2 * Copyright (c) 2005-2007, Joseph Koshy
3 * Copyright (c) 2007 The FreeBSD Foundation
6 * Portions of this software were developed by A. Joseph Koshy under
7 * sponsorship from the FreeBSD Foundation and Google, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * Transform a hwpmc(4) log into human readable form, and into
33 * gprof(1) compatible profiles.
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
39 #include <sys/param.h>
40 #include <sys/endian.h>
41 #include <sys/cpuset.h>
43 #include <sys/imgact_aout.h>
44 #include <sys/imgact_elf.h>
47 #include <sys/queue.h>
48 #include <sys/socket.h>
52 #include <netinet/in.h>
73 #include "pmcstat_log.h"
74 #include "pmcstat_top.h"
76 #define PMCSTAT_ALLOCATE 1
81 * pmcstat_initialize_logging() initialize this module, called first
82 * pmcstat_shutdown_logging() orderly shutdown, called last
83 * pmcstat_open_log() open an eventlog for processing
84 * pmcstat_process_log() print/convert an event log
85 * pmcstat_display_log() top mode display for the log
86 * pmcstat_close_log() finish processing an event log
88 * IMPLEMENTATION NOTES
90 * We correlate each 'callchain' or 'sample' entry seen in the event
91 * log back to an executable object in the system. Executable objects
93 * - program executables,
94 * - shared libraries loaded by the runtime loader,
95 * - dlopen()'ed objects loaded by the program,
96 * - the runtime loader itself,
97 * - the kernel and kernel modules.
99 * Each process that we know about is treated as a set of regions that
100 * map to executable objects. Processes are described by
101 * 'pmcstat_process' structures. Executable objects are tracked by
102 * 'pmcstat_image' structures. The kernel and kernel modules are
103 * common to all processes (they reside at the same virtual addresses
104 * for all processes). Individual processes can have their text
105 * segments and shared libraries loaded at process-specific locations.
107 * A given executable object can be in use by multiple processes
108 * (e.g., libc.so) and loaded at a different address in each.
109 * pmcstat_pcmap structures track per-image mappings.
111 * The sample log could have samples from multiple PMCs; we
112 * generate one 'gmon.out' profile per PMC.
114 * IMPLEMENTATION OF GMON OUTPUT
116 * Each executable object gets one 'gmon.out' profile, per PMC in
117 * use. Creation of 'gmon.out' profiles is done lazily. The
118 * 'gmon.out' profiles generated for a given sampling PMC are
119 * aggregates of all the samples for that particular executable
122 * IMPLEMENTATION OF SYSTEM-WIDE CALLGRAPH OUTPUT
124 * Each active pmcid has its own callgraph structure, described by a
125 * 'struct pmcstat_callgraph'. Given a process id and a list of pc
126 * values, we map each pc value to a tuple (image, symbol), where
127 * 'image' denotes an executable object and 'symbol' is the closest
128 * symbol that precedes the pc value. Each pc value in the list is
129 * also given a 'rank' that reflects its depth in the call stack.
132 struct pmcstat_pmcs pmcstat_pmcs = LIST_HEAD_INITIALIZER(pmcstat_pmcs);
135 * All image descriptors are kept in a hash table.
137 struct pmcstat_image_hash_list pmcstat_image_hash[PMCSTAT_NHASH];
140 * All process descriptors are kept in a hash table.
142 struct pmcstat_process_hash_list pmcstat_process_hash[PMCSTAT_NHASH];
144 struct pmcstat_stats pmcstat_stats; /* statistics */
145 static int ps_samples_period; /* samples count between top refresh. */
147 struct pmcstat_process *pmcstat_kernproc; /* kernel 'process' */
149 #include "pmcpl_gprof.h"
150 #include "pmcpl_callgraph.h"
151 #include "pmcpl_annotate.h"
152 #include "pmcpl_annotate_cg.h"
153 #include "pmcpl_calltree.h"
155 static struct pmc_plugins {
156 const char *pl_name; /* name */
159 int (*pl_configure)(char *opt);
161 /* init and shutdown */
162 int (*pl_init)(void);
163 void (*pl_shutdown)(FILE *mf);
165 /* sample processing */
166 void (*pl_process)(struct pmcstat_process *pp,
167 struct pmcstat_pmcrecord *pmcr, uint32_t nsamples,
168 uintfptr_t *cc, int usermode, uint32_t cpu);
171 void (*pl_initimage)(struct pmcstat_image *pi);
172 void (*pl_shutdownimage)(struct pmcstat_image *pi);
175 void (*pl_newpmc)(pmcstat_interned_string ps,
176 struct pmcstat_pmcrecord *pr);
179 void (*pl_topdisplay)(void);
182 int (*pl_topkeypress)(int c, WINDOW *w);
189 .pl_name = "callgraph",
190 .pl_init = pmcpl_cg_init,
191 .pl_shutdown = pmcpl_cg_shutdown,
192 .pl_process = pmcpl_cg_process,
193 .pl_topkeypress = pmcpl_cg_topkeypress,
194 .pl_topdisplay = pmcpl_cg_topdisplay
198 .pl_shutdown = pmcpl_gmon_shutdown,
199 .pl_process = pmcpl_gmon_process,
200 .pl_initimage = pmcpl_gmon_initimage,
201 .pl_shutdownimage = pmcpl_gmon_shutdownimage,
202 .pl_newpmc = pmcpl_gmon_newpmc
205 .pl_name = "annotate",
206 .pl_process = pmcpl_annotate_process
209 .pl_name = "calltree",
210 .pl_configure = pmcpl_ct_configure,
211 .pl_init = pmcpl_ct_init,
212 .pl_shutdown = pmcpl_ct_shutdown,
213 .pl_process = pmcpl_ct_process,
214 .pl_topkeypress = pmcpl_ct_topkeypress,
215 .pl_topdisplay = pmcpl_ct_topdisplay
218 .pl_name = "annotate_cg",
219 .pl_process = pmcpl_annotate_cg_process
227 static int pmcstat_mergepmc;
229 int pmcstat_pmcinfilter = 0; /* PMC filter for top mode. */
230 float pmcstat_threshold = 0.5; /* Cost filter for top mode. */
236 static struct pmcstat_image *pmcstat_image_from_path(pmcstat_interned_string
237 _path, int _iskernelmodule);
238 static void pmcstat_image_get_aout_params(struct pmcstat_image *_image);
239 static void pmcstat_image_get_elf_params(struct pmcstat_image *_image);
240 static void pmcstat_image_link(struct pmcstat_process *_pp,
241 struct pmcstat_image *_i, uintfptr_t _lpc);
243 static void pmcstat_pmcid_add(pmc_id_t _pmcid,
244 pmcstat_interned_string _name);
246 static void pmcstat_process_aout_exec(struct pmcstat_process *_pp,
247 struct pmcstat_image *_image, uintfptr_t _entryaddr);
248 static void pmcstat_process_elf_exec(struct pmcstat_process *_pp,
249 struct pmcstat_image *_image, uintfptr_t _entryaddr);
250 static void pmcstat_process_exec(struct pmcstat_process *_pp,
251 pmcstat_interned_string _path, uintfptr_t _entryaddr);
252 static struct pmcstat_process *pmcstat_process_lookup(pid_t _pid,
254 static int pmcstat_string_compute_hash(const char *_string);
255 static void pmcstat_string_initialize(void);
256 static int pmcstat_string_lookup_hash(pmcstat_interned_string _is);
257 static void pmcstat_string_shutdown(void);
258 static void pmcstat_stats_reset(int _reset_global);
261 * A simple implementation of interned strings. Each interned string
262 * is assigned a unique address, so that subsequent string compares
263 * can be done by a simple pointer comparison instead of using
264 * strcmp(). This speeds up hash table lookups and saves memory if
265 * duplicate strings are the norm.
267 struct pmcstat_string {
268 LIST_ENTRY(pmcstat_string) ps_next; /* hash link */
274 static LIST_HEAD(,pmcstat_string) pmcstat_string_hash[PMCSTAT_NHASH];
282 * PMC Top mode pause state.
284 static int pmcstat_pause;
287 pmcstat_stats_reset(int reset_global)
289 struct pmcstat_pmcrecord *pr;
291 /* Flush PMCs stats. */
292 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
294 pr->pr_dubious_frames = 0;
296 ps_samples_period = 0;
298 /* Flush global stats. */
300 bzero(&pmcstat_stats, sizeof(struct pmcstat_stats));
304 * Compute a 'hash' value for a string.
308 pmcstat_string_compute_hash(const char *s)
312 for (hash = 2166136261; *s; s++)
313 hash = (hash ^ *s) * 16777619;
315 return (hash & PMCSTAT_HASH_MASK);
319 * Intern a copy of string 's', and return a pointer to the
320 * interned structure.
323 pmcstat_interned_string
324 pmcstat_string_intern(const char *s)
326 struct pmcstat_string *ps;
327 const struct pmcstat_string *cps;
330 if ((cps = pmcstat_string_lookup(s)) != NULL)
333 hash = pmcstat_string_compute_hash(s);
336 if ((ps = malloc(sizeof(*ps))) == NULL)
337 err(EX_OSERR, "ERROR: Could not intern string");
340 ps->ps_string = strdup(s);
341 LIST_INSERT_HEAD(&pmcstat_string_hash[hash], ps, ps_next);
342 return ((pmcstat_interned_string) ps);
346 pmcstat_string_unintern(pmcstat_interned_string str)
350 s = ((const struct pmcstat_string *) str)->ps_string;
354 pmcstat_interned_string
355 pmcstat_string_lookup(const char *s)
357 struct pmcstat_string *ps;
360 hash = pmcstat_string_compute_hash(s);
363 LIST_FOREACH(ps, &pmcstat_string_hash[hash], ps_next)
364 if (ps->ps_len == len && ps->ps_hash == hash &&
365 strcmp(ps->ps_string, s) == 0)
371 pmcstat_string_lookup_hash(pmcstat_interned_string s)
373 const struct pmcstat_string *ps;
375 ps = (const struct pmcstat_string *) s;
376 return (ps->ps_hash);
380 * Initialize the string interning facility.
384 pmcstat_string_initialize(void)
388 for (i = 0; i < PMCSTAT_NHASH; i++)
389 LIST_INIT(&pmcstat_string_hash[i]);
393 * Destroy the string table, free'ing up space.
397 pmcstat_string_shutdown(void)
400 struct pmcstat_string *ps, *pstmp;
402 for (i = 0; i < PMCSTAT_NHASH; i++)
403 LIST_FOREACH_SAFE(ps, &pmcstat_string_hash[i], ps_next,
405 LIST_REMOVE(ps, ps_next);
412 * Determine whether a given executable image is an A.OUT object, and
413 * if so, fill in its parameters from the text file.
414 * Sets image->pi_type.
418 pmcstat_image_get_aout_params(struct pmcstat_image *image)
424 char buffer[PATH_MAX];
426 path = pmcstat_string_unintern(image->pi_execpath);
427 assert(path != NULL);
429 if (image->pi_iskernelmodule)
431 "ERROR: a.out kernel modules are unsupported \"%s\"", path);
433 (void) snprintf(buffer, sizeof(buffer), "%s%s",
434 args.pa_fsroot, path);
436 if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
437 (nbytes = read(fd, &ex, sizeof(ex))) < 0) {
438 if (args.pa_verbosity >= 2)
439 warn("WARNING: Cannot determine type of \"%s\"",
441 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
449 if ((unsigned) nbytes != sizeof(ex) ||
453 image->pi_type = PMCSTAT_IMAGE_AOUT;
455 /* TODO: the rest of a.out processing */
465 pmcstat_symbol_compare(const void *a, const void *b)
467 const struct pmcstat_symbol *sym1, *sym2;
469 sym1 = (const struct pmcstat_symbol *) a;
470 sym2 = (const struct pmcstat_symbol *) b;
472 if (sym1->ps_end <= sym2->ps_start)
474 if (sym1->ps_start >= sym2->ps_end)
480 * Map an address to a symbol in an image.
483 struct pmcstat_symbol *
484 pmcstat_symbol_search(struct pmcstat_image *image, uintfptr_t addr)
486 struct pmcstat_symbol sym;
488 if (image->pi_symbols == NULL)
493 sym.ps_end = addr + 1;
495 return (bsearch((void *) &sym, image->pi_symbols,
496 image->pi_symcount, sizeof(struct pmcstat_symbol),
497 pmcstat_symbol_compare));
501 * Add the list of symbols in the given section to the list associated
505 pmcstat_image_add_symbols(struct pmcstat_image *image, Elf *e,
506 Elf_Scn *scn, GElf_Shdr *sh)
509 size_t n, newsyms, nshsyms, nfuncsyms;
510 struct pmcstat_symbol *symptr;
515 if ((data = elf_getdata(scn, NULL)) == NULL)
519 * Determine the number of functions named in this
523 nshsyms = sh->sh_size / sh->sh_entsize;
524 for (n = nfuncsyms = 0; n < nshsyms; n++) {
525 if (gelf_getsym(data, (int) n, &sym) != &sym)
527 if (GELF_ST_TYPE(sym.st_info) == STT_FUNC)
535 * Allocate space for the new entries.
537 firsttime = image->pi_symbols == NULL;
538 symptr = realloc(image->pi_symbols,
539 sizeof(*symptr) * (image->pi_symcount + nfuncsyms));
540 if (symptr == image->pi_symbols) /* realloc() failed. */
542 image->pi_symbols = symptr;
545 * Append new symbols to the end of the current table.
547 symptr += image->pi_symcount;
549 for (n = newsyms = 0; n < nshsyms; n++) {
550 if (gelf_getsym(data, (int) n, &sym) != &sym)
552 if (GELF_ST_TYPE(sym.st_info) != STT_FUNC)
554 if (sym.st_shndx == STN_UNDEF)
557 if (!firsttime && pmcstat_symbol_search(image, sym.st_value))
558 continue; /* We've seen this symbol already. */
560 if ((fnname = elf_strptr(e, sh->sh_link, sym.st_name))
564 /* Remove spurious ARM function name. */
565 if (fnname[0] == '$' &&
566 (fnname[1] == 'a' || fnname[1] == 't' ||
572 symptr->ps_name = pmcstat_string_intern(fnname);
573 symptr->ps_start = sym.st_value - image->pi_vaddr;
574 symptr->ps_end = symptr->ps_start + sym.st_size;
580 image->pi_symcount += newsyms;
581 if (image->pi_symcount == 0)
584 assert(newsyms <= nfuncsyms);
587 * Return space to the system if there were duplicates.
589 if (newsyms < nfuncsyms)
590 image->pi_symbols = realloc(image->pi_symbols,
591 sizeof(*symptr) * image->pi_symcount);
594 * Keep the list of symbols sorted.
596 qsort(image->pi_symbols, image->pi_symcount, sizeof(*symptr),
597 pmcstat_symbol_compare);
600 * Deal with function symbols that have a size of 'zero' by
601 * making them extend to the next higher address. These
602 * symbols are usually defined in assembly code.
604 for (symptr = image->pi_symbols;
605 symptr < image->pi_symbols + (image->pi_symcount - 1);
607 if (symptr->ps_start == symptr->ps_end)
608 symptr->ps_end = (symptr+1)->ps_start;
612 * Examine an ELF file to determine the size of its text segment.
613 * Sets image->pi_type if anything conclusive can be determined about
618 pmcstat_image_get_elf_params(struct pmcstat_image *image)
622 const char *path, *elfbase;
624 uintfptr_t minva, maxva;
630 enum pmcstat_image_type image_type;
631 char buffer[PATH_MAX];
633 assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
635 image->pi_start = minva = ~(uintfptr_t) 0;
636 image->pi_end = maxva = (uintfptr_t) 0;
637 image->pi_type = image_type = PMCSTAT_IMAGE_INDETERMINABLE;
638 image->pi_isdynamic = 0;
639 image->pi_dynlinkerpath = NULL;
642 path = pmcstat_string_unintern(image->pi_execpath);
643 assert(path != NULL);
646 * Look for kernel modules under FSROOT/KERNELPATH/NAME,
647 * and user mode executable objects under FSROOT/PATHNAME.
649 if (image->pi_iskernelmodule)
650 (void) snprintf(buffer, sizeof(buffer), "%s%s/%s",
651 args.pa_fsroot, args.pa_kernel, path);
653 (void) snprintf(buffer, sizeof(buffer), "%s%s",
654 args.pa_fsroot, path);
657 if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
658 (e = elf_begin(fd, ELF_C_READ, NULL)) == NULL ||
659 (elf_kind(e) != ELF_K_ELF)) {
660 if (args.pa_verbosity >= 2)
661 warnx("WARNING: Cannot determine the type of \"%s\".",
666 if (gelf_getehdr(e, &eh) != &eh) {
668 "WARNING: Cannot retrieve the ELF Header for \"%s\": %s.",
669 buffer, elf_errmsg(-1));
673 if (eh.e_type != ET_EXEC && eh.e_type != ET_DYN &&
674 !(image->pi_iskernelmodule && eh.e_type == ET_REL)) {
675 warnx("WARNING: \"%s\" is of an unsupported ELF type.",
680 image_type = eh.e_ident[EI_CLASS] == ELFCLASS32 ?
681 PMCSTAT_IMAGE_ELF32 : PMCSTAT_IMAGE_ELF64;
684 * Determine the virtual address where an executable would be
685 * loaded. Additionally, for dynamically linked executables,
686 * save the pathname to the runtime linker.
688 if (eh.e_type == ET_EXEC) {
689 if (elf_getphnum(e, &nph) == 0) {
691 "WARNING: Could not determine the number of program headers in \"%s\": %s.",
696 for (i = 0; i < eh.e_phnum; i++) {
697 if (gelf_getphdr(e, i, &ph) != &ph) {
699 "WARNING: Retrieval of PHDR entry #%ju in \"%s\" failed: %s.",
700 (uintmax_t) i, buffer, elf_errmsg(-1));
705 image->pi_isdynamic = 1;
708 if ((elfbase = elf_rawfile(e, NULL)) == NULL) {
710 "WARNING: Cannot retrieve the interpreter for \"%s\": %s.",
711 buffer, elf_errmsg(-1));
714 image->pi_dynlinkerpath =
715 pmcstat_string_intern(elfbase +
719 if ((ph.p_flags & PF_X) != 0 &&
720 (ph.p_offset & (-ph.p_align)) == 0)
721 image->pi_vaddr = ph.p_vaddr & (-ph.p_align);
728 * Get the min and max VA associated with this ELF object.
730 if (elf_getshnum(e, &nsh) == 0) {
732 "WARNING: Could not determine the number of sections for \"%s\": %s.",
733 buffer, elf_errmsg(-1));
737 for (i = 0; i < nsh; i++) {
738 if ((scn = elf_getscn(e, i)) == NULL ||
739 gelf_getshdr(scn, &sh) != &sh) {
741 "WARNING: Could not retrieve section header #%ju in \"%s\": %s.",
742 (uintmax_t) i, buffer, elf_errmsg(-1));
745 if (sh.sh_flags & SHF_EXECINSTR) {
746 minva = min(minva, sh.sh_addr);
747 maxva = max(maxva, sh.sh_addr + sh.sh_size);
749 if (sh.sh_type == SHT_SYMTAB || sh.sh_type == SHT_DYNSYM)
750 pmcstat_image_add_symbols(image, e, scn, &sh);
753 image->pi_start = minva;
754 image->pi_end = maxva;
755 image->pi_type = image_type;
756 image->pi_fullpath = pmcstat_string_intern(buffer);
758 /* Build display name
761 for (p = buffer; *p; p++)
764 image->pi_name = pmcstat_string_intern(endp);
774 * Given an image descriptor, determine whether it is an ELF, or AOUT.
775 * If no handler claims the image, set its type to 'INDETERMINABLE'.
779 pmcstat_image_determine_type(struct pmcstat_image *image)
781 assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
783 /* Try each kind of handler in turn */
784 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
785 pmcstat_image_get_elf_params(image);
786 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
787 pmcstat_image_get_aout_params(image);
790 * Otherwise, remember that we tried to determine
791 * the object's type and had failed.
793 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
794 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
798 * Locate an image descriptor given an interned path, adding a fresh
799 * descriptor to the cache if necessary. This function also finds a
800 * suitable name for this image's sample file.
802 * We defer filling in the file format specific parts of the image
803 * structure till the time we actually see a sample that would fall
807 static struct pmcstat_image *
808 pmcstat_image_from_path(pmcstat_interned_string internedpath,
812 struct pmcstat_image *pi;
814 hash = pmcstat_string_lookup_hash(internedpath);
816 /* First, look for an existing entry. */
817 LIST_FOREACH(pi, &pmcstat_image_hash[hash], pi_next)
818 if (pi->pi_execpath == internedpath &&
819 pi->pi_iskernelmodule == iskernelmodule)
823 * Allocate a new entry and place it at the head of the hash
826 pi = malloc(sizeof(*pi));
830 pi->pi_type = PMCSTAT_IMAGE_UNKNOWN;
831 pi->pi_execpath = internedpath;
836 pi->pi_isdynamic = 0;
837 pi->pi_iskernelmodule = iskernelmodule;
838 pi->pi_dynlinkerpath = NULL;
839 pi->pi_symbols = NULL;
841 pi->pi_addr2line = NULL;
843 if (plugins[args.pa_pplugin].pl_initimage != NULL)
844 plugins[args.pa_pplugin].pl_initimage(pi);
845 if (plugins[args.pa_plugin].pl_initimage != NULL)
846 plugins[args.pa_plugin].pl_initimage(pi);
848 LIST_INSERT_HEAD(&pmcstat_image_hash[hash], pi, pi_next);
854 * Record the fact that PC values from 'start' to 'end' come from
859 pmcstat_image_link(struct pmcstat_process *pp, struct pmcstat_image *image,
862 struct pmcstat_pcmap *pcm, *pcmnew;
865 assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN &&
866 image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE);
868 if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
869 err(EX_OSERR, "ERROR: Cannot create a map entry");
872 * Adjust the map entry to only cover the text portion
876 offset = start - image->pi_vaddr;
877 pcmnew->ppm_lowpc = image->pi_start + offset;
878 pcmnew->ppm_highpc = image->pi_end + offset;
879 pcmnew->ppm_image = image;
881 assert(pcmnew->ppm_lowpc < pcmnew->ppm_highpc);
883 /* Overlapped mmap()'s are assumed to never occur. */
884 TAILQ_FOREACH(pcm, &pp->pp_map, ppm_next)
885 if (pcm->ppm_lowpc >= pcmnew->ppm_highpc)
889 TAILQ_INSERT_TAIL(&pp->pp_map, pcmnew, ppm_next);
891 TAILQ_INSERT_BEFORE(pcm, pcmnew, ppm_next);
895 * Unmap images in the range [start..end) associated with process
900 pmcstat_image_unmap(struct pmcstat_process *pp, uintfptr_t start,
903 struct pmcstat_pcmap *pcm, *pcmtmp, *pcmnew;
910 * - we could have the range completely in the middle of an
911 * existing pcmap; in this case we have to split the pcmap
912 * structure into two (i.e., generate a 'hole').
913 * - we could have the range covering multiple pcmaps; these
914 * will have to be removed.
915 * - we could have either 'start' or 'end' falling in the
916 * middle of a pcmap; in this case shorten the entry.
918 TAILQ_FOREACH_SAFE(pcm, &pp->pp_map, ppm_next, pcmtmp) {
919 assert(pcm->ppm_lowpc < pcm->ppm_highpc);
920 if (pcm->ppm_highpc <= start)
922 if (pcm->ppm_lowpc >= end)
924 if (pcm->ppm_lowpc >= start && pcm->ppm_highpc <= end) {
926 * The current pcmap is completely inside the
927 * unmapped range: remove it entirely.
929 TAILQ_REMOVE(&pp->pp_map, pcm, ppm_next);
931 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc > end) {
933 * Split this pcmap into two; curtail the
934 * current map to end at [start-1], and start
935 * the new one at [end].
937 if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
939 "ERROR: Cannot split a map entry");
941 pcmnew->ppm_image = pcm->ppm_image;
943 pcmnew->ppm_lowpc = end;
944 pcmnew->ppm_highpc = pcm->ppm_highpc;
946 pcm->ppm_highpc = start;
948 TAILQ_INSERT_AFTER(&pp->pp_map, pcm, pcmnew, ppm_next);
951 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc <= end)
952 pcm->ppm_highpc = start;
953 else if (pcm->ppm_lowpc >= start && pcm->ppm_highpc > end)
954 pcm->ppm_lowpc = end;
961 * Resolve file name and line number for the given address.
964 pmcstat_image_addr2line(struct pmcstat_image *image, uintfptr_t addr,
965 char *sourcefile, size_t sourcefile_len, unsigned *sourceline,
966 char *funcname, size_t funcname_len)
968 static int addr2line_warn = 0;
970 char *sep, cmdline[PATH_MAX], imagepath[PATH_MAX];
974 if (image->pi_addr2line == NULL) {
975 /* Try default debug file location. */
976 snprintf(imagepath, sizeof(imagepath),
977 "/usr/lib/debug/%s%s.debug",
979 pmcstat_string_unintern(image->pi_fullpath));
980 fd = open(imagepath, O_RDONLY);
982 /* Old kernel symbol path. */
983 snprintf(imagepath, sizeof(imagepath), "%s%s.symbols",
985 pmcstat_string_unintern(image->pi_fullpath));
986 fd = open(imagepath, O_RDONLY);
988 snprintf(imagepath, sizeof(imagepath), "%s%s",
990 pmcstat_string_unintern(
991 image->pi_fullpath));
997 * New addr2line support recursive inline function with -i
998 * but the format does not add a marker when no more entries
1001 snprintf(cmdline, sizeof(cmdline), "addr2line -Cfe \"%s\"",
1003 image->pi_addr2line = popen(cmdline, "r+");
1004 if (image->pi_addr2line == NULL) {
1005 if (!addr2line_warn) {
1008 "WARNING: addr2line is needed for source code information."
1015 if (feof(image->pi_addr2line) || ferror(image->pi_addr2line)) {
1016 warnx("WARNING: addr2line pipe error");
1017 pclose(image->pi_addr2line);
1018 image->pi_addr2line = NULL;
1022 fprintf(image->pi_addr2line, "%p\n", (void *)addr);
1024 if (fgets(funcname, funcname_len, image->pi_addr2line) == NULL) {
1025 warnx("WARNING: addr2line function name read error");
1028 sep = strchr(funcname, '\n');
1032 if (fgets(sourcefile, sourcefile_len, image->pi_addr2line) == NULL) {
1033 warnx("WARNING: addr2line source file read error");
1036 sep = strchr(sourcefile, ':');
1038 warnx("WARNING: addr2line source line separator missing");
1050 * Add a {pmcid,name} mapping.
1054 pmcstat_pmcid_add(pmc_id_t pmcid, pmcstat_interned_string ps)
1056 struct pmcstat_pmcrecord *pr, *prm;
1058 /* Replace an existing name for the PMC. */
1060 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1061 if (pr->pr_pmcid == pmcid) {
1062 pr->pr_pmcname = ps;
1064 } else if (pr->pr_pmcname == ps)
1068 * Otherwise, allocate a new descriptor and call the
1071 if ((pr = malloc(sizeof(*pr))) == NULL)
1072 err(EX_OSERR, "ERROR: Cannot allocate pmc record");
1074 pr->pr_pmcid = pmcid;
1075 pr->pr_pmcname = ps;
1076 pr->pr_pmcin = pmcstat_npmcs++;
1078 pr->pr_dubious_frames = 0;
1079 pr->pr_merge = prm == NULL ? pr : prm;
1081 LIST_INSERT_HEAD(&pmcstat_pmcs, pr, pr_next);
1083 if (plugins[args.pa_pplugin].pl_newpmc != NULL)
1084 plugins[args.pa_pplugin].pl_newpmc(ps, pr);
1085 if (plugins[args.pa_plugin].pl_newpmc != NULL)
1086 plugins[args.pa_plugin].pl_newpmc(ps, pr);
1090 * Given a pmcid in use, find its human-readable name.
1094 pmcstat_pmcid_to_name(pmc_id_t pmcid)
1096 struct pmcstat_pmcrecord *pr;
1098 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1099 if (pr->pr_pmcid == pmcid)
1100 return (pmcstat_string_unintern(pr->pr_pmcname));
1106 * Convert PMC index to name.
1110 pmcstat_pmcindex_to_name(int pmcin)
1112 struct pmcstat_pmcrecord *pr;
1114 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1115 if (pr->pr_pmcin == pmcin)
1116 return pmcstat_string_unintern(pr->pr_pmcname);
1122 * Return PMC record with given index.
1125 struct pmcstat_pmcrecord *
1126 pmcstat_pmcindex_to_pmcr(int pmcin)
1128 struct pmcstat_pmcrecord *pr;
1130 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1131 if (pr->pr_pmcin == pmcin)
1138 * Get PMC record by id, apply merge policy.
1141 static struct pmcstat_pmcrecord *
1142 pmcstat_lookup_pmcid(pmc_id_t pmcid)
1144 struct pmcstat_pmcrecord *pr;
1146 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
1147 if (pr->pr_pmcid == pmcid) {
1148 if (pmcstat_mergepmc)
1149 return pr->pr_merge;
1158 * Associate an AOUT image with a process.
1162 pmcstat_process_aout_exec(struct pmcstat_process *pp,
1163 struct pmcstat_image *image, uintfptr_t entryaddr)
1168 /* TODO Implement a.out handling */
1172 * Associate an ELF image with a process.
1176 pmcstat_process_elf_exec(struct pmcstat_process *pp,
1177 struct pmcstat_image *image, uintfptr_t entryaddr)
1180 struct pmcstat_image *rtldimage;
1182 assert(image->pi_type == PMCSTAT_IMAGE_ELF32 ||
1183 image->pi_type == PMCSTAT_IMAGE_ELF64);
1185 /* Create a map entry for the base executable. */
1186 pmcstat_image_link(pp, image, image->pi_vaddr);
1189 * For dynamically linked executables we need to determine
1190 * where the dynamic linker was mapped to for this process,
1191 * Subsequent executable objects that are mapped in by the
1192 * dynamic linker will be tracked by log events of type
1193 * PMCLOG_TYPE_MAP_IN.
1196 if (image->pi_isdynamic) {
1199 * The runtime loader gets loaded just after the maximum
1200 * possible heap address. Like so:
1202 * [ TEXT DATA BSS HEAP -->*RTLD SHLIBS <--STACK]
1204 * 0 VM_MAXUSER_ADDRESS
1207 * The exact address where the loader gets mapped in
1208 * will vary according to the size of the executable
1209 * and the limits on the size of the process'es data
1210 * segment at the time of exec(). The entry address
1211 * recorded at process exec time corresponds to the
1212 * 'start' address inside the dynamic linker. From
1213 * this we can figure out the address where the
1214 * runtime loader's file object had been mapped to.
1216 rtldimage = pmcstat_image_from_path(image->pi_dynlinkerpath, 0);
1217 if (rtldimage == NULL) {
1218 warnx("WARNING: Cannot find image for \"%s\".",
1219 pmcstat_string_unintern(image->pi_dynlinkerpath));
1220 pmcstat_stats.ps_exec_errors++;
1224 if (rtldimage->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1225 pmcstat_image_get_elf_params(rtldimage);
1227 if (rtldimage->pi_type != PMCSTAT_IMAGE_ELF32 &&
1228 rtldimage->pi_type != PMCSTAT_IMAGE_ELF64) {
1229 warnx("WARNING: rtld not an ELF object \"%s\".",
1230 pmcstat_string_unintern(image->pi_dynlinkerpath));
1234 libstart = entryaddr - rtldimage->pi_entry;
1235 pmcstat_image_link(pp, rtldimage, libstart);
1240 * Find the process descriptor corresponding to a PID. If 'allocate'
1241 * is zero, we return a NULL if a pid descriptor could not be found or
1242 * a process descriptor process. If 'allocate' is non-zero, then we
1243 * will attempt to allocate a fresh process descriptor. Zombie
1244 * process descriptors are only removed if a fresh allocation for the
1245 * same PID is requested.
1248 static struct pmcstat_process *
1249 pmcstat_process_lookup(pid_t pid, int allocate)
1252 struct pmcstat_pcmap *ppm, *ppmtmp;
1253 struct pmcstat_process *pp, *pptmp;
1255 hash = (uint32_t) pid & PMCSTAT_HASH_MASK; /* simplicity wins */
1257 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[hash], pp_next, pptmp)
1258 if (pp->pp_pid == pid) {
1259 /* Found a descriptor, check and process zombies */
1260 if (allocate && pp->pp_isactive == 0) {
1262 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next,
1264 TAILQ_REMOVE(&pp->pp_map, ppm,
1268 /* remove process entry */
1269 LIST_REMOVE(pp, pp_next);
1279 if ((pp = malloc(sizeof(*pp))) == NULL)
1280 err(EX_OSERR, "ERROR: Cannot allocate pid descriptor");
1283 pp->pp_isactive = 1;
1285 TAILQ_INIT(&pp->pp_map);
1287 LIST_INSERT_HEAD(&pmcstat_process_hash[hash], pp, pp_next);
1292 * Associate an image and a process.
1296 pmcstat_process_exec(struct pmcstat_process *pp,
1297 pmcstat_interned_string path, uintfptr_t entryaddr)
1299 struct pmcstat_image *image;
1301 if ((image = pmcstat_image_from_path(path, 0)) == NULL) {
1302 pmcstat_stats.ps_exec_errors++;
1306 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1307 pmcstat_image_determine_type(image);
1309 assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN);
1311 switch (image->pi_type) {
1312 case PMCSTAT_IMAGE_ELF32:
1313 case PMCSTAT_IMAGE_ELF64:
1314 pmcstat_stats.ps_exec_elf++;
1315 pmcstat_process_elf_exec(pp, image, entryaddr);
1318 case PMCSTAT_IMAGE_AOUT:
1319 pmcstat_stats.ps_exec_aout++;
1320 pmcstat_process_aout_exec(pp, image, entryaddr);
1323 case PMCSTAT_IMAGE_INDETERMINABLE:
1324 pmcstat_stats.ps_exec_indeterminable++;
1329 "ERROR: Unsupported executable type for \"%s\"",
1330 pmcstat_string_unintern(path));
1336 * Find the map entry associated with process 'p' at PC value 'pc'.
1339 struct pmcstat_pcmap *
1340 pmcstat_process_find_map(struct pmcstat_process *p, uintfptr_t pc)
1342 struct pmcstat_pcmap *ppm;
1344 TAILQ_FOREACH(ppm, &p->pp_map, ppm_next) {
1345 if (pc >= ppm->ppm_lowpc && pc < ppm->ppm_highpc)
1347 if (pc < ppm->ppm_lowpc)
1355 * Convert a hwpmc(4) log to profile information. A system-wide
1356 * callgraph is generated if FLAG_DO_CALLGRAPHS is set. gmon.out
1357 * files usable by gprof(1) are created if FLAG_DO_GPROF is set.
1360 pmcstat_analyze_log(void)
1362 uint32_t cpu, cpuflags;
1365 struct pmcstat_image *image;
1366 struct pmcstat_process *pp, *ppnew;
1367 struct pmcstat_pcmap *ppm, *ppmtmp;
1368 struct pmclog_ev ev;
1369 struct pmcstat_pmcrecord *pmcr;
1370 pmcstat_interned_string image_path;
1372 assert(args.pa_flags & FLAG_DO_ANALYSIS);
1374 if (elf_version(EV_CURRENT) == EV_NONE)
1375 err(EX_UNAVAILABLE, "Elf library intialization failed");
1377 while (pmclog_read(args.pa_logparser, &ev) == 0) {
1378 assert(ev.pl_state == PMCLOG_OK);
1380 switch (ev.pl_type) {
1381 case PMCLOG_TYPE_INITIALIZE:
1382 if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1383 PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1385 "WARNING: Log version 0x%x does not match compiled version 0x%x.",
1386 ev.pl_u.pl_i.pl_version, PMC_VERSION_MAJOR);
1389 case PMCLOG_TYPE_MAP_IN:
1391 * Introduce an address range mapping for a
1392 * userland process or the kernel (pid == -1).
1394 * We always allocate a process descriptor so
1395 * that subsequent samples seen for this
1396 * address range are mapped to the current
1397 * object being mapped in.
1399 pid = ev.pl_u.pl_mi.pl_pid;
1401 pp = pmcstat_kernproc;
1403 pp = pmcstat_process_lookup(pid,
1408 image_path = pmcstat_string_intern(ev.pl_u.pl_mi.
1410 image = pmcstat_image_from_path(image_path, pid == -1);
1411 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1412 pmcstat_image_determine_type(image);
1413 if (image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE)
1414 pmcstat_image_link(pp, image,
1415 ev.pl_u.pl_mi.pl_start);
1418 case PMCLOG_TYPE_MAP_OUT:
1420 * Remove an address map.
1422 pid = ev.pl_u.pl_mo.pl_pid;
1424 pp = pmcstat_kernproc;
1426 pp = pmcstat_process_lookup(pid, 0);
1428 if (pp == NULL) /* unknown process */
1431 pmcstat_image_unmap(pp, ev.pl_u.pl_mo.pl_start,
1432 ev.pl_u.pl_mo.pl_end);
1435 case PMCLOG_TYPE_PCSAMPLE:
1437 * Note: the `PCSAMPLE' log entry is not
1438 * generated by hpwmc(4) after version 2.
1442 * We bring in the gmon file for the image
1443 * currently associated with the PMC & pid
1444 * pair and increment the appropriate entry
1447 pmcstat_stats.ps_samples_total++;
1448 ps_samples_period++;
1450 pc = ev.pl_u.pl_s.pl_pc;
1451 pp = pmcstat_process_lookup(ev.pl_u.pl_s.pl_pid,
1454 /* Get PMC record. */
1455 pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_s.pl_pmcid);
1456 assert(pmcr != NULL);
1460 * Call the plugins processing
1461 * TODO: move pmcstat_process_find_map inside plugins
1464 if (plugins[args.pa_pplugin].pl_process != NULL)
1465 plugins[args.pa_pplugin].pl_process(
1467 pmcstat_process_find_map(pp, pc) != NULL, 0);
1468 plugins[args.pa_plugin].pl_process(
1470 pmcstat_process_find_map(pp, pc) != NULL, 0);
1473 case PMCLOG_TYPE_CALLCHAIN:
1474 pmcstat_stats.ps_samples_total++;
1475 ps_samples_period++;
1477 cpuflags = ev.pl_u.pl_cc.pl_cpuflags;
1478 cpu = PMC_CALLCHAIN_CPUFLAGS_TO_CPU(cpuflags);
1480 /* Filter on the CPU id. */
1481 if (!CPU_ISSET(cpu, &(args.pa_cpumask))) {
1482 pmcstat_stats.ps_samples_skipped++;
1486 pp = pmcstat_process_lookup(ev.pl_u.pl_cc.pl_pid,
1489 /* Get PMC record. */
1490 pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_cc.pl_pmcid);
1491 assert(pmcr != NULL);
1495 * Call the plugins processing
1498 if (plugins[args.pa_pplugin].pl_process != NULL)
1499 plugins[args.pa_pplugin].pl_process(
1501 ev.pl_u.pl_cc.pl_npc,
1502 ev.pl_u.pl_cc.pl_pc,
1503 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1505 plugins[args.pa_plugin].pl_process(
1507 ev.pl_u.pl_cc.pl_npc,
1508 ev.pl_u.pl_cc.pl_pc,
1509 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1513 case PMCLOG_TYPE_PMCALLOCATE:
1515 * Record the association pmc id between this
1518 pmcstat_pmcid_add(ev.pl_u.pl_a.pl_pmcid,
1519 pmcstat_string_intern(ev.pl_u.pl_a.pl_evname));
1522 case PMCLOG_TYPE_PMCALLOCATEDYN:
1524 * Record the association pmc id between this
1527 pmcstat_pmcid_add(ev.pl_u.pl_ad.pl_pmcid,
1528 pmcstat_string_intern(ev.pl_u.pl_ad.pl_evname));
1531 case PMCLOG_TYPE_PROCEXEC:
1534 * Change the executable image associated with
1537 pp = pmcstat_process_lookup(ev.pl_u.pl_x.pl_pid,
1540 /* delete the current process map */
1541 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
1542 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
1547 * Associate this process image.
1549 image_path = pmcstat_string_intern(
1550 ev.pl_u.pl_x.pl_pathname);
1551 assert(image_path != NULL);
1552 pmcstat_process_exec(pp, image_path,
1553 ev.pl_u.pl_x.pl_entryaddr);
1556 case PMCLOG_TYPE_PROCEXIT:
1559 * Due to the way the log is generated, the
1560 * last few samples corresponding to a process
1561 * may appear in the log after the process
1562 * exit event is recorded. Thus we keep the
1563 * process' descriptor and associated data
1564 * structures around, but mark the process as
1567 pp = pmcstat_process_lookup(ev.pl_u.pl_e.pl_pid, 0);
1570 pp->pp_isactive = 0; /* mark as a zombie */
1573 case PMCLOG_TYPE_SYSEXIT:
1574 pp = pmcstat_process_lookup(ev.pl_u.pl_se.pl_pid, 0);
1577 pp->pp_isactive = 0; /* make a zombie */
1580 case PMCLOG_TYPE_PROCFORK:
1583 * Allocate a process descriptor for the new
1587 pmcstat_process_lookup(ev.pl_u.pl_f.pl_newpid,
1591 * If we had been tracking the parent, clone
1594 pp = pmcstat_process_lookup(ev.pl_u.pl_f.pl_oldpid, 0);
1597 TAILQ_FOREACH(ppm, &pp->pp_map, ppm_next)
1598 pmcstat_image_link(ppnew, ppm->ppm_image,
1602 default: /* other types of entries are not relevant */
1607 if (ev.pl_state == PMCLOG_EOF)
1608 return (PMCSTAT_FINISHED);
1609 else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1610 return (PMCSTAT_RUNNING);
1613 "ERROR: event parsing failed (record %jd, offset 0x%jx)",
1614 (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1618 * Print log entries as text.
1622 pmcstat_print_log(void)
1624 struct pmclog_ev ev;
1627 while (pmclog_read(args.pa_logparser, &ev) == 0) {
1628 assert(ev.pl_state == PMCLOG_OK);
1629 switch (ev.pl_type) {
1630 case PMCLOG_TYPE_CALLCHAIN:
1631 PMCSTAT_PRINT_ENTRY("callchain",
1632 "%d 0x%x %d %d %c", ev.pl_u.pl_cc.pl_pid,
1633 ev.pl_u.pl_cc.pl_pmcid,
1634 PMC_CALLCHAIN_CPUFLAGS_TO_CPU(ev.pl_u.pl_cc. \
1635 pl_cpuflags), ev.pl_u.pl_cc.pl_npc,
1636 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(ev.pl_u.pl_cc.\
1637 pl_cpuflags) ? 'u' : 's');
1638 for (npc = 0; npc < ev.pl_u.pl_cc.pl_npc; npc++)
1639 PMCSTAT_PRINT_ENTRY("...", "%p",
1640 (void *) ev.pl_u.pl_cc.pl_pc[npc]);
1642 case PMCLOG_TYPE_CLOSELOG:
1643 PMCSTAT_PRINT_ENTRY("closelog",);
1645 case PMCLOG_TYPE_DROPNOTIFY:
1646 PMCSTAT_PRINT_ENTRY("drop",);
1648 case PMCLOG_TYPE_INITIALIZE:
1649 PMCSTAT_PRINT_ENTRY("initlog","0x%x \"%s\"",
1650 ev.pl_u.pl_i.pl_version,
1651 pmc_name_of_cputype(ev.pl_u.pl_i.pl_arch));
1652 if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1653 PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1655 "WARNING: Log version 0x%x != expected version 0x%x.",
1656 ev.pl_u.pl_i.pl_version, PMC_VERSION);
1658 case PMCLOG_TYPE_MAP_IN:
1659 PMCSTAT_PRINT_ENTRY("map-in","%d %p \"%s\"",
1660 ev.pl_u.pl_mi.pl_pid,
1661 (void *) ev.pl_u.pl_mi.pl_start,
1662 ev.pl_u.pl_mi.pl_pathname);
1664 case PMCLOG_TYPE_MAP_OUT:
1665 PMCSTAT_PRINT_ENTRY("map-out","%d %p %p",
1666 ev.pl_u.pl_mo.pl_pid,
1667 (void *) ev.pl_u.pl_mo.pl_start,
1668 (void *) ev.pl_u.pl_mo.pl_end);
1670 case PMCLOG_TYPE_PCSAMPLE:
1671 PMCSTAT_PRINT_ENTRY("sample","0x%x %d %p %c",
1672 ev.pl_u.pl_s.pl_pmcid,
1673 ev.pl_u.pl_s.pl_pid,
1674 (void *) ev.pl_u.pl_s.pl_pc,
1675 ev.pl_u.pl_s.pl_usermode ? 'u' : 's');
1677 case PMCLOG_TYPE_PMCALLOCATE:
1678 PMCSTAT_PRINT_ENTRY("allocate","0x%x \"%s\" 0x%x",
1679 ev.pl_u.pl_a.pl_pmcid,
1680 ev.pl_u.pl_a.pl_evname,
1681 ev.pl_u.pl_a.pl_flags);
1683 case PMCLOG_TYPE_PMCALLOCATEDYN:
1684 PMCSTAT_PRINT_ENTRY("allocatedyn","0x%x \"%s\" 0x%x",
1685 ev.pl_u.pl_ad.pl_pmcid,
1686 ev.pl_u.pl_ad.pl_evname,
1687 ev.pl_u.pl_ad.pl_flags);
1689 case PMCLOG_TYPE_PMCATTACH:
1690 PMCSTAT_PRINT_ENTRY("attach","0x%x %d \"%s\"",
1691 ev.pl_u.pl_t.pl_pmcid,
1692 ev.pl_u.pl_t.pl_pid,
1693 ev.pl_u.pl_t.pl_pathname);
1695 case PMCLOG_TYPE_PMCDETACH:
1696 PMCSTAT_PRINT_ENTRY("detach","0x%x %d",
1697 ev.pl_u.pl_d.pl_pmcid,
1698 ev.pl_u.pl_d.pl_pid);
1700 case PMCLOG_TYPE_PROCCSW:
1701 PMCSTAT_PRINT_ENTRY("cswval","0x%x %d %jd",
1702 ev.pl_u.pl_c.pl_pmcid,
1703 ev.pl_u.pl_c.pl_pid,
1704 ev.pl_u.pl_c.pl_value);
1706 case PMCLOG_TYPE_PROCEXEC:
1707 PMCSTAT_PRINT_ENTRY("exec","0x%x %d %p \"%s\"",
1708 ev.pl_u.pl_x.pl_pmcid,
1709 ev.pl_u.pl_x.pl_pid,
1710 (void *) ev.pl_u.pl_x.pl_entryaddr,
1711 ev.pl_u.pl_x.pl_pathname);
1713 case PMCLOG_TYPE_PROCEXIT:
1714 PMCSTAT_PRINT_ENTRY("exitval","0x%x %d %jd",
1715 ev.pl_u.pl_e.pl_pmcid,
1716 ev.pl_u.pl_e.pl_pid,
1717 ev.pl_u.pl_e.pl_value);
1719 case PMCLOG_TYPE_PROCFORK:
1720 PMCSTAT_PRINT_ENTRY("fork","%d %d",
1721 ev.pl_u.pl_f.pl_oldpid,
1722 ev.pl_u.pl_f.pl_newpid);
1724 case PMCLOG_TYPE_USERDATA:
1725 PMCSTAT_PRINT_ENTRY("userdata","0x%x",
1726 ev.pl_u.pl_u.pl_userdata);
1728 case PMCLOG_TYPE_SYSEXIT:
1729 PMCSTAT_PRINT_ENTRY("exit","%d",
1730 ev.pl_u.pl_se.pl_pid);
1733 fprintf(args.pa_printfile, "unknown event (type %d).\n",
1738 if (ev.pl_state == PMCLOG_EOF)
1739 return (PMCSTAT_FINISHED);
1740 else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1741 return (PMCSTAT_RUNNING);
1744 "ERROR: event parsing failed (record %jd, offset 0x%jx).",
1745 (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1750 * Public Interfaces.
1754 * Close a logfile, after first flushing all in-module queued data.
1758 pmcstat_close_log(void)
1760 /* If a local logfile is configured ask the kernel to stop
1761 * and flush data. Kernel will close the file when data is flushed
1762 * so keep the status to EXITING.
1764 if (args.pa_logfd != -1) {
1765 if (pmc_close_logfile() < 0)
1766 err(EX_OSERR, "ERROR: logging failed");
1769 return (args.pa_flags & FLAG_HAS_PIPE ? PMCSTAT_EXITING :
1776 * Open a log file, for reading or writing.
1778 * The function returns the fd of a successfully opened log or -1 in
1783 pmcstat_open_log(const char *path, int mode)
1787 const char *p, *errstr;
1788 struct addrinfo hints, *res, *res0;
1789 char hostname[MAXHOSTNAMELEN];
1795 * If 'path' is "-" then open one of stdin or stdout depending
1796 * on the value of 'mode'.
1798 * If 'path' contains a ':' and does not start with a '/' or '.',
1799 * and is being opened for writing, treat it as a "host:port"
1800 * specification and open a network socket.
1802 * Otherwise, treat 'path' as a file name and open that.
1804 if (path[0] == '-' && path[1] == '\0')
1805 fd = (mode == PMCSTAT_OPEN_FOR_READ) ? 0 : 1;
1806 else if (path[0] != '/' &&
1807 path[0] != '.' && strchr(path, ':') != NULL) {
1809 p = strrchr(path, ':');
1811 if (p == path || hlen >= sizeof(hostname)) {
1812 errstr = strerror(EINVAL);
1816 assert(hlen < sizeof(hostname));
1817 (void) strncpy(hostname, path, hlen);
1818 hostname[hlen] = '\0';
1820 (void) memset(&hints, 0, sizeof(hints));
1821 hints.ai_family = AF_UNSPEC;
1822 hints.ai_socktype = SOCK_STREAM;
1823 if ((error = getaddrinfo(hostname, p+1, &hints, &res0)) != 0) {
1824 errstr = gai_strerror(error);
1829 for (res = res0; res; res = res->ai_next) {
1830 if ((fd = socket(res->ai_family, res->ai_socktype,
1831 res->ai_protocol)) < 0) {
1832 errstr = strerror(errno);
1835 if (mode == PMCSTAT_OPEN_FOR_READ) {
1836 if (bind(fd, res->ai_addr, res->ai_addrlen) < 0) {
1837 errstr = strerror(errno);
1843 cfd = accept(fd, NULL, NULL);
1846 errstr = strerror(errno);
1852 if (connect(fd, res->ai_addr, res->ai_addrlen) < 0) {
1853 errstr = strerror(errno);
1864 } else if ((fd = open(path, mode == PMCSTAT_OPEN_FOR_READ ?
1865 O_RDONLY : (O_WRONLY|O_CREAT|O_TRUNC),
1866 S_IRUSR|S_IWUSR|S_IRGRP|S_IROTH)) < 0)
1867 errstr = strerror(errno);
1871 errx(EX_OSERR, "ERROR: Cannot open \"%s\" for %s: %s.", path,
1872 (mode == PMCSTAT_OPEN_FOR_READ ? "reading" : "writing"),
1879 * Process a log file in offline analysis mode.
1883 pmcstat_process_log(void)
1887 * If analysis has not been asked for, just print the log to
1888 * the current output file.
1890 if (args.pa_flags & FLAG_DO_PRINT)
1891 return (pmcstat_print_log());
1893 return (pmcstat_analyze_log());
1897 * Refresh top display.
1901 pmcstat_refresh_top(void)
1906 struct pmcstat_pmcrecord *pmcpr;
1908 /* If in pause mode do not refresh display. */
1912 /* Wait until PMC pop in the log. */
1913 pmcpr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1917 /* Format PMC name. */
1918 if (pmcstat_mergepmc)
1919 snprintf(pmcname, sizeof(pmcname), "[%s]",
1920 pmcstat_string_unintern(pmcpr->pr_pmcname));
1922 snprintf(pmcname, sizeof(pmcname), "%s.%d",
1923 pmcstat_string_unintern(pmcpr->pr_pmcname),
1924 pmcstat_pmcinfilter);
1926 /* Format samples count. */
1927 if (ps_samples_period > 0)
1928 v = (pmcpr->pr_samples * 100.0) / ps_samples_period;
1931 v_attrs = PMCSTAT_ATTRPERCENT(v);
1933 PMCSTAT_PRINTBEGIN();
1934 PMCSTAT_PRINTW("PMC: %s Samples: %u ",
1937 PMCSTAT_ATTRON(v_attrs);
1938 PMCSTAT_PRINTW("(%.1f%%) ", v);
1939 PMCSTAT_ATTROFF(v_attrs);
1940 PMCSTAT_PRINTW(", %u unresolved\n\n",
1941 pmcpr->pr_dubious_frames);
1942 if (plugins[args.pa_plugin].pl_topdisplay != NULL)
1943 plugins[args.pa_plugin].pl_topdisplay();
1948 * Find the next pmc index to display.
1952 pmcstat_changefilter(void)
1955 struct pmcstat_pmcrecord *pmcr;
1958 * Find the next merge target.
1960 if (pmcstat_mergepmc) {
1961 pmcin = pmcstat_pmcinfilter;
1964 pmcr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1965 if (pmcr == NULL || pmcr == pmcr->pr_merge)
1968 pmcstat_pmcinfilter++;
1969 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
1970 pmcstat_pmcinfilter = 0;
1972 } while (pmcstat_pmcinfilter != pmcin);
1977 * Top mode keypress.
1981 pmcstat_keypress_log(void)
1986 w = newwin(1, 0, 1, 0);
1988 wprintw(w, "Key: %c => ", c);
1991 wprintw(w, "enter mode 'd' or 'a' => ");
1994 args.pa_topmode = PMCSTAT_TOP_DELTA;
1995 wprintw(w, "switching to delta mode");
1997 args.pa_topmode = PMCSTAT_TOP_ACCUM;
1998 wprintw(w, "switching to accumulation mode");
2002 pmcstat_mergepmc = !pmcstat_mergepmc;
2004 * Changing merge state require data reset.
2006 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2007 plugins[args.pa_plugin].pl_shutdown(NULL);
2008 pmcstat_stats_reset(0);
2009 if (plugins[args.pa_plugin].pl_init != NULL)
2010 plugins[args.pa_plugin].pl_init();
2012 /* Update filter to be on a merge target. */
2013 pmcstat_changefilter();
2014 wprintw(w, "merge PMC %s", pmcstat_mergepmc ? "on" : "off");
2017 /* Close current plugin. */
2018 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2019 plugins[args.pa_plugin].pl_shutdown(NULL);
2021 /* Find next top display available. */
2024 if (plugins[args.pa_plugin].pl_name == NULL)
2026 } while (plugins[args.pa_plugin].pl_topdisplay == NULL);
2028 /* Open new plugin. */
2029 pmcstat_stats_reset(0);
2030 if (plugins[args.pa_plugin].pl_init != NULL)
2031 plugins[args.pa_plugin].pl_init();
2032 wprintw(w, "switching to plugin %s",
2033 plugins[args.pa_plugin].pl_name);
2036 pmcstat_pmcinfilter++;
2037 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
2038 pmcstat_pmcinfilter = 0;
2039 pmcstat_changefilter();
2040 wprintw(w, "switching to PMC %s.%d",
2041 pmcstat_pmcindex_to_name(pmcstat_pmcinfilter),
2042 pmcstat_pmcinfilter);
2045 pmcstat_pause = !pmcstat_pause;
2047 wprintw(w, "pause => press space again to continue");
2050 wprintw(w, "exiting...");
2054 if (plugins[args.pa_plugin].pl_topkeypress != NULL)
2055 if (plugins[args.pa_plugin].pl_topkeypress(c, w))
2070 pmcstat_display_log(void)
2073 pmcstat_refresh_top();
2075 /* Reset everythings if delta mode. */
2076 if (args.pa_topmode == PMCSTAT_TOP_DELTA) {
2077 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2078 plugins[args.pa_plugin].pl_shutdown(NULL);
2079 pmcstat_stats_reset(0);
2080 if (plugins[args.pa_plugin].pl_init != NULL)
2081 plugins[args.pa_plugin].pl_init();
2087 * Configure a plugins.
2091 pmcstat_pluginconfigure_log(char *opt)
2094 if (strncmp(opt, "threshold=", 10) == 0) {
2095 pmcstat_threshold = atof(opt+10);
2097 if (plugins[args.pa_plugin].pl_configure != NULL) {
2098 if (!plugins[args.pa_plugin].pl_configure(opt))
2100 "ERROR: unknown option <%s>.", opt);
2106 * Initialize module.
2110 pmcstat_initialize_logging(void)
2114 /* use a convenient format for 'ldd' output */
2115 if (setenv("LD_TRACE_LOADED_OBJECTS_FMT1","%o \"%p\" %x\n",1) != 0)
2116 err(EX_OSERR, "ERROR: Cannot setenv");
2118 /* Initialize hash tables */
2119 pmcstat_string_initialize();
2120 for (i = 0; i < PMCSTAT_NHASH; i++) {
2121 LIST_INIT(&pmcstat_image_hash[i]);
2122 LIST_INIT(&pmcstat_process_hash[i]);
2126 * Create a fake 'process' entry for the kernel with pid -1.
2127 * hwpmc(4) will subsequently inform us about where the kernel
2128 * and any loaded kernel modules are mapped.
2130 if ((pmcstat_kernproc = pmcstat_process_lookup((pid_t) -1,
2131 PMCSTAT_ALLOCATE)) == NULL)
2132 err(EX_OSERR, "ERROR: Cannot initialize logging");
2137 /* Merge PMC with same name. */
2138 pmcstat_mergepmc = args.pa_mergepmc;
2141 * Initialize plugins
2144 if (plugins[args.pa_pplugin].pl_init != NULL)
2145 plugins[args.pa_pplugin].pl_init();
2146 if (plugins[args.pa_plugin].pl_init != NULL)
2147 plugins[args.pa_plugin].pl_init();
2155 pmcstat_shutdown_logging(void)
2159 struct pmcstat_image *pi, *pitmp;
2160 struct pmcstat_process *pp, *pptmp;
2161 struct pmcstat_pcmap *ppm, *ppmtmp;
2163 /* determine where to send the map file */
2165 if (args.pa_mapfilename != NULL)
2166 mf = (strcmp(args.pa_mapfilename, "-") == 0) ?
2167 args.pa_printfile : fopen(args.pa_mapfilename, "w");
2169 if (mf == NULL && args.pa_flags & FLAG_DO_GPROF &&
2170 args.pa_verbosity >= 2)
2171 mf = args.pa_printfile;
2174 (void) fprintf(mf, "MAP:\n");
2177 * Shutdown the plugins
2180 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2181 plugins[args.pa_plugin].pl_shutdown(mf);
2182 if (plugins[args.pa_pplugin].pl_shutdown != NULL)
2183 plugins[args.pa_pplugin].pl_shutdown(mf);
2185 for (i = 0; i < PMCSTAT_NHASH; i++) {
2186 LIST_FOREACH_SAFE(pi, &pmcstat_image_hash[i], pi_next,
2188 if (plugins[args.pa_plugin].pl_shutdownimage != NULL)
2189 plugins[args.pa_plugin].pl_shutdownimage(pi);
2190 if (plugins[args.pa_pplugin].pl_shutdownimage != NULL)
2191 plugins[args.pa_pplugin].pl_shutdownimage(pi);
2193 free(pi->pi_symbols);
2194 if (pi->pi_addr2line != NULL)
2195 pclose(pi->pi_addr2line);
2196 LIST_REMOVE(pi, pi_next);
2200 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[i], pp_next,
2202 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
2203 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
2206 LIST_REMOVE(pp, pp_next);
2211 pmcstat_string_shutdown();
2214 * Print errors unless -q was specified. Print all statistics
2217 #define PRINT(N,V) do { \
2218 if (pmcstat_stats.ps_##V || args.pa_verbosity >= 2) \
2219 (void) fprintf(args.pa_printfile, " %-40s %d\n",\
2220 N, pmcstat_stats.ps_##V); \
2223 if (args.pa_verbosity >= 1 && (args.pa_flags & FLAG_DO_ANALYSIS)) {
2224 (void) fprintf(args.pa_printfile, "CONVERSION STATISTICS:\n");
2225 PRINT("#exec/a.out", exec_aout);
2226 PRINT("#exec/elf", exec_elf);
2227 PRINT("#exec/unknown", exec_indeterminable);
2228 PRINT("#exec handling errors", exec_errors);
2229 PRINT("#samples/total", samples_total);
2230 PRINT("#samples/unclaimed", samples_unknown_offset);
2231 PRINT("#samples/unknown-object", samples_indeterminable);
2232 PRINT("#samples/unknown-function", samples_unknown_function);
2233 PRINT("#callchain/dubious-frames", callchain_dubious_frames);