Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Wed, 22 Sep 2021 18:22:45 -0700 (PDT)
From:   Hugh Dickins <hughd@google.com>
To:     Peter Xu <peterx@redhat.com>
cc:     Hugh Dickins <hughd@google.com>,
        Axel Rasmussen <axelrasmussen@google.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Linux MM <linux-mm@kvack.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Andrea Arcangeli <aarcange@redhat.com>,
        Nadav Amit <nadav.amit@gmail.com>
Subject: Re: [PATCH] mm/khugepaged: Detecting uffd-wp vma more efficiently
In-Reply-To: <YUu/6lWX92WLUzpt@t490s>
Message-ID: <472a3497-ba70-ac6b-5828-bc5c4c93e9ab@google.com>
References: <20210922175156.130228-1-peterx@redhat.com> <CAJHvVch3g_UY-akMdu0O9413iCb1H83DLhR1Am8WnfUyV=s1=g@mail.gmail.com> <YUueOUfoamxOvEyO@t490s> <24224366-293a-879-95db-f69abcb0cb70@google.com> <YUu/6lWX92WLUzpt@t490s>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Precedence: bulk

On Wed, 22 Sep 2021, Peter Xu wrote:
> On Wed, Sep 22, 2021 at 04:18:09PM -0700, Hugh Dickins wrote:
> > On Wed, 22 Sep 2021, Peter Xu wrote:
> > > 
> > > Not installing pmd means uffd-minor can still trap any further faults just like
> > > before, afaiu.
> > > 
> > > There's a very trivial detail that the pmd missing case will have a very slight
> > > code path change when the next page fault happens: in __handle_mm_fault() we'll
> > > first try to go into create_huge_pmd() once, however since shmem didn't provide
> > > huge_fault(), we'll go the VM_FAULT_FALLBACK path, and things will go like
> > > before when faulting on a small pte.  The next UFFDIO_CONTINUE will allocate
> > > that missing pmd again, however it'll install a 4K page only.
> > 
> > I think you're mistaken there.
> > 
> > I can't tell you much about ->huge_fault(), something introduced for
> > DAX I believe; but shmem has managed pmd mappings without it, since
> > before ->huge_fault() was ever added.
> 
> Right, I wanted to express we didn't go into there, hence no way to allocate
> pmd there.
> 
> > 
> > Look for the call to do_set_pmd() in finish_fault(): I think you'll
> > find that is the way shmem's huge pmds get in.
> > 
> > Earlier in the thread you suggested "shmem_getpage() only returns
> > small pages": but it can very well return PageTransCompound pages,
> > head or tail, which arrive at this do_set_pmd().
> 
> But note that uffd-minor will trap the shmem fault() even if pmd_none:
> 
> 	page = pagecache_get_page(mapping, index,
> 					FGP_ENTRY | FGP_HEAD | FGP_LOCK, 0);
> 
> 	if (page && vma && userfaultfd_minor(vma)) {
> 		if (!xa_is_value(page)) {
> 			unlock_page(page);
> 			put_page(page);
> 		}
> 		*fault_type = handle_userfault(vmf, VM_UFFD_MINOR);
> 		return 0;
> 	}
> 
> That's why I think it'll be fine, because it should only be UFFDIO_CONTINUE
> that installs the pte (alongside with allocating the pmd).
> 
> Or did I miss something?

No, I think I misunderstood you before: thanks for re-explaining.
(And Axel's !userfaultfd_minor() check before calling do_fault_around()
plays an important part in making sure that it does reach shmem_fault().)

Hugh