Re: [syzbot] WARNING in kvm_mmu_notifier_invalidate_range_start (2)

Re: [syzbot] WARNING in kvm_mmu_notifier_invalidate_range_start (2)

[Paolo Bonzini]

On 3/21/22 11:25, syzbot wrote:
> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=12a2d0a9700000
> C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=13d34fd9700000
> 
> The issue was bisected to:
> 
> commit ed922739c9199bf515a3e7fec3e319ce1edeef2a
> Author: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
> Date:   Mon Dec 6 19:54:28 2021 +0000
> 
>      KVM: Use interval tree to do fast hva lookup in memslots
> 
> bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=142aa59d700000
> final oops:     https://syzkaller.appspot.com/x/report.txt?x=162aa59d700000
> console output: https://syzkaller.appspot.com/x/log.txt?x=122aa59d700000

It bisects here just because the patch introduces the warning; the issue 
is a mmu_notifier_invalidate_range_start with an empty range.  The 
offending system call

mremap(&(0x7f000000d000/0x2000)=nil, 0xfffffffffffffe74, 0x1000, 0x3, 
&(0x7f0000007000/0x1000)=nil)

really means old_len == 0 (it's page-aligned at the beginning of 
sys_mremap), and flags includes MREMAP_FIXED so it goes down to 
mremap_to and from there to move_page_tables.  No function on this path 
attempts to special case old_len == 0, the immediate fix would be

diff --git a/mm/mremap.c b/mm/mremap.c
index 002eec83e91e..0e175aef536e 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -486,6 +486,9 @@ unsigned long move_page_tables(struct vm_area_struct
  	pmd_t *old_pmd, *new_pmd;
  	pud_t *old_pud, *new_pud;

+	if (!len)
+		return 0;
+
  	old_end = old_addr + len;
  	flush_cache_range(vma, old_addr, old_end);

but there are several other ways to fix this elsewhere in the call chain:

- check for old_len == 0 somewhere in mremap_to

- skip the call in __mmu_notifier_invalidate_range_start and 
__mmu_notifier_invalidate_range_end, if people agree not to play 
whack-a-mole with the callers of mmu_notifier_invalidate_range_*.

- remove the warning in KVM

Thanks,

Paolo

Re: [syzbot] WARNING in kvm_mmu_notifier_invalidate_range_start (2)

[Maciej S. Szmigiero]

On 21.03.2022 12:01, Paolo Bonzini wrote:
> On 3/21/22 11:25, syzbot wrote:
>> syz repro:      https://urldefense.com/v3/__https://syzkaller.appspot.com/x/repro.syz?x=12a2d0a9700000__;!!ACWV5N9M2RV99hQ!bJGc10O9acwj6GeDIyIdP0zHAuWUpAyb7E4gom6naJO0VKxLGw2oijJnPqByG7ye0Uq2ZA$ C reproducer:   https://urldefense.com/v3/__https://syzkaller.appspot.com/x/repro.c?x=13d34fd9700000__;!!ACWV5N9M2RV99hQ!bJGc10O9acwj6GeDIyIdP0zHAuWUpAyb7E4gom6naJO0VKxLGw2oijJnPqByG7xoEv26SQ$
>> The issue was bisected to:
>>
>> commit ed922739c9199bf515a3e7fec3e319ce1edeef2a
>> Author: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
>> Date:   Mon Dec 6 19:54:28 2021 +0000
>>
>>      KVM: Use interval tree to do fast hva lookup in memslots
>>
>> bisection log:  https://urldefense.com/v3/__https://syzkaller.appspot.com/x/bisect.txt?x=142aa59d700000__;!!ACWV5N9M2RV99hQ!bJGc10O9acwj6GeDIyIdP0zHAuWUpAyb7E4gom6naJO0VKxLGw2oijJnPqByG7xEhtZ-FQ$ final oops:     https://urldefense.com/v3/__https://syzkaller.appspot.com/x/report.txt?x=162aa59d700000__;!!ACWV5N9M2RV99hQ!bJGc10O9acwj6GeDIyIdP0zHAuWUpAyb7E4gom6naJO0VKxLGw2oijJnPqByG7zcn2K3LQ$ console output: https://urldefense.com/v3/__https://syzkaller.appspot.com/x/log.txt?x=122aa59d700000__;!!ACWV5N9M2RV99hQ!bJGc10O9acwj6GeDIyIdP0zHAuWUpAyb7E4gom6naJO0VKxLGw2oijJnPqByG7wzducgVQ$ 
> 
> It bisects here just because the patch introduces the warning; the issue is a mmu_notifier_invalidate_range_start with an empty range.  The offending system call
> 
> mremap(&(0x7f000000d000/0x2000)=nil, 0xfffffffffffffe74, 0x1000, 0x3, &(0x7f0000007000/0x1000)=nil)
> 
> really means old_len == 0 (it's page-aligned at the beginning of sys_mremap), and flags includes MREMAP_FIXED so it goes down to mremap_to and from there to move_page_tables.  No function on this path attempts to special case old_len == 0, the immediate fix would be
> 
> diff --git a/mm/mremap.c b/mm/mremap.c
> index 002eec83e91e..0e175aef536e 100644
> --- a/mm/mremap.c
> +++ b/mm/mremap.c
> @@ -486,6 +486,9 @@ unsigned long move_page_tables(struct vm_area_struct
>       pmd_t *old_pmd, *new_pmd;
>       pud_t *old_pud, *new_pud;
> 
> +    if (!len)
> +        return 0;
> +
>       old_end = old_addr + len;
>       flush_cache_range(vma, old_addr, old_end);
> 
> but there are several other ways to fix this elsewhere in the call chain:
> 
> - check for old_len == 0 somewhere in mremap_to
> 
> - skip the call in __mmu_notifier_invalidate_range_start and __mmu_notifier_invalidate_range_end, if people agree not to play whack-a-mole with the callers of mmu_notifier_invalidate_range_*.
> 
> - remove the warning in KVM

This probably depends whether it is actually legal to call MMU notifiers
with a zero range, the first time this warning triggered it was the caller
that was fixed [1].

By the way, the warning-on-zero-range was added during memslots patch set
review process [2], but I think it ultimately does make sense.

> Thanks,
> 
> Paolo
> 

Thanks,
Maciej

[1]: https://lore.kernel.org/kvm/20211228234257.1926057-1-seanjc@google.com/
[2]: https://lore.kernel.org/kvm/YKWaFwgMNSaQQuQP@google.com/

Re: [syzbot] WARNING in kvm_mmu_notifier_invalidate_range_start (2)

[Sean Christopherson]

On Mon, Mar 21, 2022, Maciej S. Szmigiero wrote:
> On 21.03.2022 12:01, Paolo Bonzini wrote:
> > On 3/21/22 11:25, syzbot wrote:
> > diff --git a/mm/mremap.c b/mm/mremap.c
> > index 002eec83e91e..0e175aef536e 100644
> > --- a/mm/mremap.c
> > +++ b/mm/mremap.c
> > @@ -486,6 +486,9 @@ unsigned long move_page_tables(struct vm_area_struct
> >       pmd_t *old_pmd, *new_pmd;
> >       pud_t *old_pud, *new_pud;
> > 
> > +    if (!len)
> > +        return 0;
> > +
> >       old_end = old_addr + len;
> >       flush_cache_range(vma, old_addr, old_end);
> > 
> > but there are several other ways to fix this elsewhere in the call chain:
> > 
> > - check for old_len == 0 somewhere in mremap_to
> > 
> > - skip the call in __mmu_notifier_invalidate_range_start and
> >   __mmu_notifier_invalidate_range_end, if people agree not to play
> >   whack-a-mole with the callers of mmu_notifier_invalidate_range_*.
> > 
> > - remove the warning in KVM
> 
> This probably depends whether it is actually legal to call MMU notifiers
> with a zero range, the first time this warning triggered it was the caller
> that was fixed [1].
> 
> By the way, the warning-on-zero-range was added during memslots patch set
> review process [2], but I think it ultimately does make sense.

My vote is to play whack-a-mole.  This particular flavor isn't all that interesting,
but the HugeTLB bug was a genuine off-by-one error.  Given the low (so far) number
of unique reports, IMO the benefits of detecting buggy callers outweighs the cost of
having to fix/address benign paths where userspace is doing something silly.