?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'powerpc-5.11-3' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc...
[linux/fpc-iii.git] / tools / perf / util / intel-pt-decoder / intel-pt-decoder.c
blob697513f3515499e545b07e0b64fe721ba8839b36
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * intel_pt_decoder.c: Intel Processor Trace support
4 * Copyright (c) 2013-2014, Intel Corporation.
5 */
6
7 #ifndef _GNU_SOURCE
8 #define _GNU_SOURCE
9 #endif
10 #include <stdlib.h>
11 #include <stdbool.h>
12 #include <string.h>
13 #include <errno.h>
14 #include <stdint.h>
15 #include <inttypes.h>
16 #include <linux/compiler.h>
17 #include <linux/string.h>
18 #include <linux/zalloc.h>
19
20 #include "../auxtrace.h"
21
22 #include "intel-pt-insn-decoder.h"
23 #include "intel-pt-pkt-decoder.h"
24 #include "intel-pt-decoder.h"
25 #include "intel-pt-log.h"
26
27 #define INTEL_PT_BLK_SIZE 1024
28
29 #define BIT63 (((uint64_t)1 << 63))
30
31 #define INTEL_PT_RETURN 1
32
33 /* Maximum number of loops with no packets consumed i.e. stuck in a loop */
34 #define INTEL_PT_MAX_LOOPS 10000
35
36 struct intel_pt_blk {
37 struct intel_pt_blk *prev;
38 uint64_t ip[INTEL_PT_BLK_SIZE];
39 };
40
41 struct intel_pt_stack {
42 struct intel_pt_blk *blk;
43 struct intel_pt_blk *spare;
44 int pos;
45 };
46
47 enum intel_pt_pkt_state {
48 INTEL_PT_STATE_NO_PSB,
49 INTEL_PT_STATE_NO_IP,
50 INTEL_PT_STATE_ERR_RESYNC,
51 INTEL_PT_STATE_IN_SYNC,
52 INTEL_PT_STATE_TNT_CONT,
53 INTEL_PT_STATE_TNT,
54 INTEL_PT_STATE_TIP,
55 INTEL_PT_STATE_TIP_PGD,
56 INTEL_PT_STATE_FUP,
57 INTEL_PT_STATE_FUP_NO_TIP,
58 INTEL_PT_STATE_RESAMPLE,
59 };
60
61 static inline bool intel_pt_sample_time(enum intel_pt_pkt_state pkt_state)
62 {
63 switch (pkt_state) {
64 case INTEL_PT_STATE_NO_PSB:
65 case INTEL_PT_STATE_NO_IP:
66 case INTEL_PT_STATE_ERR_RESYNC:
67 case INTEL_PT_STATE_IN_SYNC:
68 case INTEL_PT_STATE_TNT_CONT:
69 case INTEL_PT_STATE_RESAMPLE:
70 return true;
71 case INTEL_PT_STATE_TNT:
72 case INTEL_PT_STATE_TIP:
73 case INTEL_PT_STATE_TIP_PGD:
74 case INTEL_PT_STATE_FUP:
75 case INTEL_PT_STATE_FUP_NO_TIP:
76 return false;
77 default:
78 return true;
79 };
80 }
81
82 #ifdef INTEL_PT_STRICT
83 #define INTEL_PT_STATE_ERR1 INTEL_PT_STATE_NO_PSB
84 #define INTEL_PT_STATE_ERR2 INTEL_PT_STATE_NO_PSB
85 #define INTEL_PT_STATE_ERR3 INTEL_PT_STATE_NO_PSB
86 #define INTEL_PT_STATE_ERR4 INTEL_PT_STATE_NO_PSB
87 #else
88 #define INTEL_PT_STATE_ERR1 (decoder->pkt_state)
89 #define INTEL_PT_STATE_ERR2 INTEL_PT_STATE_NO_IP
90 #define INTEL_PT_STATE_ERR3 INTEL_PT_STATE_ERR_RESYNC
91 #define INTEL_PT_STATE_ERR4 INTEL_PT_STATE_IN_SYNC
92 #endif
93
94 struct intel_pt_decoder {
95 int (*get_trace)(struct intel_pt_buffer *buffer, void *data);
96 int (*walk_insn)(struct intel_pt_insn *intel_pt_insn,
97 uint64_t *insn_cnt_ptr, uint64_t *ip, uint64_t to_ip,
98 uint64_t max_insn_cnt, void *data);
99 bool (*pgd_ip)(uint64_t ip, void *data);
100 int (*lookahead)(void *data, intel_pt_lookahead_cb_t cb, void *cb_data);
101 void *data;
102 struct intel_pt_state state;
103 const unsigned char *buf;
104 size_t len;
105 bool return_compression;
106 bool branch_enable;
107 bool mtc_insn;
108 bool pge;
109 bool have_tma;
110 bool have_cyc;
111 bool fixup_last_mtc;
112 bool have_last_ip;
113 bool in_psb;
114 bool hop;
115 bool hop_psb_fup;
116 bool leap;
117 enum intel_pt_param_flags flags;
118 uint64_t pos;
119 uint64_t last_ip;
120 uint64_t ip;
121 uint64_t cr3;
122 uint64_t timestamp;
123 uint64_t tsc_timestamp;
124 uint64_t ref_timestamp;
125 uint64_t buf_timestamp;
126 uint64_t sample_timestamp;
127 uint64_t ret_addr;
128 uint64_t ctc_timestamp;
129 uint64_t ctc_delta;
130 uint64_t cycle_cnt;
131 uint64_t cyc_ref_timestamp;
132 uint32_t last_mtc;
133 uint32_t tsc_ctc_ratio_n;
134 uint32_t tsc_ctc_ratio_d;
135 uint32_t tsc_ctc_mult;
136 uint32_t tsc_slip;
137 uint32_t ctc_rem_mask;
138 int mtc_shift;
139 struct intel_pt_stack stack;
140 enum intel_pt_pkt_state pkt_state;
141 enum intel_pt_pkt_ctx pkt_ctx;
142 enum intel_pt_pkt_ctx prev_pkt_ctx;
143 enum intel_pt_blk_type blk_type;
144 int blk_type_pos;
145 struct intel_pt_pkt packet;
146 struct intel_pt_pkt tnt;
147 int pkt_step;
148 int pkt_len;
149 int last_packet_type;
150 unsigned int cbr;
151 unsigned int cbr_seen;
152 unsigned int max_non_turbo_ratio;
153 double max_non_turbo_ratio_fp;
154 double cbr_cyc_to_tsc;
155 double calc_cyc_to_tsc;
156 bool have_calc_cyc_to_tsc;
157 int exec_mode;
158 unsigned int insn_bytes;
159 uint64_t period;
160 enum intel_pt_period_type period_type;
161 uint64_t tot_insn_cnt;
162 uint64_t period_insn_cnt;
163 uint64_t period_mask;
164 uint64_t period_ticks;
165 uint64_t last_masked_timestamp;
166 uint64_t tot_cyc_cnt;
167 uint64_t sample_tot_cyc_cnt;
168 uint64_t base_cyc_cnt;
169 uint64_t cyc_cnt_timestamp;
170 double tsc_to_cyc;
171 bool continuous_period;
172 bool overflow;
173 bool set_fup_tx_flags;
174 bool set_fup_ptw;
175 bool set_fup_mwait;
176 bool set_fup_pwre;
177 bool set_fup_exstop;
178 bool set_fup_bep;
179 bool sample_cyc;
180 unsigned int fup_tx_flags;
181 unsigned int tx_flags;
182 uint64_t fup_ptw_payload;
183 uint64_t fup_mwait_payload;
184 uint64_t fup_pwre_payload;
185 uint64_t cbr_payload;
186 uint64_t timestamp_insn_cnt;
187 uint64_t sample_insn_cnt;
188 uint64_t stuck_ip;
189 int no_progress;
190 int stuck_ip_prd;
191 int stuck_ip_cnt;
192 const unsigned char *next_buf;
193 size_t next_len;
194 unsigned char temp_buf[INTEL_PT_PKT_MAX_SZ];
195 };
196
197 static uint64_t intel_pt_lower_power_of_2(uint64_t x)
198 {
199 int i;
200
201 for (i = 0; x != 1; i++)
202 x >>= 1;
203
204 return x << i;
205 }
206
207 static void intel_pt_setup_period(struct intel_pt_decoder *decoder)
208 {
209 if (decoder->period_type == INTEL_PT_PERIOD_TICKS) {
210 uint64_t period;
211
212 period = intel_pt_lower_power_of_2(decoder->period);
213 decoder->period_mask = ~(period - 1);
214 decoder->period_ticks = period;
215 }
216 }
217
218 static uint64_t multdiv(uint64_t t, uint32_t n, uint32_t d)
219 {
220 if (!d)
221 return 0;
222 return (t / d) * n + ((t % d) * n) / d;
223 }
224
225 struct intel_pt_decoder *intel_pt_decoder_new(struct intel_pt_params *params)
226 {
227 struct intel_pt_decoder *decoder;
228
229 if (!params->get_trace || !params->walk_insn)
230 return NULL;
231
232 decoder = zalloc(sizeof(struct intel_pt_decoder));
233 if (!decoder)
234 return NULL;
235
236 decoder->get_trace = params->get_trace;
237 decoder->walk_insn = params->walk_insn;
238 decoder->pgd_ip = params->pgd_ip;
239 decoder->lookahead = params->lookahead;
240 decoder->data = params->data;
241 decoder->return_compression = params->return_compression;
242 decoder->branch_enable = params->branch_enable;
243 decoder->hop = params->quick >= 1;
244 decoder->leap = params->quick >= 2;
245
246 decoder->flags = params->flags;
247
248 decoder->period = params->period;
249 decoder->period_type = params->period_type;
250
251 decoder->max_non_turbo_ratio = params->max_non_turbo_ratio;
252 decoder->max_non_turbo_ratio_fp = params->max_non_turbo_ratio;
253
254 intel_pt_setup_period(decoder);
255
256 decoder->mtc_shift = params->mtc_period;
257 decoder->ctc_rem_mask = (1 << decoder->mtc_shift) - 1;
258
259 decoder->tsc_ctc_ratio_n = params->tsc_ctc_ratio_n;
260 decoder->tsc_ctc_ratio_d = params->tsc_ctc_ratio_d;
261
262 if (!decoder->tsc_ctc_ratio_n)
263 decoder->tsc_ctc_ratio_d = 0;
264
265 if (decoder->tsc_ctc_ratio_d) {
266 if (!(decoder->tsc_ctc_ratio_n % decoder->tsc_ctc_ratio_d))
267 decoder->tsc_ctc_mult = decoder->tsc_ctc_ratio_n /
268 decoder->tsc_ctc_ratio_d;
269 }
270
271 /*
272 * A TSC packet can slip past MTC packets so that the timestamp appears
273 * to go backwards. One estimate is that can be up to about 40 CPU
274 * cycles, which is certainly less than 0x1000 TSC ticks, but accept
275 * slippage an order of magnitude more to be on the safe side.
276 */
277 decoder->tsc_slip = 0x10000;
278
279 intel_pt_log("timestamp: mtc_shift %u\n", decoder->mtc_shift);
280 intel_pt_log("timestamp: tsc_ctc_ratio_n %u\n", decoder->tsc_ctc_ratio_n);
281 intel_pt_log("timestamp: tsc_ctc_ratio_d %u\n", decoder->tsc_ctc_ratio_d);
282 intel_pt_log("timestamp: tsc_ctc_mult %u\n", decoder->tsc_ctc_mult);
283 intel_pt_log("timestamp: tsc_slip %#x\n", decoder->tsc_slip);
284
285 if (decoder->hop)
286 intel_pt_log("Hop mode: decoding FUP and TIPs, but not TNT\n");
287
288 return decoder;
289 }
290
291 static void intel_pt_pop_blk(struct intel_pt_stack *stack)
292 {
293 struct intel_pt_blk *blk = stack->blk;
294
295 stack->blk = blk->prev;
296 if (!stack->spare)
297 stack->spare = blk;
298 else
299 free(blk);
300 }
301
302 static uint64_t intel_pt_pop(struct intel_pt_stack *stack)
303 {
304 if (!stack->pos) {
305 if (!stack->blk)
306 return 0;
307 intel_pt_pop_blk(stack);
308 if (!stack->blk)
309 return 0;
310 stack->pos = INTEL_PT_BLK_SIZE;
311 }
312 return stack->blk->ip[--stack->pos];
313 }
314
315 static int intel_pt_alloc_blk(struct intel_pt_stack *stack)
316 {
317 struct intel_pt_blk *blk;
318
319 if (stack->spare) {
320 blk = stack->spare;
321 stack->spare = NULL;
322 } else {
323 blk = malloc(sizeof(struct intel_pt_blk));
324 if (!blk)
325 return -ENOMEM;
326 }
327
328 blk->prev = stack->blk;
329 stack->blk = blk;
330 stack->pos = 0;
331 return 0;
332 }
333
334 static int intel_pt_push(struct intel_pt_stack *stack, uint64_t ip)
335 {
336 int err;
337
338 if (!stack->blk || stack->pos == INTEL_PT_BLK_SIZE) {
339 err = intel_pt_alloc_blk(stack);
340 if (err)
341 return err;
342 }
343
344 stack->blk->ip[stack->pos++] = ip;
345 return 0;
346 }
347
348 static void intel_pt_clear_stack(struct intel_pt_stack *stack)
349 {
350 while (stack->blk)
351 intel_pt_pop_blk(stack);
352 stack->pos = 0;
353 }
354
355 static void intel_pt_free_stack(struct intel_pt_stack *stack)
356 {
357 intel_pt_clear_stack(stack);
358 zfree(&stack->blk);
359 zfree(&stack->spare);
360 }
361
362 void intel_pt_decoder_free(struct intel_pt_decoder *decoder)
363 {
364 intel_pt_free_stack(&decoder->stack);
365 free(decoder);
366 }
367
368 static int intel_pt_ext_err(int code)
369 {
370 switch (code) {
371 case -ENOMEM:
372 return INTEL_PT_ERR_NOMEM;
373 case -ENOSYS:
374 return INTEL_PT_ERR_INTERN;
375 case -EBADMSG:
376 return INTEL_PT_ERR_BADPKT;
377 case -ENODATA:
378 return INTEL_PT_ERR_NODATA;
379 case -EILSEQ:
380 return INTEL_PT_ERR_NOINSN;
381 case -ENOENT:
382 return INTEL_PT_ERR_MISMAT;
383 case -EOVERFLOW:
384 return INTEL_PT_ERR_OVR;
385 case -ENOSPC:
386 return INTEL_PT_ERR_LOST;
387 case -ELOOP:
388 return INTEL_PT_ERR_NELOOP;
389 default:
390 return INTEL_PT_ERR_UNK;
391 }
392 }
393
394 static const char *intel_pt_err_msgs[] = {
395 [INTEL_PT_ERR_NOMEM] = "Memory allocation failed",
396 [INTEL_PT_ERR_INTERN] = "Internal error",
397 [INTEL_PT_ERR_BADPKT] = "Bad packet",
398 [INTEL_PT_ERR_NODATA] = "No more data",
399 [INTEL_PT_ERR_NOINSN] = "Failed to get instruction",
400 [INTEL_PT_ERR_MISMAT] = "Trace doesn't match instruction",
401 [INTEL_PT_ERR_OVR] = "Overflow packet",
402 [INTEL_PT_ERR_LOST] = "Lost trace data",
403 [INTEL_PT_ERR_UNK] = "Unknown error!",
404 [INTEL_PT_ERR_NELOOP] = "Never-ending loop",
405 };
406
407 int intel_pt__strerror(int code, char *buf, size_t buflen)
408 {
409 if (code < 1 || code >= INTEL_PT_ERR_MAX)
410 code = INTEL_PT_ERR_UNK;
411 strlcpy(buf, intel_pt_err_msgs[code], buflen);
412 return 0;
413 }
414
415 static uint64_t intel_pt_calc_ip(const struct intel_pt_pkt *packet,
416 uint64_t last_ip)
417 {
418 uint64_t ip;
419
420 switch (packet->count) {
421 case 1:
422 ip = (last_ip & (uint64_t)0xffffffffffff0000ULL) |
423 packet->payload;
424 break;
425 case 2:
426 ip = (last_ip & (uint64_t)0xffffffff00000000ULL) |
427 packet->payload;
428 break;
429 case 3:
430 ip = packet->payload;
431 /* Sign-extend 6-byte ip */
432 if (ip & (uint64_t)0x800000000000ULL)
433 ip |= (uint64_t)0xffff000000000000ULL;
434 break;
435 case 4:
436 ip = (last_ip & (uint64_t)0xffff000000000000ULL) |
437 packet->payload;
438 break;
439 case 6:
440 ip = packet->payload;
441 break;
442 default:
443 return 0;
444 }
445
446 return ip;
447 }
448
449 static inline void intel_pt_set_last_ip(struct intel_pt_decoder *decoder)
450 {
451 decoder->last_ip = intel_pt_calc_ip(&decoder->packet, decoder->last_ip);
452 decoder->have_last_ip = true;
453 }
454
455 static inline void intel_pt_set_ip(struct intel_pt_decoder *decoder)
456 {
457 intel_pt_set_last_ip(decoder);
458 decoder->ip = decoder->last_ip;
459 }
460
461 static void intel_pt_decoder_log_packet(struct intel_pt_decoder *decoder)
462 {
463 intel_pt_log_packet(&decoder->packet, decoder->pkt_len, decoder->pos,
464 decoder->buf);
465 }
466
467 static int intel_pt_bug(struct intel_pt_decoder *decoder)
468 {
469 intel_pt_log("ERROR: Internal error\n");
470 decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
471 return -ENOSYS;
472 }
473
474 static inline void intel_pt_clear_tx_flags(struct intel_pt_decoder *decoder)
475 {
476 decoder->tx_flags = 0;
477 }
478
479 static inline void intel_pt_update_in_tx(struct intel_pt_decoder *decoder)
480 {
481 decoder->tx_flags = decoder->packet.payload & INTEL_PT_IN_TX;
482 }
483
484 static int intel_pt_bad_packet(struct intel_pt_decoder *decoder)
485 {
486 intel_pt_clear_tx_flags(decoder);
487 decoder->have_tma = false;
488 decoder->pkt_len = 1;
489 decoder->pkt_step = 1;
490 intel_pt_decoder_log_packet(decoder);
491 if (decoder->pkt_state != INTEL_PT_STATE_NO_PSB) {
492 intel_pt_log("ERROR: Bad packet\n");
493 decoder->pkt_state = INTEL_PT_STATE_ERR1;
494 }
495 return -EBADMSG;
496 }
497
498 static inline void intel_pt_update_sample_time(struct intel_pt_decoder *decoder)
499 {
500 decoder->sample_timestamp = decoder->timestamp;
501 decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
502 }
503
504 static void intel_pt_reposition(struct intel_pt_decoder *decoder)
505 {
506 decoder->ip = 0;
507 decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
508 decoder->timestamp = 0;
509 decoder->have_tma = false;
510 }
511
512 static int intel_pt_get_data(struct intel_pt_decoder *decoder, bool reposition)
513 {
514 struct intel_pt_buffer buffer = { .buf = 0, };
515 int ret;
516
517 decoder->pkt_step = 0;
518
519 intel_pt_log("Getting more data\n");
520 ret = decoder->get_trace(&buffer, decoder->data);
521 if (ret)
522 return ret;
523 decoder->buf = buffer.buf;
524 decoder->len = buffer.len;
525 if (!decoder->len) {
526 intel_pt_log("No more data\n");
527 return -ENODATA;
528 }
529 decoder->buf_timestamp = buffer.ref_timestamp;
530 if (!buffer.consecutive || reposition) {
531 intel_pt_reposition(decoder);
532 decoder->ref_timestamp = buffer.ref_timestamp;
533 decoder->state.trace_nr = buffer.trace_nr;
534 intel_pt_log("Reference timestamp 0x%" PRIx64 "\n",
535 decoder->ref_timestamp);
536 return -ENOLINK;
537 }
538
539 return 0;
540 }
541
542 static int intel_pt_get_next_data(struct intel_pt_decoder *decoder,
543 bool reposition)
544 {
545 if (!decoder->next_buf)
546 return intel_pt_get_data(decoder, reposition);
547
548 decoder->buf = decoder->next_buf;
549 decoder->len = decoder->next_len;
550 decoder->next_buf = 0;
551 decoder->next_len = 0;
552 return 0;
553 }
554
555 static int intel_pt_get_split_packet(struct intel_pt_decoder *decoder)
556 {
557 unsigned char *buf = decoder->temp_buf;
558 size_t old_len, len, n;
559 int ret;
560
561 old_len = decoder->len;
562 len = decoder->len;
563 memcpy(buf, decoder->buf, len);
564
565 ret = intel_pt_get_data(decoder, false);
566 if (ret) {
567 decoder->pos += old_len;
568 return ret < 0 ? ret : -EINVAL;
569 }
570
571 n = INTEL_PT_PKT_MAX_SZ - len;
572 if (n > decoder->len)
573 n = decoder->len;
574 memcpy(buf + len, decoder->buf, n);
575 len += n;
576
577 decoder->prev_pkt_ctx = decoder->pkt_ctx;
578 ret = intel_pt_get_packet(buf, len, &decoder->packet, &decoder->pkt_ctx);
579 if (ret < (int)old_len) {
580 decoder->next_buf = decoder->buf;
581 decoder->next_len = decoder->len;
582 decoder->buf = buf;
583 decoder->len = old_len;
584 return intel_pt_bad_packet(decoder);
585 }
586
587 decoder->next_buf = decoder->buf + (ret - old_len);
588 decoder->next_len = decoder->len - (ret - old_len);
589
590 decoder->buf = buf;
591 decoder->len = ret;
592
593 return ret;
594 }
595
596 struct intel_pt_pkt_info {
597 struct intel_pt_decoder *decoder;
598 struct intel_pt_pkt packet;
599 uint64_t pos;
600 int pkt_len;
601 int last_packet_type;
602 void *data;
603 };
604
605 typedef int (*intel_pt_pkt_cb_t)(struct intel_pt_pkt_info *pkt_info);
606
607 /* Lookahead packets in current buffer */
608 static int intel_pt_pkt_lookahead(struct intel_pt_decoder *decoder,
609 intel_pt_pkt_cb_t cb, void *data)
610 {
611 struct intel_pt_pkt_info pkt_info;
612 const unsigned char *buf = decoder->buf;
613 enum intel_pt_pkt_ctx pkt_ctx = decoder->pkt_ctx;
614 size_t len = decoder->len;
615 int ret;
616
617 pkt_info.decoder = decoder;
618 pkt_info.pos = decoder->pos;
619 pkt_info.pkt_len = decoder->pkt_step;
620 pkt_info.last_packet_type = decoder->last_packet_type;
621 pkt_info.data = data;
622
623 while (1) {
624 do {
625 pkt_info.pos += pkt_info.pkt_len;
626 buf += pkt_info.pkt_len;
627 len -= pkt_info.pkt_len;
628
629 if (!len)
630 return INTEL_PT_NEED_MORE_BYTES;
631
632 ret = intel_pt_get_packet(buf, len, &pkt_info.packet,
633 &pkt_ctx);
634 if (!ret)
635 return INTEL_PT_NEED_MORE_BYTES;
636 if (ret < 0)
637 return ret;
638
639 pkt_info.pkt_len = ret;
640 } while (pkt_info.packet.type == INTEL_PT_PAD);
641
642 ret = cb(&pkt_info);
643 if (ret)
644 return 0;
645
646 pkt_info.last_packet_type = pkt_info.packet.type;
647 }
648 }
649
650 struct intel_pt_calc_cyc_to_tsc_info {
651 uint64_t cycle_cnt;
652 unsigned int cbr;
653 uint32_t last_mtc;
654 uint64_t ctc_timestamp;
655 uint64_t ctc_delta;
656 uint64_t tsc_timestamp;
657 uint64_t timestamp;
658 bool have_tma;
659 bool fixup_last_mtc;
660 bool from_mtc;
661 double cbr_cyc_to_tsc;
662 };
663
664 /*
665 * MTC provides a 8-bit slice of CTC but the TMA packet only provides the lower
666 * 16 bits of CTC. If mtc_shift > 8 then some of the MTC bits are not in the CTC
667 * provided by the TMA packet. Fix-up the last_mtc calculated from the TMA
668 * packet by copying the missing bits from the current MTC assuming the least
669 * difference between the two, and that the current MTC comes after last_mtc.
670 */
671 static void intel_pt_fixup_last_mtc(uint32_t mtc, int mtc_shift,
672 uint32_t *last_mtc)
673 {
674 uint32_t first_missing_bit = 1U << (16 - mtc_shift);
675 uint32_t mask = ~(first_missing_bit - 1);
676
677 *last_mtc |= mtc & mask;
678 if (*last_mtc >= mtc) {
679 *last_mtc -= first_missing_bit;
680 *last_mtc &= 0xff;
681 }
682 }
683
684 static int intel_pt_calc_cyc_cb(struct intel_pt_pkt_info *pkt_info)
685 {
686 struct intel_pt_decoder *decoder = pkt_info->decoder;
687 struct intel_pt_calc_cyc_to_tsc_info *data = pkt_info->data;
688 uint64_t timestamp;
689 double cyc_to_tsc;
690 unsigned int cbr;
691 uint32_t mtc, mtc_delta, ctc, fc, ctc_rem;
692
693 switch (pkt_info->packet.type) {
694 case INTEL_PT_TNT:
695 case INTEL_PT_TIP_PGE:
696 case INTEL_PT_TIP:
697 case INTEL_PT_FUP:
698 case INTEL_PT_PSB:
699 case INTEL_PT_PIP:
700 case INTEL_PT_MODE_EXEC:
701 case INTEL_PT_MODE_TSX:
702 case INTEL_PT_PSBEND:
703 case INTEL_PT_PAD:
704 case INTEL_PT_VMCS:
705 case INTEL_PT_MNT:
706 case INTEL_PT_PTWRITE:
707 case INTEL_PT_PTWRITE_IP:
708 case INTEL_PT_BBP:
709 case INTEL_PT_BIP:
710 case INTEL_PT_BEP:
711 case INTEL_PT_BEP_IP:
712 return 0;
713
714 case INTEL_PT_MTC:
715 if (!data->have_tma)
716 return 0;
717
718 mtc = pkt_info->packet.payload;
719 if (decoder->mtc_shift > 8 && data->fixup_last_mtc) {
720 data->fixup_last_mtc = false;
721 intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
722 &data->last_mtc);
723 }
724 if (mtc > data->last_mtc)
725 mtc_delta = mtc - data->last_mtc;
726 else
727 mtc_delta = mtc + 256 - data->last_mtc;
728 data->ctc_delta += mtc_delta << decoder->mtc_shift;
729 data->last_mtc = mtc;
730
731 if (decoder->tsc_ctc_mult) {
732 timestamp = data->ctc_timestamp +
733 data->ctc_delta * decoder->tsc_ctc_mult;
734 } else {
735 timestamp = data->ctc_timestamp +
736 multdiv(data->ctc_delta,
737 decoder->tsc_ctc_ratio_n,
738 decoder->tsc_ctc_ratio_d);
739 }
740
741 if (timestamp < data->timestamp)
742 return 1;
743
744 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
745 data->timestamp = timestamp;
746 return 0;
747 }
748
749 break;
750
751 case INTEL_PT_TSC:
752 /*
753 * For now, do not support using TSC packets - refer
754 * intel_pt_calc_cyc_to_tsc().
755 */
756 if (data->from_mtc)
757 return 1;
758 timestamp = pkt_info->packet.payload |
759 (data->timestamp & (0xffULL << 56));
760 if (data->from_mtc && timestamp < data->timestamp &&
761 data->timestamp - timestamp < decoder->tsc_slip)
762 return 1;
763 if (timestamp < data->timestamp)
764 timestamp += (1ULL << 56);
765 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
766 if (data->from_mtc)
767 return 1;
768 data->tsc_timestamp = timestamp;
769 data->timestamp = timestamp;
770 return 0;
771 }
772 break;
773
774 case INTEL_PT_TMA:
775 if (data->from_mtc)
776 return 1;
777
778 if (!decoder->tsc_ctc_ratio_d)
779 return 0;
780
781 ctc = pkt_info->packet.payload;
782 fc = pkt_info->packet.count;
783 ctc_rem = ctc & decoder->ctc_rem_mask;
784
785 data->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
786
787 data->ctc_timestamp = data->tsc_timestamp - fc;
788 if (decoder->tsc_ctc_mult) {
789 data->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
790 } else {
791 data->ctc_timestamp -=
792 multdiv(ctc_rem, decoder->tsc_ctc_ratio_n,
793 decoder->tsc_ctc_ratio_d);
794 }
795
796 data->ctc_delta = 0;
797 data->have_tma = true;
798 data->fixup_last_mtc = true;
799
800 return 0;
801
802 case INTEL_PT_CYC:
803 data->cycle_cnt += pkt_info->packet.payload;
804 return 0;
805
806 case INTEL_PT_CBR:
807 cbr = pkt_info->packet.payload;
808 if (data->cbr && data->cbr != cbr)
809 return 1;
810 data->cbr = cbr;
811 data->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
812 return 0;
813
814 case INTEL_PT_TIP_PGD:
815 case INTEL_PT_TRACESTOP:
816 case INTEL_PT_EXSTOP:
817 case INTEL_PT_EXSTOP_IP:
818 case INTEL_PT_MWAIT:
819 case INTEL_PT_PWRE:
820 case INTEL_PT_PWRX:
821 case INTEL_PT_OVF:
822 case INTEL_PT_BAD: /* Does not happen */
823 default:
824 return 1;
825 }
826
827 if (!data->cbr && decoder->cbr) {
828 data->cbr = decoder->cbr;
829 data->cbr_cyc_to_tsc = decoder->cbr_cyc_to_tsc;
830 }
831
832 if (!data->cycle_cnt)
833 return 1;
834
835 cyc_to_tsc = (double)(timestamp - decoder->timestamp) / data->cycle_cnt;
836
837 if (data->cbr && cyc_to_tsc > data->cbr_cyc_to_tsc &&
838 cyc_to_tsc / data->cbr_cyc_to_tsc > 1.25) {
839 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle too big (c.f. CBR-based value %g), pos " x64_fmt "\n",
840 cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
841 return 1;
842 }
843
844 decoder->calc_cyc_to_tsc = cyc_to_tsc;
845 decoder->have_calc_cyc_to_tsc = true;
846
847 if (data->cbr) {
848 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. CBR-based value %g, pos " x64_fmt "\n",
849 cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
850 } else {
851 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. unknown CBR-based value, pos " x64_fmt "\n",
852 cyc_to_tsc, pkt_info->pos);
853 }
854
855 return 1;
856 }
857
858 static void intel_pt_calc_cyc_to_tsc(struct intel_pt_decoder *decoder,
859 bool from_mtc)
860 {
861 struct intel_pt_calc_cyc_to_tsc_info data = {
862 .cycle_cnt = 0,
863 .cbr = 0,
864 .last_mtc = decoder->last_mtc,
865 .ctc_timestamp = decoder->ctc_timestamp,
866 .ctc_delta = decoder->ctc_delta,
867 .tsc_timestamp = decoder->tsc_timestamp,
868 .timestamp = decoder->timestamp,
869 .have_tma = decoder->have_tma,
870 .fixup_last_mtc = decoder->fixup_last_mtc,
871 .from_mtc = from_mtc,
872 .cbr_cyc_to_tsc = 0,
873 };
874
875 /*
876 * For now, do not support using TSC packets for at least the reasons:
877 * 1) timing might have stopped
878 * 2) TSC packets within PSB+ can slip against CYC packets
879 */
880 if (!from_mtc)
881 return;
882
883 intel_pt_pkt_lookahead(decoder, intel_pt_calc_cyc_cb, &data);
884 }
885
886 static int intel_pt_get_next_packet(struct intel_pt_decoder *decoder)
887 {
888 int ret;
889
890 decoder->last_packet_type = decoder->packet.type;
891
892 do {
893 decoder->pos += decoder->pkt_step;
894 decoder->buf += decoder->pkt_step;
895 decoder->len -= decoder->pkt_step;
896
897 if (!decoder->len) {
898 ret = intel_pt_get_next_data(decoder, false);
899 if (ret)
900 return ret;
901 }
902
903 decoder->prev_pkt_ctx = decoder->pkt_ctx;
904 ret = intel_pt_get_packet(decoder->buf, decoder->len,
905 &decoder->packet, &decoder->pkt_ctx);
906 if (ret == INTEL_PT_NEED_MORE_BYTES && BITS_PER_LONG == 32 &&
907 decoder->len < INTEL_PT_PKT_MAX_SZ && !decoder->next_buf) {
908 ret = intel_pt_get_split_packet(decoder);
909 if (ret < 0)
910 return ret;
911 }
912 if (ret <= 0)
913 return intel_pt_bad_packet(decoder);
914
915 decoder->pkt_len = ret;
916 decoder->pkt_step = ret;
917 intel_pt_decoder_log_packet(decoder);
918 } while (decoder->packet.type == INTEL_PT_PAD);
919
920 return 0;
921 }
922
923 static uint64_t intel_pt_next_period(struct intel_pt_decoder *decoder)
924 {
925 uint64_t timestamp, masked_timestamp;
926
927 timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
928 masked_timestamp = timestamp & decoder->period_mask;
929 if (decoder->continuous_period) {
930 if (masked_timestamp > decoder->last_masked_timestamp)
931 return 1;
932 } else {
933 timestamp += 1;
934 masked_timestamp = timestamp & decoder->period_mask;
935 if (masked_timestamp > decoder->last_masked_timestamp) {
936 decoder->last_masked_timestamp = masked_timestamp;
937 decoder->continuous_period = true;
938 }
939 }
940
941 if (masked_timestamp < decoder->last_masked_timestamp)
942 return decoder->period_ticks;
943
944 return decoder->period_ticks - (timestamp - masked_timestamp);
945 }
946
947 static uint64_t intel_pt_next_sample(struct intel_pt_decoder *decoder)
948 {
949 switch (decoder->period_type) {
950 case INTEL_PT_PERIOD_INSTRUCTIONS:
951 return decoder->period - decoder->period_insn_cnt;
952 case INTEL_PT_PERIOD_TICKS:
953 return intel_pt_next_period(decoder);
954 case INTEL_PT_PERIOD_NONE:
955 case INTEL_PT_PERIOD_MTC:
956 default:
957 return 0;
958 }
959 }
960
961 static void intel_pt_sample_insn(struct intel_pt_decoder *decoder)
962 {
963 uint64_t timestamp, masked_timestamp;
964
965 switch (decoder->period_type) {
966 case INTEL_PT_PERIOD_INSTRUCTIONS:
967 decoder->period_insn_cnt = 0;
968 break;
969 case INTEL_PT_PERIOD_TICKS:
970 timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
971 masked_timestamp = timestamp & decoder->period_mask;
972 if (masked_timestamp > decoder->last_masked_timestamp)
973 decoder->last_masked_timestamp = masked_timestamp;
974 else
975 decoder->last_masked_timestamp += decoder->period_ticks;
976 break;
977 case INTEL_PT_PERIOD_NONE:
978 case INTEL_PT_PERIOD_MTC:
979 default:
980 break;
981 }
982
983 decoder->state.type |= INTEL_PT_INSTRUCTION;
984 }
985
986 static int intel_pt_walk_insn(struct intel_pt_decoder *decoder,
987 struct intel_pt_insn *intel_pt_insn, uint64_t ip)
988 {
989 uint64_t max_insn_cnt, insn_cnt = 0;
990 int err;
991
992 if (!decoder->mtc_insn)
993 decoder->mtc_insn = true;
994
995 max_insn_cnt = intel_pt_next_sample(decoder);
996
997 err = decoder->walk_insn(intel_pt_insn, &insn_cnt, &decoder->ip, ip,
998 max_insn_cnt, decoder->data);
999
1000 decoder->tot_insn_cnt += insn_cnt;
1001 decoder->timestamp_insn_cnt += insn_cnt;
1002 decoder->sample_insn_cnt += insn_cnt;
1003 decoder->period_insn_cnt += insn_cnt;
1004
1005 if (err) {
1006 decoder->no_progress = 0;
1007 decoder->pkt_state = INTEL_PT_STATE_ERR2;
1008 intel_pt_log_at("ERROR: Failed to get instruction",
1009 decoder->ip);
1010 if (err == -ENOENT)
1011 return -ENOLINK;
1012 return -EILSEQ;
1013 }
1014
1015 if (ip && decoder->ip == ip) {
1016 err = -EAGAIN;
1017 goto out;
1018 }
1019
1020 if (max_insn_cnt && insn_cnt >= max_insn_cnt)
1021 intel_pt_sample_insn(decoder);
1022
1023 if (intel_pt_insn->branch == INTEL_PT_BR_NO_BRANCH) {
1024 decoder->state.type = INTEL_PT_INSTRUCTION;
1025 decoder->state.from_ip = decoder->ip;
1026 decoder->state.to_ip = 0;
1027 decoder->ip += intel_pt_insn->length;
1028 err = INTEL_PT_RETURN;
1029 goto out;
1030 }
1031
1032 if (intel_pt_insn->op == INTEL_PT_OP_CALL) {
1033 /* Zero-length calls are excluded */
1034 if (intel_pt_insn->branch != INTEL_PT_BR_UNCONDITIONAL ||
1035 intel_pt_insn->rel) {
1036 err = intel_pt_push(&decoder->stack, decoder->ip +
1037 intel_pt_insn->length);
1038 if (err)
1039 goto out;
1040 }
1041 } else if (intel_pt_insn->op == INTEL_PT_OP_RET) {
1042 decoder->ret_addr = intel_pt_pop(&decoder->stack);
1043 }
1044
1045 if (intel_pt_insn->branch == INTEL_PT_BR_UNCONDITIONAL) {
1046 int cnt = decoder->no_progress++;
1047
1048 decoder->state.from_ip = decoder->ip;
1049 decoder->ip += intel_pt_insn->length +
1050 intel_pt_insn->rel;
1051 decoder->state.to_ip = decoder->ip;
1052 err = INTEL_PT_RETURN;
1053
1054 /*
1055 * Check for being stuck in a loop. This can happen if a
1056 * decoder error results in the decoder erroneously setting the
1057 * ip to an address that is itself in an infinite loop that
1058 * consumes no packets. When that happens, there must be an
1059 * unconditional branch.
1060 */
1061 if (cnt) {
1062 if (cnt == 1) {
1063 decoder->stuck_ip = decoder->state.to_ip;
1064 decoder->stuck_ip_prd = 1;
1065 decoder->stuck_ip_cnt = 1;
1066 } else if (cnt > INTEL_PT_MAX_LOOPS ||
1067 decoder->state.to_ip == decoder->stuck_ip) {
1068 intel_pt_log_at("ERROR: Never-ending loop",
1069 decoder->state.to_ip);
1070 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1071 err = -ELOOP;
1072 goto out;
1073 } else if (!--decoder->stuck_ip_cnt) {
1074 decoder->stuck_ip_prd += 1;
1075 decoder->stuck_ip_cnt = decoder->stuck_ip_prd;
1076 decoder->stuck_ip = decoder->state.to_ip;
1077 }
1078 }
1079 goto out_no_progress;
1080 }
1081 out:
1082 decoder->no_progress = 0;
1083 out_no_progress:
1084 decoder->state.insn_op = intel_pt_insn->op;
1085 decoder->state.insn_len = intel_pt_insn->length;
1086 memcpy(decoder->state.insn, intel_pt_insn->buf,
1087 INTEL_PT_INSN_BUF_SZ);
1088
1089 if (decoder->tx_flags & INTEL_PT_IN_TX)
1090 decoder->state.flags |= INTEL_PT_IN_TX;
1091
1092 return err;
1093 }
1094
1095 static bool intel_pt_fup_event(struct intel_pt_decoder *decoder)
1096 {
1097 bool ret = false;
1098
1099 if (decoder->set_fup_tx_flags) {
1100 decoder->set_fup_tx_flags = false;
1101 decoder->tx_flags = decoder->fup_tx_flags;
1102 decoder->state.type = INTEL_PT_TRANSACTION;
1103 decoder->state.from_ip = decoder->ip;
1104 decoder->state.to_ip = 0;
1105 decoder->state.flags = decoder->fup_tx_flags;
1106 return true;
1107 }
1108 if (decoder->set_fup_ptw) {
1109 decoder->set_fup_ptw = false;
1110 decoder->state.type = INTEL_PT_PTW;
1111 decoder->state.flags |= INTEL_PT_FUP_IP;
1112 decoder->state.from_ip = decoder->ip;
1113 decoder->state.to_ip = 0;
1114 decoder->state.ptw_payload = decoder->fup_ptw_payload;
1115 return true;
1116 }
1117 if (decoder->set_fup_mwait) {
1118 decoder->set_fup_mwait = false;
1119 decoder->state.type = INTEL_PT_MWAIT_OP;
1120 decoder->state.from_ip = decoder->ip;
1121 decoder->state.to_ip = 0;
1122 decoder->state.mwait_payload = decoder->fup_mwait_payload;
1123 ret = true;
1124 }
1125 if (decoder->set_fup_pwre) {
1126 decoder->set_fup_pwre = false;
1127 decoder->state.type |= INTEL_PT_PWR_ENTRY;
1128 decoder->state.type &= ~INTEL_PT_BRANCH;
1129 decoder->state.from_ip = decoder->ip;
1130 decoder->state.to_ip = 0;
1131 decoder->state.pwre_payload = decoder->fup_pwre_payload;
1132 ret = true;
1133 }
1134 if (decoder->set_fup_exstop) {
1135 decoder->set_fup_exstop = false;
1136 decoder->state.type |= INTEL_PT_EX_STOP;
1137 decoder->state.type &= ~INTEL_PT_BRANCH;
1138 decoder->state.flags |= INTEL_PT_FUP_IP;
1139 decoder->state.from_ip = decoder->ip;
1140 decoder->state.to_ip = 0;
1141 ret = true;
1142 }
1143 if (decoder->set_fup_bep) {
1144 decoder->set_fup_bep = false;
1145 decoder->state.type |= INTEL_PT_BLK_ITEMS;
1146 decoder->state.type &= ~INTEL_PT_BRANCH;
1147 decoder->state.from_ip = decoder->ip;
1148 decoder->state.to_ip = 0;
1149 ret = true;
1150 }
1151 return ret;
1152 }
1153
1154 static inline bool intel_pt_fup_with_nlip(struct intel_pt_decoder *decoder,
1155 struct intel_pt_insn *intel_pt_insn,
1156 uint64_t ip, int err)
1157 {
1158 return decoder->flags & INTEL_PT_FUP_WITH_NLIP && !err &&
1159 intel_pt_insn->branch == INTEL_PT_BR_INDIRECT &&
1160 ip == decoder->ip + intel_pt_insn->length;
1161 }
1162
1163 static int intel_pt_walk_fup(struct intel_pt_decoder *decoder)
1164 {
1165 struct intel_pt_insn intel_pt_insn;
1166 uint64_t ip;
1167 int err;
1168
1169 ip = decoder->last_ip;
1170
1171 while (1) {
1172 err = intel_pt_walk_insn(decoder, &intel_pt_insn, ip);
1173 if (err == INTEL_PT_RETURN)
1174 return 0;
1175 if (err == -EAGAIN ||
1176 intel_pt_fup_with_nlip(decoder, &intel_pt_insn, ip, err)) {
1177 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1178 if (intel_pt_fup_event(decoder))
1179 return 0;
1180 return -EAGAIN;
1181 }
1182 decoder->set_fup_tx_flags = false;
1183 if (err)
1184 return err;
1185
1186 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1187 intel_pt_log_at("ERROR: Unexpected indirect branch",
1188 decoder->ip);
1189 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1190 return -ENOENT;
1191 }
1192
1193 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1194 intel_pt_log_at("ERROR: Unexpected conditional branch",
1195 decoder->ip);
1196 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1197 return -ENOENT;
1198 }
1199
1200 intel_pt_bug(decoder);
1201 }
1202 }
1203
1204 static int intel_pt_walk_tip(struct intel_pt_decoder *decoder)
1205 {
1206 struct intel_pt_insn intel_pt_insn;
1207 int err;
1208
1209 err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1210 if (err == INTEL_PT_RETURN &&
1211 decoder->pgd_ip &&
1212 decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1213 (decoder->state.type & INTEL_PT_BRANCH) &&
1214 decoder->pgd_ip(decoder->state.to_ip, decoder->data)) {
1215 /* Unconditional branch leaving filter region */
1216 decoder->no_progress = 0;
1217 decoder->pge = false;
1218 decoder->continuous_period = false;
1219 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1220 decoder->state.type |= INTEL_PT_TRACE_END;
1221 return 0;
1222 }
1223 if (err == INTEL_PT_RETURN)
1224 return 0;
1225 if (err)
1226 return err;
1227
1228 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1229 if (decoder->pkt_state == INTEL_PT_STATE_TIP_PGD) {
1230 decoder->pge = false;
1231 decoder->continuous_period = false;
1232 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1233 decoder->state.from_ip = decoder->ip;
1234 if (decoder->packet.count == 0) {
1235 decoder->state.to_ip = 0;
1236 } else {
1237 decoder->state.to_ip = decoder->last_ip;
1238 decoder->ip = decoder->last_ip;
1239 }
1240 decoder->state.type |= INTEL_PT_TRACE_END;
1241 } else {
1242 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1243 decoder->state.from_ip = decoder->ip;
1244 if (decoder->packet.count == 0) {
1245 decoder->state.to_ip = 0;
1246 } else {
1247 decoder->state.to_ip = decoder->last_ip;
1248 decoder->ip = decoder->last_ip;
1249 }
1250 }
1251 return 0;
1252 }
1253
1254 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1255 uint64_t to_ip = decoder->ip + intel_pt_insn.length +
1256 intel_pt_insn.rel;
1257
1258 if (decoder->pgd_ip &&
1259 decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1260 decoder->pgd_ip(to_ip, decoder->data)) {
1261 /* Conditional branch leaving filter region */
1262 decoder->pge = false;
1263 decoder->continuous_period = false;
1264 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1265 decoder->ip = to_ip;
1266 decoder->state.from_ip = decoder->ip;
1267 decoder->state.to_ip = to_ip;
1268 decoder->state.type |= INTEL_PT_TRACE_END;
1269 return 0;
1270 }
1271 intel_pt_log_at("ERROR: Conditional branch when expecting indirect branch",
1272 decoder->ip);
1273 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1274 return -ENOENT;
1275 }
1276
1277 return intel_pt_bug(decoder);
1278 }
1279
1280 static int intel_pt_walk_tnt(struct intel_pt_decoder *decoder)
1281 {
1282 struct intel_pt_insn intel_pt_insn;
1283 int err;
1284
1285 while (1) {
1286 err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1287 if (err == INTEL_PT_RETURN)
1288 return 0;
1289 if (err)
1290 return err;
1291
1292 if (intel_pt_insn.op == INTEL_PT_OP_RET) {
1293 if (!decoder->return_compression) {
1294 intel_pt_log_at("ERROR: RET when expecting conditional branch",
1295 decoder->ip);
1296 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1297 return -ENOENT;
1298 }
1299 if (!decoder->ret_addr) {
1300 intel_pt_log_at("ERROR: Bad RET compression (stack empty)",
1301 decoder->ip);
1302 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1303 return -ENOENT;
1304 }
1305 if (!(decoder->tnt.payload & BIT63)) {
1306 intel_pt_log_at("ERROR: Bad RET compression (TNT=N)",
1307 decoder->ip);
1308 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1309 return -ENOENT;
1310 }
1311 decoder->tnt.count -= 1;
1312 if (decoder->tnt.count)
1313 decoder->pkt_state = INTEL_PT_STATE_TNT_CONT;
1314 else
1315 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1316 decoder->tnt.payload <<= 1;
1317 decoder->state.from_ip = decoder->ip;
1318 decoder->ip = decoder->ret_addr;
1319 decoder->state.to_ip = decoder->ip;
1320 return 0;
1321 }
1322
1323 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1324 /* Handle deferred TIPs */
1325 err = intel_pt_get_next_packet(decoder);
1326 if (err)
1327 return err;
1328 if (decoder->packet.type != INTEL_PT_TIP ||
1329 decoder->packet.count == 0) {
1330 intel_pt_log_at("ERROR: Missing deferred TIP for indirect branch",
1331 decoder->ip);
1332 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1333 decoder->pkt_step = 0;
1334 return -ENOENT;
1335 }
1336 intel_pt_set_last_ip(decoder);
1337 decoder->state.from_ip = decoder->ip;
1338 decoder->state.to_ip = decoder->last_ip;
1339 decoder->ip = decoder->last_ip;
1340 return 0;
1341 }
1342
1343 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1344 decoder->tnt.count -= 1;
1345 if (decoder->tnt.count)
1346 decoder->pkt_state = INTEL_PT_STATE_TNT_CONT;
1347 else
1348 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1349 if (decoder->tnt.payload & BIT63) {
1350 decoder->tnt.payload <<= 1;
1351 decoder->state.from_ip = decoder->ip;
1352 decoder->ip += intel_pt_insn.length +
1353 intel_pt_insn.rel;
1354 decoder->state.to_ip = decoder->ip;
1355 return 0;
1356 }
1357 /* Instruction sample for a non-taken branch */
1358 if (decoder->state.type & INTEL_PT_INSTRUCTION) {
1359 decoder->tnt.payload <<= 1;
1360 decoder->state.type = INTEL_PT_INSTRUCTION;
1361 decoder->state.from_ip = decoder->ip;
1362 decoder->state.to_ip = 0;
1363 decoder->ip += intel_pt_insn.length;
1364 return 0;
1365 }
1366 decoder->sample_cyc = false;
1367 decoder->ip += intel_pt_insn.length;
1368 if (!decoder->tnt.count) {
1369 intel_pt_update_sample_time(decoder);
1370 return -EAGAIN;
1371 }
1372 decoder->tnt.payload <<= 1;
1373 continue;
1374 }
1375
1376 return intel_pt_bug(decoder);
1377 }
1378 }
1379
1380 static int intel_pt_mode_tsx(struct intel_pt_decoder *decoder, bool *no_tip)
1381 {
1382 unsigned int fup_tx_flags;
1383 int err;
1384
1385 fup_tx_flags = decoder->packet.payload &
1386 (INTEL_PT_IN_TX | INTEL_PT_ABORT_TX);
1387 err = intel_pt_get_next_packet(decoder);
1388 if (err)
1389 return err;
1390 if (decoder->packet.type == INTEL_PT_FUP) {
1391 decoder->fup_tx_flags = fup_tx_flags;
1392 decoder->set_fup_tx_flags = true;
1393 if (!(decoder->fup_tx_flags & INTEL_PT_ABORT_TX))
1394 *no_tip = true;
1395 } else {
1396 intel_pt_log_at("ERROR: Missing FUP after MODE.TSX",
1397 decoder->pos);
1398 intel_pt_update_in_tx(decoder);
1399 }
1400 return 0;
1401 }
1402
1403 static uint64_t intel_pt_8b_tsc(uint64_t timestamp, uint64_t ref_timestamp)
1404 {
1405 timestamp |= (ref_timestamp & (0xffULL << 56));
1406
1407 if (timestamp < ref_timestamp) {
1408 if (ref_timestamp - timestamp > (1ULL << 55))
1409 timestamp += (1ULL << 56);
1410 } else {
1411 if (timestamp - ref_timestamp > (1ULL << 55))
1412 timestamp -= (1ULL << 56);
1413 }
1414
1415 return timestamp;
1416 }
1417
1418 static void intel_pt_calc_tsc_timestamp(struct intel_pt_decoder *decoder)
1419 {
1420 uint64_t timestamp;
1421
1422 decoder->have_tma = false;
1423
1424 if (decoder->ref_timestamp) {
1425 timestamp = intel_pt_8b_tsc(decoder->packet.payload,
1426 decoder->ref_timestamp);
1427 decoder->tsc_timestamp = timestamp;
1428 decoder->timestamp = timestamp;
1429 decoder->ref_timestamp = 0;
1430 decoder->timestamp_insn_cnt = 0;
1431 } else if (decoder->timestamp) {
1432 timestamp = decoder->packet.payload |
1433 (decoder->timestamp & (0xffULL << 56));
1434 decoder->tsc_timestamp = timestamp;
1435 if (timestamp < decoder->timestamp &&
1436 decoder->timestamp - timestamp < decoder->tsc_slip) {
1437 intel_pt_log_to("Suppressing backwards timestamp",
1438 timestamp);
1439 timestamp = decoder->timestamp;
1440 }
1441 if (timestamp < decoder->timestamp) {
1442 intel_pt_log_to("Wraparound timestamp", timestamp);
1443 timestamp += (1ULL << 56);
1444 decoder->tsc_timestamp = timestamp;
1445 }
1446 decoder->timestamp = timestamp;
1447 decoder->timestamp_insn_cnt = 0;
1448 }
1449
1450 if (decoder->last_packet_type == INTEL_PT_CYC) {
1451 decoder->cyc_ref_timestamp = decoder->timestamp;
1452 decoder->cycle_cnt = 0;
1453 decoder->have_calc_cyc_to_tsc = false;
1454 intel_pt_calc_cyc_to_tsc(decoder, false);
1455 }
1456
1457 intel_pt_log_to("Setting timestamp", decoder->timestamp);
1458 }
1459
1460 static int intel_pt_overflow(struct intel_pt_decoder *decoder)
1461 {
1462 intel_pt_log("ERROR: Buffer overflow\n");
1463 intel_pt_clear_tx_flags(decoder);
1464 decoder->timestamp_insn_cnt = 0;
1465 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1466 decoder->overflow = true;
1467 return -EOVERFLOW;
1468 }
1469
1470 static inline void intel_pt_mtc_cyc_cnt_pge(struct intel_pt_decoder *decoder)
1471 {
1472 if (decoder->have_cyc)
1473 return;
1474
1475 decoder->cyc_cnt_timestamp = decoder->timestamp;
1476 decoder->base_cyc_cnt = decoder->tot_cyc_cnt;
1477 }
1478
1479 static inline void intel_pt_mtc_cyc_cnt_cbr(struct intel_pt_decoder *decoder)
1480 {
1481 decoder->tsc_to_cyc = decoder->cbr / decoder->max_non_turbo_ratio_fp;
1482
1483 if (decoder->pge)
1484 intel_pt_mtc_cyc_cnt_pge(decoder);
1485 }
1486
1487 static inline void intel_pt_mtc_cyc_cnt_upd(struct intel_pt_decoder *decoder)
1488 {
1489 uint64_t tot_cyc_cnt, tsc_delta;
1490
1491 if (decoder->have_cyc)
1492 return;
1493
1494 decoder->sample_cyc = true;
1495
1496 if (!decoder->pge || decoder->timestamp <= decoder->cyc_cnt_timestamp)
1497 return;
1498
1499 tsc_delta = decoder->timestamp - decoder->cyc_cnt_timestamp;
1500 tot_cyc_cnt = tsc_delta * decoder->tsc_to_cyc + decoder->base_cyc_cnt;
1501
1502 if (tot_cyc_cnt > decoder->tot_cyc_cnt)
1503 decoder->tot_cyc_cnt = tot_cyc_cnt;
1504 }
1505
1506 static void intel_pt_calc_tma(struct intel_pt_decoder *decoder)
1507 {
1508 uint32_t ctc = decoder->packet.payload;
1509 uint32_t fc = decoder->packet.count;
1510 uint32_t ctc_rem = ctc & decoder->ctc_rem_mask;
1511
1512 if (!decoder->tsc_ctc_ratio_d)
1513 return;
1514
1515 if (decoder->pge && !decoder->in_psb)
1516 intel_pt_mtc_cyc_cnt_pge(decoder);
1517 else
1518 intel_pt_mtc_cyc_cnt_upd(decoder);
1519
1520 decoder->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
1521 decoder->ctc_timestamp = decoder->tsc_timestamp - fc;
1522 if (decoder->tsc_ctc_mult) {
1523 decoder->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
1524 } else {
1525 decoder->ctc_timestamp -= multdiv(ctc_rem,
1526 decoder->tsc_ctc_ratio_n,
1527 decoder->tsc_ctc_ratio_d);
1528 }
1529 decoder->ctc_delta = 0;
1530 decoder->have_tma = true;
1531 decoder->fixup_last_mtc = true;
1532 intel_pt_log("CTC timestamp " x64_fmt " last MTC %#x CTC rem %#x\n",
1533 decoder->ctc_timestamp, decoder->last_mtc, ctc_rem);
1534 }
1535
1536 static void intel_pt_calc_mtc_timestamp(struct intel_pt_decoder *decoder)
1537 {
1538 uint64_t timestamp;
1539 uint32_t mtc, mtc_delta;
1540
1541 if (!decoder->have_tma)
1542 return;
1543
1544 mtc = decoder->packet.payload;
1545
1546 if (decoder->mtc_shift > 8 && decoder->fixup_last_mtc) {
1547 decoder->fixup_last_mtc = false;
1548 intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
1549 &decoder->last_mtc);
1550 }
1551
1552 if (mtc > decoder->last_mtc)
1553 mtc_delta = mtc - decoder->last_mtc;
1554 else
1555 mtc_delta = mtc + 256 - decoder->last_mtc;
1556
1557 decoder->ctc_delta += mtc_delta << decoder->mtc_shift;
1558
1559 if (decoder->tsc_ctc_mult) {
1560 timestamp = decoder->ctc_timestamp +
1561 decoder->ctc_delta * decoder->tsc_ctc_mult;
1562 } else {
1563 timestamp = decoder->ctc_timestamp +
1564 multdiv(decoder->ctc_delta,
1565 decoder->tsc_ctc_ratio_n,
1566 decoder->tsc_ctc_ratio_d);
1567 }
1568
1569 if (timestamp < decoder->timestamp)
1570 intel_pt_log("Suppressing MTC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1571 timestamp, decoder->timestamp);
1572 else
1573 decoder->timestamp = timestamp;
1574
1575 intel_pt_mtc_cyc_cnt_upd(decoder);
1576
1577 decoder->timestamp_insn_cnt = 0;
1578 decoder->last_mtc = mtc;
1579
1580 if (decoder->last_packet_type == INTEL_PT_CYC) {
1581 decoder->cyc_ref_timestamp = decoder->timestamp;
1582 decoder->cycle_cnt = 0;
1583 decoder->have_calc_cyc_to_tsc = false;
1584 intel_pt_calc_cyc_to_tsc(decoder, true);
1585 }
1586
1587 intel_pt_log_to("Setting timestamp", decoder->timestamp);
1588 }
1589
1590 static void intel_pt_calc_cbr(struct intel_pt_decoder *decoder)
1591 {
1592 unsigned int cbr = decoder->packet.payload & 0xff;
1593
1594 decoder->cbr_payload = decoder->packet.payload;
1595
1596 if (decoder->cbr == cbr)
1597 return;
1598
1599 decoder->cbr = cbr;
1600 decoder->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
1601
1602 intel_pt_mtc_cyc_cnt_cbr(decoder);
1603 }
1604
1605 static void intel_pt_calc_cyc_timestamp(struct intel_pt_decoder *decoder)
1606 {
1607 uint64_t timestamp = decoder->cyc_ref_timestamp;
1608
1609 decoder->have_cyc = true;
1610
1611 decoder->cycle_cnt += decoder->packet.payload;
1612 if (decoder->pge)
1613 decoder->tot_cyc_cnt += decoder->packet.payload;
1614 decoder->sample_cyc = true;
1615
1616 if (!decoder->cyc_ref_timestamp)
1617 return;
1618
1619 if (decoder->have_calc_cyc_to_tsc)
1620 timestamp += decoder->cycle_cnt * decoder->calc_cyc_to_tsc;
1621 else if (decoder->cbr)
1622 timestamp += decoder->cycle_cnt * decoder->cbr_cyc_to_tsc;
1623 else
1624 return;
1625
1626 if (timestamp < decoder->timestamp)
1627 intel_pt_log("Suppressing CYC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1628 timestamp, decoder->timestamp);
1629 else
1630 decoder->timestamp = timestamp;
1631
1632 decoder->timestamp_insn_cnt = 0;
1633
1634 intel_pt_log_to("Setting timestamp", decoder->timestamp);
1635 }
1636
1637 static void intel_pt_bbp(struct intel_pt_decoder *decoder)
1638 {
1639 if (decoder->prev_pkt_ctx == INTEL_PT_NO_CTX) {
1640 memset(decoder->state.items.mask, 0, sizeof(decoder->state.items.mask));
1641 decoder->state.items.is_32_bit = false;
1642 }
1643 decoder->blk_type = decoder->packet.payload;
1644 decoder->blk_type_pos = intel_pt_blk_type_pos(decoder->blk_type);
1645 if (decoder->blk_type == INTEL_PT_GP_REGS)
1646 decoder->state.items.is_32_bit = decoder->packet.count;
1647 if (decoder->blk_type_pos < 0) {
1648 intel_pt_log("WARNING: Unknown block type %u\n",
1649 decoder->blk_type);
1650 } else if (decoder->state.items.mask[decoder->blk_type_pos]) {
1651 intel_pt_log("WARNING: Duplicate block type %u\n",
1652 decoder->blk_type);
1653 }
1654 }
1655
1656 static void intel_pt_bip(struct intel_pt_decoder *decoder)
1657 {
1658 uint32_t id = decoder->packet.count;
1659 uint32_t bit = 1 << id;
1660 int pos = decoder->blk_type_pos;
1661
1662 if (pos < 0 || id >= INTEL_PT_BLK_ITEM_ID_CNT) {
1663 intel_pt_log("WARNING: Unknown block item %u type %d\n",
1664 id, decoder->blk_type);
1665 return;
1666 }
1667
1668 if (decoder->state.items.mask[pos] & bit) {
1669 intel_pt_log("WARNING: Duplicate block item %u type %d\n",
1670 id, decoder->blk_type);
1671 }
1672
1673 decoder->state.items.mask[pos] |= bit;
1674 decoder->state.items.val[pos][id] = decoder->packet.payload;
1675 }
1676
1677 /* Walk PSB+ packets when already in sync. */
1678 static int intel_pt_walk_psbend(struct intel_pt_decoder *decoder)
1679 {
1680 int err;
1681
1682 decoder->in_psb = true;
1683
1684 while (1) {
1685 err = intel_pt_get_next_packet(decoder);
1686 if (err)
1687 goto out;
1688
1689 switch (decoder->packet.type) {
1690 case INTEL_PT_PSBEND:
1691 err = 0;
1692 goto out;
1693
1694 case INTEL_PT_TIP_PGD:
1695 case INTEL_PT_TIP_PGE:
1696 case INTEL_PT_TIP:
1697 case INTEL_PT_TNT:
1698 case INTEL_PT_TRACESTOP:
1699 case INTEL_PT_BAD:
1700 case INTEL_PT_PSB:
1701 case INTEL_PT_PTWRITE:
1702 case INTEL_PT_PTWRITE_IP:
1703 case INTEL_PT_EXSTOP:
1704 case INTEL_PT_EXSTOP_IP:
1705 case INTEL_PT_MWAIT:
1706 case INTEL_PT_PWRE:
1707 case INTEL_PT_PWRX:
1708 case INTEL_PT_BBP:
1709 case INTEL_PT_BIP:
1710 case INTEL_PT_BEP:
1711 case INTEL_PT_BEP_IP:
1712 decoder->have_tma = false;
1713 intel_pt_log("ERROR: Unexpected packet\n");
1714 err = -EAGAIN;
1715 goto out;
1716
1717 case INTEL_PT_OVF:
1718 err = intel_pt_overflow(decoder);
1719 goto out;
1720
1721 case INTEL_PT_TSC:
1722 intel_pt_calc_tsc_timestamp(decoder);
1723 break;
1724
1725 case INTEL_PT_TMA:
1726 intel_pt_calc_tma(decoder);
1727 break;
1728
1729 case INTEL_PT_CBR:
1730 intel_pt_calc_cbr(decoder);
1731 break;
1732
1733 case INTEL_PT_MODE_EXEC:
1734 decoder->exec_mode = decoder->packet.payload;
1735 break;
1736
1737 case INTEL_PT_PIP:
1738 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1739 break;
1740
1741 case INTEL_PT_FUP:
1742 decoder->pge = true;
1743 if (decoder->packet.count) {
1744 intel_pt_set_last_ip(decoder);
1745 if (decoder->hop) {
1746 /* Act on FUP at PSBEND */
1747 decoder->ip = decoder->last_ip;
1748 decoder->hop_psb_fup = true;
1749 }
1750 }
1751 break;
1752
1753 case INTEL_PT_MODE_TSX:
1754 intel_pt_update_in_tx(decoder);
1755 break;
1756
1757 case INTEL_PT_MTC:
1758 intel_pt_calc_mtc_timestamp(decoder);
1759 if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1760 decoder->state.type |= INTEL_PT_INSTRUCTION;
1761 break;
1762
1763 case INTEL_PT_CYC:
1764 case INTEL_PT_VMCS:
1765 case INTEL_PT_MNT:
1766 case INTEL_PT_PAD:
1767 default:
1768 break;
1769 }
1770 }
1771 out:
1772 decoder->in_psb = false;
1773
1774 return err;
1775 }
1776
1777 static int intel_pt_walk_fup_tip(struct intel_pt_decoder *decoder)
1778 {
1779 int err;
1780
1781 if (decoder->tx_flags & INTEL_PT_ABORT_TX) {
1782 decoder->tx_flags = 0;
1783 decoder->state.flags &= ~INTEL_PT_IN_TX;
1784 decoder->state.flags |= INTEL_PT_ABORT_TX;
1785 } else {
1786 decoder->state.flags |= INTEL_PT_ASYNC;
1787 }
1788
1789 while (1) {
1790 err = intel_pt_get_next_packet(decoder);
1791 if (err)
1792 return err;
1793
1794 switch (decoder->packet.type) {
1795 case INTEL_PT_TNT:
1796 case INTEL_PT_FUP:
1797 case INTEL_PT_TRACESTOP:
1798 case INTEL_PT_PSB:
1799 case INTEL_PT_TSC:
1800 case INTEL_PT_TMA:
1801 case INTEL_PT_MODE_TSX:
1802 case INTEL_PT_BAD:
1803 case INTEL_PT_PSBEND:
1804 case INTEL_PT_PTWRITE:
1805 case INTEL_PT_PTWRITE_IP:
1806 case INTEL_PT_EXSTOP:
1807 case INTEL_PT_EXSTOP_IP:
1808 case INTEL_PT_MWAIT:
1809 case INTEL_PT_PWRE:
1810 case INTEL_PT_PWRX:
1811 case INTEL_PT_BBP:
1812 case INTEL_PT_BIP:
1813 case INTEL_PT_BEP:
1814 case INTEL_PT_BEP_IP:
1815 intel_pt_log("ERROR: Missing TIP after FUP\n");
1816 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1817 decoder->pkt_step = 0;
1818 return -ENOENT;
1819
1820 case INTEL_PT_CBR:
1821 intel_pt_calc_cbr(decoder);
1822 break;
1823
1824 case INTEL_PT_OVF:
1825 return intel_pt_overflow(decoder);
1826
1827 case INTEL_PT_TIP_PGD:
1828 decoder->state.from_ip = decoder->ip;
1829 if (decoder->packet.count == 0) {
1830 decoder->state.to_ip = 0;
1831 } else {
1832 intel_pt_set_ip(decoder);
1833 decoder->state.to_ip = decoder->ip;
1834 }
1835 decoder->pge = false;
1836 decoder->continuous_period = false;
1837 decoder->state.type |= INTEL_PT_TRACE_END;
1838 return 0;
1839
1840 case INTEL_PT_TIP_PGE:
1841 decoder->pge = true;
1842 intel_pt_log("Omitting PGE ip " x64_fmt "\n",
1843 decoder->ip);
1844 decoder->state.from_ip = 0;
1845 if (decoder->packet.count == 0) {
1846 decoder->state.to_ip = 0;
1847 } else {
1848 intel_pt_set_ip(decoder);
1849 decoder->state.to_ip = decoder->ip;
1850 }
1851 decoder->state.type |= INTEL_PT_TRACE_BEGIN;
1852 intel_pt_mtc_cyc_cnt_pge(decoder);
1853 return 0;
1854
1855 case INTEL_PT_TIP:
1856 decoder->state.from_ip = decoder->ip;
1857 if (decoder->packet.count == 0) {
1858 decoder->state.to_ip = 0;
1859 } else {
1860 intel_pt_set_ip(decoder);
1861 decoder->state.to_ip = decoder->ip;
1862 }
1863 return 0;
1864
1865 case INTEL_PT_PIP:
1866 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1867 break;
1868
1869 case INTEL_PT_MTC:
1870 intel_pt_calc_mtc_timestamp(decoder);
1871 if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1872 decoder->state.type |= INTEL_PT_INSTRUCTION;
1873 break;
1874
1875 case INTEL_PT_CYC:
1876 intel_pt_calc_cyc_timestamp(decoder);
1877 break;
1878
1879 case INTEL_PT_MODE_EXEC:
1880 decoder->exec_mode = decoder->packet.payload;
1881 break;
1882
1883 case INTEL_PT_VMCS:
1884 case INTEL_PT_MNT:
1885 case INTEL_PT_PAD:
1886 break;
1887
1888 default:
1889 return intel_pt_bug(decoder);
1890 }
1891 }
1892 }
1893
1894 static int intel_pt_resample(struct intel_pt_decoder *decoder)
1895 {
1896 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1897 decoder->state.type = INTEL_PT_INSTRUCTION;
1898 decoder->state.from_ip = decoder->ip;
1899 decoder->state.to_ip = 0;
1900 return 0;
1901 }
1902
1903 #define HOP_PROCESS 0
1904 #define HOP_IGNORE 1
1905 #define HOP_RETURN 2
1906 #define HOP_AGAIN 3
1907
1908 static int intel_pt_scan_for_psb(struct intel_pt_decoder *decoder);
1909
1910 /* Hop mode: Ignore TNT, do not walk code, but get ip from FUPs and TIPs */
1911 static int intel_pt_hop_trace(struct intel_pt_decoder *decoder, bool *no_tip, int *err)
1912 {
1913 /* Leap from PSB to PSB, getting ip from FUP within PSB+ */
1914 if (decoder->leap && !decoder->in_psb && decoder->packet.type != INTEL_PT_PSB) {
1915 *err = intel_pt_scan_for_psb(decoder);
1916 if (*err)
1917 return HOP_RETURN;
1918 }
1919
1920 switch (decoder->packet.type) {
1921 case INTEL_PT_TNT:
1922 return HOP_IGNORE;
1923
1924 case INTEL_PT_TIP_PGD:
1925 if (!decoder->packet.count)
1926 return HOP_IGNORE;
1927 intel_pt_set_ip(decoder);
1928 decoder->state.type |= INTEL_PT_TRACE_END;
1929 decoder->state.from_ip = 0;
1930 decoder->state.to_ip = decoder->ip;
1931 return HOP_RETURN;
1932
1933 case INTEL_PT_TIP:
1934 if (!decoder->packet.count)
1935 return HOP_IGNORE;
1936 intel_pt_set_ip(decoder);
1937 decoder->state.type = INTEL_PT_INSTRUCTION;
1938 decoder->state.from_ip = decoder->ip;
1939 decoder->state.to_ip = 0;
1940 return HOP_RETURN;
1941
1942 case INTEL_PT_FUP:
1943 if (!decoder->packet.count)
1944 return HOP_IGNORE;
1945 intel_pt_set_ip(decoder);
1946 if (intel_pt_fup_event(decoder))
1947 return HOP_RETURN;
1948 if (!decoder->branch_enable)
1949 *no_tip = true;
1950 if (*no_tip) {
1951 decoder->state.type = INTEL_PT_INSTRUCTION;
1952 decoder->state.from_ip = decoder->ip;
1953 decoder->state.to_ip = 0;
1954 return HOP_RETURN;
1955 }
1956 *err = intel_pt_walk_fup_tip(decoder);
1957 if (!*err)
1958 decoder->pkt_state = INTEL_PT_STATE_RESAMPLE;
1959 return HOP_RETURN;
1960
1961 case INTEL_PT_PSB:
1962 decoder->last_ip = 0;
1963 decoder->have_last_ip = true;
1964 decoder->hop_psb_fup = false;
1965 *err = intel_pt_walk_psbend(decoder);
1966 if (*err == -EAGAIN)
1967 return HOP_AGAIN;
1968 if (*err)
1969 return HOP_RETURN;
1970 if (decoder->hop_psb_fup) {
1971 decoder->hop_psb_fup = false;
1972 decoder->state.type = INTEL_PT_INSTRUCTION;
1973 decoder->state.from_ip = decoder->ip;
1974 decoder->state.to_ip = 0;
1975 return HOP_RETURN;
1976 }
1977 if (decoder->cbr != decoder->cbr_seen) {
1978 decoder->state.type = 0;
1979 return HOP_RETURN;
1980 }
1981 return HOP_IGNORE;
1982
1983 case INTEL_PT_BAD:
1984 case INTEL_PT_PAD:
1985 case INTEL_PT_TIP_PGE:
1986 case INTEL_PT_TSC:
1987 case INTEL_PT_TMA:
1988 case INTEL_PT_MODE_EXEC:
1989 case INTEL_PT_MODE_TSX:
1990 case INTEL_PT_MTC:
1991 case INTEL_PT_CYC:
1992 case INTEL_PT_VMCS:
1993 case INTEL_PT_PSBEND:
1994 case INTEL_PT_CBR:
1995 case INTEL_PT_TRACESTOP:
1996 case INTEL_PT_PIP:
1997 case INTEL_PT_OVF:
1998 case INTEL_PT_MNT:
1999 case INTEL_PT_PTWRITE:
2000 case INTEL_PT_PTWRITE_IP:
2001 case INTEL_PT_EXSTOP:
2002 case INTEL_PT_EXSTOP_IP:
2003 case INTEL_PT_MWAIT:
2004 case INTEL_PT_PWRE:
2005 case INTEL_PT_PWRX:
2006 case INTEL_PT_BBP:
2007 case INTEL_PT_BIP:
2008 case INTEL_PT_BEP:
2009 case INTEL_PT_BEP_IP:
2010 default:
2011 return HOP_PROCESS;
2012 }
2013 }
2014
2015 static int intel_pt_walk_trace(struct intel_pt_decoder *decoder)
2016 {
2017 bool no_tip = false;
2018 int err;
2019
2020 while (1) {
2021 err = intel_pt_get_next_packet(decoder);
2022 if (err)
2023 return err;
2024 next:
2025 if (decoder->hop) {
2026 switch (intel_pt_hop_trace(decoder, &no_tip, &err)) {
2027 case HOP_IGNORE:
2028 continue;
2029 case HOP_RETURN:
2030 return err;
2031 case HOP_AGAIN:
2032 goto next;
2033 default:
2034 break;
2035 }
2036 }
2037
2038 switch (decoder->packet.type) {
2039 case INTEL_PT_TNT:
2040 if (!decoder->packet.count)
2041 break;
2042 decoder->tnt = decoder->packet;
2043 decoder->pkt_state = INTEL_PT_STATE_TNT;
2044 err = intel_pt_walk_tnt(decoder);
2045 if (err == -EAGAIN)
2046 break;
2047 return err;
2048
2049 case INTEL_PT_TIP_PGD:
2050 if (decoder->packet.count != 0)
2051 intel_pt_set_last_ip(decoder);
2052 decoder->pkt_state = INTEL_PT_STATE_TIP_PGD;
2053 return intel_pt_walk_tip(decoder);
2054
2055 case INTEL_PT_TIP_PGE: {
2056 decoder->pge = true;
2057 intel_pt_mtc_cyc_cnt_pge(decoder);
2058 if (decoder->packet.count == 0) {
2059 intel_pt_log_at("Skipping zero TIP.PGE",
2060 decoder->pos);
2061 break;
2062 }
2063 intel_pt_set_ip(decoder);
2064 decoder->state.from_ip = 0;
2065 decoder->state.to_ip = decoder->ip;
2066 decoder->state.type |= INTEL_PT_TRACE_BEGIN;
2067 /*
2068 * In hop mode, resample to get the to_ip as an
2069 * "instruction" sample.
2070 */
2071 if (decoder->hop)
2072 decoder->pkt_state = INTEL_PT_STATE_RESAMPLE;
2073 return 0;
2074 }
2075
2076 case INTEL_PT_OVF:
2077 return intel_pt_overflow(decoder);
2078
2079 case INTEL_PT_TIP:
2080 if (decoder->packet.count != 0)
2081 intel_pt_set_last_ip(decoder);
2082 decoder->pkt_state = INTEL_PT_STATE_TIP;
2083 return intel_pt_walk_tip(decoder);
2084
2085 case INTEL_PT_FUP:
2086 if (decoder->packet.count == 0) {
2087 intel_pt_log_at("Skipping zero FUP",
2088 decoder->pos);
2089 no_tip = false;
2090 break;
2091 }
2092 intel_pt_set_last_ip(decoder);
2093 if (!decoder->branch_enable) {
2094 decoder->ip = decoder->last_ip;
2095 if (intel_pt_fup_event(decoder))
2096 return 0;
2097 no_tip = false;
2098 break;
2099 }
2100 if (decoder->set_fup_mwait)
2101 no_tip = true;
2102 if (no_tip)
2103 decoder->pkt_state = INTEL_PT_STATE_FUP_NO_TIP;
2104 else
2105 decoder->pkt_state = INTEL_PT_STATE_FUP;
2106 err = intel_pt_walk_fup(decoder);
2107 if (err != -EAGAIN)
2108 return err;
2109 if (no_tip) {
2110 no_tip = false;
2111 break;
2112 }
2113 return intel_pt_walk_fup_tip(decoder);
2114
2115 case INTEL_PT_TRACESTOP:
2116 decoder->pge = false;
2117 decoder->continuous_period = false;
2118 intel_pt_clear_tx_flags(decoder);
2119 decoder->have_tma = false;
2120 break;
2121
2122 case INTEL_PT_PSB:
2123 decoder->last_ip = 0;
2124 decoder->have_last_ip = true;
2125 intel_pt_clear_stack(&decoder->stack);
2126 err = intel_pt_walk_psbend(decoder);
2127 if (err == -EAGAIN)
2128 goto next;
2129 if (err)
2130 return err;
2131 /*
2132 * PSB+ CBR will not have changed but cater for the
2133 * possibility of another CBR change that gets caught up
2134 * in the PSB+.
2135 */
2136 if (decoder->cbr != decoder->cbr_seen) {
2137 decoder->state.type = 0;
2138 return 0;
2139 }
2140 break;
2141
2142 case INTEL_PT_PIP:
2143 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2144 break;
2145
2146 case INTEL_PT_MTC:
2147 intel_pt_calc_mtc_timestamp(decoder);
2148 if (decoder->period_type != INTEL_PT_PERIOD_MTC)
2149 break;
2150 /*
2151 * Ensure that there has been an instruction since the
2152 * last MTC.
2153 */
2154 if (!decoder->mtc_insn)
2155 break;
2156 decoder->mtc_insn = false;
2157 /* Ensure that there is a timestamp */
2158 if (!decoder->timestamp)
2159 break;
2160 decoder->state.type = INTEL_PT_INSTRUCTION;
2161 decoder->state.from_ip = decoder->ip;
2162 decoder->state.to_ip = 0;
2163 decoder->mtc_insn = false;
2164 return 0;
2165
2166 case INTEL_PT_TSC:
2167 intel_pt_calc_tsc_timestamp(decoder);
2168 break;
2169
2170 case INTEL_PT_TMA:
2171 intel_pt_calc_tma(decoder);
2172 break;
2173
2174 case INTEL_PT_CYC:
2175 intel_pt_calc_cyc_timestamp(decoder);
2176 break;
2177
2178 case INTEL_PT_CBR:
2179 intel_pt_calc_cbr(decoder);
2180 if (decoder->cbr != decoder->cbr_seen) {
2181 decoder->state.type = 0;
2182 return 0;
2183 }
2184 break;
2185
2186 case INTEL_PT_MODE_EXEC:
2187 decoder->exec_mode = decoder->packet.payload;
2188 break;
2189
2190 case INTEL_PT_MODE_TSX:
2191 /* MODE_TSX need not be followed by FUP */
2192 if (!decoder->pge || decoder->in_psb) {
2193 intel_pt_update_in_tx(decoder);
2194 break;
2195 }
2196 err = intel_pt_mode_tsx(decoder, &no_tip);
2197 if (err)
2198 return err;
2199 goto next;
2200
2201 case INTEL_PT_BAD: /* Does not happen */
2202 return intel_pt_bug(decoder);
2203
2204 case INTEL_PT_PSBEND:
2205 case INTEL_PT_VMCS:
2206 case INTEL_PT_MNT:
2207 case INTEL_PT_PAD:
2208 break;
2209
2210 case INTEL_PT_PTWRITE_IP:
2211 decoder->fup_ptw_payload = decoder->packet.payload;
2212 err = intel_pt_get_next_packet(decoder);
2213 if (err)
2214 return err;
2215 if (decoder->packet.type == INTEL_PT_FUP) {
2216 decoder->set_fup_ptw = true;
2217 no_tip = true;
2218 } else {
2219 intel_pt_log_at("ERROR: Missing FUP after PTWRITE",
2220 decoder->pos);
2221 }
2222 goto next;
2223
2224 case INTEL_PT_PTWRITE:
2225 decoder->state.type = INTEL_PT_PTW;
2226 decoder->state.from_ip = decoder->ip;
2227 decoder->state.to_ip = 0;
2228 decoder->state.ptw_payload = decoder->packet.payload;
2229 return 0;
2230
2231 case INTEL_PT_MWAIT:
2232 decoder->fup_mwait_payload = decoder->packet.payload;
2233 decoder->set_fup_mwait = true;
2234 break;
2235
2236 case INTEL_PT_PWRE:
2237 if (decoder->set_fup_mwait) {
2238 decoder->fup_pwre_payload =
2239 decoder->packet.payload;
2240 decoder->set_fup_pwre = true;
2241 break;
2242 }
2243 decoder->state.type = INTEL_PT_PWR_ENTRY;
2244 decoder->state.from_ip = decoder->ip;
2245 decoder->state.to_ip = 0;
2246 decoder->state.pwrx_payload = decoder->packet.payload;
2247 return 0;
2248
2249 case INTEL_PT_EXSTOP_IP:
2250 err = intel_pt_get_next_packet(decoder);
2251 if (err)
2252 return err;
2253 if (decoder->packet.type == INTEL_PT_FUP) {
2254 decoder->set_fup_exstop = true;
2255 no_tip = true;
2256 } else {
2257 intel_pt_log_at("ERROR: Missing FUP after EXSTOP",
2258 decoder->pos);
2259 }
2260 goto next;
2261
2262 case INTEL_PT_EXSTOP:
2263 decoder->state.type = INTEL_PT_EX_STOP;
2264 decoder->state.from_ip = decoder->ip;
2265 decoder->state.to_ip = 0;
2266 return 0;
2267
2268 case INTEL_PT_PWRX:
2269 decoder->state.type = INTEL_PT_PWR_EXIT;
2270 decoder->state.from_ip = decoder->ip;
2271 decoder->state.to_ip = 0;
2272 decoder->state.pwrx_payload = decoder->packet.payload;
2273 return 0;
2274
2275 case INTEL_PT_BBP:
2276 intel_pt_bbp(decoder);
2277 break;
2278
2279 case INTEL_PT_BIP:
2280 intel_pt_bip(decoder);
2281 break;
2282
2283 case INTEL_PT_BEP:
2284 decoder->state.type = INTEL_PT_BLK_ITEMS;
2285 decoder->state.from_ip = decoder->ip;
2286 decoder->state.to_ip = 0;
2287 return 0;
2288
2289 case INTEL_PT_BEP_IP:
2290 err = intel_pt_get_next_packet(decoder);
2291 if (err)
2292 return err;
2293 if (decoder->packet.type == INTEL_PT_FUP) {
2294 decoder->set_fup_bep = true;
2295 no_tip = true;
2296 } else {
2297 intel_pt_log_at("ERROR: Missing FUP after BEP",
2298 decoder->pos);
2299 }
2300 goto next;
2301
2302 default:
2303 return intel_pt_bug(decoder);
2304 }
2305 }
2306 }
2307
2308 static inline bool intel_pt_have_ip(struct intel_pt_decoder *decoder)
2309 {
2310 return decoder->packet.count &&
2311 (decoder->have_last_ip || decoder->packet.count == 3 ||
2312 decoder->packet.count == 6);
2313 }
2314
2315 /* Walk PSB+ packets to get in sync. */
2316 static int intel_pt_walk_psb(struct intel_pt_decoder *decoder)
2317 {
2318 int err;
2319
2320 decoder->in_psb = true;
2321
2322 while (1) {
2323 err = intel_pt_get_next_packet(decoder);
2324 if (err)
2325 goto out;
2326
2327 switch (decoder->packet.type) {
2328 case INTEL_PT_TIP_PGD:
2329 decoder->continuous_period = false;
2330 __fallthrough;
2331 case INTEL_PT_TIP_PGE:
2332 case INTEL_PT_TIP:
2333 case INTEL_PT_PTWRITE:
2334 case INTEL_PT_PTWRITE_IP:
2335 case INTEL_PT_EXSTOP:
2336 case INTEL_PT_EXSTOP_IP:
2337 case INTEL_PT_MWAIT:
2338 case INTEL_PT_PWRE:
2339 case INTEL_PT_PWRX:
2340 case INTEL_PT_BBP:
2341 case INTEL_PT_BIP:
2342 case INTEL_PT_BEP:
2343 case INTEL_PT_BEP_IP:
2344 intel_pt_log("ERROR: Unexpected packet\n");
2345 err = -ENOENT;
2346 goto out;
2347
2348 case INTEL_PT_FUP:
2349 decoder->pge = true;
2350 if (intel_pt_have_ip(decoder)) {
2351 uint64_t current_ip = decoder->ip;
2352
2353 intel_pt_set_ip(decoder);
2354 if (current_ip)
2355 intel_pt_log_to("Setting IP",
2356 decoder->ip);
2357 }
2358 break;
2359
2360 case INTEL_PT_MTC:
2361 intel_pt_calc_mtc_timestamp(decoder);
2362 break;
2363
2364 case INTEL_PT_TSC:
2365 intel_pt_calc_tsc_timestamp(decoder);
2366 break;
2367
2368 case INTEL_PT_TMA:
2369 intel_pt_calc_tma(decoder);
2370 break;
2371
2372 case INTEL_PT_CYC:
2373 intel_pt_calc_cyc_timestamp(decoder);
2374 break;
2375
2376 case INTEL_PT_CBR:
2377 intel_pt_calc_cbr(decoder);
2378 break;
2379
2380 case INTEL_PT_PIP:
2381 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2382 break;
2383
2384 case INTEL_PT_MODE_EXEC:
2385 decoder->exec_mode = decoder->packet.payload;
2386 break;
2387
2388 case INTEL_PT_MODE_TSX:
2389 intel_pt_update_in_tx(decoder);
2390 break;
2391
2392 case INTEL_PT_TRACESTOP:
2393 decoder->pge = false;
2394 decoder->continuous_period = false;
2395 intel_pt_clear_tx_flags(decoder);
2396 __fallthrough;
2397
2398 case INTEL_PT_TNT:
2399 decoder->have_tma = false;
2400 intel_pt_log("ERROR: Unexpected packet\n");
2401 if (decoder->ip)
2402 decoder->pkt_state = INTEL_PT_STATE_ERR4;
2403 else
2404 decoder->pkt_state = INTEL_PT_STATE_ERR3;
2405 err = -ENOENT;
2406 goto out;
2407
2408 case INTEL_PT_BAD: /* Does not happen */
2409 err = intel_pt_bug(decoder);
2410 goto out;
2411
2412 case INTEL_PT_OVF:
2413 err = intel_pt_overflow(decoder);
2414 goto out;
2415
2416 case INTEL_PT_PSBEND:
2417 err = 0;
2418 goto out;
2419
2420 case INTEL_PT_PSB:
2421 case INTEL_PT_VMCS:
2422 case INTEL_PT_MNT:
2423 case INTEL_PT_PAD:
2424 default:
2425 break;
2426 }
2427 }
2428 out:
2429 decoder->in_psb = false;
2430
2431 return err;
2432 }
2433
2434 static int intel_pt_walk_to_ip(struct intel_pt_decoder *decoder)
2435 {
2436 int err;
2437
2438 while (1) {
2439 err = intel_pt_get_next_packet(decoder);
2440 if (err)
2441 return err;
2442
2443 switch (decoder->packet.type) {
2444 case INTEL_PT_TIP_PGD:
2445 decoder->continuous_period = false;
2446 decoder->pge = false;
2447 if (intel_pt_have_ip(decoder))
2448 intel_pt_set_ip(decoder);
2449 if (!decoder->ip)
2450 break;
2451 decoder->state.type |= INTEL_PT_TRACE_END;
2452 return 0;
2453
2454 case INTEL_PT_TIP_PGE:
2455 decoder->pge = true;
2456 intel_pt_mtc_cyc_cnt_pge(decoder);
2457 if (intel_pt_have_ip(decoder))
2458 intel_pt_set_ip(decoder);
2459 if (!decoder->ip)
2460 break;
2461 decoder->state.type |= INTEL_PT_TRACE_BEGIN;
2462 return 0;
2463
2464 case INTEL_PT_TIP:
2465 decoder->pge = true;
2466 if (intel_pt_have_ip(decoder))
2467 intel_pt_set_ip(decoder);
2468 if (!decoder->ip)
2469 break;
2470 return 0;
2471
2472 case INTEL_PT_FUP:
2473 if (intel_pt_have_ip(decoder))
2474 intel_pt_set_ip(decoder);
2475 if (decoder->ip)
2476 return 0;
2477 break;
2478
2479 case INTEL_PT_MTC:
2480 intel_pt_calc_mtc_timestamp(decoder);
2481 break;
2482
2483 case INTEL_PT_TSC:
2484 intel_pt_calc_tsc_timestamp(decoder);
2485 break;
2486
2487 case INTEL_PT_TMA:
2488 intel_pt_calc_tma(decoder);
2489 break;
2490
2491 case INTEL_PT_CYC:
2492 intel_pt_calc_cyc_timestamp(decoder);
2493 break;
2494
2495 case INTEL_PT_CBR:
2496 intel_pt_calc_cbr(decoder);
2497 break;
2498
2499 case INTEL_PT_PIP:
2500 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2501 break;
2502
2503 case INTEL_PT_MODE_EXEC:
2504 decoder->exec_mode = decoder->packet.payload;
2505 break;
2506
2507 case INTEL_PT_MODE_TSX:
2508 intel_pt_update_in_tx(decoder);
2509 break;
2510
2511 case INTEL_PT_OVF:
2512 return intel_pt_overflow(decoder);
2513
2514 case INTEL_PT_BAD: /* Does not happen */
2515 return intel_pt_bug(decoder);
2516
2517 case INTEL_PT_TRACESTOP:
2518 decoder->pge = false;
2519 decoder->continuous_period = false;
2520 intel_pt_clear_tx_flags(decoder);
2521 decoder->have_tma = false;
2522 break;
2523
2524 case INTEL_PT_PSB:
2525 decoder->last_ip = 0;
2526 decoder->have_last_ip = true;
2527 intel_pt_clear_stack(&decoder->stack);
2528 err = intel_pt_walk_psb(decoder);
2529 if (err)
2530 return err;
2531 if (decoder->ip) {
2532 /* Do not have a sample */
2533 decoder->state.type = 0;
2534 return 0;
2535 }
2536 break;
2537
2538 case INTEL_PT_TNT:
2539 case INTEL_PT_PSBEND:
2540 case INTEL_PT_VMCS:
2541 case INTEL_PT_MNT:
2542 case INTEL_PT_PAD:
2543 case INTEL_PT_PTWRITE:
2544 case INTEL_PT_PTWRITE_IP:
2545 case INTEL_PT_EXSTOP:
2546 case INTEL_PT_EXSTOP_IP:
2547 case INTEL_PT_MWAIT:
2548 case INTEL_PT_PWRE:
2549 case INTEL_PT_PWRX:
2550 case INTEL_PT_BBP:
2551 case INTEL_PT_BIP:
2552 case INTEL_PT_BEP:
2553 case INTEL_PT_BEP_IP:
2554 default:
2555 break;
2556 }
2557 }
2558 }
2559
2560 static int intel_pt_sync_ip(struct intel_pt_decoder *decoder)
2561 {
2562 int err;
2563
2564 decoder->set_fup_tx_flags = false;
2565 decoder->set_fup_ptw = false;
2566 decoder->set_fup_mwait = false;
2567 decoder->set_fup_pwre = false;
2568 decoder->set_fup_exstop = false;
2569 decoder->set_fup_bep = false;
2570
2571 if (!decoder->branch_enable) {
2572 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2573 decoder->overflow = false;
2574 decoder->state.type = 0; /* Do not have a sample */
2575 return 0;
2576 }
2577
2578 intel_pt_log("Scanning for full IP\n");
2579 err = intel_pt_walk_to_ip(decoder);
2580 if (err)
2581 return err;
2582
2583 /* In hop mode, resample to get the to_ip as an "instruction" sample */
2584 if (decoder->hop)
2585 decoder->pkt_state = INTEL_PT_STATE_RESAMPLE;
2586 else
2587 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2588 decoder->overflow = false;
2589
2590 decoder->state.from_ip = 0;
2591 decoder->state.to_ip = decoder->ip;
2592 intel_pt_log_to("Setting IP", decoder->ip);
2593
2594 return 0;
2595 }
2596
2597 static int intel_pt_part_psb(struct intel_pt_decoder *decoder)
2598 {
2599 const unsigned char *end = decoder->buf + decoder->len;
2600 size_t i;
2601
2602 for (i = INTEL_PT_PSB_LEN - 1; i; i--) {
2603 if (i > decoder->len)
2604 continue;
2605 if (!memcmp(end - i, INTEL_PT_PSB_STR, i))
2606 return i;
2607 }
2608 return 0;
2609 }
2610
2611 static int intel_pt_rest_psb(struct intel_pt_decoder *decoder, int part_psb)
2612 {
2613 size_t rest_psb = INTEL_PT_PSB_LEN - part_psb;
2614 const char *psb = INTEL_PT_PSB_STR;
2615
2616 if (rest_psb > decoder->len ||
2617 memcmp(decoder->buf, psb + part_psb, rest_psb))
2618 return 0;
2619
2620 return rest_psb;
2621 }
2622
2623 static int intel_pt_get_split_psb(struct intel_pt_decoder *decoder,
2624 int part_psb)
2625 {
2626 int rest_psb, ret;
2627
2628 decoder->pos += decoder->len;
2629 decoder->len = 0;
2630
2631 ret = intel_pt_get_next_data(decoder, false);
2632 if (ret)
2633 return ret;
2634
2635 rest_psb = intel_pt_rest_psb(decoder, part_psb);
2636 if (!rest_psb)
2637 return 0;
2638
2639 decoder->pos -= part_psb;
2640 decoder->next_buf = decoder->buf + rest_psb;
2641 decoder->next_len = decoder->len - rest_psb;
2642 memcpy(decoder->temp_buf, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2643 decoder->buf = decoder->temp_buf;
2644 decoder->len = INTEL_PT_PSB_LEN;
2645
2646 return 0;
2647 }
2648
2649 static int intel_pt_scan_for_psb(struct intel_pt_decoder *decoder)
2650 {
2651 unsigned char *next;
2652 int ret;
2653
2654 intel_pt_log("Scanning for PSB\n");
2655 while (1) {
2656 if (!decoder->len) {
2657 ret = intel_pt_get_next_data(decoder, false);
2658 if (ret)
2659 return ret;
2660 }
2661
2662 next = memmem(decoder->buf, decoder->len, INTEL_PT_PSB_STR,
2663 INTEL_PT_PSB_LEN);
2664 if (!next) {
2665 int part_psb;
2666
2667 part_psb = intel_pt_part_psb(decoder);
2668 if (part_psb) {
2669 ret = intel_pt_get_split_psb(decoder, part_psb);
2670 if (ret)
2671 return ret;
2672 } else {
2673 decoder->pos += decoder->len;
2674 decoder->len = 0;
2675 }
2676 continue;
2677 }
2678
2679 decoder->pkt_step = next - decoder->buf;
2680 return intel_pt_get_next_packet(decoder);
2681 }
2682 }
2683
2684 static int intel_pt_sync(struct intel_pt_decoder *decoder)
2685 {
2686 int err;
2687
2688 decoder->pge = false;
2689 decoder->continuous_period = false;
2690 decoder->have_last_ip = false;
2691 decoder->last_ip = 0;
2692 decoder->ip = 0;
2693 intel_pt_clear_stack(&decoder->stack);
2694
2695 leap:
2696 err = intel_pt_scan_for_psb(decoder);
2697 if (err)
2698 return err;
2699
2700 decoder->have_last_ip = true;
2701 decoder->pkt_state = INTEL_PT_STATE_NO_IP;
2702
2703 err = intel_pt_walk_psb(decoder);
2704 if (err)
2705 return err;
2706
2707 if (decoder->ip) {
2708 decoder->state.type = 0; /* Do not have a sample */
2709 /*
2710 * In hop mode, resample to get the PSB FUP ip as an
2711 * "instruction" sample.
2712 */
2713 if (decoder->hop)
2714 decoder->pkt_state = INTEL_PT_STATE_RESAMPLE;
2715 else
2716 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2717 } else if (decoder->leap) {
2718 /*
2719 * In leap mode, only PSB+ is decoded, so keeping leaping to the
2720 * next PSB until there is an ip.
2721 */
2722 goto leap;
2723 } else {
2724 return intel_pt_sync_ip(decoder);
2725 }
2726
2727 return 0;
2728 }
2729
2730 static uint64_t intel_pt_est_timestamp(struct intel_pt_decoder *decoder)
2731 {
2732 uint64_t est = decoder->sample_insn_cnt << 1;
2733
2734 if (!decoder->cbr || !decoder->max_non_turbo_ratio)
2735 goto out;
2736
2737 est *= decoder->max_non_turbo_ratio;
2738 est /= decoder->cbr;
2739 out:
2740 return decoder->sample_timestamp + est;
2741 }
2742
2743 const struct intel_pt_state *intel_pt_decode(struct intel_pt_decoder *decoder)
2744 {
2745 int err;
2746
2747 do {
2748 decoder->state.type = INTEL_PT_BRANCH;
2749 decoder->state.flags = 0;
2750
2751 switch (decoder->pkt_state) {
2752 case INTEL_PT_STATE_NO_PSB:
2753 err = intel_pt_sync(decoder);
2754 break;
2755 case INTEL_PT_STATE_NO_IP:
2756 decoder->have_last_ip = false;
2757 decoder->last_ip = 0;
2758 decoder->ip = 0;
2759 __fallthrough;
2760 case INTEL_PT_STATE_ERR_RESYNC:
2761 err = intel_pt_sync_ip(decoder);
2762 break;
2763 case INTEL_PT_STATE_IN_SYNC:
2764 err = intel_pt_walk_trace(decoder);
2765 break;
2766 case INTEL_PT_STATE_TNT:
2767 case INTEL_PT_STATE_TNT_CONT:
2768 err = intel_pt_walk_tnt(decoder);
2769 if (err == -EAGAIN)
2770 err = intel_pt_walk_trace(decoder);
2771 break;
2772 case INTEL_PT_STATE_TIP:
2773 case INTEL_PT_STATE_TIP_PGD:
2774 err = intel_pt_walk_tip(decoder);
2775 break;
2776 case INTEL_PT_STATE_FUP:
2777 err = intel_pt_walk_fup(decoder);
2778 if (err == -EAGAIN)
2779 err = intel_pt_walk_fup_tip(decoder);
2780 break;
2781 case INTEL_PT_STATE_FUP_NO_TIP:
2782 err = intel_pt_walk_fup(decoder);
2783 if (err == -EAGAIN)
2784 err = intel_pt_walk_trace(decoder);
2785 break;
2786 case INTEL_PT_STATE_RESAMPLE:
2787 err = intel_pt_resample(decoder);
2788 break;
2789 default:
2790 err = intel_pt_bug(decoder);
2791 break;
2792 }
2793 } while (err == -ENOLINK);
2794
2795 if (err) {
2796 decoder->state.err = intel_pt_ext_err(err);
2797 decoder->state.from_ip = decoder->ip;
2798 intel_pt_update_sample_time(decoder);
2799 decoder->sample_tot_cyc_cnt = decoder->tot_cyc_cnt;
2800 } else {
2801 decoder->state.err = 0;
2802 if (decoder->cbr != decoder->cbr_seen) {
2803 decoder->cbr_seen = decoder->cbr;
2804 if (!decoder->state.type) {
2805 decoder->state.from_ip = decoder->ip;
2806 decoder->state.to_ip = 0;
2807 }
2808 decoder->state.type |= INTEL_PT_CBR_CHG;
2809 decoder->state.cbr_payload = decoder->cbr_payload;
2810 decoder->state.cbr = decoder->cbr;
2811 }
2812 if (intel_pt_sample_time(decoder->pkt_state)) {
2813 intel_pt_update_sample_time(decoder);
2814 if (decoder->sample_cyc)
2815 decoder->sample_tot_cyc_cnt = decoder->tot_cyc_cnt;
2816 }
2817 }
2818
2819 decoder->state.timestamp = decoder->sample_timestamp;
2820 decoder->state.est_timestamp = intel_pt_est_timestamp(decoder);
2821 decoder->state.cr3 = decoder->cr3;
2822 decoder->state.tot_insn_cnt = decoder->tot_insn_cnt;
2823 decoder->state.tot_cyc_cnt = decoder->sample_tot_cyc_cnt;
2824
2825 return &decoder->state;
2826 }
2827
2828 /**
2829 * intel_pt_next_psb - move buffer pointer to the start of the next PSB packet.
2830 * @buf: pointer to buffer pointer
2831 * @len: size of buffer
2832 *
2833 * Updates the buffer pointer to point to the start of the next PSB packet if
2834 * there is one, otherwise the buffer pointer is unchanged. If @buf is updated,
2835 * @len is adjusted accordingly.
2836 *
2837 * Return: %true if a PSB packet is found, %false otherwise.
2838 */
2839 static bool intel_pt_next_psb(unsigned char **buf, size_t *len)
2840 {
2841 unsigned char *next;
2842
2843 next = memmem(*buf, *len, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2844 if (next) {
2845 *len -= next - *buf;
2846 *buf = next;
2847 return true;
2848 }
2849 return false;
2850 }
2851
2852 /**
2853 * intel_pt_step_psb - move buffer pointer to the start of the following PSB
2854 * packet.
2855 * @buf: pointer to buffer pointer
2856 * @len: size of buffer
2857 *
2858 * Updates the buffer pointer to point to the start of the following PSB packet
2859 * (skipping the PSB at @buf itself) if there is one, otherwise the buffer
2860 * pointer is unchanged. If @buf is updated, @len is adjusted accordingly.
2861 *
2862 * Return: %true if a PSB packet is found, %false otherwise.
2863 */
2864 static bool intel_pt_step_psb(unsigned char **buf, size_t *len)
2865 {
2866 unsigned char *next;
2867
2868 if (!*len)
2869 return false;
2870
2871 next = memmem(*buf + 1, *len - 1, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2872 if (next) {
2873 *len -= next - *buf;
2874 *buf = next;
2875 return true;
2876 }
2877 return false;
2878 }
2879
2880 /**
2881 * intel_pt_last_psb - find the last PSB packet in a buffer.
2882 * @buf: buffer
2883 * @len: size of buffer
2884 *
2885 * This function finds the last PSB in a buffer.
2886 *
2887 * Return: A pointer to the last PSB in @buf if found, %NULL otherwise.
2888 */
2889 static unsigned char *intel_pt_last_psb(unsigned char *buf, size_t len)
2890 {
2891 const char *n = INTEL_PT_PSB_STR;
2892 unsigned char *p;
2893 size_t k;
2894
2895 if (len < INTEL_PT_PSB_LEN)
2896 return NULL;
2897
2898 k = len - INTEL_PT_PSB_LEN + 1;
2899 while (1) {
2900 p = memrchr(buf, n[0], k);
2901 if (!p)
2902 return NULL;
2903 if (!memcmp(p + 1, n + 1, INTEL_PT_PSB_LEN - 1))
2904 return p;
2905 k = p - buf;
2906 if (!k)
2907 return NULL;
2908 }
2909 }
2910
2911 /**
2912 * intel_pt_next_tsc - find and return next TSC.
2913 * @buf: buffer
2914 * @len: size of buffer
2915 * @tsc: TSC value returned
2916 * @rem: returns remaining size when TSC is found
2917 *
2918 * Find a TSC packet in @buf and return the TSC value. This function assumes
2919 * that @buf starts at a PSB and that PSB+ will contain TSC and so stops if a
2920 * PSBEND packet is found.
2921 *
2922 * Return: %true if TSC is found, false otherwise.
2923 */
2924 static bool intel_pt_next_tsc(unsigned char *buf, size_t len, uint64_t *tsc,
2925 size_t *rem)
2926 {
2927 enum intel_pt_pkt_ctx ctx = INTEL_PT_NO_CTX;
2928 struct intel_pt_pkt packet;
2929 int ret;
2930
2931 while (len) {
2932 ret = intel_pt_get_packet(buf, len, &packet, &ctx);
2933 if (ret <= 0)
2934 return false;
2935 if (packet.type == INTEL_PT_TSC) {
2936 *tsc = packet.payload;
2937 *rem = len;
2938 return true;
2939 }
2940 if (packet.type == INTEL_PT_PSBEND)
2941 return false;
2942 buf += ret;
2943 len -= ret;
2944 }
2945 return false;
2946 }
2947
2948 /**
2949 * intel_pt_tsc_cmp - compare 7-byte TSCs.
2950 * @tsc1: first TSC to compare
2951 * @tsc2: second TSC to compare
2952 *
2953 * This function compares 7-byte TSC values allowing for the possibility that
2954 * TSC wrapped around. Generally it is not possible to know if TSC has wrapped
2955 * around so for that purpose this function assumes the absolute difference is
2956 * less than half the maximum difference.
2957 *
2958 * Return: %-1 if @tsc1 is before @tsc2, %0 if @tsc1 == @tsc2, %1 if @tsc1 is
2959 * after @tsc2.
2960 */
2961 static int intel_pt_tsc_cmp(uint64_t tsc1, uint64_t tsc2)
2962 {
2963 const uint64_t halfway = (1ULL << 55);
2964
2965 if (tsc1 == tsc2)
2966 return 0;
2967
2968 if (tsc1 < tsc2) {
2969 if (tsc2 - tsc1 < halfway)
2970 return -1;
2971 else
2972 return 1;
2973 } else {
2974 if (tsc1 - tsc2 < halfway)
2975 return 1;
2976 else
2977 return -1;
2978 }
2979 }
2980
2981 #define MAX_PADDING (PERF_AUXTRACE_RECORD_ALIGNMENT - 1)
2982
2983 /**
2984 * adj_for_padding - adjust overlap to account for padding.
2985 * @buf_b: second buffer
2986 * @buf_a: first buffer
2987 * @len_a: size of first buffer
2988 *
2989 * @buf_a might have up to 7 bytes of padding appended. Adjust the overlap
2990 * accordingly.
2991 *
2992 * Return: A pointer into @buf_b from where non-overlapped data starts
2993 */
2994 static unsigned char *adj_for_padding(unsigned char *buf_b,
2995 unsigned char *buf_a, size_t len_a)
2996 {
2997 unsigned char *p = buf_b - MAX_PADDING;
2998 unsigned char *q = buf_a + len_a - MAX_PADDING;
2999 int i;
3000
3001 for (i = MAX_PADDING; i; i--, p++, q++) {
3002 if (*p != *q)
3003 break;
3004 }
3005
3006 return p;
3007 }
3008
3009 /**
3010 * intel_pt_find_overlap_tsc - determine start of non-overlapped trace data
3011 * using TSC.
3012 * @buf_a: first buffer
3013 * @len_a: size of first buffer
3014 * @buf_b: second buffer
3015 * @len_b: size of second buffer
3016 * @consecutive: returns true if there is data in buf_b that is consecutive
3017 * to buf_a
3018 *
3019 * If the trace contains TSC we can look at the last TSC of @buf_a and the
3020 * first TSC of @buf_b in order to determine if the buffers overlap, and then
3021 * walk forward in @buf_b until a later TSC is found. A precondition is that
3022 * @buf_a and @buf_b are positioned at a PSB.
3023 *
3024 * Return: A pointer into @buf_b from where non-overlapped data starts, or
3025 * @buf_b + @len_b if there is no non-overlapped data.
3026 */
3027 static unsigned char *intel_pt_find_overlap_tsc(unsigned char *buf_a,
3028 size_t len_a,
3029 unsigned char *buf_b,
3030 size_t len_b, bool *consecutive)
3031 {
3032 uint64_t tsc_a, tsc_b;
3033 unsigned char *p;
3034 size_t len, rem_a, rem_b;
3035
3036 p = intel_pt_last_psb(buf_a, len_a);
3037 if (!p)
3038 return buf_b; /* No PSB in buf_a => no overlap */
3039
3040 len = len_a - (p - buf_a);
3041 if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a)) {
3042 /* The last PSB+ in buf_a is incomplete, so go back one more */
3043 len_a -= len;
3044 p = intel_pt_last_psb(buf_a, len_a);
3045 if (!p)
3046 return buf_b; /* No full PSB+ => assume no overlap */
3047 len = len_a - (p - buf_a);
3048 if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a))
3049 return buf_b; /* No TSC in buf_a => assume no overlap */
3050 }
3051
3052 while (1) {
3053 /* Ignore PSB+ with no TSC */
3054 if (intel_pt_next_tsc(buf_b, len_b, &tsc_b, &rem_b)) {
3055 int cmp = intel_pt_tsc_cmp(tsc_a, tsc_b);
3056
3057 /* Same TSC, so buffers are consecutive */
3058 if (!cmp && rem_b >= rem_a) {
3059 unsigned char *start;
3060
3061 *consecutive = true;
3062 start = buf_b + len_b - (rem_b - rem_a);
3063 return adj_for_padding(start, buf_a, len_a);
3064 }
3065 if (cmp < 0)
3066 return buf_b; /* tsc_a < tsc_b => no overlap */
3067 }
3068
3069 if (!intel_pt_step_psb(&buf_b, &len_b))
3070 return buf_b + len_b; /* No PSB in buf_b => no data */
3071 }
3072 }
3073
3074 /**
3075 * intel_pt_find_overlap - determine start of non-overlapped trace data.
3076 * @buf_a: first buffer
3077 * @len_a: size of first buffer
3078 * @buf_b: second buffer
3079 * @len_b: size of second buffer
3080 * @have_tsc: can use TSC packets to detect overlap
3081 * @consecutive: returns true if there is data in buf_b that is consecutive
3082 * to buf_a
3083 *
3084 * When trace samples or snapshots are recorded there is the possibility that
3085 * the data overlaps. Note that, for the purposes of decoding, data is only
3086 * useful if it begins with a PSB packet.
3087 *
3088 * Return: A pointer into @buf_b from where non-overlapped data starts, or
3089 * @buf_b + @len_b if there is no non-overlapped data.
3090 */
3091 unsigned char *intel_pt_find_overlap(unsigned char *buf_a, size_t len_a,
3092 unsigned char *buf_b, size_t len_b,
3093 bool have_tsc, bool *consecutive)
3094 {
3095 unsigned char *found;
3096
3097 /* Buffer 'b' must start at PSB so throw away everything before that */
3098 if (!intel_pt_next_psb(&buf_b, &len_b))
3099 return buf_b + len_b; /* No PSB */
3100
3101 if (!intel_pt_next_psb(&buf_a, &len_a))
3102 return buf_b; /* No overlap */
3103
3104 if (have_tsc) {
3105 found = intel_pt_find_overlap_tsc(buf_a, len_a, buf_b, len_b,
3106 consecutive);
3107 if (found)
3108 return found;
3109 }
3110
3111 /*
3112 * Buffer 'b' cannot end within buffer 'a' so, for comparison purposes,
3113 * we can ignore the first part of buffer 'a'.
3114 */
3115 while (len_b < len_a) {
3116 if (!intel_pt_step_psb(&buf_a, &len_a))
3117 return buf_b; /* No overlap */
3118 }
3119
3120 /* Now len_b >= len_a */
3121 while (1) {
3122 /* Potential overlap so check the bytes */
3123 found = memmem(buf_a, len_a, buf_b, len_a);
3124 if (found) {
3125 *consecutive = true;
3126 return adj_for_padding(buf_b + len_a, buf_a, len_a);
3127 }
3128
3129 /* Try again at next PSB in buffer 'a' */
3130 if (!intel_pt_step_psb(&buf_a, &len_a))
3131 return buf_b; /* No overlap */
3132 }
3133 }
3134
3135 /**
3136 * struct fast_forward_data - data used by intel_pt_ff_cb().
3137 * @timestamp: timestamp to fast forward towards
3138 * @buf_timestamp: buffer timestamp of last buffer with trace data earlier than
3139 * the fast forward timestamp.
3140 */
3141 struct fast_forward_data {
3142 uint64_t timestamp;
3143 uint64_t buf_timestamp;
3144 };
3145
3146 /**
3147 * intel_pt_ff_cb - fast forward lookahead callback.
3148 * @buffer: Intel PT trace buffer
3149 * @data: opaque pointer to fast forward data (struct fast_forward_data)
3150 *
3151 * Determine if @buffer trace is past the fast forward timestamp.
3152 *
3153 * Return: 1 (stop lookahead) if @buffer trace is past the fast forward
3154 * timestamp, and 0 otherwise.
3155 */
3156 static int intel_pt_ff_cb(struct intel_pt_buffer *buffer, void *data)
3157 {
3158 struct fast_forward_data *d = data;
3159 unsigned char *buf;
3160 uint64_t tsc;
3161 size_t rem;
3162 size_t len;
3163
3164 buf = (unsigned char *)buffer->buf;
3165 len = buffer->len;
3166
3167 if (!intel_pt_next_psb(&buf, &len) ||
3168 !intel_pt_next_tsc(buf, len, &tsc, &rem))
3169 return 0;
3170
3171 tsc = intel_pt_8b_tsc(tsc, buffer->ref_timestamp);
3172
3173 intel_pt_log("Buffer 1st timestamp " x64_fmt " ref timestamp " x64_fmt "\n",
3174 tsc, buffer->ref_timestamp);
3175
3176 /*
3177 * If the buffer contains a timestamp earlier that the fast forward
3178 * timestamp, then record it, else stop.
3179 */
3180 if (tsc < d->timestamp)
3181 d->buf_timestamp = buffer->ref_timestamp;
3182 else
3183 return 1;
3184
3185 return 0;
3186 }
3187
3188 /**
3189 * intel_pt_fast_forward - reposition decoder forwards.
3190 * @decoder: Intel PT decoder
3191 * @timestamp: timestamp to fast forward towards
3192 *
3193 * Reposition decoder at the last PSB with a timestamp earlier than @timestamp.
3194 *
3195 * Return: 0 on success or negative error code on failure.
3196 */
3197 int intel_pt_fast_forward(struct intel_pt_decoder *decoder, uint64_t timestamp)
3198 {
3199 struct fast_forward_data d = { .timestamp = timestamp };
3200 unsigned char *buf;
3201 size_t len;
3202 int err;
3203
3204 intel_pt_log("Fast forward towards timestamp " x64_fmt "\n", timestamp);
3205
3206 /* Find buffer timestamp of buffer to fast forward to */
3207 err = decoder->lookahead(decoder->data, intel_pt_ff_cb, &d);
3208 if (err < 0)
3209 return err;
3210
3211 /* Walk to buffer with same buffer timestamp */
3212 if (d.buf_timestamp) {
3213 do {
3214 decoder->pos += decoder->len;
3215 decoder->len = 0;
3216 err = intel_pt_get_next_data(decoder, true);
3217 /* -ENOLINK means non-consecutive trace */
3218 if (err && err != -ENOLINK)
3219 return err;
3220 } while (decoder->buf_timestamp != d.buf_timestamp);
3221 }
3222
3223 if (!decoder->buf)
3224 return 0;
3225
3226 buf = (unsigned char *)decoder->buf;
3227 len = decoder->len;
3228
3229 if (!intel_pt_next_psb(&buf, &len))
3230 return 0;
3231
3232 /*
3233 * Walk PSBs while the PSB timestamp is less than the fast forward
3234 * timestamp.
3235 */
3236 do {
3237 uint64_t tsc;
3238 size_t rem;
3239
3240 if (!intel_pt_next_tsc(buf, len, &tsc, &rem))
3241 break;
3242 tsc = intel_pt_8b_tsc(tsc, decoder->buf_timestamp);
3243 /*
3244 * A TSC packet can slip past MTC packets but, after fast
3245 * forward, decoding starts at the TSC timestamp. That means
3246 * the timestamps may not be exactly the same as the timestamps
3247 * that would have been decoded without fast forward.
3248 */
3249 if (tsc < timestamp) {
3250 intel_pt_log("Fast forward to next PSB timestamp " x64_fmt "\n", tsc);
3251 decoder->pos += decoder->len - len;
3252 decoder->buf = buf;
3253 decoder->len = len;
3254 intel_pt_reposition(decoder);
3255 } else {
3256 break;
3257 }
3258 } while (intel_pt_step_psb(&buf, &len));
3259
3260 return 0;
3261 }
Linux with added FPC-III support