2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2005-2007, Joseph Koshy
5 * Copyright (c) 2007 The FreeBSD Foundation
8 * Portions of this software were developed by A. Joseph Koshy under
9 * sponsorship from the FreeBSD Foundation and Google, Inc.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * Transform a hwpmc(4) log into human readable form, and into
35 * gprof(1) compatible profiles.
38 #include <sys/cdefs.h>
39 __FBSDID("$FreeBSD$");
41 #include <sys/param.h>
42 #include <sys/endian.h>
44 #include <sys/imgact_aout.h>
45 #include <sys/imgact_elf.h>
48 #include <sys/queue.h>
49 #include <sys/socket.h>
53 #include <netinet/in.h>
75 #include "pmcstat_log.h"
76 #include "pmcstat_top.h"
77 #include "pmcpl_callgraph.h"
79 #define min(A,B) ((A) < (B) ? (A) : (B))
80 #define max(A,B) ((A) > (B) ? (A) : (B))
82 /* Get the sample value in percent related to nsamples. */
83 #define PMCPL_CG_COUNTP(a) \
84 ((a)->pcg_count * 100.0 / nsamples)
87 * The toplevel CG nodes (i.e., with rank == 0) are placed in a hash table.
90 struct pmcstat_cgnode_hash_list pmcstat_cgnode_hash[PMCSTAT_NHASH];
91 int pmcstat_cgnode_hash_count;
93 static pmcstat_interned_string pmcstat_previous_filename_printed;
95 static struct pmcstat_cgnode *
96 pmcstat_cgnode_allocate(struct pmcstat_image *image, uintfptr_t pc)
98 struct pmcstat_cgnode *cg;
100 if ((cg = malloc(sizeof(*cg))) == NULL)
101 err(EX_OSERR, "ERROR: Cannot allocate callgraph node");
103 cg->pcg_image = image;
107 cg->pcg_nchildren = 0;
108 LIST_INIT(&cg->pcg_children);
114 * Free a node and its children.
117 pmcstat_cgnode_free(struct pmcstat_cgnode *cg)
119 struct pmcstat_cgnode *cgc, *cgtmp;
121 LIST_FOREACH_SAFE(cgc, &cg->pcg_children, pcg_sibling, cgtmp)
122 pmcstat_cgnode_free(cgc);
127 * Look for a callgraph node associated with pmc `pmcid' in the global
128 * hash table that corresponds to the given `pc' value in the process
131 static struct pmcstat_cgnode *
132 pmcstat_cgnode_hash_lookup_pc(struct pmcstat_process *pp, pmc_id_t pmcid,
133 uintfptr_t pc, int usermode)
135 struct pmcstat_pcmap *ppm;
136 struct pmcstat_symbol *sym;
137 struct pmcstat_image *image;
138 struct pmcstat_cgnode *cg;
139 struct pmcstat_cgnode_hash *h;
140 uintfptr_t loadaddress;
141 unsigned int i, hash;
143 ppm = pmcstat_process_find_map(usermode ? pp : pmcstat_kernproc, pc);
147 image = ppm->ppm_image;
149 loadaddress = ppm->ppm_lowpc + image->pi_vaddr - image->pi_start;
150 pc -= loadaddress; /* Convert to an offset in the image. */
153 * Try determine the function at this offset. If we can't
154 * find a function round leave the `pc' value alone.
156 if (!(args.pa_flags & (FLAG_SKIP_TOP_FN_RES | FLAG_SHOW_OFFSET))) {
157 if ((sym = pmcstat_symbol_search(image, pc)) != NULL)
160 pmcstat_stats.ps_samples_unknown_function++;
163 for (hash = i = 0; i < sizeof(uintfptr_t); i++)
164 hash += (pc >> i) & 0xFF;
166 hash &= PMCSTAT_HASH_MASK;
169 LIST_FOREACH(h, &pmcstat_cgnode_hash[hash], pch_next)
171 if (h->pch_pmcid != pmcid)
178 if (cg->pcg_image == image && cg->pcg_func == pc)
183 * We haven't seen this (pmcid, pc) tuple yet, so allocate a
184 * new callgraph node and a new hash table entry for it.
186 cg = pmcstat_cgnode_allocate(image, pc);
187 if ((h = malloc(sizeof(*h))) == NULL)
188 err(EX_OSERR, "ERROR: Could not allocate callgraph node");
190 h->pch_pmcid = pmcid;
192 LIST_INSERT_HEAD(&pmcstat_cgnode_hash[hash], h, pch_next);
194 pmcstat_cgnode_hash_count++;
200 * Compare two callgraph nodes for sorting.
203 pmcstat_cgnode_compare(const void *a, const void *b)
205 const struct pmcstat_cgnode *const *pcg1, *const *pcg2, *cg1, *cg2;
207 pcg1 = (const struct pmcstat_cgnode *const *) a;
209 pcg2 = (const struct pmcstat_cgnode *const *) b;
212 /* Sort in reverse order */
213 if (cg1->pcg_count < cg2->pcg_count)
215 if (cg1->pcg_count > cg2->pcg_count)
221 * Find (allocating if a needed) a callgraph node in the given
222 * parent with the same (image, pcoffset) pair.
225 static struct pmcstat_cgnode *
226 pmcstat_cgnode_find(struct pmcstat_cgnode *parent, struct pmcstat_image *image,
229 struct pmcstat_cgnode *child;
231 LIST_FOREACH(child, &parent->pcg_children, pcg_sibling) {
232 if (child->pcg_image == image &&
233 child->pcg_func == pcoffset)
238 * Allocate a new structure.
241 child = pmcstat_cgnode_allocate(image, pcoffset);
244 * Link it into the parent.
246 LIST_INSERT_HEAD(&parent->pcg_children, child, pcg_sibling);
247 parent->pcg_nchildren++;
253 * Print one callgraph node. The output format is:
255 * indentation %(parent's samples) #nsamples function@object
258 pmcstat_cgnode_print(struct pmcstat_cgnode *cg, int depth, uint32_t total)
262 struct pmcstat_symbol *sym;
263 struct pmcstat_cgnode **sortbuffer, **cgn, *pcg;
268 (void) fprintf(args.pa_graphfile, "%*s", depth, space);
270 if (cg->pcg_count == total)
271 (void) fprintf(args.pa_graphfile, "100.0%% ");
273 (void) fprintf(args.pa_graphfile, "%05.2f%% ",
274 100.0 * cg->pcg_count / total);
276 n = fprintf(args.pa_graphfile, " [%u] ", cg->pcg_count);
278 /* #samples is a 12 character wide field. */
280 (void) fprintf(args.pa_graphfile, "%*s", 12 - n, space);
283 (void) fprintf(args.pa_graphfile, "%*s", depth, space);
285 sym = pmcstat_symbol_search(cg->pcg_image, cg->pcg_func);
287 (void) fprintf(args.pa_graphfile, "%s",
288 pmcstat_string_unintern(sym->ps_name));
290 (void) fprintf(args.pa_graphfile, "%p",
291 (void *) (cg->pcg_image->pi_vaddr + cg->pcg_func));
293 if (pmcstat_previous_filename_printed !=
294 cg->pcg_image->pi_fullpath) {
295 pmcstat_previous_filename_printed = cg->pcg_image->pi_fullpath;
296 (void) fprintf(args.pa_graphfile, " @ %s\n",
297 pmcstat_string_unintern(
298 pmcstat_previous_filename_printed));
300 (void) fprintf(args.pa_graphfile, "\n");
302 if (cg->pcg_nchildren == 0)
305 if ((sortbuffer = (struct pmcstat_cgnode **)
306 malloc(sizeof(struct pmcstat_cgnode *) *
307 cg->pcg_nchildren)) == NULL)
308 err(EX_OSERR, "ERROR: Cannot print callgraph");
311 LIST_FOREACH(pcg, &cg->pcg_children, pcg_sibling)
314 assert(cgn - sortbuffer == (int) cg->pcg_nchildren);
316 qsort(sortbuffer, cg->pcg_nchildren, sizeof(struct pmcstat_cgnode *),
317 pmcstat_cgnode_compare);
319 for (cgn = sortbuffer, n = 0; n < cg->pcg_nchildren; n++, cgn++)
320 pmcstat_cgnode_print(*cgn, depth+1, cg->pcg_count);
326 * Record a callchain.
330 pmcpl_cg_process(struct pmcstat_process *pp, struct pmcstat_pmcrecord *pmcr,
331 uint32_t nsamples, uintfptr_t *cc, int usermode, uint32_t cpu)
333 uintfptr_t pc, loadaddress;
335 struct pmcstat_image *image;
336 struct pmcstat_pcmap *ppm;
337 struct pmcstat_symbol *sym;
338 struct pmcstat_cgnode *parent, *child;
339 struct pmcstat_process *km;
345 * Find the callgraph node recorded in the global hash table
346 * for this (pmcid, pc).
350 pmcid = pmcr->pr_pmcid;
351 child = parent = pmcstat_cgnode_hash_lookup_pc(pp, pmcid, pc, usermode);
352 if (parent == NULL) {
353 pmcstat_stats.ps_callchain_dubious_frames++;
354 pmcr->pr_dubious_frames++;
361 * For each return address in the call chain record, subject
362 * to the maximum depth desired.
363 * - Find the image associated with the sample. Stop if there
364 * there is no valid image at that address.
365 * - Find the function that overlaps the return address.
366 * - If found: use the start address of the function.
367 * If not found (say an object's symbol table is not present or
368 * is incomplete), round down to th gprof bucket granularity.
369 * - Convert return virtual address to an offset in the image.
370 * - Look for a child with the same {offset,image} tuple,
371 * inserting one if needed.
372 * - Increment the count of occurrences of the child.
374 km = pmcstat_kernproc;
376 for (n = 1; n < (uint32_t) args.pa_graphdepth && n < nsamples; n++,
380 ppm = pmcstat_process_find_map(usermode ? pp : km, pc);
382 /* Detect full frame capture (kernel + user). */
384 ppm = pmcstat_process_find_map(pp, pc);
392 image = ppm->ppm_image;
393 loadaddress = ppm->ppm_lowpc + image->pi_vaddr -
397 if ((sym = pmcstat_symbol_search(image, pc)) != NULL)
400 child = pmcstat_cgnode_find(parent, image, pc);
406 * Printing a callgraph for a PMC.
409 pmcstat_callgraph_print_for_pmcid(struct pmcstat_pmcrecord *pmcr)
414 struct pmcstat_cgnode **sortbuffer, **cgn;
415 struct pmcstat_cgnode_hash *pch;
418 * We pull out all callgraph nodes in the top-level hash table
419 * with a matching PMC id. We then sort these based on the
420 * frequency of occurrence. Each callgraph node is then
425 pmcid = pmcr->pr_pmcid;
426 if ((sortbuffer = (struct pmcstat_cgnode **)
427 malloc(sizeof(struct pmcstat_cgnode *) *
428 pmcstat_cgnode_hash_count)) == NULL)
429 err(EX_OSERR, "ERROR: Cannot sort callgraph");
432 for (n = 0; n < PMCSTAT_NHASH; n++)
433 LIST_FOREACH(pch, &pmcstat_cgnode_hash[n], pch_next)
434 if (pch->pch_pmcid == pmcid) {
435 nsamples += pch->pch_cgnode->pcg_count;
436 *cgn++ = pch->pch_cgnode;
439 nentries = cgn - sortbuffer;
440 assert(nentries <= pmcstat_cgnode_hash_count);
447 qsort(sortbuffer, nentries, sizeof(struct pmcstat_cgnode *),
448 pmcstat_cgnode_compare);
450 (void) fprintf(args.pa_graphfile,
451 "@ %s [%u samples]\n\n",
452 pmcstat_string_unintern(pmcr->pr_pmcname),
455 for (cgn = sortbuffer, n = 0; n < nentries; n++, cgn++) {
456 pmcstat_previous_filename_printed = NULL;
457 pmcstat_cgnode_print(*cgn, 0, nsamples);
458 (void) fprintf(args.pa_graphfile, "\n");
465 * Print out callgraphs.
469 pmcstat_callgraph_print(void)
471 struct pmcstat_pmcrecord *pmcr;
473 LIST_FOREACH(pmcr, &pmcstat_pmcs, pr_next)
474 pmcstat_callgraph_print_for_pmcid(pmcr);
478 pmcstat_cgnode_topprint(struct pmcstat_cgnode *cg,
479 int depth __unused, uint32_t nsamples)
481 int v_attrs, vs_len, ns_len, width, len, n, nchildren;
484 struct pmcstat_symbol *sym;
485 struct pmcstat_cgnode **sortbuffer, **cgn, *pcg;
488 v = PMCPL_CG_COUNTP(cg);
489 snprintf(vs, sizeof(vs), "%.1f", v);
490 v_attrs = PMCSTAT_ATTRPERCENT(v);
493 sym = pmcstat_symbol_search(cg->pcg_image, cg->pcg_func);
495 snprintf(ns, sizeof(ns), "%p",
496 (void *)(cg->pcg_image->pi_vaddr + cg->pcg_func));
498 switch (args.pa_flags & (FLAG_SKIP_TOP_FN_RES | FLAG_SHOW_OFFSET)) {
499 case FLAG_SKIP_TOP_FN_RES | FLAG_SHOW_OFFSET:
500 case FLAG_SKIP_TOP_FN_RES:
501 snprintf(ns, sizeof(ns), "%p",
502 (void *)(cg->pcg_image->pi_vaddr + cg->pcg_func));
504 case FLAG_SHOW_OFFSET:
505 snprintf(ns, sizeof(ns), "%s+%#0" PRIx64,
506 pmcstat_string_unintern(sym->ps_name),
507 cg->pcg_func - sym->ps_start);
510 snprintf(ns, sizeof(ns), "%s",
511 pmcstat_string_unintern(sym->ps_name));
516 PMCSTAT_ATTRON(v_attrs);
517 PMCSTAT_PRINTW("%5.5s", vs);
518 PMCSTAT_ATTROFF(v_attrs);
519 PMCSTAT_PRINTW(" %-10.10s %-30.30s",
520 pmcstat_string_unintern(cg->pcg_image->pi_name),
523 nchildren = cg->pcg_nchildren;
524 if (nchildren == 0) {
525 PMCSTAT_PRINTW("\n");
529 width = pmcstat_displaywidth - 40;
531 if ((sortbuffer = (struct pmcstat_cgnode **)
532 malloc(sizeof(struct pmcstat_cgnode *) *
534 err(EX_OSERR, "ERROR: Cannot print callgraph");
537 LIST_FOREACH(pcg, &cg->pcg_children, pcg_sibling)
540 assert(cgn - sortbuffer == (int)nchildren);
542 qsort(sortbuffer, nchildren, sizeof(struct pmcstat_cgnode *),
543 pmcstat_cgnode_compare);
545 /* Count how many callers. */
546 for (cgn = sortbuffer, n = 0; n < nchildren; n++, cgn++) {
549 v = PMCPL_CG_COUNTP(pcg);
550 if (v < pmcstat_threshold)
555 for (cgn = sortbuffer, n = 0; n < nchildren; n++, cgn++) {
560 v = PMCPL_CG_COUNTP(pcg);
561 vs_len = snprintf(vs, sizeof(vs), ":%.1f", v);
562 v_attrs = PMCSTAT_ATTRPERCENT(v);
567 sym = pmcstat_symbol_search(pcg->pcg_image, pcg->pcg_func);
569 ns_len = snprintf(ns, sizeof(ns), "%s",
570 pmcstat_string_unintern(sym->ps_name));
572 ns_len = snprintf(ns, sizeof(ns), "%p",
573 (void *)pcg->pcg_func);
575 len = ns_len + vs_len + 1;
576 if (width - len < 0) {
577 PMCSTAT_PRINTW(" ...");
582 PMCSTAT_PRINTW(" %s", ns);
584 PMCSTAT_ATTRON(v_attrs);
585 PMCSTAT_PRINTW("%s", vs);
586 PMCSTAT_ATTROFF(v_attrs);
589 PMCSTAT_PRINTW("\n");
598 pmcpl_cg_topdisplay(void)
602 struct pmcstat_cgnode **sortbuffer, **cgn;
603 struct pmcstat_cgnode_hash *pch;
604 struct pmcstat_pmcrecord *pmcr;
606 pmcr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
608 err(EX_SOFTWARE, "ERROR: invalid pmcindex");
611 * We pull out all callgraph nodes in the top-level hash table
612 * with a matching PMC index. We then sort these based on the
613 * frequency of occurrence. Each callgraph node is then
619 if ((sortbuffer = (struct pmcstat_cgnode **)
620 malloc(sizeof(struct pmcstat_cgnode *) *
621 pmcstat_cgnode_hash_count)) == NULL)
622 err(EX_OSERR, "ERROR: Cannot sort callgraph");
625 for (n = 0; n < PMCSTAT_NHASH; n++)
626 LIST_FOREACH(pch, &pmcstat_cgnode_hash[n], pch_next)
627 if (pmcr == NULL || pch->pch_pmcid == pmcr->pr_pmcid) {
628 nsamples += pch->pch_cgnode->pcg_count;
629 *cgn++ = pch->pch_cgnode;
632 nentries = cgn - sortbuffer;
633 assert(nentries <= pmcstat_cgnode_hash_count);
640 qsort(sortbuffer, nentries, sizeof(struct pmcstat_cgnode *),
641 pmcstat_cgnode_compare);
643 PMCSTAT_PRINTW("%5.5s %-10.10s %-30.30s %s\n",
644 "%SAMP", "IMAGE", "FUNCTION", "CALLERS");
646 nentries = min(pmcstat_displayheight - 2, nentries);
648 for (cgn = sortbuffer, n = 0; n < nentries; n++, cgn++) {
649 if (PMCPL_CG_COUNTP(*cgn) < pmcstat_threshold)
651 pmcstat_cgnode_topprint(*cgn, 0, nsamples);
658 * Handle top mode keypress.
662 pmcpl_cg_topkeypress(int c, void *arg)
678 pmcstat_cgnode_hash_count = 0;
679 pmcstat_previous_filename_printed = NULL;
681 for (i = 0; i < PMCSTAT_NHASH; i++) {
682 LIST_INIT(&pmcstat_cgnode_hash[i]);
689 pmcpl_cg_shutdown(FILE *mf)
692 struct pmcstat_cgnode_hash *pch, *pchtmp;
696 if (args.pa_flags & FLAG_DO_CALLGRAPHS)
697 pmcstat_callgraph_print();
702 for (i = 0; i < PMCSTAT_NHASH; i++) {
703 LIST_FOREACH_SAFE(pch, &pmcstat_cgnode_hash[i], pch_next,
705 pmcstat_cgnode_free(pch->pch_cgnode);
706 LIST_REMOVE(pch, pch_next);