[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CA+EESO4d_iriWeLit6RbODxxMPLnns64cuo+gkQhbYebZdHDdA@mail.gmail.com>
Date: Mon, 4 Aug 2025 23:30:35 -0700
From: Lokesh Gidra <lokeshgidra@...gle.com>
To: Barry Song <21cnbao@...il.com>
Cc: akpm@...ux-foundation.org, aarcange@...hat.com, linux-mm@...ck.org,
linux-kernel@...r.kernel.org, ngeoffray@...gle.com,
Suren Baghdasaryan <surenb@...gle.com>, Kalesh Singh <kaleshsingh@...gle.com>,
Barry Song <v-songbaohua@...o.com>, David Hildenbrand <david@...hat.com>, Peter Xu <peterx@...hat.com>
Subject: Re: [PATCH] userfaultfd: opportunistic TLB-flush batching for present
pages in MOVE
On Mon, Aug 4, 2025 at 9:35 PM Barry Song <21cnbao@...il.com> wrote:
>
> On Thu, Jul 31, 2025 at 6:47 PM Lokesh Gidra <lokeshgidra@...gle.com> wrote:
> >
> > MOVE ioctl's runtime is dominated by TLB-flush cost, which is required
> > for moving present pages. Mitigate this cost by opportunistically
> > batching present contiguous pages for TLB flushing.
> >
> > Without batching, in our testing on an arm64 Android device with UFFD GC,
> > which uses MOVE ioctl for compaction, we observed that out of the total
> > time spent in move_pages_pte(), over 40% is in ptep_clear_flush(), and
> > ~20% in vm_normal_folio().
> >
> > With batching, the proportion of vm_normal_folio() increases to over
> > 70% of move_pages_pte() without any changes to vm_normal_folio().
> > Furthermore, time spent within move_pages_pte() is only ~20%, which
> > includes TLB-flush overhead.
> >
> > Cc: Suren Baghdasaryan <surenb@...gle.com>
> > Cc: Kalesh Singh <kaleshsingh@...gle.com>
> > Cc: Barry Song <v-songbaohua@...o.com>
> > Cc: David Hildenbrand <david@...hat.com>
> > Cc: Peter Xu <peterx@...hat.com>
> > Signed-off-by: Lokesh Gidra <lokeshgidra@...gle.com>
> > ---
> > mm/userfaultfd.c | 179 +++++++++++++++++++++++++++++++++--------------
> > 1 file changed, 127 insertions(+), 52 deletions(-)
> >
> > diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
> > index 8253978ee0fb..2465fb234671 100644
> > --- a/mm/userfaultfd.c
> > +++ b/mm/userfaultfd.c
> > @@ -1026,18 +1026,62 @@ static inline bool is_pte_pages_stable(pte_t *dst_pte, pte_t *src_pte,
> > pmd_same(dst_pmdval, pmdp_get_lockless(dst_pmd));
> > }
> >
> > -static int move_present_pte(struct mm_struct *mm,
> > - struct vm_area_struct *dst_vma,
> > - struct vm_area_struct *src_vma,
> > - unsigned long dst_addr, unsigned long src_addr,
> > - pte_t *dst_pte, pte_t *src_pte,
> > - pte_t orig_dst_pte, pte_t orig_src_pte,
> > - pmd_t *dst_pmd, pmd_t dst_pmdval,
> > - spinlock_t *dst_ptl, spinlock_t *src_ptl,
> > - struct folio *src_folio)
> > +/*
> > + * Checks if the two ptes and the corresponding folio are eligible for batched
> > + * move. If so, then returns pointer to the folio, after locking it. Otherwise,
> > + * returns NULL.
> > + */
> > +static struct folio *check_ptes_for_batched_move(struct vm_area_struct *src_vma,
> > + unsigned long src_addr,
> > + pte_t *src_pte, pte_t *dst_pte)
> > +{
> > + pte_t orig_dst_pte, orig_src_pte;
> > + struct folio *folio;
> > +
> > + orig_dst_pte = ptep_get(dst_pte);
> > + if (!pte_none(orig_dst_pte))
> > + return NULL;
> > +
> > + orig_src_pte = ptep_get(src_pte);
> > + if (pte_none(orig_src_pte))
> > + return NULL;
> > + if (!pte_present(orig_src_pte) || is_zero_pfn(pte_pfn(orig_src_pte)))
> > + return NULL;
> > +
> > + folio = vm_normal_folio(src_vma, src_addr, orig_src_pte);
> > + if (!folio || !folio_trylock(folio))
> > + return NULL;
> > + if (!PageAnonExclusive(&folio->page) || folio_test_large(folio)) {
> > + folio_unlock(folio);
> > + return NULL;
> > + }
> > + return folio;
> > +}
> > +
> > +static long move_present_ptes(struct mm_struct *mm,
> > + struct vm_area_struct *dst_vma,
> > + struct vm_area_struct *src_vma,
> > + unsigned long dst_addr, unsigned long src_addr,
> > + pte_t *dst_pte, pte_t *src_pte,
> > + pte_t orig_dst_pte, pte_t orig_src_pte,
> > + pmd_t *dst_pmd, pmd_t dst_pmdval,
> > + spinlock_t *dst_ptl, spinlock_t *src_ptl,
> > + struct folio *src_folio, unsigned long len)
> > {
> > int err = 0;
> > + unsigned long src_start = src_addr;
> > + unsigned long addr_end;
> > +
> > + if (len > PAGE_SIZE) {
> > + addr_end = (dst_addr + PMD_SIZE) & PMD_MASK;
> > + if (dst_addr + len > addr_end)
> > + len = addr_end - dst_addr;
> >
> > + addr_end = (src_addr + PMD_SIZE) & PMD_MASK;
> > + if (src_addr + len > addr_end)
> > + len = addr_end - src_addr;
> > + }
> > + flush_cache_range(src_vma, src_addr, src_addr + len);
> > double_pt_lock(dst_ptl, src_ptl);
> >
> > if (!is_pte_pages_stable(dst_pte, src_pte, orig_dst_pte, orig_src_pte,
> > @@ -1051,31 +1095,60 @@ static int move_present_pte(struct mm_struct *mm,
> > err = -EBUSY;
> > goto out;
> > }
> > + /* Avoid batching overhead for single page case */
> > + if (len > PAGE_SIZE) {
> > + flush_tlb_batched_pending(mm);
>
> What’s confusing to me is that they track the unmapping of multiple
> consecutive PTEs and defer TLB invalidation until later.
> In contrast, you’re not tracking anything and instead call
> flush_tlb_range() directly, which triggers the flush immediately.
>
> It seems you might be combining two different batching approaches.
These changes I made are in line with how mremap() does batching. See
move_ptes() in mm/mremap.c
>From the comment in flush_tlb_batched_pending() [1] it seems necessary
in this case too. Please correct me if I'm wrong. I'll be happy to
remove it if it's not required.
[1] https://elixir.bootlin.com/linux/v6.16/source/mm/rmap.c#L728
> From what I can tell, you're essentially using flush_range
> as a replacement for flushing each entry individually.
That's correct. The idea is to reduce the number of IPIs required for
flushing the TLB entries. Since it is quite common that the ioctl is
invoked with several pages in one go, this greatly benefits.
>
> > + arch_enter_lazy_mmu_mode();
> > + orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
> > + } else
> > + orig_src_pte = ptep_clear_flush(src_vma, src_addr, src_pte);
> > +
> > + addr_end = src_start + len;
> > + do {
> > + /* Folio got pinned from under us. Put it back and fail the move. */
> > + if (folio_maybe_dma_pinned(src_folio)) {
> > + set_pte_at(mm, src_addr, src_pte, orig_src_pte);
> > + err = -EBUSY;
> > + break;
> > + }
> >
> > - orig_src_pte = ptep_clear_flush(src_vma, src_addr, src_pte);
> > - /* Folio got pinned from under us. Put it back and fail the move. */
> > - if (folio_maybe_dma_pinned(src_folio)) {
> > - set_pte_at(mm, src_addr, src_pte, orig_src_pte);
> > - err = -EBUSY;
> > - goto out;
> > - }
> > -
> > - folio_move_anon_rmap(src_folio, dst_vma);
> > - src_folio->index = linear_page_index(dst_vma, dst_addr);
> > + folio_move_anon_rmap(src_folio, dst_vma);
> > + src_folio->index = linear_page_index(dst_vma, dst_addr);
> >
> > - orig_dst_pte = folio_mk_pte(src_folio, dst_vma->vm_page_prot);
> > - /* Set soft dirty bit so userspace can notice the pte was moved */
> > + orig_dst_pte = folio_mk_pte(src_folio, dst_vma->vm_page_prot);
> > + /* Set soft dirty bit so userspace can notice the pte was moved */
> > #ifdef CONFIG_MEM_SOFT_DIRTY
> > - orig_dst_pte = pte_mksoft_dirty(orig_dst_pte);
> > + orig_dst_pte = pte_mksoft_dirty(orig_dst_pte);
> > #endif
> > - if (pte_dirty(orig_src_pte))
> > - orig_dst_pte = pte_mkdirty(orig_dst_pte);
> > - orig_dst_pte = pte_mkwrite(orig_dst_pte, dst_vma);
> > + if (pte_dirty(orig_src_pte))
> > + orig_dst_pte = pte_mkdirty(orig_dst_pte);
> > + orig_dst_pte = pte_mkwrite(orig_dst_pte, dst_vma);
> > + set_pte_at(mm, dst_addr, dst_pte, orig_dst_pte);
> > +
> > + src_addr += PAGE_SIZE;
> > + if (src_addr == addr_end)
> > + break;
> > + src_pte++;
> > + dst_pte++;
> >
> > - set_pte_at(mm, dst_addr, dst_pte, orig_dst_pte);
> > + folio_unlock(src_folio);
> > + src_folio = check_ptes_for_batched_move(src_vma, src_addr, src_pte, dst_pte);
> > + if (!src_folio)
> > + break;
> > + orig_src_pte = ptep_get_and_clear(mm, src_addr, src_pte);
> > + dst_addr += PAGE_SIZE;
> > + } while (true);
> > +
> > + if (len > PAGE_SIZE) {
> > + arch_leave_lazy_mmu_mode();
> > + if (src_addr > src_start)
> > + flush_tlb_range(src_vma, src_start, src_addr);
> > + }
>
> Can't we just remove the `if (len > PAGE_SIZE)` check and unify the
> handling for both single-page and multi-page cases?
We certainly can. Initially it seemed to me that lazy/batched
invalidation has its own overhead and I wanted to avoid that in the
single-page case because the ioctl does get called for single pages
quite a bit. That too in time sensitive code paths. However, on a
deeper relook now, I noticed it's not really that different.
I'll unify in the next patch. Thanks for the suggestion.
>
>
> Thanks
> Barry
Powered by blists - more mailing lists