Root/tools/perf/builtin-kmem.c

1#include "builtin.h"
2#include "perf.h"
3
4#include "util/util.h"
5#include "util/cache.h"
6#include "util/symbol.h"
7#include "util/thread.h"
8#include "util/header.h"
9#include "util/session.h"
10
11#include "util/parse-options.h"
12#include "util/trace-event.h"
13
14#include "util/debug.h"
15
16#include <linux/rbtree.h>
17
18struct alloc_stat;
19typedef int (*sort_fn_t)(struct alloc_stat *, struct alloc_stat *);
20
21static char const *input_name = "perf.data";
22
23static int alloc_flag;
24static int caller_flag;
25
26static int alloc_lines = -1;
27static int caller_lines = -1;
28
29static bool raw_ip;
30
31static char default_sort_order[] = "frag,hit,bytes";
32
33static int *cpunode_map;
34static int max_cpu_num;
35
36struct alloc_stat {
37    u64 call_site;
38    u64 ptr;
39    u64 bytes_req;
40    u64 bytes_alloc;
41    u32 hit;
42    u32 pingpong;
43
44    short alloc_cpu;
45
46    struct rb_node node;
47};
48
49static struct rb_root root_alloc_stat;
50static struct rb_root root_alloc_sorted;
51static struct rb_root root_caller_stat;
52static struct rb_root root_caller_sorted;
53
54static unsigned long total_requested, total_allocated;
55static unsigned long nr_allocs, nr_cross_allocs;
56
57#define PATH_SYS_NODE "/sys/devices/system/node"
58
59static void init_cpunode_map(void)
60{
61    FILE *fp;
62    int i;
63
64    fp = fopen("/sys/devices/system/cpu/kernel_max", "r");
65    if (!fp) {
66        max_cpu_num = 4096;
67        return;
68    }
69
70    if (fscanf(fp, "%d", &max_cpu_num) < 1)
71        die("Failed to read 'kernel_max' from sysfs");
72    max_cpu_num++;
73
74    cpunode_map = calloc(max_cpu_num, sizeof(int));
75    if (!cpunode_map)
76        die("calloc");
77    for (i = 0; i < max_cpu_num; i++)
78        cpunode_map[i] = -1;
79    fclose(fp);
80}
81
82static void setup_cpunode_map(void)
83{
84    struct dirent *dent1, *dent2;
85    DIR *dir1, *dir2;
86    unsigned int cpu, mem;
87    char buf[PATH_MAX];
88
89    init_cpunode_map();
90
91    dir1 = opendir(PATH_SYS_NODE);
92    if (!dir1)
93        return;
94
95    while ((dent1 = readdir(dir1)) != NULL) {
96        if (dent1->d_type != DT_DIR ||
97            sscanf(dent1->d_name, "node%u", &mem) < 1)
98            continue;
99
100        snprintf(buf, PATH_MAX, "%s/%s", PATH_SYS_NODE, dent1->d_name);
101        dir2 = opendir(buf);
102        if (!dir2)
103            continue;
104        while ((dent2 = readdir(dir2)) != NULL) {
105            if (dent2->d_type != DT_LNK ||
106                sscanf(dent2->d_name, "cpu%u", &cpu) < 1)
107                continue;
108            cpunode_map[cpu] = mem;
109        }
110    }
111}
112
113static void insert_alloc_stat(unsigned long call_site, unsigned long ptr,
114                  int bytes_req, int bytes_alloc, int cpu)
115{
116    struct rb_node **node = &root_alloc_stat.rb_node;
117    struct rb_node *parent = NULL;
118    struct alloc_stat *data = NULL;
119
120    while (*node) {
121        parent = *node;
122        data = rb_entry(*node, struct alloc_stat, node);
123
124        if (ptr > data->ptr)
125            node = &(*node)->rb_right;
126        else if (ptr < data->ptr)
127            node = &(*node)->rb_left;
128        else
129            break;
130    }
131
132    if (data && data->ptr == ptr) {
133        data->hit++;
134        data->bytes_req += bytes_req;
135        data->bytes_alloc += bytes_alloc;
136    } else {
137        data = malloc(sizeof(*data));
138        if (!data)
139            die("malloc");
140        data->ptr = ptr;
141        data->pingpong = 0;
142        data->hit = 1;
143        data->bytes_req = bytes_req;
144        data->bytes_alloc = bytes_alloc;
145
146        rb_link_node(&data->node, parent, node);
147        rb_insert_color(&data->node, &root_alloc_stat);
148    }
149    data->call_site = call_site;
150    data->alloc_cpu = cpu;
151}
152
153static void insert_caller_stat(unsigned long call_site,
154                  int bytes_req, int bytes_alloc)
155{
156    struct rb_node **node = &root_caller_stat.rb_node;
157    struct rb_node *parent = NULL;
158    struct alloc_stat *data = NULL;
159
160    while (*node) {
161        parent = *node;
162        data = rb_entry(*node, struct alloc_stat, node);
163
164        if (call_site > data->call_site)
165            node = &(*node)->rb_right;
166        else if (call_site < data->call_site)
167            node = &(*node)->rb_left;
168        else
169            break;
170    }
171
172    if (data && data->call_site == call_site) {
173        data->hit++;
174        data->bytes_req += bytes_req;
175        data->bytes_alloc += bytes_alloc;
176    } else {
177        data = malloc(sizeof(*data));
178        if (!data)
179            die("malloc");
180        data->call_site = call_site;
181        data->pingpong = 0;
182        data->hit = 1;
183        data->bytes_req = bytes_req;
184        data->bytes_alloc = bytes_alloc;
185
186        rb_link_node(&data->node, parent, node);
187        rb_insert_color(&data->node, &root_caller_stat);
188    }
189}
190
191static void process_alloc_event(void *data,
192                struct event *event,
193                int cpu,
194                u64 timestamp __used,
195                struct thread *thread __used,
196                int node)
197{
198    unsigned long call_site;
199    unsigned long ptr;
200    int bytes_req;
201    int bytes_alloc;
202    int node1, node2;
203
204    ptr = raw_field_value(event, "ptr", data);
205    call_site = raw_field_value(event, "call_site", data);
206    bytes_req = raw_field_value(event, "bytes_req", data);
207    bytes_alloc = raw_field_value(event, "bytes_alloc", data);
208
209    insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, cpu);
210    insert_caller_stat(call_site, bytes_req, bytes_alloc);
211
212    total_requested += bytes_req;
213    total_allocated += bytes_alloc;
214
215    if (node) {
216        node1 = cpunode_map[cpu];
217        node2 = raw_field_value(event, "node", data);
218        if (node1 != node2)
219            nr_cross_allocs++;
220    }
221    nr_allocs++;
222}
223
224static int ptr_cmp(struct alloc_stat *, struct alloc_stat *);
225static int callsite_cmp(struct alloc_stat *, struct alloc_stat *);
226
227static struct alloc_stat *search_alloc_stat(unsigned long ptr,
228                        unsigned long call_site,
229                        struct rb_root *root,
230                        sort_fn_t sort_fn)
231{
232    struct rb_node *node = root->rb_node;
233    struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
234
235    while (node) {
236        struct alloc_stat *data;
237        int cmp;
238
239        data = rb_entry(node, struct alloc_stat, node);
240
241        cmp = sort_fn(&key, data);
242        if (cmp < 0)
243            node = node->rb_left;
244        else if (cmp > 0)
245            node = node->rb_right;
246        else
247            return data;
248    }
249    return NULL;
250}
251
252static void process_free_event(void *data,
253                   struct event *event,
254                   int cpu,
255                   u64 timestamp __used,
256                   struct thread *thread __used)
257{
258    unsigned long ptr;
259    struct alloc_stat *s_alloc, *s_caller;
260
261    ptr = raw_field_value(event, "ptr", data);
262
263    s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
264    if (!s_alloc)
265        return;
266
267    if (cpu != s_alloc->alloc_cpu) {
268        s_alloc->pingpong++;
269
270        s_caller = search_alloc_stat(0, s_alloc->call_site,
271                         &root_caller_stat, callsite_cmp);
272        assert(s_caller);
273        s_caller->pingpong++;
274    }
275    s_alloc->alloc_cpu = -1;
276}
277
278static void
279process_raw_event(event_t *raw_event __used, void *data,
280          int cpu, u64 timestamp, struct thread *thread)
281{
282    struct event *event;
283    int type;
284
285    type = trace_parse_common_type(data);
286    event = trace_find_event(type);
287
288    if (!strcmp(event->name, "kmalloc") ||
289        !strcmp(event->name, "kmem_cache_alloc")) {
290        process_alloc_event(data, event, cpu, timestamp, thread, 0);
291        return;
292    }
293
294    if (!strcmp(event->name, "kmalloc_node") ||
295        !strcmp(event->name, "kmem_cache_alloc_node")) {
296        process_alloc_event(data, event, cpu, timestamp, thread, 1);
297        return;
298    }
299
300    if (!strcmp(event->name, "kfree") ||
301        !strcmp(event->name, "kmem_cache_free")) {
302        process_free_event(data, event, cpu, timestamp, thread);
303        return;
304    }
305}
306
307static int process_sample_event(event_t *event, struct perf_session *session)
308{
309    struct sample_data data;
310    struct thread *thread;
311
312    memset(&data, 0, sizeof(data));
313    data.time = -1;
314    data.cpu = -1;
315    data.period = 1;
316
317    event__parse_sample(event, session->sample_type, &data);
318
319    dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
320            data.pid, data.tid, data.ip, data.period);
321
322    thread = perf_session__findnew(session, event->ip.pid);
323    if (thread == NULL) {
324        pr_debug("problem processing %d event, skipping it.\n",
325             event->header.type);
326        return -1;
327    }
328
329    dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
330
331    process_raw_event(event, data.raw_data, data.cpu,
332              data.time, thread);
333
334    return 0;
335}
336
337static struct perf_event_ops event_ops = {
338    .sample = process_sample_event,
339    .comm = event__process_comm,
340};
341
342static double fragmentation(unsigned long n_req, unsigned long n_alloc)
343{
344    if (n_alloc == 0)
345        return 0.0;
346    else
347        return 100.0 - (100.0 * n_req / n_alloc);
348}
349
350static void __print_result(struct rb_root *root, struct perf_session *session,
351               int n_lines, int is_caller)
352{
353    struct rb_node *next;
354
355    printf("%.102s\n", graph_dotted_line);
356    printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
357    printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
358    printf("%.102s\n", graph_dotted_line);
359
360    next = rb_first(root);
361
362    while (next && n_lines--) {
363        struct alloc_stat *data = rb_entry(next, struct alloc_stat,
364                           node);
365        struct symbol *sym = NULL;
366        char buf[BUFSIZ];
367        u64 addr;
368
369        if (is_caller) {
370            addr = data->call_site;
371            if (!raw_ip)
372                sym = map_groups__find_function(&session->kmaps, addr, NULL);
373        } else
374            addr = data->ptr;
375
376        if (sym != NULL)
377            snprintf(buf, sizeof(buf), "%s+%Lx", sym->name,
378                 addr - sym->start);
379        else
380            snprintf(buf, sizeof(buf), "%#Lx", addr);
381        printf(" %-34s |", buf);
382
383        printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %8lu | %6.3f%%\n",
384               (unsigned long long)data->bytes_alloc,
385               (unsigned long)data->bytes_alloc / data->hit,
386               (unsigned long long)data->bytes_req,
387               (unsigned long)data->bytes_req / data->hit,
388               (unsigned long)data->hit,
389               (unsigned long)data->pingpong,
390               fragmentation(data->bytes_req, data->bytes_alloc));
391
392        next = rb_next(next);
393    }
394
395    if (n_lines == -1)
396        printf(" ... | ... | ... | ... | ... | ... \n");
397
398    printf("%.102s\n", graph_dotted_line);
399}
400
401static void print_summary(void)
402{
403    printf("\nSUMMARY\n=======\n");
404    printf("Total bytes requested: %lu\n", total_requested);
405    printf("Total bytes allocated: %lu\n", total_allocated);
406    printf("Total bytes wasted on internal fragmentation: %lu\n",
407           total_allocated - total_requested);
408    printf("Internal fragmentation: %f%%\n",
409           fragmentation(total_requested, total_allocated));
410    printf("Cross CPU allocations: %lu/%lu\n", nr_cross_allocs, nr_allocs);
411}
412
413static void print_result(struct perf_session *session)
414{
415    if (caller_flag)
416        __print_result(&root_caller_sorted, session, caller_lines, 1);
417    if (alloc_flag)
418        __print_result(&root_alloc_sorted, session, alloc_lines, 0);
419    print_summary();
420}
421
422struct sort_dimension {
423    const char name[20];
424    sort_fn_t cmp;
425    struct list_head list;
426};
427
428static LIST_HEAD(caller_sort);
429static LIST_HEAD(alloc_sort);
430
431static void sort_insert(struct rb_root *root, struct alloc_stat *data,
432            struct list_head *sort_list)
433{
434    struct rb_node **new = &(root->rb_node);
435    struct rb_node *parent = NULL;
436    struct sort_dimension *sort;
437
438    while (*new) {
439        struct alloc_stat *this;
440        int cmp = 0;
441
442        this = rb_entry(*new, struct alloc_stat, node);
443        parent = *new;
444
445        list_for_each_entry(sort, sort_list, list) {
446            cmp = sort->cmp(data, this);
447            if (cmp)
448                break;
449        }
450
451        if (cmp > 0)
452            new = &((*new)->rb_left);
453        else
454            new = &((*new)->rb_right);
455    }
456
457    rb_link_node(&data->node, parent, new);
458    rb_insert_color(&data->node, root);
459}
460
461static void __sort_result(struct rb_root *root, struct rb_root *root_sorted,
462              struct list_head *sort_list)
463{
464    struct rb_node *node;
465    struct alloc_stat *data;
466
467    for (;;) {
468        node = rb_first(root);
469        if (!node)
470            break;
471
472        rb_erase(node, root);
473        data = rb_entry(node, struct alloc_stat, node);
474        sort_insert(root_sorted, data, sort_list);
475    }
476}
477
478static void sort_result(void)
479{
480    __sort_result(&root_alloc_stat, &root_alloc_sorted, &alloc_sort);
481    __sort_result(&root_caller_stat, &root_caller_sorted, &caller_sort);
482}
483
484static int __cmd_kmem(void)
485{
486    int err = -EINVAL;
487    struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
488    if (session == NULL)
489        return -ENOMEM;
490
491    if (!perf_session__has_traces(session, "kmem record"))
492        goto out_delete;
493
494    setup_pager();
495    err = perf_session__process_events(session, &event_ops);
496    if (err != 0)
497        goto out_delete;
498    sort_result();
499    print_result(session);
500out_delete:
501    perf_session__delete(session);
502    return err;
503}
504
505static const char * const kmem_usage[] = {
506    "perf kmem [<options>] {record|stat}",
507    NULL
508};
509
510static int ptr_cmp(struct alloc_stat *l, struct alloc_stat *r)
511{
512    if (l->ptr < r->ptr)
513        return -1;
514    else if (l->ptr > r->ptr)
515        return 1;
516    return 0;
517}
518
519static struct sort_dimension ptr_sort_dimension = {
520    .name = "ptr",
521    .cmp = ptr_cmp,
522};
523
524static int callsite_cmp(struct alloc_stat *l, struct alloc_stat *r)
525{
526    if (l->call_site < r->call_site)
527        return -1;
528    else if (l->call_site > r->call_site)
529        return 1;
530    return 0;
531}
532
533static struct sort_dimension callsite_sort_dimension = {
534    .name = "callsite",
535    .cmp = callsite_cmp,
536};
537
538static int hit_cmp(struct alloc_stat *l, struct alloc_stat *r)
539{
540    if (l->hit < r->hit)
541        return -1;
542    else if (l->hit > r->hit)
543        return 1;
544    return 0;
545}
546
547static struct sort_dimension hit_sort_dimension = {
548    .name = "hit",
549    .cmp = hit_cmp,
550};
551
552static int bytes_cmp(struct alloc_stat *l, struct alloc_stat *r)
553{
554    if (l->bytes_alloc < r->bytes_alloc)
555        return -1;
556    else if (l->bytes_alloc > r->bytes_alloc)
557        return 1;
558    return 0;
559}
560
561static struct sort_dimension bytes_sort_dimension = {
562    .name = "bytes",
563    .cmp = bytes_cmp,
564};
565
566static int frag_cmp(struct alloc_stat *l, struct alloc_stat *r)
567{
568    double x, y;
569
570    x = fragmentation(l->bytes_req, l->bytes_alloc);
571    y = fragmentation(r->bytes_req, r->bytes_alloc);
572
573    if (x < y)
574        return -1;
575    else if (x > y)
576        return 1;
577    return 0;
578}
579
580static struct sort_dimension frag_sort_dimension = {
581    .name = "frag",
582    .cmp = frag_cmp,
583};
584
585static int pingpong_cmp(struct alloc_stat *l, struct alloc_stat *r)
586{
587    if (l->pingpong < r->pingpong)
588        return -1;
589    else if (l->pingpong > r->pingpong)
590        return 1;
591    return 0;
592}
593
594static struct sort_dimension pingpong_sort_dimension = {
595    .name = "pingpong",
596    .cmp = pingpong_cmp,
597};
598
599static struct sort_dimension *avail_sorts[] = {
600    &ptr_sort_dimension,
601    &callsite_sort_dimension,
602    &hit_sort_dimension,
603    &bytes_sort_dimension,
604    &frag_sort_dimension,
605    &pingpong_sort_dimension,
606};
607
608#define NUM_AVAIL_SORTS \
609    (int)(sizeof(avail_sorts) / sizeof(struct sort_dimension *))
610
611static int sort_dimension__add(const char *tok, struct list_head *list)
612{
613    struct sort_dimension *sort;
614    int i;
615
616    for (i = 0; i < NUM_AVAIL_SORTS; i++) {
617        if (!strcmp(avail_sorts[i]->name, tok)) {
618            sort = malloc(sizeof(*sort));
619            if (!sort)
620                die("malloc");
621            memcpy(sort, avail_sorts[i], sizeof(*sort));
622            list_add_tail(&sort->list, list);
623            return 0;
624        }
625    }
626
627    return -1;
628}
629
630static int setup_sorting(struct list_head *sort_list, const char *arg)
631{
632    char *tok;
633    char *str = strdup(arg);
634
635    if (!str)
636        die("strdup");
637
638    while (true) {
639        tok = strsep(&str, ",");
640        if (!tok)
641            break;
642        if (sort_dimension__add(tok, sort_list) < 0) {
643            error("Unknown --sort key: '%s'", tok);
644            return -1;
645        }
646    }
647
648    free(str);
649    return 0;
650}
651
652static int parse_sort_opt(const struct option *opt __used,
653              const char *arg, int unset __used)
654{
655    if (!arg)
656        return -1;
657
658    if (caller_flag > alloc_flag)
659        return setup_sorting(&caller_sort, arg);
660    else
661        return setup_sorting(&alloc_sort, arg);
662
663    return 0;
664}
665
666static int parse_caller_opt(const struct option *opt __used,
667              const char *arg __used, int unset __used)
668{
669    caller_flag = (alloc_flag + 1);
670    return 0;
671}
672
673static int parse_alloc_opt(const struct option *opt __used,
674              const char *arg __used, int unset __used)
675{
676    alloc_flag = (caller_flag + 1);
677    return 0;
678}
679
680static int parse_line_opt(const struct option *opt __used,
681              const char *arg, int unset __used)
682{
683    int lines;
684
685    if (!arg)
686        return -1;
687
688    lines = strtoul(arg, NULL, 10);
689
690    if (caller_flag > alloc_flag)
691        caller_lines = lines;
692    else
693        alloc_lines = lines;
694
695    return 0;
696}
697
698static const struct option kmem_options[] = {
699    OPT_STRING('i', "input", &input_name, "file",
700           "input file name"),
701    OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
702               "show per-callsite statistics",
703               parse_caller_opt),
704    OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
705               "show per-allocation statistics",
706               parse_alloc_opt),
707    OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
708             "sort by keys: ptr, call_site, bytes, hit, pingpong, frag",
709             parse_sort_opt),
710    OPT_CALLBACK('l', "line", NULL, "num",
711             "show n lines",
712             parse_line_opt),
713    OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
714    OPT_END()
715};
716
717static const char *record_args[] = {
718    "record",
719    "-a",
720    "-R",
721    "-M",
722    "-f",
723    "-c", "1",
724    "-e", "kmem:kmalloc",
725    "-e", "kmem:kmalloc_node",
726    "-e", "kmem:kfree",
727    "-e", "kmem:kmem_cache_alloc",
728    "-e", "kmem:kmem_cache_alloc_node",
729    "-e", "kmem:kmem_cache_free",
730};
731
732static int __cmd_record(int argc, const char **argv)
733{
734    unsigned int rec_argc, i, j;
735    const char **rec_argv;
736
737    rec_argc = ARRAY_SIZE(record_args) + argc - 1;
738    rec_argv = calloc(rec_argc + 1, sizeof(char *));
739
740    for (i = 0; i < ARRAY_SIZE(record_args); i++)
741        rec_argv[i] = strdup(record_args[i]);
742
743    for (j = 1; j < (unsigned int)argc; j++, i++)
744        rec_argv[i] = argv[j];
745
746    return cmd_record(i, rec_argv, NULL);
747}
748
749int cmd_kmem(int argc, const char **argv, const char *prefix __used)
750{
751    argc = parse_options(argc, argv, kmem_options, kmem_usage, 0);
752
753    if (!argc)
754        usage_with_options(kmem_usage, kmem_options);
755
756    symbol__init();
757
758    if (!strncmp(argv[0], "rec", 3)) {
759        return __cmd_record(argc, argv);
760    } else if (!strcmp(argv[0], "stat")) {
761        setup_cpunode_map();
762
763        if (list_empty(&caller_sort))
764            setup_sorting(&caller_sort, default_sort_order);
765        if (list_empty(&alloc_sort))
766            setup_sorting(&alloc_sort, default_sort_order);
767
768        return __cmd_kmem();
769    } else
770        usage_with_options(kmem_usage, kmem_options);
771
772    return 0;
773}
774
775

Archive Download this file

Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master

Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9



interactive