Home | History | Annotate | Download | only in util
      1 #include <linux/kernel.h>
      2 #include <traceevent/event-parse.h>
      3 
      4 #include <byteswap.h>
      5 #include <unistd.h>
      6 #include <sys/types.h>
      7 #include <sys/mman.h>
      8 
      9 #include "evlist.h"
     10 #include "evsel.h"
     11 #include "session.h"
     12 #include "tool.h"
     13 #include "sort.h"
     14 #include "util.h"
     15 #include "cpumap.h"
     16 #include "perf_regs.h"
     17 #include "vdso.h"
     18 
     19 static int perf_session__open(struct perf_session *self, bool force)
     20 {
     21 	struct stat input_stat;
     22 
     23 	if (!strcmp(self->filename, "-")) {
     24 		self->fd_pipe = true;
     25 		self->fd = STDIN_FILENO;
     26 
     27 		if (perf_session__read_header(self) < 0)
     28 			pr_err("incompatible file format (rerun with -v to learn more)");
     29 
     30 		return 0;
     31 	}
     32 
     33 	self->fd = open(self->filename, O_RDONLY);
     34 	if (self->fd < 0) {
     35 		int err = errno;
     36 
     37 		pr_err("failed to open %s: %s", self->filename, strerror(err));
     38 		if (err == ENOENT && !strcmp(self->filename, "perf.data"))
     39 			pr_err("  (try 'perf record' first)");
     40 		pr_err("\n");
     41 		return -errno;
     42 	}
     43 
     44 	if (fstat(self->fd, &input_stat) < 0)
     45 		goto out_close;
     46 
     47 	if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
     48 		pr_err("file %s not owned by current user or root\n",
     49 		       self->filename);
     50 		goto out_close;
     51 	}
     52 
     53 	if (!input_stat.st_size) {
     54 		pr_info("zero-sized file (%s), nothing to do!\n",
     55 			self->filename);
     56 		goto out_close;
     57 	}
     58 
     59 	if (perf_session__read_header(self) < 0) {
     60 		pr_err("incompatible file format (rerun with -v to learn more)");
     61 		goto out_close;
     62 	}
     63 
     64 	if (!perf_evlist__valid_sample_type(self->evlist)) {
     65 		pr_err("non matching sample_type");
     66 		goto out_close;
     67 	}
     68 
     69 	if (!perf_evlist__valid_sample_id_all(self->evlist)) {
     70 		pr_err("non matching sample_id_all");
     71 		goto out_close;
     72 	}
     73 
     74 	if (!perf_evlist__valid_read_format(self->evlist)) {
     75 		pr_err("non matching read_format");
     76 		goto out_close;
     77 	}
     78 
     79 	self->size = input_stat.st_size;
     80 	return 0;
     81 
     82 out_close:
     83 	close(self->fd);
     84 	self->fd = -1;
     85 	return -1;
     86 }
     87 
     88 void perf_session__set_id_hdr_size(struct perf_session *session)
     89 {
     90 	u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
     91 
     92 	machines__set_id_hdr_size(&session->machines, id_hdr_size);
     93 }
     94 
     95 int perf_session__create_kernel_maps(struct perf_session *self)
     96 {
     97 	int ret = machine__create_kernel_maps(&self->machines.host);
     98 
     99 	if (ret >= 0)
    100 		ret = machines__create_guest_kernel_maps(&self->machines);
    101 	return ret;
    102 }
    103 
    104 static void perf_session__destroy_kernel_maps(struct perf_session *self)
    105 {
    106 	machines__destroy_kernel_maps(&self->machines);
    107 }
    108 
    109 struct perf_session *perf_session__new(const char *filename, int mode,
    110 				       bool force, bool repipe,
    111 				       struct perf_tool *tool)
    112 {
    113 	struct perf_session *self;
    114 	struct stat st;
    115 	size_t len;
    116 
    117 	if (!filename || !strlen(filename)) {
    118 		if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
    119 			filename = "-";
    120 		else
    121 			filename = "perf.data";
    122 	}
    123 
    124 	len = strlen(filename);
    125 	self = zalloc(sizeof(*self) + len);
    126 
    127 	if (self == NULL)
    128 		goto out;
    129 
    130 	memcpy(self->filename, filename, len);
    131 	self->repipe = repipe;
    132 	INIT_LIST_HEAD(&self->ordered_samples.samples);
    133 	INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
    134 	INIT_LIST_HEAD(&self->ordered_samples.to_free);
    135 	machines__init(&self->machines);
    136 
    137 	if (mode == O_RDONLY) {
    138 		if (perf_session__open(self, force) < 0)
    139 			goto out_delete;
    140 		perf_session__set_id_hdr_size(self);
    141 	} else if (mode == O_WRONLY) {
    142 		/*
    143 		 * In O_RDONLY mode this will be performed when reading the
    144 		 * kernel MMAP event, in perf_event__process_mmap().
    145 		 */
    146 		if (perf_session__create_kernel_maps(self) < 0)
    147 			goto out_delete;
    148 	}
    149 
    150 	if (tool && tool->ordering_requires_timestamps &&
    151 	    tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
    152 		dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
    153 		tool->ordered_samples = false;
    154 	}
    155 
    156 out:
    157 	return self;
    158 out_delete:
    159 	perf_session__delete(self);
    160 	return NULL;
    161 }
    162 
    163 static void perf_session__delete_dead_threads(struct perf_session *session)
    164 {
    165 	machine__delete_dead_threads(&session->machines.host);
    166 }
    167 
    168 static void perf_session__delete_threads(struct perf_session *session)
    169 {
    170 	machine__delete_threads(&session->machines.host);
    171 }
    172 
    173 static void perf_session_env__delete(struct perf_session_env *env)
    174 {
    175 	free(env->hostname);
    176 	free(env->os_release);
    177 	free(env->version);
    178 	free(env->arch);
    179 	free(env->cpu_desc);
    180 	free(env->cpuid);
    181 
    182 	free(env->cmdline);
    183 	free(env->sibling_cores);
    184 	free(env->sibling_threads);
    185 	free(env->numa_nodes);
    186 	free(env->pmu_mappings);
    187 }
    188 
    189 void perf_session__delete(struct perf_session *self)
    190 {
    191 	perf_session__destroy_kernel_maps(self);
    192 	perf_session__delete_dead_threads(self);
    193 	perf_session__delete_threads(self);
    194 	perf_session_env__delete(&self->header.env);
    195 	machines__exit(&self->machines);
    196 	close(self->fd);
    197 	free(self);
    198 	vdso__exit();
    199 }
    200 
    201 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
    202 						 __maybe_unused,
    203 						 union perf_event *event
    204 						 __maybe_unused,
    205 						 struct perf_session *session
    206 						__maybe_unused)
    207 {
    208 	dump_printf(": unhandled!\n");
    209 	return 0;
    210 }
    211 
    212 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
    213 					 union perf_event *event __maybe_unused,
    214 					 struct perf_evlist **pevlist
    215 					 __maybe_unused)
    216 {
    217 	dump_printf(": unhandled!\n");
    218 	return 0;
    219 }
    220 
    221 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
    222 				     union perf_event *event __maybe_unused,
    223 				     struct perf_sample *sample __maybe_unused,
    224 				     struct perf_evsel *evsel __maybe_unused,
    225 				     struct machine *machine __maybe_unused)
    226 {
    227 	dump_printf(": unhandled!\n");
    228 	return 0;
    229 }
    230 
    231 static int process_event_stub(struct perf_tool *tool __maybe_unused,
    232 			      union perf_event *event __maybe_unused,
    233 			      struct perf_sample *sample __maybe_unused,
    234 			      struct machine *machine __maybe_unused)
    235 {
    236 	dump_printf(": unhandled!\n");
    237 	return 0;
    238 }
    239 
    240 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
    241 				       union perf_event *event __maybe_unused,
    242 				       struct perf_session *perf_session
    243 				       __maybe_unused)
    244 {
    245 	dump_printf(": unhandled!\n");
    246 	return 0;
    247 }
    248 
    249 static int process_finished_round(struct perf_tool *tool,
    250 				  union perf_event *event,
    251 				  struct perf_session *session);
    252 
    253 void perf_tool__fill_defaults(struct perf_tool *tool)
    254 {
    255 	if (tool->sample == NULL)
    256 		tool->sample = process_event_sample_stub;
    257 	if (tool->mmap == NULL)
    258 		tool->mmap = process_event_stub;
    259 	if (tool->mmap2 == NULL)
    260 		tool->mmap2 = process_event_stub;
    261 	if (tool->comm == NULL)
    262 		tool->comm = process_event_stub;
    263 	if (tool->fork == NULL)
    264 		tool->fork = process_event_stub;
    265 	if (tool->exit == NULL)
    266 		tool->exit = process_event_stub;
    267 	if (tool->lost == NULL)
    268 		tool->lost = perf_event__process_lost;
    269 	if (tool->read == NULL)
    270 		tool->read = process_event_sample_stub;
    271 	if (tool->throttle == NULL)
    272 		tool->throttle = process_event_stub;
    273 	if (tool->unthrottle == NULL)
    274 		tool->unthrottle = process_event_stub;
    275 	if (tool->attr == NULL)
    276 		tool->attr = process_event_synth_attr_stub;
    277 	if (tool->tracing_data == NULL)
    278 		tool->tracing_data = process_event_synth_tracing_data_stub;
    279 	if (tool->build_id == NULL)
    280 		tool->build_id = process_finished_round_stub;
    281 	if (tool->finished_round == NULL) {
    282 		if (tool->ordered_samples)
    283 			tool->finished_round = process_finished_round;
    284 		else
    285 			tool->finished_round = process_finished_round_stub;
    286 	}
    287 }
    288 
    289 void mem_bswap_32(void *src, int byte_size)
    290 {
    291 	u32 *m = src;
    292 	while (byte_size > 0) {
    293 		*m = bswap_32(*m);
    294 		byte_size -= sizeof(u32);
    295 		++m;
    296 	}
    297 }
    298 
    299 void mem_bswap_64(void *src, int byte_size)
    300 {
    301 	u64 *m = src;
    302 
    303 	while (byte_size > 0) {
    304 		*m = bswap_64(*m);
    305 		byte_size -= sizeof(u64);
    306 		++m;
    307 	}
    308 }
    309 
    310 static void swap_sample_id_all(union perf_event *event, void *data)
    311 {
    312 	void *end = (void *) event + event->header.size;
    313 	int size = end - data;
    314 
    315 	BUG_ON(size % sizeof(u64));
    316 	mem_bswap_64(data, size);
    317 }
    318 
    319 static void perf_event__all64_swap(union perf_event *event,
    320 				   bool sample_id_all __maybe_unused)
    321 {
    322 	struct perf_event_header *hdr = &event->header;
    323 	mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
    324 }
    325 
    326 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
    327 {
    328 	event->comm.pid = bswap_32(event->comm.pid);
    329 	event->comm.tid = bswap_32(event->comm.tid);
    330 
    331 	if (sample_id_all) {
    332 		void *data = &event->comm.comm;
    333 
    334 		data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
    335 		swap_sample_id_all(event, data);
    336 	}
    337 }
    338 
    339 static void perf_event__mmap_swap(union perf_event *event,
    340 				  bool sample_id_all)
    341 {
    342 	event->mmap.pid	  = bswap_32(event->mmap.pid);
    343 	event->mmap.tid	  = bswap_32(event->mmap.tid);
    344 	event->mmap.start = bswap_64(event->mmap.start);
    345 	event->mmap.len	  = bswap_64(event->mmap.len);
    346 	event->mmap.pgoff = bswap_64(event->mmap.pgoff);
    347 
    348 	if (sample_id_all) {
    349 		void *data = &event->mmap.filename;
    350 
    351 		data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
    352 		swap_sample_id_all(event, data);
    353 	}
    354 }
    355 
    356 static void perf_event__mmap2_swap(union perf_event *event,
    357 				  bool sample_id_all)
    358 {
    359 	event->mmap2.pid   = bswap_32(event->mmap2.pid);
    360 	event->mmap2.tid   = bswap_32(event->mmap2.tid);
    361 	event->mmap2.start = bswap_64(event->mmap2.start);
    362 	event->mmap2.len   = bswap_64(event->mmap2.len);
    363 	event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
    364 	event->mmap2.maj   = bswap_32(event->mmap2.maj);
    365 	event->mmap2.min   = bswap_32(event->mmap2.min);
    366 	event->mmap2.ino   = bswap_64(event->mmap2.ino);
    367 
    368 	if (sample_id_all) {
    369 		void *data = &event->mmap2.filename;
    370 
    371 		data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
    372 		swap_sample_id_all(event, data);
    373 	}
    374 }
    375 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
    376 {
    377 	event->fork.pid	 = bswap_32(event->fork.pid);
    378 	event->fork.tid	 = bswap_32(event->fork.tid);
    379 	event->fork.ppid = bswap_32(event->fork.ppid);
    380 	event->fork.ptid = bswap_32(event->fork.ptid);
    381 	event->fork.time = bswap_64(event->fork.time);
    382 
    383 	if (sample_id_all)
    384 		swap_sample_id_all(event, &event->fork + 1);
    385 }
    386 
    387 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
    388 {
    389 	event->read.pid		 = bswap_32(event->read.pid);
    390 	event->read.tid		 = bswap_32(event->read.tid);
    391 	event->read.value	 = bswap_64(event->read.value);
    392 	event->read.time_enabled = bswap_64(event->read.time_enabled);
    393 	event->read.time_running = bswap_64(event->read.time_running);
    394 	event->read.id		 = bswap_64(event->read.id);
    395 
    396 	if (sample_id_all)
    397 		swap_sample_id_all(event, &event->read + 1);
    398 }
    399 
    400 static u8 revbyte(u8 b)
    401 {
    402 	int rev = (b >> 4) | ((b & 0xf) << 4);
    403 	rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
    404 	rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
    405 	return (u8) rev;
    406 }
    407 
    408 /*
    409  * XXX this is hack in attempt to carry flags bitfield
    410  * throught endian village. ABI says:
    411  *
    412  * Bit-fields are allocated from right to left (least to most significant)
    413  * on little-endian implementations and from left to right (most to least
    414  * significant) on big-endian implementations.
    415  *
    416  * The above seems to be byte specific, so we need to reverse each
    417  * byte of the bitfield. 'Internet' also says this might be implementation
    418  * specific and we probably need proper fix and carry perf_event_attr
    419  * bitfield flags in separate data file FEAT_ section. Thought this seems
    420  * to work for now.
    421  */
    422 static void swap_bitfield(u8 *p, unsigned len)
    423 {
    424 	unsigned i;
    425 
    426 	for (i = 0; i < len; i++) {
    427 		*p = revbyte(*p);
    428 		p++;
    429 	}
    430 }
    431 
    432 /* exported for swapping attributes in file header */
    433 void perf_event__attr_swap(struct perf_event_attr *attr)
    434 {
    435 	attr->type		= bswap_32(attr->type);
    436 	attr->size		= bswap_32(attr->size);
    437 	attr->config		= bswap_64(attr->config);
    438 	attr->sample_period	= bswap_64(attr->sample_period);
    439 	attr->sample_type	= bswap_64(attr->sample_type);
    440 	attr->read_format	= bswap_64(attr->read_format);
    441 	attr->wakeup_events	= bswap_32(attr->wakeup_events);
    442 	attr->bp_type		= bswap_32(attr->bp_type);
    443 	attr->bp_addr		= bswap_64(attr->bp_addr);
    444 	attr->bp_len		= bswap_64(attr->bp_len);
    445 
    446 	swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
    447 }
    448 
    449 static void perf_event__hdr_attr_swap(union perf_event *event,
    450 				      bool sample_id_all __maybe_unused)
    451 {
    452 	size_t size;
    453 
    454 	perf_event__attr_swap(&event->attr.attr);
    455 
    456 	size = event->header.size;
    457 	size -= (void *)&event->attr.id - (void *)event;
    458 	mem_bswap_64(event->attr.id, size);
    459 }
    460 
    461 static void perf_event__event_type_swap(union perf_event *event,
    462 					bool sample_id_all __maybe_unused)
    463 {
    464 	event->event_type.event_type.event_id =
    465 		bswap_64(event->event_type.event_type.event_id);
    466 }
    467 
    468 static void perf_event__tracing_data_swap(union perf_event *event,
    469 					  bool sample_id_all __maybe_unused)
    470 {
    471 	event->tracing_data.size = bswap_32(event->tracing_data.size);
    472 }
    473 
    474 typedef void (*perf_event__swap_op)(union perf_event *event,
    475 				    bool sample_id_all);
    476 
    477 static perf_event__swap_op perf_event__swap_ops[] = {
    478 	[PERF_RECORD_MMAP]		  = perf_event__mmap_swap,
    479 	[PERF_RECORD_MMAP2]		  = perf_event__mmap2_swap,
    480 	[PERF_RECORD_COMM]		  = perf_event__comm_swap,
    481 	[PERF_RECORD_FORK]		  = perf_event__task_swap,
    482 	[PERF_RECORD_EXIT]		  = perf_event__task_swap,
    483 	[PERF_RECORD_LOST]		  = perf_event__all64_swap,
    484 	[PERF_RECORD_READ]		  = perf_event__read_swap,
    485 	[PERF_RECORD_SAMPLE]		  = perf_event__all64_swap,
    486 	[PERF_RECORD_HEADER_ATTR]	  = perf_event__hdr_attr_swap,
    487 	[PERF_RECORD_HEADER_EVENT_TYPE]	  = perf_event__event_type_swap,
    488 	[PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
    489 	[PERF_RECORD_HEADER_BUILD_ID]	  = NULL,
    490 	[PERF_RECORD_HEADER_MAX]	  = NULL,
    491 };
    492 
    493 struct sample_queue {
    494 	u64			timestamp;
    495 	u64			file_offset;
    496 	union perf_event	*event;
    497 	struct list_head	list;
    498 };
    499 
    500 static void perf_session_free_sample_buffers(struct perf_session *session)
    501 {
    502 	struct ordered_samples *os = &session->ordered_samples;
    503 
    504 	while (!list_empty(&os->to_free)) {
    505 		struct sample_queue *sq;
    506 
    507 		sq = list_entry(os->to_free.next, struct sample_queue, list);
    508 		list_del(&sq->list);
    509 		free(sq);
    510 	}
    511 }
    512 
    513 static int perf_session_deliver_event(struct perf_session *session,
    514 				      union perf_event *event,
    515 				      struct perf_sample *sample,
    516 				      struct perf_tool *tool,
    517 				      u64 file_offset);
    518 
    519 static int flush_sample_queue(struct perf_session *s,
    520 		       struct perf_tool *tool)
    521 {
    522 	struct ordered_samples *os = &s->ordered_samples;
    523 	struct list_head *head = &os->samples;
    524 	struct sample_queue *tmp, *iter;
    525 	struct perf_sample sample;
    526 	u64 limit = os->next_flush;
    527 	u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
    528 	unsigned idx = 0, progress_next = os->nr_samples / 16;
    529 	bool show_progress = limit == ULLONG_MAX;
    530 	int ret;
    531 
    532 	if (!tool->ordered_samples || !limit)
    533 		return 0;
    534 
    535 	list_for_each_entry_safe(iter, tmp, head, list) {
    536 		if (session_done())
    537 			return 0;
    538 
    539 		if (iter->timestamp > limit)
    540 			break;
    541 
    542 		ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
    543 		if (ret)
    544 			pr_err("Can't parse sample, err = %d\n", ret);
    545 		else {
    546 			ret = perf_session_deliver_event(s, iter->event, &sample, tool,
    547 							 iter->file_offset);
    548 			if (ret)
    549 				return ret;
    550 		}
    551 
    552 		os->last_flush = iter->timestamp;
    553 		list_del(&iter->list);
    554 		list_add(&iter->list, &os->sample_cache);
    555 		if (show_progress && (++idx >= progress_next)) {
    556 			progress_next += os->nr_samples / 16;
    557 			ui_progress__update(idx, os->nr_samples,
    558 					    "Processing time ordered events...");
    559 		}
    560 	}
    561 
    562 	if (list_empty(head)) {
    563 		os->last_sample = NULL;
    564 	} else if (last_ts <= limit) {
    565 		os->last_sample =
    566 			list_entry(head->prev, struct sample_queue, list);
    567 	}
    568 
    569 	os->nr_samples = 0;
    570 
    571 	return 0;
    572 }
    573 
    574 /*
    575  * When perf record finishes a pass on every buffers, it records this pseudo
    576  * event.
    577  * We record the max timestamp t found in the pass n.
    578  * Assuming these timestamps are monotonic across cpus, we know that if
    579  * a buffer still has events with timestamps below t, they will be all
    580  * available and then read in the pass n + 1.
    581  * Hence when we start to read the pass n + 2, we can safely flush every
    582  * events with timestamps below t.
    583  *
    584  *    ============ PASS n =================
    585  *       CPU 0         |   CPU 1
    586  *                     |
    587  *    cnt1 timestamps  |   cnt2 timestamps
    588  *          1          |         2
    589  *          2          |         3
    590  *          -          |         4  <--- max recorded
    591  *
    592  *    ============ PASS n + 1 ==============
    593  *       CPU 0         |   CPU 1
    594  *                     |
    595  *    cnt1 timestamps  |   cnt2 timestamps
    596  *          3          |         5
    597  *          4          |         6
    598  *          5          |         7 <---- max recorded
    599  *
    600  *      Flush every events below timestamp 4
    601  *
    602  *    ============ PASS n + 2 ==============
    603  *       CPU 0         |   CPU 1
    604  *                     |
    605  *    cnt1 timestamps  |   cnt2 timestamps
    606  *          6          |         8
    607  *          7          |         9
    608  *          -          |         10
    609  *
    610  *      Flush every events below timestamp 7
    611  *      etc...
    612  */
    613 static int process_finished_round(struct perf_tool *tool,
    614 				  union perf_event *event __maybe_unused,
    615 				  struct perf_session *session)
    616 {
    617 	int ret = flush_sample_queue(session, tool);
    618 	if (!ret)
    619 		session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
    620 
    621 	return ret;
    622 }
    623 
    624 /* The queue is ordered by time */
    625 static void __queue_event(struct sample_queue *new, struct perf_session *s)
    626 {
    627 	struct ordered_samples *os = &s->ordered_samples;
    628 	struct sample_queue *sample = os->last_sample;
    629 	u64 timestamp = new->timestamp;
    630 	struct list_head *p;
    631 
    632 	++os->nr_samples;
    633 	os->last_sample = new;
    634 
    635 	if (!sample) {
    636 		list_add(&new->list, &os->samples);
    637 		os->max_timestamp = timestamp;
    638 		return;
    639 	}
    640 
    641 	/*
    642 	 * last_sample might point to some random place in the list as it's
    643 	 * the last queued event. We expect that the new event is close to
    644 	 * this.
    645 	 */
    646 	if (sample->timestamp <= timestamp) {
    647 		while (sample->timestamp <= timestamp) {
    648 			p = sample->list.next;
    649 			if (p == &os->samples) {
    650 				list_add_tail(&new->list, &os->samples);
    651 				os->max_timestamp = timestamp;
    652 				return;
    653 			}
    654 			sample = list_entry(p, struct sample_queue, list);
    655 		}
    656 		list_add_tail(&new->list, &sample->list);
    657 	} else {
    658 		while (sample->timestamp > timestamp) {
    659 			p = sample->list.prev;
    660 			if (p == &os->samples) {
    661 				list_add(&new->list, &os->samples);
    662 				return;
    663 			}
    664 			sample = list_entry(p, struct sample_queue, list);
    665 		}
    666 		list_add(&new->list, &sample->list);
    667 	}
    668 }
    669 
    670 #define MAX_SAMPLE_BUFFER	(64 * 1024 / sizeof(struct sample_queue))
    671 
    672 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
    673 				    struct perf_sample *sample, u64 file_offset)
    674 {
    675 	struct ordered_samples *os = &s->ordered_samples;
    676 	struct list_head *sc = &os->sample_cache;
    677 	u64 timestamp = sample->time;
    678 	struct sample_queue *new;
    679 
    680 	if (!timestamp || timestamp == ~0ULL)
    681 		return -ETIME;
    682 
    683 	if (timestamp < s->ordered_samples.last_flush) {
    684 		printf("Warning: Timestamp below last timeslice flush\n");
    685 		return -EINVAL;
    686 	}
    687 
    688 	if (!list_empty(sc)) {
    689 		new = list_entry(sc->next, struct sample_queue, list);
    690 		list_del(&new->list);
    691 	} else if (os->sample_buffer) {
    692 		new = os->sample_buffer + os->sample_buffer_idx;
    693 		if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
    694 			os->sample_buffer = NULL;
    695 	} else {
    696 		os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
    697 		if (!os->sample_buffer)
    698 			return -ENOMEM;
    699 		list_add(&os->sample_buffer->list, &os->to_free);
    700 		os->sample_buffer_idx = 2;
    701 		new = os->sample_buffer + 1;
    702 	}
    703 
    704 	new->timestamp = timestamp;
    705 	new->file_offset = file_offset;
    706 	new->event = event;
    707 
    708 	__queue_event(new, s);
    709 
    710 	return 0;
    711 }
    712 
    713 static void callchain__printf(struct perf_sample *sample)
    714 {
    715 	unsigned int i;
    716 
    717 	printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
    718 
    719 	for (i = 0; i < sample->callchain->nr; i++)
    720 		printf("..... %2d: %016" PRIx64 "\n",
    721 		       i, sample->callchain->ips[i]);
    722 }
    723 
    724 static void branch_stack__printf(struct perf_sample *sample)
    725 {
    726 	uint64_t i;
    727 
    728 	printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
    729 
    730 	for (i = 0; i < sample->branch_stack->nr; i++)
    731 		printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
    732 			i, sample->branch_stack->entries[i].from,
    733 			sample->branch_stack->entries[i].to);
    734 }
    735 
    736 static void regs_dump__printf(u64 mask, u64 *regs)
    737 {
    738 	unsigned rid, i = 0;
    739 
    740 	for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
    741 		u64 val = regs[i++];
    742 
    743 		printf(".... %-5s 0x%" PRIx64 "\n",
    744 		       perf_reg_name(rid), val);
    745 	}
    746 }
    747 
    748 static void regs_user__printf(struct perf_sample *sample, u64 mask)
    749 {
    750 	struct regs_dump *user_regs = &sample->user_regs;
    751 
    752 	if (user_regs->regs) {
    753 		printf("... user regs: mask 0x%" PRIx64 "\n", mask);
    754 		regs_dump__printf(mask, user_regs->regs);
    755 	}
    756 }
    757 
    758 static void stack_user__printf(struct stack_dump *dump)
    759 {
    760 	printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
    761 	       dump->size, dump->offset);
    762 }
    763 
    764 static void perf_session__print_tstamp(struct perf_session *session,
    765 				       union perf_event *event,
    766 				       struct perf_sample *sample)
    767 {
    768 	u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
    769 
    770 	if (event->header.type != PERF_RECORD_SAMPLE &&
    771 	    !perf_evlist__sample_id_all(session->evlist)) {
    772 		fputs("-1 -1 ", stdout);
    773 		return;
    774 	}
    775 
    776 	if ((sample_type & PERF_SAMPLE_CPU))
    777 		printf("%u ", sample->cpu);
    778 
    779 	if (sample_type & PERF_SAMPLE_TIME)
    780 		printf("%" PRIu64 " ", sample->time);
    781 }
    782 
    783 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
    784 {
    785 	printf("... sample_read:\n");
    786 
    787 	if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
    788 		printf("...... time enabled %016" PRIx64 "\n",
    789 		       sample->read.time_enabled);
    790 
    791 	if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
    792 		printf("...... time running %016" PRIx64 "\n",
    793 		       sample->read.time_running);
    794 
    795 	if (read_format & PERF_FORMAT_GROUP) {
    796 		u64 i;
    797 
    798 		printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
    799 
    800 		for (i = 0; i < sample->read.group.nr; i++) {
    801 			struct sample_read_value *value;
    802 
    803 			value = &sample->read.group.values[i];
    804 			printf("..... id %016" PRIx64
    805 			       ", value %016" PRIx64 "\n",
    806 			       value->id, value->value);
    807 		}
    808 	} else
    809 		printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
    810 			sample->read.one.id, sample->read.one.value);
    811 }
    812 
    813 static void dump_event(struct perf_session *session, union perf_event *event,
    814 		       u64 file_offset, struct perf_sample *sample)
    815 {
    816 	if (!dump_trace)
    817 		return;
    818 
    819 	printf("\n%#" PRIx64 " [%#x]: event: %d\n",
    820 	       file_offset, event->header.size, event->header.type);
    821 
    822 	trace_event(event);
    823 
    824 	if (sample)
    825 		perf_session__print_tstamp(session, event, sample);
    826 
    827 	printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
    828 	       event->header.size, perf_event__name(event->header.type));
    829 }
    830 
    831 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
    832 			struct perf_sample *sample)
    833 {
    834 	u64 sample_type;
    835 
    836 	if (!dump_trace)
    837 		return;
    838 
    839 	printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
    840 	       event->header.misc, sample->pid, sample->tid, sample->ip,
    841 	       sample->period, sample->addr);
    842 
    843 	sample_type = evsel->attr.sample_type;
    844 
    845 	if (sample_type & PERF_SAMPLE_CALLCHAIN)
    846 		callchain__printf(sample);
    847 
    848 	if (sample_type & PERF_SAMPLE_BRANCH_STACK)
    849 		branch_stack__printf(sample);
    850 
    851 	if (sample_type & PERF_SAMPLE_REGS_USER)
    852 		regs_user__printf(sample, evsel->attr.sample_regs_user);
    853 
    854 	if (sample_type & PERF_SAMPLE_STACK_USER)
    855 		stack_user__printf(&sample->user_stack);
    856 
    857 	if (sample_type & PERF_SAMPLE_WEIGHT)
    858 		printf("... weight: %" PRIu64 "\n", sample->weight);
    859 
    860 	if (sample_type & PERF_SAMPLE_DATA_SRC)
    861 		printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
    862 
    863 	if (sample_type & PERF_SAMPLE_READ)
    864 		sample_read__printf(sample, evsel->attr.read_format);
    865 }
    866 
    867 static struct machine *
    868 	perf_session__find_machine_for_cpumode(struct perf_session *session,
    869 					       union perf_event *event,
    870 					       struct perf_sample *sample)
    871 {
    872 	const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
    873 
    874 	if (perf_guest &&
    875 	    ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
    876 	     (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
    877 		u32 pid;
    878 
    879 		if (event->header.type == PERF_RECORD_MMAP
    880 		    || event->header.type == PERF_RECORD_MMAP2)
    881 			pid = event->mmap.pid;
    882 		else
    883 			pid = sample->pid;
    884 
    885 		return perf_session__findnew_machine(session, pid);
    886 	}
    887 
    888 	return &session->machines.host;
    889 }
    890 
    891 static int deliver_sample_value(struct perf_session *session,
    892 				struct perf_tool *tool,
    893 				union perf_event *event,
    894 				struct perf_sample *sample,
    895 				struct sample_read_value *v,
    896 				struct machine *machine)
    897 {
    898 	struct perf_sample_id *sid;
    899 
    900 	sid = perf_evlist__id2sid(session->evlist, v->id);
    901 	if (sid) {
    902 		sample->id     = v->id;
    903 		sample->period = v->value - sid->period;
    904 		sid->period    = v->value;
    905 	}
    906 
    907 	if (!sid || sid->evsel == NULL) {
    908 		++session->stats.nr_unknown_id;
    909 		return 0;
    910 	}
    911 
    912 	return tool->sample(tool, event, sample, sid->evsel, machine);
    913 }
    914 
    915 static int deliver_sample_group(struct perf_session *session,
    916 				struct perf_tool *tool,
    917 				union  perf_event *event,
    918 				struct perf_sample *sample,
    919 				struct machine *machine)
    920 {
    921 	int ret = -EINVAL;
    922 	u64 i;
    923 
    924 	for (i = 0; i < sample->read.group.nr; i++) {
    925 		ret = deliver_sample_value(session, tool, event, sample,
    926 					   &sample->read.group.values[i],
    927 					   machine);
    928 		if (ret)
    929 			break;
    930 	}
    931 
    932 	return ret;
    933 }
    934 
    935 static int
    936 perf_session__deliver_sample(struct perf_session *session,
    937 			     struct perf_tool *tool,
    938 			     union  perf_event *event,
    939 			     struct perf_sample *sample,
    940 			     struct perf_evsel *evsel,
    941 			     struct machine *machine)
    942 {
    943 	/* We know evsel != NULL. */
    944 	u64 sample_type = evsel->attr.sample_type;
    945 	u64 read_format = evsel->attr.read_format;
    946 
    947 	/* Standard sample delievery. */
    948 	if (!(sample_type & PERF_SAMPLE_READ))
    949 		return tool->sample(tool, event, sample, evsel, machine);
    950 
    951 	/* For PERF_SAMPLE_READ we have either single or group mode. */
    952 	if (read_format & PERF_FORMAT_GROUP)
    953 		return deliver_sample_group(session, tool, event, sample,
    954 					    machine);
    955 	else
    956 		return deliver_sample_value(session, tool, event, sample,
    957 					    &sample->read.one, machine);
    958 }
    959 
    960 static int perf_session_deliver_event(struct perf_session *session,
    961 				      union perf_event *event,
    962 				      struct perf_sample *sample,
    963 				      struct perf_tool *tool,
    964 				      u64 file_offset)
    965 {
    966 	struct perf_evsel *evsel;
    967 	struct machine *machine;
    968 
    969 	dump_event(session, event, file_offset, sample);
    970 
    971 	evsel = perf_evlist__id2evsel(session->evlist, sample->id);
    972 	if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
    973 		/*
    974 		 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
    975 		 * because the tools right now may apply filters, discarding
    976 		 * some of the samples. For consistency, in the future we
    977 		 * should have something like nr_filtered_samples and remove
    978 		 * the sample->period from total_sample_period, etc, KISS for
    979 		 * now tho.
    980 		 *
    981 		 * Also testing against NULL allows us to handle files without
    982 		 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
    983 		 * future probably it'll be a good idea to restrict event
    984 		 * processing via perf_session to files with both set.
    985 		 */
    986 		hists__inc_nr_events(&evsel->hists, event->header.type);
    987 	}
    988 
    989 	machine = perf_session__find_machine_for_cpumode(session, event,
    990 							 sample);
    991 
    992 	switch (event->header.type) {
    993 	case PERF_RECORD_SAMPLE:
    994 		dump_sample(evsel, event, sample);
    995 		if (evsel == NULL) {
    996 			++session->stats.nr_unknown_id;
    997 			return 0;
    998 		}
    999 		if (machine == NULL) {
   1000 			++session->stats.nr_unprocessable_samples;
   1001 			return 0;
   1002 		}
   1003 		return perf_session__deliver_sample(session, tool, event,
   1004 						    sample, evsel, machine);
   1005 	case PERF_RECORD_MMAP:
   1006 		return tool->mmap(tool, event, sample, machine);
   1007 	case PERF_RECORD_MMAP2:
   1008 		return tool->mmap2(tool, event, sample, machine);
   1009 	case PERF_RECORD_COMM:
   1010 		return tool->comm(tool, event, sample, machine);
   1011 	case PERF_RECORD_FORK:
   1012 		return tool->fork(tool, event, sample, machine);
   1013 	case PERF_RECORD_EXIT:
   1014 		return tool->exit(tool, event, sample, machine);
   1015 	case PERF_RECORD_LOST:
   1016 		if (tool->lost == perf_event__process_lost)
   1017 			session->stats.total_lost += event->lost.lost;
   1018 		return tool->lost(tool, event, sample, machine);
   1019 	case PERF_RECORD_READ:
   1020 		return tool->read(tool, event, sample, evsel, machine);
   1021 	case PERF_RECORD_THROTTLE:
   1022 		return tool->throttle(tool, event, sample, machine);
   1023 	case PERF_RECORD_UNTHROTTLE:
   1024 		return tool->unthrottle(tool, event, sample, machine);
   1025 	default:
   1026 		++session->stats.nr_unknown_events;
   1027 		return -1;
   1028 	}
   1029 }
   1030 
   1031 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
   1032 					    struct perf_tool *tool, u64 file_offset)
   1033 {
   1034 	int err;
   1035 
   1036 	dump_event(session, event, file_offset, NULL);
   1037 
   1038 	/* These events are processed right away */
   1039 	switch (event->header.type) {
   1040 	case PERF_RECORD_HEADER_ATTR:
   1041 		err = tool->attr(tool, event, &session->evlist);
   1042 		if (err == 0)
   1043 			perf_session__set_id_hdr_size(session);
   1044 		return err;
   1045 	case PERF_RECORD_HEADER_TRACING_DATA:
   1046 		/* setup for reading amidst mmap */
   1047 		lseek(session->fd, file_offset, SEEK_SET);
   1048 		return tool->tracing_data(tool, event, session);
   1049 	case PERF_RECORD_HEADER_BUILD_ID:
   1050 		return tool->build_id(tool, event, session);
   1051 	case PERF_RECORD_FINISHED_ROUND:
   1052 		return tool->finished_round(tool, event, session);
   1053 	default:
   1054 		return -EINVAL;
   1055 	}
   1056 }
   1057 
   1058 static void event_swap(union perf_event *event, bool sample_id_all)
   1059 {
   1060 	perf_event__swap_op swap;
   1061 
   1062 	swap = perf_event__swap_ops[event->header.type];
   1063 	if (swap)
   1064 		swap(event, sample_id_all);
   1065 }
   1066 
   1067 static int perf_session__process_event(struct perf_session *session,
   1068 				       union perf_event *event,
   1069 				       struct perf_tool *tool,
   1070 				       u64 file_offset)
   1071 {
   1072 	struct perf_sample sample;
   1073 	int ret;
   1074 
   1075 	if (session->header.needs_swap)
   1076 		event_swap(event, perf_evlist__sample_id_all(session->evlist));
   1077 
   1078 	if (event->header.type >= PERF_RECORD_HEADER_MAX)
   1079 		return -EINVAL;
   1080 
   1081 	events_stats__inc(&session->stats, event->header.type);
   1082 
   1083 	if (event->header.type >= PERF_RECORD_USER_TYPE_START)
   1084 		return perf_session__process_user_event(session, event, tool, file_offset);
   1085 
   1086 	/*
   1087 	 * For all kernel events we get the sample data
   1088 	 */
   1089 	ret = perf_evlist__parse_sample(session->evlist, event, &sample);
   1090 	if (ret)
   1091 		return ret;
   1092 
   1093 	if (tool->ordered_samples) {
   1094 		ret = perf_session_queue_event(session, event, &sample,
   1095 					       file_offset);
   1096 		if (ret != -ETIME)
   1097 			return ret;
   1098 	}
   1099 
   1100 	return perf_session_deliver_event(session, event, &sample, tool,
   1101 					  file_offset);
   1102 }
   1103 
   1104 void perf_event_header__bswap(struct perf_event_header *self)
   1105 {
   1106 	self->type = bswap_32(self->type);
   1107 	self->misc = bswap_16(self->misc);
   1108 	self->size = bswap_16(self->size);
   1109 }
   1110 
   1111 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
   1112 {
   1113 	return machine__findnew_thread(&session->machines.host, 0, pid);
   1114 }
   1115 
   1116 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
   1117 {
   1118 	struct thread *thread = perf_session__findnew(self, 0);
   1119 
   1120 	if (thread == NULL || thread__set_comm(thread, "swapper")) {
   1121 		pr_err("problem inserting idle task.\n");
   1122 		thread = NULL;
   1123 	}
   1124 
   1125 	return thread;
   1126 }
   1127 
   1128 static void perf_session__warn_about_errors(const struct perf_session *session,
   1129 					    const struct perf_tool *tool)
   1130 {
   1131 	if (tool->lost == perf_event__process_lost &&
   1132 	    session->stats.nr_events[PERF_RECORD_LOST] != 0) {
   1133 		ui__warning("Processed %d events and lost %d chunks!\n\n"
   1134 			    "Check IO/CPU overload!\n\n",
   1135 			    session->stats.nr_events[0],
   1136 			    session->stats.nr_events[PERF_RECORD_LOST]);
   1137 	}
   1138 
   1139 	if (session->stats.nr_unknown_events != 0) {
   1140 		ui__warning("Found %u unknown events!\n\n"
   1141 			    "Is this an older tool processing a perf.data "
   1142 			    "file generated by a more recent tool?\n\n"
   1143 			    "If that is not the case, consider "
   1144 			    "reporting to linux-kernel (at) vger.kernel.org.\n\n",
   1145 			    session->stats.nr_unknown_events);
   1146 	}
   1147 
   1148 	if (session->stats.nr_unknown_id != 0) {
   1149 		ui__warning("%u samples with id not present in the header\n",
   1150 			    session->stats.nr_unknown_id);
   1151 	}
   1152 
   1153  	if (session->stats.nr_invalid_chains != 0) {
   1154  		ui__warning("Found invalid callchains!\n\n"
   1155  			    "%u out of %u events were discarded for this reason.\n\n"
   1156  			    "Consider reporting to linux-kernel (at) vger.kernel.org.\n\n",
   1157  			    session->stats.nr_invalid_chains,
   1158  			    session->stats.nr_events[PERF_RECORD_SAMPLE]);
   1159  	}
   1160 
   1161 	if (session->stats.nr_unprocessable_samples != 0) {
   1162 		ui__warning("%u unprocessable samples recorded.\n"
   1163 			    "Do you have a KVM guest running and not using 'perf kvm'?\n",
   1164 			    session->stats.nr_unprocessable_samples);
   1165 	}
   1166 }
   1167 
   1168 volatile int session_done;
   1169 
   1170 static int __perf_session__process_pipe_events(struct perf_session *self,
   1171 					       struct perf_tool *tool)
   1172 {
   1173 	union perf_event *event;
   1174 	uint32_t size, cur_size = 0;
   1175 	void *buf = NULL;
   1176 	int skip = 0;
   1177 	u64 head;
   1178 	int err;
   1179 	void *p;
   1180 
   1181 	perf_tool__fill_defaults(tool);
   1182 
   1183 	head = 0;
   1184 	cur_size = sizeof(union perf_event);
   1185 
   1186 	buf = malloc(cur_size);
   1187 	if (!buf)
   1188 		return -errno;
   1189 more:
   1190 	event = buf;
   1191 	err = readn(self->fd, event, sizeof(struct perf_event_header));
   1192 	if (err <= 0) {
   1193 		if (err == 0)
   1194 			goto done;
   1195 
   1196 		pr_err("failed to read event header\n");
   1197 		goto out_err;
   1198 	}
   1199 
   1200 	if (self->header.needs_swap)
   1201 		perf_event_header__bswap(&event->header);
   1202 
   1203 	size = event->header.size;
   1204 	if (size < sizeof(struct perf_event_header)) {
   1205 		pr_err("bad event header size\n");
   1206 		goto out_err;
   1207 	}
   1208 
   1209 	if (size > cur_size) {
   1210 		void *new = realloc(buf, size);
   1211 		if (!new) {
   1212 			pr_err("failed to allocate memory to read event\n");
   1213 			goto out_err;
   1214 		}
   1215 		buf = new;
   1216 		cur_size = size;
   1217 		event = buf;
   1218 	}
   1219 	p = event;
   1220 	p += sizeof(struct perf_event_header);
   1221 
   1222 	if (size - sizeof(struct perf_event_header)) {
   1223 		err = readn(self->fd, p, size - sizeof(struct perf_event_header));
   1224 		if (err <= 0) {
   1225 			if (err == 0) {
   1226 				pr_err("unexpected end of event stream\n");
   1227 				goto done;
   1228 			}
   1229 
   1230 			pr_err("failed to read event data\n");
   1231 			goto out_err;
   1232 		}
   1233 	}
   1234 
   1235 	if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
   1236 		pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
   1237 		       head, event->header.size, event->header.type);
   1238 		err = -EINVAL;
   1239 		goto out_err;
   1240 	}
   1241 
   1242 	head += size;
   1243 
   1244 	if (skip > 0)
   1245 		head += skip;
   1246 
   1247 	if (!session_done())
   1248 		goto more;
   1249 done:
   1250 	err = 0;
   1251 out_err:
   1252 	free(buf);
   1253 	perf_session__warn_about_errors(self, tool);
   1254 	perf_session_free_sample_buffers(self);
   1255 	return err;
   1256 }
   1257 
   1258 static union perf_event *
   1259 fetch_mmaped_event(struct perf_session *session,
   1260 		   u64 head, size_t mmap_size, char *buf)
   1261 {
   1262 	union perf_event *event;
   1263 
   1264 	/*
   1265 	 * Ensure we have enough space remaining to read
   1266 	 * the size of the event in the headers.
   1267 	 */
   1268 	if (head + sizeof(event->header) > mmap_size)
   1269 		return NULL;
   1270 
   1271 	event = (union perf_event *)(buf + head);
   1272 
   1273 	if (session->header.needs_swap)
   1274 		perf_event_header__bswap(&event->header);
   1275 
   1276 	if (head + event->header.size > mmap_size) {
   1277 		/* We're not fetching the event so swap back again */
   1278 		if (session->header.needs_swap)
   1279 			perf_event_header__bswap(&event->header);
   1280 		return NULL;
   1281 	}
   1282 
   1283 	return event;
   1284 }
   1285 
   1286 /*
   1287  * On 64bit we can mmap the data file in one go. No need for tiny mmap
   1288  * slices. On 32bit we use 32MB.
   1289  */
   1290 #if BITS_PER_LONG == 64
   1291 #define MMAP_SIZE ULLONG_MAX
   1292 #define NUM_MMAPS 1
   1293 #else
   1294 #define MMAP_SIZE (32 * 1024 * 1024ULL)
   1295 #define NUM_MMAPS 128
   1296 #endif
   1297 
   1298 int __perf_session__process_events(struct perf_session *session,
   1299 				   u64 data_offset, u64 data_size,
   1300 				   u64 file_size, struct perf_tool *tool)
   1301 {
   1302 	u64 head, page_offset, file_offset, file_pos, progress_next;
   1303 	int err, mmap_prot, mmap_flags, map_idx = 0;
   1304 	size_t	mmap_size;
   1305 	char *buf, *mmaps[NUM_MMAPS];
   1306 	union perf_event *event;
   1307 	uint32_t size;
   1308 
   1309 	perf_tool__fill_defaults(tool);
   1310 
   1311 	page_offset = page_size * (data_offset / page_size);
   1312 	file_offset = page_offset;
   1313 	head = data_offset - page_offset;
   1314 
   1315 	if (data_size && (data_offset + data_size < file_size))
   1316 		file_size = data_offset + data_size;
   1317 
   1318 	progress_next = file_size / 16;
   1319 
   1320 	mmap_size = MMAP_SIZE;
   1321 	if (mmap_size > file_size)
   1322 		mmap_size = file_size;
   1323 
   1324 	memset(mmaps, 0, sizeof(mmaps));
   1325 
   1326 	mmap_prot  = PROT_READ;
   1327 	mmap_flags = MAP_SHARED;
   1328 
   1329 	if (session->header.needs_swap) {
   1330 		mmap_prot  |= PROT_WRITE;
   1331 		mmap_flags = MAP_PRIVATE;
   1332 	}
   1333 remap:
   1334 	buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
   1335 		   file_offset);
   1336 	if (buf == MAP_FAILED) {
   1337 		pr_err("failed to mmap file\n");
   1338 		err = -errno;
   1339 		goto out_err;
   1340 	}
   1341 	mmaps[map_idx] = buf;
   1342 	map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
   1343 	file_pos = file_offset + head;
   1344 
   1345 more:
   1346 	event = fetch_mmaped_event(session, head, mmap_size, buf);
   1347 	if (!event) {
   1348 		if (mmaps[map_idx]) {
   1349 			munmap(mmaps[map_idx], mmap_size);
   1350 			mmaps[map_idx] = NULL;
   1351 		}
   1352 
   1353 		page_offset = page_size * (head / page_size);
   1354 		file_offset += page_offset;
   1355 		head -= page_offset;
   1356 		goto remap;
   1357 	}
   1358 
   1359 	size = event->header.size;
   1360 
   1361 	if (size < sizeof(struct perf_event_header) ||
   1362 	    perf_session__process_event(session, event, tool, file_pos) < 0) {
   1363 		pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
   1364 		       file_offset + head, event->header.size,
   1365 		       event->header.type);
   1366 		err = -EINVAL;
   1367 		goto out_err;
   1368 	}
   1369 
   1370 	head += size;
   1371 	file_pos += size;
   1372 
   1373 	if (file_pos >= progress_next) {
   1374 		progress_next += file_size / 16;
   1375 		ui_progress__update(file_pos, file_size,
   1376 				    "Processing events...");
   1377 	}
   1378 
   1379 	err = 0;
   1380 	if (session_done())
   1381 		goto out_err;
   1382 
   1383 	if (file_pos < file_size)
   1384 		goto more;
   1385 
   1386 	/* do the final flush for ordered samples */
   1387 	session->ordered_samples.next_flush = ULLONG_MAX;
   1388 	err = flush_sample_queue(session, tool);
   1389 out_err:
   1390 	ui_progress__finish();
   1391 	perf_session__warn_about_errors(session, tool);
   1392 	perf_session_free_sample_buffers(session);
   1393 	return err;
   1394 }
   1395 
   1396 int perf_session__process_events(struct perf_session *self,
   1397 				 struct perf_tool *tool)
   1398 {
   1399 	int err;
   1400 
   1401 	if (perf_session__register_idle_thread(self) == NULL)
   1402 		return -ENOMEM;
   1403 
   1404 	if (!self->fd_pipe)
   1405 		err = __perf_session__process_events(self,
   1406 						     self->header.data_offset,
   1407 						     self->header.data_size,
   1408 						     self->size, tool);
   1409 	else
   1410 		err = __perf_session__process_pipe_events(self, tool);
   1411 
   1412 	return err;
   1413 }
   1414 
   1415 bool perf_session__has_traces(struct perf_session *session, const char *msg)
   1416 {
   1417 	struct perf_evsel *evsel;
   1418 
   1419 	list_for_each_entry(evsel, &session->evlist->entries, node) {
   1420 		if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
   1421 			return true;
   1422 	}
   1423 
   1424 	pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
   1425 	return false;
   1426 }
   1427 
   1428 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
   1429 				     const char *symbol_name, u64 addr)
   1430 {
   1431 	char *bracket;
   1432 	enum map_type i;
   1433 	struct ref_reloc_sym *ref;
   1434 
   1435 	ref = zalloc(sizeof(struct ref_reloc_sym));
   1436 	if (ref == NULL)
   1437 		return -ENOMEM;
   1438 
   1439 	ref->name = strdup(symbol_name);
   1440 	if (ref->name == NULL) {
   1441 		free(ref);
   1442 		return -ENOMEM;
   1443 	}
   1444 
   1445 	bracket = strchr(ref->name, ']');
   1446 	if (bracket)
   1447 		*bracket = '\0';
   1448 
   1449 	ref->addr = addr;
   1450 
   1451 	for (i = 0; i < MAP__NR_TYPES; ++i) {
   1452 		struct kmap *kmap = map__kmap(maps[i]);
   1453 		kmap->ref_reloc_sym = ref;
   1454 	}
   1455 
   1456 	return 0;
   1457 }
   1458 
   1459 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
   1460 {
   1461 	return machines__fprintf_dsos(&self->machines, fp);
   1462 }
   1463 
   1464 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
   1465 					  bool (skip)(struct dso *dso, int parm), int parm)
   1466 {
   1467 	return machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
   1468 }
   1469 
   1470 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
   1471 {
   1472 	struct perf_evsel *pos;
   1473 	size_t ret = fprintf(fp, "Aggregated stats:\n");
   1474 
   1475 	ret += events_stats__fprintf(&session->stats, fp);
   1476 
   1477 	list_for_each_entry(pos, &session->evlist->entries, node) {
   1478 		ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
   1479 		ret += events_stats__fprintf(&pos->hists.stats, fp);
   1480 	}
   1481 
   1482 	return ret;
   1483 }
   1484 
   1485 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
   1486 {
   1487 	/*
   1488 	 * FIXME: Here we have to actually print all the machines in this
   1489 	 * session, not just the host...
   1490 	 */
   1491 	return machine__fprintf(&session->machines.host, fp);
   1492 }
   1493 
   1494 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
   1495 					      unsigned int type)
   1496 {
   1497 	struct perf_evsel *pos;
   1498 
   1499 	list_for_each_entry(pos, &session->evlist->entries, node) {
   1500 		if (pos->attr.type == type)
   1501 			return pos;
   1502 	}
   1503 	return NULL;
   1504 }
   1505 
   1506 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
   1507 			  struct perf_sample *sample, struct machine *machine,
   1508 			  unsigned int print_opts, unsigned int stack_depth)
   1509 {
   1510 	struct addr_location al;
   1511 	struct callchain_cursor_node *node;
   1512 	int print_ip = print_opts & PRINT_IP_OPT_IP;
   1513 	int print_sym = print_opts & PRINT_IP_OPT_SYM;
   1514 	int print_dso = print_opts & PRINT_IP_OPT_DSO;
   1515 	int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
   1516 	int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
   1517 	char s = print_oneline ? ' ' : '\t';
   1518 
   1519 	if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
   1520 		error("problem processing %d event, skipping it.\n",
   1521 			event->header.type);
   1522 		return;
   1523 	}
   1524 
   1525 	if (symbol_conf.use_callchain && sample->callchain) {
   1526 
   1527 		if (machine__resolve_callchain(machine, evsel, al.thread,
   1528 					       sample, NULL, NULL) != 0) {
   1529 			if (verbose)
   1530 				error("Failed to resolve callchain. Skipping\n");
   1531 			return;
   1532 		}
   1533 		callchain_cursor_commit(&callchain_cursor);
   1534 
   1535 		while (stack_depth) {
   1536 			node = callchain_cursor_current(&callchain_cursor);
   1537 			if (!node)
   1538 				break;
   1539 
   1540 			if (print_ip)
   1541 				printf("%c%16" PRIx64, s, node->ip);
   1542 
   1543 			if (print_sym) {
   1544 				printf(" ");
   1545 				if (print_symoffset) {
   1546 					al.addr = node->ip;
   1547 					al.map  = node->map;
   1548 					symbol__fprintf_symname_offs(node->sym, &al, stdout);
   1549 				} else
   1550 					symbol__fprintf_symname(node->sym, stdout);
   1551 			}
   1552 
   1553 			if (print_dso) {
   1554 				printf(" (");
   1555 				map__fprintf_dsoname(node->map, stdout);
   1556 				printf(")");
   1557 			}
   1558 
   1559 			if (!print_oneline)
   1560 				printf("\n");
   1561 
   1562 			callchain_cursor_advance(&callchain_cursor);
   1563 
   1564 			stack_depth--;
   1565 		}
   1566 
   1567 	} else {
   1568 		if (print_ip)
   1569 			printf("%16" PRIx64, sample->ip);
   1570 
   1571 		if (print_sym) {
   1572 			printf(" ");
   1573 			if (print_symoffset)
   1574 				symbol__fprintf_symname_offs(al.sym, &al,
   1575 							     stdout);
   1576 			else
   1577 				symbol__fprintf_symname(al.sym, stdout);
   1578 		}
   1579 
   1580 		if (print_dso) {
   1581 			printf(" (");
   1582 			map__fprintf_dsoname(al.map, stdout);
   1583 			printf(")");
   1584 		}
   1585 	}
   1586 }
   1587 
   1588 int perf_session__cpu_bitmap(struct perf_session *session,
   1589 			     const char *cpu_list, unsigned long *cpu_bitmap)
   1590 {
   1591 	int i;
   1592 	struct cpu_map *map;
   1593 
   1594 	for (i = 0; i < PERF_TYPE_MAX; ++i) {
   1595 		struct perf_evsel *evsel;
   1596 
   1597 		evsel = perf_session__find_first_evtype(session, i);
   1598 		if (!evsel)
   1599 			continue;
   1600 
   1601 		if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
   1602 			pr_err("File does not contain CPU events. "
   1603 			       "Remove -c option to proceed.\n");
   1604 			return -1;
   1605 		}
   1606 	}
   1607 
   1608 	map = cpu_map__new(cpu_list);
   1609 	if (map == NULL) {
   1610 		pr_err("Invalid cpu_list\n");
   1611 		return -1;
   1612 	}
   1613 
   1614 	for (i = 0; i < map->nr; i++) {
   1615 		int cpu = map->map[i];
   1616 
   1617 		if (cpu >= MAX_NR_CPUS) {
   1618 			pr_err("Requested CPU %d too large. "
   1619 			       "Consider raising MAX_NR_CPUS\n", cpu);
   1620 			return -1;
   1621 		}
   1622 
   1623 		set_bit(cpu, cpu_bitmap);
   1624 	}
   1625 
   1626 	return 0;
   1627 }
   1628 
   1629 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
   1630 				bool full)
   1631 {
   1632 	struct stat st;
   1633 	int ret;
   1634 
   1635 	if (session == NULL || fp == NULL)
   1636 		return;
   1637 
   1638 	ret = fstat(session->fd, &st);
   1639 	if (ret == -1)
   1640 		return;
   1641 
   1642 	fprintf(fp, "# ========\n");
   1643 	fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
   1644 	perf_header__fprintf_info(session, fp, full);
   1645 	fprintf(fp, "# ========\n#\n");
   1646 }
   1647 
   1648 
   1649 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
   1650 					     const struct perf_evsel_str_handler *assocs,
   1651 					     size_t nr_assocs)
   1652 {
   1653 	struct perf_evsel *evsel;
   1654 	size_t i;
   1655 	int err;
   1656 
   1657 	for (i = 0; i < nr_assocs; i++) {
   1658 		/*
   1659 		 * Adding a handler for an event not in the session,
   1660 		 * just ignore it.
   1661 		 */
   1662 		evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
   1663 		if (evsel == NULL)
   1664 			continue;
   1665 
   1666 		err = -EEXIST;
   1667 		if (evsel->handler.func != NULL)
   1668 			goto out;
   1669 		evsel->handler.func = assocs[i].handler;
   1670 	}
   1671 
   1672 	err = 0;
   1673 out:
   1674 	return err;
   1675 }
   1676