2011-06-08 19:19:12

by Eric B Munson

[permalink] [raw]
Subject: [PATCH] Add debugging boundary check to pfn_to_page

Bugzilla 36192 showed a problem where pages were being accessed outside of
a node boundary. It would be helpful in diagnosing this kind of problem to
have pfn_to_page complain when a page is accessed outside of the node boundary.
This patch adds a new debug config option which adds a WARN_ON in pfn_to_page
that will complain when pages are accessed outside of the node boundary.

Signed-of-by: Eric B Munson <[email protected]>
---
include/asm-generic/memory_model.h | 19 +++++++++++++++----
lib/Kconfig.debug | 10 ++++++++++
2 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/include/asm-generic/memory_model.h b/include/asm-generic/memory_model.h
index fb2d63f..a0f1d19 100644
--- a/include/asm-generic/memory_model.h
+++ b/include/asm-generic/memory_model.h
@@ -62,11 +62,22 @@
(unsigned long)(__pg - __section_mem_map_addr(__nr_to_section(__sec))); \
})

-#define __pfn_to_page(pfn) \
-({ unsigned long __pfn = (pfn); \
- struct mem_section *__sec = __pfn_to_section(__pfn); \
- __section_mem_map_addr(__sec) + __pfn; \
+#ifdef CONFIG_DEBUG_MEMORY_MODEL
+#define __pfn_to_page(pfn) \
+({ unsigned long __pfn = (pfn); \
+ struct mem_section *__sec = __pfn_to_section(__pfn); \
+ struct page *__page = __section_mem_map_addr(__sec) + __pfn; \
+ WARN_ON(__page->flags == 0); \
+ __page; \
})
+#else
+#define __pfn_to_page(pfn) \
+({ unsigned long __pfn = (pfn); \
+ struct mem_section *__sec = __pfn_to_section(__pfn); \
+ __section_mem_map_addr(__sec) + __pfn; \
+})
+#endif /* CONFIG_DEBUG_MEMORY_MODEL */
+
#endif /* CONFIG_FLATMEM/DISCONTIGMEM/SPARSEMEM */

#define page_to_pfn __page_to_pfn
diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
index dd373c8..d932cbf 100644
--- a/lib/Kconfig.debug
+++ b/lib/Kconfig.debug
@@ -777,6 +777,16 @@ config DEBUG_MEMORY_INIT

If unsure, say Y

+config DEBUG_MEMORY_MODEL
+ bool "Debug memory model" if SPARSEMEM || DISCONTIGMEM
+ depends on SPARSEMEM || DISCONTIGMEM
+ help
+ Enable this to check that page accesses are done within node
+ boundaries. The check will warn each time a page is requested
+ outside node boundaries.
+
+ If unsure, say N
+
config DEBUG_LIST
bool "Debug linked list manipulation"
depends on DEBUG_KERNEL
--
1.7.4.1


2011-06-08 19:31:41

by Randy Dunlap

[permalink] [raw]
Subject: Re: [PATCH] Add debugging boundary check to pfn_to_page

On 06/08/11 12:18, Eric B Munson wrote:
> Bugzilla 36192 showed a problem where pages were being accessed outside of
> a node boundary. It would be helpful in diagnosing this kind of problem to
> have pfn_to_page complain when a page is accessed outside of the node boundary.
> This patch adds a new debug config option which adds a WARN_ON in pfn_to_page
> that will complain when pages are accessed outside of the node boundary.
>
> Signed-of-by: Eric B Munson <[email protected]>
> ---
> include/asm-generic/memory_model.h | 19 +++++++++++++++----
> lib/Kconfig.debug | 10 ++++++++++
> 2 files changed, 25 insertions(+), 4 deletions(-)
>
> diff --git a/include/asm-generic/memory_model.h b/include/asm-generic/memory_model.h
> index fb2d63f..a0f1d19 100644
> --- a/include/asm-generic/memory_model.h
> +++ b/include/asm-generic/memory_model.h
> @@ -62,11 +62,22 @@
> (unsigned long)(__pg - __section_mem_map_addr(__nr_to_section(__sec))); \
> })
>
> -#define __pfn_to_page(pfn) \
> -({ unsigned long __pfn = (pfn); \
> - struct mem_section *__sec = __pfn_to_section(__pfn); \
> - __section_mem_map_addr(__sec) + __pfn; \
> +#ifdef CONFIG_DEBUG_MEMORY_MODEL
> +#define __pfn_to_page(pfn) \
> +({ unsigned long __pfn = (pfn); \
> + struct mem_section *__sec = __pfn_to_section(__pfn); \
> + struct page *__page = __section_mem_map_addr(__sec) + __pfn; \
> + WARN_ON(__page->flags == 0); \
> + __page; \
> })
> +#else
> +#define __pfn_to_page(pfn) \
> +({ unsigned long __pfn = (pfn); \
> + struct mem_section *__sec = __pfn_to_section(__pfn); \
> + __section_mem_map_addr(__sec) + __pfn; \
> +})
> +#endif /* CONFIG_DEBUG_MEMORY_MODEL */
> +
> #endif /* CONFIG_FLATMEM/DISCONTIGMEM/SPARSEMEM */
>
> #define page_to_pfn __page_to_pfn
> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
> index dd373c8..d932cbf 100644
> --- a/lib/Kconfig.debug
> +++ b/lib/Kconfig.debug
> @@ -777,6 +777,16 @@ config DEBUG_MEMORY_INIT
>
> If unsure, say Y
>
> +config DEBUG_MEMORY_MODEL
> + bool "Debug memory model" if SPARSEMEM || DISCONTIGMEM
> + depends on SPARSEMEM || DISCONTIGMEM

bool <prompt> <expr>
creates the dependency, so the "depends on" line is not needed.

> + help
> + Enable this to check that page accesses are done within node
> + boundaries. The check will warn each time a page is requested
> + outside node boundaries.
> +
> + If unsure, say N
> +
> config DEBUG_LIST
> bool "Debug linked list manipulation"
> depends on DEBUG_KERNEL


--
~Randy
*** Remember to use Documentation/SubmitChecklist when testing your code ***

2011-06-08 19:56:40

by Paul E. McKenney

[permalink] [raw]
Subject: Re: [PATCH] Add debugging boundary check to pfn_to_page

On Wed, Jun 08, 2011 at 03:18:54PM -0400, Eric B Munson wrote:
> Bugzilla 36192 showed a problem where pages were being accessed outside of
> a node boundary. It would be helpful in diagnosing this kind of problem to
> have pfn_to_page complain when a page is accessed outside of the node boundary.
> This patch adds a new debug config option which adds a WARN_ON in pfn_to_page
> that will complain when pages are accessed outside of the node boundary.
>
> Signed-of-by: Eric B Munson <[email protected]>
> ---
> include/asm-generic/memory_model.h | 19 +++++++++++++++----
> lib/Kconfig.debug | 10 ++++++++++
> 2 files changed, 25 insertions(+), 4 deletions(-)
>
> diff --git a/include/asm-generic/memory_model.h b/include/asm-generic/memory_model.h
> index fb2d63f..a0f1d19 100644
> --- a/include/asm-generic/memory_model.h
> +++ b/include/asm-generic/memory_model.h
> @@ -62,11 +62,22 @@
> (unsigned long)(__pg - __section_mem_map_addr(__nr_to_section(__sec))); \
> })
>
> -#define __pfn_to_page(pfn) \
> -({ unsigned long __pfn = (pfn); \
> - struct mem_section *__sec = __pfn_to_section(__pfn); \
> - __section_mem_map_addr(__sec) + __pfn; \
> +#ifdef CONFIG_DEBUG_MEMORY_MODEL
> +#define __pfn_to_page(pfn) \
> +({ unsigned long __pfn = (pfn); \
> + struct mem_section *__sec = __pfn_to_section(__pfn); \
> + struct page *__page = __section_mem_map_addr(__sec) + __pfn; \
> + WARN_ON(__page->flags == 0); \
> + __page; \
> })
> +#else
> +#define __pfn_to_page(pfn) \
> +({ unsigned long __pfn = (pfn); \
> + struct mem_section *__sec = __pfn_to_section(__pfn); \
> + __section_mem_map_addr(__sec) + __pfn; \
> +})
> +#endif /* CONFIG_DEBUG_MEMORY_MODEL */
> +

The following variant would avoid the duplicate code, FWIW.

#define __pfn_to_page_nodebug(pfn) \
({ unsigned long __pfn = (pfn); \
struct mem_section *__sec = __pfn_to_section(__pfn); \
__section_mem_map_addr(__sec) + __pfn; \
})
#ifdef CONFIG_DEBUG_MEMORY_MODEL
#define __pfn_to_page(pfn) \
({ \
struct page *__page = __pfn_to_page_nodebug(pfn); \
WARN_ON(__page->flags == 0); \
__page; \
})
#else
#define __pfn_to_page(pfn) __pfn_to_page_nodebug(pfn)
#endif /* CONFIG_DEBUG_MEMORY_MODEL */

Thanx, Paul

> #endif /* CONFIG_FLATMEM/DISCONTIGMEM/SPARSEMEM */
>
> #define page_to_pfn __page_to_pfn
> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
> index dd373c8..d932cbf 100644
> --- a/lib/Kconfig.debug
> +++ b/lib/Kconfig.debug
> @@ -777,6 +777,16 @@ config DEBUG_MEMORY_INIT
>
> If unsure, say Y
>
> +config DEBUG_MEMORY_MODEL
> + bool "Debug memory model" if SPARSEMEM || DISCONTIGMEM
> + depends on SPARSEMEM || DISCONTIGMEM
> + help
> + Enable this to check that page accesses are done within node
> + boundaries. The check will warn each time a page is requested
> + outside node boundaries.
> +
> + If unsure, say N
> +
> config DEBUG_LIST
> bool "Debug linked list manipulation"
> depends on DEBUG_KERNEL
> --
> 1.7.4.1
>

2011-06-08 23:53:00

by Dave Hansen

[permalink] [raw]
Subject: Re: [PATCH] Add debugging boundary check to pfn_to_page

On Wed, 2011-06-08 at 15:18 -0400, Eric B Munson wrote:
> -#define __pfn_to_page(pfn) \
> -({ unsigned long __pfn = (pfn); \
> - struct mem_section *__sec = __pfn_to_section(__pfn); \
> - __section_mem_map_addr(__sec) + __pfn; \
> +#ifdef CONFIG_DEBUG_MEMORY_MODEL
> +#define __pfn_to_page(pfn) \
> +({ unsigned long __pfn = (pfn); \
> + struct mem_section *__sec = __pfn_to_section(__pfn); \
> + struct page *__page = __section_mem_map_addr(__sec) + __pfn; \
> + WARN_ON(__page->flags == 0); \
> + __page; \

What was the scenario you're trying to catch here? If you give a really
crummy __pfn, you'll probably go off the end of one of the mem_section[]
arrays, and get garbage back for __sec. You might also get a NULL back
from __section_mem_map_addr() if the section is possibly valid, but just
not present on this particular system.

I _think_ the only kind of bug this will catch is if you have a valid
section, with a valid section_mem_map[] but still manage to find
yourself with an 'struct page' unclaimed by any zone and thus
uninitialized.

You could catch a lot more cases by being a bit more paranoid:

void check_pfn(unsigned long pfn)
{
int nid;

// hacked in from pfn_to_nid:
// Don't actually do this, add a new helper near pfn_to_nid()
// Can this even fit in the physnode_map?
if (pfn / PAGES_PER_ELEMENT > ARRAY_SIZE(physnode_map))
WARN();

// Is there a valid nid there?
nid = pfn_to_nid(pfn);
if (nid == -1)
WARN();

// check against NODE_DATA(nid)->node_start_pfn;
// check against NODE_DATA(nid)->node_spanned_pages;
}
> })
> +#else
> +#define __pfn_to_page(pfn) \
> +({ unsigned long __pfn = (pfn); \
> + struct mem_section *__sec = __pfn_to_section(__pfn); \
> + __section_mem_map_addr(__sec) + __pfn; \
> +})
> +#endif /* CONFIG_DEBUG_MEMORY_MODEL */

Instead of making a completely new __pfn_to_page() in the debugging
case, I'd probably do something like this:

#ifdef CONFIG_DEBUG_MEMORY_MODEL
#define check_foo(foo) {\
some_check_here(foo);\
WARN_ON(foo->flags);\
}
#else
#define check_foo(foo) do{}while(0)
#endif;

#define __pfn_to_page(pfn) \
({ unsigned long __pfn = (pfn); \
struct mem_section *__sec = __pfn_to_section(__pfn); \
struct page *__page = __section_mem_map_addr(__sec) + __pfn; \
check_foo(page) \
__page; \
})

That'll make sure that the two copies of __pfn_to_page() don't
accidentally diverge. It also makes it a lot easier to read, I think.

-- Dave

2011-06-10 13:27:56

by Eric B Munson

[permalink] [raw]
Subject: Re: [PATCH] Add debugging boundary check to pfn_to_page

On Wed, 08 Jun 2011, Dave Hansen wrote:

> On Wed, 2011-06-08 at 15:18 -0400, Eric B Munson wrote:
> > -#define __pfn_to_page(pfn) \
> > -({ unsigned long __pfn = (pfn); \
> > - struct mem_section *__sec = __pfn_to_section(__pfn); \
> > - __section_mem_map_addr(__sec) + __pfn; \
> > +#ifdef CONFIG_DEBUG_MEMORY_MODEL
> > +#define __pfn_to_page(pfn) \
> > +({ unsigned long __pfn = (pfn); \
> > + struct mem_section *__sec = __pfn_to_section(__pfn); \
> > + struct page *__page = __section_mem_map_addr(__sec) + __pfn; \
> > + WARN_ON(__page->flags == 0); \
> > + __page; \
>
> What was the scenario you're trying to catch here? If you give a really
> crummy __pfn, you'll probably go off the end of one of the mem_section[]
> arrays, and get garbage back for __sec. You might also get a NULL back
> from __section_mem_map_addr() if the section is possibly valid, but just
> not present on this particular system.
>
> I _think_ the only kind of bug this will catch is if you have a valid
> section, with a valid section_mem_map[] but still manage to find
> yourself with an 'struct page' unclaimed by any zone and thus
> uninitialized.

This is the case I was going after. I will rework for a V2 based on the
feedback here.

>
> You could catch a lot more cases by being a bit more paranoid:
>
> void check_pfn(unsigned long pfn)
> {
> int nid;
>
> // hacked in from pfn_to_nid:
> // Don't actually do this, add a new helper near pfn_to_nid()
> // Can this even fit in the physnode_map?
> if (pfn / PAGES_PER_ELEMENT > ARRAY_SIZE(physnode_map))
> WARN();
>
> // Is there a valid nid there?
> nid = pfn_to_nid(pfn);
> if (nid == -1)
> WARN();
>
> // check against NODE_DATA(nid)->node_start_pfn;
> // check against NODE_DATA(nid)->node_spanned_pages;
> }
> > })
> > +#else
> > +#define __pfn_to_page(pfn) \
> > +({ unsigned long __pfn = (pfn); \
> > + struct mem_section *__sec = __pfn_to_section(__pfn); \
> > + __section_mem_map_addr(__sec) + __pfn; \
> > +})
> > +#endif /* CONFIG_DEBUG_MEMORY_MODEL */
>
> Instead of making a completely new __pfn_to_page() in the debugging
> case, I'd probably do something like this:
>
> #ifdef CONFIG_DEBUG_MEMORY_MODEL
> #define check_foo(foo) {\
> some_check_here(foo);\
> WARN_ON(foo->flags);\
> }
> #else
> #define check_foo(foo) do{}while(0)
> #endif;
>
> #define __pfn_to_page(pfn) \
> ({ unsigned long __pfn = (pfn); \
> struct mem_section *__sec = __pfn_to_section(__pfn); \
> struct page *__page = __section_mem_map_addr(__sec) + __pfn; \
> check_foo(page) \
> __page; \
> })
>
> That'll make sure that the two copies of __pfn_to_page() don't
> accidentally diverge. It also makes it a lot easier to read, I think.
>
> -- Dave
>


Attachments:
(No filename) (3.22 kB)
signature.asc (490.00 B)
Digital signature
Download all attachments