain@arm.com>]

On 11/12/25 8:58 pm, Ryan Roberts wrote:
> On 10/12/2025 22:28, Vishal Moola (Oracle) wrote:
>> On Wed, Dec 10, 2025 at 01:21:22PM +0000, Ryan Roberts wrote:
>>> Hi Vishal,
>>>
>>>
>>> On 21/10/2025 20:44, Vishal Moola (Oracle) wrote:
>>>> Sometimes, vm_area_alloc_pages() will want many pages from the buddy
>>>> allocator. Rather than making requests to the buddy allocator for at
>>>> most 100 pages at a time, we can eagerly request large order pages a
>>>> smaller number of times.
>>>>
>>>> We still split the large order pages down to order-0 as the rest of the
>>>> vmalloc code (and some callers) depend on it. We still defer to the bulk
>>>> allocator and fallback path in case of order-0 pages or failure.
>>>>
>>>> Running 1000 iterations of allocations on a small 4GB system finds:
>>>>
>>>> 1000 2mb allocations:
>>>> 	[Baseline]			[This patch]
>>>> 	real    46.310s			real    0m34.582
>>>> 	user    0.001s			user    0.006s
>>>> 	sys     46.058s			sys     0m34.365s
>>>>
>>>> 10000 200kb allocations:
>>>> 	[Baseline]			[This patch]
>>>> 	real    56.104s			real    0m43.696
>>>> 	user    0.001s			user    0.003s
>>>> 	sys     55.375s			sys     0m42.995s
>>> I'm seeing some big vmalloc micro benchmark regressions on arm64, for which
>>> bisect is pointing to this patch.
>> Ulad had similar findings/concerns[1]. Tldr: The numbers you are seeing
>> are expected for how the test module is currently written.
> Hmm... simplistically, I'd say that either the tests are bad, in which case they
> should be deleted, or they are good, in which case we shouldn't ignore the
> regressions. Having tests that we learn to ignore is the worst of both worlds.

AFAICR the test does some million-odd iterations by default, which is the real problem.
On my RFC [1] I notice that reducing the iterations reduces the regression - till
some multiple of ten thousand iterations, the regression is zero. Doing this
alloc->free a million freaking times messes up the buddy badly.

[1] https://lore.kernel.org/all/20251112110807.69958-1-dev.jain@arm.com/

>
> But I see your point about the allocation pattern not being very realistic.
>
>>> The tests are all originally from the vmalloc_test module. Note that (R)
>>> indicates a statistically significant regression and (I) indicates a
>>> statistically improvement.
>>>
>>> p is number of pages in the allocation, h is huge. So it looks like the
>>> regressions are all coming for the non-huge case, where we want to split to
>>> order-0.
>>>
>>> +---------------------------------+----------------------------------------------------------+------------+------------------------+
>>> | Benchmark                       | Result Class                                             |     6-18-0 |   6-18-0-gc2f2b01b74be |
>>> +=================================+==========================================================+============+========================+
>>> | micromm/vmalloc                 | fix_align_alloc_test: p:1, h:0, l:500000 (usec)          |  514126.58 |            (R) -42.20% |
>>> |                                 | fix_size_alloc_test: p:1, h:0, l:500000 (usec)           |  320458.33 |                 -0.02% |
>>> |                                 | fix_size_alloc_test: p:4, h:0, l:500000 (usec)           |  399680.33 |            (R) -23.43% |
>>> |                                 | fix_size_alloc_test: p:16, h:0, l:500000 (usec)          |  788723.25 |            (R) -23.66% |
>>> |                                 | fix_size_alloc_test: p:16, h:1, l:500000 (usec)          |  979839.58 |                 -1.05% |
>>> |                                 | fix_size_alloc_test: p:64, h:0, l:100000 (usec)          |  481454.58 |            (R) -23.99% |
>>> |                                 | fix_size_alloc_test: p:64, h:1, l:100000 (usec)          |  615924.00 |              (I) 2.56% |
>>> |                                 | fix_size_alloc_test: p:256, h:0, l:100000 (usec)         | 1799224.08 |            (R) -23.28% |
>>> |                                 | fix_size_alloc_test: p:256, h:1, l:100000 (usec)         | 2313859.25 |              (I) 3.43% |
>>> |                                 | fix_size_alloc_test: p:512, h:0, l:100000 (usec)         | 3541904.75 |            (R) -23.86% |
>>> |                                 | fix_size_alloc_test: p:512, h:1, l:100000 (usec)         | 3597577.25 |             (R) -2.97% |
>>> |                                 | full_fit_alloc_test: p:1, h:0, l:500000 (usec)           |  487021.83 |              (I) 4.95% |
>>> |                                 | kvfree_rcu_1_arg_vmalloc_test: p:1, h:0, l:500000 (usec) |  344466.33 |                 -0.65% |
>>> |                                 | kvfree_rcu_2_arg_vmalloc_test: p:1, h:0, l:500000 (usec) |  342484.25 |                 -1.58% |
>>> |                                 | long_busy_list_alloc_test: p:1, h:0, l:500000 (usec)     | 4034901.17 |            (R) -25.35% |
>>> |                                 | pcpu_alloc_test: p:1, h:0, l:500000 (usec)               |  195973.42 |                  0.57% |
>>> |                                 | random_size_align_alloc_test: p:1, h:0, l:500000 (usec)  |  643489.33 |            (R) -47.63% |
>>> |                                 | random_size_alloc_test: p:1, h:0, l:500000 (usec)        | 2029261.33 |            (R) -27.88% |
>>> |                                 | vm_map_ram_test: p:1, h:0, l:500000 (usec)               |   83557.08 |                 -0.22% |
>>> +---------------------------------+----------------------------------------------------------+------------+------------------------+
>>>
>>> I have a couple of thoughts from looking at the patch:
>>>
>>>   - Perhaps split_page() is the bulk of the cost? Previously for this case we
>>>     were allocating order-0 so there was no split to do. For h=1, split would
>>>     have already been called so that would explain why no regression for that
>>>     case?
>> For h=1, this patch shouldn't change (as long as nr_pages <
>> arch_vmap_{pte,pmd}_supported_shift). This is why you don't see regressions
>> in those cases.
> arm64 supports 64K contigous-mappings with vmalloc so once nr_pages >= 16 we can
> take the huge path.
>
>>>   - I guess we are bypassing the pcpu cache? Could this be having an effect? Dev
>>>     (cc'ed) did some similar investigation a while back and saw increased vmalloc
>>>     latencies when bypassing pcpu cache.
>> I'd say this is more a case of this test module targeting the pcpu
>> cache. The module allocates then frees one at a time, which promotes
>> reusing pcpu pages. [1] Has some numbers after modifying the test such
>> that all the allocations are made before freeing any.
> OK fair enough.
>
> We are seeing a bunch of other regressions in higher level benchmarks too; but
> haven't yet concluded what's causing those. I'll report back if this patch looks
> connected.
>
> Thanks,
> Ryan
>
>
>>>   - Philosophically is allocating physically contiguous memory when it is not
>>>     strictly needed the right thing to do? Large physically contiguous blocks are
>>>     a scarce resource so we don't want to waste them. Although I guess it could
>>>     be argued that this actually preserves the contiguous blocks because the
>>>     lifetime of all the pages is tied together. Anyway, I doubt this is the
>> This was the primary incentive for this patch :)
>>
>>>     reason for the slow down, since those benchmarks are not under memory
>>>     pressure.
>>>
>>> Anyway, it would be good to resolve the performance regressions if we can.
>> Imo, the appropriate way to address these is to modify the test module
>> as seen in [1].
>>
>> [1] https://lore.kernel.org/linux-mm/aPJ6lLf24TfW_1n7@milan/
>