perf annotate: Split allocation of annotated_source struct
[linux/fpc-iii.git] / tools / perf / util / intel-pt-decoder / intel-pt-decoder.c
blobf9157aed12890f3f3dc417dba9c55aeea7ac537e
1 /*
2 * intel_pt_decoder.c: Intel Processor Trace support
3 * Copyright (c) 2013-2014, Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
16 #ifndef _GNU_SOURCE
17 #define _GNU_SOURCE
18 #endif
19 #include <stdlib.h>
20 #include <stdbool.h>
21 #include <string.h>
22 #include <errno.h>
23 #include <stdint.h>
24 #include <inttypes.h>
25 #include <linux/compiler.h>
27 #include "../cache.h"
28 #include "../util.h"
30 #include "intel-pt-insn-decoder.h"
31 #include "intel-pt-pkt-decoder.h"
32 #include "intel-pt-decoder.h"
33 #include "intel-pt-log.h"
35 #define INTEL_PT_BLK_SIZE 1024
37 #define BIT63 (((uint64_t)1 << 63))
39 #define INTEL_PT_RETURN 1
41 /* Maximum number of loops with no packets consumed i.e. stuck in a loop */
42 #define INTEL_PT_MAX_LOOPS 10000
44 struct intel_pt_blk {
45 struct intel_pt_blk *prev;
46 uint64_t ip[INTEL_PT_BLK_SIZE];
49 struct intel_pt_stack {
50 struct intel_pt_blk *blk;
51 struct intel_pt_blk *spare;
52 int pos;
55 enum intel_pt_pkt_state {
56 INTEL_PT_STATE_NO_PSB,
57 INTEL_PT_STATE_NO_IP,
58 INTEL_PT_STATE_ERR_RESYNC,
59 INTEL_PT_STATE_IN_SYNC,
60 INTEL_PT_STATE_TNT,
61 INTEL_PT_STATE_TIP,
62 INTEL_PT_STATE_TIP_PGD,
63 INTEL_PT_STATE_FUP,
64 INTEL_PT_STATE_FUP_NO_TIP,
67 static inline bool intel_pt_sample_time(enum intel_pt_pkt_state pkt_state)
69 switch (pkt_state) {
70 case INTEL_PT_STATE_NO_PSB:
71 case INTEL_PT_STATE_NO_IP:
72 case INTEL_PT_STATE_ERR_RESYNC:
73 case INTEL_PT_STATE_IN_SYNC:
74 case INTEL_PT_STATE_TNT:
75 return true;
76 case INTEL_PT_STATE_TIP:
77 case INTEL_PT_STATE_TIP_PGD:
78 case INTEL_PT_STATE_FUP:
79 case INTEL_PT_STATE_FUP_NO_TIP:
80 return false;
81 default:
82 return true;
86 #ifdef INTEL_PT_STRICT
87 #define INTEL_PT_STATE_ERR1 INTEL_PT_STATE_NO_PSB
88 #define INTEL_PT_STATE_ERR2 INTEL_PT_STATE_NO_PSB
89 #define INTEL_PT_STATE_ERR3 INTEL_PT_STATE_NO_PSB
90 #define INTEL_PT_STATE_ERR4 INTEL_PT_STATE_NO_PSB
91 #else
92 #define INTEL_PT_STATE_ERR1 (decoder->pkt_state)
93 #define INTEL_PT_STATE_ERR2 INTEL_PT_STATE_NO_IP
94 #define INTEL_PT_STATE_ERR3 INTEL_PT_STATE_ERR_RESYNC
95 #define INTEL_PT_STATE_ERR4 INTEL_PT_STATE_IN_SYNC
96 #endif
98 struct intel_pt_decoder {
99 int (*get_trace)(struct intel_pt_buffer *buffer, void *data);
100 int (*walk_insn)(struct intel_pt_insn *intel_pt_insn,
101 uint64_t *insn_cnt_ptr, uint64_t *ip, uint64_t to_ip,
102 uint64_t max_insn_cnt, void *data);
103 bool (*pgd_ip)(uint64_t ip, void *data);
104 void *data;
105 struct intel_pt_state state;
106 const unsigned char *buf;
107 size_t len;
108 bool return_compression;
109 bool branch_enable;
110 bool mtc_insn;
111 bool pge;
112 bool have_tma;
113 bool have_cyc;
114 bool fixup_last_mtc;
115 bool have_last_ip;
116 uint64_t pos;
117 uint64_t last_ip;
118 uint64_t ip;
119 uint64_t cr3;
120 uint64_t timestamp;
121 uint64_t tsc_timestamp;
122 uint64_t ref_timestamp;
123 uint64_t sample_timestamp;
124 uint64_t ret_addr;
125 uint64_t ctc_timestamp;
126 uint64_t ctc_delta;
127 uint64_t cycle_cnt;
128 uint64_t cyc_ref_timestamp;
129 uint32_t last_mtc;
130 uint32_t tsc_ctc_ratio_n;
131 uint32_t tsc_ctc_ratio_d;
132 uint32_t tsc_ctc_mult;
133 uint32_t tsc_slip;
134 uint32_t ctc_rem_mask;
135 int mtc_shift;
136 struct intel_pt_stack stack;
137 enum intel_pt_pkt_state pkt_state;
138 struct intel_pt_pkt packet;
139 struct intel_pt_pkt tnt;
140 int pkt_step;
141 int pkt_len;
142 int last_packet_type;
143 unsigned int cbr;
144 unsigned int cbr_seen;
145 unsigned int max_non_turbo_ratio;
146 double max_non_turbo_ratio_fp;
147 double cbr_cyc_to_tsc;
148 double calc_cyc_to_tsc;
149 bool have_calc_cyc_to_tsc;
150 int exec_mode;
151 unsigned int insn_bytes;
152 uint64_t period;
153 enum intel_pt_period_type period_type;
154 uint64_t tot_insn_cnt;
155 uint64_t period_insn_cnt;
156 uint64_t period_mask;
157 uint64_t period_ticks;
158 uint64_t last_masked_timestamp;
159 bool continuous_period;
160 bool overflow;
161 bool set_fup_tx_flags;
162 bool set_fup_ptw;
163 bool set_fup_mwait;
164 bool set_fup_pwre;
165 bool set_fup_exstop;
166 unsigned int fup_tx_flags;
167 unsigned int tx_flags;
168 uint64_t fup_ptw_payload;
169 uint64_t fup_mwait_payload;
170 uint64_t fup_pwre_payload;
171 uint64_t cbr_payload;
172 uint64_t timestamp_insn_cnt;
173 uint64_t sample_insn_cnt;
174 uint64_t stuck_ip;
175 int no_progress;
176 int stuck_ip_prd;
177 int stuck_ip_cnt;
178 const unsigned char *next_buf;
179 size_t next_len;
180 unsigned char temp_buf[INTEL_PT_PKT_MAX_SZ];
183 static uint64_t intel_pt_lower_power_of_2(uint64_t x)
185 int i;
187 for (i = 0; x != 1; i++)
188 x >>= 1;
190 return x << i;
193 static void intel_pt_setup_period(struct intel_pt_decoder *decoder)
195 if (decoder->period_type == INTEL_PT_PERIOD_TICKS) {
196 uint64_t period;
198 period = intel_pt_lower_power_of_2(decoder->period);
199 decoder->period_mask = ~(period - 1);
200 decoder->period_ticks = period;
204 static uint64_t multdiv(uint64_t t, uint32_t n, uint32_t d)
206 if (!d)
207 return 0;
208 return (t / d) * n + ((t % d) * n) / d;
211 struct intel_pt_decoder *intel_pt_decoder_new(struct intel_pt_params *params)
213 struct intel_pt_decoder *decoder;
215 if (!params->get_trace || !params->walk_insn)
216 return NULL;
218 decoder = zalloc(sizeof(struct intel_pt_decoder));
219 if (!decoder)
220 return NULL;
222 decoder->get_trace = params->get_trace;
223 decoder->walk_insn = params->walk_insn;
224 decoder->pgd_ip = params->pgd_ip;
225 decoder->data = params->data;
226 decoder->return_compression = params->return_compression;
227 decoder->branch_enable = params->branch_enable;
229 decoder->period = params->period;
230 decoder->period_type = params->period_type;
232 decoder->max_non_turbo_ratio = params->max_non_turbo_ratio;
233 decoder->max_non_turbo_ratio_fp = params->max_non_turbo_ratio;
235 intel_pt_setup_period(decoder);
237 decoder->mtc_shift = params->mtc_period;
238 decoder->ctc_rem_mask = (1 << decoder->mtc_shift) - 1;
240 decoder->tsc_ctc_ratio_n = params->tsc_ctc_ratio_n;
241 decoder->tsc_ctc_ratio_d = params->tsc_ctc_ratio_d;
243 if (!decoder->tsc_ctc_ratio_n)
244 decoder->tsc_ctc_ratio_d = 0;
246 if (decoder->tsc_ctc_ratio_d) {
247 if (!(decoder->tsc_ctc_ratio_n % decoder->tsc_ctc_ratio_d))
248 decoder->tsc_ctc_mult = decoder->tsc_ctc_ratio_n /
249 decoder->tsc_ctc_ratio_d;
252 * Allow for timestamps appearing to backwards because a TSC
253 * packet has slipped past a MTC packet, so allow 2 MTC ticks
254 * or ...
256 decoder->tsc_slip = multdiv(2 << decoder->mtc_shift,
257 decoder->tsc_ctc_ratio_n,
258 decoder->tsc_ctc_ratio_d);
260 /* ... or 0x100 paranoia */
261 if (decoder->tsc_slip < 0x100)
262 decoder->tsc_slip = 0x100;
264 intel_pt_log("timestamp: mtc_shift %u\n", decoder->mtc_shift);
265 intel_pt_log("timestamp: tsc_ctc_ratio_n %u\n", decoder->tsc_ctc_ratio_n);
266 intel_pt_log("timestamp: tsc_ctc_ratio_d %u\n", decoder->tsc_ctc_ratio_d);
267 intel_pt_log("timestamp: tsc_ctc_mult %u\n", decoder->tsc_ctc_mult);
268 intel_pt_log("timestamp: tsc_slip %#x\n", decoder->tsc_slip);
270 return decoder;
273 static void intel_pt_pop_blk(struct intel_pt_stack *stack)
275 struct intel_pt_blk *blk = stack->blk;
277 stack->blk = blk->prev;
278 if (!stack->spare)
279 stack->spare = blk;
280 else
281 free(blk);
284 static uint64_t intel_pt_pop(struct intel_pt_stack *stack)
286 if (!stack->pos) {
287 if (!stack->blk)
288 return 0;
289 intel_pt_pop_blk(stack);
290 if (!stack->blk)
291 return 0;
292 stack->pos = INTEL_PT_BLK_SIZE;
294 return stack->blk->ip[--stack->pos];
297 static int intel_pt_alloc_blk(struct intel_pt_stack *stack)
299 struct intel_pt_blk *blk;
301 if (stack->spare) {
302 blk = stack->spare;
303 stack->spare = NULL;
304 } else {
305 blk = malloc(sizeof(struct intel_pt_blk));
306 if (!blk)
307 return -ENOMEM;
310 blk->prev = stack->blk;
311 stack->blk = blk;
312 stack->pos = 0;
313 return 0;
316 static int intel_pt_push(struct intel_pt_stack *stack, uint64_t ip)
318 int err;
320 if (!stack->blk || stack->pos == INTEL_PT_BLK_SIZE) {
321 err = intel_pt_alloc_blk(stack);
322 if (err)
323 return err;
326 stack->blk->ip[stack->pos++] = ip;
327 return 0;
330 static void intel_pt_clear_stack(struct intel_pt_stack *stack)
332 while (stack->blk)
333 intel_pt_pop_blk(stack);
334 stack->pos = 0;
337 static void intel_pt_free_stack(struct intel_pt_stack *stack)
339 intel_pt_clear_stack(stack);
340 zfree(&stack->blk);
341 zfree(&stack->spare);
344 void intel_pt_decoder_free(struct intel_pt_decoder *decoder)
346 intel_pt_free_stack(&decoder->stack);
347 free(decoder);
350 static int intel_pt_ext_err(int code)
352 switch (code) {
353 case -ENOMEM:
354 return INTEL_PT_ERR_NOMEM;
355 case -ENOSYS:
356 return INTEL_PT_ERR_INTERN;
357 case -EBADMSG:
358 return INTEL_PT_ERR_BADPKT;
359 case -ENODATA:
360 return INTEL_PT_ERR_NODATA;
361 case -EILSEQ:
362 return INTEL_PT_ERR_NOINSN;
363 case -ENOENT:
364 return INTEL_PT_ERR_MISMAT;
365 case -EOVERFLOW:
366 return INTEL_PT_ERR_OVR;
367 case -ENOSPC:
368 return INTEL_PT_ERR_LOST;
369 case -ELOOP:
370 return INTEL_PT_ERR_NELOOP;
371 default:
372 return INTEL_PT_ERR_UNK;
376 static const char *intel_pt_err_msgs[] = {
377 [INTEL_PT_ERR_NOMEM] = "Memory allocation failed",
378 [INTEL_PT_ERR_INTERN] = "Internal error",
379 [INTEL_PT_ERR_BADPKT] = "Bad packet",
380 [INTEL_PT_ERR_NODATA] = "No more data",
381 [INTEL_PT_ERR_NOINSN] = "Failed to get instruction",
382 [INTEL_PT_ERR_MISMAT] = "Trace doesn't match instruction",
383 [INTEL_PT_ERR_OVR] = "Overflow packet",
384 [INTEL_PT_ERR_LOST] = "Lost trace data",
385 [INTEL_PT_ERR_UNK] = "Unknown error!",
386 [INTEL_PT_ERR_NELOOP] = "Never-ending loop",
389 int intel_pt__strerror(int code, char *buf, size_t buflen)
391 if (code < 1 || code >= INTEL_PT_ERR_MAX)
392 code = INTEL_PT_ERR_UNK;
393 strlcpy(buf, intel_pt_err_msgs[code], buflen);
394 return 0;
397 static uint64_t intel_pt_calc_ip(const struct intel_pt_pkt *packet,
398 uint64_t last_ip)
400 uint64_t ip;
402 switch (packet->count) {
403 case 1:
404 ip = (last_ip & (uint64_t)0xffffffffffff0000ULL) |
405 packet->payload;
406 break;
407 case 2:
408 ip = (last_ip & (uint64_t)0xffffffff00000000ULL) |
409 packet->payload;
410 break;
411 case 3:
412 ip = packet->payload;
413 /* Sign-extend 6-byte ip */
414 if (ip & (uint64_t)0x800000000000ULL)
415 ip |= (uint64_t)0xffff000000000000ULL;
416 break;
417 case 4:
418 ip = (last_ip & (uint64_t)0xffff000000000000ULL) |
419 packet->payload;
420 break;
421 case 6:
422 ip = packet->payload;
423 break;
424 default:
425 return 0;
428 return ip;
431 static inline void intel_pt_set_last_ip(struct intel_pt_decoder *decoder)
433 decoder->last_ip = intel_pt_calc_ip(&decoder->packet, decoder->last_ip);
434 decoder->have_last_ip = true;
437 static inline void intel_pt_set_ip(struct intel_pt_decoder *decoder)
439 intel_pt_set_last_ip(decoder);
440 decoder->ip = decoder->last_ip;
443 static void intel_pt_decoder_log_packet(struct intel_pt_decoder *decoder)
445 intel_pt_log_packet(&decoder->packet, decoder->pkt_len, decoder->pos,
446 decoder->buf);
449 static int intel_pt_bug(struct intel_pt_decoder *decoder)
451 intel_pt_log("ERROR: Internal error\n");
452 decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
453 return -ENOSYS;
456 static inline void intel_pt_clear_tx_flags(struct intel_pt_decoder *decoder)
458 decoder->tx_flags = 0;
461 static inline void intel_pt_update_in_tx(struct intel_pt_decoder *decoder)
463 decoder->tx_flags = decoder->packet.payload & INTEL_PT_IN_TX;
466 static int intel_pt_bad_packet(struct intel_pt_decoder *decoder)
468 intel_pt_clear_tx_flags(decoder);
469 decoder->have_tma = false;
470 decoder->pkt_len = 1;
471 decoder->pkt_step = 1;
472 intel_pt_decoder_log_packet(decoder);
473 if (decoder->pkt_state != INTEL_PT_STATE_NO_PSB) {
474 intel_pt_log("ERROR: Bad packet\n");
475 decoder->pkt_state = INTEL_PT_STATE_ERR1;
477 return -EBADMSG;
480 static int intel_pt_get_data(struct intel_pt_decoder *decoder)
482 struct intel_pt_buffer buffer = { .buf = 0, };
483 int ret;
485 decoder->pkt_step = 0;
487 intel_pt_log("Getting more data\n");
488 ret = decoder->get_trace(&buffer, decoder->data);
489 if (ret)
490 return ret;
491 decoder->buf = buffer.buf;
492 decoder->len = buffer.len;
493 if (!decoder->len) {
494 intel_pt_log("No more data\n");
495 return -ENODATA;
497 if (!buffer.consecutive) {
498 decoder->ip = 0;
499 decoder->pkt_state = INTEL_PT_STATE_NO_PSB;
500 decoder->ref_timestamp = buffer.ref_timestamp;
501 decoder->timestamp = 0;
502 decoder->have_tma = false;
503 decoder->state.trace_nr = buffer.trace_nr;
504 intel_pt_log("Reference timestamp 0x%" PRIx64 "\n",
505 decoder->ref_timestamp);
506 return -ENOLINK;
509 return 0;
512 static int intel_pt_get_next_data(struct intel_pt_decoder *decoder)
514 if (!decoder->next_buf)
515 return intel_pt_get_data(decoder);
517 decoder->buf = decoder->next_buf;
518 decoder->len = decoder->next_len;
519 decoder->next_buf = 0;
520 decoder->next_len = 0;
521 return 0;
524 static int intel_pt_get_split_packet(struct intel_pt_decoder *decoder)
526 unsigned char *buf = decoder->temp_buf;
527 size_t old_len, len, n;
528 int ret;
530 old_len = decoder->len;
531 len = decoder->len;
532 memcpy(buf, decoder->buf, len);
534 ret = intel_pt_get_data(decoder);
535 if (ret) {
536 decoder->pos += old_len;
537 return ret < 0 ? ret : -EINVAL;
540 n = INTEL_PT_PKT_MAX_SZ - len;
541 if (n > decoder->len)
542 n = decoder->len;
543 memcpy(buf + len, decoder->buf, n);
544 len += n;
546 ret = intel_pt_get_packet(buf, len, &decoder->packet);
547 if (ret < (int)old_len) {
548 decoder->next_buf = decoder->buf;
549 decoder->next_len = decoder->len;
550 decoder->buf = buf;
551 decoder->len = old_len;
552 return intel_pt_bad_packet(decoder);
555 decoder->next_buf = decoder->buf + (ret - old_len);
556 decoder->next_len = decoder->len - (ret - old_len);
558 decoder->buf = buf;
559 decoder->len = ret;
561 return ret;
564 struct intel_pt_pkt_info {
565 struct intel_pt_decoder *decoder;
566 struct intel_pt_pkt packet;
567 uint64_t pos;
568 int pkt_len;
569 int last_packet_type;
570 void *data;
573 typedef int (*intel_pt_pkt_cb_t)(struct intel_pt_pkt_info *pkt_info);
575 /* Lookahead packets in current buffer */
576 static int intel_pt_pkt_lookahead(struct intel_pt_decoder *decoder,
577 intel_pt_pkt_cb_t cb, void *data)
579 struct intel_pt_pkt_info pkt_info;
580 const unsigned char *buf = decoder->buf;
581 size_t len = decoder->len;
582 int ret;
584 pkt_info.decoder = decoder;
585 pkt_info.pos = decoder->pos;
586 pkt_info.pkt_len = decoder->pkt_step;
587 pkt_info.last_packet_type = decoder->last_packet_type;
588 pkt_info.data = data;
590 while (1) {
591 do {
592 pkt_info.pos += pkt_info.pkt_len;
593 buf += pkt_info.pkt_len;
594 len -= pkt_info.pkt_len;
596 if (!len)
597 return INTEL_PT_NEED_MORE_BYTES;
599 ret = intel_pt_get_packet(buf, len, &pkt_info.packet);
600 if (!ret)
601 return INTEL_PT_NEED_MORE_BYTES;
602 if (ret < 0)
603 return ret;
605 pkt_info.pkt_len = ret;
606 } while (pkt_info.packet.type == INTEL_PT_PAD);
608 ret = cb(&pkt_info);
609 if (ret)
610 return 0;
612 pkt_info.last_packet_type = pkt_info.packet.type;
616 struct intel_pt_calc_cyc_to_tsc_info {
617 uint64_t cycle_cnt;
618 unsigned int cbr;
619 uint32_t last_mtc;
620 uint64_t ctc_timestamp;
621 uint64_t ctc_delta;
622 uint64_t tsc_timestamp;
623 uint64_t timestamp;
624 bool have_tma;
625 bool fixup_last_mtc;
626 bool from_mtc;
627 double cbr_cyc_to_tsc;
631 * MTC provides a 8-bit slice of CTC but the TMA packet only provides the lower
632 * 16 bits of CTC. If mtc_shift > 8 then some of the MTC bits are not in the CTC
633 * provided by the TMA packet. Fix-up the last_mtc calculated from the TMA
634 * packet by copying the missing bits from the current MTC assuming the least
635 * difference between the two, and that the current MTC comes after last_mtc.
637 static void intel_pt_fixup_last_mtc(uint32_t mtc, int mtc_shift,
638 uint32_t *last_mtc)
640 uint32_t first_missing_bit = 1U << (16 - mtc_shift);
641 uint32_t mask = ~(first_missing_bit - 1);
643 *last_mtc |= mtc & mask;
644 if (*last_mtc >= mtc) {
645 *last_mtc -= first_missing_bit;
646 *last_mtc &= 0xff;
650 static int intel_pt_calc_cyc_cb(struct intel_pt_pkt_info *pkt_info)
652 struct intel_pt_decoder *decoder = pkt_info->decoder;
653 struct intel_pt_calc_cyc_to_tsc_info *data = pkt_info->data;
654 uint64_t timestamp;
655 double cyc_to_tsc;
656 unsigned int cbr;
657 uint32_t mtc, mtc_delta, ctc, fc, ctc_rem;
659 switch (pkt_info->packet.type) {
660 case INTEL_PT_TNT:
661 case INTEL_PT_TIP_PGE:
662 case INTEL_PT_TIP:
663 case INTEL_PT_FUP:
664 case INTEL_PT_PSB:
665 case INTEL_PT_PIP:
666 case INTEL_PT_MODE_EXEC:
667 case INTEL_PT_MODE_TSX:
668 case INTEL_PT_PSBEND:
669 case INTEL_PT_PAD:
670 case INTEL_PT_VMCS:
671 case INTEL_PT_MNT:
672 case INTEL_PT_PTWRITE:
673 case INTEL_PT_PTWRITE_IP:
674 return 0;
676 case INTEL_PT_MTC:
677 if (!data->have_tma)
678 return 0;
680 mtc = pkt_info->packet.payload;
681 if (decoder->mtc_shift > 8 && data->fixup_last_mtc) {
682 data->fixup_last_mtc = false;
683 intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
684 &data->last_mtc);
686 if (mtc > data->last_mtc)
687 mtc_delta = mtc - data->last_mtc;
688 else
689 mtc_delta = mtc + 256 - data->last_mtc;
690 data->ctc_delta += mtc_delta << decoder->mtc_shift;
691 data->last_mtc = mtc;
693 if (decoder->tsc_ctc_mult) {
694 timestamp = data->ctc_timestamp +
695 data->ctc_delta * decoder->tsc_ctc_mult;
696 } else {
697 timestamp = data->ctc_timestamp +
698 multdiv(data->ctc_delta,
699 decoder->tsc_ctc_ratio_n,
700 decoder->tsc_ctc_ratio_d);
703 if (timestamp < data->timestamp)
704 return 1;
706 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
707 data->timestamp = timestamp;
708 return 0;
711 break;
713 case INTEL_PT_TSC:
715 * For now, do not support using TSC packets - refer
716 * intel_pt_calc_cyc_to_tsc().
718 if (data->from_mtc)
719 return 1;
720 timestamp = pkt_info->packet.payload |
721 (data->timestamp & (0xffULL << 56));
722 if (data->from_mtc && timestamp < data->timestamp &&
723 data->timestamp - timestamp < decoder->tsc_slip)
724 return 1;
725 if (timestamp < data->timestamp)
726 timestamp += (1ULL << 56);
727 if (pkt_info->last_packet_type != INTEL_PT_CYC) {
728 if (data->from_mtc)
729 return 1;
730 data->tsc_timestamp = timestamp;
731 data->timestamp = timestamp;
732 return 0;
734 break;
736 case INTEL_PT_TMA:
737 if (data->from_mtc)
738 return 1;
740 if (!decoder->tsc_ctc_ratio_d)
741 return 0;
743 ctc = pkt_info->packet.payload;
744 fc = pkt_info->packet.count;
745 ctc_rem = ctc & decoder->ctc_rem_mask;
747 data->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
749 data->ctc_timestamp = data->tsc_timestamp - fc;
750 if (decoder->tsc_ctc_mult) {
751 data->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
752 } else {
753 data->ctc_timestamp -=
754 multdiv(ctc_rem, decoder->tsc_ctc_ratio_n,
755 decoder->tsc_ctc_ratio_d);
758 data->ctc_delta = 0;
759 data->have_tma = true;
760 data->fixup_last_mtc = true;
762 return 0;
764 case INTEL_PT_CYC:
765 data->cycle_cnt += pkt_info->packet.payload;
766 return 0;
768 case INTEL_PT_CBR:
769 cbr = pkt_info->packet.payload;
770 if (data->cbr && data->cbr != cbr)
771 return 1;
772 data->cbr = cbr;
773 data->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
774 return 0;
776 case INTEL_PT_TIP_PGD:
777 case INTEL_PT_TRACESTOP:
778 case INTEL_PT_EXSTOP:
779 case INTEL_PT_EXSTOP_IP:
780 case INTEL_PT_MWAIT:
781 case INTEL_PT_PWRE:
782 case INTEL_PT_PWRX:
783 case INTEL_PT_OVF:
784 case INTEL_PT_BAD: /* Does not happen */
785 default:
786 return 1;
789 if (!data->cbr && decoder->cbr) {
790 data->cbr = decoder->cbr;
791 data->cbr_cyc_to_tsc = decoder->cbr_cyc_to_tsc;
794 if (!data->cycle_cnt)
795 return 1;
797 cyc_to_tsc = (double)(timestamp - decoder->timestamp) / data->cycle_cnt;
799 if (data->cbr && cyc_to_tsc > data->cbr_cyc_to_tsc &&
800 cyc_to_tsc / data->cbr_cyc_to_tsc > 1.25) {
801 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle too big (c.f. CBR-based value %g), pos " x64_fmt "\n",
802 cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
803 return 1;
806 decoder->calc_cyc_to_tsc = cyc_to_tsc;
807 decoder->have_calc_cyc_to_tsc = true;
809 if (data->cbr) {
810 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. CBR-based value %g, pos " x64_fmt "\n",
811 cyc_to_tsc, data->cbr_cyc_to_tsc, pkt_info->pos);
812 } else {
813 intel_pt_log("Timestamp: calculated %g TSC ticks per cycle c.f. unknown CBR-based value, pos " x64_fmt "\n",
814 cyc_to_tsc, pkt_info->pos);
817 return 1;
820 static void intel_pt_calc_cyc_to_tsc(struct intel_pt_decoder *decoder,
821 bool from_mtc)
823 struct intel_pt_calc_cyc_to_tsc_info data = {
824 .cycle_cnt = 0,
825 .cbr = 0,
826 .last_mtc = decoder->last_mtc,
827 .ctc_timestamp = decoder->ctc_timestamp,
828 .ctc_delta = decoder->ctc_delta,
829 .tsc_timestamp = decoder->tsc_timestamp,
830 .timestamp = decoder->timestamp,
831 .have_tma = decoder->have_tma,
832 .fixup_last_mtc = decoder->fixup_last_mtc,
833 .from_mtc = from_mtc,
834 .cbr_cyc_to_tsc = 0,
838 * For now, do not support using TSC packets for at least the reasons:
839 * 1) timing might have stopped
840 * 2) TSC packets within PSB+ can slip against CYC packets
842 if (!from_mtc)
843 return;
845 intel_pt_pkt_lookahead(decoder, intel_pt_calc_cyc_cb, &data);
848 static int intel_pt_get_next_packet(struct intel_pt_decoder *decoder)
850 int ret;
852 decoder->last_packet_type = decoder->packet.type;
854 do {
855 decoder->pos += decoder->pkt_step;
856 decoder->buf += decoder->pkt_step;
857 decoder->len -= decoder->pkt_step;
859 if (!decoder->len) {
860 ret = intel_pt_get_next_data(decoder);
861 if (ret)
862 return ret;
865 ret = intel_pt_get_packet(decoder->buf, decoder->len,
866 &decoder->packet);
867 if (ret == INTEL_PT_NEED_MORE_BYTES &&
868 decoder->len < INTEL_PT_PKT_MAX_SZ && !decoder->next_buf) {
869 ret = intel_pt_get_split_packet(decoder);
870 if (ret < 0)
871 return ret;
873 if (ret <= 0)
874 return intel_pt_bad_packet(decoder);
876 decoder->pkt_len = ret;
877 decoder->pkt_step = ret;
878 intel_pt_decoder_log_packet(decoder);
879 } while (decoder->packet.type == INTEL_PT_PAD);
881 return 0;
884 static uint64_t intel_pt_next_period(struct intel_pt_decoder *decoder)
886 uint64_t timestamp, masked_timestamp;
888 timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
889 masked_timestamp = timestamp & decoder->period_mask;
890 if (decoder->continuous_period) {
891 if (masked_timestamp != decoder->last_masked_timestamp)
892 return 1;
893 } else {
894 timestamp += 1;
895 masked_timestamp = timestamp & decoder->period_mask;
896 if (masked_timestamp != decoder->last_masked_timestamp) {
897 decoder->last_masked_timestamp = masked_timestamp;
898 decoder->continuous_period = true;
901 return decoder->period_ticks - (timestamp - masked_timestamp);
904 static uint64_t intel_pt_next_sample(struct intel_pt_decoder *decoder)
906 switch (decoder->period_type) {
907 case INTEL_PT_PERIOD_INSTRUCTIONS:
908 return decoder->period - decoder->period_insn_cnt;
909 case INTEL_PT_PERIOD_TICKS:
910 return intel_pt_next_period(decoder);
911 case INTEL_PT_PERIOD_NONE:
912 case INTEL_PT_PERIOD_MTC:
913 default:
914 return 0;
918 static void intel_pt_sample_insn(struct intel_pt_decoder *decoder)
920 uint64_t timestamp, masked_timestamp;
922 switch (decoder->period_type) {
923 case INTEL_PT_PERIOD_INSTRUCTIONS:
924 decoder->period_insn_cnt = 0;
925 break;
926 case INTEL_PT_PERIOD_TICKS:
927 timestamp = decoder->timestamp + decoder->timestamp_insn_cnt;
928 masked_timestamp = timestamp & decoder->period_mask;
929 decoder->last_masked_timestamp = masked_timestamp;
930 break;
931 case INTEL_PT_PERIOD_NONE:
932 case INTEL_PT_PERIOD_MTC:
933 default:
934 break;
937 decoder->state.type |= INTEL_PT_INSTRUCTION;
940 static int intel_pt_walk_insn(struct intel_pt_decoder *decoder,
941 struct intel_pt_insn *intel_pt_insn, uint64_t ip)
943 uint64_t max_insn_cnt, insn_cnt = 0;
944 int err;
946 if (!decoder->mtc_insn)
947 decoder->mtc_insn = true;
949 max_insn_cnt = intel_pt_next_sample(decoder);
951 err = decoder->walk_insn(intel_pt_insn, &insn_cnt, &decoder->ip, ip,
952 max_insn_cnt, decoder->data);
954 decoder->tot_insn_cnt += insn_cnt;
955 decoder->timestamp_insn_cnt += insn_cnt;
956 decoder->sample_insn_cnt += insn_cnt;
957 decoder->period_insn_cnt += insn_cnt;
959 if (err) {
960 decoder->no_progress = 0;
961 decoder->pkt_state = INTEL_PT_STATE_ERR2;
962 intel_pt_log_at("ERROR: Failed to get instruction",
963 decoder->ip);
964 if (err == -ENOENT)
965 return -ENOLINK;
966 return -EILSEQ;
969 if (ip && decoder->ip == ip) {
970 err = -EAGAIN;
971 goto out;
974 if (max_insn_cnt && insn_cnt >= max_insn_cnt)
975 intel_pt_sample_insn(decoder);
977 if (intel_pt_insn->branch == INTEL_PT_BR_NO_BRANCH) {
978 decoder->state.type = INTEL_PT_INSTRUCTION;
979 decoder->state.from_ip = decoder->ip;
980 decoder->state.to_ip = 0;
981 decoder->ip += intel_pt_insn->length;
982 err = INTEL_PT_RETURN;
983 goto out;
986 if (intel_pt_insn->op == INTEL_PT_OP_CALL) {
987 /* Zero-length calls are excluded */
988 if (intel_pt_insn->branch != INTEL_PT_BR_UNCONDITIONAL ||
989 intel_pt_insn->rel) {
990 err = intel_pt_push(&decoder->stack, decoder->ip +
991 intel_pt_insn->length);
992 if (err)
993 goto out;
995 } else if (intel_pt_insn->op == INTEL_PT_OP_RET) {
996 decoder->ret_addr = intel_pt_pop(&decoder->stack);
999 if (intel_pt_insn->branch == INTEL_PT_BR_UNCONDITIONAL) {
1000 int cnt = decoder->no_progress++;
1002 decoder->state.from_ip = decoder->ip;
1003 decoder->ip += intel_pt_insn->length +
1004 intel_pt_insn->rel;
1005 decoder->state.to_ip = decoder->ip;
1006 err = INTEL_PT_RETURN;
1009 * Check for being stuck in a loop. This can happen if a
1010 * decoder error results in the decoder erroneously setting the
1011 * ip to an address that is itself in an infinite loop that
1012 * consumes no packets. When that happens, there must be an
1013 * unconditional branch.
1015 if (cnt) {
1016 if (cnt == 1) {
1017 decoder->stuck_ip = decoder->state.to_ip;
1018 decoder->stuck_ip_prd = 1;
1019 decoder->stuck_ip_cnt = 1;
1020 } else if (cnt > INTEL_PT_MAX_LOOPS ||
1021 decoder->state.to_ip == decoder->stuck_ip) {
1022 intel_pt_log_at("ERROR: Never-ending loop",
1023 decoder->state.to_ip);
1024 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1025 err = -ELOOP;
1026 goto out;
1027 } else if (!--decoder->stuck_ip_cnt) {
1028 decoder->stuck_ip_prd += 1;
1029 decoder->stuck_ip_cnt = decoder->stuck_ip_prd;
1030 decoder->stuck_ip = decoder->state.to_ip;
1033 goto out_no_progress;
1035 out:
1036 decoder->no_progress = 0;
1037 out_no_progress:
1038 decoder->state.insn_op = intel_pt_insn->op;
1039 decoder->state.insn_len = intel_pt_insn->length;
1040 memcpy(decoder->state.insn, intel_pt_insn->buf,
1041 INTEL_PT_INSN_BUF_SZ);
1043 if (decoder->tx_flags & INTEL_PT_IN_TX)
1044 decoder->state.flags |= INTEL_PT_IN_TX;
1046 return err;
1049 static bool intel_pt_fup_event(struct intel_pt_decoder *decoder)
1051 bool ret = false;
1053 if (decoder->set_fup_tx_flags) {
1054 decoder->set_fup_tx_flags = false;
1055 decoder->tx_flags = decoder->fup_tx_flags;
1056 decoder->state.type = INTEL_PT_TRANSACTION;
1057 decoder->state.from_ip = decoder->ip;
1058 decoder->state.to_ip = 0;
1059 decoder->state.flags = decoder->fup_tx_flags;
1060 return true;
1062 if (decoder->set_fup_ptw) {
1063 decoder->set_fup_ptw = false;
1064 decoder->state.type = INTEL_PT_PTW;
1065 decoder->state.flags |= INTEL_PT_FUP_IP;
1066 decoder->state.from_ip = decoder->ip;
1067 decoder->state.to_ip = 0;
1068 decoder->state.ptw_payload = decoder->fup_ptw_payload;
1069 return true;
1071 if (decoder->set_fup_mwait) {
1072 decoder->set_fup_mwait = false;
1073 decoder->state.type = INTEL_PT_MWAIT_OP;
1074 decoder->state.from_ip = decoder->ip;
1075 decoder->state.to_ip = 0;
1076 decoder->state.mwait_payload = decoder->fup_mwait_payload;
1077 ret = true;
1079 if (decoder->set_fup_pwre) {
1080 decoder->set_fup_pwre = false;
1081 decoder->state.type |= INTEL_PT_PWR_ENTRY;
1082 decoder->state.type &= ~INTEL_PT_BRANCH;
1083 decoder->state.from_ip = decoder->ip;
1084 decoder->state.to_ip = 0;
1085 decoder->state.pwre_payload = decoder->fup_pwre_payload;
1086 ret = true;
1088 if (decoder->set_fup_exstop) {
1089 decoder->set_fup_exstop = false;
1090 decoder->state.type |= INTEL_PT_EX_STOP;
1091 decoder->state.type &= ~INTEL_PT_BRANCH;
1092 decoder->state.flags |= INTEL_PT_FUP_IP;
1093 decoder->state.from_ip = decoder->ip;
1094 decoder->state.to_ip = 0;
1095 ret = true;
1097 return ret;
1100 static int intel_pt_walk_fup(struct intel_pt_decoder *decoder)
1102 struct intel_pt_insn intel_pt_insn;
1103 uint64_t ip;
1104 int err;
1106 ip = decoder->last_ip;
1108 while (1) {
1109 err = intel_pt_walk_insn(decoder, &intel_pt_insn, ip);
1110 if (err == INTEL_PT_RETURN)
1111 return 0;
1112 if (err == -EAGAIN) {
1113 if (intel_pt_fup_event(decoder))
1114 return 0;
1115 return err;
1117 decoder->set_fup_tx_flags = false;
1118 if (err)
1119 return err;
1121 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1122 intel_pt_log_at("ERROR: Unexpected indirect branch",
1123 decoder->ip);
1124 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1125 return -ENOENT;
1128 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1129 intel_pt_log_at("ERROR: Unexpected conditional branch",
1130 decoder->ip);
1131 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1132 return -ENOENT;
1135 intel_pt_bug(decoder);
1139 static int intel_pt_walk_tip(struct intel_pt_decoder *decoder)
1141 struct intel_pt_insn intel_pt_insn;
1142 int err;
1144 err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1145 if (err == INTEL_PT_RETURN &&
1146 decoder->pgd_ip &&
1147 decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1148 (decoder->state.type & INTEL_PT_BRANCH) &&
1149 decoder->pgd_ip(decoder->state.to_ip, decoder->data)) {
1150 /* Unconditional branch leaving filter region */
1151 decoder->no_progress = 0;
1152 decoder->pge = false;
1153 decoder->continuous_period = false;
1154 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1155 decoder->state.to_ip = 0;
1156 return 0;
1158 if (err == INTEL_PT_RETURN)
1159 return 0;
1160 if (err)
1161 return err;
1163 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1164 if (decoder->pkt_state == INTEL_PT_STATE_TIP_PGD) {
1165 decoder->pge = false;
1166 decoder->continuous_period = false;
1167 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1168 decoder->state.from_ip = decoder->ip;
1169 decoder->state.to_ip = 0;
1170 if (decoder->packet.count != 0)
1171 decoder->ip = decoder->last_ip;
1172 } else {
1173 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1174 decoder->state.from_ip = decoder->ip;
1175 if (decoder->packet.count == 0) {
1176 decoder->state.to_ip = 0;
1177 } else {
1178 decoder->state.to_ip = decoder->last_ip;
1179 decoder->ip = decoder->last_ip;
1182 return 0;
1185 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1186 uint64_t to_ip = decoder->ip + intel_pt_insn.length +
1187 intel_pt_insn.rel;
1189 if (decoder->pgd_ip &&
1190 decoder->pkt_state == INTEL_PT_STATE_TIP_PGD &&
1191 decoder->pgd_ip(to_ip, decoder->data)) {
1192 /* Conditional branch leaving filter region */
1193 decoder->pge = false;
1194 decoder->continuous_period = false;
1195 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1196 decoder->ip = to_ip;
1197 decoder->state.from_ip = decoder->ip;
1198 decoder->state.to_ip = 0;
1199 return 0;
1201 intel_pt_log_at("ERROR: Conditional branch when expecting indirect branch",
1202 decoder->ip);
1203 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1204 return -ENOENT;
1207 return intel_pt_bug(decoder);
1210 static int intel_pt_walk_tnt(struct intel_pt_decoder *decoder)
1212 struct intel_pt_insn intel_pt_insn;
1213 int err;
1215 while (1) {
1216 err = intel_pt_walk_insn(decoder, &intel_pt_insn, 0);
1217 if (err == INTEL_PT_RETURN)
1218 return 0;
1219 if (err)
1220 return err;
1222 if (intel_pt_insn.op == INTEL_PT_OP_RET) {
1223 if (!decoder->return_compression) {
1224 intel_pt_log_at("ERROR: RET when expecting conditional branch",
1225 decoder->ip);
1226 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1227 return -ENOENT;
1229 if (!decoder->ret_addr) {
1230 intel_pt_log_at("ERROR: Bad RET compression (stack empty)",
1231 decoder->ip);
1232 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1233 return -ENOENT;
1235 if (!(decoder->tnt.payload & BIT63)) {
1236 intel_pt_log_at("ERROR: Bad RET compression (TNT=N)",
1237 decoder->ip);
1238 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1239 return -ENOENT;
1241 decoder->tnt.count -= 1;
1242 if (!decoder->tnt.count)
1243 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1244 decoder->tnt.payload <<= 1;
1245 decoder->state.from_ip = decoder->ip;
1246 decoder->ip = decoder->ret_addr;
1247 decoder->state.to_ip = decoder->ip;
1248 return 0;
1251 if (intel_pt_insn.branch == INTEL_PT_BR_INDIRECT) {
1252 /* Handle deferred TIPs */
1253 err = intel_pt_get_next_packet(decoder);
1254 if (err)
1255 return err;
1256 if (decoder->packet.type != INTEL_PT_TIP ||
1257 decoder->packet.count == 0) {
1258 intel_pt_log_at("ERROR: Missing deferred TIP for indirect branch",
1259 decoder->ip);
1260 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1261 decoder->pkt_step = 0;
1262 return -ENOENT;
1264 intel_pt_set_last_ip(decoder);
1265 decoder->state.from_ip = decoder->ip;
1266 decoder->state.to_ip = decoder->last_ip;
1267 decoder->ip = decoder->last_ip;
1268 return 0;
1271 if (intel_pt_insn.branch == INTEL_PT_BR_CONDITIONAL) {
1272 decoder->tnt.count -= 1;
1273 if (!decoder->tnt.count)
1274 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
1275 if (decoder->tnt.payload & BIT63) {
1276 decoder->tnt.payload <<= 1;
1277 decoder->state.from_ip = decoder->ip;
1278 decoder->ip += intel_pt_insn.length +
1279 intel_pt_insn.rel;
1280 decoder->state.to_ip = decoder->ip;
1281 return 0;
1283 /* Instruction sample for a non-taken branch */
1284 if (decoder->state.type & INTEL_PT_INSTRUCTION) {
1285 decoder->tnt.payload <<= 1;
1286 decoder->state.type = INTEL_PT_INSTRUCTION;
1287 decoder->state.from_ip = decoder->ip;
1288 decoder->state.to_ip = 0;
1289 decoder->ip += intel_pt_insn.length;
1290 return 0;
1292 decoder->ip += intel_pt_insn.length;
1293 if (!decoder->tnt.count)
1294 return -EAGAIN;
1295 decoder->tnt.payload <<= 1;
1296 continue;
1299 return intel_pt_bug(decoder);
1303 static int intel_pt_mode_tsx(struct intel_pt_decoder *decoder, bool *no_tip)
1305 unsigned int fup_tx_flags;
1306 int err;
1308 fup_tx_flags = decoder->packet.payload &
1309 (INTEL_PT_IN_TX | INTEL_PT_ABORT_TX);
1310 err = intel_pt_get_next_packet(decoder);
1311 if (err)
1312 return err;
1313 if (decoder->packet.type == INTEL_PT_FUP) {
1314 decoder->fup_tx_flags = fup_tx_flags;
1315 decoder->set_fup_tx_flags = true;
1316 if (!(decoder->fup_tx_flags & INTEL_PT_ABORT_TX))
1317 *no_tip = true;
1318 } else {
1319 intel_pt_log_at("ERROR: Missing FUP after MODE.TSX",
1320 decoder->pos);
1321 intel_pt_update_in_tx(decoder);
1323 return 0;
1326 static void intel_pt_calc_tsc_timestamp(struct intel_pt_decoder *decoder)
1328 uint64_t timestamp;
1330 decoder->have_tma = false;
1332 if (decoder->ref_timestamp) {
1333 timestamp = decoder->packet.payload |
1334 (decoder->ref_timestamp & (0xffULL << 56));
1335 if (timestamp < decoder->ref_timestamp) {
1336 if (decoder->ref_timestamp - timestamp > (1ULL << 55))
1337 timestamp += (1ULL << 56);
1338 } else {
1339 if (timestamp - decoder->ref_timestamp > (1ULL << 55))
1340 timestamp -= (1ULL << 56);
1342 decoder->tsc_timestamp = timestamp;
1343 decoder->timestamp = timestamp;
1344 decoder->ref_timestamp = 0;
1345 decoder->timestamp_insn_cnt = 0;
1346 } else if (decoder->timestamp) {
1347 timestamp = decoder->packet.payload |
1348 (decoder->timestamp & (0xffULL << 56));
1349 decoder->tsc_timestamp = timestamp;
1350 if (timestamp < decoder->timestamp &&
1351 decoder->timestamp - timestamp < decoder->tsc_slip) {
1352 intel_pt_log_to("Suppressing backwards timestamp",
1353 timestamp);
1354 timestamp = decoder->timestamp;
1356 if (timestamp < decoder->timestamp) {
1357 intel_pt_log_to("Wraparound timestamp", timestamp);
1358 timestamp += (1ULL << 56);
1359 decoder->tsc_timestamp = timestamp;
1361 decoder->timestamp = timestamp;
1362 decoder->timestamp_insn_cnt = 0;
1365 if (decoder->last_packet_type == INTEL_PT_CYC) {
1366 decoder->cyc_ref_timestamp = decoder->timestamp;
1367 decoder->cycle_cnt = 0;
1368 decoder->have_calc_cyc_to_tsc = false;
1369 intel_pt_calc_cyc_to_tsc(decoder, false);
1372 intel_pt_log_to("Setting timestamp", decoder->timestamp);
1375 static int intel_pt_overflow(struct intel_pt_decoder *decoder)
1377 intel_pt_log("ERROR: Buffer overflow\n");
1378 intel_pt_clear_tx_flags(decoder);
1379 decoder->have_tma = false;
1380 decoder->cbr = 0;
1381 decoder->timestamp_insn_cnt = 0;
1382 decoder->pkt_state = INTEL_PT_STATE_ERR_RESYNC;
1383 decoder->overflow = true;
1384 return -EOVERFLOW;
1387 static void intel_pt_calc_tma(struct intel_pt_decoder *decoder)
1389 uint32_t ctc = decoder->packet.payload;
1390 uint32_t fc = decoder->packet.count;
1391 uint32_t ctc_rem = ctc & decoder->ctc_rem_mask;
1393 if (!decoder->tsc_ctc_ratio_d)
1394 return;
1396 decoder->last_mtc = (ctc >> decoder->mtc_shift) & 0xff;
1397 decoder->ctc_timestamp = decoder->tsc_timestamp - fc;
1398 if (decoder->tsc_ctc_mult) {
1399 decoder->ctc_timestamp -= ctc_rem * decoder->tsc_ctc_mult;
1400 } else {
1401 decoder->ctc_timestamp -= multdiv(ctc_rem,
1402 decoder->tsc_ctc_ratio_n,
1403 decoder->tsc_ctc_ratio_d);
1405 decoder->ctc_delta = 0;
1406 decoder->have_tma = true;
1407 decoder->fixup_last_mtc = true;
1408 intel_pt_log("CTC timestamp " x64_fmt " last MTC %#x CTC rem %#x\n",
1409 decoder->ctc_timestamp, decoder->last_mtc, ctc_rem);
1412 static void intel_pt_calc_mtc_timestamp(struct intel_pt_decoder *decoder)
1414 uint64_t timestamp;
1415 uint32_t mtc, mtc_delta;
1417 if (!decoder->have_tma)
1418 return;
1420 mtc = decoder->packet.payload;
1422 if (decoder->mtc_shift > 8 && decoder->fixup_last_mtc) {
1423 decoder->fixup_last_mtc = false;
1424 intel_pt_fixup_last_mtc(mtc, decoder->mtc_shift,
1425 &decoder->last_mtc);
1428 if (mtc > decoder->last_mtc)
1429 mtc_delta = mtc - decoder->last_mtc;
1430 else
1431 mtc_delta = mtc + 256 - decoder->last_mtc;
1433 decoder->ctc_delta += mtc_delta << decoder->mtc_shift;
1435 if (decoder->tsc_ctc_mult) {
1436 timestamp = decoder->ctc_timestamp +
1437 decoder->ctc_delta * decoder->tsc_ctc_mult;
1438 } else {
1439 timestamp = decoder->ctc_timestamp +
1440 multdiv(decoder->ctc_delta,
1441 decoder->tsc_ctc_ratio_n,
1442 decoder->tsc_ctc_ratio_d);
1445 if (timestamp < decoder->timestamp)
1446 intel_pt_log("Suppressing MTC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1447 timestamp, decoder->timestamp);
1448 else
1449 decoder->timestamp = timestamp;
1451 decoder->timestamp_insn_cnt = 0;
1452 decoder->last_mtc = mtc;
1454 if (decoder->last_packet_type == INTEL_PT_CYC) {
1455 decoder->cyc_ref_timestamp = decoder->timestamp;
1456 decoder->cycle_cnt = 0;
1457 decoder->have_calc_cyc_to_tsc = false;
1458 intel_pt_calc_cyc_to_tsc(decoder, true);
1462 static void intel_pt_calc_cbr(struct intel_pt_decoder *decoder)
1464 unsigned int cbr = decoder->packet.payload & 0xff;
1466 decoder->cbr_payload = decoder->packet.payload;
1468 if (decoder->cbr == cbr)
1469 return;
1471 decoder->cbr = cbr;
1472 decoder->cbr_cyc_to_tsc = decoder->max_non_turbo_ratio_fp / cbr;
1475 static void intel_pt_calc_cyc_timestamp(struct intel_pt_decoder *decoder)
1477 uint64_t timestamp = decoder->cyc_ref_timestamp;
1479 decoder->have_cyc = true;
1481 decoder->cycle_cnt += decoder->packet.payload;
1483 if (!decoder->cyc_ref_timestamp)
1484 return;
1486 if (decoder->have_calc_cyc_to_tsc)
1487 timestamp += decoder->cycle_cnt * decoder->calc_cyc_to_tsc;
1488 else if (decoder->cbr)
1489 timestamp += decoder->cycle_cnt * decoder->cbr_cyc_to_tsc;
1490 else
1491 return;
1493 if (timestamp < decoder->timestamp)
1494 intel_pt_log("Suppressing CYC timestamp " x64_fmt " less than current timestamp " x64_fmt "\n",
1495 timestamp, decoder->timestamp);
1496 else
1497 decoder->timestamp = timestamp;
1499 decoder->timestamp_insn_cnt = 0;
1502 /* Walk PSB+ packets when already in sync. */
1503 static int intel_pt_walk_psbend(struct intel_pt_decoder *decoder)
1505 int err;
1507 while (1) {
1508 err = intel_pt_get_next_packet(decoder);
1509 if (err)
1510 return err;
1512 switch (decoder->packet.type) {
1513 case INTEL_PT_PSBEND:
1514 return 0;
1516 case INTEL_PT_TIP_PGD:
1517 case INTEL_PT_TIP_PGE:
1518 case INTEL_PT_TIP:
1519 case INTEL_PT_TNT:
1520 case INTEL_PT_TRACESTOP:
1521 case INTEL_PT_BAD:
1522 case INTEL_PT_PSB:
1523 case INTEL_PT_PTWRITE:
1524 case INTEL_PT_PTWRITE_IP:
1525 case INTEL_PT_EXSTOP:
1526 case INTEL_PT_EXSTOP_IP:
1527 case INTEL_PT_MWAIT:
1528 case INTEL_PT_PWRE:
1529 case INTEL_PT_PWRX:
1530 decoder->have_tma = false;
1531 intel_pt_log("ERROR: Unexpected packet\n");
1532 return -EAGAIN;
1534 case INTEL_PT_OVF:
1535 return intel_pt_overflow(decoder);
1537 case INTEL_PT_TSC:
1538 intel_pt_calc_tsc_timestamp(decoder);
1539 break;
1541 case INTEL_PT_TMA:
1542 intel_pt_calc_tma(decoder);
1543 break;
1545 case INTEL_PT_CBR:
1546 intel_pt_calc_cbr(decoder);
1547 break;
1549 case INTEL_PT_MODE_EXEC:
1550 decoder->exec_mode = decoder->packet.payload;
1551 break;
1553 case INTEL_PT_PIP:
1554 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1555 break;
1557 case INTEL_PT_FUP:
1558 decoder->pge = true;
1559 if (decoder->packet.count)
1560 intel_pt_set_last_ip(decoder);
1561 break;
1563 case INTEL_PT_MODE_TSX:
1564 intel_pt_update_in_tx(decoder);
1565 break;
1567 case INTEL_PT_MTC:
1568 intel_pt_calc_mtc_timestamp(decoder);
1569 if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1570 decoder->state.type |= INTEL_PT_INSTRUCTION;
1571 break;
1573 case INTEL_PT_CYC:
1574 case INTEL_PT_VMCS:
1575 case INTEL_PT_MNT:
1576 case INTEL_PT_PAD:
1577 default:
1578 break;
1583 static int intel_pt_walk_fup_tip(struct intel_pt_decoder *decoder)
1585 int err;
1587 if (decoder->tx_flags & INTEL_PT_ABORT_TX) {
1588 decoder->tx_flags = 0;
1589 decoder->state.flags &= ~INTEL_PT_IN_TX;
1590 decoder->state.flags |= INTEL_PT_ABORT_TX;
1591 } else {
1592 decoder->state.flags |= INTEL_PT_ASYNC;
1595 while (1) {
1596 err = intel_pt_get_next_packet(decoder);
1597 if (err)
1598 return err;
1600 switch (decoder->packet.type) {
1601 case INTEL_PT_TNT:
1602 case INTEL_PT_FUP:
1603 case INTEL_PT_TRACESTOP:
1604 case INTEL_PT_PSB:
1605 case INTEL_PT_TSC:
1606 case INTEL_PT_TMA:
1607 case INTEL_PT_CBR:
1608 case INTEL_PT_MODE_TSX:
1609 case INTEL_PT_BAD:
1610 case INTEL_PT_PSBEND:
1611 case INTEL_PT_PTWRITE:
1612 case INTEL_PT_PTWRITE_IP:
1613 case INTEL_PT_EXSTOP:
1614 case INTEL_PT_EXSTOP_IP:
1615 case INTEL_PT_MWAIT:
1616 case INTEL_PT_PWRE:
1617 case INTEL_PT_PWRX:
1618 intel_pt_log("ERROR: Missing TIP after FUP\n");
1619 decoder->pkt_state = INTEL_PT_STATE_ERR3;
1620 decoder->pkt_step = 0;
1621 return -ENOENT;
1623 case INTEL_PT_OVF:
1624 return intel_pt_overflow(decoder);
1626 case INTEL_PT_TIP_PGD:
1627 decoder->state.from_ip = decoder->ip;
1628 decoder->state.to_ip = 0;
1629 if (decoder->packet.count != 0) {
1630 intel_pt_set_ip(decoder);
1631 intel_pt_log("Omitting PGD ip " x64_fmt "\n",
1632 decoder->ip);
1634 decoder->pge = false;
1635 decoder->continuous_period = false;
1636 return 0;
1638 case INTEL_PT_TIP_PGE:
1639 decoder->pge = true;
1640 intel_pt_log("Omitting PGE ip " x64_fmt "\n",
1641 decoder->ip);
1642 decoder->state.from_ip = 0;
1643 if (decoder->packet.count == 0) {
1644 decoder->state.to_ip = 0;
1645 } else {
1646 intel_pt_set_ip(decoder);
1647 decoder->state.to_ip = decoder->ip;
1649 return 0;
1651 case INTEL_PT_TIP:
1652 decoder->state.from_ip = decoder->ip;
1653 if (decoder->packet.count == 0) {
1654 decoder->state.to_ip = 0;
1655 } else {
1656 intel_pt_set_ip(decoder);
1657 decoder->state.to_ip = decoder->ip;
1659 return 0;
1661 case INTEL_PT_PIP:
1662 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1663 break;
1665 case INTEL_PT_MTC:
1666 intel_pt_calc_mtc_timestamp(decoder);
1667 if (decoder->period_type == INTEL_PT_PERIOD_MTC)
1668 decoder->state.type |= INTEL_PT_INSTRUCTION;
1669 break;
1671 case INTEL_PT_CYC:
1672 intel_pt_calc_cyc_timestamp(decoder);
1673 break;
1675 case INTEL_PT_MODE_EXEC:
1676 decoder->exec_mode = decoder->packet.payload;
1677 break;
1679 case INTEL_PT_VMCS:
1680 case INTEL_PT_MNT:
1681 case INTEL_PT_PAD:
1682 break;
1684 default:
1685 return intel_pt_bug(decoder);
1690 static int intel_pt_walk_trace(struct intel_pt_decoder *decoder)
1692 bool no_tip = false;
1693 int err;
1695 while (1) {
1696 err = intel_pt_get_next_packet(decoder);
1697 if (err)
1698 return err;
1699 next:
1700 switch (decoder->packet.type) {
1701 case INTEL_PT_TNT:
1702 if (!decoder->packet.count)
1703 break;
1704 decoder->tnt = decoder->packet;
1705 decoder->pkt_state = INTEL_PT_STATE_TNT;
1706 err = intel_pt_walk_tnt(decoder);
1707 if (err == -EAGAIN)
1708 break;
1709 return err;
1711 case INTEL_PT_TIP_PGD:
1712 if (decoder->packet.count != 0)
1713 intel_pt_set_last_ip(decoder);
1714 decoder->pkt_state = INTEL_PT_STATE_TIP_PGD;
1715 return intel_pt_walk_tip(decoder);
1717 case INTEL_PT_TIP_PGE: {
1718 decoder->pge = true;
1719 if (decoder->packet.count == 0) {
1720 intel_pt_log_at("Skipping zero TIP.PGE",
1721 decoder->pos);
1722 break;
1724 intel_pt_set_ip(decoder);
1725 decoder->state.from_ip = 0;
1726 decoder->state.to_ip = decoder->ip;
1727 return 0;
1730 case INTEL_PT_OVF:
1731 return intel_pt_overflow(decoder);
1733 case INTEL_PT_TIP:
1734 if (decoder->packet.count != 0)
1735 intel_pt_set_last_ip(decoder);
1736 decoder->pkt_state = INTEL_PT_STATE_TIP;
1737 return intel_pt_walk_tip(decoder);
1739 case INTEL_PT_FUP:
1740 if (decoder->packet.count == 0) {
1741 intel_pt_log_at("Skipping zero FUP",
1742 decoder->pos);
1743 no_tip = false;
1744 break;
1746 intel_pt_set_last_ip(decoder);
1747 if (!decoder->branch_enable) {
1748 decoder->ip = decoder->last_ip;
1749 if (intel_pt_fup_event(decoder))
1750 return 0;
1751 no_tip = false;
1752 break;
1754 if (decoder->set_fup_mwait)
1755 no_tip = true;
1756 err = intel_pt_walk_fup(decoder);
1757 if (err != -EAGAIN) {
1758 if (err)
1759 return err;
1760 if (no_tip)
1761 decoder->pkt_state =
1762 INTEL_PT_STATE_FUP_NO_TIP;
1763 else
1764 decoder->pkt_state = INTEL_PT_STATE_FUP;
1765 return 0;
1767 if (no_tip) {
1768 no_tip = false;
1769 break;
1771 return intel_pt_walk_fup_tip(decoder);
1773 case INTEL_PT_TRACESTOP:
1774 decoder->pge = false;
1775 decoder->continuous_period = false;
1776 intel_pt_clear_tx_flags(decoder);
1777 decoder->have_tma = false;
1778 break;
1780 case INTEL_PT_PSB:
1781 decoder->last_ip = 0;
1782 decoder->have_last_ip = true;
1783 intel_pt_clear_stack(&decoder->stack);
1784 err = intel_pt_walk_psbend(decoder);
1785 if (err == -EAGAIN)
1786 goto next;
1787 if (err)
1788 return err;
1789 break;
1791 case INTEL_PT_PIP:
1792 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
1793 break;
1795 case INTEL_PT_MTC:
1796 intel_pt_calc_mtc_timestamp(decoder);
1797 if (decoder->period_type != INTEL_PT_PERIOD_MTC)
1798 break;
1800 * Ensure that there has been an instruction since the
1801 * last MTC.
1803 if (!decoder->mtc_insn)
1804 break;
1805 decoder->mtc_insn = false;
1806 /* Ensure that there is a timestamp */
1807 if (!decoder->timestamp)
1808 break;
1809 decoder->state.type = INTEL_PT_INSTRUCTION;
1810 decoder->state.from_ip = decoder->ip;
1811 decoder->state.to_ip = 0;
1812 decoder->mtc_insn = false;
1813 return 0;
1815 case INTEL_PT_TSC:
1816 intel_pt_calc_tsc_timestamp(decoder);
1817 break;
1819 case INTEL_PT_TMA:
1820 intel_pt_calc_tma(decoder);
1821 break;
1823 case INTEL_PT_CYC:
1824 intel_pt_calc_cyc_timestamp(decoder);
1825 break;
1827 case INTEL_PT_CBR:
1828 intel_pt_calc_cbr(decoder);
1829 if (!decoder->branch_enable &&
1830 decoder->cbr != decoder->cbr_seen) {
1831 decoder->cbr_seen = decoder->cbr;
1832 decoder->state.type = INTEL_PT_CBR_CHG;
1833 decoder->state.from_ip = decoder->ip;
1834 decoder->state.to_ip = 0;
1835 decoder->state.cbr_payload =
1836 decoder->packet.payload;
1837 return 0;
1839 break;
1841 case INTEL_PT_MODE_EXEC:
1842 decoder->exec_mode = decoder->packet.payload;
1843 break;
1845 case INTEL_PT_MODE_TSX:
1846 /* MODE_TSX need not be followed by FUP */
1847 if (!decoder->pge) {
1848 intel_pt_update_in_tx(decoder);
1849 break;
1851 err = intel_pt_mode_tsx(decoder, &no_tip);
1852 if (err)
1853 return err;
1854 goto next;
1856 case INTEL_PT_BAD: /* Does not happen */
1857 return intel_pt_bug(decoder);
1859 case INTEL_PT_PSBEND:
1860 case INTEL_PT_VMCS:
1861 case INTEL_PT_MNT:
1862 case INTEL_PT_PAD:
1863 break;
1865 case INTEL_PT_PTWRITE_IP:
1866 decoder->fup_ptw_payload = decoder->packet.payload;
1867 err = intel_pt_get_next_packet(decoder);
1868 if (err)
1869 return err;
1870 if (decoder->packet.type == INTEL_PT_FUP) {
1871 decoder->set_fup_ptw = true;
1872 no_tip = true;
1873 } else {
1874 intel_pt_log_at("ERROR: Missing FUP after PTWRITE",
1875 decoder->pos);
1877 goto next;
1879 case INTEL_PT_PTWRITE:
1880 decoder->state.type = INTEL_PT_PTW;
1881 decoder->state.from_ip = decoder->ip;
1882 decoder->state.to_ip = 0;
1883 decoder->state.ptw_payload = decoder->packet.payload;
1884 return 0;
1886 case INTEL_PT_MWAIT:
1887 decoder->fup_mwait_payload = decoder->packet.payload;
1888 decoder->set_fup_mwait = true;
1889 break;
1891 case INTEL_PT_PWRE:
1892 if (decoder->set_fup_mwait) {
1893 decoder->fup_pwre_payload =
1894 decoder->packet.payload;
1895 decoder->set_fup_pwre = true;
1896 break;
1898 decoder->state.type = INTEL_PT_PWR_ENTRY;
1899 decoder->state.from_ip = decoder->ip;
1900 decoder->state.to_ip = 0;
1901 decoder->state.pwrx_payload = decoder->packet.payload;
1902 return 0;
1904 case INTEL_PT_EXSTOP_IP:
1905 err = intel_pt_get_next_packet(decoder);
1906 if (err)
1907 return err;
1908 if (decoder->packet.type == INTEL_PT_FUP) {
1909 decoder->set_fup_exstop = true;
1910 no_tip = true;
1911 } else {
1912 intel_pt_log_at("ERROR: Missing FUP after EXSTOP",
1913 decoder->pos);
1915 goto next;
1917 case INTEL_PT_EXSTOP:
1918 decoder->state.type = INTEL_PT_EX_STOP;
1919 decoder->state.from_ip = decoder->ip;
1920 decoder->state.to_ip = 0;
1921 return 0;
1923 case INTEL_PT_PWRX:
1924 decoder->state.type = INTEL_PT_PWR_EXIT;
1925 decoder->state.from_ip = decoder->ip;
1926 decoder->state.to_ip = 0;
1927 decoder->state.pwrx_payload = decoder->packet.payload;
1928 return 0;
1930 default:
1931 return intel_pt_bug(decoder);
1936 static inline bool intel_pt_have_ip(struct intel_pt_decoder *decoder)
1938 return decoder->packet.count &&
1939 (decoder->have_last_ip || decoder->packet.count == 3 ||
1940 decoder->packet.count == 6);
1943 /* Walk PSB+ packets to get in sync. */
1944 static int intel_pt_walk_psb(struct intel_pt_decoder *decoder)
1946 int err;
1948 while (1) {
1949 err = intel_pt_get_next_packet(decoder);
1950 if (err)
1951 return err;
1953 switch (decoder->packet.type) {
1954 case INTEL_PT_TIP_PGD:
1955 decoder->continuous_period = false;
1956 __fallthrough;
1957 case INTEL_PT_TIP_PGE:
1958 case INTEL_PT_TIP:
1959 case INTEL_PT_PTWRITE:
1960 case INTEL_PT_PTWRITE_IP:
1961 case INTEL_PT_EXSTOP:
1962 case INTEL_PT_EXSTOP_IP:
1963 case INTEL_PT_MWAIT:
1964 case INTEL_PT_PWRE:
1965 case INTEL_PT_PWRX:
1966 intel_pt_log("ERROR: Unexpected packet\n");
1967 return -ENOENT;
1969 case INTEL_PT_FUP:
1970 decoder->pge = true;
1971 if (intel_pt_have_ip(decoder)) {
1972 uint64_t current_ip = decoder->ip;
1974 intel_pt_set_ip(decoder);
1975 if (current_ip)
1976 intel_pt_log_to("Setting IP",
1977 decoder->ip);
1979 break;
1981 case INTEL_PT_MTC:
1982 intel_pt_calc_mtc_timestamp(decoder);
1983 break;
1985 case INTEL_PT_TSC:
1986 intel_pt_calc_tsc_timestamp(decoder);
1987 break;
1989 case INTEL_PT_TMA:
1990 intel_pt_calc_tma(decoder);
1991 break;
1993 case INTEL_PT_CYC:
1994 intel_pt_calc_cyc_timestamp(decoder);
1995 break;
1997 case INTEL_PT_CBR:
1998 intel_pt_calc_cbr(decoder);
1999 break;
2001 case INTEL_PT_PIP:
2002 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2003 break;
2005 case INTEL_PT_MODE_EXEC:
2006 decoder->exec_mode = decoder->packet.payload;
2007 break;
2009 case INTEL_PT_MODE_TSX:
2010 intel_pt_update_in_tx(decoder);
2011 break;
2013 case INTEL_PT_TRACESTOP:
2014 decoder->pge = false;
2015 decoder->continuous_period = false;
2016 intel_pt_clear_tx_flags(decoder);
2017 __fallthrough;
2019 case INTEL_PT_TNT:
2020 decoder->have_tma = false;
2021 intel_pt_log("ERROR: Unexpected packet\n");
2022 if (decoder->ip)
2023 decoder->pkt_state = INTEL_PT_STATE_ERR4;
2024 else
2025 decoder->pkt_state = INTEL_PT_STATE_ERR3;
2026 return -ENOENT;
2028 case INTEL_PT_BAD: /* Does not happen */
2029 return intel_pt_bug(decoder);
2031 case INTEL_PT_OVF:
2032 return intel_pt_overflow(decoder);
2034 case INTEL_PT_PSBEND:
2035 return 0;
2037 case INTEL_PT_PSB:
2038 case INTEL_PT_VMCS:
2039 case INTEL_PT_MNT:
2040 case INTEL_PT_PAD:
2041 default:
2042 break;
2047 static int intel_pt_walk_to_ip(struct intel_pt_decoder *decoder)
2049 int err;
2051 while (1) {
2052 err = intel_pt_get_next_packet(decoder);
2053 if (err)
2054 return err;
2056 switch (decoder->packet.type) {
2057 case INTEL_PT_TIP_PGD:
2058 decoder->continuous_period = false;
2059 __fallthrough;
2060 case INTEL_PT_TIP_PGE:
2061 case INTEL_PT_TIP:
2062 decoder->pge = decoder->packet.type != INTEL_PT_TIP_PGD;
2063 if (intel_pt_have_ip(decoder))
2064 intel_pt_set_ip(decoder);
2065 if (decoder->ip)
2066 return 0;
2067 break;
2069 case INTEL_PT_FUP:
2070 if (intel_pt_have_ip(decoder))
2071 intel_pt_set_ip(decoder);
2072 if (decoder->ip)
2073 return 0;
2074 break;
2076 case INTEL_PT_MTC:
2077 intel_pt_calc_mtc_timestamp(decoder);
2078 break;
2080 case INTEL_PT_TSC:
2081 intel_pt_calc_tsc_timestamp(decoder);
2082 break;
2084 case INTEL_PT_TMA:
2085 intel_pt_calc_tma(decoder);
2086 break;
2088 case INTEL_PT_CYC:
2089 intel_pt_calc_cyc_timestamp(decoder);
2090 break;
2092 case INTEL_PT_CBR:
2093 intel_pt_calc_cbr(decoder);
2094 break;
2096 case INTEL_PT_PIP:
2097 decoder->cr3 = decoder->packet.payload & (BIT63 - 1);
2098 break;
2100 case INTEL_PT_MODE_EXEC:
2101 decoder->exec_mode = decoder->packet.payload;
2102 break;
2104 case INTEL_PT_MODE_TSX:
2105 intel_pt_update_in_tx(decoder);
2106 break;
2108 case INTEL_PT_OVF:
2109 return intel_pt_overflow(decoder);
2111 case INTEL_PT_BAD: /* Does not happen */
2112 return intel_pt_bug(decoder);
2114 case INTEL_PT_TRACESTOP:
2115 decoder->pge = false;
2116 decoder->continuous_period = false;
2117 intel_pt_clear_tx_flags(decoder);
2118 decoder->have_tma = false;
2119 break;
2121 case INTEL_PT_PSB:
2122 decoder->last_ip = 0;
2123 decoder->have_last_ip = true;
2124 intel_pt_clear_stack(&decoder->stack);
2125 err = intel_pt_walk_psb(decoder);
2126 if (err)
2127 return err;
2128 if (decoder->ip) {
2129 /* Do not have a sample */
2130 decoder->state.type = 0;
2131 return 0;
2133 break;
2135 case INTEL_PT_TNT:
2136 case INTEL_PT_PSBEND:
2137 case INTEL_PT_VMCS:
2138 case INTEL_PT_MNT:
2139 case INTEL_PT_PAD:
2140 case INTEL_PT_PTWRITE:
2141 case INTEL_PT_PTWRITE_IP:
2142 case INTEL_PT_EXSTOP:
2143 case INTEL_PT_EXSTOP_IP:
2144 case INTEL_PT_MWAIT:
2145 case INTEL_PT_PWRE:
2146 case INTEL_PT_PWRX:
2147 default:
2148 break;
2153 static int intel_pt_sync_ip(struct intel_pt_decoder *decoder)
2155 int err;
2157 decoder->set_fup_tx_flags = false;
2158 decoder->set_fup_ptw = false;
2159 decoder->set_fup_mwait = false;
2160 decoder->set_fup_pwre = false;
2161 decoder->set_fup_exstop = false;
2163 if (!decoder->branch_enable) {
2164 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2165 decoder->overflow = false;
2166 decoder->state.type = 0; /* Do not have a sample */
2167 return 0;
2170 intel_pt_log("Scanning for full IP\n");
2171 err = intel_pt_walk_to_ip(decoder);
2172 if (err)
2173 return err;
2175 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2176 decoder->overflow = false;
2178 decoder->state.from_ip = 0;
2179 decoder->state.to_ip = decoder->ip;
2180 intel_pt_log_to("Setting IP", decoder->ip);
2182 return 0;
2185 static int intel_pt_part_psb(struct intel_pt_decoder *decoder)
2187 const unsigned char *end = decoder->buf + decoder->len;
2188 size_t i;
2190 for (i = INTEL_PT_PSB_LEN - 1; i; i--) {
2191 if (i > decoder->len)
2192 continue;
2193 if (!memcmp(end - i, INTEL_PT_PSB_STR, i))
2194 return i;
2196 return 0;
2199 static int intel_pt_rest_psb(struct intel_pt_decoder *decoder, int part_psb)
2201 size_t rest_psb = INTEL_PT_PSB_LEN - part_psb;
2202 const char *psb = INTEL_PT_PSB_STR;
2204 if (rest_psb > decoder->len ||
2205 memcmp(decoder->buf, psb + part_psb, rest_psb))
2206 return 0;
2208 return rest_psb;
2211 static int intel_pt_get_split_psb(struct intel_pt_decoder *decoder,
2212 int part_psb)
2214 int rest_psb, ret;
2216 decoder->pos += decoder->len;
2217 decoder->len = 0;
2219 ret = intel_pt_get_next_data(decoder);
2220 if (ret)
2221 return ret;
2223 rest_psb = intel_pt_rest_psb(decoder, part_psb);
2224 if (!rest_psb)
2225 return 0;
2227 decoder->pos -= part_psb;
2228 decoder->next_buf = decoder->buf + rest_psb;
2229 decoder->next_len = decoder->len - rest_psb;
2230 memcpy(decoder->temp_buf, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2231 decoder->buf = decoder->temp_buf;
2232 decoder->len = INTEL_PT_PSB_LEN;
2234 return 0;
2237 static int intel_pt_scan_for_psb(struct intel_pt_decoder *decoder)
2239 unsigned char *next;
2240 int ret;
2242 intel_pt_log("Scanning for PSB\n");
2243 while (1) {
2244 if (!decoder->len) {
2245 ret = intel_pt_get_next_data(decoder);
2246 if (ret)
2247 return ret;
2250 next = memmem(decoder->buf, decoder->len, INTEL_PT_PSB_STR,
2251 INTEL_PT_PSB_LEN);
2252 if (!next) {
2253 int part_psb;
2255 part_psb = intel_pt_part_psb(decoder);
2256 if (part_psb) {
2257 ret = intel_pt_get_split_psb(decoder, part_psb);
2258 if (ret)
2259 return ret;
2260 } else {
2261 decoder->pos += decoder->len;
2262 decoder->len = 0;
2264 continue;
2267 decoder->pkt_step = next - decoder->buf;
2268 return intel_pt_get_next_packet(decoder);
2272 static int intel_pt_sync(struct intel_pt_decoder *decoder)
2274 int err;
2276 decoder->pge = false;
2277 decoder->continuous_period = false;
2278 decoder->have_last_ip = false;
2279 decoder->last_ip = 0;
2280 decoder->ip = 0;
2281 intel_pt_clear_stack(&decoder->stack);
2283 err = intel_pt_scan_for_psb(decoder);
2284 if (err)
2285 return err;
2287 decoder->have_last_ip = true;
2288 decoder->pkt_state = INTEL_PT_STATE_NO_IP;
2290 err = intel_pt_walk_psb(decoder);
2291 if (err)
2292 return err;
2294 if (decoder->ip) {
2295 decoder->state.type = 0; /* Do not have a sample */
2296 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2297 } else {
2298 return intel_pt_sync_ip(decoder);
2301 return 0;
2304 static uint64_t intel_pt_est_timestamp(struct intel_pt_decoder *decoder)
2306 uint64_t est = decoder->sample_insn_cnt << 1;
2308 if (!decoder->cbr || !decoder->max_non_turbo_ratio)
2309 goto out;
2311 est *= decoder->max_non_turbo_ratio;
2312 est /= decoder->cbr;
2313 out:
2314 return decoder->sample_timestamp + est;
2317 const struct intel_pt_state *intel_pt_decode(struct intel_pt_decoder *decoder)
2319 int err;
2321 do {
2322 decoder->state.type = INTEL_PT_BRANCH;
2323 decoder->state.flags = 0;
2325 switch (decoder->pkt_state) {
2326 case INTEL_PT_STATE_NO_PSB:
2327 err = intel_pt_sync(decoder);
2328 break;
2329 case INTEL_PT_STATE_NO_IP:
2330 decoder->have_last_ip = false;
2331 decoder->last_ip = 0;
2332 decoder->ip = 0;
2333 __fallthrough;
2334 case INTEL_PT_STATE_ERR_RESYNC:
2335 err = intel_pt_sync_ip(decoder);
2336 break;
2337 case INTEL_PT_STATE_IN_SYNC:
2338 err = intel_pt_walk_trace(decoder);
2339 break;
2340 case INTEL_PT_STATE_TNT:
2341 err = intel_pt_walk_tnt(decoder);
2342 if (err == -EAGAIN)
2343 err = intel_pt_walk_trace(decoder);
2344 break;
2345 case INTEL_PT_STATE_TIP:
2346 case INTEL_PT_STATE_TIP_PGD:
2347 err = intel_pt_walk_tip(decoder);
2348 break;
2349 case INTEL_PT_STATE_FUP:
2350 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2351 err = intel_pt_walk_fup(decoder);
2352 if (err == -EAGAIN)
2353 err = intel_pt_walk_fup_tip(decoder);
2354 else if (!err)
2355 decoder->pkt_state = INTEL_PT_STATE_FUP;
2356 break;
2357 case INTEL_PT_STATE_FUP_NO_TIP:
2358 decoder->pkt_state = INTEL_PT_STATE_IN_SYNC;
2359 err = intel_pt_walk_fup(decoder);
2360 if (err == -EAGAIN)
2361 err = intel_pt_walk_trace(decoder);
2362 break;
2363 default:
2364 err = intel_pt_bug(decoder);
2365 break;
2367 } while (err == -ENOLINK);
2369 if (err) {
2370 decoder->state.err = intel_pt_ext_err(err);
2371 decoder->state.from_ip = decoder->ip;
2372 decoder->sample_timestamp = decoder->timestamp;
2373 decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
2374 } else {
2375 decoder->state.err = 0;
2376 if (decoder->cbr != decoder->cbr_seen && decoder->state.type) {
2377 decoder->cbr_seen = decoder->cbr;
2378 decoder->state.type |= INTEL_PT_CBR_CHG;
2379 decoder->state.cbr_payload = decoder->cbr_payload;
2381 if (intel_pt_sample_time(decoder->pkt_state)) {
2382 decoder->sample_timestamp = decoder->timestamp;
2383 decoder->sample_insn_cnt = decoder->timestamp_insn_cnt;
2387 decoder->state.timestamp = decoder->sample_timestamp;
2388 decoder->state.est_timestamp = intel_pt_est_timestamp(decoder);
2389 decoder->state.cr3 = decoder->cr3;
2390 decoder->state.tot_insn_cnt = decoder->tot_insn_cnt;
2392 return &decoder->state;
2396 * intel_pt_next_psb - move buffer pointer to the start of the next PSB packet.
2397 * @buf: pointer to buffer pointer
2398 * @len: size of buffer
2400 * Updates the buffer pointer to point to the start of the next PSB packet if
2401 * there is one, otherwise the buffer pointer is unchanged. If @buf is updated,
2402 * @len is adjusted accordingly.
2404 * Return: %true if a PSB packet is found, %false otherwise.
2406 static bool intel_pt_next_psb(unsigned char **buf, size_t *len)
2408 unsigned char *next;
2410 next = memmem(*buf, *len, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2411 if (next) {
2412 *len -= next - *buf;
2413 *buf = next;
2414 return true;
2416 return false;
2420 * intel_pt_step_psb - move buffer pointer to the start of the following PSB
2421 * packet.
2422 * @buf: pointer to buffer pointer
2423 * @len: size of buffer
2425 * Updates the buffer pointer to point to the start of the following PSB packet
2426 * (skipping the PSB at @buf itself) if there is one, otherwise the buffer
2427 * pointer is unchanged. If @buf is updated, @len is adjusted accordingly.
2429 * Return: %true if a PSB packet is found, %false otherwise.
2431 static bool intel_pt_step_psb(unsigned char **buf, size_t *len)
2433 unsigned char *next;
2435 if (!*len)
2436 return false;
2438 next = memmem(*buf + 1, *len - 1, INTEL_PT_PSB_STR, INTEL_PT_PSB_LEN);
2439 if (next) {
2440 *len -= next - *buf;
2441 *buf = next;
2442 return true;
2444 return false;
2448 * intel_pt_last_psb - find the last PSB packet in a buffer.
2449 * @buf: buffer
2450 * @len: size of buffer
2452 * This function finds the last PSB in a buffer.
2454 * Return: A pointer to the last PSB in @buf if found, %NULL otherwise.
2456 static unsigned char *intel_pt_last_psb(unsigned char *buf, size_t len)
2458 const char *n = INTEL_PT_PSB_STR;
2459 unsigned char *p;
2460 size_t k;
2462 if (len < INTEL_PT_PSB_LEN)
2463 return NULL;
2465 k = len - INTEL_PT_PSB_LEN + 1;
2466 while (1) {
2467 p = memrchr(buf, n[0], k);
2468 if (!p)
2469 return NULL;
2470 if (!memcmp(p + 1, n + 1, INTEL_PT_PSB_LEN - 1))
2471 return p;
2472 k = p - buf;
2473 if (!k)
2474 return NULL;
2479 * intel_pt_next_tsc - find and return next TSC.
2480 * @buf: buffer
2481 * @len: size of buffer
2482 * @tsc: TSC value returned
2483 * @rem: returns remaining size when TSC is found
2485 * Find a TSC packet in @buf and return the TSC value. This function assumes
2486 * that @buf starts at a PSB and that PSB+ will contain TSC and so stops if a
2487 * PSBEND packet is found.
2489 * Return: %true if TSC is found, false otherwise.
2491 static bool intel_pt_next_tsc(unsigned char *buf, size_t len, uint64_t *tsc,
2492 size_t *rem)
2494 struct intel_pt_pkt packet;
2495 int ret;
2497 while (len) {
2498 ret = intel_pt_get_packet(buf, len, &packet);
2499 if (ret <= 0)
2500 return false;
2501 if (packet.type == INTEL_PT_TSC) {
2502 *tsc = packet.payload;
2503 *rem = len;
2504 return true;
2506 if (packet.type == INTEL_PT_PSBEND)
2507 return false;
2508 buf += ret;
2509 len -= ret;
2511 return false;
2515 * intel_pt_tsc_cmp - compare 7-byte TSCs.
2516 * @tsc1: first TSC to compare
2517 * @tsc2: second TSC to compare
2519 * This function compares 7-byte TSC values allowing for the possibility that
2520 * TSC wrapped around. Generally it is not possible to know if TSC has wrapped
2521 * around so for that purpose this function assumes the absolute difference is
2522 * less than half the maximum difference.
2524 * Return: %-1 if @tsc1 is before @tsc2, %0 if @tsc1 == @tsc2, %1 if @tsc1 is
2525 * after @tsc2.
2527 static int intel_pt_tsc_cmp(uint64_t tsc1, uint64_t tsc2)
2529 const uint64_t halfway = (1ULL << 55);
2531 if (tsc1 == tsc2)
2532 return 0;
2534 if (tsc1 < tsc2) {
2535 if (tsc2 - tsc1 < halfway)
2536 return -1;
2537 else
2538 return 1;
2539 } else {
2540 if (tsc1 - tsc2 < halfway)
2541 return 1;
2542 else
2543 return -1;
2548 * intel_pt_find_overlap_tsc - determine start of non-overlapped trace data
2549 * using TSC.
2550 * @buf_a: first buffer
2551 * @len_a: size of first buffer
2552 * @buf_b: second buffer
2553 * @len_b: size of second buffer
2554 * @consecutive: returns true if there is data in buf_b that is consecutive
2555 * to buf_a
2557 * If the trace contains TSC we can look at the last TSC of @buf_a and the
2558 * first TSC of @buf_b in order to determine if the buffers overlap, and then
2559 * walk forward in @buf_b until a later TSC is found. A precondition is that
2560 * @buf_a and @buf_b are positioned at a PSB.
2562 * Return: A pointer into @buf_b from where non-overlapped data starts, or
2563 * @buf_b + @len_b if there is no non-overlapped data.
2565 static unsigned char *intel_pt_find_overlap_tsc(unsigned char *buf_a,
2566 size_t len_a,
2567 unsigned char *buf_b,
2568 size_t len_b, bool *consecutive)
2570 uint64_t tsc_a, tsc_b;
2571 unsigned char *p;
2572 size_t len, rem_a, rem_b;
2574 p = intel_pt_last_psb(buf_a, len_a);
2575 if (!p)
2576 return buf_b; /* No PSB in buf_a => no overlap */
2578 len = len_a - (p - buf_a);
2579 if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a)) {
2580 /* The last PSB+ in buf_a is incomplete, so go back one more */
2581 len_a -= len;
2582 p = intel_pt_last_psb(buf_a, len_a);
2583 if (!p)
2584 return buf_b; /* No full PSB+ => assume no overlap */
2585 len = len_a - (p - buf_a);
2586 if (!intel_pt_next_tsc(p, len, &tsc_a, &rem_a))
2587 return buf_b; /* No TSC in buf_a => assume no overlap */
2590 while (1) {
2591 /* Ignore PSB+ with no TSC */
2592 if (intel_pt_next_tsc(buf_b, len_b, &tsc_b, &rem_b)) {
2593 int cmp = intel_pt_tsc_cmp(tsc_a, tsc_b);
2595 /* Same TSC, so buffers are consecutive */
2596 if (!cmp && rem_b >= rem_a) {
2597 *consecutive = true;
2598 return buf_b + len_b - (rem_b - rem_a);
2600 if (cmp < 0)
2601 return buf_b; /* tsc_a < tsc_b => no overlap */
2604 if (!intel_pt_step_psb(&buf_b, &len_b))
2605 return buf_b + len_b; /* No PSB in buf_b => no data */
2610 * intel_pt_find_overlap - determine start of non-overlapped trace data.
2611 * @buf_a: first buffer
2612 * @len_a: size of first buffer
2613 * @buf_b: second buffer
2614 * @len_b: size of second buffer
2615 * @have_tsc: can use TSC packets to detect overlap
2616 * @consecutive: returns true if there is data in buf_b that is consecutive
2617 * to buf_a
2619 * When trace samples or snapshots are recorded there is the possibility that
2620 * the data overlaps. Note that, for the purposes of decoding, data is only
2621 * useful if it begins with a PSB packet.
2623 * Return: A pointer into @buf_b from where non-overlapped data starts, or
2624 * @buf_b + @len_b if there is no non-overlapped data.
2626 unsigned char *intel_pt_find_overlap(unsigned char *buf_a, size_t len_a,
2627 unsigned char *buf_b, size_t len_b,
2628 bool have_tsc, bool *consecutive)
2630 unsigned char *found;
2632 /* Buffer 'b' must start at PSB so throw away everything before that */
2633 if (!intel_pt_next_psb(&buf_b, &len_b))
2634 return buf_b + len_b; /* No PSB */
2636 if (!intel_pt_next_psb(&buf_a, &len_a))
2637 return buf_b; /* No overlap */
2639 if (have_tsc) {
2640 found = intel_pt_find_overlap_tsc(buf_a, len_a, buf_b, len_b,
2641 consecutive);
2642 if (found)
2643 return found;
2647 * Buffer 'b' cannot end within buffer 'a' so, for comparison purposes,
2648 * we can ignore the first part of buffer 'a'.
2650 while (len_b < len_a) {
2651 if (!intel_pt_step_psb(&buf_a, &len_a))
2652 return buf_b; /* No overlap */
2655 /* Now len_b >= len_a */
2656 while (1) {
2657 /* Potential overlap so check the bytes */
2658 found = memmem(buf_a, len_a, buf_b, len_a);
2659 if (found) {
2660 *consecutive = true;
2661 return buf_b + len_a;
2664 /* Try again at next PSB in buffer 'a' */
2665 if (!intel_pt_step_psb(&buf_a, &len_a))
2666 return buf_b; /* No overlap */