diff mbox series

mm: add stackdepot information on page->private for tracking

Message ID 1665026390-16096-1-git-send-email-zhaoyang.huang@unisoc.com (mailing list archive)
State New
Headers show
Series mm: add stackdepot information on page->private for tracking | expand

Commit Message

zhaoyang.huang Oct. 6, 2022, 3:19 a.m. UTC
From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>

Private is vacant for most of Non-LRU pages while the user has explicitly
operation on page->private via set_page_private, I would like introduce
stackdepot information on page->private for a simplified tracking mechanism
which could be help for kernel driver's memory leak.

Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
---
 mm/page_alloc.c | 28 +++++++++++++++++++++++++++-
 1 file changed, 27 insertions(+), 1 deletion(-)

Comments

Vlastimil Babka Oct. 7, 2022, 10:08 a.m. UTC | #1
On 10/6/22 05:19, zhaoyang.huang wrote:
> From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> 
> Private is vacant for most of Non-LRU pages while the user has explicitly
> operation on page->private via set_page_private, I would like introduce
> stackdepot information on page->private for a simplified tracking mechanism
> which could be help for kernel driver's memory leak.
> 
> Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>

This duplicates the existing page_owner functionality in a way that
unconditionally adds overhead to all kernels that have CONFIG_STACKDEPOT
enabled build-time (and also misses the need to initialize stackdepot properly).

Also wouldn't be suprised if some existing page->private users were actually
confused by the field suddenly being non-zero without their own action.

> ---
>  mm/page_alloc.c | 28 +++++++++++++++++++++++++++-
>  1 file changed, 27 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index e5486d4..b79a503 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -75,6 +75,7 @@
>  #include <linux/khugepaged.h>
>  #include <linux/buffer_head.h>
>  #include <linux/delayacct.h>
> +#include <linux/stackdepot.h>
>  #include <asm/sections.h>
>  #include <asm/tlbflush.h>
>  #include <asm/div64.h>
> @@ -2464,6 +2465,25 @@ static inline bool should_skip_init(gfp_t flags)
>  	return (flags & __GFP_SKIP_ZERO);
>  }
>  
> +#ifdef CONFIG_STACKDEPOT
> +static noinline depot_stack_handle_t set_track_prepare(void)
> +{
> +       depot_stack_handle_t trace_handle;
> +       unsigned long entries[16];
> +       unsigned int nr_entries;
> +
> +       nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
> +       trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
> +
> +       return trace_handle;
> +}
> +#else
> +static inline depot_stack_handle_t set_track_prepare(void)
> +{
> +       return 0;
> +}
> +#endif
> +
>  inline void post_alloc_hook(struct page *page, unsigned int order,
>  				gfp_t gfp_flags)
>  {
> @@ -2471,8 +2491,14 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
>  			!should_skip_init(gfp_flags);
>  	bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
>  	int i;
> +	depot_stack_handle_t stack_handle = set_track_prepare();
>  
> -	set_page_private(page, 0);
> +	/*
> +	 * Don't worry, user will cover private directly without checking
> +	 * this field and has ability to trace the page. This also will not
> +	 * affect expected state when freeing
> +	 */
> +	set_page_private(page, stack_handle);
>  	set_page_refcounted(page);
>  
>  	arch_alloc_page(page, order);
Zhaoyang Huang Oct. 9, 2022, 2:25 a.m. UTC | #2
On Fri, Oct 7, 2022 at 6:08 PM Vlastimil Babka <vbabka@suse.cz> wrote:
>
> On 10/6/22 05:19, zhaoyang.huang wrote:
> > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
> >
> > Private is vacant for most of Non-LRU pages while the user has explicitly
> > operation on page->private via set_page_private, I would like introduce
> > stackdepot information on page->private for a simplified tracking mechanism
> > which could be help for kernel driver's memory leak.
> >
> > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
>
> This duplicates the existing page_owner functionality in a way that
> unconditionally adds overhead to all kernels that have CONFIG_STACKDEPOT
> enabled build-time (and also misses the need to initialize stackdepot properly).
Sure. This patch could be deemed as a light and complement of the page
owner which depends on proc fs in lived system for showing the result.
This patch could be mainly helpful on RAM dump as it is hard to find
page_ext for page owners. I also would like to make this optional via
defconfig item.
>
> Also wouldn't be suprised if some existing page->private users were actually
> confused by the field suddenly being non-zero without their own action.
IMO, the existing page->private users will cover this field directly
without distrubed by handle.

>
> > ---
> >  mm/page_alloc.c | 28 +++++++++++++++++++++++++++-
> >  1 file changed, 27 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index e5486d4..b79a503 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -75,6 +75,7 @@
> >  #include <linux/khugepaged.h>
> >  #include <linux/buffer_head.h>
> >  #include <linux/delayacct.h>
> > +#include <linux/stackdepot.h>
> >  #include <asm/sections.h>
> >  #include <asm/tlbflush.h>
> >  #include <asm/div64.h>
> > @@ -2464,6 +2465,25 @@ static inline bool should_skip_init(gfp_t flags)
> >       return (flags & __GFP_SKIP_ZERO);
> >  }
> >
> > +#ifdef CONFIG_STACKDEPOT
> > +static noinline depot_stack_handle_t set_track_prepare(void)
> > +{
> > +       depot_stack_handle_t trace_handle;
> > +       unsigned long entries[16];
> > +       unsigned int nr_entries;
> > +
> > +       nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
> > +       trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
> > +
> > +       return trace_handle;
> > +}
> > +#else
> > +static inline depot_stack_handle_t set_track_prepare(void)
> > +{
> > +       return 0;
> > +}
> > +#endif
> > +
> >  inline void post_alloc_hook(struct page *page, unsigned int order,
> >                               gfp_t gfp_flags)
> >  {
> > @@ -2471,8 +2491,14 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
> >                       !should_skip_init(gfp_flags);
> >       bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
> >       int i;
> > +     depot_stack_handle_t stack_handle = set_track_prepare();
> >
> > -     set_page_private(page, 0);
> > +     /*
> > +      * Don't worry, user will cover private directly without checking
> > +      * this field and has ability to trace the page. This also will not
> > +      * affect expected state when freeing
> > +      */
> > +     set_page_private(page, stack_handle);
> >       set_page_refcounted(page);
> >
> >       arch_alloc_page(page, order);
>
Vlastimil Babka Oct. 10, 2022, 8:57 a.m. UTC | #3
On 10/9/22 04:25, Zhaoyang Huang wrote:
> On Fri, Oct 7, 2022 at 6:08 PM Vlastimil Babka <vbabka@suse.cz> wrote:
>>
>> On 10/6/22 05:19, zhaoyang.huang wrote:
>> > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
>> >
>> > Private is vacant for most of Non-LRU pages while the user has explicitly
>> > operation on page->private via set_page_private, I would like introduce
>> > stackdepot information on page->private for a simplified tracking mechanism
>> > which could be help for kernel driver's memory leak.
>> >
>> > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
>>
>> This duplicates the existing page_owner functionality in a way that
>> unconditionally adds overhead to all kernels that have CONFIG_STACKDEPOT
>> enabled build-time (and also misses the need to initialize stackdepot properly).
> Sure. This patch could be deemed as a light and complement of the page
> owner which depends on proc fs in lived system for showing the result.
> This patch could be mainly helpful on RAM dump as it is hard to find
> page_ext for page owners. I also would like to make this optional via
> defconfig item.

I'm still not convinced we need this, between existing page_owner and the
proposed code tagging framework.

https://lore.kernel.org/all/20220830214919.53220-1-surenb@google.com/

For finding page_ext in crash dumps, it's possible with a scriptable
debugger like drgn or crash-python.

>>
>> Also wouldn't be suprised if some existing page->private users were actually
>> confused by the field suddenly being non-zero without their own action.
> IMO, the existing page->private users will cover this field directly
> without distrubed by handle.

Well the bot wasn't happy so far
https://lore.kernel.org/all/202210072204.cfea59d3-oliver.sang@intel.com/

>>
>> > ---
>> >  mm/page_alloc.c | 28 +++++++++++++++++++++++++++-
>> >  1 file changed, 27 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> > index e5486d4..b79a503 100644
>> > --- a/mm/page_alloc.c
>> > +++ b/mm/page_alloc.c
>> > @@ -75,6 +75,7 @@
>> >  #include <linux/khugepaged.h>
>> >  #include <linux/buffer_head.h>
>> >  #include <linux/delayacct.h>
>> > +#include <linux/stackdepot.h>
>> >  #include <asm/sections.h>
>> >  #include <asm/tlbflush.h>
>> >  #include <asm/div64.h>
>> > @@ -2464,6 +2465,25 @@ static inline bool should_skip_init(gfp_t flags)
>> >       return (flags & __GFP_SKIP_ZERO);
>> >  }
>> >
>> > +#ifdef CONFIG_STACKDEPOT
>> > +static noinline depot_stack_handle_t set_track_prepare(void)
>> > +{
>> > +       depot_stack_handle_t trace_handle;
>> > +       unsigned long entries[16];
>> > +       unsigned int nr_entries;
>> > +
>> > +       nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
>> > +       trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
>> > +
>> > +       return trace_handle;
>> > +}
>> > +#else
>> > +static inline depot_stack_handle_t set_track_prepare(void)
>> > +{
>> > +       return 0;
>> > +}
>> > +#endif
>> > +
>> >  inline void post_alloc_hook(struct page *page, unsigned int order,
>> >                               gfp_t gfp_flags)
>> >  {
>> > @@ -2471,8 +2491,14 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
>> >                       !should_skip_init(gfp_flags);
>> >       bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
>> >       int i;
>> > +     depot_stack_handle_t stack_handle = set_track_prepare();
>> >
>> > -     set_page_private(page, 0);
>> > +     /*
>> > +      * Don't worry, user will cover private directly without checking
>> > +      * this field and has ability to trace the page. This also will not
>> > +      * affect expected state when freeing
>> > +      */
>> > +     set_page_private(page, stack_handle);
>> >       set_page_refcounted(page);
>> >
>> >       arch_alloc_page(page, order);
>>
diff mbox series

Patch

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index e5486d4..b79a503 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -75,6 +75,7 @@ 
 #include <linux/khugepaged.h>
 #include <linux/buffer_head.h>
 #include <linux/delayacct.h>
+#include <linux/stackdepot.h>
 #include <asm/sections.h>
 #include <asm/tlbflush.h>
 #include <asm/div64.h>
@@ -2464,6 +2465,25 @@  static inline bool should_skip_init(gfp_t flags)
 	return (flags & __GFP_SKIP_ZERO);
 }
 
+#ifdef CONFIG_STACKDEPOT
+static noinline depot_stack_handle_t set_track_prepare(void)
+{
+       depot_stack_handle_t trace_handle;
+       unsigned long entries[16];
+       unsigned int nr_entries;
+
+       nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
+       trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
+
+       return trace_handle;
+}
+#else
+static inline depot_stack_handle_t set_track_prepare(void)
+{
+       return 0;
+}
+#endif
+
 inline void post_alloc_hook(struct page *page, unsigned int order,
 				gfp_t gfp_flags)
 {
@@ -2471,8 +2491,14 @@  inline void post_alloc_hook(struct page *page, unsigned int order,
 			!should_skip_init(gfp_flags);
 	bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
 	int i;
+	depot_stack_handle_t stack_handle = set_track_prepare();
 
-	set_page_private(page, 0);
+	/*
+	 * Don't worry, user will cover private directly without checking
+	 * this field and has ability to trace the page. This also will not
+	 * affect expected state when freeing
+	 */
+	set_page_private(page, stack_handle);
 	set_page_refcounted(page);
 
 	arch_alloc_page(page, order);