Re: [PATCH v4 2/3] mm/khugepaged: Fix GUP-fast interaction by sending IPI

From: Peter Xu
Date: Mon Nov 28 2022 - 15:16:51 EST


On Mon, Nov 28, 2022 at 08:56:54PM +0100, Jann Horn wrote:
> On Mon, Nov 28, 2022 at 8:54 PM Yang Shi <shy828301@xxxxxxxxx> wrote:
> >
> > On Mon, Nov 28, 2022 at 10:03 AM Jann Horn <jannh@xxxxxxxxxx> wrote:
> > >
> > > Since commit 70cbc3cc78a99 ("mm: gup: fix the fast GUP race against THP
> > > collapse"), the lockless_pages_from_mm() fastpath rechecks the pmd_t to
> > > ensure that the page table was not removed by khugepaged in between.
> > >
> > > However, lockless_pages_from_mm() still requires that the page table is not
> > > concurrently freed or reused to store non-PTE data. Otherwise, problems
> > > can occur because:
> > >
> > > - deposited page tables can be freed when a THP page somewhere in the
> > > mm is removed
> > > - some architectures store non-PTE information inside deposited page
> > > tables (see radix__pgtable_trans_huge_deposit())
> > >
> > > Additionally, lockless_pages_from_mm() is also somewhat brittle with
> > > regards to page tables being repeatedly moved back and forth, but
> > > that shouldn't be an issue in practice.
> > >
> > > Fix it by sending IPIs (if the architecture uses
> > > semi-RCU-style page table freeing) before freeing/reusing page tables.
> > >
> > > As noted in mm/gup.c, on configs that define CONFIG_HAVE_FAST_GUP,
> > > there are two possible cases:
> > >
> > > 1. CONFIG_MMU_GATHER_RCU_TABLE_FREE is set, causing
> > > tlb_remove_table_sync_one() to send an IPI to synchronize with
> > > lockless_pages_from_mm().
> > > 2. CONFIG_MMU_GATHER_RCU_TABLE_FREE is unset, indicating that all
> > > TLB flushes are already guaranteed to send IPIs.
> > > tlb_remove_table_sync_one() will do nothing, but we've already
> > > run pmdp_collapse_flush(), which did a TLB flush, which must have
> > > involved IPIs.
> >
> > I'm trying to catch up with the discussion after the holiday break. I
> > understand you switched from always allocating a new page table page
> > (we decided before) to sending IPIs to serialize against fast-GUP,
> > this is fine to me.
> >
> > So the code now looks like:
> > pmdp_collapse_flush()
> > sending IPI
> >
> > But the missing part is how we reached "TLB flushes are already
> > guaranteed to send IPIs" when CONFIG_MMU_GATHER_RCU_TABLE_FREE is
> > unset? ARM64 doesn't do it IIRC. Or did I miss something?
>
> From arch/arm64/Kconfig:
>
> select MMU_GATHER_RCU_TABLE_FREE
>
> CONFIG_MMU_GATHER_RCU_TABLE_FREE is not a config option that the user
> can freely toggle; it is an option selected by the architecture.

True.

I think I understand what Yang is confused about and I had the same
question (asked in the old threads but didn't yet got a confirmation),
since I think arm64 didn't use IPI for tlb is also true (according to the
arm64 version of __flush_tlb_range), so PPC doesn't seem to be the only one.

I mentioned PPC only because I saw the comment in mmu_gather.c:

* Architectures that do not have this (PPC) need to delay the freeing by some
* other means, this is that means.

So I think it's obsolete.

In short, IIUC there's just an implicit dependency that any
!MMU_GATHER_RCU_TABLE_FREE arch must require IPI for tlb flush (not vice
versa, hence arm64 can have RCU_TABLE_FREE), or something could be broken.

--
Peter Xu