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_calltree.h"
154 static struct pmc_plugins {
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 static 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 comparison 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.
278 static int pmcstat_pause;
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)
425 "ERROR: a.out kernel modules are 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 if (args.pa_verbosity >= 2)
433 warn("WARNING: Cannot determine type of \"%s\"",
435 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
443 if ((unsigned) nbytes != sizeof(ex) ||
447 image->pi_type = PMCSTAT_IMAGE_AOUT;
449 /* TODO: the rest of a.out processing */
459 pmcstat_symbol_compare(const void *a, const void *b)
461 const struct pmcstat_symbol *sym1, *sym2;
463 sym1 = (const struct pmcstat_symbol *) a;
464 sym2 = (const struct pmcstat_symbol *) b;
466 if (sym1->ps_end <= sym2->ps_start)
468 if (sym1->ps_start >= sym2->ps_end)
474 * Map an address to a symbol in an image.
477 struct pmcstat_symbol *
478 pmcstat_symbol_search(struct pmcstat_image *image, uintfptr_t addr)
480 struct pmcstat_symbol sym;
482 if (image->pi_symbols == NULL)
487 sym.ps_end = addr + 1;
489 return (bsearch((void *) &sym, image->pi_symbols,
490 image->pi_symcount, sizeof(struct pmcstat_symbol),
491 pmcstat_symbol_compare));
495 * Add the list of symbols in the given section to the list associated
499 pmcstat_image_add_symbols(struct pmcstat_image *image, Elf *e,
500 Elf_Scn *scn, GElf_Shdr *sh)
503 size_t n, newsyms, nshsyms, nfuncsyms;
504 struct pmcstat_symbol *symptr;
509 if ((data = elf_getdata(scn, NULL)) == NULL)
513 * Determine the number of functions named in this
517 nshsyms = sh->sh_size / sh->sh_entsize;
518 for (n = nfuncsyms = 0; n < nshsyms; n++) {
519 if (gelf_getsym(data, (int) n, &sym) != &sym)
521 if (GELF_ST_TYPE(sym.st_info) == STT_FUNC)
529 * Allocate space for the new entries.
531 firsttime = image->pi_symbols == NULL;
532 symptr = realloc(image->pi_symbols,
533 sizeof(*symptr) * (image->pi_symcount + nfuncsyms));
534 if (symptr == image->pi_symbols) /* realloc() failed. */
536 image->pi_symbols = symptr;
539 * Append new symbols to the end of the current table.
541 symptr += image->pi_symcount;
543 for (n = newsyms = 0; n < nshsyms; n++) {
544 if (gelf_getsym(data, (int) n, &sym) != &sym)
546 if (GELF_ST_TYPE(sym.st_info) != STT_FUNC)
548 if (sym.st_shndx == STN_UNDEF)
551 if (!firsttime && pmcstat_symbol_search(image, sym.st_value))
552 continue; /* We've seen this symbol already. */
554 if ((fnname = elf_strptr(e, sh->sh_link, sym.st_name))
558 /* Remove spurious ARM function name. */
559 if (fnname[0] == '$' &&
560 (fnname[1] == 'a' || fnname[1] == 't' ||
566 symptr->ps_name = pmcstat_string_intern(fnname);
567 symptr->ps_start = sym.st_value - image->pi_vaddr;
568 symptr->ps_end = symptr->ps_start + sym.st_size;
574 image->pi_symcount += newsyms;
575 if (image->pi_symcount == 0)
578 assert(newsyms <= nfuncsyms);
581 * Return space to the system if there were duplicates.
583 if (newsyms < nfuncsyms)
584 image->pi_symbols = realloc(image->pi_symbols,
585 sizeof(*symptr) * image->pi_symcount);
588 * Keep the list of symbols sorted.
590 qsort(image->pi_symbols, image->pi_symcount, sizeof(*symptr),
591 pmcstat_symbol_compare);
594 * Deal with function symbols that have a size of 'zero' by
595 * making them extend to the next higher address. These
596 * symbols are usually defined in assembly code.
598 for (symptr = image->pi_symbols;
599 symptr < image->pi_symbols + (image->pi_symcount - 1);
601 if (symptr->ps_start == symptr->ps_end)
602 symptr->ps_end = (symptr+1)->ps_start;
606 * Examine an ELF file to determine the size of its text segment.
607 * Sets image->pi_type if anything conclusive can be determined about
612 pmcstat_image_get_elf_params(struct pmcstat_image *image)
616 const char *path, *elfbase;
618 uintfptr_t minva, maxva;
624 enum pmcstat_image_type image_type;
625 char buffer[PATH_MAX];
627 assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
629 image->pi_start = minva = ~(uintfptr_t) 0;
630 image->pi_end = maxva = (uintfptr_t) 0;
631 image->pi_type = image_type = PMCSTAT_IMAGE_INDETERMINABLE;
632 image->pi_isdynamic = 0;
633 image->pi_dynlinkerpath = NULL;
636 path = pmcstat_string_unintern(image->pi_execpath);
637 assert(path != NULL);
640 * Look for kernel modules under FSROOT/KERNELPATH/NAME,
641 * and user mode executable objects under FSROOT/PATHNAME.
643 if (image->pi_iskernelmodule)
644 (void) snprintf(buffer, sizeof(buffer), "%s%s/%s",
645 args.pa_fsroot, args.pa_kernel, path);
647 (void) snprintf(buffer, sizeof(buffer), "%s%s",
648 args.pa_fsroot, path);
651 if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
652 (e = elf_begin(fd, ELF_C_READ, NULL)) == NULL ||
653 (elf_kind(e) != ELF_K_ELF)) {
654 if (args.pa_verbosity >= 2)
655 warnx("WARNING: Cannot determine the type of \"%s\".",
660 if (gelf_getehdr(e, &eh) != &eh) {
662 "WARNING: Cannot retrieve the ELF Header for \"%s\": %s.",
663 buffer, elf_errmsg(-1));
667 if (eh.e_type != ET_EXEC && eh.e_type != ET_DYN &&
668 !(image->pi_iskernelmodule && eh.e_type == ET_REL)) {
669 warnx("WARNING: \"%s\" is of an unsupported ELF type.",
674 image_type = eh.e_ident[EI_CLASS] == ELFCLASS32 ?
675 PMCSTAT_IMAGE_ELF32 : PMCSTAT_IMAGE_ELF64;
678 * Determine the virtual address where an executable would be
679 * loaded. Additionally, for dynamically linked executables,
680 * save the pathname to the runtime linker.
682 if (eh.e_type == ET_EXEC) {
683 if (elf_getphnum(e, &nph) == 0) {
685 "WARNING: Could not determine the number of program headers in \"%s\": %s.",
690 for (i = 0; i < eh.e_phnum; i++) {
691 if (gelf_getphdr(e, i, &ph) != &ph) {
693 "WARNING: Retrieval of PHDR entry #%ju in \"%s\" failed: %s.",
694 (uintmax_t) i, buffer, elf_errmsg(-1));
699 image->pi_isdynamic = 1;
702 if ((elfbase = elf_rawfile(e, NULL)) == NULL) {
704 "WARNING: Cannot retrieve the interpreter for \"%s\": %s.",
705 buffer, elf_errmsg(-1));
708 image->pi_dynlinkerpath =
709 pmcstat_string_intern(elfbase +
713 if ((ph.p_offset & (-ph.p_align)) == 0)
714 image->pi_vaddr = ph.p_vaddr & (-ph.p_align);
721 * Get the min and max VA associated with this ELF object.
723 if (elf_getshnum(e, &nsh) == 0) {
725 "WARNING: Could not determine the number of sections for \"%s\": %s.",
726 buffer, elf_errmsg(-1));
730 for (i = 0; i < nsh; i++) {
731 if ((scn = elf_getscn(e, i)) == NULL ||
732 gelf_getshdr(scn, &sh) != &sh) {
734 "WARNING: Could not retrieve section header #%ju in \"%s\": %s.",
735 (uintmax_t) i, buffer, elf_errmsg(-1));
738 if (sh.sh_flags & SHF_EXECINSTR) {
739 minva = min(minva, sh.sh_addr);
740 maxva = max(maxva, sh.sh_addr + sh.sh_size);
742 if (sh.sh_type == SHT_SYMTAB || sh.sh_type == SHT_DYNSYM)
743 pmcstat_image_add_symbols(image, e, scn, &sh);
746 image->pi_start = minva;
747 image->pi_end = maxva;
748 image->pi_type = image_type;
749 image->pi_fullpath = pmcstat_string_intern(buffer);
751 /* Build display name
754 for (p = buffer; *p; p++)
757 image->pi_name = pmcstat_string_intern(endp);
767 * Given an image descriptor, determine whether it is an ELF, or AOUT.
768 * If no handler claims the image, set its type to 'INDETERMINABLE'.
772 pmcstat_image_determine_type(struct pmcstat_image *image)
774 assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
776 /* Try each kind of handler in turn */
777 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
778 pmcstat_image_get_elf_params(image);
779 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
780 pmcstat_image_get_aout_params(image);
783 * Otherwise, remember that we tried to determine
784 * the object's type and had failed.
786 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
787 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
791 * Locate an image descriptor given an interned path, adding a fresh
792 * descriptor to the cache if necessary. This function also finds a
793 * suitable name for this image's sample file.
795 * We defer filling in the file format specific parts of the image
796 * structure till the time we actually see a sample that would fall
800 static struct pmcstat_image *
801 pmcstat_image_from_path(pmcstat_interned_string internedpath,
805 struct pmcstat_image *pi;
807 hash = pmcstat_string_lookup_hash(internedpath);
809 /* First, look for an existing entry. */
810 LIST_FOREACH(pi, &pmcstat_image_hash[hash], pi_next)
811 if (pi->pi_execpath == internedpath &&
812 pi->pi_iskernelmodule == iskernelmodule)
816 * Allocate a new entry and place it at the head of the hash
819 pi = malloc(sizeof(*pi));
823 pi->pi_type = PMCSTAT_IMAGE_UNKNOWN;
824 pi->pi_execpath = internedpath;
829 pi->pi_isdynamic = 0;
830 pi->pi_iskernelmodule = iskernelmodule;
831 pi->pi_dynlinkerpath = NULL;
832 pi->pi_symbols = NULL;
834 pi->pi_addr2line = NULL;
836 if (plugins[args.pa_pplugin].pl_initimage != NULL)
837 plugins[args.pa_pplugin].pl_initimage(pi);
838 if (plugins[args.pa_plugin].pl_initimage != NULL)
839 plugins[args.pa_plugin].pl_initimage(pi);
841 LIST_INSERT_HEAD(&pmcstat_image_hash[hash], pi, pi_next);
847 * Record the fact that PC values from 'start' to 'end' come from
852 pmcstat_image_link(struct pmcstat_process *pp, struct pmcstat_image *image,
855 struct pmcstat_pcmap *pcm, *pcmnew;
858 assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN &&
859 image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE);
861 if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
862 err(EX_OSERR, "ERROR: Cannot create a map entry");
865 * Adjust the map entry to only cover the text portion
869 offset = start - image->pi_vaddr;
870 pcmnew->ppm_lowpc = image->pi_start + offset;
871 pcmnew->ppm_highpc = image->pi_end + offset;
872 pcmnew->ppm_image = image;
874 assert(pcmnew->ppm_lowpc < pcmnew->ppm_highpc);
876 /* Overlapped mmap()'s are assumed to never occur. */
877 TAILQ_FOREACH(pcm, &pp->pp_map, ppm_next)
878 if (pcm->ppm_lowpc >= pcmnew->ppm_highpc)
882 TAILQ_INSERT_TAIL(&pp->pp_map, pcmnew, ppm_next);
884 TAILQ_INSERT_BEFORE(pcm, pcmnew, ppm_next);
888 * Unmap images in the range [start..end) associated with process
893 pmcstat_image_unmap(struct pmcstat_process *pp, uintfptr_t start,
896 struct pmcstat_pcmap *pcm, *pcmtmp, *pcmnew;
903 * - we could have the range completely in the middle of an
904 * existing pcmap; in this case we have to split the pcmap
905 * structure into two (i.e., generate a 'hole').
906 * - we could have the range covering multiple pcmaps; these
907 * will have to be removed.
908 * - we could have either 'start' or 'end' falling in the
909 * middle of a pcmap; in this case shorten the entry.
911 TAILQ_FOREACH_SAFE(pcm, &pp->pp_map, ppm_next, pcmtmp) {
912 assert(pcm->ppm_lowpc < pcm->ppm_highpc);
913 if (pcm->ppm_highpc <= start)
915 if (pcm->ppm_lowpc >= end)
917 if (pcm->ppm_lowpc >= start && pcm->ppm_highpc <= end) {
919 * The current pcmap is completely inside the
920 * unmapped range: remove it entirely.
922 TAILQ_REMOVE(&pp->pp_map, pcm, ppm_next);
924 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc > end) {
926 * Split this pcmap into two; curtail the
927 * current map to end at [start-1], and start
928 * the new one at [end].
930 if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
932 "ERROR: Cannot split a map entry");
934 pcmnew->ppm_image = pcm->ppm_image;
936 pcmnew->ppm_lowpc = end;
937 pcmnew->ppm_highpc = pcm->ppm_highpc;
939 pcm->ppm_highpc = start;
941 TAILQ_INSERT_AFTER(&pp->pp_map, pcm, pcmnew, ppm_next);
944 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc <= end)
945 pcm->ppm_highpc = start;
946 else if (pcm->ppm_lowpc >= start && pcm->ppm_highpc > end)
947 pcm->ppm_lowpc = end;
954 * Resolve file name and line number for the given address.
957 pmcstat_image_addr2line(struct pmcstat_image *image, uintfptr_t addr,
958 char *sourcefile, size_t sourcefile_len, unsigned *sourceline,
959 char *funcname, size_t funcname_len)
961 static int addr2line_warn = 0;
964 char *sep, cmdline[PATH_MAX], imagepath[PATH_MAX];
967 if (image->pi_addr2line == NULL) {
968 snprintf(imagepath, sizeof(imagepath), "%s%s.symbols",
970 pmcstat_string_unintern(image->pi_fullpath));
971 fd = open(imagepath, O_RDONLY);
973 snprintf(imagepath, sizeof(imagepath), "%s%s",
975 pmcstat_string_unintern(image->pi_fullpath));
979 * New addr2line support recursive inline function with -i
980 * but the format does not add a marker when no more entries
983 snprintf(cmdline, sizeof(cmdline), "addr2line -Cfe \"%s\"",
985 image->pi_addr2line = popen(cmdline, "r+");
986 if (image->pi_addr2line == NULL) {
987 if (!addr2line_warn) {
990 "WARNING: addr2line is needed for source code information."
997 if (feof(image->pi_addr2line) || ferror(image->pi_addr2line)) {
998 warnx("WARNING: addr2line pipe error");
999 pclose(image->pi_addr2line);
1000 image->pi_addr2line = NULL;
1004 fprintf(image->pi_addr2line, "%p\n", (void *)addr);
1006 if (fgets(funcname, funcname_len, image->pi_addr2line) == NULL) {
1007 warnx("WARNING: addr2line function name read error");
1010 sep = strchr(funcname, '\n');
1014 if (fgets(sourcefile, sourcefile_len, image->pi_addr2line) == NULL) {
1015 warnx("WARNING: addr2line source file read error");
1018 sep = strchr(sourcefile, ':');
1020 warnx("WARNING: addr2line source line separator missing");
1032 * Add a {pmcid,name} mapping.
1036 pmcstat_pmcid_add(pmc_id_t pmcid, pmcstat_interned_string ps)
1038 struct pmcstat_pmcrecord *pr, *prm;
1040 /* Replace an existing name for the PMC. */
1042 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1043 if (pr->pr_pmcid == pmcid) {
1044 pr->pr_pmcname = ps;
1046 } else if (pr->pr_pmcname == ps)
1050 * Otherwise, allocate a new descriptor and call the
1053 if ((pr = malloc(sizeof(*pr))) == NULL)
1054 err(EX_OSERR, "ERROR: Cannot allocate pmc record");
1056 pr->pr_pmcid = pmcid;
1057 pr->pr_pmcname = ps;
1058 pr->pr_pmcin = pmcstat_npmcs++;
1060 pr->pr_dubious_frames = 0;
1061 pr->pr_merge = prm == NULL ? pr : prm;
1063 LIST_INSERT_HEAD(&pmcstat_pmcs, pr, pr_next);
1065 if (plugins[args.pa_pplugin].pl_newpmc != NULL)
1066 plugins[args.pa_pplugin].pl_newpmc(ps, pr);
1067 if (plugins[args.pa_plugin].pl_newpmc != NULL)
1068 plugins[args.pa_plugin].pl_newpmc(ps, pr);
1072 * Given a pmcid in use, find its human-readable name.
1076 pmcstat_pmcid_to_name(pmc_id_t pmcid)
1078 struct pmcstat_pmcrecord *pr;
1080 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1081 if (pr->pr_pmcid == pmcid)
1082 return (pmcstat_string_unintern(pr->pr_pmcname));
1088 * Convert PMC index to name.
1092 pmcstat_pmcindex_to_name(int pmcin)
1094 struct pmcstat_pmcrecord *pr;
1096 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1097 if (pr->pr_pmcin == pmcin)
1098 return pmcstat_string_unintern(pr->pr_pmcname);
1104 * Return PMC record with given index.
1107 struct pmcstat_pmcrecord *
1108 pmcstat_pmcindex_to_pmcr(int pmcin)
1110 struct pmcstat_pmcrecord *pr;
1112 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1113 if (pr->pr_pmcin == pmcin)
1120 * Get PMC record by id, apply merge policy.
1123 static struct pmcstat_pmcrecord *
1124 pmcstat_lookup_pmcid(pmc_id_t pmcid)
1126 struct pmcstat_pmcrecord *pr;
1128 LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
1129 if (pr->pr_pmcid == pmcid) {
1130 if (pmcstat_mergepmc)
1131 return pr->pr_merge;
1140 * Associate an AOUT image with a process.
1144 pmcstat_process_aout_exec(struct pmcstat_process *pp,
1145 struct pmcstat_image *image, uintfptr_t entryaddr)
1150 /* TODO Implement a.out handling */
1154 * Associate an ELF image with a process.
1158 pmcstat_process_elf_exec(struct pmcstat_process *pp,
1159 struct pmcstat_image *image, uintfptr_t entryaddr)
1162 struct pmcstat_image *rtldimage;
1164 assert(image->pi_type == PMCSTAT_IMAGE_ELF32 ||
1165 image->pi_type == PMCSTAT_IMAGE_ELF64);
1167 /* Create a map entry for the base executable. */
1168 pmcstat_image_link(pp, image, image->pi_vaddr);
1171 * For dynamically linked executables we need to determine
1172 * where the dynamic linker was mapped to for this process,
1173 * Subsequent executable objects that are mapped in by the
1174 * dynamic linker will be tracked by log events of type
1175 * PMCLOG_TYPE_MAP_IN.
1178 if (image->pi_isdynamic) {
1181 * The runtime loader gets loaded just after the maximum
1182 * possible heap address. Like so:
1184 * [ TEXT DATA BSS HEAP -->*RTLD SHLIBS <--STACK]
1186 * 0 VM_MAXUSER_ADDRESS
1189 * The exact address where the loader gets mapped in
1190 * will vary according to the size of the executable
1191 * and the limits on the size of the process'es data
1192 * segment at the time of exec(). The entry address
1193 * recorded at process exec time corresponds to the
1194 * 'start' address inside the dynamic linker. From
1195 * this we can figure out the address where the
1196 * runtime loader's file object had been mapped to.
1198 rtldimage = pmcstat_image_from_path(image->pi_dynlinkerpath, 0);
1199 if (rtldimage == NULL) {
1200 warnx("WARNING: Cannot find image for \"%s\".",
1201 pmcstat_string_unintern(image->pi_dynlinkerpath));
1202 pmcstat_stats.ps_exec_errors++;
1206 if (rtldimage->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1207 pmcstat_image_get_elf_params(rtldimage);
1209 if (rtldimage->pi_type != PMCSTAT_IMAGE_ELF32 &&
1210 rtldimage->pi_type != PMCSTAT_IMAGE_ELF64) {
1211 warnx("WARNING: rtld not an ELF object \"%s\".",
1212 pmcstat_string_unintern(image->pi_dynlinkerpath));
1216 libstart = entryaddr - rtldimage->pi_entry;
1217 pmcstat_image_link(pp, rtldimage, libstart);
1222 * Find the process descriptor corresponding to a PID. If 'allocate'
1223 * is zero, we return a NULL if a pid descriptor could not be found or
1224 * a process descriptor process. If 'allocate' is non-zero, then we
1225 * will attempt to allocate a fresh process descriptor. Zombie
1226 * process descriptors are only removed if a fresh allocation for the
1227 * same PID is requested.
1230 static struct pmcstat_process *
1231 pmcstat_process_lookup(pid_t pid, int allocate)
1234 struct pmcstat_pcmap *ppm, *ppmtmp;
1235 struct pmcstat_process *pp, *pptmp;
1237 hash = (uint32_t) pid & PMCSTAT_HASH_MASK; /* simplicity wins */
1239 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[hash], pp_next, pptmp)
1240 if (pp->pp_pid == pid) {
1241 /* Found a descriptor, check and process zombies */
1242 if (allocate && pp->pp_isactive == 0) {
1244 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next,
1246 TAILQ_REMOVE(&pp->pp_map, ppm,
1250 /* remove process entry */
1251 LIST_REMOVE(pp, pp_next);
1261 if ((pp = malloc(sizeof(*pp))) == NULL)
1262 err(EX_OSERR, "ERROR: Cannot allocate pid descriptor");
1265 pp->pp_isactive = 1;
1267 TAILQ_INIT(&pp->pp_map);
1269 LIST_INSERT_HEAD(&pmcstat_process_hash[hash], pp, pp_next);
1274 * Associate an image and a process.
1278 pmcstat_process_exec(struct pmcstat_process *pp,
1279 pmcstat_interned_string path, uintfptr_t entryaddr)
1281 struct pmcstat_image *image;
1283 if ((image = pmcstat_image_from_path(path, 0)) == NULL) {
1284 pmcstat_stats.ps_exec_errors++;
1288 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1289 pmcstat_image_determine_type(image);
1291 assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN);
1293 switch (image->pi_type) {
1294 case PMCSTAT_IMAGE_ELF32:
1295 case PMCSTAT_IMAGE_ELF64:
1296 pmcstat_stats.ps_exec_elf++;
1297 pmcstat_process_elf_exec(pp, image, entryaddr);
1300 case PMCSTAT_IMAGE_AOUT:
1301 pmcstat_stats.ps_exec_aout++;
1302 pmcstat_process_aout_exec(pp, image, entryaddr);
1305 case PMCSTAT_IMAGE_INDETERMINABLE:
1306 pmcstat_stats.ps_exec_indeterminable++;
1311 "ERROR: Unsupported executable type for \"%s\"",
1312 pmcstat_string_unintern(path));
1318 * Find the map entry associated with process 'p' at PC value 'pc'.
1321 struct pmcstat_pcmap *
1322 pmcstat_process_find_map(struct pmcstat_process *p, uintfptr_t pc)
1324 struct pmcstat_pcmap *ppm;
1326 TAILQ_FOREACH(ppm, &p->pp_map, ppm_next) {
1327 if (pc >= ppm->ppm_lowpc && pc < ppm->ppm_highpc)
1329 if (pc < ppm->ppm_lowpc)
1337 * Convert a hwpmc(4) log to profile information. A system-wide
1338 * callgraph is generated if FLAG_DO_CALLGRAPHS is set. gmon.out
1339 * files usable by gprof(1) are created if FLAG_DO_GPROF is set.
1342 pmcstat_analyze_log(void)
1344 uint32_t cpu, cpuflags;
1347 struct pmcstat_image *image;
1348 struct pmcstat_process *pp, *ppnew;
1349 struct pmcstat_pcmap *ppm, *ppmtmp;
1350 struct pmclog_ev ev;
1351 struct pmcstat_pmcrecord *pmcr;
1352 pmcstat_interned_string image_path;
1354 assert(args.pa_flags & FLAG_DO_ANALYSIS);
1356 if (elf_version(EV_CURRENT) == EV_NONE)
1357 err(EX_UNAVAILABLE, "Elf library intialization failed");
1359 while (pmclog_read(args.pa_logparser, &ev) == 0) {
1360 assert(ev.pl_state == PMCLOG_OK);
1362 switch (ev.pl_type) {
1363 case PMCLOG_TYPE_INITIALIZE:
1364 if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1365 PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1367 "WARNING: Log version 0x%x does not match compiled version 0x%x.",
1368 ev.pl_u.pl_i.pl_version, PMC_VERSION_MAJOR);
1371 case PMCLOG_TYPE_MAP_IN:
1373 * Introduce an address range mapping for a
1374 * userland process or the kernel (pid == -1).
1376 * We always allocate a process descriptor so
1377 * that subsequent samples seen for this
1378 * address range are mapped to the current
1379 * object being mapped in.
1381 pid = ev.pl_u.pl_mi.pl_pid;
1383 pp = pmcstat_kernproc;
1385 pp = pmcstat_process_lookup(pid,
1390 image_path = pmcstat_string_intern(ev.pl_u.pl_mi.
1392 image = pmcstat_image_from_path(image_path, pid == -1);
1393 if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1394 pmcstat_image_determine_type(image);
1395 if (image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE)
1396 pmcstat_image_link(pp, image,
1397 ev.pl_u.pl_mi.pl_start);
1400 case PMCLOG_TYPE_MAP_OUT:
1402 * Remove an address map.
1404 pid = ev.pl_u.pl_mo.pl_pid;
1406 pp = pmcstat_kernproc;
1408 pp = pmcstat_process_lookup(pid, 0);
1410 if (pp == NULL) /* unknown process */
1413 pmcstat_image_unmap(pp, ev.pl_u.pl_mo.pl_start,
1414 ev.pl_u.pl_mo.pl_end);
1417 case PMCLOG_TYPE_PCSAMPLE:
1419 * Note: the `PCSAMPLE' log entry is not
1420 * generated by hpwmc(4) after version 2.
1424 * We bring in the gmon file for the image
1425 * currently associated with the PMC & pid
1426 * pair and increment the appropriate entry
1429 pmcstat_stats.ps_samples_total++;
1430 ps_samples_period++;
1432 pc = ev.pl_u.pl_s.pl_pc;
1433 pp = pmcstat_process_lookup(ev.pl_u.pl_s.pl_pid,
1436 /* Get PMC record. */
1437 pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_s.pl_pmcid);
1438 assert(pmcr != NULL);
1442 * Call the plugins processing
1443 * TODO: move pmcstat_process_find_map inside plugins
1446 if (plugins[args.pa_pplugin].pl_process != NULL)
1447 plugins[args.pa_pplugin].pl_process(
1449 pmcstat_process_find_map(pp, pc) != NULL, 0);
1450 plugins[args.pa_plugin].pl_process(
1452 pmcstat_process_find_map(pp, pc) != NULL, 0);
1455 case PMCLOG_TYPE_CALLCHAIN:
1456 pmcstat_stats.ps_samples_total++;
1457 ps_samples_period++;
1459 cpuflags = ev.pl_u.pl_cc.pl_cpuflags;
1460 cpu = PMC_CALLCHAIN_CPUFLAGS_TO_CPU(cpuflags);
1462 /* Filter on the CPU id. */
1463 if (!CPU_ISSET(cpu, &(args.pa_cpumask))) {
1464 pmcstat_stats.ps_samples_skipped++;
1468 pp = pmcstat_process_lookup(ev.pl_u.pl_cc.pl_pid,
1471 /* Get PMC record. */
1472 pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_cc.pl_pmcid);
1473 assert(pmcr != NULL);
1477 * Call the plugins processing
1480 if (plugins[args.pa_pplugin].pl_process != NULL)
1481 plugins[args.pa_pplugin].pl_process(
1483 ev.pl_u.pl_cc.pl_npc,
1484 ev.pl_u.pl_cc.pl_pc,
1485 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1487 plugins[args.pa_plugin].pl_process(
1489 ev.pl_u.pl_cc.pl_npc,
1490 ev.pl_u.pl_cc.pl_pc,
1491 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1495 case PMCLOG_TYPE_PMCALLOCATE:
1497 * Record the association pmc id between this
1500 pmcstat_pmcid_add(ev.pl_u.pl_a.pl_pmcid,
1501 pmcstat_string_intern(ev.pl_u.pl_a.pl_evname));
1504 case PMCLOG_TYPE_PMCALLOCATEDYN:
1506 * Record the association pmc id between this
1509 pmcstat_pmcid_add(ev.pl_u.pl_ad.pl_pmcid,
1510 pmcstat_string_intern(ev.pl_u.pl_ad.pl_evname));
1513 case PMCLOG_TYPE_PROCEXEC:
1516 * Change the executable image associated with
1519 pp = pmcstat_process_lookup(ev.pl_u.pl_x.pl_pid,
1522 /* delete the current process map */
1523 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
1524 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
1528 /* associate this process image */
1529 image_path = pmcstat_string_intern(
1530 ev.pl_u.pl_x.pl_pathname);
1531 assert(image_path != NULL);
1532 pmcstat_process_exec(pp, image_path,
1533 ev.pl_u.pl_x.pl_entryaddr);
1536 case PMCLOG_TYPE_PROCEXIT:
1539 * Due to the way the log is generated, the
1540 * last few samples corresponding to a process
1541 * may appear in the log after the process
1542 * exit event is recorded. Thus we keep the
1543 * process' descriptor and associated data
1544 * structures around, but mark the process as
1547 pp = pmcstat_process_lookup(ev.pl_u.pl_e.pl_pid, 0);
1550 pp->pp_isactive = 0; /* mark as a zombie */
1553 case PMCLOG_TYPE_SYSEXIT:
1554 pp = pmcstat_process_lookup(ev.pl_u.pl_se.pl_pid, 0);
1557 pp->pp_isactive = 0; /* make a zombie */
1560 case PMCLOG_TYPE_PROCFORK:
1563 * Allocate a process descriptor for the new
1567 pmcstat_process_lookup(ev.pl_u.pl_f.pl_newpid,
1571 * If we had been tracking the parent, clone
1574 pp = pmcstat_process_lookup(ev.pl_u.pl_f.pl_oldpid, 0);
1577 TAILQ_FOREACH(ppm, &pp->pp_map, ppm_next)
1578 pmcstat_image_link(ppnew, ppm->ppm_image,
1582 default: /* other types of entries are not relevant */
1587 if (ev.pl_state == PMCLOG_EOF)
1588 return (PMCSTAT_FINISHED);
1589 else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1590 return (PMCSTAT_RUNNING);
1593 "ERROR: event parsing failed (record %jd, offset 0x%jx)",
1594 (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1598 * Print log entries as text.
1602 pmcstat_print_log(void)
1604 struct pmclog_ev ev;
1607 while (pmclog_read(args.pa_logparser, &ev) == 0) {
1608 assert(ev.pl_state == PMCLOG_OK);
1609 switch (ev.pl_type) {
1610 case PMCLOG_TYPE_CALLCHAIN:
1611 PMCSTAT_PRINT_ENTRY("callchain",
1612 "%d 0x%x %d %d %c", ev.pl_u.pl_cc.pl_pid,
1613 ev.pl_u.pl_cc.pl_pmcid,
1614 PMC_CALLCHAIN_CPUFLAGS_TO_CPU(ev.pl_u.pl_cc. \
1615 pl_cpuflags), ev.pl_u.pl_cc.pl_npc,
1616 PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(ev.pl_u.pl_cc.\
1617 pl_cpuflags) ? 'u' : 's');
1618 for (npc = 0; npc < ev.pl_u.pl_cc.pl_npc; npc++)
1619 PMCSTAT_PRINT_ENTRY("...", "%p",
1620 (void *) ev.pl_u.pl_cc.pl_pc[npc]);
1622 case PMCLOG_TYPE_CLOSELOG:
1623 PMCSTAT_PRINT_ENTRY("closelog",);
1625 case PMCLOG_TYPE_DROPNOTIFY:
1626 PMCSTAT_PRINT_ENTRY("drop",);
1628 case PMCLOG_TYPE_INITIALIZE:
1629 PMCSTAT_PRINT_ENTRY("initlog","0x%x \"%s\"",
1630 ev.pl_u.pl_i.pl_version,
1631 pmc_name_of_cputype(ev.pl_u.pl_i.pl_arch));
1632 if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1633 PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1635 "WARNING: Log version 0x%x != expected version 0x%x.",
1636 ev.pl_u.pl_i.pl_version, PMC_VERSION);
1638 case PMCLOG_TYPE_MAP_IN:
1639 PMCSTAT_PRINT_ENTRY("map-in","%d %p \"%s\"",
1640 ev.pl_u.pl_mi.pl_pid,
1641 (void *) ev.pl_u.pl_mi.pl_start,
1642 ev.pl_u.pl_mi.pl_pathname);
1644 case PMCLOG_TYPE_MAP_OUT:
1645 PMCSTAT_PRINT_ENTRY("map-out","%d %p %p",
1646 ev.pl_u.pl_mo.pl_pid,
1647 (void *) ev.pl_u.pl_mo.pl_start,
1648 (void *) ev.pl_u.pl_mo.pl_end);
1650 case PMCLOG_TYPE_PCSAMPLE:
1651 PMCSTAT_PRINT_ENTRY("sample","0x%x %d %p %c",
1652 ev.pl_u.pl_s.pl_pmcid,
1653 ev.pl_u.pl_s.pl_pid,
1654 (void *) ev.pl_u.pl_s.pl_pc,
1655 ev.pl_u.pl_s.pl_usermode ? 'u' : 's');
1657 case PMCLOG_TYPE_PMCALLOCATE:
1658 PMCSTAT_PRINT_ENTRY("allocate","0x%x \"%s\" 0x%x",
1659 ev.pl_u.pl_a.pl_pmcid,
1660 ev.pl_u.pl_a.pl_evname,
1661 ev.pl_u.pl_a.pl_flags);
1663 case PMCLOG_TYPE_PMCALLOCATEDYN:
1664 PMCSTAT_PRINT_ENTRY("allocatedyn","0x%x \"%s\" 0x%x",
1665 ev.pl_u.pl_ad.pl_pmcid,
1666 ev.pl_u.pl_ad.pl_evname,
1667 ev.pl_u.pl_ad.pl_flags);
1669 case PMCLOG_TYPE_PMCATTACH:
1670 PMCSTAT_PRINT_ENTRY("attach","0x%x %d \"%s\"",
1671 ev.pl_u.pl_t.pl_pmcid,
1672 ev.pl_u.pl_t.pl_pid,
1673 ev.pl_u.pl_t.pl_pathname);
1675 case PMCLOG_TYPE_PMCDETACH:
1676 PMCSTAT_PRINT_ENTRY("detach","0x%x %d",
1677 ev.pl_u.pl_d.pl_pmcid,
1678 ev.pl_u.pl_d.pl_pid);
1680 case PMCLOG_TYPE_PROCCSW:
1681 PMCSTAT_PRINT_ENTRY("cswval","0x%x %d %jd",
1682 ev.pl_u.pl_c.pl_pmcid,
1683 ev.pl_u.pl_c.pl_pid,
1684 ev.pl_u.pl_c.pl_value);
1686 case PMCLOG_TYPE_PROCEXEC:
1687 PMCSTAT_PRINT_ENTRY("exec","0x%x %d %p \"%s\"",
1688 ev.pl_u.pl_x.pl_pmcid,
1689 ev.pl_u.pl_x.pl_pid,
1690 (void *) ev.pl_u.pl_x.pl_entryaddr,
1691 ev.pl_u.pl_x.pl_pathname);
1693 case PMCLOG_TYPE_PROCEXIT:
1694 PMCSTAT_PRINT_ENTRY("exitval","0x%x %d %jd",
1695 ev.pl_u.pl_e.pl_pmcid,
1696 ev.pl_u.pl_e.pl_pid,
1697 ev.pl_u.pl_e.pl_value);
1699 case PMCLOG_TYPE_PROCFORK:
1700 PMCSTAT_PRINT_ENTRY("fork","%d %d",
1701 ev.pl_u.pl_f.pl_oldpid,
1702 ev.pl_u.pl_f.pl_newpid);
1704 case PMCLOG_TYPE_USERDATA:
1705 PMCSTAT_PRINT_ENTRY("userdata","0x%x",
1706 ev.pl_u.pl_u.pl_userdata);
1708 case PMCLOG_TYPE_SYSEXIT:
1709 PMCSTAT_PRINT_ENTRY("exit","%d",
1710 ev.pl_u.pl_se.pl_pid);
1713 fprintf(args.pa_printfile, "unknown event (type %d).\n",
1718 if (ev.pl_state == PMCLOG_EOF)
1719 return (PMCSTAT_FINISHED);
1720 else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1721 return (PMCSTAT_RUNNING);
1724 "ERROR: event parsing failed (record %jd, offset 0x%jx).",
1725 (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1730 * Public Interfaces.
1734 * Close a logfile, after first flushing all in-module queued data.
1738 pmcstat_close_log(void)
1740 /* If a local logfile is configured ask the kernel to stop
1741 * and flush data. Kernel will close the file when data is flushed
1742 * so keep the status to EXITING.
1744 if (args.pa_logfd != -1) {
1745 if (pmc_close_logfile() < 0)
1746 err(EX_OSERR, "ERROR: logging failed");
1749 return (args.pa_flags & FLAG_HAS_PIPE ? PMCSTAT_EXITING :
1756 * Open a log file, for reading or writing.
1758 * The function returns the fd of a successfully opened log or -1 in
1763 pmcstat_open_log(const char *path, int mode)
1767 const char *p, *errstr;
1768 struct addrinfo hints, *res, *res0;
1769 char hostname[MAXHOSTNAMELEN];
1775 * If 'path' is "-" then open one of stdin or stdout depending
1776 * on the value of 'mode'.
1778 * If 'path' contains a ':' and does not start with a '/' or '.',
1779 * and is being opened for writing, treat it as a "host:port"
1780 * specification and open a network socket.
1782 * Otherwise, treat 'path' as a file name and open that.
1784 if (path[0] == '-' && path[1] == '\0')
1785 fd = (mode == PMCSTAT_OPEN_FOR_READ) ? 0 : 1;
1786 else if (path[0] != '/' &&
1787 path[0] != '.' && strchr(path, ':') != NULL) {
1789 p = strrchr(path, ':');
1791 if (p == path || hlen >= sizeof(hostname)) {
1792 errstr = strerror(EINVAL);
1796 assert(hlen < sizeof(hostname));
1797 (void) strncpy(hostname, path, hlen);
1798 hostname[hlen] = '\0';
1800 (void) memset(&hints, 0, sizeof(hints));
1801 hints.ai_family = AF_UNSPEC;
1802 hints.ai_socktype = SOCK_STREAM;
1803 if ((error = getaddrinfo(hostname, p+1, &hints, &res0)) != 0) {
1804 errstr = gai_strerror(error);
1809 for (res = res0; res; res = res->ai_next) {
1810 if ((fd = socket(res->ai_family, res->ai_socktype,
1811 res->ai_protocol)) < 0) {
1812 errstr = strerror(errno);
1815 if (mode == PMCSTAT_OPEN_FOR_READ) {
1816 if (bind(fd, res->ai_addr, res->ai_addrlen) < 0) {
1817 errstr = strerror(errno);
1823 cfd = accept(fd, NULL, NULL);
1826 errstr = strerror(errno);
1832 if (connect(fd, res->ai_addr, res->ai_addrlen) < 0) {
1833 errstr = strerror(errno);
1844 } else if ((fd = open(path, mode == PMCSTAT_OPEN_FOR_READ ?
1845 O_RDONLY : (O_WRONLY|O_CREAT|O_TRUNC),
1846 S_IRUSR|S_IWUSR|S_IRGRP|S_IROTH)) < 0)
1847 errstr = strerror(errno);
1851 errx(EX_OSERR, "ERROR: Cannot open \"%s\" for %s: %s.", path,
1852 (mode == PMCSTAT_OPEN_FOR_READ ? "reading" : "writing"),
1859 * Process a log file in offline analysis mode.
1863 pmcstat_process_log(void)
1867 * If analysis has not been asked for, just print the log to
1868 * the current output file.
1870 if (args.pa_flags & FLAG_DO_PRINT)
1871 return (pmcstat_print_log());
1873 return (pmcstat_analyze_log());
1877 * Refresh top display.
1881 pmcstat_refresh_top(void)
1886 struct pmcstat_pmcrecord *pmcpr;
1888 /* If in pause mode do not refresh display. */
1892 /* Wait until PMC pop in the log. */
1893 pmcpr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1897 /* Format PMC name. */
1898 if (pmcstat_mergepmc)
1899 snprintf(pmcname, sizeof(pmcname), "[%s]",
1900 pmcstat_string_unintern(pmcpr->pr_pmcname));
1902 snprintf(pmcname, sizeof(pmcname), "%s.%d",
1903 pmcstat_string_unintern(pmcpr->pr_pmcname),
1904 pmcstat_pmcinfilter);
1906 /* Format samples count. */
1907 if (ps_samples_period > 0)
1908 v = (pmcpr->pr_samples * 100.0) / ps_samples_period;
1911 v_attrs = PMCSTAT_ATTRPERCENT(v);
1913 PMCSTAT_PRINTBEGIN();
1914 PMCSTAT_PRINTW("PMC: %s Samples: %u ",
1917 PMCSTAT_ATTRON(v_attrs);
1918 PMCSTAT_PRINTW("(%.1f%%) ", v);
1919 PMCSTAT_ATTROFF(v_attrs);
1920 PMCSTAT_PRINTW(", %u unresolved\n\n",
1921 pmcpr->pr_dubious_frames);
1922 if (plugins[args.pa_plugin].pl_topdisplay != NULL)
1923 plugins[args.pa_plugin].pl_topdisplay();
1928 * Find the next pmc index to display.
1932 pmcstat_changefilter(void)
1935 struct pmcstat_pmcrecord *pmcr;
1938 * Find the next merge target.
1940 if (pmcstat_mergepmc) {
1941 pmcin = pmcstat_pmcinfilter;
1944 pmcr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1945 if (pmcr == NULL || pmcr == pmcr->pr_merge)
1948 pmcstat_pmcinfilter++;
1949 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
1950 pmcstat_pmcinfilter = 0;
1952 } while (pmcstat_pmcinfilter != pmcin);
1957 * Top mode keypress.
1961 pmcstat_keypress_log(void)
1966 w = newwin(1, 0, 1, 0);
1968 wprintw(w, "Key: %c => ", c);
1971 wprintw(w, "enter mode 'd' or 'a' => ");
1974 args.pa_topmode = PMCSTAT_TOP_DELTA;
1975 wprintw(w, "switching to delta mode");
1977 args.pa_topmode = PMCSTAT_TOP_ACCUM;
1978 wprintw(w, "switching to accumulation mode");
1982 pmcstat_mergepmc = !pmcstat_mergepmc;
1984 * Changing merge state require data reset.
1986 if (plugins[args.pa_plugin].pl_shutdown != NULL)
1987 plugins[args.pa_plugin].pl_shutdown(NULL);
1988 pmcstat_stats_reset(0);
1989 if (plugins[args.pa_plugin].pl_init != NULL)
1990 plugins[args.pa_plugin].pl_init();
1992 /* Update filter to be on a merge target. */
1993 pmcstat_changefilter();
1994 wprintw(w, "merge PMC %s", pmcstat_mergepmc ? "on" : "off");
1997 /* Close current plugin. */
1998 if (plugins[args.pa_plugin].pl_shutdown != NULL)
1999 plugins[args.pa_plugin].pl_shutdown(NULL);
2001 /* Find next top display available. */
2004 if (plugins[args.pa_plugin].pl_name == NULL)
2006 } while (plugins[args.pa_plugin].pl_topdisplay == NULL);
2008 /* Open new plugin. */
2009 pmcstat_stats_reset(0);
2010 if (plugins[args.pa_plugin].pl_init != NULL)
2011 plugins[args.pa_plugin].pl_init();
2012 wprintw(w, "switching to plugin %s",
2013 plugins[args.pa_plugin].pl_name);
2016 pmcstat_pmcinfilter++;
2017 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
2018 pmcstat_pmcinfilter = 0;
2019 pmcstat_changefilter();
2020 wprintw(w, "switching to PMC %s.%d",
2021 pmcstat_pmcindex_to_name(pmcstat_pmcinfilter),
2022 pmcstat_pmcinfilter);
2025 pmcstat_pause = !pmcstat_pause;
2027 wprintw(w, "pause => press space again to continue");
2030 wprintw(w, "exiting...");
2034 if (plugins[args.pa_plugin].pl_topkeypress != NULL)
2035 if (plugins[args.pa_plugin].pl_topkeypress(c, w))
2050 pmcstat_display_log(void)
2053 pmcstat_refresh_top();
2055 /* Reset everythings if delta mode. */
2056 if (args.pa_topmode == PMCSTAT_TOP_DELTA) {
2057 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2058 plugins[args.pa_plugin].pl_shutdown(NULL);
2059 pmcstat_stats_reset(0);
2060 if (plugins[args.pa_plugin].pl_init != NULL)
2061 plugins[args.pa_plugin].pl_init();
2067 * Configure a plugins.
2071 pmcstat_pluginconfigure_log(char *opt)
2074 if (strncmp(opt, "threshold=", 10) == 0) {
2075 pmcstat_threshold = atof(opt+10);
2077 if (plugins[args.pa_plugin].pl_configure != NULL) {
2078 if (!plugins[args.pa_plugin].pl_configure(opt))
2080 "ERROR: unknown option <%s>.", opt);
2086 * Initialize module.
2090 pmcstat_initialize_logging(void)
2094 /* use a convenient format for 'ldd' output */
2095 if (setenv("LD_TRACE_LOADED_OBJECTS_FMT1","%o \"%p\" %x\n",1) != 0)
2096 err(EX_OSERR, "ERROR: Cannot setenv");
2098 /* Initialize hash tables */
2099 pmcstat_string_initialize();
2100 for (i = 0; i < PMCSTAT_NHASH; i++) {
2101 LIST_INIT(&pmcstat_image_hash[i]);
2102 LIST_INIT(&pmcstat_process_hash[i]);
2106 * Create a fake 'process' entry for the kernel with pid -1.
2107 * hwpmc(4) will subsequently inform us about where the kernel
2108 * and any loaded kernel modules are mapped.
2110 if ((pmcstat_kernproc = pmcstat_process_lookup((pid_t) -1,
2111 PMCSTAT_ALLOCATE)) == NULL)
2112 err(EX_OSERR, "ERROR: Cannot initialize logging");
2117 /* Merge PMC with same name. */
2118 pmcstat_mergepmc = args.pa_mergepmc;
2121 * Initialize plugins
2124 if (plugins[args.pa_pplugin].pl_init != NULL)
2125 plugins[args.pa_pplugin].pl_init();
2126 if (plugins[args.pa_plugin].pl_init != NULL)
2127 plugins[args.pa_plugin].pl_init();
2135 pmcstat_shutdown_logging(void)
2139 struct pmcstat_image *pi, *pitmp;
2140 struct pmcstat_process *pp, *pptmp;
2141 struct pmcstat_pcmap *ppm, *ppmtmp;
2143 /* determine where to send the map file */
2145 if (args.pa_mapfilename != NULL)
2146 mf = (strcmp(args.pa_mapfilename, "-") == 0) ?
2147 args.pa_printfile : fopen(args.pa_mapfilename, "w");
2149 if (mf == NULL && args.pa_flags & FLAG_DO_GPROF &&
2150 args.pa_verbosity >= 2)
2151 mf = args.pa_printfile;
2154 (void) fprintf(mf, "MAP:\n");
2157 * Shutdown the plugins
2160 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2161 plugins[args.pa_plugin].pl_shutdown(mf);
2162 if (plugins[args.pa_pplugin].pl_shutdown != NULL)
2163 plugins[args.pa_pplugin].pl_shutdown(mf);
2165 for (i = 0; i < PMCSTAT_NHASH; i++) {
2166 LIST_FOREACH_SAFE(pi, &pmcstat_image_hash[i], pi_next,
2168 if (plugins[args.pa_plugin].pl_shutdownimage != NULL)
2169 plugins[args.pa_plugin].pl_shutdownimage(pi);
2170 if (plugins[args.pa_pplugin].pl_shutdownimage != NULL)
2171 plugins[args.pa_pplugin].pl_shutdownimage(pi);
2173 free(pi->pi_symbols);
2174 if (pi->pi_addr2line != NULL)
2175 pclose(pi->pi_addr2line);
2176 LIST_REMOVE(pi, pi_next);
2180 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[i], pp_next,
2182 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
2183 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
2186 LIST_REMOVE(pp, pp_next);
2191 pmcstat_string_shutdown();
2194 * Print errors unless -q was specified. Print all statistics
2197 #define PRINT(N,V) do { \
2198 if (pmcstat_stats.ps_##V || args.pa_verbosity >= 2) \
2199 (void) fprintf(args.pa_printfile, " %-40s %d\n",\
2200 N, pmcstat_stats.ps_##V); \
2203 if (args.pa_verbosity >= 1 && (args.pa_flags & FLAG_DO_ANALYSIS)) {
2204 (void) fprintf(args.pa_printfile, "CONVERSION STATISTICS:\n");
2205 PRINT("#exec/a.out", exec_aout);
2206 PRINT("#exec/elf", exec_elf);
2207 PRINT("#exec/unknown", exec_indeterminable);
2208 PRINT("#exec handling errors", exec_errors);
2209 PRINT("#samples/total", samples_total);
2210 PRINT("#samples/unclaimed", samples_unknown_offset);
2211 PRINT("#samples/unknown-object", samples_indeterminable);
2212 PRINT("#samples/unknown-function", samples_unknown_function);
2213 PRINT("#callchain/dubious-frames", callchain_dubious_frames);