Re: [PATCH v4 2/3] vmalloc: Optimize vfree

[David Hildenbrand <david@kernel.org>]

On 3/30/26 18:15, Muhammad Usama Anjum wrote:
> On 30/03/2026 3:38 pm, David Hildenbrand (Arm) wrote:
>> On 3/27/26 13:57, Muhammad Usama Anjum wrote:
>>> From: Ryan Roberts <ryan.roberts@arm.com>
>>>
>>> Whenever vmalloc allocates high order pages (e.g. for a huge mapping) it
>>> must immediately split_page() to order-0 so that it remains compatible
>>> with users that want to access the underlying struct page.
>>> Commit a06157804399 ("mm/vmalloc: request large order pages from buddy
>>> allocator") recently made it much more likely for vmalloc to allocate
>>> high order pages which are subsequently split to order-0.
>>>
>>> Unfortunately this had the side effect of causing performance
>>> regressions for tight vmalloc/vfree loops (e.g. test_vmalloc.ko
>>> benchmarks). See Closes: tag. This happens because the high order pages
>>> must be gotten from the buddy but then because they are split to
>>> order-0, when they are freed they are freed to the order-0 pcp.
>>> Previously allocation was for order-0 pages so they were recycled from
>>> the pcp.
>>>
>>> It would be preferable if when vmalloc allocates an (e.g.) order-3 page
>>> that it also frees that order-3 page to the order-3 pcp, then the
>>> regression could be removed.
>>>
>>> So let's do exactly that; update stats separately first as coalescing is
>>> hard to do correctly without complexity. Use free_pages_bulk() which uses
>>> the new __free_contig_range() API to batch-free contiguous ranges of pfns.
>>> This not only removes the regression, but significantly improves
>>> performance of vfree beyond the baseline.
>>>
>>> A selection of test_vmalloc benchmarks running on arm64 server class
>>> system. mm-new is the baseline. Commit a06157804399 ("mm/vmalloc: request
>>> large order pages from buddy allocator") was added in v6.19-rc1 where we
>>> see regressions. Then with this change performance is much better. (>0
>>> is faster, <0 is slower, (R)/(I) = statistically significant
>>> Regression/Improvement):
>>>
>>> +-----------------+----------------------------------------------------------+-------------------+--------------------+
>>> | Benchmark       | Result Class                                             |   mm-new          |  this series       |
>>> +=================+==========================================================+===================+====================+
>>> | micromm/vmalloc | fix_align_alloc_test: p:1, h:0, l:500000 (usec)          |        1331843.33 |         (I) 67.17% |
>>> |                 | fix_size_alloc_test: p:1, h:0, l:500000 (usec)           |         415907.33 |             -5.14% |
>>> |                 | fix_size_alloc_test: p:4, h:0, l:500000 (usec)           |         755448.00 |         (I) 53.55% |
>>> |                 | fix_size_alloc_test: p:16, h:0, l:500000 (usec)          |        1591331.33 |         (I) 57.26% |
>>> |                 | fix_size_alloc_test: p:16, h:1, l:500000 (usec)          |        1594345.67 |         (I) 68.46% |
>>> |                 | fix_size_alloc_test: p:64, h:0, l:100000 (usec)          |        1071826.00 |         (I) 79.27% |
>>> |                 | fix_size_alloc_test: p:64, h:1, l:100000 (usec)          |        1018385.00 |         (I) 84.17% |
>>> |                 | fix_size_alloc_test: p:256, h:0, l:100000 (usec)         |        3970899.67 |         (I) 77.01% |
>>> |                 | fix_size_alloc_test: p:256, h:1, l:100000 (usec)         |        3821788.67 |         (I) 89.44% |
>>> |                 | fix_size_alloc_test: p:512, h:0, l:100000 (usec)         |        7795968.00 |         (I) 82.67% |
>>> |                 | fix_size_alloc_test: p:512, h:1, l:100000 (usec)         |        6530169.67 |        (I) 118.09% |
>>> |                 | full_fit_alloc_test: p:1, h:0, l:500000 (usec)           |         626808.33 |             -0.98% |
>>> |                 | kvfree_rcu_1_arg_vmalloc_test: p:1, h:0, l:500000 (usec) |         532145.67 |             -1.68% |
>>> |                 | kvfree_rcu_2_arg_vmalloc_test: p:1, h:0, l:500000 (usec) |         537032.67 |             -0.96% |
>>> |                 | long_busy_list_alloc_test: p:1, h:0, l:500000 (usec)     |        8805069.00 |         (I) 74.58% |
>>> |                 | pcpu_alloc_test: p:1, h:0, l:500000 (usec)               |         500824.67 |              4.35% |
>>> |                 | random_size_align_alloc_test: p:1, h:0, l:500000 (usec)  |        1637554.67 |         (I) 76.99% |
>>> |                 | random_size_alloc_test: p:1, h:0, l:500000 (usec)        |        4556288.67 |         (I) 72.23% |
>>> |                 | vm_map_ram_test: p:1, h:0, l:500000 (usec)               |         107371.00 |             -0.70% |
>>> +-----------------+----------------------------------------------------------+-------------------+--------------------+
>>>
>>> Fixes: a06157804399 ("mm/vmalloc: request large order pages from buddy allocator")
>>> Closes: https://lore.kernel.org/all/66919a28-bc81-49c9-b68f-dd7c73395a0d@arm.com/
>>> Acked-by: Zi Yan <ziy@nvidia.com>
>>> Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
>>> Co-developed-by: Muhammad Usama Anjum <usama.anjum@arm.com>
>>> Signed-off-by: Muhammad Usama Anjum <usama.anjum@arm.com>
>>> ---
>>> Changes since v3:
>>> - Add kerneldoc comment and update description
>>> - Add tag
>>>
>>> Changes since v2:
>>> - Remove BUG_ON in favour of simple implementation as this has never
>>>   been seen to output any bug in the past as well
>>> - Move the free loop to separate function, free_pages_bulk()
>>> - Update stats, lruvec_stat in separate loop
>>>
>>> Changes since v1:
>>> - Rebase on mm-new
>>> - Rerun benchmarks
>>> ---
>>>  include/linux/gfp.h |  2 ++
>>>  mm/page_alloc.c     | 38 ++++++++++++++++++++++++++++++++++++++
>>>  mm/vmalloc.c        | 16 +++++-----------
>>>  3 files changed, 45 insertions(+), 11 deletions(-)
>>>
>>> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
>>> index 7c1f9da7c8e56..71f9097ab99a0 100644
>>> --- a/include/linux/gfp.h
>>> +++ b/include/linux/gfp.h
>>> @@ -239,6 +239,8 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
>>>  				struct page **page_array);
>>>  #define __alloc_pages_bulk(...)			alloc_hooks(alloc_pages_bulk_noprof(__VA_ARGS__))
>>>  
>>> +void free_pages_bulk(struct page **page_array, unsigned long nr_pages);
>>> +
>>>  unsigned long alloc_pages_bulk_mempolicy_noprof(gfp_t gfp,
>>>  				unsigned long nr_pages,
>>>  				struct page **page_array);
>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>>> index 18a96b51aa0be..64be8a9019dca 100644
>>> --- a/mm/page_alloc.c
>>> +++ b/mm/page_alloc.c
>>> @@ -5175,6 +5175,44 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
>>>  }
>>>  EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
>>>  
>>> +/*
>>> + * free_pages_bulk - Free an array of order-0 pages
>>> + * @page_array: Array of pages to free
>>> + * @nr_pages: The number of pages in the array
>>> + *
>>> + * Free the order-0 pages. Adjacent entries whose PFNs form a contiguous
>>> + * run are released with a single __free_contig_range() call.
>>> + *
>>> + * This assumes page_array is sorted in ascending PFN order. Without that,
>>> + * the function still frees all pages, but contiguous runs may not be
>>> + * detected and the freeing pattern can degrade to freeing one page at a
>>> + * time.
>>> + *
>>> + * Context: Sleepable process context only; calls cond_resched()
>>> + */
>>> +void free_pages_bulk(struct page **page_array, unsigned long nr_pages)
>>> +{
>>> +	unsigned long start_pfn = 0, pfn;
>>> +	unsigned long i, nr_contig = 0;
>>> +
>>> +	for (i = 0; i < nr_pages; i++) {
>>> +		pfn = page_to_pfn(page_array[i]);
>>> +		if (!nr_contig) {
>>> +			start_pfn = pfn;
>>> +			nr_contig = 1;
>>> +		} else if (start_pfn + nr_contig != pfn) {
>>> +			__free_contig_range(start_pfn, nr_contig);
>>> +			start_pfn = pfn;
>>> +			nr_contig = 1;
>>> +			cond_resched();
>>> +		} else {
>>> +			nr_contig++;
>>> +		}
>>> +	}
>>
>> What happened to the idea of using num_pages_contiguous()? I think that
>> should generate more efficient code (all we're doing is comparing
>> pointers really on SPARSEMEM_VMEMMAP) and the end result looks more
>> readable?
> Sorry, I misunderstood you in the light of duplicate usage of
> num_pages_contiguous(). I'll update the implementation.


A note that it should even speedup SPARSEMEM && !SPARSEMEM_VMEMMAP,
where a page_to_pfn() is expensive and best minimized/avoided :)

-- 
Cheers,

David