Re: [syzbot] [mm?] WARNING in __page_table_check_ptes_set

[David Hildenbrand]

On 22.04.24 12:38, David Hildenbrand wrote:
> On 22.04.24 12:07, David Hildenbrand wrote:
> > On 21.04.24 22:16, syzbot wrote:
> > > Hello,
> > >
> > > syzbot found the following issue on:
> > >
> > > HEAD commit:    4eab35893071 Add linux-next specific files for 20240417
> > > git tree:       linux-next
> > > console+strace: https://syzkaller.appspot.com/x/log.txt?x=1727a61b180000
> > > kernel config:  https://syzkaller.appspot.com/x/.config?x=27920e47287645ff
> > > dashboard link: https://syzkaller.appspot.com/bug?extid=d8426b591c36b21c750e
> > > compiler:       Debian clang version 15.0.6, GNU ld (GNU Binutils for Debian) 2.40
> > > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=156da22d180000
> > > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=163dfec7180000
> > >
> > > Downloadable assets:
> > > disk image: https://storage.googleapis.com/syzbot-assets/9f7d6c097fb4/disk-4eab3589.raw.xz
> > > vmlinux: https://storage.googleapis.com/syzbot-assets/287b16352982/vmlinux-4eab3589.xz
> > > kernel image: https://storage.googleapis.com/syzbot-assets/23839c65c573/bzImage-4eab3589.xz
> > >
> > > IMPORTANT: if you fix the issue, please add the following tag to the commit:
> > > Reported-by: syzbot+d8426b591c36b21c750e@xxxxxxxxxxxxxxxxxxxxxxxxx
> > >
> > > ------------[ cut here ]------------
> > > WARNING: CPU: 0 PID: 5084 at mm/page_table_check.c:199 __page_table_check_pte mm/page_table_check.c:199 [inline]
> > > WARNING: CPU: 0 PID: 5084 at mm/page_table_check.c:199 __page_table_check_ptes_set+0x1db/0x420
> >
> > I think this is
> >
> > if (pte_present(pte) && pte_uffd_wp(pte))
> > 	WARN_ON_ONCE(pte_write(pte));
> >
> > mm/page_table_check.c:213
> > > Modules linked in:
> > > CPU: 0 PID: 5084 Comm: syz-executor382 Not tainted 6.9.0-rc4-next-20240417-syzkaller #0
> > > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 03/27/2024
> > > RIP: 0010:__page_table_check_pte mm/page_table_check.c:199 [inline]
> > > RIP: 0010:__page_table_check_ptes_set+0x1db/0x420 mm/page_table_check.c:213
> > > Code: 48 8b 7c 24 40 48 c7 c6 80 19 46 8e e8 ee df 8e ff 41 83 fc 1d 74 18 41 83 fc 1a 75 1d e8 5d da 8e ff eb 10 e8 56 da 8e ff 90 <0f> 0b 90 eb 10 e8 4b da 8e ff 90 0f 0b 90 eb 05 e8 40 da 8e ff 48
> > > RSP: 0018:ffffc9000366f740 EFLAGS: 00010293
> > > RAX: ffffffff8207833a RBX: ffffc9000366f7c0 RCX: ffff888022af3c00
> > > RDX: 0000000000000000 RSI: 0000000000000004 RDI: 0000000000000000
> > > RBP: ffffc9000366f830 R08: ffffffff820782af R09: 1ffffd40000a6a10
> > > R10: dffffc0000000000 R11: fffff940000a6a11 R12: 0000000000000000
> > > R13: 0000000014d42c67 R14: 0000000000000001 R15: 0000000000000000
> > > FS:  0000555567f79380(0000) GS:ffff8880b9400000(0000) knlGS:0000000000000000
> > > CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > CR2: 000000000066c7e0 CR3: 0000000078cb0000 CR4: 00000000003506f0
> > > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > > Call Trace:
> > >     <TASK>
> > >     page_table_check_ptes_set include/linux/page_table_check.h:74 [inline]
> > >     set_ptes include/linux/pgtable.h:267 [inline]
> > >     __ptep_modify_prot_commit include/linux/pgtable.h:1269 [inline]
> > >     ptep_modify_prot_commit include/linux/pgtable.h:1302 [inline]
> > >     change_pte_range mm/mprotect.c:194 [inline]
> > >     change_pmd_range mm/mprotect.c:424 [inline]
> > >     change_pud_range mm/mprotect.c:457 [inline]
> > >     change_p4d_range mm/mprotect.c:480 [inline]
> > >     change_protection_range mm/mprotect.c:508 [inline]
> > >     change_protection+0x2770/0x3cc0 mm/mprotect.c:542
> > >     mprotect_fixup+0x740/0xa90 mm/mprotect.c:655
> > >     do_mprotect_pkey+0x90d/0xe00 mm/mprotect.c:820
> > >     __do_sys_mprotect mm/mprotect.c:841 [inline]
> > >     __se_sys_mprotect mm/mprotect.c:838 [inline]
> > >     __x64_sys_mprotect+0x80/0x90 mm/mprotect.c:838
> > >     do_syscall_x64 arch/x86/entry/common.c:52 [inline]
> > >     do_syscall_64+0xf5/0x240 arch/x86/entry/common.c:83
> > >     entry_SYSCALL_64_after_hwframe+0x77/0x7f
> > > RIP: 0033:0x7f45514bf429
> > > Code: 48 83 c4 28 c3 e8 37 17 00 00 0f 1f 80 00 00 00 00 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 b8 ff ff ff f7 d8 64 89 01 48
> > > RSP: 002b:00007ffe52191598 EFLAGS: 00000246 ORIG_RAX: 000000000000000a
> > > RAX: ffffffffffffffda RBX: 00007ffe52191768 RCX: 00007f45514bf429
> > > RDX: 000000000000000f RSI: 0000000000004000 RDI: 0000000020ffc000
> > > RBP: 00007f4551532610 R08: 00007ffe52191768 R09: 00007ffe52191768
> > > R10: 00007ffe52191768 R11: 0000000000000246 R12: 0000000000000001
> > > R13: 00007ffe52191758 R14: 0000000000000001 R15: 0000000000000001
> > >     </TASK>
> >
> > Did we find a real issue that involves mprotect()?
> >
> > At least can_change_pte_writable() should always return "false" for
> > userfaultfd_pte_wp().
> >
> > Do we maybe have a uffd-wp PTE outside of a UFFD_WP VMA?
> >
> > Or was the PTE already writable and we only detect it now as we call
> > mprotect()? (missed to detect it earlier?)
>
> Staring at the reproducer, we do
>
>
>     syscall(__NR_mmap, /*addr=*/0x1ffff000ul, /*len=*/0x1000ul, /*prot=*/0ul,
>             /*flags=MAP_FIXED|MAP_ANONYMOUS|MAP_PRIVATE*/ 0x32ul, /*fd=*/-1,
>             /*offset=*/0ul);
>     syscall(__NR_mmap, /*addr=*/0x20000000ul, /*len=*/0x1000000ul,
>             /*prot=PROT_WRITE|PROT_READ|PROT_EXEC*/ 7ul,
>             /*flags=MAP_FIXED|MAP_ANONYMOUS|MAP_PRIVATE*/ 0x32ul, /*fd=*/-1,
>             /*offset=*/0ul);
>
> -> Writable anonymous memmory
>
>     syscall(__NR_mmap, /*addr=*/0x21000000ul, /*len=*/0x1000ul, /*prot=*/0ul,
>             /*flags=MAP_FIXED|MAP_ANONYMOUS|MAP_PRIVATE*/ 0x32ul, /*fd=*/-1,
>             /*offset=*/0ul);
>     intptr_t res = 0;
>     res = syscall(__NR_userfaultfd,
>                   /*flags=UFFD_USER_MODE_ONLY|O_NONBLOCK*/ 0x801ul);
>     if (res != -1)
>       r[0] = res;
>     *(uint64_t*)0x200004c0 = 0xaa;
>     *(uint64_t*)0x200004c8 = 0;
>     *(uint64_t*)0x200004d0 = 0;
>     syscall(__NR_ioctl, /*fd=*/r[0], /*cmd=*/0xc018aa3f, /*arg=*/0x200004c0ul);
>
> -> _UFFDIO_API handshake?
>
>     syscall(__NR_mprotect, /*addr=*/0x20ffc000ul, /*len=*/0x3000ul,
>             /*prot=PROT_SEM|PROT_EXEC*/ 0xcul);
>
> -> Protect target range R/O. I assume: no page populated yet?
> -> 3 pages starting at 0x20ffc000ul;
>
>     *(uint64_t*)0x20000180 = 0x20ffc000;
>     *(uint64_t*)0x20000188 = 0x3000;
>     *(uint64_t*)0x20000190 = 3;
>     *(uint64_t*)0x20000198 = 0;
>     syscall(__NR_ioctl, /*fd=*/r[0], /*cmd=*/0xc020aa00, /*arg=*/0x20000180ul);
>
> -> _UFFDIO_REGISTER (aa00)
> -> _range = 3 pages starting at 0x20ffc000ul
> -> _mode = UFFDIO_REGISTER_MODE_WP | UFFDIO_REGISTER_MODE_MINOR
>
>     *(uint64_t*)0x20000000 = 0x20ffd000;
>     *(uint64_t*)0x20000008 = 0x20ffb000;
>     *(uint64_t*)0x20000010 = 0x1000;
>     *(uint64_t*)0x20000018 = 3;
>     *(uint64_t*)0x20000020 = 0;
>     syscall(__NR_ioctl, /*fd=*/r[0], /*cmd=*/0xc028aa03, /*arg=*/0x20000000ul);
>
> -> _UFFDIO_COPY (aa03)
> -> dst = 0x20ffd000
> -> src = 0x20ffb000
> -> len = 0x1000 (single page)
> -> mode = UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP
>
> -> We are copying into the R/O range. src should be R/W and trigger a page fault
>      on access where we get a fresh page.
>
>     *(uint16_t*)0x200000c0 = 1;
>     *(uint64_t*)0x200000c8 = 0x20000040;
>     *(uint16_t*)0x20000040 = 6;
>     *(uint8_t*)0x20000042 = 0;
>     *(uint8_t*)0x20000043 = 0;
>     *(uint32_t*)0x20000044 = 0x7fffffff;
>     res = syscall(__NR_seccomp, /*op=*/1ul, /*flags=*/0ul, /*arg=*/0x200000c0ul);
>     if (res != -1)
>       r[1] = res;
>     syscall(__NR_open_tree, /*dfd=*/-1, /*filename=*/0ul, /*flags=*/0ul);
>
> -> No idea what happens here and if it is relevant. If __NR_seccomp failed, we would
>      no set r[1].
>
>     syscall(__NR_close_range, /*fd=*/r[1], /*max_fd=*/-1, /*flags=*/0ul);
>
> -> Is that closing uffd as well, especially if __NR_seccomp failed?
>
>     syscall(__NR_mprotect, /*addr=*/0x20ffc000ul, /*len=*/0x4000ul,
>             /*prot=PROT_SEM|PROT_WRITE|PROT_READ|PROT_EXEC*/ 0xful);
>
> -> Restore write permissions. This seems to fire the uffd-wp page table check I assume.

I think the issue is that userfaultfd_release() will clear the VMA UFFD_WP flag,
but it will not clear PTE uffd-wp bits. So we have leftover PTE uffd-wp bits at
the time we wr-unprotect.

I thought we removed that lazy handling, but looks like we didn't consider the
"close uffd" case in:

commit f369b07c861435bd812a9d14493f71b34132ed6f
Author: Peter Xu <peterx@xxxxxxxxxx>
Date:   Thu Aug 11 16:13:40 2022 -0400

    mm/uffd: reset write protection when unregister with wp-mode


close should behave just like unregister.


Simplified+readable reproducer:

#define _GNU_SOURCE

#include <stdint.h>
#include <fcntl.h>
#include <sys/syscall.h>
#include <sys/mman.h>
#include <sys/types.h>
#include <sys/ioctl.h>
#include <linux/userfaultfd.h>
#include <unistd.h>

int main(void)
{
        void *src = mmap(0, 4096, PROT_READ, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
        void *dst = mmap(0, 4096, PROT_READ, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
        struct uffdio_register uffdio_register = {};
        struct uffdio_copy uffdio_copy = {};
        struct uffdio_api uffdio_api = {};
        int uffd;

        uffd = syscall(SYS_userfaultfd, O_CLOEXEC | O_NONBLOCK | UFFD_USER_MODE_ONLY);
        uffdio_api.api = UFFD_API;
        ioctl(uffd, UFFDIO_API, &uffdio_api);

        uffdio_register.range.start = (uintptr_t)dst;
        uffdio_register.range.len = 4096;
        uffdio_register.mode = UFFDIO_REGISTER_MODE_WP;
        ioctl(uffd, UFFDIO_REGISTER, &uffdio_register);

        uffdio_copy.dst = (uintptr_t)dst;
        uffdio_copy.src = (uintptr_t)src;
        uffdio_copy.len = 4096;
        uffdio_copy.mode = UFFDIO_COPY_MODE_DONTWAKE|UFFDIO_COPY_MODE_WP;
        ioctl(uffd, UFFDIO_COPY, &uffdio_copy);

        close(uffd);

        mprotect(dst, 4096, PROT_READ|PROT_WRITE);
        return 0;
}


--
Cheers,

David / dhildenb