4 #include "util/evsel.h"
6 #include "util/cache.h"
7 #include "util/symbol.h"
8 #include "util/thread.h"
9 #include "util/header.h"
10 #include "util/session.h"
11 #include "util/tool.h"
13 #include "util/parse-options.h"
14 #include "util/trace-event.h"
16 #include "util/debug.h"
18 #include <linux/rbtree.h>
21 typedef int (*sort_fn_t
)(struct alloc_stat
*, struct alloc_stat
*);
23 static const char *input_name
;
25 static int alloc_flag
;
26 static int caller_flag
;
28 static int alloc_lines
= -1;
29 static int caller_lines
= -1;
33 static char default_sort_order
[] = "frag,hit,bytes";
35 static int *cpunode_map
;
36 static int max_cpu_num
;
51 static struct rb_root root_alloc_stat
;
52 static struct rb_root root_alloc_sorted
;
53 static struct rb_root root_caller_stat
;
54 static struct rb_root root_caller_sorted
;
56 static unsigned long total_requested
, total_allocated
;
57 static unsigned long nr_allocs
, nr_cross_allocs
;
59 #define PATH_SYS_NODE "/sys/devices/system/node"
61 static void init_cpunode_map(void)
66 fp
= fopen("/sys/devices/system/cpu/kernel_max", "r");
72 if (fscanf(fp
, "%d", &max_cpu_num
) < 1)
73 die("Failed to read 'kernel_max' from sysfs");
76 cpunode_map
= calloc(max_cpu_num
, sizeof(int));
79 for (i
= 0; i
< max_cpu_num
; i
++)
84 static void setup_cpunode_map(void)
86 struct dirent
*dent1
, *dent2
;
88 unsigned int cpu
, mem
;
93 dir1
= opendir(PATH_SYS_NODE
);
97 while ((dent1
= readdir(dir1
)) != NULL
) {
98 if (dent1
->d_type
!= DT_DIR
||
99 sscanf(dent1
->d_name
, "node%u", &mem
) < 1)
102 snprintf(buf
, PATH_MAX
, "%s/%s", PATH_SYS_NODE
, dent1
->d_name
);
106 while ((dent2
= readdir(dir2
)) != NULL
) {
107 if (dent2
->d_type
!= DT_LNK
||
108 sscanf(dent2
->d_name
, "cpu%u", &cpu
) < 1)
110 cpunode_map
[cpu
] = mem
;
117 static void insert_alloc_stat(unsigned long call_site
, unsigned long ptr
,
118 int bytes_req
, int bytes_alloc
, int cpu
)
120 struct rb_node
**node
= &root_alloc_stat
.rb_node
;
121 struct rb_node
*parent
= NULL
;
122 struct alloc_stat
*data
= NULL
;
126 data
= rb_entry(*node
, struct alloc_stat
, node
);
129 node
= &(*node
)->rb_right
;
130 else if (ptr
< data
->ptr
)
131 node
= &(*node
)->rb_left
;
136 if (data
&& data
->ptr
== ptr
) {
138 data
->bytes_req
+= bytes_req
;
139 data
->bytes_alloc
+= bytes_alloc
;
141 data
= malloc(sizeof(*data
));
147 data
->bytes_req
= bytes_req
;
148 data
->bytes_alloc
= bytes_alloc
;
150 rb_link_node(&data
->node
, parent
, node
);
151 rb_insert_color(&data
->node
, &root_alloc_stat
);
153 data
->call_site
= call_site
;
154 data
->alloc_cpu
= cpu
;
157 static void insert_caller_stat(unsigned long call_site
,
158 int bytes_req
, int bytes_alloc
)
160 struct rb_node
**node
= &root_caller_stat
.rb_node
;
161 struct rb_node
*parent
= NULL
;
162 struct alloc_stat
*data
= NULL
;
166 data
= rb_entry(*node
, struct alloc_stat
, node
);
168 if (call_site
> data
->call_site
)
169 node
= &(*node
)->rb_right
;
170 else if (call_site
< data
->call_site
)
171 node
= &(*node
)->rb_left
;
176 if (data
&& data
->call_site
== call_site
) {
178 data
->bytes_req
+= bytes_req
;
179 data
->bytes_alloc
+= bytes_alloc
;
181 data
= malloc(sizeof(*data
));
184 data
->call_site
= call_site
;
187 data
->bytes_req
= bytes_req
;
188 data
->bytes_alloc
= bytes_alloc
;
190 rb_link_node(&data
->node
, parent
, node
);
191 rb_insert_color(&data
->node
, &root_caller_stat
);
195 static void process_alloc_event(void *data
,
196 struct event_format
*event
,
198 u64 timestamp __used
,
199 struct thread
*thread __used
,
202 unsigned long call_site
;
208 ptr
= raw_field_value(event
, "ptr", data
);
209 call_site
= raw_field_value(event
, "call_site", data
);
210 bytes_req
= raw_field_value(event
, "bytes_req", data
);
211 bytes_alloc
= raw_field_value(event
, "bytes_alloc", data
);
213 insert_alloc_stat(call_site
, ptr
, bytes_req
, bytes_alloc
, cpu
);
214 insert_caller_stat(call_site
, bytes_req
, bytes_alloc
);
216 total_requested
+= bytes_req
;
217 total_allocated
+= bytes_alloc
;
220 node1
= cpunode_map
[cpu
];
221 node2
= raw_field_value(event
, "node", data
);
228 static int ptr_cmp(struct alloc_stat
*, struct alloc_stat
*);
229 static int callsite_cmp(struct alloc_stat
*, struct alloc_stat
*);
231 static struct alloc_stat
*search_alloc_stat(unsigned long ptr
,
232 unsigned long call_site
,
233 struct rb_root
*root
,
236 struct rb_node
*node
= root
->rb_node
;
237 struct alloc_stat key
= { .ptr
= ptr
, .call_site
= call_site
};
240 struct alloc_stat
*data
;
243 data
= rb_entry(node
, struct alloc_stat
, node
);
245 cmp
= sort_fn(&key
, data
);
247 node
= node
->rb_left
;
249 node
= node
->rb_right
;
256 static void process_free_event(void *data
,
257 struct event_format
*event
,
259 u64 timestamp __used
,
260 struct thread
*thread __used
)
263 struct alloc_stat
*s_alloc
, *s_caller
;
265 ptr
= raw_field_value(event
, "ptr", data
);
267 s_alloc
= search_alloc_stat(ptr
, 0, &root_alloc_stat
, ptr_cmp
);
271 if (cpu
!= s_alloc
->alloc_cpu
) {
274 s_caller
= search_alloc_stat(0, s_alloc
->call_site
,
275 &root_caller_stat
, callsite_cmp
);
277 s_caller
->pingpong
++;
279 s_alloc
->alloc_cpu
= -1;
282 static void process_raw_event(struct perf_evsel
*evsel
, void *data
,
283 int cpu
, u64 timestamp
, struct thread
*thread
)
285 struct event_format
*event
= evsel
->tp_format
;
287 if (!strcmp(event
->name
, "kmalloc") ||
288 !strcmp(event
->name
, "kmem_cache_alloc")) {
289 process_alloc_event(data
, event
, cpu
, timestamp
, thread
, 0);
293 if (!strcmp(event
->name
, "kmalloc_node") ||
294 !strcmp(event
->name
, "kmem_cache_alloc_node")) {
295 process_alloc_event(data
, event
, cpu
, timestamp
, thread
, 1);
299 if (!strcmp(event
->name
, "kfree") ||
300 !strcmp(event
->name
, "kmem_cache_free")) {
301 process_free_event(data
, event
, cpu
, timestamp
, thread
);
306 static int process_sample_event(struct perf_tool
*tool __used
,
307 union perf_event
*event
,
308 struct perf_sample
*sample
,
309 struct perf_evsel
*evsel
,
310 struct machine
*machine
)
312 struct thread
*thread
= machine__findnew_thread(machine
, event
->ip
.pid
);
314 if (thread
== NULL
) {
315 pr_debug("problem processing %d event, skipping it.\n",
320 dump_printf(" ... thread: %s:%d\n", thread
->comm
, thread
->pid
);
322 process_raw_event(evsel
, sample
->raw_data
, sample
->cpu
,
323 sample
->time
, thread
);
328 static struct perf_tool perf_kmem
= {
329 .sample
= process_sample_event
,
330 .comm
= perf_event__process_comm
,
331 .ordered_samples
= true,
334 static double fragmentation(unsigned long n_req
, unsigned long n_alloc
)
339 return 100.0 - (100.0 * n_req
/ n_alloc
);
342 static void __print_result(struct rb_root
*root
, struct perf_session
*session
,
343 int n_lines
, int is_caller
)
345 struct rb_node
*next
;
346 struct machine
*machine
;
348 printf("%.102s\n", graph_dotted_line
);
349 printf(" %-34s |", is_caller
? "Callsite": "Alloc Ptr");
350 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
351 printf("%.102s\n", graph_dotted_line
);
353 next
= rb_first(root
);
355 machine
= perf_session__find_host_machine(session
);
357 pr_err("__print_result: couldn't find kernel information\n");
360 while (next
&& n_lines
--) {
361 struct alloc_stat
*data
= rb_entry(next
, struct alloc_stat
,
363 struct symbol
*sym
= NULL
;
369 addr
= data
->call_site
;
371 sym
= machine__find_kernel_function(machine
, addr
, &map
, NULL
);
376 snprintf(buf
, sizeof(buf
), "%s+%" PRIx64
"", sym
->name
,
377 addr
- map
->unmap_ip(map
, sym
->start
));
379 snprintf(buf
, sizeof(buf
), "%#" PRIx64
"", addr
);
380 printf(" %-34s |", buf
);
382 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
383 (unsigned long long)data
->bytes_alloc
,
384 (unsigned long)data
->bytes_alloc
/ data
->hit
,
385 (unsigned long long)data
->bytes_req
,
386 (unsigned long)data
->bytes_req
/ data
->hit
,
387 (unsigned long)data
->hit
,
388 (unsigned long)data
->pingpong
,
389 fragmentation(data
->bytes_req
, data
->bytes_alloc
));
391 next
= rb_next(next
);
395 printf(" ... | ... | ... | ... | ... | ... \n");
397 printf("%.102s\n", graph_dotted_line
);
400 static void print_summary(void)
402 printf("\nSUMMARY\n=======\n");
403 printf("Total bytes requested: %lu\n", total_requested
);
404 printf("Total bytes allocated: %lu\n", total_allocated
);
405 printf("Total bytes wasted on internal fragmentation: %lu\n",
406 total_allocated
- total_requested
);
407 printf("Internal fragmentation: %f%%\n",
408 fragmentation(total_requested
, total_allocated
));
409 printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs
, nr_allocs
);
412 static void print_result(struct perf_session
*session
)
415 __print_result(&root_caller_sorted
, session
, caller_lines
, 1);
417 __print_result(&root_alloc_sorted
, session
, alloc_lines
, 0);
421 struct sort_dimension
{
424 struct list_head list
;
427 static LIST_HEAD(caller_sort
);
428 static LIST_HEAD(alloc_sort
);
430 static void sort_insert(struct rb_root
*root
, struct alloc_stat
*data
,
431 struct list_head
*sort_list
)
433 struct rb_node
**new = &(root
->rb_node
);
434 struct rb_node
*parent
= NULL
;
435 struct sort_dimension
*sort
;
438 struct alloc_stat
*this;
441 this = rb_entry(*new, struct alloc_stat
, node
);
444 list_for_each_entry(sort
, sort_list
, list
) {
445 cmp
= sort
->cmp(data
, this);
451 new = &((*new)->rb_left
);
453 new = &((*new)->rb_right
);
456 rb_link_node(&data
->node
, parent
, new);
457 rb_insert_color(&data
->node
, root
);
460 static void __sort_result(struct rb_root
*root
, struct rb_root
*root_sorted
,
461 struct list_head
*sort_list
)
463 struct rb_node
*node
;
464 struct alloc_stat
*data
;
467 node
= rb_first(root
);
471 rb_erase(node
, root
);
472 data
= rb_entry(node
, struct alloc_stat
, node
);
473 sort_insert(root_sorted
, data
, sort_list
);
477 static void sort_result(void)
479 __sort_result(&root_alloc_stat
, &root_alloc_sorted
, &alloc_sort
);
480 __sort_result(&root_caller_stat
, &root_caller_sorted
, &caller_sort
);
483 static int __cmd_kmem(void)
486 struct perf_session
*session
;
488 session
= perf_session__new(input_name
, O_RDONLY
, 0, false, &perf_kmem
);
492 if (perf_session__create_kernel_maps(session
) < 0)
495 if (!perf_session__has_traces(session
, "kmem record"))
499 err
= perf_session__process_events(session
, &perf_kmem
);
503 print_result(session
);
505 perf_session__delete(session
);
509 static const char * const kmem_usage
[] = {
510 "perf kmem [<options>] {record|stat}",
514 static int ptr_cmp(struct alloc_stat
*l
, struct alloc_stat
*r
)
518 else if (l
->ptr
> r
->ptr
)
523 static struct sort_dimension ptr_sort_dimension
= {
528 static int callsite_cmp(struct alloc_stat
*l
, struct alloc_stat
*r
)
530 if (l
->call_site
< r
->call_site
)
532 else if (l
->call_site
> r
->call_site
)
537 static struct sort_dimension callsite_sort_dimension
= {
542 static int hit_cmp(struct alloc_stat
*l
, struct alloc_stat
*r
)
546 else if (l
->hit
> r
->hit
)
551 static struct sort_dimension hit_sort_dimension
= {
556 static int bytes_cmp(struct alloc_stat
*l
, struct alloc_stat
*r
)
558 if (l
->bytes_alloc
< r
->bytes_alloc
)
560 else if (l
->bytes_alloc
> r
->bytes_alloc
)
565 static struct sort_dimension bytes_sort_dimension
= {
570 static int frag_cmp(struct alloc_stat
*l
, struct alloc_stat
*r
)
574 x
= fragmentation(l
->bytes_req
, l
->bytes_alloc
);
575 y
= fragmentation(r
->bytes_req
, r
->bytes_alloc
);
584 static struct sort_dimension frag_sort_dimension
= {
589 static int pingpong_cmp(struct alloc_stat
*l
, struct alloc_stat
*r
)
591 if (l
->pingpong
< r
->pingpong
)
593 else if (l
->pingpong
> r
->pingpong
)
598 static struct sort_dimension pingpong_sort_dimension
= {
603 static struct sort_dimension
*avail_sorts
[] = {
605 &callsite_sort_dimension
,
607 &bytes_sort_dimension
,
608 &frag_sort_dimension
,
609 &pingpong_sort_dimension
,
612 #define NUM_AVAIL_SORTS \
613 (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
615 static int sort_dimension__add(const char *tok
, struct list_head
*list
)
617 struct sort_dimension
*sort
;
620 for (i
= 0; i
< NUM_AVAIL_SORTS
; i
++) {
621 if (!strcmp(avail_sorts
[i
]->name
, tok
)) {
622 sort
= malloc(sizeof(*sort
));
625 memcpy(sort
, avail_sorts
[i
], sizeof(*sort
));
626 list_add_tail(&sort
->list
, list
);
634 static int setup_sorting(struct list_head
*sort_list
, const char *arg
)
637 char *str
= strdup(arg
);
643 tok
= strsep(&str
, ",");
646 if (sort_dimension__add(tok
, sort_list
) < 0) {
647 error("Unknown --sort key: '%s'", tok
);
657 static int parse_sort_opt(const struct option
*opt __used
,
658 const char *arg
, int unset __used
)
663 if (caller_flag
> alloc_flag
)
664 return setup_sorting(&caller_sort
, arg
);
666 return setup_sorting(&alloc_sort
, arg
);
671 static int parse_caller_opt(const struct option
*opt __used
,
672 const char *arg __used
, int unset __used
)
674 caller_flag
= (alloc_flag
+ 1);
678 static int parse_alloc_opt(const struct option
*opt __used
,
679 const char *arg __used
, int unset __used
)
681 alloc_flag
= (caller_flag
+ 1);
685 static int parse_line_opt(const struct option
*opt __used
,
686 const char *arg
, int unset __used
)
693 lines
= strtoul(arg
, NULL
, 10);
695 if (caller_flag
> alloc_flag
)
696 caller_lines
= lines
;
703 static const struct option kmem_options
[] = {
704 OPT_STRING('i', "input", &input_name
, "file",
706 OPT_CALLBACK_NOOPT(0, "caller", NULL
, NULL
,
707 "show per-callsite statistics",
709 OPT_CALLBACK_NOOPT(0, "alloc", NULL
, NULL
,
710 "show per-allocation statistics",
712 OPT_CALLBACK('s', "sort", NULL
, "key[,key2...]",
713 "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
715 OPT_CALLBACK('l', "line", NULL
, "num",
718 OPT_BOOLEAN(0, "raw-ip", &raw_ip
, "show raw ip instead of symbol"),
722 static const char *record_args
[] = {
728 "-e", "kmem:kmalloc",
729 "-e", "kmem:kmalloc_node",
731 "-e", "kmem:kmem_cache_alloc",
732 "-e", "kmem:kmem_cache_alloc_node",
733 "-e", "kmem:kmem_cache_free",
736 static int __cmd_record(int argc
, const char **argv
)
738 unsigned int rec_argc
, i
, j
;
739 const char **rec_argv
;
741 rec_argc
= ARRAY_SIZE(record_args
) + argc
- 1;
742 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
744 if (rec_argv
== NULL
)
747 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
748 rec_argv
[i
] = strdup(record_args
[i
]);
750 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
751 rec_argv
[i
] = argv
[j
];
753 return cmd_record(i
, rec_argv
, NULL
);
756 int cmd_kmem(int argc
, const char **argv
, const char *prefix __used
)
758 argc
= parse_options(argc
, argv
, kmem_options
, kmem_usage
, 0);
761 usage_with_options(kmem_usage
, kmem_options
);
765 if (!strncmp(argv
[0], "rec", 3)) {
766 return __cmd_record(argc
, argv
);
767 } else if (!strcmp(argv
[0], "stat")) {
770 if (list_empty(&caller_sort
))
771 setup_sorting(&caller_sort
, default_sort_order
);
772 if (list_empty(&alloc_sort
))
773 setup_sorting(&alloc_sort
, default_sort_order
);
777 usage_with_options(kmem_usage
, kmem_options
);