1 // SPDX-License-Identifier: GPL-2.0
6 #include "util/evlist.h"
7 #include "util/evsel.h"
8 #include "util/config.h"
10 #include "util/symbol.h"
11 #include "util/thread.h"
12 #include "util/header.h"
13 #include "util/session.h"
14 #include "util/tool.h"
15 #include "util/callchain.h"
16 #include "util/time-utils.h"
17 #include <linux/err.h>
19 #include <subcmd/pager.h>
20 #include <subcmd/parse-options.h>
21 #include "util/trace-event.h"
22 #include "util/data.h"
23 #include "util/cpumap.h"
25 #include "util/debug.h"
26 #include "util/string2.h"
28 #include <linux/kernel.h>
29 #include <linux/rbtree.h>
30 #include <linux/string.h>
31 #include <linux/zalloc.h>
37 #include <linux/ctype.h>
42 static long kmem_page_size;
46 } kmem_default = KMEM_SLAB; /* for backward compatibility */
49 typedef int (*sort_fn_t)(void *, void *);
51 static int alloc_flag;
52 static int caller_flag;
54 static int alloc_lines = -1;
55 static int caller_lines = -1;
73 static struct rb_root root_alloc_stat;
74 static struct rb_root root_alloc_sorted;
75 static struct rb_root root_caller_stat;
76 static struct rb_root root_caller_sorted;
78 static unsigned long total_requested, total_allocated, total_freed;
79 static unsigned long nr_allocs, nr_cross_allocs;
81 /* filters for controlling start and stop of time of analysis */
82 static struct perf_time_interval ptime;
85 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
86 int bytes_req, int bytes_alloc, int cpu)
88 struct rb_node **node = &root_alloc_stat.rb_node;
89 struct rb_node *parent = NULL;
90 struct alloc_stat *data = NULL;
94 data = rb_entry(*node, struct alloc_stat, node);
97 node = &(*node)->rb_right;
98 else if (ptr < data->ptr)
99 node = &(*node)->rb_left;
104 if (data && data->ptr == ptr) {
106 data->bytes_req += bytes_req;
107 data->bytes_alloc += bytes_alloc;
109 data = malloc(sizeof(*data));
111 pr_err("%s: malloc failed\n", __func__);
117 data->bytes_req = bytes_req;
118 data->bytes_alloc = bytes_alloc;
120 rb_link_node(&data->node, parent, node);
121 rb_insert_color(&data->node, &root_alloc_stat);
123 data->call_site = call_site;
124 data->alloc_cpu = cpu;
125 data->last_alloc = bytes_alloc;
130 static int insert_caller_stat(unsigned long call_site,
131 int bytes_req, int bytes_alloc)
133 struct rb_node **node = &root_caller_stat.rb_node;
134 struct rb_node *parent = NULL;
135 struct alloc_stat *data = NULL;
139 data = rb_entry(*node, struct alloc_stat, node);
141 if (call_site > data->call_site)
142 node = &(*node)->rb_right;
143 else if (call_site < data->call_site)
144 node = &(*node)->rb_left;
149 if (data && data->call_site == call_site) {
151 data->bytes_req += bytes_req;
152 data->bytes_alloc += bytes_alloc;
154 data = malloc(sizeof(*data));
156 pr_err("%s: malloc failed\n", __func__);
159 data->call_site = call_site;
162 data->bytes_req = bytes_req;
163 data->bytes_alloc = bytes_alloc;
165 rb_link_node(&data->node, parent, node);
166 rb_insert_color(&data->node, &root_caller_stat);
172 static int perf_evsel__process_alloc_event(struct evsel *evsel,
173 struct perf_sample *sample)
175 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
176 call_site = perf_evsel__intval(evsel, sample, "call_site");
177 int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
178 bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
180 if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
181 insert_caller_stat(call_site, bytes_req, bytes_alloc))
184 total_requested += bytes_req;
185 total_allocated += bytes_alloc;
191 static int perf_evsel__process_alloc_node_event(struct evsel *evsel,
192 struct perf_sample *sample)
194 int ret = perf_evsel__process_alloc_event(evsel, sample);
197 int node1 = cpu__get_node(sample->cpu),
198 node2 = perf_evsel__intval(evsel, sample, "node");
207 static int ptr_cmp(void *, void *);
208 static int slab_callsite_cmp(void *, void *);
210 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
211 unsigned long call_site,
212 struct rb_root *root,
215 struct rb_node *node = root->rb_node;
216 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
219 struct alloc_stat *data;
222 data = rb_entry(node, struct alloc_stat, node);
224 cmp = sort_fn(&key, data);
226 node = node->rb_left;
228 node = node->rb_right;
235 static int perf_evsel__process_free_event(struct evsel *evsel,
236 struct perf_sample *sample)
238 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
239 struct alloc_stat *s_alloc, *s_caller;
241 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
245 total_freed += s_alloc->last_alloc;
247 if ((short)sample->cpu != s_alloc->alloc_cpu) {
250 s_caller = search_alloc_stat(0, s_alloc->call_site,
255 s_caller->pingpong++;
257 s_alloc->alloc_cpu = -1;
262 static u64 total_page_alloc_bytes;
263 static u64 total_page_free_bytes;
264 static u64 total_page_nomatch_bytes;
265 static u64 total_page_fail_bytes;
266 static unsigned long nr_page_allocs;
267 static unsigned long nr_page_frees;
268 static unsigned long nr_page_fails;
269 static unsigned long nr_page_nomatch;
272 static bool live_page;
273 static struct perf_session *kmem_session;
275 #define MAX_MIGRATE_TYPES 6
276 #define MAX_PAGE_ORDER 11
278 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
286 unsigned migrate_type;
293 static struct rb_root page_live_tree;
294 static struct rb_root page_alloc_tree;
295 static struct rb_root page_alloc_sorted;
296 static struct rb_root page_caller_tree;
297 static struct rb_root page_caller_sorted;
305 static int nr_alloc_funcs;
306 static struct alloc_func *alloc_func_list;
308 static int funcmp(const void *a, const void *b)
310 const struct alloc_func *fa = a;
311 const struct alloc_func *fb = b;
313 if (fa->start > fb->start)
319 static int callcmp(const void *a, const void *b)
321 const struct alloc_func *fa = a;
322 const struct alloc_func *fb = b;
324 if (fb->start <= fa->start && fa->end < fb->end)
327 if (fa->start > fb->start)
333 static int build_alloc_func_list(void)
336 struct map *kernel_map;
338 struct rb_node *node;
339 struct alloc_func *func;
340 struct machine *machine = &kmem_session->machines.host;
341 regex_t alloc_func_regex;
342 static const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
344 ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
348 regerror(ret, &alloc_func_regex, err, sizeof(err));
349 pr_err("Invalid regex: %s\n%s", pattern, err);
353 kernel_map = machine__kernel_map(machine);
354 if (map__load(kernel_map) < 0) {
355 pr_err("cannot load kernel map\n");
359 map__for_each_symbol(kernel_map, sym, node) {
360 if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
363 func = realloc(alloc_func_list,
364 (nr_alloc_funcs + 1) * sizeof(*func));
368 pr_debug("alloc func: %s\n", sym->name);
369 func[nr_alloc_funcs].start = sym->start;
370 func[nr_alloc_funcs].end = sym->end;
371 func[nr_alloc_funcs].name = sym->name;
373 alloc_func_list = func;
377 qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
379 regfree(&alloc_func_regex);
384 * Find first non-memory allocation function from callchain.
385 * The allocation functions are in the 'alloc_func_list'.
387 static u64 find_callsite(struct evsel *evsel, struct perf_sample *sample)
389 struct addr_location al;
390 struct machine *machine = &kmem_session->machines.host;
391 struct callchain_cursor_node *node;
393 if (alloc_func_list == NULL) {
394 if (build_alloc_func_list() < 0)
398 al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
399 sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
401 callchain_cursor_commit(&callchain_cursor);
403 struct alloc_func key, *caller;
406 node = callchain_cursor_current(&callchain_cursor);
410 key.start = key.end = node->ip;
411 caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
412 sizeof(key), callcmp);
416 addr = map__unmap_ip(node->map, node->ip);
422 pr_debug3("skipping alloc function: %s\n", caller->name);
424 callchain_cursor_advance(&callchain_cursor);
428 pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
432 struct sort_dimension {
435 struct list_head list;
438 static LIST_HEAD(page_alloc_sort_input);
439 static LIST_HEAD(page_caller_sort_input);
441 static struct page_stat *
442 __page_stat__findnew_page(struct page_stat *pstat, bool create)
444 struct rb_node **node = &page_live_tree.rb_node;
445 struct rb_node *parent = NULL;
446 struct page_stat *data;
452 data = rb_entry(*node, struct page_stat, node);
454 cmp = data->page - pstat->page;
456 node = &parent->rb_left;
458 node = &parent->rb_right;
466 data = zalloc(sizeof(*data));
468 data->page = pstat->page;
469 data->order = pstat->order;
470 data->gfp_flags = pstat->gfp_flags;
471 data->migrate_type = pstat->migrate_type;
473 rb_link_node(&data->node, parent, node);
474 rb_insert_color(&data->node, &page_live_tree);
480 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
482 return __page_stat__findnew_page(pstat, false);
485 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
487 return __page_stat__findnew_page(pstat, true);
490 static struct page_stat *
491 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
493 struct rb_node **node = &page_alloc_tree.rb_node;
494 struct rb_node *parent = NULL;
495 struct page_stat *data;
496 struct sort_dimension *sort;
502 data = rb_entry(*node, struct page_stat, node);
504 list_for_each_entry(sort, &page_alloc_sort_input, list) {
505 cmp = sort->cmp(pstat, data);
511 node = &parent->rb_left;
513 node = &parent->rb_right;
521 data = zalloc(sizeof(*data));
523 data->page = pstat->page;
524 data->order = pstat->order;
525 data->gfp_flags = pstat->gfp_flags;
526 data->migrate_type = pstat->migrate_type;
528 rb_link_node(&data->node, parent, node);
529 rb_insert_color(&data->node, &page_alloc_tree);
535 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
537 return __page_stat__findnew_alloc(pstat, false);
540 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
542 return __page_stat__findnew_alloc(pstat, true);
545 static struct page_stat *
546 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
548 struct rb_node **node = &page_caller_tree.rb_node;
549 struct rb_node *parent = NULL;
550 struct page_stat *data;
551 struct sort_dimension *sort;
557 data = rb_entry(*node, struct page_stat, node);
559 list_for_each_entry(sort, &page_caller_sort_input, list) {
560 cmp = sort->cmp(pstat, data);
566 node = &parent->rb_left;
568 node = &parent->rb_right;
576 data = zalloc(sizeof(*data));
578 data->callsite = pstat->callsite;
579 data->order = pstat->order;
580 data->gfp_flags = pstat->gfp_flags;
581 data->migrate_type = pstat->migrate_type;
583 rb_link_node(&data->node, parent, node);
584 rb_insert_color(&data->node, &page_caller_tree);
590 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
592 return __page_stat__findnew_caller(pstat, false);
595 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
597 return __page_stat__findnew_caller(pstat, true);
600 static bool valid_page(u64 pfn_or_page)
602 if (use_pfn && pfn_or_page == -1UL)
604 if (!use_pfn && pfn_or_page == 0)
612 char *human_readable;
615 static struct gfp_flag *gfps;
618 static int gfpcmp(const void *a, const void *b)
620 const struct gfp_flag *fa = a;
621 const struct gfp_flag *fb = b;
623 return fa->flags - fb->flags;
626 /* see include/trace/events/mmflags.h */
627 static const struct {
628 const char *original;
630 } gfp_compact_table[] = {
631 { "GFP_TRANSHUGE", "THP" },
632 { "GFP_TRANSHUGE_LIGHT", "THL" },
633 { "GFP_HIGHUSER_MOVABLE", "HUM" },
634 { "GFP_HIGHUSER", "HU" },
636 { "GFP_KERNEL_ACCOUNT", "KAC" },
637 { "GFP_KERNEL", "K" },
638 { "GFP_NOFS", "NF" },
639 { "GFP_ATOMIC", "A" },
640 { "GFP_NOIO", "NI" },
641 { "GFP_NOWAIT", "NW" },
643 { "__GFP_HIGHMEM", "HM" },
644 { "GFP_DMA32", "D32" },
645 { "__GFP_HIGH", "H" },
646 { "__GFP_ATOMIC", "_A" },
649 { "__GFP_NOWARN", "NWR" },
650 { "__GFP_RETRY_MAYFAIL", "R" },
651 { "__GFP_NOFAIL", "NF" },
652 { "__GFP_NORETRY", "NR" },
653 { "__GFP_COMP", "C" },
654 { "__GFP_ZERO", "Z" },
655 { "__GFP_NOMEMALLOC", "NMA" },
656 { "__GFP_MEMALLOC", "MA" },
657 { "__GFP_HARDWALL", "HW" },
658 { "__GFP_THISNODE", "TN" },
659 { "__GFP_RECLAIMABLE", "RC" },
660 { "__GFP_MOVABLE", "M" },
661 { "__GFP_ACCOUNT", "AC" },
662 { "__GFP_WRITE", "WR" },
663 { "__GFP_RECLAIM", "R" },
664 { "__GFP_DIRECT_RECLAIM", "DR" },
665 { "__GFP_KSWAPD_RECLAIM", "KR" },
668 static size_t max_gfp_len;
670 static char *compact_gfp_flags(char *gfp_flags)
672 char *orig_flags = strdup(gfp_flags);
673 char *new_flags = NULL;
674 char *str, *pos = NULL;
677 if (orig_flags == NULL)
680 str = strtok_r(orig_flags, "|", &pos);
686 for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
687 if (strcmp(gfp_compact_table[i].original, str))
690 cpt = gfp_compact_table[i].compact;
691 new = realloc(new_flags, len + strlen(cpt) + 2);
700 strcpy(new_flags, cpt);
702 strcat(new_flags, "|");
703 strcat(new_flags, cpt);
710 str = strtok_r(NULL, "|", &pos);
713 if (max_gfp_len < len)
720 static char *compact_gfp_string(unsigned long gfp_flags)
722 struct gfp_flag key = {
725 struct gfp_flag *gfp;
727 gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
729 return gfp->compact_str;
734 static int parse_gfp_flags(struct evsel *evsel, struct perf_sample *sample,
735 unsigned int gfp_flags)
737 struct tep_record record = {
739 .data = sample->raw_data,
740 .size = sample->raw_size,
742 struct trace_seq seq;
743 char *str, *pos = NULL;
746 struct gfp_flag key = {
750 if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
754 trace_seq_init(&seq);
755 tep_print_event(evsel->tp_format->tep,
756 &seq, &record, "%s", TEP_PRINT_INFO);
758 str = strtok_r(seq.buffer, " ", &pos);
760 if (!strncmp(str, "gfp_flags=", 10)) {
761 struct gfp_flag *new;
763 new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
770 new->flags = gfp_flags;
771 new->human_readable = strdup(str + 10);
772 new->compact_str = compact_gfp_flags(str + 10);
773 if (!new->human_readable || !new->compact_str)
776 qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
779 str = strtok_r(NULL, " ", &pos);
782 trace_seq_destroy(&seq);
786 static int perf_evsel__process_page_alloc_event(struct evsel *evsel,
787 struct perf_sample *sample)
790 unsigned int order = perf_evsel__intval(evsel, sample, "order");
791 unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
792 unsigned int migrate_type = perf_evsel__intval(evsel, sample,
794 u64 bytes = kmem_page_size << order;
796 struct page_stat *pstat;
797 struct page_stat this = {
799 .gfp_flags = gfp_flags,
800 .migrate_type = migrate_type,
804 page = perf_evsel__intval(evsel, sample, "pfn");
806 page = perf_evsel__intval(evsel, sample, "page");
809 total_page_alloc_bytes += bytes;
811 if (!valid_page(page)) {
813 total_page_fail_bytes += bytes;
818 if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
821 callsite = find_callsite(evsel, sample);
824 * This is to find the current page (with correct gfp flags and
825 * migrate type) at free event.
828 pstat = page_stat__findnew_page(&this);
833 pstat->alloc_bytes += bytes;
834 pstat->callsite = callsite;
837 pstat = page_stat__findnew_alloc(&this);
842 pstat->alloc_bytes += bytes;
843 pstat->callsite = callsite;
846 this.callsite = callsite;
847 pstat = page_stat__findnew_caller(&this);
852 pstat->alloc_bytes += bytes;
854 order_stats[order][migrate_type]++;
859 static int perf_evsel__process_page_free_event(struct evsel *evsel,
860 struct perf_sample *sample)
863 unsigned int order = perf_evsel__intval(evsel, sample, "order");
864 u64 bytes = kmem_page_size << order;
865 struct page_stat *pstat;
866 struct page_stat this = {
871 page = perf_evsel__intval(evsel, sample, "pfn");
873 page = perf_evsel__intval(evsel, sample, "page");
876 total_page_free_bytes += bytes;
879 pstat = page_stat__find_page(&this);
881 pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
885 total_page_nomatch_bytes += bytes;
890 this.gfp_flags = pstat->gfp_flags;
891 this.migrate_type = pstat->migrate_type;
892 this.callsite = pstat->callsite;
894 rb_erase(&pstat->node, &page_live_tree);
898 order_stats[this.order][this.migrate_type]--;
900 pstat = page_stat__find_alloc(&this);
905 pstat->free_bytes += bytes;
908 pstat = page_stat__find_caller(&this);
913 pstat->free_bytes += bytes;
917 pstat->alloc_bytes -= bytes;
919 if (pstat->nr_alloc == 0) {
920 rb_erase(&pstat->node, &page_caller_tree);
928 static bool perf_kmem__skip_sample(struct perf_sample *sample)
930 /* skip sample based on time? */
931 if (perf_time__skip_sample(&ptime, sample->time))
937 typedef int (*tracepoint_handler)(struct evsel *evsel,
938 struct perf_sample *sample);
940 static int process_sample_event(struct perf_tool *tool __maybe_unused,
941 union perf_event *event,
942 struct perf_sample *sample,
944 struct machine *machine)
947 struct thread *thread = machine__findnew_thread(machine, sample->pid,
950 if (thread == NULL) {
951 pr_debug("problem processing %d event, skipping it.\n",
956 if (perf_kmem__skip_sample(sample))
959 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
961 if (evsel->handler != NULL) {
962 tracepoint_handler f = evsel->handler;
963 err = f(evsel, sample);
971 static struct perf_tool perf_kmem = {
972 .sample = process_sample_event,
973 .comm = perf_event__process_comm,
974 .mmap = perf_event__process_mmap,
975 .mmap2 = perf_event__process_mmap2,
976 .namespaces = perf_event__process_namespaces,
977 .ordered_events = true,
980 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
985 return 100.0 - (100.0 * n_req / n_alloc);
988 static void __print_slab_result(struct rb_root *root,
989 struct perf_session *session,
990 int n_lines, int is_caller)
992 struct rb_node *next;
993 struct machine *machine = &session->machines.host;
995 printf("%.105s\n", graph_dotted_line);
996 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
997 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
998 printf("%.105s\n", graph_dotted_line);
1000 next = rb_first(root);
1002 while (next && n_lines--) {
1003 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
1005 struct symbol *sym = NULL;
1011 addr = data->call_site;
1013 sym = machine__find_kernel_symbol(machine, addr, &map);
1018 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1019 addr - map->unmap_ip(map, sym->start));
1021 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1022 printf(" %-34s |", buf);
1024 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1025 (unsigned long long)data->bytes_alloc,
1026 (unsigned long)data->bytes_alloc / data->hit,
1027 (unsigned long long)data->bytes_req,
1028 (unsigned long)data->bytes_req / data->hit,
1029 (unsigned long)data->hit,
1030 (unsigned long)data->pingpong,
1031 fragmentation(data->bytes_req, data->bytes_alloc));
1033 next = rb_next(next);
1037 printf(" ... | ... | ... | ... | ... | ... \n");
1039 printf("%.105s\n", graph_dotted_line);
1042 static const char * const migrate_type_str[] = {
1051 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1053 struct rb_node *next = rb_first(&page_alloc_sorted);
1054 struct machine *machine = &session->machines.host;
1056 int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1058 printf("\n%.105s\n", graph_dotted_line);
1059 printf(" %-16s | %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n",
1060 use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1061 gfp_len, "GFP flags");
1062 printf("%.105s\n", graph_dotted_line);
1065 format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1067 format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1069 while (next && n_lines--) {
1070 struct page_stat *data;
1076 data = rb_entry(next, struct page_stat, node);
1077 sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1081 scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1083 printf(format, (unsigned long long)data->page,
1084 (unsigned long long)data->alloc_bytes / 1024,
1085 data->nr_alloc, data->order,
1086 migrate_type_str[data->migrate_type],
1087 gfp_len, compact_gfp_string(data->gfp_flags), caller);
1089 next = rb_next(next);
1092 if (n_lines == -1) {
1093 printf(" ... | ... | ... | ... | ... | %-*s | ...\n",
1097 printf("%.105s\n", graph_dotted_line);
1100 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1102 struct rb_node *next = rb_first(&page_caller_sorted);
1103 struct machine *machine = &session->machines.host;
1104 int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1106 printf("\n%.105s\n", graph_dotted_line);
1107 printf(" %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n",
1108 live_page ? "Live" : "Total", gfp_len, "GFP flags");
1109 printf("%.105s\n", graph_dotted_line);
1111 while (next && n_lines--) {
1112 struct page_stat *data;
1118 data = rb_entry(next, struct page_stat, node);
1119 sym = machine__find_kernel_symbol(machine, data->callsite, &map);
1123 scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1125 printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1126 (unsigned long long)data->alloc_bytes / 1024,
1127 data->nr_alloc, data->order,
1128 migrate_type_str[data->migrate_type],
1129 gfp_len, compact_gfp_string(data->gfp_flags), caller);
1131 next = rb_next(next);
1134 if (n_lines == -1) {
1135 printf(" ... | ... | ... | ... | %-*s | ...\n",
1139 printf("%.105s\n", graph_dotted_line);
1142 static void print_gfp_flags(void)
1147 printf("# GFP flags\n");
1148 printf("# ---------\n");
1149 for (i = 0; i < nr_gfps; i++) {
1150 printf("# %08x: %*s: %s\n", gfps[i].flags,
1151 (int) max_gfp_len, gfps[i].compact_str,
1152 gfps[i].human_readable);
1156 static void print_slab_summary(void)
1158 printf("\nSUMMARY (SLAB allocator)");
1159 printf("\n========================\n");
1160 printf("Total bytes requested: %'lu\n", total_requested);
1161 printf("Total bytes allocated: %'lu\n", total_allocated);
1162 printf("Total bytes freed: %'lu\n", total_freed);
1163 if (total_allocated > total_freed) {
1164 printf("Net total bytes allocated: %'lu\n",
1165 total_allocated - total_freed);
1167 printf("Total bytes wasted on internal fragmentation: %'lu\n",
1168 total_allocated - total_requested);
1169 printf("Internal fragmentation: %f%%\n",
1170 fragmentation(total_requested, total_allocated));
1171 printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1174 static void print_page_summary(void)
1177 u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1178 u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1180 printf("\nSUMMARY (page allocator)");
1181 printf("\n========================\n");
1182 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1183 nr_page_allocs, total_page_alloc_bytes / 1024);
1184 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free requests",
1185 nr_page_frees, total_page_free_bytes / 1024);
1188 printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1189 nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1190 printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1191 nr_page_allocs - nr_alloc_freed,
1192 (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1193 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1194 nr_page_nomatch, total_page_nomatch_bytes / 1024);
1197 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1198 nr_page_fails, total_page_fail_bytes / 1024);
1201 printf("%5s %12s %12s %12s %12s %12s\n", "Order", "Unmovable",
1202 "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1203 printf("%.5s %.12s %.12s %.12s %.12s %.12s\n", graph_dotted_line,
1204 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1205 graph_dotted_line, graph_dotted_line);
1207 for (o = 0; o < MAX_PAGE_ORDER; o++) {
1209 for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1210 if (order_stats[o][m])
1211 printf(" %'12d", order_stats[o][m]);
1213 printf(" %12c", '.');
1219 static void print_slab_result(struct perf_session *session)
1222 __print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1224 __print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1225 print_slab_summary();
1228 static void print_page_result(struct perf_session *session)
1230 if (caller_flag || alloc_flag)
1233 __print_page_caller_result(session, caller_lines);
1235 __print_page_alloc_result(session, alloc_lines);
1236 print_page_summary();
1239 static void print_result(struct perf_session *session)
1242 print_slab_result(session);
1244 print_page_result(session);
1247 static LIST_HEAD(slab_caller_sort);
1248 static LIST_HEAD(slab_alloc_sort);
1249 static LIST_HEAD(page_caller_sort);
1250 static LIST_HEAD(page_alloc_sort);
1252 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1253 struct list_head *sort_list)
1255 struct rb_node **new = &(root->rb_node);
1256 struct rb_node *parent = NULL;
1257 struct sort_dimension *sort;
1260 struct alloc_stat *this;
1263 this = rb_entry(*new, struct alloc_stat, node);
1266 list_for_each_entry(sort, sort_list, list) {
1267 cmp = sort->cmp(data, this);
1273 new = &((*new)->rb_left);
1275 new = &((*new)->rb_right);
1278 rb_link_node(&data->node, parent, new);
1279 rb_insert_color(&data->node, root);
1282 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1283 struct list_head *sort_list)
1285 struct rb_node *node;
1286 struct alloc_stat *data;
1289 node = rb_first(root);
1293 rb_erase(node, root);
1294 data = rb_entry(node, struct alloc_stat, node);
1295 sort_slab_insert(root_sorted, data, sort_list);
1299 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1300 struct list_head *sort_list)
1302 struct rb_node **new = &root->rb_node;
1303 struct rb_node *parent = NULL;
1304 struct sort_dimension *sort;
1307 struct page_stat *this;
1310 this = rb_entry(*new, struct page_stat, node);
1313 list_for_each_entry(sort, sort_list, list) {
1314 cmp = sort->cmp(data, this);
1320 new = &parent->rb_left;
1322 new = &parent->rb_right;
1325 rb_link_node(&data->node, parent, new);
1326 rb_insert_color(&data->node, root);
1329 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1330 struct list_head *sort_list)
1332 struct rb_node *node;
1333 struct page_stat *data;
1336 node = rb_first(root);
1340 rb_erase(node, root);
1341 data = rb_entry(node, struct page_stat, node);
1342 sort_page_insert(root_sorted, data, sort_list);
1346 static void sort_result(void)
1349 __sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1351 __sort_slab_result(&root_caller_stat, &root_caller_sorted,
1356 __sort_page_result(&page_live_tree, &page_alloc_sorted,
1359 __sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1362 __sort_page_result(&page_caller_tree, &page_caller_sorted,
1367 static int __cmd_kmem(struct perf_session *session)
1370 struct evsel *evsel;
1371 const struct evsel_str_handler kmem_tracepoints[] = {
1372 /* slab allocator */
1373 { "kmem:kmalloc", perf_evsel__process_alloc_event, },
1374 { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, },
1375 { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, },
1376 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1377 { "kmem:kfree", perf_evsel__process_free_event, },
1378 { "kmem:kmem_cache_free", perf_evsel__process_free_event, },
1379 /* page allocator */
1380 { "kmem:mm_page_alloc", perf_evsel__process_page_alloc_event, },
1381 { "kmem:mm_page_free", perf_evsel__process_page_free_event, },
1384 if (!perf_session__has_traces(session, "kmem record"))
1387 if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1388 pr_err("Initializing perf session tracepoint handlers failed\n");
1392 evlist__for_each_entry(session->evlist, evsel) {
1393 if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1394 perf_evsel__field(evsel, "pfn")) {
1401 err = perf_session__process_events(session);
1403 pr_err("error during process events: %d\n", err);
1407 print_result(session);
1412 /* slab sort keys */
1413 static int ptr_cmp(void *a, void *b)
1415 struct alloc_stat *l = a;
1416 struct alloc_stat *r = b;
1418 if (l->ptr < r->ptr)
1420 else if (l->ptr > r->ptr)
1425 static struct sort_dimension ptr_sort_dimension = {
1430 static int slab_callsite_cmp(void *a, void *b)
1432 struct alloc_stat *l = a;
1433 struct alloc_stat *r = b;
1435 if (l->call_site < r->call_site)
1437 else if (l->call_site > r->call_site)
1442 static struct sort_dimension callsite_sort_dimension = {
1444 .cmp = slab_callsite_cmp,
1447 static int hit_cmp(void *a, void *b)
1449 struct alloc_stat *l = a;
1450 struct alloc_stat *r = b;
1452 if (l->hit < r->hit)
1454 else if (l->hit > r->hit)
1459 static struct sort_dimension hit_sort_dimension = {
1464 static int bytes_cmp(void *a, void *b)
1466 struct alloc_stat *l = a;
1467 struct alloc_stat *r = b;
1469 if (l->bytes_alloc < r->bytes_alloc)
1471 else if (l->bytes_alloc > r->bytes_alloc)
1476 static struct sort_dimension bytes_sort_dimension = {
1481 static int frag_cmp(void *a, void *b)
1484 struct alloc_stat *l = a;
1485 struct alloc_stat *r = b;
1487 x = fragmentation(l->bytes_req, l->bytes_alloc);
1488 y = fragmentation(r->bytes_req, r->bytes_alloc);
1497 static struct sort_dimension frag_sort_dimension = {
1502 static int pingpong_cmp(void *a, void *b)
1504 struct alloc_stat *l = a;
1505 struct alloc_stat *r = b;
1507 if (l->pingpong < r->pingpong)
1509 else if (l->pingpong > r->pingpong)
1514 static struct sort_dimension pingpong_sort_dimension = {
1516 .cmp = pingpong_cmp,
1519 /* page sort keys */
1520 static int page_cmp(void *a, void *b)
1522 struct page_stat *l = a;
1523 struct page_stat *r = b;
1525 if (l->page < r->page)
1527 else if (l->page > r->page)
1532 static struct sort_dimension page_sort_dimension = {
1537 static int page_callsite_cmp(void *a, void *b)
1539 struct page_stat *l = a;
1540 struct page_stat *r = b;
1542 if (l->callsite < r->callsite)
1544 else if (l->callsite > r->callsite)
1549 static struct sort_dimension page_callsite_sort_dimension = {
1551 .cmp = page_callsite_cmp,
1554 static int page_hit_cmp(void *a, void *b)
1556 struct page_stat *l = a;
1557 struct page_stat *r = b;
1559 if (l->nr_alloc < r->nr_alloc)
1561 else if (l->nr_alloc > r->nr_alloc)
1566 static struct sort_dimension page_hit_sort_dimension = {
1568 .cmp = page_hit_cmp,
1571 static int page_bytes_cmp(void *a, void *b)
1573 struct page_stat *l = a;
1574 struct page_stat *r = b;
1576 if (l->alloc_bytes < r->alloc_bytes)
1578 else if (l->alloc_bytes > r->alloc_bytes)
1583 static struct sort_dimension page_bytes_sort_dimension = {
1585 .cmp = page_bytes_cmp,
1588 static int page_order_cmp(void *a, void *b)
1590 struct page_stat *l = a;
1591 struct page_stat *r = b;
1593 if (l->order < r->order)
1595 else if (l->order > r->order)
1600 static struct sort_dimension page_order_sort_dimension = {
1602 .cmp = page_order_cmp,
1605 static int migrate_type_cmp(void *a, void *b)
1607 struct page_stat *l = a;
1608 struct page_stat *r = b;
1610 /* for internal use to find free'd page */
1611 if (l->migrate_type == -1U)
1614 if (l->migrate_type < r->migrate_type)
1616 else if (l->migrate_type > r->migrate_type)
1621 static struct sort_dimension migrate_type_sort_dimension = {
1623 .cmp = migrate_type_cmp,
1626 static int gfp_flags_cmp(void *a, void *b)
1628 struct page_stat *l = a;
1629 struct page_stat *r = b;
1631 /* for internal use to find free'd page */
1632 if (l->gfp_flags == -1U)
1635 if (l->gfp_flags < r->gfp_flags)
1637 else if (l->gfp_flags > r->gfp_flags)
1642 static struct sort_dimension gfp_flags_sort_dimension = {
1644 .cmp = gfp_flags_cmp,
1647 static struct sort_dimension *slab_sorts[] = {
1648 &ptr_sort_dimension,
1649 &callsite_sort_dimension,
1650 &hit_sort_dimension,
1651 &bytes_sort_dimension,
1652 &frag_sort_dimension,
1653 &pingpong_sort_dimension,
1656 static struct sort_dimension *page_sorts[] = {
1657 &page_sort_dimension,
1658 &page_callsite_sort_dimension,
1659 &page_hit_sort_dimension,
1660 &page_bytes_sort_dimension,
1661 &page_order_sort_dimension,
1662 &migrate_type_sort_dimension,
1663 &gfp_flags_sort_dimension,
1666 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1668 struct sort_dimension *sort;
1671 for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1672 if (!strcmp(slab_sorts[i]->name, tok)) {
1673 sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1675 pr_err("%s: memdup failed\n", __func__);
1678 list_add_tail(&sort->list, list);
1686 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1688 struct sort_dimension *sort;
1691 for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1692 if (!strcmp(page_sorts[i]->name, tok)) {
1693 sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1695 pr_err("%s: memdup failed\n", __func__);
1698 list_add_tail(&sort->list, list);
1706 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1709 char *str = strdup(arg);
1713 pr_err("%s: strdup failed\n", __func__);
1718 tok = strsep(&pos, ",");
1721 if (slab_sort_dimension__add(tok, sort_list) < 0) {
1722 pr_err("Unknown slab --sort key: '%s'", tok);
1732 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1735 char *str = strdup(arg);
1739 pr_err("%s: strdup failed\n", __func__);
1744 tok = strsep(&pos, ",");
1747 if (page_sort_dimension__add(tok, sort_list) < 0) {
1748 pr_err("Unknown page --sort key: '%s'", tok);
1758 static int parse_sort_opt(const struct option *opt __maybe_unused,
1759 const char *arg, int unset __maybe_unused)
1764 if (kmem_page > kmem_slab ||
1765 (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1766 if (caller_flag > alloc_flag)
1767 return setup_page_sorting(&page_caller_sort, arg);
1769 return setup_page_sorting(&page_alloc_sort, arg);
1771 if (caller_flag > alloc_flag)
1772 return setup_slab_sorting(&slab_caller_sort, arg);
1774 return setup_slab_sorting(&slab_alloc_sort, arg);
1780 static int parse_caller_opt(const struct option *opt __maybe_unused,
1781 const char *arg __maybe_unused,
1782 int unset __maybe_unused)
1784 caller_flag = (alloc_flag + 1);
1788 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1789 const char *arg __maybe_unused,
1790 int unset __maybe_unused)
1792 alloc_flag = (caller_flag + 1);
1796 static int parse_slab_opt(const struct option *opt __maybe_unused,
1797 const char *arg __maybe_unused,
1798 int unset __maybe_unused)
1800 kmem_slab = (kmem_page + 1);
1804 static int parse_page_opt(const struct option *opt __maybe_unused,
1805 const char *arg __maybe_unused,
1806 int unset __maybe_unused)
1808 kmem_page = (kmem_slab + 1);
1812 static int parse_line_opt(const struct option *opt __maybe_unused,
1813 const char *arg, int unset __maybe_unused)
1820 lines = strtoul(arg, NULL, 10);
1822 if (caller_flag > alloc_flag)
1823 caller_lines = lines;
1825 alloc_lines = lines;
1830 static int __cmd_record(int argc, const char **argv)
1832 const char * const record_args[] = {
1833 "record", "-a", "-R", "-c", "1",
1835 const char * const slab_events[] = {
1836 "-e", "kmem:kmalloc",
1837 "-e", "kmem:kmalloc_node",
1839 "-e", "kmem:kmem_cache_alloc",
1840 "-e", "kmem:kmem_cache_alloc_node",
1841 "-e", "kmem:kmem_cache_free",
1843 const char * const page_events[] = {
1844 "-e", "kmem:mm_page_alloc",
1845 "-e", "kmem:mm_page_free",
1847 unsigned int rec_argc, i, j;
1848 const char **rec_argv;
1850 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1852 rec_argc += ARRAY_SIZE(slab_events);
1854 rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1856 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1858 if (rec_argv == NULL)
1861 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1862 rec_argv[i] = strdup(record_args[i]);
1865 for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1866 rec_argv[i] = strdup(slab_events[j]);
1869 rec_argv[i++] = strdup("-g");
1871 for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1872 rec_argv[i] = strdup(page_events[j]);
1875 for (j = 1; j < (unsigned int)argc; j++, i++)
1876 rec_argv[i] = argv[j];
1878 return cmd_record(i, rec_argv);
1881 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1883 if (!strcmp(var, "kmem.default")) {
1884 if (!strcmp(value, "slab"))
1885 kmem_default = KMEM_SLAB;
1886 else if (!strcmp(value, "page"))
1887 kmem_default = KMEM_PAGE;
1889 pr_err("invalid default value ('slab' or 'page' required): %s\n",
1897 int cmd_kmem(int argc, const char **argv)
1899 const char * const default_slab_sort = "frag,hit,bytes";
1900 const char * const default_page_sort = "bytes,hit";
1901 struct perf_data data = {
1902 .mode = PERF_DATA_MODE_READ,
1904 const struct option kmem_options[] = {
1905 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1906 OPT_INCR('v', "verbose", &verbose,
1907 "be more verbose (show symbol address, etc)"),
1908 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1909 "show per-callsite statistics", parse_caller_opt),
1910 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1911 "show per-allocation statistics", parse_alloc_opt),
1912 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1913 "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1914 "page, order, migtype, gfp", parse_sort_opt),
1915 OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1916 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1917 OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1918 OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1920 OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1922 OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1923 OPT_STRING(0, "time", &time_str, "str",
1924 "Time span of interest (start,stop)"),
1927 const char *const kmem_subcommands[] = { "record", "stat", NULL };
1928 const char *kmem_usage[] = {
1932 struct perf_session *session;
1933 static const char errmsg[] = "No %s allocation events found. Have you run 'perf kmem record --%s'?\n";
1934 int ret = perf_config(kmem_config, NULL);
1939 argc = parse_options_subcommand(argc, argv, kmem_options,
1940 kmem_subcommands, kmem_usage, 0);
1943 usage_with_options(kmem_usage, kmem_options);
1945 if (kmem_slab == 0 && kmem_page == 0) {
1946 if (kmem_default == KMEM_SLAB)
1952 if (!strncmp(argv[0], "rec", 3)) {
1954 return __cmd_record(argc, argv);
1957 data.path = input_name;
1959 kmem_session = session = perf_session__new(&data, false, &perf_kmem);
1960 if (IS_ERR(session))
1961 return PTR_ERR(session);
1966 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1968 pr_err(errmsg, "slab", "slab");
1974 struct evsel *evsel;
1976 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1977 "kmem:mm_page_alloc");
1978 if (evsel == NULL) {
1979 pr_err(errmsg, "page", "page");
1983 kmem_page_size = tep_get_page_size(evsel->tp_format->tep);
1984 symbol_conf.use_callchain = true;
1987 symbol__init(&session->header.env);
1989 if (perf_time__parse_str(&ptime, time_str) != 0) {
1990 pr_err("Invalid time string\n");
1995 if (!strcmp(argv[0], "stat")) {
1996 setlocale(LC_ALL, "");
1998 if (cpu__setup_cpunode_map())
2001 if (list_empty(&slab_caller_sort))
2002 setup_slab_sorting(&slab_caller_sort, default_slab_sort);
2003 if (list_empty(&slab_alloc_sort))
2004 setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2005 if (list_empty(&page_caller_sort))
2006 setup_page_sorting(&page_caller_sort, default_page_sort);
2007 if (list_empty(&page_alloc_sort))
2008 setup_page_sorting(&page_alloc_sort, default_page_sort);
2011 setup_page_sorting(&page_alloc_sort_input,
2012 "page,order,migtype,gfp");
2013 setup_page_sorting(&page_caller_sort_input,
2014 "callsite,order,migtype,gfp");
2016 ret = __cmd_kmem(session);
2018 usage_with_options(kmem_usage, kmem_options);
2021 perf_session__delete(session);