On Fri, Dec 19, 2025 at 05:24:36AM +0800, Barry Song wrote:
> On Thu, Dec 18, 2025 at 9:55 PM Uladzislau Rezki <[email protected]> wrote:
> >
> > On Thu, Dec 18, 2025 at 02:01:56PM +0100, David Hildenbrand (Red Hat) wrote:
> > > On 12/15/25 06:30, Barry Song wrote:
> > > > From: Barry Song <[email protected]>
> > > >
> > > > In many cases, the pages passed to vmap() may include high-order
> > > > pages allocated with __GFP_COMP flags. For example, the systemheap
> > > > often allocates pages in descending order: order 8, then 4, then 0.
> > > > Currently, vmap() iterates over every page individually—even pages
> > > > inside a high-order block are handled one by one.
> > > >
> > > > This patch detects high-order pages and maps them as a single
> > > > contiguous block whenever possible.
> > > >
> > > > An alternative would be to implement a new API, vmap_sg(), but that
> > > > change seems to be large in scope.
> > > >
> > > > When vmapping a 128MB dma-buf using the systemheap, this patch
> > > > makes system_heap_do_vmap() roughly 17× faster.
> > > >
> > > > W/ patch:
> > > > [ 10.404769] system_heap_do_vmap took 2494000 ns
> > > > [ 12.525921] system_heap_do_vmap took 2467008 ns
> > > > [ 14.517348] system_heap_do_vmap took 2471008 ns
> > > > [ 16.593406] system_heap_do_vmap took 2444000 ns
> > > > [ 19.501341] system_heap_do_vmap took 2489008 ns
> > > >
> > > > W/o patch:
> > > > [ 7.413756] system_heap_do_vmap took 42626000 ns
> > > > [ 9.425610] system_heap_do_vmap took 42500992 ns
> > > > [ 11.810898] system_heap_do_vmap took 42215008 ns
> > > > [ 14.336790] system_heap_do_vmap took 42134992 ns
> > > > [ 16.373890] system_heap_do_vmap took 42750000 ns
> > > >
> > >
> > > That's quite a speedup.
> > >
> > > > Cc: David Hildenbrand <[email protected]>
> > > > Cc: Uladzislau Rezki <[email protected]>
> > > > Cc: Sumit Semwal <[email protected]>
> > > > Cc: John Stultz <[email protected]>
> > > > Cc: Maxime Ripard <[email protected]>
> > > > Tested-by: Tangquan Zheng <[email protected]>
> > > > Signed-off-by: Barry Song <[email protected]>
> > > > ---
> > > > * diff with rfc:
> > > > Many code refinements based on David's suggestions, thanks!
> > > > Refine comment and changelog according to Uladzislau, thanks!
> > > > rfc link:
> > > >
> > > > https://lore.kernel.org/linux-mm/[email protected]/
> > > >
> > > > mm/vmalloc.c | 45 +++++++++++++++++++++++++++++++++++++++------
> > > > 1 file changed, 39 insertions(+), 6 deletions(-)
> > > >
> > > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > > > index 41dd01e8430c..8d577767a9e5 100644
> > > > --- a/mm/vmalloc.c
> > > > +++ b/mm/vmalloc.c
> > > > @@ -642,6 +642,29 @@ static int vmap_small_pages_range_noflush(unsigned
> > > > long addr, unsigned long end,
> > > > return err;
> > > > }
> > > > +static inline int get_vmap_batch_order(struct page **pages,
> > > > + unsigned int stride, unsigned int max_steps, unsigned int
> > > > idx)
> > > > +{
> > > > + int nr_pages = 1;
> > >
> > > unsigned int, maybe
>
> Right
>
> > >
> > > Why are you initializing nr_pages when you overwrite it below?
>
> Right, initializing nr_pages can be dropped.
>
> > >
> > > > +
> > > > + /*
> > > > + * Currently, batching is only supported in vmap_pages_range
> > > > + * when page_shift == PAGE_SHIFT.
> > >
> > > I don't know the code so realizing how we go from page_shift to stride too
> > > me a second. Maybe only talk about stride here?
> > >
> > > OTOH, is "stride" really the right terminology?
> > >
> > > we calculate it as
> > >
> > > stride = 1U << (page_shift - PAGE_SHIFT);
> > >
> > > page_shift - PAGE_SHIFT should give us an "order". So is this a
> > > "granularity" in nr_pages?
>
> This is the case where vmalloc() may realize that it has
> high-order pages and therefore calls
> vmap_pages_range_noflush() with a page_shift larger than
> PAGE_SHIFT. For vmap(), we take a pages array, so
> page_shift is always PAGE_SHIFT.
>
> > >
> > > Again, I don't know this code, so sorry for the question.
> > >
> > To me "stride" also sounds unclear.
>
> Thanks, David and Uladzislau. On second thought, this stride may be
> redundant, and it should be possible to drop it entirely. This results
> in the code below:
>
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 41dd01e8430c..3962bdcb43e5 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -642,6 +642,20 @@ static int vmap_small_pages_range_noflush(unsigned long
> addr, unsigned long end,
> return err;
> }
>
> +static inline int get_vmap_batch_order(struct page **pages,
> + unsigned int max_steps, unsigned int idx)
> +{
> + unsigned int nr_pages = compound_nr(pages[idx]);
> +
> + if (nr_pages == 1 || max_steps < nr_pages)
> + return 0;
> +
> + if (num_pages_contiguous(&pages[idx], nr_pages) == nr_pages)
> + return compound_order(pages[idx]);
> + return 0;
> +}
> +
>
> /*
> * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
> * flush caches.
> @@ -658,20 +672,35 @@ int __vmap_pages_range_noflush(unsigned long addr,
> unsigned long end,
>
> WARN_ON(page_shift < PAGE_SHIFT);
>
> + /*
> + * For vmap(), users may allocate pages from high orders down to
> + * order 0, while always using PAGE_SHIFT as the page_shift.
> + * We first check whether the initial page is a compound page. If so,
> + * there may be an opportunity to batch multiple pages together.
> + */
> if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
> - page_shift == PAGE_SHIFT)
> + (page_shift == PAGE_SHIFT && !PageCompound(pages[0])))
> return vmap_small_pages_range_noflush(addr, end, prot, pages);
Hm.. If first few pages are order-0 and the rest are compound
then we do nothing.
>
> - for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
> + for (i = 0; i < nr; ) {
> + unsigned int shift = page_shift;
> int err;
>
> - err = vmap_range_noflush(addr, addr + (1UL << page_shift),
> + /*
> + * For vmap() cases, page_shift is always PAGE_SHIFT, even
> + * if the pages are physically contiguous, they may still
> + * be mapped in a batch.
> + */
> + if (page_shift == PAGE_SHIFT)
> + shift += get_vmap_batch_order(pages, nr - i, i);
> + err = vmap_range_noflush(addr, addr + (1UL << shift),
> page_to_phys(pages[i]), prot,
> - page_shift);
> + shift);
> if (err)
> return err;
>
> - addr += 1UL << page_shift;
> + addr += 1UL << shift;
> + i += 1U << shift;
> }
>
> return 0;
>
> Does this look clearer?
>
The concern is we mix it with a huge page mapping path. If we want to batch
v-mapping for page_shift == PAGE_SHIFT case, where "pages" array may contain
compound pages(folio)(corner case to me), i think we should split it.
--
Uladzislau Rezki
