1 // SPDX-License-Identifier: GPL-2.0
5 #include "util/evlist.h"
6 #include "util/evsel.h"
8 #include "util/config.h"
9 #include "util/symbol.h"
10 #include "util/thread.h"
11 #include "util/header.h"
12 #include "util/session.h"
13 #include "util/tool.h"
14 #include "util/callchain.h"
15 #include "util/time-utils.h"
17 #include <subcmd/parse-options.h>
18 #include "util/trace-event.h"
19 #include "util/data.h"
20 #include "util/cpumap.h"
22 #include "util/debug.h"
24 #include <linux/kernel.h>
25 #include <linux/rbtree.h>
26 #include <linux/string.h>
32 #include "sane_ctype.h"
37 static long kmem_page_size;
41 } kmem_default = KMEM_SLAB; /* for backward compatibility */
44 typedef int (*sort_fn_t)(void *, void *);
46 static int alloc_flag;
47 static int caller_flag;
49 static int alloc_lines = -1;
50 static int caller_lines = -1;
68 static struct rb_root root_alloc_stat;
69 static struct rb_root root_alloc_sorted;
70 static struct rb_root root_caller_stat;
71 static struct rb_root root_caller_sorted;
73 static unsigned long total_requested, total_allocated, total_freed;
74 static unsigned long nr_allocs, nr_cross_allocs;
76 /* filters for controlling start and stop of time of analysis */
77 static struct perf_time_interval ptime;
80 static int insert_alloc_stat(unsigned long call_site, unsigned long ptr,
81 int bytes_req, int bytes_alloc, int cpu)
83 struct rb_node **node = &root_alloc_stat.rb_node;
84 struct rb_node *parent = NULL;
85 struct alloc_stat *data = NULL;
89 data = rb_entry(*node, struct alloc_stat, node);
92 node = &(*node)->rb_right;
93 else if (ptr < data->ptr)
94 node = &(*node)->rb_left;
99 if (data && data->ptr == ptr) {
101 data->bytes_req += bytes_req;
102 data->bytes_alloc += bytes_alloc;
104 data = malloc(sizeof(*data));
106 pr_err("%s: malloc failed\n", __func__);
112 data->bytes_req = bytes_req;
113 data->bytes_alloc = bytes_alloc;
115 rb_link_node(&data->node, parent, node);
116 rb_insert_color(&data->node, &root_alloc_stat);
118 data->call_site = call_site;
119 data->alloc_cpu = cpu;
120 data->last_alloc = bytes_alloc;
125 static int insert_caller_stat(unsigned long call_site,
126 int bytes_req, int bytes_alloc)
128 struct rb_node **node = &root_caller_stat.rb_node;
129 struct rb_node *parent = NULL;
130 struct alloc_stat *data = NULL;
134 data = rb_entry(*node, struct alloc_stat, node);
136 if (call_site > data->call_site)
137 node = &(*node)->rb_right;
138 else if (call_site < data->call_site)
139 node = &(*node)->rb_left;
144 if (data && data->call_site == call_site) {
146 data->bytes_req += bytes_req;
147 data->bytes_alloc += bytes_alloc;
149 data = malloc(sizeof(*data));
151 pr_err("%s: malloc failed\n", __func__);
154 data->call_site = call_site;
157 data->bytes_req = bytes_req;
158 data->bytes_alloc = bytes_alloc;
160 rb_link_node(&data->node, parent, node);
161 rb_insert_color(&data->node, &root_caller_stat);
167 static int perf_evsel__process_alloc_event(struct perf_evsel *evsel,
168 struct perf_sample *sample)
170 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr"),
171 call_site = perf_evsel__intval(evsel, sample, "call_site");
172 int bytes_req = perf_evsel__intval(evsel, sample, "bytes_req"),
173 bytes_alloc = perf_evsel__intval(evsel, sample, "bytes_alloc");
175 if (insert_alloc_stat(call_site, ptr, bytes_req, bytes_alloc, sample->cpu) ||
176 insert_caller_stat(call_site, bytes_req, bytes_alloc))
179 total_requested += bytes_req;
180 total_allocated += bytes_alloc;
186 static int perf_evsel__process_alloc_node_event(struct perf_evsel *evsel,
187 struct perf_sample *sample)
189 int ret = perf_evsel__process_alloc_event(evsel, sample);
192 int node1 = cpu__get_node(sample->cpu),
193 node2 = perf_evsel__intval(evsel, sample, "node");
202 static int ptr_cmp(void *, void *);
203 static int slab_callsite_cmp(void *, void *);
205 static struct alloc_stat *search_alloc_stat(unsigned long ptr,
206 unsigned long call_site,
207 struct rb_root *root,
210 struct rb_node *node = root->rb_node;
211 struct alloc_stat key = { .ptr = ptr, .call_site = call_site };
214 struct alloc_stat *data;
217 data = rb_entry(node, struct alloc_stat, node);
219 cmp = sort_fn(&key, data);
221 node = node->rb_left;
223 node = node->rb_right;
230 static int perf_evsel__process_free_event(struct perf_evsel *evsel,
231 struct perf_sample *sample)
233 unsigned long ptr = perf_evsel__intval(evsel, sample, "ptr");
234 struct alloc_stat *s_alloc, *s_caller;
236 s_alloc = search_alloc_stat(ptr, 0, &root_alloc_stat, ptr_cmp);
240 total_freed += s_alloc->last_alloc;
242 if ((short)sample->cpu != s_alloc->alloc_cpu) {
245 s_caller = search_alloc_stat(0, s_alloc->call_site,
250 s_caller->pingpong++;
252 s_alloc->alloc_cpu = -1;
257 static u64 total_page_alloc_bytes;
258 static u64 total_page_free_bytes;
259 static u64 total_page_nomatch_bytes;
260 static u64 total_page_fail_bytes;
261 static unsigned long nr_page_allocs;
262 static unsigned long nr_page_frees;
263 static unsigned long nr_page_fails;
264 static unsigned long nr_page_nomatch;
267 static bool live_page;
268 static struct perf_session *kmem_session;
270 #define MAX_MIGRATE_TYPES 6
271 #define MAX_PAGE_ORDER 11
273 static int order_stats[MAX_PAGE_ORDER][MAX_MIGRATE_TYPES];
281 unsigned migrate_type;
288 static struct rb_root page_live_tree;
289 static struct rb_root page_alloc_tree;
290 static struct rb_root page_alloc_sorted;
291 static struct rb_root page_caller_tree;
292 static struct rb_root page_caller_sorted;
300 static int nr_alloc_funcs;
301 static struct alloc_func *alloc_func_list;
303 static int funcmp(const void *a, const void *b)
305 const struct alloc_func *fa = a;
306 const struct alloc_func *fb = b;
308 if (fa->start > fb->start)
314 static int callcmp(const void *a, const void *b)
316 const struct alloc_func *fa = a;
317 const struct alloc_func *fb = b;
319 if (fb->start <= fa->start && fa->end < fb->end)
322 if (fa->start > fb->start)
328 static int build_alloc_func_list(void)
331 struct map *kernel_map;
333 struct rb_node *node;
334 struct alloc_func *func;
335 struct machine *machine = &kmem_session->machines.host;
336 regex_t alloc_func_regex;
337 const char pattern[] = "^_?_?(alloc|get_free|get_zeroed)_pages?";
339 ret = regcomp(&alloc_func_regex, pattern, REG_EXTENDED);
343 regerror(ret, &alloc_func_regex, err, sizeof(err));
344 pr_err("Invalid regex: %s\n%s", pattern, err);
348 kernel_map = machine__kernel_map(machine);
349 if (map__load(kernel_map) < 0) {
350 pr_err("cannot load kernel map\n");
354 map__for_each_symbol(kernel_map, sym, node) {
355 if (regexec(&alloc_func_regex, sym->name, 0, NULL, 0))
358 func = realloc(alloc_func_list,
359 (nr_alloc_funcs + 1) * sizeof(*func));
363 pr_debug("alloc func: %s\n", sym->name);
364 func[nr_alloc_funcs].start = sym->start;
365 func[nr_alloc_funcs].end = sym->end;
366 func[nr_alloc_funcs].name = sym->name;
368 alloc_func_list = func;
372 qsort(alloc_func_list, nr_alloc_funcs, sizeof(*func), funcmp);
374 regfree(&alloc_func_regex);
379 * Find first non-memory allocation function from callchain.
380 * The allocation functions are in the 'alloc_func_list'.
382 static u64 find_callsite(struct perf_evsel *evsel, struct perf_sample *sample)
384 struct addr_location al;
385 struct machine *machine = &kmem_session->machines.host;
386 struct callchain_cursor_node *node;
388 if (alloc_func_list == NULL) {
389 if (build_alloc_func_list() < 0)
393 al.thread = machine__findnew_thread(machine, sample->pid, sample->tid);
394 sample__resolve_callchain(sample, &callchain_cursor, NULL, evsel, &al, 16);
396 callchain_cursor_commit(&callchain_cursor);
398 struct alloc_func key, *caller;
401 node = callchain_cursor_current(&callchain_cursor);
405 key.start = key.end = node->ip;
406 caller = bsearch(&key, alloc_func_list, nr_alloc_funcs,
407 sizeof(key), callcmp);
411 addr = map__unmap_ip(node->map, node->ip);
417 pr_debug3("skipping alloc function: %s\n", caller->name);
419 callchain_cursor_advance(&callchain_cursor);
423 pr_debug2("unknown callsite: %"PRIx64 "\n", sample->ip);
427 struct sort_dimension {
430 struct list_head list;
433 static LIST_HEAD(page_alloc_sort_input);
434 static LIST_HEAD(page_caller_sort_input);
436 static struct page_stat *
437 __page_stat__findnew_page(struct page_stat *pstat, bool create)
439 struct rb_node **node = &page_live_tree.rb_node;
440 struct rb_node *parent = NULL;
441 struct page_stat *data;
447 data = rb_entry(*node, struct page_stat, node);
449 cmp = data->page - pstat->page;
451 node = &parent->rb_left;
453 node = &parent->rb_right;
461 data = zalloc(sizeof(*data));
463 data->page = pstat->page;
464 data->order = pstat->order;
465 data->gfp_flags = pstat->gfp_flags;
466 data->migrate_type = pstat->migrate_type;
468 rb_link_node(&data->node, parent, node);
469 rb_insert_color(&data->node, &page_live_tree);
475 static struct page_stat *page_stat__find_page(struct page_stat *pstat)
477 return __page_stat__findnew_page(pstat, false);
480 static struct page_stat *page_stat__findnew_page(struct page_stat *pstat)
482 return __page_stat__findnew_page(pstat, true);
485 static struct page_stat *
486 __page_stat__findnew_alloc(struct page_stat *pstat, bool create)
488 struct rb_node **node = &page_alloc_tree.rb_node;
489 struct rb_node *parent = NULL;
490 struct page_stat *data;
491 struct sort_dimension *sort;
497 data = rb_entry(*node, struct page_stat, node);
499 list_for_each_entry(sort, &page_alloc_sort_input, list) {
500 cmp = sort->cmp(pstat, data);
506 node = &parent->rb_left;
508 node = &parent->rb_right;
516 data = zalloc(sizeof(*data));
518 data->page = pstat->page;
519 data->order = pstat->order;
520 data->gfp_flags = pstat->gfp_flags;
521 data->migrate_type = pstat->migrate_type;
523 rb_link_node(&data->node, parent, node);
524 rb_insert_color(&data->node, &page_alloc_tree);
530 static struct page_stat *page_stat__find_alloc(struct page_stat *pstat)
532 return __page_stat__findnew_alloc(pstat, false);
535 static struct page_stat *page_stat__findnew_alloc(struct page_stat *pstat)
537 return __page_stat__findnew_alloc(pstat, true);
540 static struct page_stat *
541 __page_stat__findnew_caller(struct page_stat *pstat, bool create)
543 struct rb_node **node = &page_caller_tree.rb_node;
544 struct rb_node *parent = NULL;
545 struct page_stat *data;
546 struct sort_dimension *sort;
552 data = rb_entry(*node, struct page_stat, node);
554 list_for_each_entry(sort, &page_caller_sort_input, list) {
555 cmp = sort->cmp(pstat, data);
561 node = &parent->rb_left;
563 node = &parent->rb_right;
571 data = zalloc(sizeof(*data));
573 data->callsite = pstat->callsite;
574 data->order = pstat->order;
575 data->gfp_flags = pstat->gfp_flags;
576 data->migrate_type = pstat->migrate_type;
578 rb_link_node(&data->node, parent, node);
579 rb_insert_color(&data->node, &page_caller_tree);
585 static struct page_stat *page_stat__find_caller(struct page_stat *pstat)
587 return __page_stat__findnew_caller(pstat, false);
590 static struct page_stat *page_stat__findnew_caller(struct page_stat *pstat)
592 return __page_stat__findnew_caller(pstat, true);
595 static bool valid_page(u64 pfn_or_page)
597 if (use_pfn && pfn_or_page == -1UL)
599 if (!use_pfn && pfn_or_page == 0)
607 char *human_readable;
610 static struct gfp_flag *gfps;
613 static int gfpcmp(const void *a, const void *b)
615 const struct gfp_flag *fa = a;
616 const struct gfp_flag *fb = b;
618 return fa->flags - fb->flags;
621 /* see include/trace/events/mmflags.h */
622 static const struct {
623 const char *original;
625 } gfp_compact_table[] = {
626 { "GFP_TRANSHUGE", "THP" },
627 { "GFP_TRANSHUGE_LIGHT", "THL" },
628 { "GFP_HIGHUSER_MOVABLE", "HUM" },
629 { "GFP_HIGHUSER", "HU" },
631 { "GFP_KERNEL_ACCOUNT", "KAC" },
632 { "GFP_KERNEL", "K" },
633 { "GFP_NOFS", "NF" },
634 { "GFP_ATOMIC", "A" },
635 { "GFP_NOIO", "NI" },
636 { "GFP_NOWAIT", "NW" },
638 { "__GFP_HIGHMEM", "HM" },
639 { "GFP_DMA32", "D32" },
640 { "__GFP_HIGH", "H" },
641 { "__GFP_ATOMIC", "_A" },
644 { "__GFP_COLD", "CO" },
645 { "__GFP_NOWARN", "NWR" },
646 { "__GFP_RETRY_MAYFAIL", "R" },
647 { "__GFP_NOFAIL", "NF" },
648 { "__GFP_NORETRY", "NR" },
649 { "__GFP_COMP", "C" },
650 { "__GFP_ZERO", "Z" },
651 { "__GFP_NOMEMALLOC", "NMA" },
652 { "__GFP_MEMALLOC", "MA" },
653 { "__GFP_HARDWALL", "HW" },
654 { "__GFP_THISNODE", "TN" },
655 { "__GFP_RECLAIMABLE", "RC" },
656 { "__GFP_MOVABLE", "M" },
657 { "__GFP_ACCOUNT", "AC" },
658 { "__GFP_NOTRACK", "NT" },
659 { "__GFP_WRITE", "WR" },
660 { "__GFP_RECLAIM", "R" },
661 { "__GFP_DIRECT_RECLAIM", "DR" },
662 { "__GFP_KSWAPD_RECLAIM", "KR" },
665 static size_t max_gfp_len;
667 static char *compact_gfp_flags(char *gfp_flags)
669 char *orig_flags = strdup(gfp_flags);
670 char *new_flags = NULL;
671 char *str, *pos = NULL;
674 if (orig_flags == NULL)
677 str = strtok_r(orig_flags, "|", &pos);
683 for (i = 0; i < ARRAY_SIZE(gfp_compact_table); i++) {
684 if (strcmp(gfp_compact_table[i].original, str))
687 cpt = gfp_compact_table[i].compact;
688 new = realloc(new_flags, len + strlen(cpt) + 2);
697 strcpy(new_flags, cpt);
699 strcat(new_flags, "|");
700 strcat(new_flags, cpt);
707 str = strtok_r(NULL, "|", &pos);
710 if (max_gfp_len < len)
717 static char *compact_gfp_string(unsigned long gfp_flags)
719 struct gfp_flag key = {
722 struct gfp_flag *gfp;
724 gfp = bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp);
726 return gfp->compact_str;
731 static int parse_gfp_flags(struct perf_evsel *evsel, struct perf_sample *sample,
732 unsigned int gfp_flags)
734 struct pevent_record record = {
736 .data = sample->raw_data,
737 .size = sample->raw_size,
739 struct trace_seq seq;
740 char *str, *pos = NULL;
743 struct gfp_flag key = {
747 if (bsearch(&key, gfps, nr_gfps, sizeof(*gfps), gfpcmp))
751 trace_seq_init(&seq);
752 pevent_event_info(&seq, evsel->tp_format, &record);
754 str = strtok_r(seq.buffer, " ", &pos);
756 if (!strncmp(str, "gfp_flags=", 10)) {
757 struct gfp_flag *new;
759 new = realloc(gfps, (nr_gfps + 1) * sizeof(*gfps));
766 new->flags = gfp_flags;
767 new->human_readable = strdup(str + 10);
768 new->compact_str = compact_gfp_flags(str + 10);
769 if (!new->human_readable || !new->compact_str)
772 qsort(gfps, nr_gfps, sizeof(*gfps), gfpcmp);
775 str = strtok_r(NULL, " ", &pos);
778 trace_seq_destroy(&seq);
782 static int perf_evsel__process_page_alloc_event(struct perf_evsel *evsel,
783 struct perf_sample *sample)
786 unsigned int order = perf_evsel__intval(evsel, sample, "order");
787 unsigned int gfp_flags = perf_evsel__intval(evsel, sample, "gfp_flags");
788 unsigned int migrate_type = perf_evsel__intval(evsel, sample,
790 u64 bytes = kmem_page_size << order;
792 struct page_stat *pstat;
793 struct page_stat this = {
795 .gfp_flags = gfp_flags,
796 .migrate_type = migrate_type,
800 page = perf_evsel__intval(evsel, sample, "pfn");
802 page = perf_evsel__intval(evsel, sample, "page");
805 total_page_alloc_bytes += bytes;
807 if (!valid_page(page)) {
809 total_page_fail_bytes += bytes;
814 if (parse_gfp_flags(evsel, sample, gfp_flags) < 0)
817 callsite = find_callsite(evsel, sample);
820 * This is to find the current page (with correct gfp flags and
821 * migrate type) at free event.
824 pstat = page_stat__findnew_page(&this);
829 pstat->alloc_bytes += bytes;
830 pstat->callsite = callsite;
833 pstat = page_stat__findnew_alloc(&this);
838 pstat->alloc_bytes += bytes;
839 pstat->callsite = callsite;
842 this.callsite = callsite;
843 pstat = page_stat__findnew_caller(&this);
848 pstat->alloc_bytes += bytes;
850 order_stats[order][migrate_type]++;
855 static int perf_evsel__process_page_free_event(struct perf_evsel *evsel,
856 struct perf_sample *sample)
859 unsigned int order = perf_evsel__intval(evsel, sample, "order");
860 u64 bytes = kmem_page_size << order;
861 struct page_stat *pstat;
862 struct page_stat this = {
867 page = perf_evsel__intval(evsel, sample, "pfn");
869 page = perf_evsel__intval(evsel, sample, "page");
872 total_page_free_bytes += bytes;
875 pstat = page_stat__find_page(&this);
877 pr_debug2("missing free at page %"PRIx64" (order: %d)\n",
881 total_page_nomatch_bytes += bytes;
886 this.gfp_flags = pstat->gfp_flags;
887 this.migrate_type = pstat->migrate_type;
888 this.callsite = pstat->callsite;
890 rb_erase(&pstat->node, &page_live_tree);
894 order_stats[this.order][this.migrate_type]--;
896 pstat = page_stat__find_alloc(&this);
901 pstat->free_bytes += bytes;
904 pstat = page_stat__find_caller(&this);
909 pstat->free_bytes += bytes;
913 pstat->alloc_bytes -= bytes;
915 if (pstat->nr_alloc == 0) {
916 rb_erase(&pstat->node, &page_caller_tree);
924 static bool perf_kmem__skip_sample(struct perf_sample *sample)
926 /* skip sample based on time? */
927 if (perf_time__skip_sample(&ptime, sample->time))
933 typedef int (*tracepoint_handler)(struct perf_evsel *evsel,
934 struct perf_sample *sample);
936 static int process_sample_event(struct perf_tool *tool __maybe_unused,
937 union perf_event *event,
938 struct perf_sample *sample,
939 struct perf_evsel *evsel,
940 struct machine *machine)
943 struct thread *thread = machine__findnew_thread(machine, sample->pid,
946 if (thread == NULL) {
947 pr_debug("problem processing %d event, skipping it.\n",
952 if (perf_kmem__skip_sample(sample))
955 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid);
957 if (evsel->handler != NULL) {
958 tracepoint_handler f = evsel->handler;
959 err = f(evsel, sample);
967 static struct perf_tool perf_kmem = {
968 .sample = process_sample_event,
969 .comm = perf_event__process_comm,
970 .mmap = perf_event__process_mmap,
971 .mmap2 = perf_event__process_mmap2,
972 .namespaces = perf_event__process_namespaces,
973 .ordered_events = true,
976 static double fragmentation(unsigned long n_req, unsigned long n_alloc)
981 return 100.0 - (100.0 * n_req / n_alloc);
984 static void __print_slab_result(struct rb_root *root,
985 struct perf_session *session,
986 int n_lines, int is_caller)
988 struct rb_node *next;
989 struct machine *machine = &session->machines.host;
991 printf("%.105s\n", graph_dotted_line);
992 printf(" %-34s |", is_caller ? "Callsite": "Alloc Ptr");
993 printf(" Total_alloc/Per | Total_req/Per | Hit | Ping-pong | Frag\n");
994 printf("%.105s\n", graph_dotted_line);
996 next = rb_first(root);
998 while (next && n_lines--) {
999 struct alloc_stat *data = rb_entry(next, struct alloc_stat,
1001 struct symbol *sym = NULL;
1007 addr = data->call_site;
1009 sym = machine__find_kernel_function(machine, addr, &map);
1014 snprintf(buf, sizeof(buf), "%s+%" PRIx64 "", sym->name,
1015 addr - map->unmap_ip(map, sym->start));
1017 snprintf(buf, sizeof(buf), "%#" PRIx64 "", addr);
1018 printf(" %-34s |", buf);
1020 printf(" %9llu/%-5lu | %9llu/%-5lu | %8lu | %9lu | %6.3f%%\n",
1021 (unsigned long long)data->bytes_alloc,
1022 (unsigned long)data->bytes_alloc / data->hit,
1023 (unsigned long long)data->bytes_req,
1024 (unsigned long)data->bytes_req / data->hit,
1025 (unsigned long)data->hit,
1026 (unsigned long)data->pingpong,
1027 fragmentation(data->bytes_req, data->bytes_alloc));
1029 next = rb_next(next);
1033 printf(" ... | ... | ... | ... | ... | ... \n");
1035 printf("%.105s\n", graph_dotted_line);
1038 static const char * const migrate_type_str[] = {
1047 static void __print_page_alloc_result(struct perf_session *session, int n_lines)
1049 struct rb_node *next = rb_first(&page_alloc_sorted);
1050 struct machine *machine = &session->machines.host;
1052 int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1054 printf("\n%.105s\n", graph_dotted_line);
1055 printf(" %-16s | %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n",
1056 use_pfn ? "PFN" : "Page", live_page ? "Live" : "Total",
1057 gfp_len, "GFP flags");
1058 printf("%.105s\n", graph_dotted_line);
1061 format = " %16llu | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1063 format = " %016llx | %'16llu | %'9d | %5d | %8s | %-*s | %s\n";
1065 while (next && n_lines--) {
1066 struct page_stat *data;
1072 data = rb_entry(next, struct page_stat, node);
1073 sym = machine__find_kernel_function(machine, data->callsite, &map);
1077 scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1079 printf(format, (unsigned long long)data->page,
1080 (unsigned long long)data->alloc_bytes / 1024,
1081 data->nr_alloc, data->order,
1082 migrate_type_str[data->migrate_type],
1083 gfp_len, compact_gfp_string(data->gfp_flags), caller);
1085 next = rb_next(next);
1088 if (n_lines == -1) {
1089 printf(" ... | ... | ... | ... | ... | %-*s | ...\n",
1093 printf("%.105s\n", graph_dotted_line);
1096 static void __print_page_caller_result(struct perf_session *session, int n_lines)
1098 struct rb_node *next = rb_first(&page_caller_sorted);
1099 struct machine *machine = &session->machines.host;
1100 int gfp_len = max(strlen("GFP flags"), max_gfp_len);
1102 printf("\n%.105s\n", graph_dotted_line);
1103 printf(" %5s alloc (KB) | Hits | Order | Mig.type | %-*s | Callsite\n",
1104 live_page ? "Live" : "Total", gfp_len, "GFP flags");
1105 printf("%.105s\n", graph_dotted_line);
1107 while (next && n_lines--) {
1108 struct page_stat *data;
1114 data = rb_entry(next, struct page_stat, node);
1115 sym = machine__find_kernel_function(machine, data->callsite, &map);
1119 scnprintf(buf, sizeof(buf), "%"PRIx64, data->callsite);
1121 printf(" %'16llu | %'9d | %5d | %8s | %-*s | %s\n",
1122 (unsigned long long)data->alloc_bytes / 1024,
1123 data->nr_alloc, data->order,
1124 migrate_type_str[data->migrate_type],
1125 gfp_len, compact_gfp_string(data->gfp_flags), caller);
1127 next = rb_next(next);
1130 if (n_lines == -1) {
1131 printf(" ... | ... | ... | ... | %-*s | ...\n",
1135 printf("%.105s\n", graph_dotted_line);
1138 static void print_gfp_flags(void)
1143 printf("# GFP flags\n");
1144 printf("# ---------\n");
1145 for (i = 0; i < nr_gfps; i++) {
1146 printf("# %08x: %*s: %s\n", gfps[i].flags,
1147 (int) max_gfp_len, gfps[i].compact_str,
1148 gfps[i].human_readable);
1152 static void print_slab_summary(void)
1154 printf("\nSUMMARY (SLAB allocator)");
1155 printf("\n========================\n");
1156 printf("Total bytes requested: %'lu\n", total_requested);
1157 printf("Total bytes allocated: %'lu\n", total_allocated);
1158 printf("Total bytes freed: %'lu\n", total_freed);
1159 if (total_allocated > total_freed) {
1160 printf("Net total bytes allocated: %'lu\n",
1161 total_allocated - total_freed);
1163 printf("Total bytes wasted on internal fragmentation: %'lu\n",
1164 total_allocated - total_requested);
1165 printf("Internal fragmentation: %f%%\n",
1166 fragmentation(total_requested, total_allocated));
1167 printf("Cross CPU allocations: %'lu/%'lu\n", nr_cross_allocs, nr_allocs);
1170 static void print_page_summary(void)
1173 u64 nr_alloc_freed = nr_page_frees - nr_page_nomatch;
1174 u64 total_alloc_freed_bytes = total_page_free_bytes - total_page_nomatch_bytes;
1176 printf("\nSUMMARY (page allocator)");
1177 printf("\n========================\n");
1178 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation requests",
1179 nr_page_allocs, total_page_alloc_bytes / 1024);
1180 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free requests",
1181 nr_page_frees, total_page_free_bytes / 1024);
1184 printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc+freed requests",
1185 nr_alloc_freed, (total_alloc_freed_bytes) / 1024);
1186 printf("%-30s: %'16"PRIu64" [ %'16"PRIu64" KB ]\n", "Total alloc-only requests",
1187 nr_page_allocs - nr_alloc_freed,
1188 (total_page_alloc_bytes - total_alloc_freed_bytes) / 1024);
1189 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total free-only requests",
1190 nr_page_nomatch, total_page_nomatch_bytes / 1024);
1193 printf("%-30s: %'16lu [ %'16"PRIu64" KB ]\n", "Total allocation failures",
1194 nr_page_fails, total_page_fail_bytes / 1024);
1197 printf("%5s %12s %12s %12s %12s %12s\n", "Order", "Unmovable",
1198 "Reclaimable", "Movable", "Reserved", "CMA/Isolated");
1199 printf("%.5s %.12s %.12s %.12s %.12s %.12s\n", graph_dotted_line,
1200 graph_dotted_line, graph_dotted_line, graph_dotted_line,
1201 graph_dotted_line, graph_dotted_line);
1203 for (o = 0; o < MAX_PAGE_ORDER; o++) {
1205 for (m = 0; m < MAX_MIGRATE_TYPES - 1; m++) {
1206 if (order_stats[o][m])
1207 printf(" %'12d", order_stats[o][m]);
1209 printf(" %12c", '.');
1215 static void print_slab_result(struct perf_session *session)
1218 __print_slab_result(&root_caller_sorted, session, caller_lines, 1);
1220 __print_slab_result(&root_alloc_sorted, session, alloc_lines, 0);
1221 print_slab_summary();
1224 static void print_page_result(struct perf_session *session)
1226 if (caller_flag || alloc_flag)
1229 __print_page_caller_result(session, caller_lines);
1231 __print_page_alloc_result(session, alloc_lines);
1232 print_page_summary();
1235 static void print_result(struct perf_session *session)
1238 print_slab_result(session);
1240 print_page_result(session);
1243 static LIST_HEAD(slab_caller_sort);
1244 static LIST_HEAD(slab_alloc_sort);
1245 static LIST_HEAD(page_caller_sort);
1246 static LIST_HEAD(page_alloc_sort);
1248 static void sort_slab_insert(struct rb_root *root, struct alloc_stat *data,
1249 struct list_head *sort_list)
1251 struct rb_node **new = &(root->rb_node);
1252 struct rb_node *parent = NULL;
1253 struct sort_dimension *sort;
1256 struct alloc_stat *this;
1259 this = rb_entry(*new, struct alloc_stat, node);
1262 list_for_each_entry(sort, sort_list, list) {
1263 cmp = sort->cmp(data, this);
1269 new = &((*new)->rb_left);
1271 new = &((*new)->rb_right);
1274 rb_link_node(&data->node, parent, new);
1275 rb_insert_color(&data->node, root);
1278 static void __sort_slab_result(struct rb_root *root, struct rb_root *root_sorted,
1279 struct list_head *sort_list)
1281 struct rb_node *node;
1282 struct alloc_stat *data;
1285 node = rb_first(root);
1289 rb_erase(node, root);
1290 data = rb_entry(node, struct alloc_stat, node);
1291 sort_slab_insert(root_sorted, data, sort_list);
1295 static void sort_page_insert(struct rb_root *root, struct page_stat *data,
1296 struct list_head *sort_list)
1298 struct rb_node **new = &root->rb_node;
1299 struct rb_node *parent = NULL;
1300 struct sort_dimension *sort;
1303 struct page_stat *this;
1306 this = rb_entry(*new, struct page_stat, node);
1309 list_for_each_entry(sort, sort_list, list) {
1310 cmp = sort->cmp(data, this);
1316 new = &parent->rb_left;
1318 new = &parent->rb_right;
1321 rb_link_node(&data->node, parent, new);
1322 rb_insert_color(&data->node, root);
1325 static void __sort_page_result(struct rb_root *root, struct rb_root *root_sorted,
1326 struct list_head *sort_list)
1328 struct rb_node *node;
1329 struct page_stat *data;
1332 node = rb_first(root);
1336 rb_erase(node, root);
1337 data = rb_entry(node, struct page_stat, node);
1338 sort_page_insert(root_sorted, data, sort_list);
1342 static void sort_result(void)
1345 __sort_slab_result(&root_alloc_stat, &root_alloc_sorted,
1347 __sort_slab_result(&root_caller_stat, &root_caller_sorted,
1352 __sort_page_result(&page_live_tree, &page_alloc_sorted,
1355 __sort_page_result(&page_alloc_tree, &page_alloc_sorted,
1358 __sort_page_result(&page_caller_tree, &page_caller_sorted,
1363 static int __cmd_kmem(struct perf_session *session)
1366 struct perf_evsel *evsel;
1367 const struct perf_evsel_str_handler kmem_tracepoints[] = {
1368 /* slab allocator */
1369 { "kmem:kmalloc", perf_evsel__process_alloc_event, },
1370 { "kmem:kmem_cache_alloc", perf_evsel__process_alloc_event, },
1371 { "kmem:kmalloc_node", perf_evsel__process_alloc_node_event, },
1372 { "kmem:kmem_cache_alloc_node", perf_evsel__process_alloc_node_event, },
1373 { "kmem:kfree", perf_evsel__process_free_event, },
1374 { "kmem:kmem_cache_free", perf_evsel__process_free_event, },
1375 /* page allocator */
1376 { "kmem:mm_page_alloc", perf_evsel__process_page_alloc_event, },
1377 { "kmem:mm_page_free", perf_evsel__process_page_free_event, },
1380 if (!perf_session__has_traces(session, "kmem record"))
1383 if (perf_session__set_tracepoints_handlers(session, kmem_tracepoints)) {
1384 pr_err("Initializing perf session tracepoint handlers failed\n");
1388 evlist__for_each_entry(session->evlist, evsel) {
1389 if (!strcmp(perf_evsel__name(evsel), "kmem:mm_page_alloc") &&
1390 perf_evsel__field(evsel, "pfn")) {
1397 err = perf_session__process_events(session);
1399 pr_err("error during process events: %d\n", err);
1403 print_result(session);
1408 /* slab sort keys */
1409 static int ptr_cmp(void *a, void *b)
1411 struct alloc_stat *l = a;
1412 struct alloc_stat *r = b;
1414 if (l->ptr < r->ptr)
1416 else if (l->ptr > r->ptr)
1421 static struct sort_dimension ptr_sort_dimension = {
1426 static int slab_callsite_cmp(void *a, void *b)
1428 struct alloc_stat *l = a;
1429 struct alloc_stat *r = b;
1431 if (l->call_site < r->call_site)
1433 else if (l->call_site > r->call_site)
1438 static struct sort_dimension callsite_sort_dimension = {
1440 .cmp = slab_callsite_cmp,
1443 static int hit_cmp(void *a, void *b)
1445 struct alloc_stat *l = a;
1446 struct alloc_stat *r = b;
1448 if (l->hit < r->hit)
1450 else if (l->hit > r->hit)
1455 static struct sort_dimension hit_sort_dimension = {
1460 static int bytes_cmp(void *a, void *b)
1462 struct alloc_stat *l = a;
1463 struct alloc_stat *r = b;
1465 if (l->bytes_alloc < r->bytes_alloc)
1467 else if (l->bytes_alloc > r->bytes_alloc)
1472 static struct sort_dimension bytes_sort_dimension = {
1477 static int frag_cmp(void *a, void *b)
1480 struct alloc_stat *l = a;
1481 struct alloc_stat *r = b;
1483 x = fragmentation(l->bytes_req, l->bytes_alloc);
1484 y = fragmentation(r->bytes_req, r->bytes_alloc);
1493 static struct sort_dimension frag_sort_dimension = {
1498 static int pingpong_cmp(void *a, void *b)
1500 struct alloc_stat *l = a;
1501 struct alloc_stat *r = b;
1503 if (l->pingpong < r->pingpong)
1505 else if (l->pingpong > r->pingpong)
1510 static struct sort_dimension pingpong_sort_dimension = {
1512 .cmp = pingpong_cmp,
1515 /* page sort keys */
1516 static int page_cmp(void *a, void *b)
1518 struct page_stat *l = a;
1519 struct page_stat *r = b;
1521 if (l->page < r->page)
1523 else if (l->page > r->page)
1528 static struct sort_dimension page_sort_dimension = {
1533 static int page_callsite_cmp(void *a, void *b)
1535 struct page_stat *l = a;
1536 struct page_stat *r = b;
1538 if (l->callsite < r->callsite)
1540 else if (l->callsite > r->callsite)
1545 static struct sort_dimension page_callsite_sort_dimension = {
1547 .cmp = page_callsite_cmp,
1550 static int page_hit_cmp(void *a, void *b)
1552 struct page_stat *l = a;
1553 struct page_stat *r = b;
1555 if (l->nr_alloc < r->nr_alloc)
1557 else if (l->nr_alloc > r->nr_alloc)
1562 static struct sort_dimension page_hit_sort_dimension = {
1564 .cmp = page_hit_cmp,
1567 static int page_bytes_cmp(void *a, void *b)
1569 struct page_stat *l = a;
1570 struct page_stat *r = b;
1572 if (l->alloc_bytes < r->alloc_bytes)
1574 else if (l->alloc_bytes > r->alloc_bytes)
1579 static struct sort_dimension page_bytes_sort_dimension = {
1581 .cmp = page_bytes_cmp,
1584 static int page_order_cmp(void *a, void *b)
1586 struct page_stat *l = a;
1587 struct page_stat *r = b;
1589 if (l->order < r->order)
1591 else if (l->order > r->order)
1596 static struct sort_dimension page_order_sort_dimension = {
1598 .cmp = page_order_cmp,
1601 static int migrate_type_cmp(void *a, void *b)
1603 struct page_stat *l = a;
1604 struct page_stat *r = b;
1606 /* for internal use to find free'd page */
1607 if (l->migrate_type == -1U)
1610 if (l->migrate_type < r->migrate_type)
1612 else if (l->migrate_type > r->migrate_type)
1617 static struct sort_dimension migrate_type_sort_dimension = {
1619 .cmp = migrate_type_cmp,
1622 static int gfp_flags_cmp(void *a, void *b)
1624 struct page_stat *l = a;
1625 struct page_stat *r = b;
1627 /* for internal use to find free'd page */
1628 if (l->gfp_flags == -1U)
1631 if (l->gfp_flags < r->gfp_flags)
1633 else if (l->gfp_flags > r->gfp_flags)
1638 static struct sort_dimension gfp_flags_sort_dimension = {
1640 .cmp = gfp_flags_cmp,
1643 static struct sort_dimension *slab_sorts[] = {
1644 &ptr_sort_dimension,
1645 &callsite_sort_dimension,
1646 &hit_sort_dimension,
1647 &bytes_sort_dimension,
1648 &frag_sort_dimension,
1649 &pingpong_sort_dimension,
1652 static struct sort_dimension *page_sorts[] = {
1653 &page_sort_dimension,
1654 &page_callsite_sort_dimension,
1655 &page_hit_sort_dimension,
1656 &page_bytes_sort_dimension,
1657 &page_order_sort_dimension,
1658 &migrate_type_sort_dimension,
1659 &gfp_flags_sort_dimension,
1662 static int slab_sort_dimension__add(const char *tok, struct list_head *list)
1664 struct sort_dimension *sort;
1667 for (i = 0; i < (int)ARRAY_SIZE(slab_sorts); i++) {
1668 if (!strcmp(slab_sorts[i]->name, tok)) {
1669 sort = memdup(slab_sorts[i], sizeof(*slab_sorts[i]));
1671 pr_err("%s: memdup failed\n", __func__);
1674 list_add_tail(&sort->list, list);
1682 static int page_sort_dimension__add(const char *tok, struct list_head *list)
1684 struct sort_dimension *sort;
1687 for (i = 0; i < (int)ARRAY_SIZE(page_sorts); i++) {
1688 if (!strcmp(page_sorts[i]->name, tok)) {
1689 sort = memdup(page_sorts[i], sizeof(*page_sorts[i]));
1691 pr_err("%s: memdup failed\n", __func__);
1694 list_add_tail(&sort->list, list);
1702 static int setup_slab_sorting(struct list_head *sort_list, const char *arg)
1705 char *str = strdup(arg);
1709 pr_err("%s: strdup failed\n", __func__);
1714 tok = strsep(&pos, ",");
1717 if (slab_sort_dimension__add(tok, sort_list) < 0) {
1718 pr_err("Unknown slab --sort key: '%s'", tok);
1728 static int setup_page_sorting(struct list_head *sort_list, const char *arg)
1731 char *str = strdup(arg);
1735 pr_err("%s: strdup failed\n", __func__);
1740 tok = strsep(&pos, ",");
1743 if (page_sort_dimension__add(tok, sort_list) < 0) {
1744 pr_err("Unknown page --sort key: '%s'", tok);
1754 static int parse_sort_opt(const struct option *opt __maybe_unused,
1755 const char *arg, int unset __maybe_unused)
1760 if (kmem_page > kmem_slab ||
1761 (kmem_page == 0 && kmem_slab == 0 && kmem_default == KMEM_PAGE)) {
1762 if (caller_flag > alloc_flag)
1763 return setup_page_sorting(&page_caller_sort, arg);
1765 return setup_page_sorting(&page_alloc_sort, arg);
1767 if (caller_flag > alloc_flag)
1768 return setup_slab_sorting(&slab_caller_sort, arg);
1770 return setup_slab_sorting(&slab_alloc_sort, arg);
1776 static int parse_caller_opt(const struct option *opt __maybe_unused,
1777 const char *arg __maybe_unused,
1778 int unset __maybe_unused)
1780 caller_flag = (alloc_flag + 1);
1784 static int parse_alloc_opt(const struct option *opt __maybe_unused,
1785 const char *arg __maybe_unused,
1786 int unset __maybe_unused)
1788 alloc_flag = (caller_flag + 1);
1792 static int parse_slab_opt(const struct option *opt __maybe_unused,
1793 const char *arg __maybe_unused,
1794 int unset __maybe_unused)
1796 kmem_slab = (kmem_page + 1);
1800 static int parse_page_opt(const struct option *opt __maybe_unused,
1801 const char *arg __maybe_unused,
1802 int unset __maybe_unused)
1804 kmem_page = (kmem_slab + 1);
1808 static int parse_line_opt(const struct option *opt __maybe_unused,
1809 const char *arg, int unset __maybe_unused)
1816 lines = strtoul(arg, NULL, 10);
1818 if (caller_flag > alloc_flag)
1819 caller_lines = lines;
1821 alloc_lines = lines;
1826 static int __cmd_record(int argc, const char **argv)
1828 const char * const record_args[] = {
1829 "record", "-a", "-R", "-c", "1",
1831 const char * const slab_events[] = {
1832 "-e", "kmem:kmalloc",
1833 "-e", "kmem:kmalloc_node",
1835 "-e", "kmem:kmem_cache_alloc",
1836 "-e", "kmem:kmem_cache_alloc_node",
1837 "-e", "kmem:kmem_cache_free",
1839 const char * const page_events[] = {
1840 "-e", "kmem:mm_page_alloc",
1841 "-e", "kmem:mm_page_free",
1843 unsigned int rec_argc, i, j;
1844 const char **rec_argv;
1846 rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1848 rec_argc += ARRAY_SIZE(slab_events);
1850 rec_argc += ARRAY_SIZE(page_events) + 1; /* for -g */
1852 rec_argv = calloc(rec_argc + 1, sizeof(char *));
1854 if (rec_argv == NULL)
1857 for (i = 0; i < ARRAY_SIZE(record_args); i++)
1858 rec_argv[i] = strdup(record_args[i]);
1861 for (j = 0; j < ARRAY_SIZE(slab_events); j++, i++)
1862 rec_argv[i] = strdup(slab_events[j]);
1865 rec_argv[i++] = strdup("-g");
1867 for (j = 0; j < ARRAY_SIZE(page_events); j++, i++)
1868 rec_argv[i] = strdup(page_events[j]);
1871 for (j = 1; j < (unsigned int)argc; j++, i++)
1872 rec_argv[i] = argv[j];
1874 return cmd_record(i, rec_argv);
1877 static int kmem_config(const char *var, const char *value, void *cb __maybe_unused)
1879 if (!strcmp(var, "kmem.default")) {
1880 if (!strcmp(value, "slab"))
1881 kmem_default = KMEM_SLAB;
1882 else if (!strcmp(value, "page"))
1883 kmem_default = KMEM_PAGE;
1885 pr_err("invalid default value ('slab' or 'page' required): %s\n",
1893 int cmd_kmem(int argc, const char **argv)
1895 const char * const default_slab_sort = "frag,hit,bytes";
1896 const char * const default_page_sort = "bytes,hit";
1897 struct perf_data data = {
1898 .mode = PERF_DATA_MODE_READ,
1900 const struct option kmem_options[] = {
1901 OPT_STRING('i', "input", &input_name, "file", "input file name"),
1902 OPT_INCR('v', "verbose", &verbose,
1903 "be more verbose (show symbol address, etc)"),
1904 OPT_CALLBACK_NOOPT(0, "caller", NULL, NULL,
1905 "show per-callsite statistics", parse_caller_opt),
1906 OPT_CALLBACK_NOOPT(0, "alloc", NULL, NULL,
1907 "show per-allocation statistics", parse_alloc_opt),
1908 OPT_CALLBACK('s', "sort", NULL, "key[,key2...]",
1909 "sort by keys: ptr, callsite, bytes, hit, pingpong, frag, "
1910 "page, order, migtype, gfp", parse_sort_opt),
1911 OPT_CALLBACK('l', "line", NULL, "num", "show n lines", parse_line_opt),
1912 OPT_BOOLEAN(0, "raw-ip", &raw_ip, "show raw ip instead of symbol"),
1913 OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
1914 OPT_CALLBACK_NOOPT(0, "slab", NULL, NULL, "Analyze slab allocator",
1916 OPT_CALLBACK_NOOPT(0, "page", NULL, NULL, "Analyze page allocator",
1918 OPT_BOOLEAN(0, "live", &live_page, "Show live page stat"),
1919 OPT_STRING(0, "time", &time_str, "str",
1920 "Time span of interest (start,stop)"),
1923 const char *const kmem_subcommands[] = { "record", "stat", NULL };
1924 const char *kmem_usage[] = {
1928 struct perf_session *session;
1929 const char errmsg[] = "No %s allocation events found. Have you run 'perf kmem record --%s'?\n";
1930 int ret = perf_config(kmem_config, NULL);
1935 argc = parse_options_subcommand(argc, argv, kmem_options,
1936 kmem_subcommands, kmem_usage, 0);
1939 usage_with_options(kmem_usage, kmem_options);
1941 if (kmem_slab == 0 && kmem_page == 0) {
1942 if (kmem_default == KMEM_SLAB)
1948 if (!strncmp(argv[0], "rec", 3)) {
1950 return __cmd_record(argc, argv);
1953 data.file.path = input_name;
1955 kmem_session = session = perf_session__new(&data, false, &perf_kmem);
1956 if (session == NULL)
1962 if (!perf_evlist__find_tracepoint_by_name(session->evlist,
1964 pr_err(errmsg, "slab", "slab");
1970 struct perf_evsel *evsel;
1972 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
1973 "kmem:mm_page_alloc");
1974 if (evsel == NULL) {
1975 pr_err(errmsg, "page", "page");
1979 kmem_page_size = pevent_get_page_size(evsel->tp_format->pevent);
1980 symbol_conf.use_callchain = true;
1983 symbol__init(&session->header.env);
1985 if (perf_time__parse_str(&ptime, time_str) != 0) {
1986 pr_err("Invalid time string\n");
1991 if (!strcmp(argv[0], "stat")) {
1992 setlocale(LC_ALL, "");
1994 if (cpu__setup_cpunode_map())
1997 if (list_empty(&slab_caller_sort))
1998 setup_slab_sorting(&slab_caller_sort, default_slab_sort);
1999 if (list_empty(&slab_alloc_sort))
2000 setup_slab_sorting(&slab_alloc_sort, default_slab_sort);
2001 if (list_empty(&page_caller_sort))
2002 setup_page_sorting(&page_caller_sort, default_page_sort);
2003 if (list_empty(&page_alloc_sort))
2004 setup_page_sorting(&page_alloc_sort, default_page_sort);
2007 setup_page_sorting(&page_alloc_sort_input,
2008 "page,order,migtype,gfp");
2009 setup_page_sorting(&page_caller_sort_input,
2010 "callsite,order,migtype,gfp");
2012 ret = __cmd_kmem(session);
2014 usage_with_options(kmem_usage, kmem_options);
2017 perf_session__delete(session);