2021-03-30 15:35:36

by Andrey Konovalov

Subject: [PATCH mm] mm, kasan: fix for "integrate page_alloc init with HW_TAGS"

My commit "integrate page_alloc init with HW_TAGS" changed the order of
kernel_unpoison_pages() and kernel_init_free_pages() calls. This leads
to __GFP_ZERO allocations being incorrectly poisoned when page poisoning
is enabled.

Fix by restoring the initial order. Also add a warning comment.

Reported-by: Vlastimil Babka <[email protected]>
Reported-by: Sergei Trofimovich <[email protected]>
Signed-off-by: Andrey Konovalov <[email protected]>
---
mm/page_alloc.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 033bd92e8398..1fc5061f8ca1 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2328,6 +2328,12 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
arch_alloc_page(page, order);
debug_pagealloc_map_pages(page, 1 << order);

+ /*
+ * Page unpoisoning must happen before memory initialization.
+ * Otherwise, a __GFP_ZERO allocation will not be initialized.
+ */
+ kernel_unpoison_pages(page, 1 << order);
+
/*
* As memory initialization might be integrated into KASAN,
* kasan_alloc_pages and kernel_init_free_pages must be
@@ -2338,7 +2344,6 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
if (init && !kasan_has_integrated_init())
kernel_init_free_pages(page, 1 << order);

- kernel_unpoison_pages(page, 1 << order);
set_page_owner(page, order, gfp_flags);
}

--
2.31.0.291.g576ba9dcdaf-goog

2021-03-30 15:58:28

by Vlastimil Babka

Subject: Re: [PATCH mm] mm, kasan: fix for "integrate page_alloc init with HW_TAGS"

On 3/30/21 5:31 PM, Andrey Konovalov wrote:
> My commit "integrate page_alloc init with HW_TAGS" changed the order of
> kernel_unpoison_pages() and kernel_init_free_pages() calls. This leads
> to __GFP_ZERO allocations being incorrectly poisoned when page poisoning
> is enabled.

Correction: This leads to check_poison_mem() complain about memory corruption
because the poison pattern has already been overwritten by zeroes.

> Fix by restoring the initial order. Also add a warning comment.
>
> Reported-by: Vlastimil Babka <[email protected]>
> Reported-by: Sergei Trofimovich <[email protected]>
> Signed-off-by: Andrey Konovalov <[email protected]>

I expect this will be folded to your patch in mmotm anyway, so the changelog is
not as important...

> ---
> mm/page_alloc.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 033bd92e8398..1fc5061f8ca1 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2328,6 +2328,12 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
> arch_alloc_page(page, order);
> debug_pagealloc_map_pages(page, 1 << order);
>
> + /*
> + * Page unpoisoning must happen before memory initialization.
> + * Otherwise, a __GFP_ZERO allocation will not be initialized.

... but the comment should be corrected too:
"Otherwise, a __GFP_ZERO allocation will trigger a memory corruption report
during unpoisoning."

Thanks.

> + */
> + kernel_unpoison_pages(page, 1 << order);
> +
> /*
> * As memory initialization might be integrated into KASAN,
> * kasan_alloc_pages and kernel_init_free_pages must be
> @@ -2338,7 +2344,6 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
> if (init && !kasan_has_integrated_init())
> kernel_init_free_pages(page, 1 << order);
>
> - kernel_unpoison_pages(page, 1 << order);
> set_page_owner(page, order, gfp_flags);
> }
>
>

2021-03-30 16:41:34

by Andrey Konovalov

Subject: Re: [PATCH mm] mm, kasan: fix for "integrate page_alloc init with HW_TAGS"

On Tue, Mar 30, 2021 at 5:54 PM Vlastimil Babka <[email protected]> wrote:
>
> On 3/30/21 5:31 PM, Andrey Konovalov wrote:
> > My commit "integrate page_alloc init with HW_TAGS" changed the order of
> > kernel_unpoison_pages() and kernel_init_free_pages() calls. This leads
> > to __GFP_ZERO allocations being incorrectly poisoned when page poisoning
> > is enabled.
>
> Correction: This leads to check_poison_mem() complain about memory corruption
> because the poison pattern has already been overwritten by zeroes.

Ah, indeed. Will send v2. Thanks!