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 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_calltree.h"
155 const char *pl_name; /* name */
158 int (*pl_configure)(char *opt);
160 /* init and shutdown */
161 int (*pl_init)(void);
162 void (*pl_shutdown)(FILE *mf);
164 /* sample processing */
165 void (*pl_process)(struct pmcstat_process *pp,
166 struct pmcstat_pmcrecord *pmcr, uint32_t nsamples,
167 uintfptr_t *cc, int usermode, uint32_t cpu);
170 void (*pl_initimage)(struct pmcstat_image *pi);
171 void (*pl_shutdownimage)(struct pmcstat_image *pi);
174 void (*pl_newpmc)(pmcstat_interned_string ps,
175 struct pmcstat_pmcrecord *pr);
178 void (*pl_topdisplay)(void);
181 int (*pl_topkeypress)(int c, WINDOW *w);
188 .pl_name = "callgraph",
189 .pl_init = pmcpl_cg_init,
190 .pl_shutdown = pmcpl_cg_shutdown,
191 .pl_process = pmcpl_cg_process,
192 .pl_topkeypress = pmcpl_cg_topkeypress,
193 .pl_topdisplay = pmcpl_cg_topdisplay
197 .pl_shutdown = pmcpl_gmon_shutdown,
198 .pl_process = pmcpl_gmon_process,
199 .pl_initimage = pmcpl_gmon_initimage,
200 .pl_shutdownimage = pmcpl_gmon_shutdownimage,
201 .pl_newpmc = pmcpl_gmon_newpmc
204 .pl_name = "annotate",
205 .pl_process = pmcpl_annotate_process
208 .pl_name = "calltree",
209 .pl_configure = pmcpl_ct_configure,
210 .pl_init = pmcpl_ct_init,
211 .pl_shutdown = pmcpl_ct_shutdown,
212 .pl_process = pmcpl_ct_process,
213 .pl_topkeypress = pmcpl_ct_topkeypress,
214 .pl_topdisplay = pmcpl_ct_topdisplay
221 int pmcstat_mergepmc;
223 int pmcstat_pmcinfilter = 0; /* PMC filter for top mode. */
224 float pmcstat_threshold = 0.5; /* Cost filter for top mode. */
230 static struct pmcstat_image *pmcstat_image_from_path(pmcstat_interned_string
231 _path, int _iskernelmodule);
232 static void pmcstat_image_get_aout_params(struct pmcstat_image *_image);
233 static void pmcstat_image_get_elf_params(struct pmcstat_image *_image);
234 static void pmcstat_image_link(struct pmcstat_process *_pp,
235 struct pmcstat_image *_i, uintfptr_t _lpc);
237 static void pmcstat_pmcid_add(pmc_id_t _pmcid,
238 pmcstat_interned_string _name);
240 static void pmcstat_process_aout_exec(struct pmcstat_process *_pp,
241 struct pmcstat_image *_image, uintfptr_t _entryaddr);
242 static void pmcstat_process_elf_exec(struct pmcstat_process *_pp,
243 struct pmcstat_image *_image, uintfptr_t _entryaddr);
244 static void pmcstat_process_exec(struct pmcstat_process *_pp,
245 pmcstat_interned_string _path, uintfptr_t _entryaddr);
246 static struct pmcstat_process *pmcstat_process_lookup(pid_t _pid,
248 static int pmcstat_string_compute_hash(const char *_string);
249 static void pmcstat_string_initialize(void);
250 static int pmcstat_string_lookup_hash(pmcstat_interned_string _is);
251 static void pmcstat_string_shutdown(void);
252 static void pmcstat_stats_reset(int _reset_global);
255 * A simple implementation of interned strings. Each interned string
256 * is assigned a unique address, so that subsequent string compares
257 * can be done by a simple pointer comparision instead of using
258 * strcmp(). This speeds up hash table lookups and saves memory if
259 * duplicate strings are the norm.
261 struct pmcstat_string {
262 LIST_ENTRY(pmcstat_string) ps_next; /* hash link */
268 static LIST_HEAD(,pmcstat_string) pmcstat_string_hash[PMCSTAT_NHASH];
276 * PMC Top mode pause state.
281 pmcstat_stats_reset(int reset_global)
283 struct pmcstat_pmcrecord *pr;
285 /* Flush PMCs stats. */
286 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
288 pr->pr_dubious_frames = 0;
290 ps_samples_period = 0;
292 /* Flush global stats. */
294 bzero(&pmcstat_stats, sizeof(struct pmcstat_stats));
298 * Compute a 'hash' value for a string.
302 pmcstat_string_compute_hash(const char *s)
306 for (hash = 0; *s; s++)
309 return (hash & PMCSTAT_HASH_MASK);
313 * Intern a copy of string 's', and return a pointer to the
314 * interned structure.
317 pmcstat_interned_string
318 pmcstat_string_intern(const char *s)
320 struct pmcstat_string *ps;
321 const struct pmcstat_string *cps;
324 if ((cps = pmcstat_string_lookup(s)) != NULL)
327 hash = pmcstat_string_compute_hash(s);
330 if ((ps = malloc(sizeof(*ps))) == NULL)
331 err(EX_OSERR, "ERROR: Could not intern string");
334 ps->ps_string = strdup(s);
335 LIST_INSERT_HEAD(&pmcstat_string_hash[hash], ps, ps_next);
336 return ((pmcstat_interned_string) ps);
340 pmcstat_string_unintern(pmcstat_interned_string str)
344 s = ((const struct pmcstat_string *) str)->ps_string;
348 pmcstat_interned_string
349 pmcstat_string_lookup(const char *s)
351 struct pmcstat_string *ps;
354 hash = pmcstat_string_compute_hash(s);
357 LIST_FOREACH(ps, &pmcstat_string_hash[hash], ps_next)
358 if (ps->ps_len == len && ps->ps_hash == hash &&
359 strcmp(ps->ps_string, s) == 0)
365 pmcstat_string_lookup_hash(pmcstat_interned_string s)
367 const struct pmcstat_string *ps;
369 ps = (const struct pmcstat_string *) s;
370 return (ps->ps_hash);
374 * Initialize the string interning facility.
378 pmcstat_string_initialize(void)
382 for (i = 0; i < PMCSTAT_NHASH; i++)
383 LIST_INIT(&pmcstat_string_hash[i]);
387 * Destroy the string table, free'ing up space.
391 pmcstat_string_shutdown(void)
394 struct pmcstat_string *ps, *pstmp;
396 for (i = 0; i < PMCSTAT_NHASH; i++)
397 LIST_FOREACH_SAFE(ps, &pmcstat_string_hash[i], ps_next,
399 LIST_REMOVE(ps, ps_next);
406 * Determine whether a given executable image is an A.OUT object, and
407 * if so, fill in its parameters from the text file.
408 * Sets image->pi_type.
412 pmcstat_image_get_aout_params(struct pmcstat_image *image)
418 char buffer[PATH_MAX];
420 path = pmcstat_string_unintern(image->pi_execpath);
421 assert(path != NULL);
423 if (image->pi_iskernelmodule)
424 errx(EX_SOFTWARE, "ERROR: a.out kernel modules are "
425 "unsupported \"%s\"", path);
427 (void) snprintf(buffer, sizeof(buffer), "%s%s",
428 args.pa_fsroot, path);
430 if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
431 (nbytes = read(fd, &ex, sizeof(ex))) < 0) {
432 warn("WARNING: Cannot determine type of \"%s\"", path);
433 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
441 if ((unsigned) nbytes != sizeof(ex) ||
445 image->pi_type = PMCSTAT_IMAGE_AOUT;
447 /* TODO: the rest of a.out processing */
457 pmcstat_symbol_compare(const void *a, const void *b)
459 const struct pmcstat_symbol *sym1, *sym2;
461 sym1 = (const struct pmcstat_symbol *) a;
462 sym2 = (const struct pmcstat_symbol *) b;
464 if (sym1->ps_end <= sym2->ps_start)
466 if (sym1->ps_start >= sym2->ps_end)
472 * Map an address to a symbol in an image.
475 struct pmcstat_symbol *
476 pmcstat_symbol_search(struct pmcstat_image *image, uintfptr_t addr)
478 struct pmcstat_symbol sym;
480 if (image->pi_symbols == NULL)
485 sym.ps_end = addr + 1;
487 return (bsearch((void *) &sym, image->pi_symbols,
488 image->pi_symcount, sizeof(struct pmcstat_symbol),
489 pmcstat_symbol_compare));
493 * Add the list of symbols in the given section to the list associated
497 pmcstat_image_add_symbols(struct pmcstat_image *image, Elf *e,
498 Elf_Scn *scn, GElf_Shdr *sh)
501 size_t n, newsyms, nshsyms, nfuncsyms;
502 struct pmcstat_symbol *symptr;
507 if ((data = elf_getdata(scn, NULL)) == NULL)
511 * Determine the number of functions named in this
515 nshsyms = sh->sh_size / sh->sh_entsize;
516 for (n = nfuncsyms = 0; n < nshsyms; n++) {
517 if (gelf_getsym(data, (int) n, &sym) != &sym)
519 if (GELF_ST_TYPE(sym.st_info) == STT_FUNC)
527 * Allocate space for the new entries.
529 firsttime = image->pi_symbols == NULL;
530 symptr = realloc(image->pi_symbols,
531 sizeof(*symptr) * (image->pi_symcount + nfuncsyms));
532 if (symptr == image->pi_symbols) /* realloc() failed. */
534 image->pi_symbols = symptr;
537 * Append new symbols to the end of the current table.
539 symptr += image->pi_symcount;
541 for (n = newsyms = 0; n < nshsyms; n++) {
542 if (gelf_getsym(data, (int) n, &sym) != &sym)
544 if (GELF_ST_TYPE(sym.st_info) != STT_FUNC)
546 if (sym.st_shndx == STN_UNDEF)
549 if (!firsttime && pmcstat_symbol_search(image, sym.st_value))
550 continue; /* We've seen this symbol already. */
552 if ((fnname = elf_strptr(e, sh->sh_link, sym.st_name))
556 symptr->ps_name = pmcstat_string_intern(fnname);
557 symptr->ps_start = sym.st_value - image->pi_vaddr;
558 symptr->ps_end = symptr->ps_start + sym.st_size;
564 image->pi_symcount += newsyms;
566 assert(newsyms <= nfuncsyms);
569 * Return space to the system if there were duplicates.
571 if (newsyms < nfuncsyms)
572 image->pi_symbols = realloc(image->pi_symbols,
573 sizeof(*symptr) * image->pi_symcount);
576 * Keep the list of symbols sorted.
578 qsort(image->pi_symbols, image->pi_symcount, sizeof(*symptr),
579 pmcstat_symbol_compare);
582 * Deal with function symbols that have a size of 'zero' by
583 * making them extend to the next higher address. These
584 * symbols are usually defined in assembly code.
586 for (symptr = image->pi_symbols;
587 symptr < image->pi_symbols + (image->pi_symcount - 1);
589 if (symptr->ps_start == symptr->ps_end)
590 symptr->ps_end = (symptr+1)->ps_start;
594 * Examine an ELF file to determine the size of its text segment.
595 * Sets image->pi_type if anything conclusive can be determined about
600 pmcstat_image_get_elf_params(struct pmcstat_image *image)
604 const char *path, *elfbase;
606 uintfptr_t minva, maxva;
612 enum pmcstat_image_type image_type;
613 char buffer[PATH_MAX];
615 assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
617 image->pi_start = minva = ~(uintfptr_t) 0;
618 image->pi_end = maxva = (uintfptr_t) 0;
619 image->pi_type = image_type = PMCSTAT_IMAGE_INDETERMINABLE;
620 image->pi_isdynamic = 0;
621 image->pi_dynlinkerpath = NULL;
624 path = pmcstat_string_unintern(image->pi_execpath);
625 assert(path != NULL);
628 * Look for kernel modules under FSROOT/KERNELPATH/NAME,
629 * and user mode executable objects under FSROOT/PATHNAME.
631 if (image->pi_iskernelmodule)
632 (void) snprintf(buffer, sizeof(buffer), "%s%s/%s",
633 args.pa_fsroot, args.pa_kernel, path);
635 (void) snprintf(buffer, sizeof(buffer), "%s%s",
636 args.pa_fsroot, path);
639 if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
640 (e = elf_begin(fd, ELF_C_READ, NULL)) == NULL ||
641 (elf_kind(e) != ELF_K_ELF)) {
642 warnx("WARNING: Cannot determine the type of \"%s\".",
647 if (gelf_getehdr(e, &eh) != &eh) {
648 warnx("WARNING: Cannot retrieve the ELF Header for "
649 "\"%s\": %s.", buffer, elf_errmsg(-1));
653 if (eh.e_type != ET_EXEC && eh.e_type != ET_DYN &&
654 !(image->pi_iskernelmodule && eh.e_type == ET_REL)) {
655 warnx("WARNING: \"%s\" is of an unsupported ELF type.",
660 image_type = eh.e_ident[EI_CLASS] == ELFCLASS32 ?
661 PMCSTAT_IMAGE_ELF32 : PMCSTAT_IMAGE_ELF64;
664 * Determine the virtual address where an executable would be
665 * loaded. Additionally, for dynamically linked executables,
666 * save the pathname to the runtime linker.
668 if (eh.e_type == ET_EXEC) {
669 if (elf_getphnum(e, &nph) == 0) {
670 warnx("WARNING: Could not determine the number of "
671 "program headers in \"%s\": %s.", buffer,
675 for (i = 0; i < eh.e_phnum; i++) {
676 if (gelf_getphdr(e, i, &ph) != &ph) {
677 warnx("WARNING: Retrieval of PHDR entry #%ju "
678 "in \"%s\" failed: %s.", (uintmax_t) i,
679 buffer, elf_errmsg(-1));
684 image->pi_isdynamic = 1;
687 if ((elfbase = elf_rawfile(e, NULL)) == NULL) {
688 warnx("WARNING: Cannot retrieve the "
689 "interpreter for \"%s\": %s.",
690 buffer, elf_errmsg(-1));
693 image->pi_dynlinkerpath =
694 pmcstat_string_intern(elfbase +
698 if (ph.p_offset == 0)
699 image->pi_vaddr = ph.p_vaddr;
706 * Get the min and max VA associated with this ELF object.
708 if (elf_getshnum(e, &nsh) == 0) {
709 warnx("WARNING: Could not determine the number of sections "
710 "for \"%s\": %s.", buffer, elf_errmsg(-1));
714 for (i = 0; i < nsh; i++) {
715 if ((scn = elf_getscn(e, i)) == NULL ||
716 gelf_getshdr(scn, &sh) != &sh) {
717 warnx("WARNING: Could not retrieve section header "
718 "#%ju in \"%s\": %s.", (uintmax_t) i, buffer,
722 if (sh.sh_flags & SHF_EXECINSTR) {
723 minva = min(minva, sh.sh_addr);
724 maxva = max(maxva, sh.sh_addr + sh.sh_size);
726 if (sh.sh_type == SHT_SYMTAB || sh.sh_type == SHT_DYNSYM)
727 pmcstat_image_add_symbols(image, e, scn, &sh);
730 image->pi_start = minva;
731 image->pi_end = maxva;
732 image->pi_type = image_type;
733 image->pi_fullpath = pmcstat_string_intern(buffer);
735 /* Build display name
738 for (p = buffer; *p; p++)
741 image->pi_name = pmcstat_string_intern(endp);
751 * Given an image descriptor, determine whether it is an ELF, or AOUT.
752 * If no handler claims the image, set its type to 'INDETERMINABLE'.
756 pmcstat_image_determine_type(struct pmcstat_image *image)
758 assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
760 /* Try each kind of handler in turn */
761 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
762 pmcstat_image_get_elf_params(image);
763 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
764 pmcstat_image_get_aout_params(image);
767 * Otherwise, remember that we tried to determine
768 * the object's type and had failed.
770 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
771 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
775 * Locate an image descriptor given an interned path, adding a fresh
776 * descriptor to the cache if necessary. This function also finds a
777 * suitable name for this image's sample file.
779 * We defer filling in the file format specific parts of the image
780 * structure till the time we actually see a sample that would fall
784 static struct pmcstat_image *
785 pmcstat_image_from_path(pmcstat_interned_string internedpath,
789 struct pmcstat_image *pi;
791 hash = pmcstat_string_lookup_hash(internedpath);
793 /* First, look for an existing entry. */
794 LIST_FOREACH(pi, &pmcstat_image_hash[hash], pi_next)
795 if (pi->pi_execpath == internedpath &&
796 pi->pi_iskernelmodule == iskernelmodule)
800 * Allocate a new entry and place it at the head of the hash
803 pi = malloc(sizeof(*pi));
807 pi->pi_type = PMCSTAT_IMAGE_UNKNOWN;
808 pi->pi_execpath = internedpath;
813 pi->pi_isdynamic = 0;
814 pi->pi_iskernelmodule = iskernelmodule;
815 pi->pi_dynlinkerpath = NULL;
816 pi->pi_symbols = NULL;
818 pi->pi_addr2line = NULL;
820 if (plugins[args.pa_pplugin].pl_initimage != NULL)
821 plugins[args.pa_pplugin].pl_initimage(pi);
822 if (plugins[args.pa_plugin].pl_initimage != NULL)
823 plugins[args.pa_plugin].pl_initimage(pi);
825 LIST_INSERT_HEAD(&pmcstat_image_hash[hash], pi, pi_next);
831 * Record the fact that PC values from 'start' to 'end' come from
836 pmcstat_image_link(struct pmcstat_process *pp, struct pmcstat_image *image,
839 struct pmcstat_pcmap *pcm, *pcmnew;
842 assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN &&
843 image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE);
845 if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
846 err(EX_OSERR, "ERROR: Cannot create a map entry");
849 * Adjust the map entry to only cover the text portion
853 offset = start - image->pi_vaddr;
854 pcmnew->ppm_lowpc = image->pi_start + offset;
855 pcmnew->ppm_highpc = image->pi_end + offset;
856 pcmnew->ppm_image = image;
858 assert(pcmnew->ppm_lowpc < pcmnew->ppm_highpc);
860 /* Overlapped mmap()'s are assumed to never occur. */
861 TAILQ_FOREACH(pcm, &pp->pp_map, ppm_next)
862 if (pcm->ppm_lowpc >= pcmnew->ppm_highpc)
866 TAILQ_INSERT_TAIL(&pp->pp_map, pcmnew, ppm_next);
868 TAILQ_INSERT_BEFORE(pcm, pcmnew, ppm_next);
872 * Unmap images in the range [start..end) associated with process
877 pmcstat_image_unmap(struct pmcstat_process *pp, uintfptr_t start,
880 struct pmcstat_pcmap *pcm, *pcmtmp, *pcmnew;
887 * - we could have the range completely in the middle of an
888 * existing pcmap; in this case we have to split the pcmap
889 * structure into two (i.e., generate a 'hole').
890 * - we could have the range covering multiple pcmaps; these
891 * will have to be removed.
892 * - we could have either 'start' or 'end' falling in the
893 * middle of a pcmap; in this case shorten the entry.
895 TAILQ_FOREACH_SAFE(pcm, &pp->pp_map, ppm_next, pcmtmp) {
896 assert(pcm->ppm_lowpc < pcm->ppm_highpc);
897 if (pcm->ppm_highpc <= start)
899 if (pcm->ppm_lowpc >= end)
901 if (pcm->ppm_lowpc >= start && pcm->ppm_highpc <= end) {
903 * The current pcmap is completely inside the
904 * unmapped range: remove it entirely.
906 TAILQ_REMOVE(&pp->pp_map, pcm, ppm_next);
908 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc > end) {
910 * Split this pcmap into two; curtail the
911 * current map to end at [start-1], and start
912 * the new one at [end].
914 if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
915 err(EX_OSERR, "ERROR: Cannot split a map "
918 pcmnew->ppm_image = pcm->ppm_image;
920 pcmnew->ppm_lowpc = end;
921 pcmnew->ppm_highpc = pcm->ppm_highpc;
923 pcm->ppm_highpc = start;
925 TAILQ_INSERT_AFTER(&pp->pp_map, pcm, pcmnew, ppm_next);
928 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc <= end)
929 pcm->ppm_highpc = start;
930 else if (pcm->ppm_lowpc >= start && pcm->ppm_highpc > end)
931 pcm->ppm_lowpc = end;
938 * Resolve file name and line number for the given address.
941 pmcstat_image_addr2line(struct pmcstat_image *image, uintfptr_t addr,
942 char *sourcefile, size_t sourcefile_len, unsigned *sourceline,
943 char *funcname, size_t funcname_len)
945 static int addr2line_warn = 0;
947 char *sep, cmdline[PATH_MAX], imagepath[PATH_MAX];
950 if (image->pi_addr2line == NULL) {
951 snprintf(imagepath, sizeof(imagepath), "%s%s.symbols",
953 pmcstat_string_unintern(image->pi_fullpath));
954 fd = open(imagepath, O_RDONLY);
956 snprintf(imagepath, sizeof(imagepath), "%s%s",
958 pmcstat_string_unintern(image->pi_fullpath));
961 snprintf(cmdline, sizeof(cmdline), "addr2line -Cfe \"%s\"",
963 image->pi_addr2line = popen(cmdline, "r+");
964 if (image->pi_addr2line == NULL) {
965 if (!addr2line_warn) {
967 warnx("WARNING: addr2line is needed"
968 "for source code information.");
974 if (feof(image->pi_addr2line) || ferror(image->pi_addr2line)) {
975 warnx("WARNING: addr2line pipe error");
976 pclose(image->pi_addr2line);
977 image->pi_addr2line = NULL;
981 fprintf(image->pi_addr2line, "%p\n", (void *)addr);
983 if (fgets(funcname, funcname_len, image->pi_addr2line) == NULL) {
984 warnx("WARNING: addr2line function name read error");
987 sep = strchr(funcname, '\n');
991 if (fgets(sourcefile, sourcefile_len, image->pi_addr2line) == NULL) {
992 warnx("WARNING: addr2line source file read error");
995 sep = strchr(sourcefile, ':');
997 warnx("WARNING: addr2line source line separator missing");
1001 *sourceline = atoi(sep+1);
1002 if (*sourceline == 0)
1009 * Add a {pmcid,name} mapping.
1013 pmcstat_pmcid_add(pmc_id_t pmcid, pmcstat_interned_string ps)
1015 struct pmcstat_pmcrecord *pr, *prm;
1017 /* Replace an existing name for the PMC. */
1019 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1020 if (pr->pr_pmcid == pmcid) {
1021 pr->pr_pmcname = ps;
1023 } else if (pr->pr_pmcname == ps)
1027 * Otherwise, allocate a new descriptor and call the
1030 if ((pr = malloc(sizeof(*pr))) == NULL)
1031 err(EX_OSERR, "ERROR: Cannot allocate pmc record");
1033 pr->pr_pmcid = pmcid;
1034 pr->pr_pmcname = ps;
1035 pr->pr_pmcin = pmcstat_npmcs++;
1037 pr->pr_dubious_frames = 0;
1038 pr->pr_merge = prm == NULL ? pr : prm;
1040 LIST_INSERT_HEAD(&pmcstat_pmcs, pr, pr_next);
1042 if (plugins[args.pa_pplugin].pl_newpmc != NULL)
1043 plugins[args.pa_pplugin].pl_newpmc(ps, pr);
1044 if (plugins[args.pa_plugin].pl_newpmc != NULL)
1045 plugins[args.pa_plugin].pl_newpmc(ps, pr);
1049 * Given a pmcid in use, find its human-readable name.
1053 pmcstat_pmcid_to_name(pmc_id_t pmcid)
1055 struct pmcstat_pmcrecord *pr;
1057 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1058 if (pr->pr_pmcid == pmcid)
1059 return (pmcstat_string_unintern(pr->pr_pmcname));
1065 * Convert PMC index to name.
1069 pmcstat_pmcindex_to_name(int pmcin)
1071 struct pmcstat_pmcrecord *pr;
1073 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1074 if (pr->pr_pmcin == pmcin)
1075 return pmcstat_string_unintern(pr->pr_pmcname);
1081 * Return PMC record with given index.
1084 struct pmcstat_pmcrecord *
1085 pmcstat_pmcindex_to_pmcr(int pmcin)
1087 struct pmcstat_pmcrecord *pr;
1089 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1090 if (pr->pr_pmcin == pmcin)
1097 * Get PMC record by id, apply merge policy.
1100 static struct pmcstat_pmcrecord *
1101 pmcstat_lookup_pmcid(pmc_id_t pmcid)
1103 struct pmcstat_pmcrecord *pr;
1105 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
1106 if (pr->pr_pmcid == pmcid) {
1107 if (pmcstat_mergepmc)
1108 return pr->pr_merge;
1117 * Associate an AOUT image with a process.
1121 pmcstat_process_aout_exec(struct pmcstat_process *pp,
1122 struct pmcstat_image *image, uintfptr_t entryaddr)
1127 /* TODO Implement a.out handling */
1131 * Associate an ELF image with a process.
1135 pmcstat_process_elf_exec(struct pmcstat_process *pp,
1136 struct pmcstat_image *image, uintfptr_t entryaddr)
1139 struct pmcstat_image *rtldimage;
1141 assert(image->pi_type == PMCSTAT_IMAGE_ELF32 ||
1142 image->pi_type == PMCSTAT_IMAGE_ELF64);
1144 /* Create a map entry for the base executable. */
1145 pmcstat_image_link(pp, image, image->pi_vaddr);
1148 * For dynamically linked executables we need to determine
1149 * where the dynamic linker was mapped to for this process,
1150 * Subsequent executable objects that are mapped in by the
1151 * dynamic linker will be tracked by log events of type
1152 * PMCLOG_TYPE_MAP_IN.
1155 if (image->pi_isdynamic) {
1158 * The runtime loader gets loaded just after the maximum
1159 * possible heap address. Like so:
1161 * [ TEXT DATA BSS HEAP -->*RTLD SHLIBS <--STACK]
1163 * 0 VM_MAXUSER_ADDRESS
1166 * The exact address where the loader gets mapped in
1167 * will vary according to the size of the executable
1168 * and the limits on the size of the process'es data
1169 * segment at the time of exec(). The entry address
1170 * recorded at process exec time corresponds to the
1171 * 'start' address inside the dynamic linker. From
1172 * this we can figure out the address where the
1173 * runtime loader's file object had been mapped to.
1175 rtldimage = pmcstat_image_from_path(image->pi_dynlinkerpath, 0);
1176 if (rtldimage == NULL) {
1177 warnx("WARNING: Cannot find image for \"%s\".",
1178 pmcstat_string_unintern(image->pi_dynlinkerpath));
1179 pmcstat_stats.ps_exec_errors++;
1183 if (rtldimage->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1184 pmcstat_image_get_elf_params(rtldimage);
1186 if (rtldimage->pi_type != PMCSTAT_IMAGE_ELF32 &&
1187 rtldimage->pi_type != PMCSTAT_IMAGE_ELF64) {
1188 warnx("WARNING: rtld not an ELF object \"%s\".",
1189 pmcstat_string_unintern(image->pi_dynlinkerpath));
1193 libstart = entryaddr - rtldimage->pi_entry;
1194 pmcstat_image_link(pp, rtldimage, libstart);
1199 * Find the process descriptor corresponding to a PID. If 'allocate'
1200 * is zero, we return a NULL if a pid descriptor could not be found or
1201 * a process descriptor process. If 'allocate' is non-zero, then we
1202 * will attempt to allocate a fresh process descriptor. Zombie
1203 * process descriptors are only removed if a fresh allocation for the
1204 * same PID is requested.
1207 static struct pmcstat_process *
1208 pmcstat_process_lookup(pid_t pid, int allocate)
1211 struct pmcstat_pcmap *ppm, *ppmtmp;
1212 struct pmcstat_process *pp, *pptmp;
1214 hash = (uint32_t) pid & PMCSTAT_HASH_MASK; /* simplicity wins */
1216 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[hash], pp_next, pptmp)
1217 if (pp->pp_pid == pid) {
1218 /* Found a descriptor, check and process zombies */
1219 if (allocate && pp->pp_isactive == 0) {
1221 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next,
1223 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
1226 /* remove process entry */
1227 LIST_REMOVE(pp, pp_next);
1237 if ((pp = malloc(sizeof(*pp))) == NULL)
1238 err(EX_OSERR, "ERROR: Cannot allocate pid descriptor");
1241 pp->pp_isactive = 1;
1243 TAILQ_INIT(&pp->pp_map);
1245 LIST_INSERT_HEAD(&pmcstat_process_hash[hash], pp, pp_next);
1250 * Associate an image and a process.
1254 pmcstat_process_exec(struct pmcstat_process *pp,
1255 pmcstat_interned_string path, uintfptr_t entryaddr)
1257 struct pmcstat_image *image;
1259 if ((image = pmcstat_image_from_path(path, 0)) == NULL) {
1260 pmcstat_stats.ps_exec_errors++;
1264 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1265 pmcstat_image_determine_type(image);
1267 assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN);
1269 switch (image->pi_type) {
1270 case PMCSTAT_IMAGE_ELF32:
1271 case PMCSTAT_IMAGE_ELF64:
1272 pmcstat_stats.ps_exec_elf++;
1273 pmcstat_process_elf_exec(pp, image, entryaddr);
1276 case PMCSTAT_IMAGE_AOUT:
1277 pmcstat_stats.ps_exec_aout++;
1278 pmcstat_process_aout_exec(pp, image, entryaddr);
1281 case PMCSTAT_IMAGE_INDETERMINABLE:
1282 pmcstat_stats.ps_exec_indeterminable++;
1286 err(EX_SOFTWARE, "ERROR: Unsupported executable type for "
1287 "\"%s\"", pmcstat_string_unintern(path));
1293 * Find the map entry associated with process 'p' at PC value 'pc'.
1296 struct pmcstat_pcmap *
1297 pmcstat_process_find_map(struct pmcstat_process *p, uintfptr_t pc)
1299 struct pmcstat_pcmap *ppm;
1301 TAILQ_FOREACH(ppm, &p->pp_map, ppm_next) {
1302 if (pc >= ppm->ppm_lowpc && pc < ppm->ppm_highpc)
1304 if (pc < ppm->ppm_lowpc)
1312 * Convert a hwpmc(4) log to profile information. A system-wide
1313 * callgraph is generated if FLAG_DO_CALLGRAPHS is set. gmon.out
1314 * files usable by gprof(1) are created if FLAG_DO_GPROF is set.
1317 pmcstat_analyze_log(void)
1319 uint32_t cpu, cpuflags;
1322 struct pmcstat_image *image;
1323 struct pmcstat_process *pp, *ppnew;
1324 struct pmcstat_pcmap *ppm, *ppmtmp;
1325 struct pmclog_ev ev;
1326 struct pmcstat_pmcrecord *pmcr;
1327 pmcstat_interned_string image_path;
1329 assert(args.pa_flags & FLAG_DO_ANALYSIS);
1331 if (elf_version(EV_CURRENT) == EV_NONE)
1332 err(EX_UNAVAILABLE, "Elf library intialization failed");
1334 while (pmclog_read(args.pa_logparser, &ev) == 0) {
1335 assert(ev.pl_state == PMCLOG_OK);
1337 switch (ev.pl_type) {
1338 case PMCLOG_TYPE_INITIALIZE:
1339 if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1340 PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1341 warnx("WARNING: Log version 0x%x does not "
1342 "match compiled version 0x%x.",
1343 ev.pl_u.pl_i.pl_version,
1347 case PMCLOG_TYPE_MAP_IN:
1349 * Introduce an address range mapping for a
1350 * userland process or the kernel (pid == -1).
1352 * We always allocate a process descriptor so
1353 * that subsequent samples seen for this
1354 * address range are mapped to the current
1355 * object being mapped in.
1357 pid = ev.pl_u.pl_mi.pl_pid;
1359 pp = pmcstat_kernproc;
1361 pp = pmcstat_process_lookup(pid,
1366 image_path = pmcstat_string_intern(ev.pl_u.pl_mi.
1368 image = pmcstat_image_from_path(image_path, pid == -1);
1369 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1370 pmcstat_image_determine_type(image);
1371 if (image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE)
1372 pmcstat_image_link(pp, image,
1373 ev.pl_u.pl_mi.pl_start);
1376 case PMCLOG_TYPE_MAP_OUT:
1378 * Remove an address map.
1380 pid = ev.pl_u.pl_mo.pl_pid;
1382 pp = pmcstat_kernproc;
1384 pp = pmcstat_process_lookup(pid, 0);
1386 if (pp == NULL) /* unknown process */
1389 pmcstat_image_unmap(pp, ev.pl_u.pl_mo.pl_start,
1390 ev.pl_u.pl_mo.pl_end);
1393 case PMCLOG_TYPE_PCSAMPLE:
1395 * Note: the `PCSAMPLE' log entry is not
1396 * generated by hpwmc(4) after version 2.
1400 * We bring in the gmon file for the image
1401 * currently associated with the PMC & pid
1402 * pair and increment the appropriate entry
1405 pmcstat_stats.ps_samples_total++;
1406 ps_samples_period++;
1408 pc = ev.pl_u.pl_s.pl_pc;
1409 pp = pmcstat_process_lookup(ev.pl_u.pl_s.pl_pid,
1412 /* Get PMC record. */
1413 pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_s.pl_pmcid);
1414 assert(pmcr != NULL);
1418 * Call the plugins processing
1419 * TODO: move pmcstat_process_find_map inside plugins
1422 if (plugins[args.pa_pplugin].pl_process != NULL)
1423 plugins[args.pa_pplugin].pl_process(
1425 pmcstat_process_find_map(pp, pc) != NULL, 0);
1426 plugins[args.pa_plugin].pl_process(
1428 pmcstat_process_find_map(pp, pc) != NULL, 0);
1431 case PMCLOG_TYPE_CALLCHAIN:
1432 pmcstat_stats.ps_samples_total++;
1433 ps_samples_period++;
1435 cpuflags = ev.pl_u.pl_cc.pl_cpuflags;
1436 cpu = PMC_CALLCHAIN_CPUFLAGS_TO_CPU(cpuflags);
1438 /* Filter on the CPU id. */
1439 if (!CPU_ISSET(cpu, &(args.pa_cpumask))) {
1440 pmcstat_stats.ps_samples_skipped++;
1444 pp = pmcstat_process_lookup(ev.pl_u.pl_cc.pl_pid,
1447 /* Get PMC record. */
1448 pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_cc.pl_pmcid);
1449 assert(pmcr != NULL);
1453 * Call the plugins processing
1456 if (plugins[args.pa_pplugin].pl_process != NULL)
1457 plugins[args.pa_pplugin].pl_process(
1459 ev.pl_u.pl_cc.pl_npc,
1460 ev.pl_u.pl_cc.pl_pc,
1461 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1463 plugins[args.pa_plugin].pl_process(
1465 ev.pl_u.pl_cc.pl_npc,
1466 ev.pl_u.pl_cc.pl_pc,
1467 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1471 case PMCLOG_TYPE_PMCALLOCATE:
1473 * Record the association pmc id between this
1476 pmcstat_pmcid_add(ev.pl_u.pl_a.pl_pmcid,
1477 pmcstat_string_intern(ev.pl_u.pl_a.pl_evname));
1480 case PMCLOG_TYPE_PROCEXEC:
1483 * Change the executable image associated with
1486 pp = pmcstat_process_lookup(ev.pl_u.pl_x.pl_pid,
1489 /* delete the current process map */
1490 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
1491 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
1495 /* associate this process image */
1496 image_path = pmcstat_string_intern(
1497 ev.pl_u.pl_x.pl_pathname);
1498 assert(image_path != NULL);
1499 pmcstat_process_exec(pp, image_path,
1500 ev.pl_u.pl_x.pl_entryaddr);
1503 case PMCLOG_TYPE_PROCEXIT:
1506 * Due to the way the log is generated, the
1507 * last few samples corresponding to a process
1508 * may appear in the log after the process
1509 * exit event is recorded. Thus we keep the
1510 * process' descriptor and associated data
1511 * structures around, but mark the process as
1514 pp = pmcstat_process_lookup(ev.pl_u.pl_e.pl_pid, 0);
1517 pp->pp_isactive = 0; /* mark as a zombie */
1520 case PMCLOG_TYPE_SYSEXIT:
1521 pp = pmcstat_process_lookup(ev.pl_u.pl_se.pl_pid, 0);
1524 pp->pp_isactive = 0; /* make a zombie */
1527 case PMCLOG_TYPE_PROCFORK:
1530 * Allocate a process descriptor for the new
1534 pmcstat_process_lookup(ev.pl_u.pl_f.pl_newpid,
1538 * If we had been tracking the parent, clone
1541 pp = pmcstat_process_lookup(ev.pl_u.pl_f.pl_oldpid, 0);
1544 TAILQ_FOREACH(ppm, &pp->pp_map, ppm_next)
1545 pmcstat_image_link(ppnew, ppm->ppm_image,
1549 default: /* other types of entries are not relevant */
1554 if (ev.pl_state == PMCLOG_EOF)
1555 return (PMCSTAT_FINISHED);
1556 else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1557 return (PMCSTAT_RUNNING);
1559 err(EX_DATAERR, "ERROR: event parsing failed (record %jd, "
1560 "offset 0x%jx)", (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1564 * Print log entries as text.
1568 pmcstat_print_log(void)
1570 struct pmclog_ev ev;
1573 while (pmclog_read(args.pa_logparser, &ev) == 0) {
1574 assert(ev.pl_state == PMCLOG_OK);
1575 switch (ev.pl_type) {
1576 case PMCLOG_TYPE_CALLCHAIN:
1577 PMCSTAT_PRINT_ENTRY("callchain",
1578 "%d 0x%x %d %d %c", ev.pl_u.pl_cc.pl_pid,
1579 ev.pl_u.pl_cc.pl_pmcid,
1580 PMC_CALLCHAIN_CPUFLAGS_TO_CPU(ev.pl_u.pl_cc. \
1581 pl_cpuflags), ev.pl_u.pl_cc.pl_npc,
1582 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(ev.pl_u.pl_cc.\
1583 pl_cpuflags) ? 'u' : 's');
1584 for (npc = 0; npc < ev.pl_u.pl_cc.pl_npc; npc++)
1585 PMCSTAT_PRINT_ENTRY("...", "%p",
1586 (void *) ev.pl_u.pl_cc.pl_pc[npc]);
1588 case PMCLOG_TYPE_CLOSELOG:
1589 PMCSTAT_PRINT_ENTRY("closelog",);
1591 case PMCLOG_TYPE_DROPNOTIFY:
1592 PMCSTAT_PRINT_ENTRY("drop",);
1594 case PMCLOG_TYPE_INITIALIZE:
1595 PMCSTAT_PRINT_ENTRY("initlog","0x%x \"%s\"",
1596 ev.pl_u.pl_i.pl_version,
1597 pmc_name_of_cputype(ev.pl_u.pl_i.pl_arch));
1598 if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1599 PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1600 warnx("WARNING: Log version 0x%x != expected "
1601 "version 0x%x.", ev.pl_u.pl_i.pl_version,
1604 case PMCLOG_TYPE_MAP_IN:
1605 PMCSTAT_PRINT_ENTRY("map-in","%d %p \"%s\"",
1606 ev.pl_u.pl_mi.pl_pid,
1607 (void *) ev.pl_u.pl_mi.pl_start,
1608 ev.pl_u.pl_mi.pl_pathname);
1610 case PMCLOG_TYPE_MAP_OUT:
1611 PMCSTAT_PRINT_ENTRY("map-out","%d %p %p",
1612 ev.pl_u.pl_mo.pl_pid,
1613 (void *) ev.pl_u.pl_mo.pl_start,
1614 (void *) ev.pl_u.pl_mo.pl_end);
1616 case PMCLOG_TYPE_PCSAMPLE:
1617 PMCSTAT_PRINT_ENTRY("sample","0x%x %d %p %c",
1618 ev.pl_u.pl_s.pl_pmcid,
1619 ev.pl_u.pl_s.pl_pid,
1620 (void *) ev.pl_u.pl_s.pl_pc,
1621 ev.pl_u.pl_s.pl_usermode ? 'u' : 's');
1623 case PMCLOG_TYPE_PMCALLOCATE:
1624 PMCSTAT_PRINT_ENTRY("allocate","0x%x \"%s\" 0x%x",
1625 ev.pl_u.pl_a.pl_pmcid,
1626 ev.pl_u.pl_a.pl_evname,
1627 ev.pl_u.pl_a.pl_flags);
1629 case PMCLOG_TYPE_PMCATTACH:
1630 PMCSTAT_PRINT_ENTRY("attach","0x%x %d \"%s\"",
1631 ev.pl_u.pl_t.pl_pmcid,
1632 ev.pl_u.pl_t.pl_pid,
1633 ev.pl_u.pl_t.pl_pathname);
1635 case PMCLOG_TYPE_PMCDETACH:
1636 PMCSTAT_PRINT_ENTRY("detach","0x%x %d",
1637 ev.pl_u.pl_d.pl_pmcid,
1638 ev.pl_u.pl_d.pl_pid);
1640 case PMCLOG_TYPE_PROCCSW:
1641 PMCSTAT_PRINT_ENTRY("cswval","0x%x %d %jd",
1642 ev.pl_u.pl_c.pl_pmcid,
1643 ev.pl_u.pl_c.pl_pid,
1644 ev.pl_u.pl_c.pl_value);
1646 case PMCLOG_TYPE_PROCEXEC:
1647 PMCSTAT_PRINT_ENTRY("exec","0x%x %d %p \"%s\"",
1648 ev.pl_u.pl_x.pl_pmcid,
1649 ev.pl_u.pl_x.pl_pid,
1650 (void *) ev.pl_u.pl_x.pl_entryaddr,
1651 ev.pl_u.pl_x.pl_pathname);
1653 case PMCLOG_TYPE_PROCEXIT:
1654 PMCSTAT_PRINT_ENTRY("exitval","0x%x %d %jd",
1655 ev.pl_u.pl_e.pl_pmcid,
1656 ev.pl_u.pl_e.pl_pid,
1657 ev.pl_u.pl_e.pl_value);
1659 case PMCLOG_TYPE_PROCFORK:
1660 PMCSTAT_PRINT_ENTRY("fork","%d %d",
1661 ev.pl_u.pl_f.pl_oldpid,
1662 ev.pl_u.pl_f.pl_newpid);
1664 case PMCLOG_TYPE_USERDATA:
1665 PMCSTAT_PRINT_ENTRY("userdata","0x%x",
1666 ev.pl_u.pl_u.pl_userdata);
1668 case PMCLOG_TYPE_SYSEXIT:
1669 PMCSTAT_PRINT_ENTRY("exit","%d",
1670 ev.pl_u.pl_se.pl_pid);
1673 fprintf(args.pa_printfile, "unknown event (type %d).\n",
1678 if (ev.pl_state == PMCLOG_EOF)
1679 return (PMCSTAT_FINISHED);
1680 else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1681 return (PMCSTAT_RUNNING);
1683 errx(EX_DATAERR, "ERROR: event parsing failed "
1684 "(record %jd, offset 0x%jx).",
1685 (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1690 * Public Interfaces.
1694 * Close a logfile, after first flushing all in-module queued data.
1698 pmcstat_close_log(void)
1700 /* If a local logfile is configured ask the kernel to stop
1701 * and flush data. Kernel will close the file when data is flushed
1702 * so keep the status to EXITING.
1704 if (args.pa_logfd != -1) {
1705 if (pmc_flush_logfile() < 0)
1706 err(EX_OSERR, "ERROR: logging failed");
1709 return (args.pa_flags & FLAG_HAS_PIPE ? PMCSTAT_EXITING :
1716 * Open a log file, for reading or writing.
1718 * The function returns the fd of a successfully opened log or -1 in
1723 pmcstat_open_log(const char *path, int mode)
1727 const char *p, *errstr;
1728 struct addrinfo hints, *res, *res0;
1729 char hostname[MAXHOSTNAMELEN];
1735 * If 'path' is "-" then open one of stdin or stdout depending
1736 * on the value of 'mode'.
1738 * If 'path' contains a ':' and does not start with a '/' or '.',
1739 * and is being opened for writing, treat it as a "host:port"
1740 * specification and open a network socket.
1742 * Otherwise, treat 'path' as a file name and open that.
1744 if (path[0] == '-' && path[1] == '\0')
1745 fd = (mode == PMCSTAT_OPEN_FOR_READ) ? 0 : 1;
1746 else if (path[0] != '/' &&
1747 path[0] != '.' && strchr(path, ':') != NULL) {
1749 p = strrchr(path, ':');
1751 if (p == path || hlen >= sizeof(hostname)) {
1752 errstr = strerror(EINVAL);
1756 assert(hlen < sizeof(hostname));
1757 (void) strncpy(hostname, path, hlen);
1758 hostname[hlen] = '\0';
1760 (void) memset(&hints, 0, sizeof(hints));
1761 hints.ai_family = AF_UNSPEC;
1762 hints.ai_socktype = SOCK_STREAM;
1763 if ((error = getaddrinfo(hostname, p+1, &hints, &res0)) != 0) {
1764 errstr = gai_strerror(error);
1769 for (res = res0; res; res = res->ai_next) {
1770 if ((fd = socket(res->ai_family, res->ai_socktype,
1771 res->ai_protocol)) < 0) {
1772 errstr = strerror(errno);
1775 if (mode == PMCSTAT_OPEN_FOR_READ) {
1776 if (bind(fd, res->ai_addr, res->ai_addrlen) < 0) {
1777 errstr = strerror(errno);
1783 cfd = accept(fd, NULL, NULL);
1786 errstr = strerror(errno);
1792 if (connect(fd, res->ai_addr, res->ai_addrlen) < 0) {
1793 errstr = strerror(errno);
1804 } else if ((fd = open(path, mode == PMCSTAT_OPEN_FOR_READ ?
1805 O_RDONLY : (O_WRONLY|O_CREAT|O_TRUNC),
1806 S_IRUSR|S_IWUSR|S_IRGRP|S_IROTH)) < 0)
1807 errstr = strerror(errno);
1811 errx(EX_OSERR, "ERROR: Cannot open \"%s\" for %s: %s.", path,
1812 (mode == PMCSTAT_OPEN_FOR_READ ? "reading" : "writing"),
1819 * Process a log file in offline analysis mode.
1823 pmcstat_process_log(void)
1827 * If analysis has not been asked for, just print the log to
1828 * the current output file.
1830 if (args.pa_flags & FLAG_DO_PRINT)
1831 return (pmcstat_print_log());
1833 return (pmcstat_analyze_log());
1837 * Refresh top display.
1841 pmcstat_refresh_top(void)
1846 struct pmcstat_pmcrecord *pmcpr;
1848 /* If in pause mode do not refresh display. */
1852 /* Wait until PMC pop in the log. */
1853 pmcpr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1857 /* Format PMC name. */
1858 if (pmcstat_mergepmc)
1859 snprintf(pmcname, sizeof(pmcname), "[%s]",
1860 pmcstat_string_unintern(pmcpr->pr_pmcname));
1862 snprintf(pmcname, sizeof(pmcname), "%s.%d",
1863 pmcstat_string_unintern(pmcpr->pr_pmcname),
1864 pmcstat_pmcinfilter);
1866 /* Format samples count. */
1867 if (ps_samples_period > 0)
1868 v = (pmcpr->pr_samples * 100.0) / ps_samples_period;
1871 v_attrs = PMCSTAT_ATTRPERCENT(v);
1873 PMCSTAT_PRINTBEGIN();
1874 PMCSTAT_PRINTW("PMC: %s Samples: %u ",
1877 PMCSTAT_ATTRON(v_attrs);
1878 PMCSTAT_PRINTW("(%.1f%%) ", v);
1879 PMCSTAT_ATTROFF(v_attrs);
1880 PMCSTAT_PRINTW(", %u unresolved\n\n",
1881 pmcpr->pr_dubious_frames);
1882 if (plugins[args.pa_plugin].pl_topdisplay != NULL)
1883 plugins[args.pa_plugin].pl_topdisplay();
1888 * Find the next pmc index to display.
1892 pmcstat_changefilter(void)
1895 struct pmcstat_pmcrecord *pmcr;
1898 * Find the next merge target.
1900 if (pmcstat_mergepmc) {
1901 pmcin = pmcstat_pmcinfilter;
1904 pmcr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1905 if (pmcr == NULL || pmcr == pmcr->pr_merge)
1908 pmcstat_pmcinfilter++;
1909 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
1910 pmcstat_pmcinfilter = 0;
1912 } while (pmcstat_pmcinfilter != pmcin);
1917 * Top mode keypress.
1921 pmcstat_keypress_log(void)
1926 w = newwin(1, 0, 1, 0);
1928 wprintw(w, "Key: %c => ", c);
1931 wprintw(w, "enter mode 'd' or 'a' => ");
1934 args.pa_topmode = PMCSTAT_TOP_DELTA;
1935 wprintw(w, "switching to delta mode");
1937 args.pa_topmode = PMCSTAT_TOP_ACCUM;
1938 wprintw(w, "switching to accumulation mode");
1942 pmcstat_mergepmc = !pmcstat_mergepmc;
1944 * Changing merge state require data reset.
1946 if (plugins[args.pa_plugin].pl_shutdown != NULL)
1947 plugins[args.pa_plugin].pl_shutdown(NULL);
1948 pmcstat_stats_reset(0);
1949 if (plugins[args.pa_plugin].pl_init != NULL)
1950 plugins[args.pa_plugin].pl_init();
1952 /* Update filter to be on a merge target. */
1953 pmcstat_changefilter();
1954 wprintw(w, "merge PMC %s", pmcstat_mergepmc ? "on" : "off");
1957 /* Close current plugin. */
1958 if (plugins[args.pa_plugin].pl_shutdown != NULL)
1959 plugins[args.pa_plugin].pl_shutdown(NULL);
1961 /* Find next top display available. */
1964 if (plugins[args.pa_plugin].pl_name == NULL)
1966 } while (plugins[args.pa_plugin].pl_topdisplay == NULL);
1968 /* Open new plugin. */
1969 pmcstat_stats_reset(0);
1970 if (plugins[args.pa_plugin].pl_init != NULL)
1971 plugins[args.pa_plugin].pl_init();
1972 wprintw(w, "switching to plugin %s",
1973 plugins[args.pa_plugin].pl_name);
1976 pmcstat_pmcinfilter++;
1977 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
1978 pmcstat_pmcinfilter = 0;
1979 pmcstat_changefilter();
1980 wprintw(w, "switching to PMC %s.%d",
1981 pmcstat_pmcindex_to_name(pmcstat_pmcinfilter),
1982 pmcstat_pmcinfilter);
1985 pmcstat_pause = !pmcstat_pause;
1987 wprintw(w, "pause => press space again to continue");
1990 wprintw(w, "exiting...");
1994 if (plugins[args.pa_plugin].pl_topkeypress != NULL)
1995 if (plugins[args.pa_plugin].pl_topkeypress(c, w))
2010 pmcstat_display_log(void)
2013 pmcstat_refresh_top();
2015 /* Reset everythings if delta mode. */
2016 if (args.pa_topmode == PMCSTAT_TOP_DELTA) {
2017 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2018 plugins[args.pa_plugin].pl_shutdown(NULL);
2019 pmcstat_stats_reset(0);
2020 if (plugins[args.pa_plugin].pl_init != NULL)
2021 plugins[args.pa_plugin].pl_init();
2027 * Configure a plugins.
2031 pmcstat_pluginconfigure_log(char *opt)
2034 if (strncmp(opt, "threshold=", 10) == 0) {
2035 pmcstat_threshold = atof(opt+10);
2037 if (plugins[args.pa_plugin].pl_configure != NULL) {
2038 if (!plugins[args.pa_plugin].pl_configure(opt))
2040 "ERROR: unknown option <%s>.", opt);
2046 * Initialize module.
2050 pmcstat_initialize_logging(void)
2054 /* use a convenient format for 'ldd' output */
2055 if (setenv("LD_TRACE_LOADED_OBJECTS_FMT1","%o \"%p\" %x\n",1) != 0)
2056 err(EX_OSERR, "ERROR: Cannot setenv");
2058 /* Initialize hash tables */
2059 pmcstat_string_initialize();
2060 for (i = 0; i < PMCSTAT_NHASH; i++) {
2061 LIST_INIT(&pmcstat_image_hash[i]);
2062 LIST_INIT(&pmcstat_process_hash[i]);
2066 * Create a fake 'process' entry for the kernel with pid -1.
2067 * hwpmc(4) will subsequently inform us about where the kernel
2068 * and any loaded kernel modules are mapped.
2070 if ((pmcstat_kernproc = pmcstat_process_lookup((pid_t) -1,
2071 PMCSTAT_ALLOCATE)) == NULL)
2072 err(EX_OSERR, "ERROR: Cannot initialize logging");
2077 /* Merge PMC with same name. */
2078 pmcstat_mergepmc = args.pa_mergepmc;
2081 * Initialize plugins
2084 if (plugins[args.pa_pplugin].pl_init != NULL)
2085 plugins[args.pa_pplugin].pl_init();
2086 if (plugins[args.pa_plugin].pl_init != NULL)
2087 plugins[args.pa_plugin].pl_init();
2095 pmcstat_shutdown_logging(void)
2099 struct pmcstat_image *pi, *pitmp;
2100 struct pmcstat_process *pp, *pptmp;
2101 struct pmcstat_pcmap *ppm, *ppmtmp;
2103 /* determine where to send the map file */
2105 if (args.pa_mapfilename != NULL)
2106 mf = (strcmp(args.pa_mapfilename, "-") == 0) ?
2107 args.pa_printfile : fopen(args.pa_mapfilename, "w");
2109 if (mf == NULL && args.pa_flags & FLAG_DO_GPROF &&
2110 args.pa_verbosity >= 2)
2111 mf = args.pa_printfile;
2114 (void) fprintf(mf, "MAP:\n");
2117 * Shutdown the plugins
2120 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2121 plugins[args.pa_plugin].pl_shutdown(mf);
2122 if (plugins[args.pa_pplugin].pl_shutdown != NULL)
2123 plugins[args.pa_pplugin].pl_shutdown(mf);
2125 for (i = 0; i < PMCSTAT_NHASH; i++) {
2126 LIST_FOREACH_SAFE(pi, &pmcstat_image_hash[i], pi_next,
2128 if (plugins[args.pa_plugin].pl_shutdownimage != NULL)
2129 plugins[args.pa_plugin].pl_shutdownimage(pi);
2130 if (plugins[args.pa_pplugin].pl_shutdownimage != NULL)
2131 plugins[args.pa_pplugin].pl_shutdownimage(pi);
2133 free(pi->pi_symbols);
2134 if (pi->pi_addr2line != NULL)
2135 pclose(pi->pi_addr2line);
2136 LIST_REMOVE(pi, pi_next);
2140 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[i], pp_next,
2142 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
2143 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
2146 LIST_REMOVE(pp, pp_next);
2151 pmcstat_string_shutdown();
2154 * Print errors unless -q was specified. Print all statistics
2157 #define PRINT(N,V) do { \
2158 if (pmcstat_stats.ps_##V || args.pa_verbosity >= 2) \
2159 (void) fprintf(args.pa_printfile, " %-40s %d\n",\
2160 N, pmcstat_stats.ps_##V); \
2163 if (args.pa_verbosity >= 1 && (args.pa_flags & FLAG_DO_ANALYSIS)) {
2164 (void) fprintf(args.pa_printfile, "CONVERSION STATISTICS:\n");
2165 PRINT("#exec/a.out", exec_aout);
2166 PRINT("#exec/elf", exec_elf);
2167 PRINT("#exec/unknown", exec_indeterminable);
2168 PRINT("#exec handling errors", exec_errors);
2169 PRINT("#samples/total", samples_total);
2170 PRINT("#samples/unclaimed", samples_unknown_offset);
2171 PRINT("#samples/unknown-object", samples_indeterminable);
2172 PRINT("#samples/unknown-function", samples_unknown_function);
2173 PRINT("#callchain/dubious-frames", callchain_dubious_frames);