2021-02-26 01:29:33

by Andrey Konovalov

Subject: [PATCH] kasan, mm: fix crash with HW_TAGS and DEBUG_PAGEALLOC

Currently, kasan_free_nondeferred_pages()->kasan_free_pages() is called
after debug_pagealloc_unmap_pages(). This causes a crash when
debug_pagealloc is enabled, as HW_TAGS KASAN can't set tags on an
unmapped page.

This patch puts kasan_free_nondeferred_pages() before
debug_pagealloc_unmap_pages().

Besides fixing the crash, this also makes the annotation order consistent
with debug_pagealloc_map_pages() preceding kasan_alloc_pages().

Signed-off-by: Andrey Konovalov <[email protected]>
---
mm/page_alloc.c | 8 ++++++--
1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index c89e7b107514..54bc237fd319 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1311,10 +1311,14 @@ static __always_inline bool free_pages_prepare(struct page *page,
*/
arch_free_page(page, order);

- debug_pagealloc_unmap_pages(page, 1 << order);
-
+ /*
+ * With hardware tag-based KASAN, memory tags must be set
+ * before unmapping the page with debug_pagealloc.
+ */
kasan_free_nondeferred_pages(page, order, fpi_flags);

+ debug_pagealloc_unmap_pages(page, 1 << order);
+
return true;
}

--
2.30.1.766.gb4fecdf3b7-goog

2021-03-04 23:31:20

by Andrew Morton

Subject: Re: [PATCH] kasan, mm: fix crash with HW_TAGS and DEBUG_PAGEALLOC

On Fri, 26 Feb 2021 02:25:37 +0100 Andrey Konovalov <[email protected]> wrote:

> Currently, kasan_free_nondeferred_pages()->kasan_free_pages() is called
> after debug_pagealloc_unmap_pages(). This causes a crash when
> debug_pagealloc is enabled, as HW_TAGS KASAN can't set tags on an
> unmapped page.
>
> This patch puts kasan_free_nondeferred_pages() before
> debug_pagealloc_unmap_pages().
>
> Besides fixing the crash, this also makes the annotation order consistent
> with debug_pagealloc_map_pages() preceding kasan_alloc_pages().
>

This bug exists in 5.12, does it not?

If so, is cc:stable appropriate and if so, do we have a suitable Fixes:
commit?

Thanks.

2021-03-05 15:47:21

by Andrey Konovalov

Subject: Re: [PATCH] kasan, mm: fix crash with HW_TAGS and DEBUG_PAGEALLOC

On Thu, Mar 4, 2021 at 12:23 AM Andrew Morton <[email protected]> wrote:
>
> On Fri, 26 Feb 2021 02:25:37 +0100 Andrey Konovalov <[email protected]> wrote:
>
> > Currently, kasan_free_nondeferred_pages()->kasan_free_pages() is called
> > after debug_pagealloc_unmap_pages(). This causes a crash when
> > debug_pagealloc is enabled, as HW_TAGS KASAN can't set tags on an
> > unmapped page.
> >
> > This patch puts kasan_free_nondeferred_pages() before
> > debug_pagealloc_unmap_pages().
> >
> > Besides fixing the crash, this also makes the annotation order consistent
> > with debug_pagealloc_map_pages() preceding kasan_alloc_pages().
> >
>
> This bug exists in 5.12, does it not?
>
> If so, is cc:stable appropriate and if so, do we have a suitable Fixes:
> commit?

Sure:

Fixes: 94ab5b61ee16 ("kasan, arm64: enable CONFIG_KASAN_HW_TAGS")
Cc: <[email protected]>

2021-03-05 17:06:55

by Andrey Konovalov

Subject: Re: [PATCH] kasan, mm: fix crash with HW_TAGS and DEBUG_PAGEALLOC

On Fri, Feb 26, 2021 at 2:25 AM Andrey Konovalov <[email protected]> wrote:
>
> Currently, kasan_free_nondeferred_pages()->kasan_free_pages() is called
> after debug_pagealloc_unmap_pages(). This causes a crash when
> debug_pagealloc is enabled, as HW_TAGS KASAN can't set tags on an
> unmapped page.
>
> This patch puts kasan_free_nondeferred_pages() before
> debug_pagealloc_unmap_pages().
>
> Besides fixing the crash, this also makes the annotation order consistent
> with debug_pagealloc_map_pages() preceding kasan_alloc_pages().
>
> Signed-off-by: Andrey Konovalov <[email protected]>
> ---
> mm/page_alloc.c | 8 ++++++--
> 1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index c89e7b107514..54bc237fd319 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1311,10 +1311,14 @@ static __always_inline bool free_pages_prepare(struct page *page,
> */
> arch_free_page(page, order);
>
> - debug_pagealloc_unmap_pages(page, 1 << order);
> -
> + /*
> + * With hardware tag-based KASAN, memory tags must be set
> + * before unmapping the page with debug_pagealloc.
> + */
> kasan_free_nondeferred_pages(page, order, fpi_flags);

Looking at this again, I think we need to move kasan_() callback above
arch_free_page(), as that can also make the page unavailable. I'll
send v2.