perf symbols: fix slowness due to -ffunction-section
PATCH 1/1 perf symbols: fix slowness due to -ffunction-section
Perf can take minutes to parse an image when -ffunction-section is used. This is especially true with the kernel image when it is compiled this way, which is the arm64 default since the patcheset "Enable deadcode elimination at link time".
Perf organize maps using a rbtree. Whenever perf finds a new symbols, it first searches this rbtree for the map it belongs to, by strcmp()'aring section names. When it finds the map with the right name, it uses it to add the symbol. With a usual image there aren't so many maps but when using -ffunction-section there's basically one map per function.
With the kernel image that's north of 40,000 maps. For most symbols perf has to parses the entire rbtree to eventually create a new map and add it. Consequently perf spends most of the time browsing a rbtree that keeps getting larger.
This performance fix introduces a secondary rbtree that indexes maps based on the section name.
--- tools/perf/util/map.c | 27 +++++++++++++++++++++++++++ tools/perf/util/map.h | 2 ++ tools/perf/util/symbol.c | 15 +++++++++++++-- 3 files changed, 42 insertions(+), 2 deletions(-)
tools/perf/util/map.c +27 -0
@@ -21,6 +21,7 @@#include "unwind.h" static void __maps__insert (struct maps *maps,struct map *map); +static void __maps__insert_name (struct maps *maps,struct map *map);static inline int is_anon_memory (const char *filename,u32 flags) {
@@ -496,6 +497,7 @@ u64 map__objdump_2mem(struct map *map, u64 ip)static void maps__init (struct maps *maps) { maps->entries =RB_ROOT ; + maps->names =RB_ROOT ;init_rwsem (&maps->lock); }
@@ -664,6 +666,7 @@ size_t map_groups__fprintf(struct map_groups *mg, FILE *fp)static void __map_groups__insert (struct map_groups *mg,struct map *map) {__maps__insert (&mg->maps, map); +__maps__insert_name (&mg->maps, map); map->groups = mg; }
@@ -824,10 +827,34 @@ static void __maps__insert(struct maps *maps, struct map *map)map__get (map); } +static void __maps__insert_name (struct maps *maps,struct map *map) +{ +struct rb_node **p = &maps->names.rb_node; +struct rb_node *parent =NULL ; +struct map *m; +int rc; + +while (*p !=NULL ) { + parent = *p; + m = rb_entry(parent, struct map, rb_node_name); + rc =strcmp (m->dso->short_name, map->dso->short_name); +if (rc <0 ) + p = &(*p)->rb_left; +else if (rc >0 ) + p = &(*p)->rb_right; +else +return ; + } +rb_link_node (&map->rb_node_name, parent, p); +rb_insert_color (&map->rb_node_name, &maps->names); +map__get (map); +} +void maps__insert (struct maps *maps,struct map *map) {down_write (&maps->lock);__maps__insert (maps, map); +__maps__insert_name (maps, map);up_write (&maps->lock); }
tools/perf/util/map.h +2 -0
@@ -25,6 +25,7 @@ struct map {struct rb_node rb_node;struct list_head node; }; +struct rb_node rb_node_name;u64 start;u64 end;bool erange_warned;
@@ -57,6 +58,7 @@ struct kmap {struct maps {struct rb_root entries; +struct rb_root names;struct rw_semaphore lock; };
tools/perf/util/symbol.c +13 -2
@@ -1680,11 +1680,22 @@ struct map *map_groups__find_by_name(struct map_groups *mg, const char *name) {struct maps *maps = &mg->maps;struct map *map; +struct rb_node *node;down_read (&maps->lock); -for (map =maps__first (maps); map; map =map__next (map)) { -if (map->dso &&strcmp (map->dso->short_name, name) ==0 ) +for (node = maps->names.rb_node; node; ) { +int rc; + + map = rb_entry(node, struct map, rb_node_name); + + rc =strcmp (map->dso->short_name, name); +if (rc <0 ) + node = node->rb_left; +else if (rc >0 ) + node = node->rb_right; +else + goto out_unlock; }
Signature · 1 line
1.8.3.1
Em Wed, Nov 21, 2018 at 09:51:19AM -0800, Eric Saint-Etienne escreveu:
Quoted 15 lines
Perf can take minutes to parse an image when -ffunction-section is used. This is especially true with the kernel image when it is compiled this way, which is the arm64 default since the patcheset "Enable deadcode elimination at link time".
Perf organize maps using a rbtree. Whenever perf finds a new symbols, it first searches this rbtree for the map it belongs to, by strcmp()'aring section names. When it finds the map with the right name, it uses it to add the symbol. With a usual image there aren't so many maps but when using -ffunction-section there's basically one map per function.
With the kernel image that's north of 40,000 maps. For most symbols perf has to parses the entire rbtree to eventually create a new map and add it. Consequently perf spends most of the time browsing a rbtree that keeps getting larger.
Quoted 7 lines
This performance fix introduces a secondary rbtree that indexes maps based on the section name.
Signed-off-by Eric Saint-Etienne <[email protected]>Reviewed-by Dave Kleikamp <[email protected]>Reviewed-by David Aldridge <[email protected]>Reviewed-by Rob Gardner <[email protected]>
Looks sane, thanks to the multiple reviewers, really appreciated,
Applied.
- Arnaldo
tip-bot for Eric Saint-Etienne
merge bot[tip:perf/core] perf symbols: Fix slowness due to -ffunction-section
In reply to Eric Saint-Etienneperf symbols: Fix slowness due to -ffunction-section
Commit-ID: 1e6285699b3034e6f4d1f091edd46d717580bf7c Gitweb: https://git.kernel.org/tip/1e6285699b3034e6f4d1f091edd46d717580bf7c Author: Eric Saint-Etienne <[email protected]> AuthorDate: Wed, 21 Nov 2018 09:51:19 -0800 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Wed, 21 Nov 2018 22:39:59 -0300
perf symbols: Fix slowness due to -ffunction-section
Perf can take minutes to parse an image when -ffunction-section is used. This is especially true with the kernel image when it is compiled this way, which is the arm64 default since the patcheset "Enable deadcode elimination at link time".
Perf organize maps using a rbtree. Whenever perf finds a new symbols, it first searches this rbtree for the map it belongs to, by strcmp()'aring section names. When it finds the map with the right name, it uses it to add the symbol. With a usual image there aren't so many maps but when using -ffunction-section there's basically one map per function. With the kernel image that's north of 40,000 maps. For most symbols perf has to parses the entire rbtree to eventually create a new map and add it. Consequently perf spends most of the time browsing a rbtree that keeps getting larger.
This performance fix introduces a secondary rbtree that indexes maps based on the section name.
--- tools/perf/util/map.c | 27 +++++++++++++++++++++++++++ tools/perf/util/map.h | 2 ++ tools/perf/util/symbol.c | 15 +++++++++++++-- 3 files changed, 42 insertions(+), 2 deletions(-)
tools/perf/util/map.c +27 -0
@@ -21,6 +21,7 @@#include "unwind.h" static void __maps__insert (struct maps *maps,struct map *map); +static void __maps__insert_name (struct maps *maps,struct map *map);static inline int is_anon_memory (const char *filename,u32 flags) {
@@ -496,6 +497,7 @@ u64 map__objdump_2mem(struct map *map, u64 ip)static void maps__init (struct maps *maps) { maps->entries =RB_ROOT ; + maps->names =RB_ROOT ;init_rwsem (&maps->lock); }
@@ -664,6 +666,7 @@ size_t map_groups__fprintf(struct map_groups *mg, FILE *fp)static void __map_groups__insert (struct map_groups *mg,struct map *map) {__maps__insert (&mg->maps, map); +__maps__insert_name (&mg->maps, map); map->groups = mg; }
@@ -824,10 +827,34 @@ static void __maps__insert(struct maps *maps, struct map *map)map__get (map); } +static void __maps__insert_name (struct maps *maps,struct map *map) +{ +struct rb_node **p = &maps->names.rb_node; +struct rb_node *parent =NULL ; +struct map *m; +int rc; + +while (*p !=NULL ) { + parent = *p; + m = rb_entry(parent, struct map, rb_node_name); + rc =strcmp (m->dso->short_name, map->dso->short_name); +if (rc <0 ) + p = &(*p)->rb_left; +else if (rc >0 ) + p = &(*p)->rb_right; +else +return ; + } +rb_link_node (&map->rb_node_name, parent, p); +rb_insert_color (&map->rb_node_name, &maps->names); +map__get (map); +} +void maps__insert (struct maps *maps,struct map *map) {down_write (&maps->lock);__maps__insert (maps, map); +__maps__insert_name (maps, map);up_write (&maps->lock); }
tools/perf/util/map.h +2 -0
@@ -25,6 +25,7 @@ struct map {struct rb_node rb_node;struct list_head node; }; +struct rb_node rb_node_name;u64 start;u64 end;bool erange_warned;
@@ -57,6 +58,7 @@ struct kmap {struct maps {struct rb_root entries; +struct rb_root names;struct rw_semaphore lock; };
tools/perf/util/symbol.c +13 -2
@@ -1680,11 +1680,22 @@ struct map *map_groups__find_by_name(struct map_groups *mg, const char *name) {struct maps *maps = &mg->maps;struct map *map; +struct rb_node *node;down_read (&maps->lock); -for (map =maps__first (maps); map; map =map__next (map)) { -if (map->dso &&strcmp (map->dso->short_name, name) ==0 ) +for (node = maps->names.rb_node; node; ) { +int rc; + + map = rb_entry(node, struct map, rb_node_name); + + rc =strcmp (map->dso->short_name, name); +if (rc <0 ) + node = node->rb_left; +else if (rc >0 ) + node = node->rb_right; +else + goto out_unlock; }