[RFC/PATCH 0/3] dwarves print: Cacheline accounting fixes

[RFC/PATCH 0/3] dwarves print: Cacheline accounting fixes

[Jiri Olsa]

hi,
hit 2 issues when used pahole
 - sometimes it did not account the hole between variables
 - when using -E to extract inner struct, cacheline boundaries were wrong

with this patchset I was able to get it work on structs
that I was investigating, but I'm not sure I covered all
parts or fixed it in proper way.

thanks,
jirka


---
Jiri Olsa (3):
      dwarves print: Fix holes accounting
      dwarves print: Pass into struct conf_fprintf class__fprintf_cacheline_boundary
      dwarves print: Keep global cacheline in struct conf_fprintf

 dwarves.h         |  1 +
 dwarves_fprintf.c | 45 ++++++++++++++++++++++++++-------------------
 2 files changed, 27 insertions(+), 19 deletions(-)
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 1/3] dwarves print: Fix holes accounting

[Jiri Olsa]

Sometimes the hole could be missing, try to bypass
this issue by comparing last and current offsets.

Signed-off-by: Jiri Olsa <jolsa-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 dwarves_fprintf.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/dwarves_fprintf.c b/dwarves_fprintf.c
index 71d8ce9f982b..df80af255c67 100644
--- a/dwarves_fprintf.c
+++ b/dwarves_fprintf.c
@@ -1268,6 +1268,17 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
 		}
 		pos = tag__class_member(tag_pos);
 
+		/*
+		 * Sometimes the hole could be missing, try to bypass
+		 * this issue by comparing last and current offsets.
+		 */
+		if (last) {
+			uint32_t tmp = last->byte_offset + last->byte_size + last->hole;
+
+			if (pos->byte_offset > tmp)
+				sum_holes += pos->byte_offset - tmp;
+		}
+
 		if (last != NULL &&
 		    pos->byte_offset != last->byte_offset &&
 		    !cconf.suppress_comments)
-- 
2.4.3

--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/3] dwarves print: Fix holes accounting

[Arnaldo Carvalho de Melo]

Em Tue, Dec 08, 2015 at 11:47:04PM +0100, Jiri Olsa escreveu:
> Sometimes the hole could be missing, try to bypass
> this issue by comparing last and current offsets.

Can you provide an example of before and after this change? What real
case triggered this?

- Arnaldo
 
> Signed-off-by: Jiri Olsa <jolsa-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> ---
>  dwarves_fprintf.c | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/dwarves_fprintf.c b/dwarves_fprintf.c
> index 71d8ce9f982b..df80af255c67 100644
> --- a/dwarves_fprintf.c
> +++ b/dwarves_fprintf.c
> @@ -1268,6 +1268,17 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
>  		}
>  		pos = tag__class_member(tag_pos);
>  
> +		/*
> +		 * Sometimes the hole could be missing, try to bypass
> +		 * this issue by comparing last and current offsets.
> +		 */
> +		if (last) {
> +			uint32_t tmp = last->byte_offset + last->byte_size + last->hole;
> +
> +			if (pos->byte_offset > tmp)
> +				sum_holes += pos->byte_offset - tmp;
> +		}
> +
>  		if (last != NULL &&
>  		    pos->byte_offset != last->byte_offset &&
>  		    !cconf.suppress_comments)
> -- 
> 2.4.3
> 
> --
> To unsubscribe from this list: send the line "unsubscribe dwarves" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/3] dwarves print: Fix holes accounting

[Arnaldo Carvalho de Melo]

Em Wed, Dec 16, 2015 at 02:58:39PM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Tue, Dec 08, 2015 at 11:47:04PM +0100, Jiri Olsa escreveu:
> > Sometimes the hole could be missing, try to bypass
> > this issue by comparing last and current offsets.
> 
> Can you provide an example of before and after this change? What real
> case triggered this?
> 
> - Arnaldo
>  
> > Signed-off-by: Jiri Olsa <jolsa-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> > ---
> >  dwarves_fprintf.c | 11 +++++++++++
> >  1 file changed, 11 insertions(+)
> > 
> > diff --git a/dwarves_fprintf.c b/dwarves_fprintf.c
> > index 71d8ce9f982b..df80af255c67 100644
> > --- a/dwarves_fprintf.c
> > +++ b/dwarves_fprintf.c
> > @@ -1268,6 +1268,17 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
> >  		}
> >  		pos = tag__class_member(tag_pos);
> >  
> > +		/*
> > +		 * Sometimes the hole could be missing, try to bypass
> > +		 * this issue by comparing last and current offsets.
> > +		 */
> > +		if (last) {
> > +			uint32_t tmp = last->byte_offset + last->byte_size + last->hole;
> > +
> > +			if (pos->byte_offset > tmp)
> > +				sum_holes += pos->byte_offset - tmp;
> > +		}
> > +

I mean, how could this be? last->hole then would have been wrong,
miscalculated, looking at where it is calculated now...

- arnaldo

> >  		if (last != NULL &&
> >  		    pos->byte_offset != last->byte_offset &&
> >  		    !cconf.suppress_comments)
> > -- 
> > 2.4.3
> > 
> > --
> > To unsubscribe from this list: send the line "unsubscribe dwarves" in
> > the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/3] dwarves print: Fix holes accounting

[Arnaldo Carvalho de Melo]

Em Wed, Dec 16, 2015 at 03:05:32PM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Wed, Dec 16, 2015 at 02:58:39PM -0300, Arnaldo Carvalho de Melo escreveu:
> > Em Tue, Dec 08, 2015 at 11:47:04PM +0100, Jiri Olsa escreveu:
> > > Sometimes the hole could be missing, try to bypass
> > > this issue by comparing last and current offsets.

> > Can you provide an example of before and after this change? What real
> > case triggered this?

<SNIP>

> > > +++ b/dwarves_fprintf.c
> > > @@ -1268,6 +1268,17 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
> > >  		}
> > >  		pos = tag__class_member(tag_pos);
> > >  
> > > +		/*
> > > +		 * Sometimes the hole could be missing, try to bypass
> > > +		 * this issue by comparing last and current offsets.
> > > +		 */
> > > +		if (last) {
> > > +			uint32_t tmp = last->byte_offset + last->byte_size + last->hole;
> > > +
> > > +			if (pos->byte_offset > tmp)
> > > +				sum_holes += pos->byte_offset - tmp;
> > > +		}
 
> I mean, how could this be? last->hole then would have been wrong,
> miscalculated, looking at where it is calculated now...

Yeah, class__find_holes() should have properly calculated last->hole so
that (last->byte_offset + last->byte_size + last->hole) ==
pos->byte_offset.

- Arnaldo
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 2/3] dwarves print: Pass into struct conf_fprintf class__fprintf_cacheline_boundary

[Jiri Olsa]

Passing into struct conf_fprintf class__fprintf_cacheline_boundary,
so the next patch is more easy to read.

Signed-off-by: Jiri Olsa <jolsa-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 dwarves_fprintf.c | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/dwarves_fprintf.c b/dwarves_fprintf.c
index df80af255c67..2d114210831a 100644
--- a/dwarves_fprintf.c
+++ b/dwarves_fprintf.c
@@ -1122,7 +1122,8 @@ static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
 						size_t sum, size_t sum_holes,
 						uint8_t *newline,
 						uint32_t *cacheline,
-						int indent, FILE *fp)
+						struct conf_fprintf *cconf,
+						FILE *fp)
 {
 	const size_t real_sum = sum + sum_holes;
 	size_t printed = 0;
@@ -1139,7 +1140,7 @@ static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
 			++printed;
 		}
 
-		printed += fprintf(fp, "%.*s", indent, tabs);
+		printed += fprintf(fp, "%.*s", cconf->indent, tabs);
 
 		if (cacheline_pos == 0)
 			printed += fprintf(fp, "/* --- cacheline %u boundary "
@@ -1287,7 +1288,7 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
 							      sum, sum_holes,
 							      &newline,
 							      &last_cacheline,
-							      cconf.indent,
+							      &cconf,
 							      fp);
 		/*
 		 * These paranoid checks doesn't make much sense on
@@ -1477,7 +1478,7 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
 							     sum, sum_holes,
 							     &newline,
 							     &last_cacheline,
-							     cconf.indent, fp);
+							     &cconf, fp);
 	if (!cconf.show_only_data_members)
 		class__vtable_fprintf(class, cu, &cconf, fp);
 
-- 
2.4.3

--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2/3] dwarves print: Pass into struct conf_fprintf class__fprintf_cacheline_boundary

[Arnaldo Carvalho de Melo]

Em Tue, Dec 08, 2015 at 11:47:05PM +0100, Jiri Olsa escreveu:
> Passing into struct conf_fprintf class__fprintf_cacheline_boundary,
> so the next patch is more easy to read.

"Passing struct conf_fprintf to class__fprintf_cacheline_boundary(), so
the next patch is more easy to read."

?

> Signed-off-by: Jiri Olsa <jolsa-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> ---
>  dwarves_fprintf.c | 9 +++++----
>  1 file changed, 5 insertions(+), 4 deletions(-)
> 
> diff --git a/dwarves_fprintf.c b/dwarves_fprintf.c
> index df80af255c67..2d114210831a 100644
> --- a/dwarves_fprintf.c
> +++ b/dwarves_fprintf.c
> @@ -1122,7 +1122,8 @@ static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
>  						size_t sum, size_t sum_holes,
>  						uint8_t *newline,
>  						uint32_t *cacheline,
> -						int indent, FILE *fp)
> +						struct conf_fprintf *cconf,
> +						FILE *fp)
>  {
>  	const size_t real_sum = sum + sum_holes;
>  	size_t printed = 0;
> @@ -1139,7 +1140,7 @@ static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
>  			++printed;
>  		}
>  
> -		printed += fprintf(fp, "%.*s", indent, tabs);
> +		printed += fprintf(fp, "%.*s", cconf->indent, tabs);
>  
>  		if (cacheline_pos == 0)
>  			printed += fprintf(fp, "/* --- cacheline %u boundary "
> @@ -1287,7 +1288,7 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
>  							      sum, sum_holes,
>  							      &newline,
>  							      &last_cacheline,
> -							      cconf.indent,
> +							      &cconf,
>  							      fp);
>  		/*
>  		 * These paranoid checks doesn't make much sense on
> @@ -1477,7 +1478,7 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
>  							     sum, sum_holes,
>  							     &newline,
>  							     &last_cacheline,
> -							     cconf.indent, fp);
> +							     &cconf, fp);
>  	if (!cconf.show_only_data_members)
>  		class__vtable_fprintf(class, cu, &cconf, fp);
>  
> -- 
> 2.4.3
> 
> --
> To unsubscribe from this list: send the line "unsubscribe dwarves" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2/3] dwarves print: Pass into struct conf_fprintf class__fprintf_cacheline_boundary

[Jiri Olsa]

On Wed, Dec 16, 2015 at 03:00:17PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Tue, Dec 08, 2015 at 11:47:05PM +0100, Jiri Olsa escreveu:
> > Passing into struct conf_fprintf class__fprintf_cacheline_boundary,
> > so the next patch is more easy to read.
> 
> "Passing struct conf_fprintf to class__fprintf_cacheline_boundary(), so
> the next patch is more easy to read."
> 
> ?

exactly ;-)

jirka
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 3/3] dwarves print: Keep global cacheline in struct conf_fprintf

[Jiri Olsa]

Together with adding cconf->base_offset to sums used
for cacheline computation, this ensures proper cacheline
number to be printed for nested struct. Note possitions
of 'cacheline' lines in following outputs.

Before:
  $ pahole -C task_struct -E ~/kernel/linux-perf-local/vmlinux
  struct task_struct {
        volatile long int          state;                                                /*     0     8 */

        --- First cacheline is ok

        struct task_struct *       last_wakee;                                           /*    56     8 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        int                        wake_cpu;                                             /*    64     4 */

        ...

        const struct sched_class  * sched_class;                                         /*    88     8 */
        struct sched_entity {
                struct load_weight {
                        long unsigned int weight;                                        /*    96     8 */
                        /* typedef u32 */ unsigned int inv_weight;                       /*   104     4 */
                } load; /*    96    16 */

                --- Second one is still in relative mode and gives wrong
                    alignment in wrt task_struct start

                unsigned int       on_rq;                                                /*   152     4 */
                /* --- cacheline 1 boundary (64 bytes) --- */
                /* typedef u64 */ long long unsigned int exec_start;                     /*   160     8 */
                /* typedef u64 */ long long unsigned int sum_exec_runtime;               /*   168     8 */
       ...

After:
  $ pahole -C task_struct -E ~/kernel/linux-perf-local/vmlinux

   struct task_struct {
        volatile long int          state;                                                /*     0     8 */

        --- First cacheline is ok

        struct task_struct *       last_wakee;                                           /*    56     8 */
        /* --- cacheline 1 boundary (64 bytes) --- */
        int                        wake_cpu;                                             /*    64     4 */

        ...

        const struct sched_class  * sched_class;                                         /*    88     8 */
        struct sched_entity {
                struct load_weight {
                        long unsigned int weight;                                        /*    96     8 */
                        /* typedef u32 */ unsigned int inv_weight;                       /*   104     4 */
                } load; /*    96    16 */
                struct rb_node {
                        long unsigned int __rb_parent_color;                             /*   112     8 */
                        struct rb_node * rb_right;                                       /*   120     8 */
                        /* --- cacheline 2 boundary (128 bytes) --- */
                        struct rb_node * rb_left;                                        /*   128     8 */
                } run_node; /*   112    24 */

                --- Second one is still in relative mode and gives wrong
                    alignment in wrt task_struct start

                /* --- cacheline 2 boundary (128 bytes) was 8 bytes ago --- */
                struct list_head {
                        struct list_head * next;                                         /*   136     8 */
                        struct list_head * prev;                                         /*   144     8 */
                } group_node; /*   136    16 */
       ...

Signed-off-by: Jiri Olsa <jolsa-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
---
 dwarves.h         |  1 +
 dwarves_fprintf.c | 25 ++++++++++---------------
 2 files changed, 11 insertions(+), 15 deletions(-)

diff --git a/dwarves.h b/dwarves.h
index 73fec8ac614a..44ebf8384b9c 100644
--- a/dwarves.h
+++ b/dwarves.h
@@ -56,6 +56,7 @@ struct conf_fprintf {
 	int32_t	   type_spacing;
 	int32_t	   name_spacing;
 	uint32_t   base_offset;
+	uint32_t   base_cacheline;
 	uint8_t	   indent;
 	uint8_t	   expand_types:1;
 	uint8_t	   expand_pointers:1;
diff --git a/dwarves_fprintf.c b/dwarves_fprintf.c
index 2d114210831a..72d8b08bf235 100644
--- a/dwarves_fprintf.c
+++ b/dwarves_fprintf.c
@@ -1118,22 +1118,21 @@ size_t function__fprintf_stats(const struct tag *tag, const struct cu *cu,
 	return printed + fprintf(fp, " */\n");
 }
 
-static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
-						size_t sum, size_t sum_holes,
+static size_t class__fprintf_cacheline_boundary(size_t sum, size_t sum_holes,
 						uint8_t *newline,
-						uint32_t *cacheline,
 						struct conf_fprintf *cconf,
 						FILE *fp)
 {
-	const size_t real_sum = sum + sum_holes;
+	const size_t real_sum = sum + sum_holes + cconf->base_offset;
 	size_t printed = 0;
+	uint32_t cacheline = real_sum / cacheline_size;
 
-	*cacheline = real_sum / cacheline_size;
-
-	if (*cacheline > last_cacheline) {
+	if (cacheline != cconf->base_cacheline) {
 		const uint32_t cacheline_pos = real_sum % cacheline_size;
 		const uint32_t cacheline_in_bytes = real_sum - cacheline_pos;
 
+		cconf->base_cacheline = cacheline;
+
 		if (*newline) {
 			fputc('\n', fp);
 			*newline = 0;
@@ -1144,12 +1143,12 @@ static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
 
 		if (cacheline_pos == 0)
 			printed += fprintf(fp, "/* --- cacheline %u boundary "
-					   "(%u bytes) --- */\n", *cacheline,
+					   "(%u bytes) --- */\n", cacheline,
 					   cacheline_in_bytes);
 		else
 			printed += fprintf(fp, "/* --- cacheline %u boundary "
 					   "(%u bytes) was %u bytes ago --- "
-					   "*/\n", *cacheline,
+					   "*/\n", cacheline,
 					   cacheline_in_bytes, cacheline_pos);
 	}
 	return printed;
@@ -1284,10 +1283,8 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
 		    pos->byte_offset != last->byte_offset &&
 		    !cconf.suppress_comments)
 			printed +=
-			    class__fprintf_cacheline_boundary(last_cacheline,
-							      sum, sum_holes,
+			    class__fprintf_cacheline_boundary(sum, sum_holes,
 							      &newline,
-							      &last_cacheline,
 							      &cconf,
 							      fp);
 		/*
@@ -1474,10 +1471,8 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
 	}
 
 	if (!cconf.suppress_comments)
-		printed += class__fprintf_cacheline_boundary(last_cacheline,
-							     sum, sum_holes,
+		printed += class__fprintf_cacheline_boundary(sum, sum_holes,
 							     &newline,
-							     &last_cacheline,
 							     &cconf, fp);
 	if (!cconf.show_only_data_members)
 		class__vtable_fprintf(class, cu, &cconf, fp);
-- 
2.4.3

--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 3/3] dwarves print: Keep global cacheline in struct conf_fprintf

[Arnaldo Carvalho de Melo]

Em Tue, Dec 08, 2015 at 11:47:06PM +0100, Jiri Olsa escreveu:
> Together with adding cconf->base_offset to sums used
> for cacheline computation, this ensures proper cacheline
> number to be printed for nested struct. Note possitions
> of 'cacheline' lines in following outputs.
> 
> Before:
>   $ pahole -C task_struct -E ~/kernel/linux-perf-local/vmlinux
>   struct task_struct {
>         volatile long int          state;                                                /*     0     8 */
> 
>         --- First cacheline is ok
> 
>         struct task_struct *       last_wakee;                                           /*    56     8 */
>         /* --- cacheline 1 boundary (64 bytes) --- */
>         int                        wake_cpu;                                             /*    64     4 */
> 
>         ...
> 
>         const struct sched_class  * sched_class;                                         /*    88     8 */
>         struct sched_entity {
>                 struct load_weight {
>                         long unsigned int weight;                                        /*    96     8 */
>                         /* typedef u32 */ unsigned int inv_weight;                       /*   104     4 */
>                 } load; /*    96    16 */
> 
>                 --- Second one is still in relative mode and gives wrong
>                     alignment in wrt task_struct start
> 
>                 unsigned int       on_rq;                                                /*   152     4 */
>                 /* --- cacheline 1 boundary (64 bytes) --- */

Ok, this is a bug...

>                 /* typedef u64 */ long long unsigned int exec_start;                     /*   160     8 */
>                 /* typedef u64 */ long long unsigned int sum_exec_runtime;               /*   168     8 */
>        ...
> 
> After:
>   $ pahole -C task_struct -E ~/kernel/linux-perf-local/vmlinux
> 
>    struct task_struct {
>         volatile long int          state;                                                /*     0     8 */
> 
>         --- First cacheline is ok
> 
>         struct task_struct *       last_wakee;                                           /*    56     8 */
>         /* --- cacheline 1 boundary (64 bytes) --- */
>         int                        wake_cpu;                                             /*    64     4 */
> 
>         ...
> 
>         const struct sched_class  * sched_class;                                         /*    88     8 */
>         struct sched_entity {
>                 struct load_weight {
>                         long unsigned int weight;                                        /*    96     8 */
>                         /* typedef u32 */ unsigned int inv_weight;                       /*   104     4 */
>                 } load; /*    96    16 */
>                 struct rb_node {
>                         long unsigned int __rb_parent_color;                             /*   112     8 */
>                         struct rb_node * rb_right;                                       /*   120     8 */
>                         /* --- cacheline 2 boundary (128 bytes) --- */

Which seems fixed here, but...

>                         struct rb_node * rb_left;                                        /*   128     8 */
>                 } run_node; /*   112    24 */
> 
>                 --- Second one is still in relative mode and gives wrong
>                     alignment in wrt task_struct start

... why this one is buggy still? I.e. it should be fixed in such a way
that when we go recursively expanding structs, we know what is the
current cacheline we're on.

- Arnaldo

> 
>                 /* --- cacheline 2 boundary (128 bytes) was 8 bytes ago --- */
>                 struct list_head {
>                         struct list_head * next;                                         /*   136     8 */
>                         struct list_head * prev;                                         /*   144     8 */
>                 } group_node; /*   136    16 */
>        ...
> 
> Signed-off-by: Jiri Olsa <jolsa-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
> ---
>  dwarves.h         |  1 +
>  dwarves_fprintf.c | 25 ++++++++++---------------
>  2 files changed, 11 insertions(+), 15 deletions(-)
> 
> diff --git a/dwarves.h b/dwarves.h
> index 73fec8ac614a..44ebf8384b9c 100644
> --- a/dwarves.h
> +++ b/dwarves.h
> @@ -56,6 +56,7 @@ struct conf_fprintf {
>  	int32_t	   type_spacing;
>  	int32_t	   name_spacing;
>  	uint32_t   base_offset;
> +	uint32_t   base_cacheline;
>  	uint8_t	   indent;
>  	uint8_t	   expand_types:1;
>  	uint8_t	   expand_pointers:1;
> diff --git a/dwarves_fprintf.c b/dwarves_fprintf.c
> index 2d114210831a..72d8b08bf235 100644
> --- a/dwarves_fprintf.c
> +++ b/dwarves_fprintf.c
> @@ -1118,22 +1118,21 @@ size_t function__fprintf_stats(const struct tag *tag, const struct cu *cu,
>  	return printed + fprintf(fp, " */\n");
>  }
>  
> -static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
> -						size_t sum, size_t sum_holes,
> +static size_t class__fprintf_cacheline_boundary(size_t sum, size_t sum_holes,
>  						uint8_t *newline,
> -						uint32_t *cacheline,
>  						struct conf_fprintf *cconf,
>  						FILE *fp)
>  {
> -	const size_t real_sum = sum + sum_holes;
> +	const size_t real_sum = sum + sum_holes + cconf->base_offset;
>  	size_t printed = 0;
> +	uint32_t cacheline = real_sum / cacheline_size;
>  
> -	*cacheline = real_sum / cacheline_size;
> -
> -	if (*cacheline > last_cacheline) {
> +	if (cacheline != cconf->base_cacheline) {
>  		const uint32_t cacheline_pos = real_sum % cacheline_size;
>  		const uint32_t cacheline_in_bytes = real_sum - cacheline_pos;
>  
> +		cconf->base_cacheline = cacheline;
> +
>  		if (*newline) {
>  			fputc('\n', fp);
>  			*newline = 0;
> @@ -1144,12 +1143,12 @@ static size_t class__fprintf_cacheline_boundary(uint32_t last_cacheline,
>  
>  		if (cacheline_pos == 0)
>  			printed += fprintf(fp, "/* --- cacheline %u boundary "
> -					   "(%u bytes) --- */\n", *cacheline,
> +					   "(%u bytes) --- */\n", cacheline,
>  					   cacheline_in_bytes);
>  		else
>  			printed += fprintf(fp, "/* --- cacheline %u boundary "
>  					   "(%u bytes) was %u bytes ago --- "
> -					   "*/\n", *cacheline,
> +					   "*/\n", cacheline,
>  					   cacheline_in_bytes, cacheline_pos);
>  	}
>  	return printed;
> @@ -1284,10 +1283,8 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
>  		    pos->byte_offset != last->byte_offset &&
>  		    !cconf.suppress_comments)
>  			printed +=
> -			    class__fprintf_cacheline_boundary(last_cacheline,
> -							      sum, sum_holes,
> +			    class__fprintf_cacheline_boundary(sum, sum_holes,
>  							      &newline,
> -							      &last_cacheline,
>  							      &cconf,
>  							      fp);
>  		/*
> @@ -1474,10 +1471,8 @@ size_t class__fprintf(struct class *class, const struct cu *cu,
>  	}
>  
>  	if (!cconf.suppress_comments)
> -		printed += class__fprintf_cacheline_boundary(last_cacheline,
> -							     sum, sum_holes,
> +		printed += class__fprintf_cacheline_boundary(sum, sum_holes,
>  							     &newline,
> -							     &last_cacheline,
>  							     &cconf, fp);
>  	if (!cconf.show_only_data_members)
>  		class__vtable_fprintf(class, cu, &cconf, fp);
> -- 
> 2.4.3
> 
> --
> To unsubscribe from this list: send the line "unsubscribe dwarves" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [RFC/PATCH 0/3] dwarves print: Cacheline accounting fixes

[Jiri Olsa]

On Tue, Dec 08, 2015 at 11:47:03PM +0100, Jiri Olsa wrote:
> hi,
> hit 2 issues when used pahole
>  - sometimes it did not account the hole between variables
>  - when using -E to extract inner struct, cacheline boundaries were wrong
> 
> with this patchset I was able to get it work on structs
> that I was investigating, but I'm not sure I covered all
> parts or fixed it in proper way.

any feedback?

thanks,
jirka
--
To unsubscribe from this list: send the line "unsubscribe dwarves" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html