Re: [PATCH] mm: Fix mmap_assert_locked() in follow_pte()

[David Hildenbrand]

On 11.07.24 07:13, Pei Li wrote:
> This patch fixes this warning by acquiring read lock before entering
> untrack_pfn() while write lock is not held.
>
> syzbot has tested the proposed patch and the reproducer did not
> trigger any issue.
>
> Reported-by: syzbot+35a4414f6e247f515443@xxxxxxxxxxxxxxxxxxxxxxxxx
> Closes: https://syzkaller.appspot.com/bug?extid=35a4414f6e247f515443
> Tested-by: syzbot+35a4414f6e247f515443@xxxxxxxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Pei Li <peili.dev@xxxxxxxxx>
> ---
> Syzbot reported the following warning in follow_pte():
>
> WARNING: CPU: 3 PID: 5192 at include/linux/rwsem.h:195 rwsem_assert_held include/linux/rwsem.h:195 [inline]
> WARNING: CPU: 3 PID: 5192 at include/linux/rwsem.h:195 mmap_assert_locked include/linux/mmap_lock.h:65 [inline]
> WARNING: CPU: 3 PID: 5192 at include/linux/rwsem.h:195 follow_pte+0x414/0x4c0 mm/memory.c:5980
>
> This is because we are assuming that mm->mmap_lock should be held when
> entering follow_pte(). This is added in commit c5541ba378e3 (mm:
> follow_pte() improvements).
>
> However, in the following call stack, we are not acquring the lock:
>   follow_phys arch/x86/mm/pat/memtype.c:957 [inline]
>   get_pat_info+0xf2/0x510 arch/x86/mm/pat/memtype.c:991
>   untrack_pfn+0xf7/0x4d0 arch/x86/mm/pat/memtype.c:1104
>   unmap_single_vma+0x1bd/0x2b0 mm/memory.c:1819
>   zap_page_range_single+0x326/0x560 mm/memory.c:1920

That implies that unmap_vmas() is called without the mmap lock in read 
mode, correct?

Do we know how this happens?

* exit_mmap() holds the mmap lock in read mode
* unmap_region is documented to hold the mmap lock in read mode

> In zap_page_range_single(), we passed mm_wr_locked as false, as we do
> not expect write lock to be held.
> In the special case where vma->vm_flags is set as VM_PFNMAP, we are
> hitting untrack_pfn() which eventually calls into follow_phys.
>
> This patch fixes this warning by acquiring read lock before entering
> untrack_pfn() while write lock is not held.
>
> syzbot has tested the proposed patch and the reproducer did not trigger any issue:
>
> Tested on:
>
> commit:         9d9a2f29 Merge tag 'mm-hotfixes-stable-2024-07-10-13-1..
> git tree:       upstream
> console output: https://syzkaller.appspot.com/x/log.txt?x=13be8021980000
> kernel config:  https://syzkaller.appspot.com/x/.config?x=3456bae478301dc8
> dashboard link: https://syzkaller.appspot.com/bug?extid=35a4414f6e247f515443
> compiler:       gcc (Debian 12.2.0-14) 12.2.0, GNU ld (GNU Binutils for Debian) 2.40
> patch:          https://syzkaller.appspot.com/x/patch.diff?x=145e3441980000
>
> Note: testing is done by a robot and is best-effort only.
> ---
>   mm/memory.c | 9 ++++++++-
>   1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/mm/memory.c b/mm/memory.c
> index d10e616d7389..75d7959b835b 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1815,9 +1815,16 @@ static void unmap_single_vma(struct mmu_gather *tlb,
>   	if (vma->vm_file)
>   		uprobe_munmap(vma, start, end);
>   
> -	if (unlikely(vma->vm_flags & VM_PFNMAP))
> +	if (unlikely(vma->vm_flags & VM_PFNMAP)) {
> +		if (!mm_wr_locked)
> +			mmap_read_lock(vma->vm_mm);
> +
>   		untrack_pfn(vma, 0, 0, mm_wr_locked);
>   
> +		if (!mm_wr_locked)
> +			mmap_read_unlock(vma->vm_mm);
> +	}
> +
>   	if (start != end) {
>   		if (unlikely(is_vm_hugetlb_page(vma))) {

I'm not sure if this is the right fix. I like to understand how we end 
up without the mmap lock at least in read mode in that path?

--
Cheers,

David / dhildenb