Received-SPF: pass (google.com: domain of linux-kernel+bounces-112031-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:40f1:3f00::1 as permitted sender) client-ip=2604:1380:40f1:3f00::1;
Precedence: bulk
MIME-Version: 1.0
References: <CABOYnLyevJeravW=QrH0JUPYEcDN160aZFb7kwndm-J2rmz0HQ@mail.gmail.com>
 <CAJfpegu8qTARQBftZSaiif0E6dbRcbBvZvW7dQf8sf_ymoogCA@mail.gmail.com>
 <c58a8dc8-5346-4247-9a0a-8b1be286e779@redhat.com> <CAJfpegt3UCsMmxd0taOY11Uaw5U=eS1fE5dn0wZX3HF0oy8-oQ@mail.gmail.com>
 <620f68b0-4fe0-4e3e-856a-dedb4bcdf3a7@redhat.com>
In-Reply-To: <620f68b0-4fe0-4e3e-856a-dedb4bcdf3a7@redhat.com>
From: Miklos Szeredi <miklos@szeredi.hu>
Date: Fri, 22 Mar 2024 22:13:55 +0100
Message-ID: <CAJfpegub5Ny9kyX+dDbRwx7kd6ZdxtOeQ9RTK8n=LGGSzA9iOQ@mail.gmail.com>
Subject: Re: BUG: unable to handle kernel paging request in fuse_copy_do
To: David Hildenbrand <david@redhat.com>
Cc: xingwei lee <xrivendell7@gmail.com>, linux-fsdevel@vger.kernel.org, 
	linux-kernel@vger.kernel.org, samsun1006219@gmail.com, 
	syzkaller-bugs@googlegroups.com, linux-mm <linux-mm@kvack.org>, 
	Mike Rapoport <rppt@kernel.org>
Content-Type: text/plain; charset="UTF-8"

On Fri, 22 Mar 2024 at 22:08, David Hildenbrand <david@redhat.com> wrote:
>
> On 22.03.24 20:46, Miklos Szeredi wrote:
> > On Fri, 22 Mar 2024 at 16:41, David Hildenbrand <david@redhat.com> wrote:
> >
> >> But at least the vmsplice() just seems to work. Which is weird, because
> >> GUP-fast should not apply (page not faulted in?)
> >
> > But it is faulted in, and that indeed seems to be the root cause.
>
> secretmem mmap() won't populate the page tables. So it's not faulted in yet.
>
> When we GUP via vmsplice, GUP-fast should not find it in the page tables
> and fallback to slow GUP.
>
> There, we seem to pass check_vma_flags(), trigger faultin_page() to
> fault it in, and then find it via follow_page_mask().
>
> ... and I wonder how we manage to skip check_vma_flags(), or otherwise
> managed to GUP it.
>
> vmsplice() should, in theory, never succeed here.
>
> Weird :/
>
> > Improved repro:
> >
> > #define _GNU_SOURCE
> >
> > #include <fcntl.h>
> > #include <unistd.h>
> > #include <stdio.h>
> > #include <errno.h>
> > #include <sys/mman.h>
> > #include <sys/syscall.h>
> >
> > int main(void)
> > {
> >          int fd1, fd2;
> >          int pip[2];
> >          struct iovec iov;
> >          char *addr;
> >          int ret;
> >
> >          fd1 = syscall(__NR_memfd_secret, 0);
> >          addr = mmap(NULL, 4096, PROT_READ | PROT_WRITE, MAP_SHARED, fd1, 0);
> >          ftruncate(fd1, 7);
> >          addr[0] = 1; /* fault in page */

Here the page is faulted in and GUP-fast will find it.  It's not in
the kernel page table, but it is in the user page table, which is what
matter for GUP.

Thanks,
Miklos