mm: fix vrealloc()'s KASAN poisoning logic

[PATCH mm/stable] mm: fix vrealloc()'s KASAN poisoning logic

Posted by Andrii Nakryiko 1 year, 2 months ago

When vrealloc() reuses already allocated vmap_area, we need to
re-annotate poisoned and unpoisoned portions of underlying memory
according to the new size.

Note, hard-coding KASAN_VMALLOC_PROT_NORMAL might not be exactly
correct, but KASAN flag logic is pretty involved and spread out
throughout __vmalloc_node_range_noprof(), so I'm using the bare minimum
flag here and leaving the rest to mm people to refactor this logic and
reuse it here.

Fixes: 3ddc2fefe6f3 ("mm: vmalloc: implement vrealloc()")
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
---
 mm/vmalloc.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/vmalloc.c b/mm/vmalloc.c
index 7ed39d104201..f009b21705c1 100644
--- a/mm/vmalloc.c
+++ b/mm/vmalloc.c
@@ -4093,7 +4093,8 @@ void *vrealloc_noprof(const void *p, size_t size, gfp_t flags)
 		/* Zero out spare memory. */
 		if (want_init_on_alloc(flags))
 			memset((void *)p + size, 0, old_size - size);
-
+		kasan_poison_vmalloc(p + size, old_size - size);
+		kasan_unpoison_vmalloc(p, size, KASAN_VMALLOC_PROT_NORMAL);
 		return (void *)p;
 	}
 
-- 
2.43.5

Re: [PATCH mm/stable] mm: fix vrealloc()'s KASAN poisoning logic

Posted by Andrew Morton 1 year, 2 months ago

On Mon, 25 Nov 2024 16:52:06 -0800 Andrii Nakryiko <andrii@kernel.org> wrote:

> When vrealloc() reuses already allocated vmap_area, we need to
> re-annotate poisoned and unpoisoned portions of underlying memory
> according to the new size.

What are the consequences of this oversight?

When fixing a flaw, please always remember to describe the visible
effects of that flaw.

> Note, hard-coding KASAN_VMALLOC_PROT_NORMAL might not be exactly
> correct, but KASAN flag logic is pretty involved and spread out
> throughout __vmalloc_node_range_noprof(), so I'm using the bare minimum
> flag here and leaving the rest to mm people to refactor this logic and
> reuse it here.
> 
> Fixes: 3ddc2fefe6f3 ("mm: vmalloc: implement vrealloc()")

Because a cc:stable might be appropriate here.  But without knowing the
effects, it's hard to determine this.

> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -4093,7 +4093,8 @@ void *vrealloc_noprof(const void *p, size_t size, gfp_t flags)
>  		/* Zero out spare memory. */
>  		if (want_init_on_alloc(flags))
>  			memset((void *)p + size, 0, old_size - size);
> -
> +		kasan_poison_vmalloc(p + size, old_size - size);
> +		kasan_unpoison_vmalloc(p, size, KASAN_VMALLOC_PROT_NORMAL);
>  		return (void *)p;
>  	}
>

Re: [PATCH mm/stable] mm: fix vrealloc()'s KASAN poisoning logic

Posted by Andrii Nakryiko 1 year, 2 months ago

On Wed, Nov 27, 2024 at 4:58 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>
> On Mon, 25 Nov 2024 16:52:06 -0800 Andrii Nakryiko <andrii@kernel.org> wrote:
>
> > When vrealloc() reuses already allocated vmap_area, we need to
> > re-annotate poisoned and unpoisoned portions of underlying memory
> > according to the new size.
>
> What are the consequences of this oversight?
>
> When fixing a flaw, please always remember to describe the visible
> effects of that flaw.
>

See [0] for false KASAN splat. I should have left a link to that, sorry.

  [0] https://lore.kernel.org/bpf/67450f9b.050a0220.21d33d.0004.GAE@google.com/

> > Note, hard-coding KASAN_VMALLOC_PROT_NORMAL might not be exactly
> > correct, but KASAN flag logic is pretty involved and spread out
> > throughout __vmalloc_node_range_noprof(), so I'm using the bare minimum
> > flag here and leaving the rest to mm people to refactor this logic and
> > reuse it here.
> >
> > Fixes: 3ddc2fefe6f3 ("mm: vmalloc: implement vrealloc()")
>
> Because a cc:stable might be appropriate here.  But without knowing the
> effects, it's hard to determine this.

This is KASAN-related, so the effect is a KASAN mis-reporting issue
where there is none.

>
> > --- a/mm/vmalloc.c
> > +++ b/mm/vmalloc.c
> > @@ -4093,7 +4093,8 @@ void *vrealloc_noprof(const void *p, size_t size, gfp_t flags)
> >               /* Zero out spare memory. */
> >               if (want_init_on_alloc(flags))
> >                       memset((void *)p + size, 0, old_size - size);
> > -
> > +             kasan_poison_vmalloc(p + size, old_size - size);
> > +             kasan_unpoison_vmalloc(p, size, KASAN_VMALLOC_PROT_NORMAL);
> >               return (void *)p;
> >       }
> >
>

Re: [PATCH mm/stable] mm: fix vrealloc()'s KASAN poisoning logic

Posted by Alexei Starovoitov 1 year, 2 months ago

Andrew,

What is the status of this urgent fix ?

vrealloc() is broken with kasan atm.

On Wed, Nov 27, 2024 at 10:16 PM Andrii Nakryiko
<andrii.nakryiko@gmail.com> wrote:
>
> On Wed, Nov 27, 2024 at 4:58 PM Andrew Morton <akpm@linux-foundation.org> wrote:
> >
> > On Mon, 25 Nov 2024 16:52:06 -0800 Andrii Nakryiko <andrii@kernel.org> wrote:
> >
> > > When vrealloc() reuses already allocated vmap_area, we need to
> > > re-annotate poisoned and unpoisoned portions of underlying memory
> > > according to the new size.
> >
> > What are the consequences of this oversight?
> >
> > When fixing a flaw, please always remember to describe the visible
> > effects of that flaw.
> >
>
> See [0] for false KASAN splat. I should have left a link to that, sorry.
>
>   [0] https://lore.kernel.org/bpf/67450f9b.050a0220.21d33d.0004.GAE@google.com/
>
> > > Note, hard-coding KASAN_VMALLOC_PROT_NORMAL might not be exactly
> > > correct, but KASAN flag logic is pretty involved and spread out
> > > throughout __vmalloc_node_range_noprof(), so I'm using the bare minimum
> > > flag here and leaving the rest to mm people to refactor this logic and
> > > reuse it here.
> > >
> > > Fixes: 3ddc2fefe6f3 ("mm: vmalloc: implement vrealloc()")
> >
> > Because a cc:stable might be appropriate here.  But without knowing the
> > effects, it's hard to determine this.
>
> This is KASAN-related, so the effect is a KASAN mis-reporting issue
> where there is none.
>
> >
> > > --- a/mm/vmalloc.c
> > > +++ b/mm/vmalloc.c
> > > @@ -4093,7 +4093,8 @@ void *vrealloc_noprof(const void *p, size_t size, gfp_t flags)
> > >               /* Zero out spare memory. */
> > >               if (want_init_on_alloc(flags))
> > >                       memset((void *)p + size, 0, old_size - size);
> > > -
> > > +             kasan_poison_vmalloc(p + size, old_size - size);
> > > +             kasan_unpoison_vmalloc(p, size, KASAN_VMALLOC_PROT_NORMAL);
> > >               return (void *)p;
> > >       }
> > >
> >

Re: [PATCH mm/stable] mm: fix vrealloc()'s KASAN poisoning logic

Posted by Andrew Morton 1 year, 2 months ago

On Wed, 4 Dec 2024 09:01:06 -0800 Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:

> Andrew,
> 
> What is the status of this urgent fix ?
> 

In mm-hotfixes for an upstream maerge later this week.

Re: [PATCH mm/stable] mm: fix vrealloc()'s KASAN poisoning logic

Posted by Alexei Starovoitov 1 year, 2 months ago

On Wed, Dec 4, 2024 at 1:50 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>
> On Wed, 4 Dec 2024 09:01:06 -0800 Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
>
> > Andrew,
> >
> > What is the status of this urgent fix ?
> >
>
> In mm-hotfixes for an upstream maerge later this week.

Awesome. Thanks!