1 // SPDX-License-Identifier: GPL-2.0
3 #include <linux/kernel.h>
4 #include <linux/types.h>
10 #include <sys/param.h>
12 #include "parse-events.h"
15 #include "thread_map.h"
23 #include "sane_ctype.h"
33 static unsigned int hex(char c)
35 if (c >= '0' && c <= '9')
37 if (c >= 'a' && c <= 'f')
42 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
45 size_t bytes_read = 0;
46 unsigned char *chunk_start = *buf;
49 while (*buf_len > 0) {
52 /* Get 2 hex digits */
60 /* Store byte and advance buf */
61 **buf = (hex(c1) << 4) | hex(c2);
72 * objdump will display raw insn as LE if code endian
73 * is LE and bytes_per_chunk > 1. In that case reverse
74 * the chunk we just read.
76 * see disassemble_bytes() at binutils/objdump.c for details
77 * how objdump chooses display endian)
79 if (bytes_read > 1 && !bigendian()) {
80 unsigned char *chunk_end = chunk_start + bytes_read - 1;
83 while (chunk_start < chunk_end) {
85 *chunk_start = *chunk_end;
95 static size_t read_objdump_line(const char *line, unsigned char *buf,
99 size_t ret, bytes_read = 0;
101 /* Skip to a colon */
102 p = strchr(line, ':');
107 /* Skip initial spaces */
115 ret = read_objdump_chunk(&p, &buf, &buf_len);
120 /* return number of successfully read bytes */
124 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
127 size_t line_len, off_last = 0;
130 u64 addr, last_addr = start_addr;
132 while (off_last < *len) {
133 size_t off, read_bytes, written_bytes;
134 unsigned char tmp[BUFSZ];
136 ret = getline(&line, &line_len, f);
140 pr_debug("getline failed\n");
145 /* read objdump data into temporary buffer */
146 read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
150 if (sscanf(line, "%"PRIx64, &addr) != 1)
152 if (addr < last_addr) {
153 pr_debug("addr going backwards, read beyond section?\n");
158 /* copy it from temporary buffer to 'buf' according
159 * to address on current objdump line */
160 off = addr - start_addr;
163 written_bytes = MIN(read_bytes, *len - off);
164 memcpy(buf + off, tmp, written_bytes);
165 off_last = off + written_bytes;
168 /* len returns number of bytes that could not be read */
176 static int read_via_objdump(const char *filename, u64 addr, void *buf,
179 char cmd[PATH_MAX * 2];
184 fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
185 ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
187 if (ret <= 0 || (size_t)ret >= sizeof(cmd))
190 pr_debug("Objdump command is: %s\n", cmd);
192 /* Ignore objdump errors */
193 strcat(cmd, " 2>/dev/null");
197 pr_debug("popen failed\n");
201 ret = read_objdump_output(f, buf, &len, addr);
203 pr_debug("objdump read too few bytes: %zd\n", len);
213 static void dump_buf(unsigned char *buf, size_t len)
217 for (i = 0; i < len; i++) {
218 pr_debug("0x%02x ", buf[i]);
225 static int read_object_code(u64 addr, size_t len, u8 cpumode,
226 struct thread *thread, struct state *state)
228 struct addr_location al;
229 unsigned char buf1[BUFSZ];
230 unsigned char buf2[BUFSZ];
233 const char *objdump_name;
234 char decomp_name[KMOD_DECOMP_LEN];
238 pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
240 if (!thread__find_map(thread, cpumode, addr, &al) || !al.map->dso) {
241 if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
242 pr_debug("Hypervisor address can not be resolved - skipping\n");
246 pr_debug("thread__find_map failed\n");
250 pr_debug("File is: %s\n", al.map->dso->long_name);
252 if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
253 !dso__is_kcore(al.map->dso)) {
254 pr_debug("Unexpected kernel address - skipping\n");
258 pr_debug("On file address is: %#"PRIx64"\n", al.addr);
263 /* Do not go off the map */
264 if (addr + len > al.map->end)
265 len = al.map->end - addr;
267 /* Read the object code using perf */
268 ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
270 if (ret_len != len) {
271 pr_debug("dso__data_read_offset failed\n");
276 * Converting addresses for use by objdump requires more information.
277 * map__load() does that. See map__rip_2objdump() for details.
279 if (map__load(al.map))
282 /* objdump struggles with kcore - try each map only once */
283 if (dso__is_kcore(al.map->dso)) {
286 for (d = 0; d < state->done_cnt; d++) {
287 if (state->done[d] == al.map->start) {
288 pr_debug("kcore map tested already");
289 pr_debug(" - skipping\n");
293 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
294 pr_debug("Too many kcore maps - skipping\n");
297 state->done[state->done_cnt++] = al.map->start;
300 objdump_name = al.map->dso->long_name;
301 if (dso__needs_decompress(al.map->dso)) {
302 if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
304 sizeof(decomp_name)) < 0) {
305 pr_debug("decompression failed\n");
310 objdump_name = decomp_name;
313 /* Read the object code using objdump */
314 objdump_addr = map__rip_2objdump(al.map, al.addr);
315 ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
318 unlink(objdump_name);
322 * The kernel maps are inaccurate - assume objdump is right in
325 if (cpumode == PERF_RECORD_MISC_KERNEL ||
326 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
329 pr_debug("Reducing len to %zu\n", len);
330 } else if (dso__is_kcore(al.map->dso)) {
332 * objdump cannot handle very large segments
333 * that may be found in kcore.
335 pr_debug("objdump failed for kcore");
336 pr_debug(" - skipping\n");
344 pr_debug("read_via_objdump failed\n");
348 /* The results should be identical */
349 if (memcmp(buf1, buf2, len)) {
350 pr_debug("Bytes read differ from those read by objdump\n");
351 pr_debug("buf1 (dso):\n");
353 pr_debug("buf2 (objdump):\n");
357 pr_debug("Bytes read match those read by objdump\n");
362 static int process_sample_event(struct machine *machine,
363 struct perf_evlist *evlist,
364 union perf_event *event, struct state *state)
366 struct perf_sample sample;
367 struct thread *thread;
370 if (perf_evlist__parse_sample(evlist, event, &sample)) {
371 pr_debug("perf_evlist__parse_sample failed\n");
375 thread = machine__findnew_thread(machine, sample.pid, sample.tid);
377 pr_debug("machine__findnew_thread failed\n");
381 ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
386 static int process_event(struct machine *machine, struct perf_evlist *evlist,
387 union perf_event *event, struct state *state)
389 if (event->header.type == PERF_RECORD_SAMPLE)
390 return process_sample_event(machine, evlist, event, state);
392 if (event->header.type == PERF_RECORD_THROTTLE ||
393 event->header.type == PERF_RECORD_UNTHROTTLE)
396 if (event->header.type < PERF_RECORD_MAX) {
399 ret = machine__process_event(machine, event, NULL);
401 pr_debug("machine__process_event failed, event type %u\n",
409 static int process_events(struct machine *machine, struct perf_evlist *evlist,
412 union perf_event *event;
413 struct perf_mmap *md;
416 for (i = 0; i < evlist->nr_mmaps; i++) {
417 md = &evlist->mmap[i];
418 if (perf_mmap__read_init(md) < 0)
421 while ((event = perf_mmap__read_event(md)) != NULL) {
422 ret = process_event(machine, evlist, event, state);
423 perf_mmap__consume(md);
427 perf_mmap__read_done(md);
432 static int comp(const void *a, const void *b)
434 return *(int *)a - *(int *)b;
437 static void do_sort_something(void)
441 for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
442 buf[i] = ARRAY_SIZE(buf) - i - 1;
444 qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
446 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
448 pr_debug("qsort failed\n");
454 static void sort_something(void)
458 for (i = 0; i < 10; i++)
462 static void syscall_something(void)
467 for (i = 0; i < 1000; i++) {
468 if (pipe(pipefd) < 0) {
469 pr_debug("pipe failed\n");
477 static void fs_something(void)
479 const char *test_file_name = "temp-perf-code-reading-test-file--";
483 for (i = 0; i < 1000; i++) {
484 f = fopen(test_file_name, "w+");
487 unlink(test_file_name);
492 static const char *do_determine_event(bool excl_kernel)
494 const char *event = excl_kernel ? "cycles:u" : "cycles";
497 char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
501 if (get_cpuid(cpuid, sizeof(cpuid)))
503 ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
504 model, cpum_cf_v, &cpum_cf_a);
505 if (ret != 5) /* Not available */
507 if (excl_kernel && (cpum_cf_a & 4))
509 if (!excl_kernel && (cpum_cf_a & 2))
512 /* Fall through: missing authorization */
514 event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
520 static void do_something(void)
530 TEST_CODE_READING_OK,
531 TEST_CODE_READING_NO_VMLINUX,
532 TEST_CODE_READING_NO_KCORE,
533 TEST_CODE_READING_NO_ACCESS,
534 TEST_CODE_READING_NO_KERNEL_OBJ,
537 static int do_test_code_reading(bool try_kcore)
539 struct machine *machine;
540 struct thread *thread;
541 struct record_opts opts = {
542 .mmap_pages = UINT_MAX,
543 .user_freq = UINT_MAX,
544 .user_interval = ULLONG_MAX,
550 struct state state = {
553 struct thread_map *threads = NULL;
554 struct cpu_map *cpus = NULL;
555 struct perf_evlist *evlist = NULL;
556 struct perf_evsel *evsel = NULL;
560 bool have_vmlinux, have_kcore, excl_kernel = false;
564 machine = machine__new_host();
565 machine->env = &perf_env;
567 ret = machine__create_kernel_maps(machine);
569 pr_debug("machine__create_kernel_maps failed\n");
573 /* Force the use of kallsyms instead of vmlinux to try kcore */
575 symbol_conf.kallsyms_name = "/proc/kallsyms";
577 /* Load kernel map */
578 map = machine__kernel_map(machine);
579 ret = map__load(map);
581 pr_debug("map__load failed\n");
584 have_vmlinux = dso__is_vmlinux(map->dso);
585 have_kcore = dso__is_kcore(map->dso);
587 /* 2nd time through we just try kcore */
588 if (try_kcore && !have_kcore)
589 return TEST_CODE_READING_NO_KCORE;
591 /* No point getting kernel events if there is no kernel object */
592 if (!have_vmlinux && !have_kcore)
595 threads = thread_map__new_by_tid(pid);
597 pr_debug("thread_map__new_by_tid failed\n");
601 ret = perf_event__synthesize_thread_map(NULL, threads,
602 perf_event__process, machine, false);
604 pr_debug("perf_event__synthesize_thread_map failed\n");
608 thread = machine__findnew_thread(machine, pid, pid);
610 pr_debug("machine__findnew_thread failed\n");
614 cpus = cpu_map__new(NULL);
616 pr_debug("cpu_map__new failed\n");
623 evlist = perf_evlist__new();
625 pr_debug("perf_evlist__new failed\n");
629 perf_evlist__set_maps(evlist, cpus, threads);
631 str = do_determine_event(excl_kernel);
632 pr_debug("Parsing event '%s'\n", str);
633 ret = parse_events(evlist, str, NULL);
635 pr_debug("parse_events failed\n");
639 perf_evlist__config(evlist, &opts, NULL);
641 evsel = perf_evlist__first(evlist);
643 evsel->attr.comm = 1;
644 evsel->attr.disabled = 1;
645 evsel->attr.enable_on_exec = 0;
647 ret = perf_evlist__open(evlist);
652 * Both cpus and threads are now owned by evlist
653 * and will be freed by following perf_evlist__set_maps
654 * call. Getting refference to keep them alive.
657 thread_map__get(threads);
658 perf_evlist__set_maps(evlist, NULL, NULL);
659 perf_evlist__delete(evlist);
666 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
667 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
675 ret = perf_evlist__mmap(evlist, UINT_MAX);
677 pr_debug("perf_evlist__mmap failed\n");
681 perf_evlist__enable(evlist);
685 perf_evlist__disable(evlist);
687 ret = process_events(machine, evlist, &state);
691 if (!have_vmlinux && !have_kcore && !try_kcore)
692 err = TEST_CODE_READING_NO_KERNEL_OBJ;
693 else if (!have_vmlinux && !try_kcore)
694 err = TEST_CODE_READING_NO_VMLINUX;
695 else if (excl_kernel)
696 err = TEST_CODE_READING_NO_ACCESS;
698 err = TEST_CODE_READING_OK;
704 perf_evlist__delete(evlist);
707 thread_map__put(threads);
709 machine__delete_threads(machine);
710 machine__delete(machine);
715 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
719 ret = do_test_code_reading(false);
721 ret = do_test_code_reading(true);
724 case TEST_CODE_READING_OK:
726 case TEST_CODE_READING_NO_VMLINUX:
727 pr_debug("no vmlinux\n");
729 case TEST_CODE_READING_NO_KCORE:
730 pr_debug("no kcore\n");
732 case TEST_CODE_READING_NO_ACCESS:
733 pr_debug("no access\n");
735 case TEST_CODE_READING_NO_KERNEL_OBJ:
736 pr_debug("no kernel obj\n");