linux-kernel.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: "Aneesh Kumar K.V" <aneesh.kumar@linux.ibm.com>
To: Yang Shi <shy828301@gmail.com>, Peter Xu <peterx@redhat.com>
Cc: david@redhat.com, kirill.shutemov@linux.intel.com,
	jhubbard@nvidia.com, jgg@nvidia.com, hughd@google.com,
	akpm@linux-foundation.org, linux-mm@kvack.org,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH] mm: gup: fix the fast GUP race against THP collapse
Date: Mon, 05 Sep 2022 14:24:58 +0530	[thread overview]
Message-ID: <87ilm2jj4t.fsf@linux.ibm.com> (raw)
In-Reply-To: <CAHbLzkqGdnwY4P8jKQR0ojm6QV6b3dBi5pwrC1UJ4dqi3EqS4w@mail.gmail.com>

Yang Shi <shy828301@gmail.com> writes:

>
> On Fri, Sep 2, 2022 at 9:00 AM Peter Xu <peterx@redhat.com> wrote:
>>
>> On Thu, Sep 01, 2022 at 04:50:45PM -0700, Yang Shi wrote:
>> > On Thu, Sep 1, 2022 at 4:26 PM Peter Xu <peterx@redhat.com> wrote:
>> > >
>> > > Hi, Yang,
>> > >
>> > > On Thu, Sep 01, 2022 at 03:27:07PM -0700, Yang Shi wrote:
>> > > > Since general RCU GUP fast was introduced in commit 2667f50e8b81 ("mm:
>> > > > introduce a general RCU get_user_pages_fast()"), a TLB flush is no longer
>> > > > sufficient to handle concurrent GUP-fast in all cases, it only handles
>> > > > traditional IPI-based GUP-fast correctly.
>> > >
>> > > If TLB flush (or, IPI broadcasts) used to work to protect against gup-fast,
>> > > I'm kind of confused why it's not sufficient even if with RCU gup?  Isn't
>> > > that'll keep working as long as interrupt disabled (which current fast-gup
>> > > will still do)?
>> >
>> > Actually the wording was copied from David's commit log for his
>> > PageAnonExclusive fix. My understanding is the IPI broadcast still
>> > works, but it may not be supported by all architectures and not
>> > preferred anymore. So we should avoid depending on IPI broadcast IIUC.
>> >
>> > >
>> > > IIUC the issue is you suspect not all archs correctly implemented
>> > > pmdp_collapse_flush(), or am I wrong?
>> >
>> > This is a possible fix, please see below for details.
>> >
>> > >
>> > > > On architectures that send
>> > > > an IPI broadcast on TLB flush, it works as expected.  But on the
>> > > > architectures that do not use IPI to broadcast TLB flush, it may have
>> > > > the below race:
>> > > >
>> > > >    CPU A                                          CPU B
>> > > > THP collapse                                     fast GUP
>> > > >                                               gup_pmd_range() <-- see valid pmd
>> > > >                                                   gup_pte_range() <-- work on pte
>> > > > pmdp_collapse_flush() <-- clear pmd and flush
>> > > > __collapse_huge_page_isolate()
>> > > >     check page pinned <-- before GUP bump refcount
>> > > >                                                       pin the page
>> > > >                                                       check PTE <-- no change
>> > > > __collapse_huge_page_copy()
>> > > >     copy data to huge page
>> > > >     ptep_clear()
>> > > > install huge pmd for the huge page
>> > > >                                                       return the stale page
>> > > > discard the stale page
>> > > >
>> > > > The race could be fixed by checking whether PMD is changed or not after
>> > > > taking the page pin in fast GUP, just like what it does for PTE.  If the
>> > > > PMD is changed it means there may be parallel THP collapse, so GUP
>> > > > should back off.
>> > >
>> > > Could the race also be fixed by impl pmdp_collapse_flush() correctly for
>> > > the archs that are missing? Do you know which arch(s) is broken with it?
>> >
>> > Yes, and this was suggested by me in the first place, but per the
>> > suggestion from John and David, this is not the preferred way. I think
>> > it is because:
>> >
>> > Firstly, using IPI to serialize against fast GUP is not recommended
>> > anymore since fast GUP does check PTE then back off so we should avoid
>> > it.
>> > Secondly, if checking PMD then backing off could solve the problem,
>> > why do we still need broadcast IPI? It doesn't sound performant.
>> >
>> > >
>> > > It's just not clear to me whether this patch is an optimization or a fix,
>> > > if it's a fix whether the IPI broadcast in ppc pmdp_collapse_flush() would
>> > > still be needed.
>> >
>> > It is a fix and the fix will make IPI broadcast not useful anymore.
>>
>> How about another patch to remove the ppc impl too?  Then it can be a two
>> patches series.
>
> BTW, I don't think we could remove the ppc implementation since it is
> different from the generic pmdp_collapse_flush(), particularly for the
> hash part IIUC.
>
> The generic version calls flush_tlb_range() -> hash__flush_tlb_range()
> for hash, but the hash call is actually no-op. The ppc version calls
> hash__pmdp_collapse_flush() -> flush_tlb_pmd_range(), which does
> something useful.
>

We should actually rename flush_tlb_pmd_range(). It actually flush the
hash page table entries.

I will do the below patch for ppc64 to clarify this better

diff --git a/arch/powerpc/include/asm/book3s/64/tlbflush-hash.h b/arch/powerpc/include/asm/book3s/64/tlbflush-hash.h
index 8b762f282190..fd30fa20c392 100644
--- a/arch/powerpc/include/asm/book3s/64/tlbflush-hash.h
+++ b/arch/powerpc/include/asm/book3s/64/tlbflush-hash.h
@@ -112,13 +112,12 @@ static inline void hash__flush_tlb_kernel_range(unsigned long start,
 
 struct mmu_gather;
 extern void hash__tlb_flush(struct mmu_gather *tlb);
-void flush_tlb_pmd_range(struct mm_struct *mm, pmd_t *pmd, unsigned long addr);
 
 #ifdef CONFIG_PPC_64S_HASH_MMU
 /* Private function for use by PCI IO mapping code */
 extern void __flush_hash_table_range(unsigned long start, unsigned long end);
-extern void flush_tlb_pmd_range(struct mm_struct *mm, pmd_t *pmd,
-				unsigned long addr);
+extern void flush_hash_table_pmd_range(struct mm_struct *mm, pmd_t *pmd,
+				       unsigned long addr);
 #else
 static inline void __flush_hash_table_range(unsigned long start, unsigned long end) { }
 #endif
diff --git a/arch/powerpc/mm/book3s64/hash_pgtable.c b/arch/powerpc/mm/book3s64/hash_pgtable.c
index ae008b9df0e6..f30131933a01 100644
--- a/arch/powerpc/mm/book3s64/hash_pgtable.c
+++ b/arch/powerpc/mm/book3s64/hash_pgtable.c
@@ -256,7 +256,7 @@ pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long addres
 	 * the __collapse_huge_page_copy can result in copying
 	 * the old content.
 	 */
-	flush_tlb_pmd_range(vma->vm_mm, &pmd, address);
+	flush_hash_table_pmd_range(vma->vm_mm, &pmd, address);
 	return pmd;
 }
 
diff --git a/arch/powerpc/mm/book3s64/hash_tlb.c b/arch/powerpc/mm/book3s64/hash_tlb.c
index eb0bccaf221e..a64ea0a7ef96 100644
--- a/arch/powerpc/mm/book3s64/hash_tlb.c
+++ b/arch/powerpc/mm/book3s64/hash_tlb.c
@@ -221,7 +221,7 @@ void __flush_hash_table_range(unsigned long start, unsigned long end)
 	local_irq_restore(flags);
 }
 
-void flush_tlb_pmd_range(struct mm_struct *mm, pmd_t *pmd, unsigned long addr)
+void flush_hash_table_pmd_range(struct mm_struct *mm, pmd_t *pmd, unsigned long addr)
 {
 	pte_t *pte;
 	pte_t *start_pte;


  parent reply	other threads:[~2022-09-05  8:56 UTC|newest]

Thread overview: 38+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-09-01 22:27 [PATCH] mm: gup: fix the fast GUP race against THP collapse Yang Shi
2022-09-01 23:26 ` Peter Xu
2022-09-01 23:50   ` Yang Shi
2022-09-02  6:39     ` David Hildenbrand
2022-09-02 15:23       ` Yang Shi
2022-09-02 15:59     ` Peter Xu
2022-09-02 16:04       ` Peter Xu
2022-09-02 17:30       ` Yang Shi
2022-09-02 17:45       ` Yang Shi
2022-09-02 20:33         ` Peter Xu
2022-09-05  8:56           ` Aneesh Kumar K.V
2022-09-05  8:54         ` Aneesh Kumar K.V [this message]
2022-09-06 19:07           ` Yang Shi
2022-09-07  4:50             ` Aneesh Kumar K V
2022-09-07 17:08               ` Yang Shi
2022-09-04 22:21       ` John Hubbard
2022-09-02  6:42 ` David Hildenbrand
2022-09-04 22:29 ` John Hubbard
2022-09-05  7:59   ` David Hildenbrand
2022-09-05 10:16     ` Baolin Wang
2022-09-05 10:24       ` David Hildenbrand
2022-09-05 11:11         ` David Hildenbrand
2022-09-05 14:35           ` Baolin Wang
2022-09-05 14:40             ` David Hildenbrand
2022-09-06  5:53               ` Baolin Wang
2022-09-06  2:12     ` John Hubbard
2022-09-06 12:50       ` David Hildenbrand
2022-09-06 13:47     ` Jason Gunthorpe
2022-09-06 13:57       ` David Hildenbrand
2022-09-06 14:30         ` Jason Gunthorpe
2022-09-06 14:44           ` David Hildenbrand
2022-09-06 15:33             ` Jason Gunthorpe
2022-09-06 19:11             ` Yang Shi
2022-09-06 23:16             ` John Hubbard
2022-09-06 19:01     ` Yang Shi
2022-09-05  9:03   ` Baolin Wang
2022-09-06 18:50   ` Yang Shi
2022-09-06 21:27     ` John Hubbard

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=87ilm2jj4t.fsf@linux.ibm.com \
    --to=aneesh.kumar@linux.ibm.com \
    --cc=akpm@linux-foundation.org \
    --cc=david@redhat.com \
    --cc=hughd@google.com \
    --cc=jgg@nvidia.com \
    --cc=jhubbard@nvidia.com \
    --cc=kirill.shutemov@linux.intel.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=peterx@redhat.com \
    --cc=shy828301@gmail.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).