2 #include <linux/kernel.h>
3 #include <linux/types.h>
11 #include "parse-events.h"
14 #include "thread_map.h"
22 #include "sane_ctype.h"
32 static unsigned int hex(char c)
34 if (c >= '0' && c <= '9')
36 if (c >= 'a' && c <= 'f')
41 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
44 size_t bytes_read = 0;
45 unsigned char *chunk_start = *buf;
48 while (*buf_len > 0) {
51 /* Get 2 hex digits */
59 /* Store byte and advance buf */
60 **buf = (hex(c1) << 4) | hex(c2);
71 * objdump will display raw insn as LE if code endian
72 * is LE and bytes_per_chunk > 1. In that case reverse
73 * the chunk we just read.
75 * see disassemble_bytes() at binutils/objdump.c for details
76 * how objdump chooses display endian)
78 if (bytes_read > 1 && !bigendian()) {
79 unsigned char *chunk_end = chunk_start + bytes_read - 1;
82 while (chunk_start < chunk_end) {
84 *chunk_start = *chunk_end;
94 static size_t read_objdump_line(const char *line, unsigned char *buf,
98 size_t ret, bytes_read = 0;
100 /* Skip to a colon */
101 p = strchr(line, ':');
106 /* Skip initial spaces */
114 ret = read_objdump_chunk(&p, &buf, &buf_len);
119 /* return number of successfully read bytes */
123 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
126 size_t line_len, off_last = 0;
129 u64 addr, last_addr = start_addr;
131 while (off_last < *len) {
132 size_t off, read_bytes, written_bytes;
133 unsigned char tmp[BUFSZ];
135 ret = getline(&line, &line_len, f);
139 pr_debug("getline failed\n");
144 /* read objdump data into temporary buffer */
145 read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
149 if (sscanf(line, "%"PRIx64, &addr) != 1)
151 if (addr < last_addr) {
152 pr_debug("addr going backwards, read beyond section?\n");
157 /* copy it from temporary buffer to 'buf' according
158 * to address on current objdump line */
159 off = addr - start_addr;
162 written_bytes = MIN(read_bytes, *len - off);
163 memcpy(buf + off, tmp, written_bytes);
164 off_last = off + written_bytes;
167 /* len returns number of bytes that could not be read */
175 static int read_via_objdump(const char *filename, u64 addr, void *buf,
178 char cmd[PATH_MAX * 2];
183 fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
184 ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
186 if (ret <= 0 || (size_t)ret >= sizeof(cmd))
189 pr_debug("Objdump command is: %s\n", cmd);
191 /* Ignore objdump errors */
192 strcat(cmd, " 2>/dev/null");
196 pr_debug("popen failed\n");
200 ret = read_objdump_output(f, buf, &len, addr);
202 pr_debug("objdump read too few bytes: %zd\n", len);
212 static void dump_buf(unsigned char *buf, size_t len)
216 for (i = 0; i < len; i++) {
217 pr_debug("0x%02x ", buf[i]);
224 static int read_object_code(u64 addr, size_t len, u8 cpumode,
225 struct thread *thread, struct state *state)
227 struct addr_location al;
228 unsigned char buf1[BUFSZ];
229 unsigned char buf2[BUFSZ];
232 const char *objdump_name;
233 char decomp_name[KMOD_DECOMP_LEN];
236 pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
238 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
239 if (!al.map || !al.map->dso) {
240 pr_debug("thread__find_addr_map failed\n");
244 pr_debug("File is: %s\n", al.map->dso->long_name);
246 if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
247 !dso__is_kcore(al.map->dso)) {
248 pr_debug("Unexpected kernel address - skipping\n");
252 pr_debug("On file address is: %#"PRIx64"\n", al.addr);
257 /* Do not go off the map */
258 if (addr + len > al.map->end)
259 len = al.map->end - addr;
261 /* Read the object code using perf */
262 ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
264 if (ret_len != len) {
265 pr_debug("dso__data_read_offset failed\n");
270 * Converting addresses for use by objdump requires more information.
271 * map__load() does that. See map__rip_2objdump() for details.
273 if (map__load(al.map))
276 /* objdump struggles with kcore - try each map only once */
277 if (dso__is_kcore(al.map->dso)) {
280 for (d = 0; d < state->done_cnt; d++) {
281 if (state->done[d] == al.map->start) {
282 pr_debug("kcore map tested already");
283 pr_debug(" - skipping\n");
287 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
288 pr_debug("Too many kcore maps - skipping\n");
291 state->done[state->done_cnt++] = al.map->start;
294 objdump_name = al.map->dso->long_name;
295 if (dso__needs_decompress(al.map->dso)) {
296 if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
298 sizeof(decomp_name)) < 0) {
299 pr_debug("decompression failed\n");
303 objdump_name = decomp_name;
306 /* Read the object code using objdump */
307 objdump_addr = map__rip_2objdump(al.map, al.addr);
308 ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
310 if (dso__needs_decompress(al.map->dso))
311 unlink(objdump_name);
315 * The kernel maps are inaccurate - assume objdump is right in
318 if (cpumode == PERF_RECORD_MISC_KERNEL ||
319 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
322 pr_debug("Reducing len to %zu\n", len);
323 } else if (dso__is_kcore(al.map->dso)) {
325 * objdump cannot handle very large segments
326 * that may be found in kcore.
328 pr_debug("objdump failed for kcore");
329 pr_debug(" - skipping\n");
337 pr_debug("read_via_objdump failed\n");
341 /* The results should be identical */
342 if (memcmp(buf1, buf2, len)) {
343 pr_debug("Bytes read differ from those read by objdump\n");
344 pr_debug("buf1 (dso):\n");
346 pr_debug("buf2 (objdump):\n");
350 pr_debug("Bytes read match those read by objdump\n");
355 static int process_sample_event(struct machine *machine,
356 struct perf_evlist *evlist,
357 union perf_event *event, struct state *state)
359 struct perf_sample sample;
360 struct thread *thread;
363 if (perf_evlist__parse_sample(evlist, event, &sample)) {
364 pr_debug("perf_evlist__parse_sample failed\n");
368 thread = machine__findnew_thread(machine, sample.pid, sample.tid);
370 pr_debug("machine__findnew_thread failed\n");
374 ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
379 static int process_event(struct machine *machine, struct perf_evlist *evlist,
380 union perf_event *event, struct state *state)
382 if (event->header.type == PERF_RECORD_SAMPLE)
383 return process_sample_event(machine, evlist, event, state);
385 if (event->header.type == PERF_RECORD_THROTTLE ||
386 event->header.type == PERF_RECORD_UNTHROTTLE)
389 if (event->header.type < PERF_RECORD_MAX) {
392 ret = machine__process_event(machine, event, NULL);
394 pr_debug("machine__process_event failed, event type %u\n",
402 static int process_events(struct machine *machine, struct perf_evlist *evlist,
405 union perf_event *event;
408 for (i = 0; i < evlist->nr_mmaps; i++) {
409 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
410 ret = process_event(machine, evlist, event, state);
411 perf_evlist__mmap_consume(evlist, i);
419 static int comp(const void *a, const void *b)
421 return *(int *)a - *(int *)b;
424 static void do_sort_something(void)
428 for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
429 buf[i] = ARRAY_SIZE(buf) - i - 1;
431 qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
433 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
435 pr_debug("qsort failed\n");
441 static void sort_something(void)
445 for (i = 0; i < 10; i++)
449 static void syscall_something(void)
454 for (i = 0; i < 1000; i++) {
455 if (pipe(pipefd) < 0) {
456 pr_debug("pipe failed\n");
464 static void fs_something(void)
466 const char *test_file_name = "temp-perf-code-reading-test-file--";
470 for (i = 0; i < 1000; i++) {
471 f = fopen(test_file_name, "w+");
474 unlink(test_file_name);
479 static void do_something(void)
489 TEST_CODE_READING_OK,
490 TEST_CODE_READING_NO_VMLINUX,
491 TEST_CODE_READING_NO_KCORE,
492 TEST_CODE_READING_NO_ACCESS,
493 TEST_CODE_READING_NO_KERNEL_OBJ,
496 static int do_test_code_reading(bool try_kcore)
498 struct machine *machine;
499 struct thread *thread;
500 struct record_opts opts = {
501 .mmap_pages = UINT_MAX,
502 .user_freq = UINT_MAX,
503 .user_interval = ULLONG_MAX,
509 struct state state = {
512 struct thread_map *threads = NULL;
513 struct cpu_map *cpus = NULL;
514 struct perf_evlist *evlist = NULL;
515 struct perf_evsel *evsel = NULL;
519 bool have_vmlinux, have_kcore, excl_kernel = false;
523 machine = machine__new_host();
525 ret = machine__create_kernel_maps(machine);
527 pr_debug("machine__create_kernel_maps failed\n");
531 /* Force the use of kallsyms instead of vmlinux to try kcore */
533 symbol_conf.kallsyms_name = "/proc/kallsyms";
535 /* Load kernel map */
536 map = machine__kernel_map(machine);
537 ret = map__load(map);
539 pr_debug("map__load failed\n");
542 have_vmlinux = dso__is_vmlinux(map->dso);
543 have_kcore = dso__is_kcore(map->dso);
545 /* 2nd time through we just try kcore */
546 if (try_kcore && !have_kcore)
547 return TEST_CODE_READING_NO_KCORE;
549 /* No point getting kernel events if there is no kernel object */
550 if (!have_vmlinux && !have_kcore)
553 threads = thread_map__new_by_tid(pid);
555 pr_debug("thread_map__new_by_tid failed\n");
559 ret = perf_event__synthesize_thread_map(NULL, threads,
560 perf_event__process, machine, false, 500);
562 pr_debug("perf_event__synthesize_thread_map failed\n");
566 thread = machine__findnew_thread(machine, pid, pid);
568 pr_debug("machine__findnew_thread failed\n");
572 cpus = cpu_map__new(NULL);
574 pr_debug("cpu_map__new failed\n");
581 evlist = perf_evlist__new();
583 pr_debug("perf_evlist__new failed\n");
587 perf_evlist__set_maps(evlist, cpus, threads);
593 pr_debug("Parsing event '%s'\n", str);
594 ret = parse_events(evlist, str, NULL);
596 pr_debug("parse_events failed\n");
600 perf_evlist__config(evlist, &opts, NULL);
602 evsel = perf_evlist__first(evlist);
604 evsel->attr.comm = 1;
605 evsel->attr.disabled = 1;
606 evsel->attr.enable_on_exec = 0;
608 ret = perf_evlist__open(evlist);
613 * Both cpus and threads are now owned by evlist
614 * and will be freed by following perf_evlist__set_maps
615 * call. Getting refference to keep them alive.
618 thread_map__get(threads);
619 perf_evlist__set_maps(evlist, NULL, NULL);
620 perf_evlist__delete(evlist);
627 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
628 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
636 ret = perf_evlist__mmap(evlist, UINT_MAX, false);
638 pr_debug("perf_evlist__mmap failed\n");
642 perf_evlist__enable(evlist);
646 perf_evlist__disable(evlist);
648 ret = process_events(machine, evlist, &state);
652 if (!have_vmlinux && !have_kcore && !try_kcore)
653 err = TEST_CODE_READING_NO_KERNEL_OBJ;
654 else if (!have_vmlinux && !try_kcore)
655 err = TEST_CODE_READING_NO_VMLINUX;
656 else if (excl_kernel)
657 err = TEST_CODE_READING_NO_ACCESS;
659 err = TEST_CODE_READING_OK;
665 perf_evlist__delete(evlist);
668 thread_map__put(threads);
670 machine__delete_threads(machine);
671 machine__delete(machine);
676 int test__code_reading(int subtest __maybe_unused)
680 ret = do_test_code_reading(false);
682 ret = do_test_code_reading(true);
685 case TEST_CODE_READING_OK:
687 case TEST_CODE_READING_NO_VMLINUX:
688 pr_debug("no vmlinux\n");
690 case TEST_CODE_READING_NO_KCORE:
691 pr_debug("no kcore\n");
693 case TEST_CODE_READING_NO_ACCESS:
694 pr_debug("no access\n");
696 case TEST_CODE_READING_NO_KERNEL_OBJ:
697 pr_debug("no kernel obj\n");