Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20230609005158.2421285-1-surenb@google.com> <20230609005158.2421285-5-surenb@google.com>
 <ZIOOmC26qh4EXUEX@x1n> <CAJuCfpHKUjAwgWbxvJQDyEnneRD03p2M6247Q6=3-oOq_FL7zA@mail.gmail.com>
 <CAJuCfpG3PrbGxpDAEkyGQXW88+otb=FsbrhPJ4ePN7Xhn0a+_A@mail.gmail.com> <ZIcfYQ1c5teMSHAX@x1n>
In-Reply-To: <ZIcfYQ1c5teMSHAX@x1n>
From:   Suren Baghdasaryan <surenb@google.com>
Date:   Mon, 12 Jun 2023 09:07:38 -0700
Message-ID: <CAJuCfpGZvhBUdfNHojXwqZbspuhy0bstjT+-JMfwgmnqTnkoHA@mail.gmail.com>
Subject: Re: [PATCH v2 4/6] mm: drop VMA lock before waiting for migration
To:     Peter Xu <peterx@redhat.com>
Cc:     akpm@linux-foundation.org, willy@infradead.org, hannes@cmpxchg.org,
        mhocko@suse.com, josef@toxicpanda.com, jack@suse.cz,
        ldufour@linux.ibm.com, laurent.dufour@fr.ibm.com,
        michel@lespinasse.org, liam.howlett@oracle.com, jglisse@google.com,
        vbabka@suse.cz, minchan@google.com, dave@stgolabs.net,
        punit.agrawal@bytedance.com, lstoakes@gmail.com, hdanton@sina.com,
        apopple@nvidia.com, ying.huang@intel.com, david@redhat.com,
        yuzhao@google.com, dhowells@redhat.com, hughd@google.com,
        viro@zeniv.linux.org.uk, brauner@kernel.org,
        pasha.tatashin@soleen.com, linux-mm@kvack.org,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org,
        kernel-team@android.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Mon, Jun 12, 2023 at 6:36=E2=80=AFAM Peter Xu <peterx@redhat.com> wrote:
>
> On Fri, Jun 09, 2023 at 06:29:43PM -0700, Suren Baghdasaryan wrote:
> > On Fri, Jun 9, 2023 at 3:30=E2=80=AFPM Suren Baghdasaryan <surenb@googl=
e.com> wrote:
> > >
> > > On Fri, Jun 9, 2023 at 1:42=E2=80=AFPM Peter Xu <peterx@redhat.com> w=
rote:
> > > >
> > > > On Thu, Jun 08, 2023 at 05:51:56PM -0700, Suren Baghdasaryan wrote:
> > > > > migration_entry_wait does not need VMA lock, therefore it can be =
dropped
> > > > > before waiting. Introduce VM_FAULT_VMA_UNLOCKED to indicate that =
VMA
> > > > > lock was dropped while in handle_mm_fault().
> > > > > Note that once VMA lock is dropped, the VMA reference can't be us=
ed as
> > > > > there are no guarantees it was not freed.
> > > >
> > > > Then vma lock behaves differently from mmap read lock, am I right? =
 Can we
> > > > still make them match on behaviors, or there's reason not to do so?
> > >
> > > I think we could match their behavior by also dropping mmap_lock here
> > > when fault is handled under mmap_lock (!(fault->flags &
> > > FAULT_FLAG_VMA_LOCK)).
> > > I missed the fact that VM_FAULT_COMPLETED can be used to skip droppin=
g
> > > mmap_lock in do_page_fault(), so indeed, I might be able to use
> > > VM_FAULT_COMPLETED to skip vma_end_read(vma) for per-vma locks as wel=
l
> > > instead of introducing FAULT_FLAG_VMA_LOCK. I think that was your ide=
a
> > > of reusing existing flags?
> > Sorry, I meant VM_FAULT_VMA_UNLOCKED, not FAULT_FLAG_VMA_LOCK in the
> > above reply.
> >
> > I took a closer look into using VM_FAULT_COMPLETED instead of
> > VM_FAULT_VMA_UNLOCKED but when we fall back from per-vma lock to
> > mmap_lock we need to retry with an indication that the per-vma lock
> > was dropped. Returning (VM_FAULT_RETRY | VM_FAULT_COMPLETE) to
> > indicate such state seems strange to me ("retry" and "complete" seem
>
> Not relevant to this migration patch, but for the whole idea I was thinki=
ng
> whether it should just work if we simply:
>
>         fault =3D handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LO=
CK, regs);
> -       vma_end_read(vma);
> +       if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
> +               vma_end_read(vma);
>
> ?

Today when we can't handle a page fault under per-vma locks we return
VM_FAULT_RETRY, in which case per-vma lock is dropped and the fault is
retried under mmap_lock. The condition you suggest above would not
drop per-vma lock for VM_FAULT_RETRY case and would break the current
fallback mechanism.
However your suggestion gave me an idea. I could indicate that per-vma
lock got dropped using vmf structure (like Matthew suggested before)
and once handle_pte_fault(vmf) returns I could check if it returned
VM_FAULT_RETRY but per-vma lock is still held. If that happens I can
call vma_end_read() before returning from __handle_mm_fault(). That
way any time handle_mm_fault() returns VM_FAULT_RETRY per-vma lock
will be already released, so your condition in do_page_fault() will
work correctly. That would eliminate the need for a new
VM_FAULT_VMA_UNLOCKED flag. WDYT?

>
> GUP may need more caution on NOWAIT, but vma lock is only in fault paths =
so
> IIUC it's fine?
>
> --
> Peter Xu
>
> --
> To unsubscribe from this group and stop receiving emails from it, send an=
 email to kernel-team+unsubscribe@android.com.
>