1*4882a593Smuzhiyun // SPDX-License-Identifier: GPL-2.0
2*4882a593Smuzhiyun /*
3*4882a593Smuzhiyun * Arm Statistical Profiling Extensions (SPE) support
4*4882a593Smuzhiyun * Copyright (c) 2017-2018, Arm Ltd.
5*4882a593Smuzhiyun */
6*4882a593Smuzhiyun
7*4882a593Smuzhiyun #include <byteswap.h>
8*4882a593Smuzhiyun #include <endian.h>
9*4882a593Smuzhiyun #include <errno.h>
10*4882a593Smuzhiyun #include <inttypes.h>
11*4882a593Smuzhiyun #include <linux/bitops.h>
12*4882a593Smuzhiyun #include <linux/kernel.h>
13*4882a593Smuzhiyun #include <linux/log2.h>
14*4882a593Smuzhiyun #include <linux/types.h>
15*4882a593Smuzhiyun #include <linux/zalloc.h>
16*4882a593Smuzhiyun #include <stdlib.h>
17*4882a593Smuzhiyun #include <unistd.h>
18*4882a593Smuzhiyun
19*4882a593Smuzhiyun #include "auxtrace.h"
20*4882a593Smuzhiyun #include "color.h"
21*4882a593Smuzhiyun #include "debug.h"
22*4882a593Smuzhiyun #include "evlist.h"
23*4882a593Smuzhiyun #include "evsel.h"
24*4882a593Smuzhiyun #include "machine.h"
25*4882a593Smuzhiyun #include "session.h"
26*4882a593Smuzhiyun #include "symbol.h"
27*4882a593Smuzhiyun #include "thread.h"
28*4882a593Smuzhiyun #include "thread-stack.h"
29*4882a593Smuzhiyun #include "tool.h"
30*4882a593Smuzhiyun #include "util/synthetic-events.h"
31*4882a593Smuzhiyun
32*4882a593Smuzhiyun #include "arm-spe.h"
33*4882a593Smuzhiyun #include "arm-spe-decoder/arm-spe-decoder.h"
34*4882a593Smuzhiyun #include "arm-spe-decoder/arm-spe-pkt-decoder.h"
35*4882a593Smuzhiyun
36*4882a593Smuzhiyun #define MAX_TIMESTAMP (~0ULL)
37*4882a593Smuzhiyun
38*4882a593Smuzhiyun struct arm_spe {
39*4882a593Smuzhiyun struct auxtrace auxtrace;
40*4882a593Smuzhiyun struct auxtrace_queues queues;
41*4882a593Smuzhiyun struct auxtrace_heap heap;
42*4882a593Smuzhiyun struct itrace_synth_opts synth_opts;
43*4882a593Smuzhiyun u32 auxtrace_type;
44*4882a593Smuzhiyun struct perf_session *session;
45*4882a593Smuzhiyun struct machine *machine;
46*4882a593Smuzhiyun u32 pmu_type;
47*4882a593Smuzhiyun
48*4882a593Smuzhiyun u8 timeless_decoding;
49*4882a593Smuzhiyun u8 data_queued;
50*4882a593Smuzhiyun
51*4882a593Smuzhiyun u64 sample_type;
52*4882a593Smuzhiyun u8 sample_flc;
53*4882a593Smuzhiyun u8 sample_llc;
54*4882a593Smuzhiyun u8 sample_tlb;
55*4882a593Smuzhiyun u8 sample_branch;
56*4882a593Smuzhiyun u8 sample_remote_access;
57*4882a593Smuzhiyun
58*4882a593Smuzhiyun u64 l1d_miss_id;
59*4882a593Smuzhiyun u64 l1d_access_id;
60*4882a593Smuzhiyun u64 llc_miss_id;
61*4882a593Smuzhiyun u64 llc_access_id;
62*4882a593Smuzhiyun u64 tlb_miss_id;
63*4882a593Smuzhiyun u64 tlb_access_id;
64*4882a593Smuzhiyun u64 branch_miss_id;
65*4882a593Smuzhiyun u64 remote_access_id;
66*4882a593Smuzhiyun
67*4882a593Smuzhiyun u64 kernel_start;
68*4882a593Smuzhiyun
69*4882a593Smuzhiyun unsigned long num_events;
70*4882a593Smuzhiyun };
71*4882a593Smuzhiyun
72*4882a593Smuzhiyun struct arm_spe_queue {
73*4882a593Smuzhiyun struct arm_spe *spe;
74*4882a593Smuzhiyun unsigned int queue_nr;
75*4882a593Smuzhiyun struct auxtrace_buffer *buffer;
76*4882a593Smuzhiyun struct auxtrace_buffer *old_buffer;
77*4882a593Smuzhiyun union perf_event *event_buf;
78*4882a593Smuzhiyun bool on_heap;
79*4882a593Smuzhiyun bool done;
80*4882a593Smuzhiyun pid_t pid;
81*4882a593Smuzhiyun pid_t tid;
82*4882a593Smuzhiyun int cpu;
83*4882a593Smuzhiyun struct arm_spe_decoder *decoder;
84*4882a593Smuzhiyun u64 time;
85*4882a593Smuzhiyun u64 timestamp;
86*4882a593Smuzhiyun struct thread *thread;
87*4882a593Smuzhiyun };
88*4882a593Smuzhiyun
arm_spe_dump(struct arm_spe * spe __maybe_unused,unsigned char * buf,size_t len)89*4882a593Smuzhiyun static void arm_spe_dump(struct arm_spe *spe __maybe_unused,
90*4882a593Smuzhiyun unsigned char *buf, size_t len)
91*4882a593Smuzhiyun {
92*4882a593Smuzhiyun struct arm_spe_pkt packet;
93*4882a593Smuzhiyun size_t pos = 0;
94*4882a593Smuzhiyun int ret, pkt_len, i;
95*4882a593Smuzhiyun char desc[ARM_SPE_PKT_DESC_MAX];
96*4882a593Smuzhiyun const char *color = PERF_COLOR_BLUE;
97*4882a593Smuzhiyun
98*4882a593Smuzhiyun color_fprintf(stdout, color,
99*4882a593Smuzhiyun ". ... ARM SPE data: size %zu bytes\n",
100*4882a593Smuzhiyun len);
101*4882a593Smuzhiyun
102*4882a593Smuzhiyun while (len) {
103*4882a593Smuzhiyun ret = arm_spe_get_packet(buf, len, &packet);
104*4882a593Smuzhiyun if (ret > 0)
105*4882a593Smuzhiyun pkt_len = ret;
106*4882a593Smuzhiyun else
107*4882a593Smuzhiyun pkt_len = 1;
108*4882a593Smuzhiyun printf(".");
109*4882a593Smuzhiyun color_fprintf(stdout, color, " %08x: ", pos);
110*4882a593Smuzhiyun for (i = 0; i < pkt_len; i++)
111*4882a593Smuzhiyun color_fprintf(stdout, color, " %02x", buf[i]);
112*4882a593Smuzhiyun for (; i < 16; i++)
113*4882a593Smuzhiyun color_fprintf(stdout, color, " ");
114*4882a593Smuzhiyun if (ret > 0) {
115*4882a593Smuzhiyun ret = arm_spe_pkt_desc(&packet, desc,
116*4882a593Smuzhiyun ARM_SPE_PKT_DESC_MAX);
117*4882a593Smuzhiyun if (ret > 0)
118*4882a593Smuzhiyun color_fprintf(stdout, color, " %s\n", desc);
119*4882a593Smuzhiyun } else {
120*4882a593Smuzhiyun color_fprintf(stdout, color, " Bad packet!\n");
121*4882a593Smuzhiyun }
122*4882a593Smuzhiyun pos += pkt_len;
123*4882a593Smuzhiyun buf += pkt_len;
124*4882a593Smuzhiyun len -= pkt_len;
125*4882a593Smuzhiyun }
126*4882a593Smuzhiyun }
127*4882a593Smuzhiyun
arm_spe_dump_event(struct arm_spe * spe,unsigned char * buf,size_t len)128*4882a593Smuzhiyun static void arm_spe_dump_event(struct arm_spe *spe, unsigned char *buf,
129*4882a593Smuzhiyun size_t len)
130*4882a593Smuzhiyun {
131*4882a593Smuzhiyun printf(".\n");
132*4882a593Smuzhiyun arm_spe_dump(spe, buf, len);
133*4882a593Smuzhiyun }
134*4882a593Smuzhiyun
arm_spe_get_trace(struct arm_spe_buffer * b,void * data)135*4882a593Smuzhiyun static int arm_spe_get_trace(struct arm_spe_buffer *b, void *data)
136*4882a593Smuzhiyun {
137*4882a593Smuzhiyun struct arm_spe_queue *speq = data;
138*4882a593Smuzhiyun struct auxtrace_buffer *buffer = speq->buffer;
139*4882a593Smuzhiyun struct auxtrace_buffer *old_buffer = speq->old_buffer;
140*4882a593Smuzhiyun struct auxtrace_queue *queue;
141*4882a593Smuzhiyun
142*4882a593Smuzhiyun queue = &speq->spe->queues.queue_array[speq->queue_nr];
143*4882a593Smuzhiyun
144*4882a593Smuzhiyun buffer = auxtrace_buffer__next(queue, buffer);
145*4882a593Smuzhiyun /* If no more data, drop the previous auxtrace_buffer and return */
146*4882a593Smuzhiyun if (!buffer) {
147*4882a593Smuzhiyun if (old_buffer)
148*4882a593Smuzhiyun auxtrace_buffer__drop_data(old_buffer);
149*4882a593Smuzhiyun b->len = 0;
150*4882a593Smuzhiyun return 0;
151*4882a593Smuzhiyun }
152*4882a593Smuzhiyun
153*4882a593Smuzhiyun speq->buffer = buffer;
154*4882a593Smuzhiyun
155*4882a593Smuzhiyun /* If the aux_buffer doesn't have data associated, try to load it */
156*4882a593Smuzhiyun if (!buffer->data) {
157*4882a593Smuzhiyun /* get the file desc associated with the perf data file */
158*4882a593Smuzhiyun int fd = perf_data__fd(speq->spe->session->data);
159*4882a593Smuzhiyun
160*4882a593Smuzhiyun buffer->data = auxtrace_buffer__get_data(buffer, fd);
161*4882a593Smuzhiyun if (!buffer->data)
162*4882a593Smuzhiyun return -ENOMEM;
163*4882a593Smuzhiyun }
164*4882a593Smuzhiyun
165*4882a593Smuzhiyun b->len = buffer->size;
166*4882a593Smuzhiyun b->buf = buffer->data;
167*4882a593Smuzhiyun
168*4882a593Smuzhiyun if (b->len) {
169*4882a593Smuzhiyun if (old_buffer)
170*4882a593Smuzhiyun auxtrace_buffer__drop_data(old_buffer);
171*4882a593Smuzhiyun speq->old_buffer = buffer;
172*4882a593Smuzhiyun } else {
173*4882a593Smuzhiyun auxtrace_buffer__drop_data(buffer);
174*4882a593Smuzhiyun return arm_spe_get_trace(b, data);
175*4882a593Smuzhiyun }
176*4882a593Smuzhiyun
177*4882a593Smuzhiyun return 0;
178*4882a593Smuzhiyun }
179*4882a593Smuzhiyun
arm_spe__alloc_queue(struct arm_spe * spe,unsigned int queue_nr)180*4882a593Smuzhiyun static struct arm_spe_queue *arm_spe__alloc_queue(struct arm_spe *spe,
181*4882a593Smuzhiyun unsigned int queue_nr)
182*4882a593Smuzhiyun {
183*4882a593Smuzhiyun struct arm_spe_params params = { .get_trace = 0, };
184*4882a593Smuzhiyun struct arm_spe_queue *speq;
185*4882a593Smuzhiyun
186*4882a593Smuzhiyun speq = zalloc(sizeof(*speq));
187*4882a593Smuzhiyun if (!speq)
188*4882a593Smuzhiyun return NULL;
189*4882a593Smuzhiyun
190*4882a593Smuzhiyun speq->event_buf = malloc(PERF_SAMPLE_MAX_SIZE);
191*4882a593Smuzhiyun if (!speq->event_buf)
192*4882a593Smuzhiyun goto out_free;
193*4882a593Smuzhiyun
194*4882a593Smuzhiyun speq->spe = spe;
195*4882a593Smuzhiyun speq->queue_nr = queue_nr;
196*4882a593Smuzhiyun speq->pid = -1;
197*4882a593Smuzhiyun speq->tid = -1;
198*4882a593Smuzhiyun speq->cpu = -1;
199*4882a593Smuzhiyun
200*4882a593Smuzhiyun /* params set */
201*4882a593Smuzhiyun params.get_trace = arm_spe_get_trace;
202*4882a593Smuzhiyun params.data = speq;
203*4882a593Smuzhiyun
204*4882a593Smuzhiyun /* create new decoder */
205*4882a593Smuzhiyun speq->decoder = arm_spe_decoder_new(¶ms);
206*4882a593Smuzhiyun if (!speq->decoder)
207*4882a593Smuzhiyun goto out_free;
208*4882a593Smuzhiyun
209*4882a593Smuzhiyun return speq;
210*4882a593Smuzhiyun
211*4882a593Smuzhiyun out_free:
212*4882a593Smuzhiyun zfree(&speq->event_buf);
213*4882a593Smuzhiyun free(speq);
214*4882a593Smuzhiyun
215*4882a593Smuzhiyun return NULL;
216*4882a593Smuzhiyun }
217*4882a593Smuzhiyun
arm_spe_cpumode(struct arm_spe * spe,u64 ip)218*4882a593Smuzhiyun static inline u8 arm_spe_cpumode(struct arm_spe *spe, u64 ip)
219*4882a593Smuzhiyun {
220*4882a593Smuzhiyun return ip >= spe->kernel_start ?
221*4882a593Smuzhiyun PERF_RECORD_MISC_KERNEL :
222*4882a593Smuzhiyun PERF_RECORD_MISC_USER;
223*4882a593Smuzhiyun }
224*4882a593Smuzhiyun
arm_spe_prep_sample(struct arm_spe * spe,struct arm_spe_queue * speq,union perf_event * event,struct perf_sample * sample)225*4882a593Smuzhiyun static void arm_spe_prep_sample(struct arm_spe *spe,
226*4882a593Smuzhiyun struct arm_spe_queue *speq,
227*4882a593Smuzhiyun union perf_event *event,
228*4882a593Smuzhiyun struct perf_sample *sample)
229*4882a593Smuzhiyun {
230*4882a593Smuzhiyun struct arm_spe_record *record = &speq->decoder->record;
231*4882a593Smuzhiyun
232*4882a593Smuzhiyun if (!spe->timeless_decoding)
233*4882a593Smuzhiyun sample->time = speq->timestamp;
234*4882a593Smuzhiyun
235*4882a593Smuzhiyun sample->ip = record->from_ip;
236*4882a593Smuzhiyun sample->cpumode = arm_spe_cpumode(spe, sample->ip);
237*4882a593Smuzhiyun sample->pid = speq->pid;
238*4882a593Smuzhiyun sample->tid = speq->tid;
239*4882a593Smuzhiyun sample->addr = record->to_ip;
240*4882a593Smuzhiyun sample->period = 1;
241*4882a593Smuzhiyun sample->cpu = speq->cpu;
242*4882a593Smuzhiyun
243*4882a593Smuzhiyun event->sample.header.type = PERF_RECORD_SAMPLE;
244*4882a593Smuzhiyun event->sample.header.misc = sample->cpumode;
245*4882a593Smuzhiyun event->sample.header.size = sizeof(struct perf_event_header);
246*4882a593Smuzhiyun }
247*4882a593Smuzhiyun
arm_spe__inject_event(union perf_event * event,struct perf_sample * sample,u64 type)248*4882a593Smuzhiyun static int arm_spe__inject_event(union perf_event *event, struct perf_sample *sample, u64 type)
249*4882a593Smuzhiyun {
250*4882a593Smuzhiyun event->header.size = perf_event__sample_event_size(sample, type, 0);
251*4882a593Smuzhiyun return perf_event__synthesize_sample(event, type, 0, sample);
252*4882a593Smuzhiyun }
253*4882a593Smuzhiyun
254*4882a593Smuzhiyun static inline int
arm_spe_deliver_synth_event(struct arm_spe * spe,struct arm_spe_queue * speq __maybe_unused,union perf_event * event,struct perf_sample * sample)255*4882a593Smuzhiyun arm_spe_deliver_synth_event(struct arm_spe *spe,
256*4882a593Smuzhiyun struct arm_spe_queue *speq __maybe_unused,
257*4882a593Smuzhiyun union perf_event *event,
258*4882a593Smuzhiyun struct perf_sample *sample)
259*4882a593Smuzhiyun {
260*4882a593Smuzhiyun int ret;
261*4882a593Smuzhiyun
262*4882a593Smuzhiyun if (spe->synth_opts.inject) {
263*4882a593Smuzhiyun ret = arm_spe__inject_event(event, sample, spe->sample_type);
264*4882a593Smuzhiyun if (ret)
265*4882a593Smuzhiyun return ret;
266*4882a593Smuzhiyun }
267*4882a593Smuzhiyun
268*4882a593Smuzhiyun ret = perf_session__deliver_synth_event(spe->session, event, sample);
269*4882a593Smuzhiyun if (ret)
270*4882a593Smuzhiyun pr_err("ARM SPE: failed to deliver event, error %d\n", ret);
271*4882a593Smuzhiyun
272*4882a593Smuzhiyun return ret;
273*4882a593Smuzhiyun }
274*4882a593Smuzhiyun
275*4882a593Smuzhiyun static int
arm_spe_synth_spe_events_sample(struct arm_spe_queue * speq,u64 spe_events_id)276*4882a593Smuzhiyun arm_spe_synth_spe_events_sample(struct arm_spe_queue *speq,
277*4882a593Smuzhiyun u64 spe_events_id)
278*4882a593Smuzhiyun {
279*4882a593Smuzhiyun struct arm_spe *spe = speq->spe;
280*4882a593Smuzhiyun union perf_event *event = speq->event_buf;
281*4882a593Smuzhiyun struct perf_sample sample = { .ip = 0, };
282*4882a593Smuzhiyun
283*4882a593Smuzhiyun arm_spe_prep_sample(spe, speq, event, &sample);
284*4882a593Smuzhiyun
285*4882a593Smuzhiyun sample.id = spe_events_id;
286*4882a593Smuzhiyun sample.stream_id = spe_events_id;
287*4882a593Smuzhiyun
288*4882a593Smuzhiyun return arm_spe_deliver_synth_event(spe, speq, event, &sample);
289*4882a593Smuzhiyun }
290*4882a593Smuzhiyun
arm_spe_sample(struct arm_spe_queue * speq)291*4882a593Smuzhiyun static int arm_spe_sample(struct arm_spe_queue *speq)
292*4882a593Smuzhiyun {
293*4882a593Smuzhiyun const struct arm_spe_record *record = &speq->decoder->record;
294*4882a593Smuzhiyun struct arm_spe *spe = speq->spe;
295*4882a593Smuzhiyun int err;
296*4882a593Smuzhiyun
297*4882a593Smuzhiyun if (spe->sample_flc) {
298*4882a593Smuzhiyun if (record->type & ARM_SPE_L1D_MISS) {
299*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(
300*4882a593Smuzhiyun speq, spe->l1d_miss_id);
301*4882a593Smuzhiyun if (err)
302*4882a593Smuzhiyun return err;
303*4882a593Smuzhiyun }
304*4882a593Smuzhiyun
305*4882a593Smuzhiyun if (record->type & ARM_SPE_L1D_ACCESS) {
306*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(
307*4882a593Smuzhiyun speq, spe->l1d_access_id);
308*4882a593Smuzhiyun if (err)
309*4882a593Smuzhiyun return err;
310*4882a593Smuzhiyun }
311*4882a593Smuzhiyun }
312*4882a593Smuzhiyun
313*4882a593Smuzhiyun if (spe->sample_llc) {
314*4882a593Smuzhiyun if (record->type & ARM_SPE_LLC_MISS) {
315*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(
316*4882a593Smuzhiyun speq, spe->llc_miss_id);
317*4882a593Smuzhiyun if (err)
318*4882a593Smuzhiyun return err;
319*4882a593Smuzhiyun }
320*4882a593Smuzhiyun
321*4882a593Smuzhiyun if (record->type & ARM_SPE_LLC_ACCESS) {
322*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(
323*4882a593Smuzhiyun speq, spe->llc_access_id);
324*4882a593Smuzhiyun if (err)
325*4882a593Smuzhiyun return err;
326*4882a593Smuzhiyun }
327*4882a593Smuzhiyun }
328*4882a593Smuzhiyun
329*4882a593Smuzhiyun if (spe->sample_tlb) {
330*4882a593Smuzhiyun if (record->type & ARM_SPE_TLB_MISS) {
331*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(
332*4882a593Smuzhiyun speq, spe->tlb_miss_id);
333*4882a593Smuzhiyun if (err)
334*4882a593Smuzhiyun return err;
335*4882a593Smuzhiyun }
336*4882a593Smuzhiyun
337*4882a593Smuzhiyun if (record->type & ARM_SPE_TLB_ACCESS) {
338*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(
339*4882a593Smuzhiyun speq, spe->tlb_access_id);
340*4882a593Smuzhiyun if (err)
341*4882a593Smuzhiyun return err;
342*4882a593Smuzhiyun }
343*4882a593Smuzhiyun }
344*4882a593Smuzhiyun
345*4882a593Smuzhiyun if (spe->sample_branch && (record->type & ARM_SPE_BRANCH_MISS)) {
346*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(speq,
347*4882a593Smuzhiyun spe->branch_miss_id);
348*4882a593Smuzhiyun if (err)
349*4882a593Smuzhiyun return err;
350*4882a593Smuzhiyun }
351*4882a593Smuzhiyun
352*4882a593Smuzhiyun if (spe->sample_remote_access &&
353*4882a593Smuzhiyun (record->type & ARM_SPE_REMOTE_ACCESS)) {
354*4882a593Smuzhiyun err = arm_spe_synth_spe_events_sample(speq,
355*4882a593Smuzhiyun spe->remote_access_id);
356*4882a593Smuzhiyun if (err)
357*4882a593Smuzhiyun return err;
358*4882a593Smuzhiyun }
359*4882a593Smuzhiyun
360*4882a593Smuzhiyun return 0;
361*4882a593Smuzhiyun }
362*4882a593Smuzhiyun
arm_spe_run_decoder(struct arm_spe_queue * speq,u64 * timestamp)363*4882a593Smuzhiyun static int arm_spe_run_decoder(struct arm_spe_queue *speq, u64 *timestamp)
364*4882a593Smuzhiyun {
365*4882a593Smuzhiyun struct arm_spe *spe = speq->spe;
366*4882a593Smuzhiyun int ret;
367*4882a593Smuzhiyun
368*4882a593Smuzhiyun if (!spe->kernel_start)
369*4882a593Smuzhiyun spe->kernel_start = machine__kernel_start(spe->machine);
370*4882a593Smuzhiyun
371*4882a593Smuzhiyun while (1) {
372*4882a593Smuzhiyun ret = arm_spe_decode(speq->decoder);
373*4882a593Smuzhiyun if (!ret) {
374*4882a593Smuzhiyun pr_debug("No data or all data has been processed.\n");
375*4882a593Smuzhiyun return 1;
376*4882a593Smuzhiyun }
377*4882a593Smuzhiyun
378*4882a593Smuzhiyun /*
379*4882a593Smuzhiyun * Error is detected when decode SPE trace data, continue to
380*4882a593Smuzhiyun * the next trace data and find out more records.
381*4882a593Smuzhiyun */
382*4882a593Smuzhiyun if (ret < 0)
383*4882a593Smuzhiyun continue;
384*4882a593Smuzhiyun
385*4882a593Smuzhiyun ret = arm_spe_sample(speq);
386*4882a593Smuzhiyun if (ret)
387*4882a593Smuzhiyun return ret;
388*4882a593Smuzhiyun
389*4882a593Smuzhiyun if (!spe->timeless_decoding && speq->timestamp >= *timestamp) {
390*4882a593Smuzhiyun *timestamp = speq->timestamp;
391*4882a593Smuzhiyun return 0;
392*4882a593Smuzhiyun }
393*4882a593Smuzhiyun }
394*4882a593Smuzhiyun
395*4882a593Smuzhiyun return 0;
396*4882a593Smuzhiyun }
397*4882a593Smuzhiyun
arm_spe__setup_queue(struct arm_spe * spe,struct auxtrace_queue * queue,unsigned int queue_nr)398*4882a593Smuzhiyun static int arm_spe__setup_queue(struct arm_spe *spe,
399*4882a593Smuzhiyun struct auxtrace_queue *queue,
400*4882a593Smuzhiyun unsigned int queue_nr)
401*4882a593Smuzhiyun {
402*4882a593Smuzhiyun struct arm_spe_queue *speq = queue->priv;
403*4882a593Smuzhiyun struct arm_spe_record *record;
404*4882a593Smuzhiyun
405*4882a593Smuzhiyun if (list_empty(&queue->head) || speq)
406*4882a593Smuzhiyun return 0;
407*4882a593Smuzhiyun
408*4882a593Smuzhiyun speq = arm_spe__alloc_queue(spe, queue_nr);
409*4882a593Smuzhiyun
410*4882a593Smuzhiyun if (!speq)
411*4882a593Smuzhiyun return -ENOMEM;
412*4882a593Smuzhiyun
413*4882a593Smuzhiyun queue->priv = speq;
414*4882a593Smuzhiyun
415*4882a593Smuzhiyun if (queue->cpu != -1)
416*4882a593Smuzhiyun speq->cpu = queue->cpu;
417*4882a593Smuzhiyun
418*4882a593Smuzhiyun if (!speq->on_heap) {
419*4882a593Smuzhiyun int ret;
420*4882a593Smuzhiyun
421*4882a593Smuzhiyun if (spe->timeless_decoding)
422*4882a593Smuzhiyun return 0;
423*4882a593Smuzhiyun
424*4882a593Smuzhiyun retry:
425*4882a593Smuzhiyun ret = arm_spe_decode(speq->decoder);
426*4882a593Smuzhiyun
427*4882a593Smuzhiyun if (!ret)
428*4882a593Smuzhiyun return 0;
429*4882a593Smuzhiyun
430*4882a593Smuzhiyun if (ret < 0)
431*4882a593Smuzhiyun goto retry;
432*4882a593Smuzhiyun
433*4882a593Smuzhiyun record = &speq->decoder->record;
434*4882a593Smuzhiyun
435*4882a593Smuzhiyun speq->timestamp = record->timestamp;
436*4882a593Smuzhiyun ret = auxtrace_heap__add(&spe->heap, queue_nr, speq->timestamp);
437*4882a593Smuzhiyun if (ret)
438*4882a593Smuzhiyun return ret;
439*4882a593Smuzhiyun speq->on_heap = true;
440*4882a593Smuzhiyun }
441*4882a593Smuzhiyun
442*4882a593Smuzhiyun return 0;
443*4882a593Smuzhiyun }
444*4882a593Smuzhiyun
arm_spe__setup_queues(struct arm_spe * spe)445*4882a593Smuzhiyun static int arm_spe__setup_queues(struct arm_spe *spe)
446*4882a593Smuzhiyun {
447*4882a593Smuzhiyun unsigned int i;
448*4882a593Smuzhiyun int ret;
449*4882a593Smuzhiyun
450*4882a593Smuzhiyun for (i = 0; i < spe->queues.nr_queues; i++) {
451*4882a593Smuzhiyun ret = arm_spe__setup_queue(spe, &spe->queues.queue_array[i], i);
452*4882a593Smuzhiyun if (ret)
453*4882a593Smuzhiyun return ret;
454*4882a593Smuzhiyun }
455*4882a593Smuzhiyun
456*4882a593Smuzhiyun return 0;
457*4882a593Smuzhiyun }
458*4882a593Smuzhiyun
arm_spe__update_queues(struct arm_spe * spe)459*4882a593Smuzhiyun static int arm_spe__update_queues(struct arm_spe *spe)
460*4882a593Smuzhiyun {
461*4882a593Smuzhiyun if (spe->queues.new_data) {
462*4882a593Smuzhiyun spe->queues.new_data = false;
463*4882a593Smuzhiyun return arm_spe__setup_queues(spe);
464*4882a593Smuzhiyun }
465*4882a593Smuzhiyun
466*4882a593Smuzhiyun return 0;
467*4882a593Smuzhiyun }
468*4882a593Smuzhiyun
arm_spe__is_timeless_decoding(struct arm_spe * spe)469*4882a593Smuzhiyun static bool arm_spe__is_timeless_decoding(struct arm_spe *spe)
470*4882a593Smuzhiyun {
471*4882a593Smuzhiyun struct evsel *evsel;
472*4882a593Smuzhiyun struct evlist *evlist = spe->session->evlist;
473*4882a593Smuzhiyun bool timeless_decoding = true;
474*4882a593Smuzhiyun
475*4882a593Smuzhiyun /*
476*4882a593Smuzhiyun * Circle through the list of event and complain if we find one
477*4882a593Smuzhiyun * with the time bit set.
478*4882a593Smuzhiyun */
479*4882a593Smuzhiyun evlist__for_each_entry(evlist, evsel) {
480*4882a593Smuzhiyun if ((evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
481*4882a593Smuzhiyun timeless_decoding = false;
482*4882a593Smuzhiyun }
483*4882a593Smuzhiyun
484*4882a593Smuzhiyun return timeless_decoding;
485*4882a593Smuzhiyun }
486*4882a593Smuzhiyun
arm_spe_set_pid_tid_cpu(struct arm_spe * spe,struct auxtrace_queue * queue)487*4882a593Smuzhiyun static void arm_spe_set_pid_tid_cpu(struct arm_spe *spe,
488*4882a593Smuzhiyun struct auxtrace_queue *queue)
489*4882a593Smuzhiyun {
490*4882a593Smuzhiyun struct arm_spe_queue *speq = queue->priv;
491*4882a593Smuzhiyun pid_t tid;
492*4882a593Smuzhiyun
493*4882a593Smuzhiyun tid = machine__get_current_tid(spe->machine, speq->cpu);
494*4882a593Smuzhiyun if (tid != -1) {
495*4882a593Smuzhiyun speq->tid = tid;
496*4882a593Smuzhiyun thread__zput(speq->thread);
497*4882a593Smuzhiyun } else
498*4882a593Smuzhiyun speq->tid = queue->tid;
499*4882a593Smuzhiyun
500*4882a593Smuzhiyun if ((!speq->thread) && (speq->tid != -1)) {
501*4882a593Smuzhiyun speq->thread = machine__find_thread(spe->machine, -1,
502*4882a593Smuzhiyun speq->tid);
503*4882a593Smuzhiyun }
504*4882a593Smuzhiyun
505*4882a593Smuzhiyun if (speq->thread) {
506*4882a593Smuzhiyun speq->pid = speq->thread->pid_;
507*4882a593Smuzhiyun if (queue->cpu == -1)
508*4882a593Smuzhiyun speq->cpu = speq->thread->cpu;
509*4882a593Smuzhiyun }
510*4882a593Smuzhiyun }
511*4882a593Smuzhiyun
arm_spe_process_queues(struct arm_spe * spe,u64 timestamp)512*4882a593Smuzhiyun static int arm_spe_process_queues(struct arm_spe *spe, u64 timestamp)
513*4882a593Smuzhiyun {
514*4882a593Smuzhiyun unsigned int queue_nr;
515*4882a593Smuzhiyun u64 ts;
516*4882a593Smuzhiyun int ret;
517*4882a593Smuzhiyun
518*4882a593Smuzhiyun while (1) {
519*4882a593Smuzhiyun struct auxtrace_queue *queue;
520*4882a593Smuzhiyun struct arm_spe_queue *speq;
521*4882a593Smuzhiyun
522*4882a593Smuzhiyun if (!spe->heap.heap_cnt)
523*4882a593Smuzhiyun return 0;
524*4882a593Smuzhiyun
525*4882a593Smuzhiyun if (spe->heap.heap_array[0].ordinal >= timestamp)
526*4882a593Smuzhiyun return 0;
527*4882a593Smuzhiyun
528*4882a593Smuzhiyun queue_nr = spe->heap.heap_array[0].queue_nr;
529*4882a593Smuzhiyun queue = &spe->queues.queue_array[queue_nr];
530*4882a593Smuzhiyun speq = queue->priv;
531*4882a593Smuzhiyun
532*4882a593Smuzhiyun auxtrace_heap__pop(&spe->heap);
533*4882a593Smuzhiyun
534*4882a593Smuzhiyun if (spe->heap.heap_cnt) {
535*4882a593Smuzhiyun ts = spe->heap.heap_array[0].ordinal + 1;
536*4882a593Smuzhiyun if (ts > timestamp)
537*4882a593Smuzhiyun ts = timestamp;
538*4882a593Smuzhiyun } else {
539*4882a593Smuzhiyun ts = timestamp;
540*4882a593Smuzhiyun }
541*4882a593Smuzhiyun
542*4882a593Smuzhiyun arm_spe_set_pid_tid_cpu(spe, queue);
543*4882a593Smuzhiyun
544*4882a593Smuzhiyun ret = arm_spe_run_decoder(speq, &ts);
545*4882a593Smuzhiyun if (ret < 0) {
546*4882a593Smuzhiyun auxtrace_heap__add(&spe->heap, queue_nr, ts);
547*4882a593Smuzhiyun return ret;
548*4882a593Smuzhiyun }
549*4882a593Smuzhiyun
550*4882a593Smuzhiyun if (!ret) {
551*4882a593Smuzhiyun ret = auxtrace_heap__add(&spe->heap, queue_nr, ts);
552*4882a593Smuzhiyun if (ret < 0)
553*4882a593Smuzhiyun return ret;
554*4882a593Smuzhiyun } else {
555*4882a593Smuzhiyun speq->on_heap = false;
556*4882a593Smuzhiyun }
557*4882a593Smuzhiyun }
558*4882a593Smuzhiyun
559*4882a593Smuzhiyun return 0;
560*4882a593Smuzhiyun }
561*4882a593Smuzhiyun
arm_spe_process_timeless_queues(struct arm_spe * spe,pid_t tid,u64 time_)562*4882a593Smuzhiyun static int arm_spe_process_timeless_queues(struct arm_spe *spe, pid_t tid,
563*4882a593Smuzhiyun u64 time_)
564*4882a593Smuzhiyun {
565*4882a593Smuzhiyun struct auxtrace_queues *queues = &spe->queues;
566*4882a593Smuzhiyun unsigned int i;
567*4882a593Smuzhiyun u64 ts = 0;
568*4882a593Smuzhiyun
569*4882a593Smuzhiyun for (i = 0; i < queues->nr_queues; i++) {
570*4882a593Smuzhiyun struct auxtrace_queue *queue = &spe->queues.queue_array[i];
571*4882a593Smuzhiyun struct arm_spe_queue *speq = queue->priv;
572*4882a593Smuzhiyun
573*4882a593Smuzhiyun if (speq && (tid == -1 || speq->tid == tid)) {
574*4882a593Smuzhiyun speq->time = time_;
575*4882a593Smuzhiyun arm_spe_set_pid_tid_cpu(spe, queue);
576*4882a593Smuzhiyun arm_spe_run_decoder(speq, &ts);
577*4882a593Smuzhiyun }
578*4882a593Smuzhiyun }
579*4882a593Smuzhiyun return 0;
580*4882a593Smuzhiyun }
581*4882a593Smuzhiyun
arm_spe_process_event(struct perf_session * session,union perf_event * event,struct perf_sample * sample,struct perf_tool * tool)582*4882a593Smuzhiyun static int arm_spe_process_event(struct perf_session *session,
583*4882a593Smuzhiyun union perf_event *event,
584*4882a593Smuzhiyun struct perf_sample *sample,
585*4882a593Smuzhiyun struct perf_tool *tool)
586*4882a593Smuzhiyun {
587*4882a593Smuzhiyun int err = 0;
588*4882a593Smuzhiyun u64 timestamp;
589*4882a593Smuzhiyun struct arm_spe *spe = container_of(session->auxtrace,
590*4882a593Smuzhiyun struct arm_spe, auxtrace);
591*4882a593Smuzhiyun
592*4882a593Smuzhiyun if (dump_trace)
593*4882a593Smuzhiyun return 0;
594*4882a593Smuzhiyun
595*4882a593Smuzhiyun if (!tool->ordered_events) {
596*4882a593Smuzhiyun pr_err("SPE trace requires ordered events\n");
597*4882a593Smuzhiyun return -EINVAL;
598*4882a593Smuzhiyun }
599*4882a593Smuzhiyun
600*4882a593Smuzhiyun if (sample->time && (sample->time != (u64) -1))
601*4882a593Smuzhiyun timestamp = sample->time;
602*4882a593Smuzhiyun else
603*4882a593Smuzhiyun timestamp = 0;
604*4882a593Smuzhiyun
605*4882a593Smuzhiyun if (timestamp || spe->timeless_decoding) {
606*4882a593Smuzhiyun err = arm_spe__update_queues(spe);
607*4882a593Smuzhiyun if (err)
608*4882a593Smuzhiyun return err;
609*4882a593Smuzhiyun }
610*4882a593Smuzhiyun
611*4882a593Smuzhiyun if (spe->timeless_decoding) {
612*4882a593Smuzhiyun if (event->header.type == PERF_RECORD_EXIT) {
613*4882a593Smuzhiyun err = arm_spe_process_timeless_queues(spe,
614*4882a593Smuzhiyun event->fork.tid,
615*4882a593Smuzhiyun sample->time);
616*4882a593Smuzhiyun }
617*4882a593Smuzhiyun } else if (timestamp) {
618*4882a593Smuzhiyun if (event->header.type == PERF_RECORD_EXIT) {
619*4882a593Smuzhiyun err = arm_spe_process_queues(spe, timestamp);
620*4882a593Smuzhiyun if (err)
621*4882a593Smuzhiyun return err;
622*4882a593Smuzhiyun }
623*4882a593Smuzhiyun }
624*4882a593Smuzhiyun
625*4882a593Smuzhiyun return err;
626*4882a593Smuzhiyun }
627*4882a593Smuzhiyun
arm_spe_process_auxtrace_event(struct perf_session * session,union perf_event * event,struct perf_tool * tool __maybe_unused)628*4882a593Smuzhiyun static int arm_spe_process_auxtrace_event(struct perf_session *session,
629*4882a593Smuzhiyun union perf_event *event,
630*4882a593Smuzhiyun struct perf_tool *tool __maybe_unused)
631*4882a593Smuzhiyun {
632*4882a593Smuzhiyun struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
633*4882a593Smuzhiyun auxtrace);
634*4882a593Smuzhiyun
635*4882a593Smuzhiyun if (!spe->data_queued) {
636*4882a593Smuzhiyun struct auxtrace_buffer *buffer;
637*4882a593Smuzhiyun off_t data_offset;
638*4882a593Smuzhiyun int fd = perf_data__fd(session->data);
639*4882a593Smuzhiyun int err;
640*4882a593Smuzhiyun
641*4882a593Smuzhiyun if (perf_data__is_pipe(session->data)) {
642*4882a593Smuzhiyun data_offset = 0;
643*4882a593Smuzhiyun } else {
644*4882a593Smuzhiyun data_offset = lseek(fd, 0, SEEK_CUR);
645*4882a593Smuzhiyun if (data_offset == -1)
646*4882a593Smuzhiyun return -errno;
647*4882a593Smuzhiyun }
648*4882a593Smuzhiyun
649*4882a593Smuzhiyun err = auxtrace_queues__add_event(&spe->queues, session, event,
650*4882a593Smuzhiyun data_offset, &buffer);
651*4882a593Smuzhiyun if (err)
652*4882a593Smuzhiyun return err;
653*4882a593Smuzhiyun
654*4882a593Smuzhiyun /* Dump here now we have copied a piped trace out of the pipe */
655*4882a593Smuzhiyun if (dump_trace) {
656*4882a593Smuzhiyun if (auxtrace_buffer__get_data(buffer, fd)) {
657*4882a593Smuzhiyun arm_spe_dump_event(spe, buffer->data,
658*4882a593Smuzhiyun buffer->size);
659*4882a593Smuzhiyun auxtrace_buffer__put_data(buffer);
660*4882a593Smuzhiyun }
661*4882a593Smuzhiyun }
662*4882a593Smuzhiyun }
663*4882a593Smuzhiyun
664*4882a593Smuzhiyun return 0;
665*4882a593Smuzhiyun }
666*4882a593Smuzhiyun
arm_spe_flush(struct perf_session * session __maybe_unused,struct perf_tool * tool __maybe_unused)667*4882a593Smuzhiyun static int arm_spe_flush(struct perf_session *session __maybe_unused,
668*4882a593Smuzhiyun struct perf_tool *tool __maybe_unused)
669*4882a593Smuzhiyun {
670*4882a593Smuzhiyun struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
671*4882a593Smuzhiyun auxtrace);
672*4882a593Smuzhiyun int ret;
673*4882a593Smuzhiyun
674*4882a593Smuzhiyun if (dump_trace)
675*4882a593Smuzhiyun return 0;
676*4882a593Smuzhiyun
677*4882a593Smuzhiyun if (!tool->ordered_events)
678*4882a593Smuzhiyun return -EINVAL;
679*4882a593Smuzhiyun
680*4882a593Smuzhiyun ret = arm_spe__update_queues(spe);
681*4882a593Smuzhiyun if (ret < 0)
682*4882a593Smuzhiyun return ret;
683*4882a593Smuzhiyun
684*4882a593Smuzhiyun if (spe->timeless_decoding)
685*4882a593Smuzhiyun return arm_spe_process_timeless_queues(spe, -1,
686*4882a593Smuzhiyun MAX_TIMESTAMP - 1);
687*4882a593Smuzhiyun
688*4882a593Smuzhiyun return arm_spe_process_queues(spe, MAX_TIMESTAMP);
689*4882a593Smuzhiyun }
690*4882a593Smuzhiyun
arm_spe_free_queue(void * priv)691*4882a593Smuzhiyun static void arm_spe_free_queue(void *priv)
692*4882a593Smuzhiyun {
693*4882a593Smuzhiyun struct arm_spe_queue *speq = priv;
694*4882a593Smuzhiyun
695*4882a593Smuzhiyun if (!speq)
696*4882a593Smuzhiyun return;
697*4882a593Smuzhiyun thread__zput(speq->thread);
698*4882a593Smuzhiyun arm_spe_decoder_free(speq->decoder);
699*4882a593Smuzhiyun zfree(&speq->event_buf);
700*4882a593Smuzhiyun free(speq);
701*4882a593Smuzhiyun }
702*4882a593Smuzhiyun
arm_spe_free_events(struct perf_session * session)703*4882a593Smuzhiyun static void arm_spe_free_events(struct perf_session *session)
704*4882a593Smuzhiyun {
705*4882a593Smuzhiyun struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
706*4882a593Smuzhiyun auxtrace);
707*4882a593Smuzhiyun struct auxtrace_queues *queues = &spe->queues;
708*4882a593Smuzhiyun unsigned int i;
709*4882a593Smuzhiyun
710*4882a593Smuzhiyun for (i = 0; i < queues->nr_queues; i++) {
711*4882a593Smuzhiyun arm_spe_free_queue(queues->queue_array[i].priv);
712*4882a593Smuzhiyun queues->queue_array[i].priv = NULL;
713*4882a593Smuzhiyun }
714*4882a593Smuzhiyun auxtrace_queues__free(queues);
715*4882a593Smuzhiyun }
716*4882a593Smuzhiyun
arm_spe_free(struct perf_session * session)717*4882a593Smuzhiyun static void arm_spe_free(struct perf_session *session)
718*4882a593Smuzhiyun {
719*4882a593Smuzhiyun struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe,
720*4882a593Smuzhiyun auxtrace);
721*4882a593Smuzhiyun
722*4882a593Smuzhiyun auxtrace_heap__free(&spe->heap);
723*4882a593Smuzhiyun arm_spe_free_events(session);
724*4882a593Smuzhiyun session->auxtrace = NULL;
725*4882a593Smuzhiyun free(spe);
726*4882a593Smuzhiyun }
727*4882a593Smuzhiyun
arm_spe_evsel_is_auxtrace(struct perf_session * session,struct evsel * evsel)728*4882a593Smuzhiyun static bool arm_spe_evsel_is_auxtrace(struct perf_session *session,
729*4882a593Smuzhiyun struct evsel *evsel)
730*4882a593Smuzhiyun {
731*4882a593Smuzhiyun struct arm_spe *spe = container_of(session->auxtrace, struct arm_spe, auxtrace);
732*4882a593Smuzhiyun
733*4882a593Smuzhiyun return evsel->core.attr.type == spe->pmu_type;
734*4882a593Smuzhiyun }
735*4882a593Smuzhiyun
736*4882a593Smuzhiyun static const char * const arm_spe_info_fmts[] = {
737*4882a593Smuzhiyun [ARM_SPE_PMU_TYPE] = " PMU Type %"PRId64"\n",
738*4882a593Smuzhiyun };
739*4882a593Smuzhiyun
arm_spe_print_info(__u64 * arr)740*4882a593Smuzhiyun static void arm_spe_print_info(__u64 *arr)
741*4882a593Smuzhiyun {
742*4882a593Smuzhiyun if (!dump_trace)
743*4882a593Smuzhiyun return;
744*4882a593Smuzhiyun
745*4882a593Smuzhiyun fprintf(stdout, arm_spe_info_fmts[ARM_SPE_PMU_TYPE], arr[ARM_SPE_PMU_TYPE]);
746*4882a593Smuzhiyun }
747*4882a593Smuzhiyun
748*4882a593Smuzhiyun struct arm_spe_synth {
749*4882a593Smuzhiyun struct perf_tool dummy_tool;
750*4882a593Smuzhiyun struct perf_session *session;
751*4882a593Smuzhiyun };
752*4882a593Smuzhiyun
arm_spe_event_synth(struct perf_tool * tool,union perf_event * event,struct perf_sample * sample __maybe_unused,struct machine * machine __maybe_unused)753*4882a593Smuzhiyun static int arm_spe_event_synth(struct perf_tool *tool,
754*4882a593Smuzhiyun union perf_event *event,
755*4882a593Smuzhiyun struct perf_sample *sample __maybe_unused,
756*4882a593Smuzhiyun struct machine *machine __maybe_unused)
757*4882a593Smuzhiyun {
758*4882a593Smuzhiyun struct arm_spe_synth *arm_spe_synth =
759*4882a593Smuzhiyun container_of(tool, struct arm_spe_synth, dummy_tool);
760*4882a593Smuzhiyun
761*4882a593Smuzhiyun return perf_session__deliver_synth_event(arm_spe_synth->session,
762*4882a593Smuzhiyun event, NULL);
763*4882a593Smuzhiyun }
764*4882a593Smuzhiyun
arm_spe_synth_event(struct perf_session * session,struct perf_event_attr * attr,u64 id)765*4882a593Smuzhiyun static int arm_spe_synth_event(struct perf_session *session,
766*4882a593Smuzhiyun struct perf_event_attr *attr, u64 id)
767*4882a593Smuzhiyun {
768*4882a593Smuzhiyun struct arm_spe_synth arm_spe_synth;
769*4882a593Smuzhiyun
770*4882a593Smuzhiyun memset(&arm_spe_synth, 0, sizeof(struct arm_spe_synth));
771*4882a593Smuzhiyun arm_spe_synth.session = session;
772*4882a593Smuzhiyun
773*4882a593Smuzhiyun return perf_event__synthesize_attr(&arm_spe_synth.dummy_tool, attr, 1,
774*4882a593Smuzhiyun &id, arm_spe_event_synth);
775*4882a593Smuzhiyun }
776*4882a593Smuzhiyun
arm_spe_set_event_name(struct evlist * evlist,u64 id,const char * name)777*4882a593Smuzhiyun static void arm_spe_set_event_name(struct evlist *evlist, u64 id,
778*4882a593Smuzhiyun const char *name)
779*4882a593Smuzhiyun {
780*4882a593Smuzhiyun struct evsel *evsel;
781*4882a593Smuzhiyun
782*4882a593Smuzhiyun evlist__for_each_entry(evlist, evsel) {
783*4882a593Smuzhiyun if (evsel->core.id && evsel->core.id[0] == id) {
784*4882a593Smuzhiyun if (evsel->name)
785*4882a593Smuzhiyun zfree(&evsel->name);
786*4882a593Smuzhiyun evsel->name = strdup(name);
787*4882a593Smuzhiyun break;
788*4882a593Smuzhiyun }
789*4882a593Smuzhiyun }
790*4882a593Smuzhiyun }
791*4882a593Smuzhiyun
792*4882a593Smuzhiyun static int
arm_spe_synth_events(struct arm_spe * spe,struct perf_session * session)793*4882a593Smuzhiyun arm_spe_synth_events(struct arm_spe *spe, struct perf_session *session)
794*4882a593Smuzhiyun {
795*4882a593Smuzhiyun struct evlist *evlist = session->evlist;
796*4882a593Smuzhiyun struct evsel *evsel;
797*4882a593Smuzhiyun struct perf_event_attr attr;
798*4882a593Smuzhiyun bool found = false;
799*4882a593Smuzhiyun u64 id;
800*4882a593Smuzhiyun int err;
801*4882a593Smuzhiyun
802*4882a593Smuzhiyun evlist__for_each_entry(evlist, evsel) {
803*4882a593Smuzhiyun if (evsel->core.attr.type == spe->pmu_type) {
804*4882a593Smuzhiyun found = true;
805*4882a593Smuzhiyun break;
806*4882a593Smuzhiyun }
807*4882a593Smuzhiyun }
808*4882a593Smuzhiyun
809*4882a593Smuzhiyun if (!found) {
810*4882a593Smuzhiyun pr_debug("No selected events with SPE trace data\n");
811*4882a593Smuzhiyun return 0;
812*4882a593Smuzhiyun }
813*4882a593Smuzhiyun
814*4882a593Smuzhiyun memset(&attr, 0, sizeof(struct perf_event_attr));
815*4882a593Smuzhiyun attr.size = sizeof(struct perf_event_attr);
816*4882a593Smuzhiyun attr.type = PERF_TYPE_HARDWARE;
817*4882a593Smuzhiyun attr.sample_type = evsel->core.attr.sample_type & PERF_SAMPLE_MASK;
818*4882a593Smuzhiyun attr.sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID |
819*4882a593Smuzhiyun PERF_SAMPLE_PERIOD;
820*4882a593Smuzhiyun if (spe->timeless_decoding)
821*4882a593Smuzhiyun attr.sample_type &= ~(u64)PERF_SAMPLE_TIME;
822*4882a593Smuzhiyun else
823*4882a593Smuzhiyun attr.sample_type |= PERF_SAMPLE_TIME;
824*4882a593Smuzhiyun
825*4882a593Smuzhiyun spe->sample_type = attr.sample_type;
826*4882a593Smuzhiyun
827*4882a593Smuzhiyun attr.exclude_user = evsel->core.attr.exclude_user;
828*4882a593Smuzhiyun attr.exclude_kernel = evsel->core.attr.exclude_kernel;
829*4882a593Smuzhiyun attr.exclude_hv = evsel->core.attr.exclude_hv;
830*4882a593Smuzhiyun attr.exclude_host = evsel->core.attr.exclude_host;
831*4882a593Smuzhiyun attr.exclude_guest = evsel->core.attr.exclude_guest;
832*4882a593Smuzhiyun attr.sample_id_all = evsel->core.attr.sample_id_all;
833*4882a593Smuzhiyun attr.read_format = evsel->core.attr.read_format;
834*4882a593Smuzhiyun
835*4882a593Smuzhiyun /* create new id val to be a fixed offset from evsel id */
836*4882a593Smuzhiyun id = evsel->core.id[0] + 1000000000;
837*4882a593Smuzhiyun
838*4882a593Smuzhiyun if (!id)
839*4882a593Smuzhiyun id = 1;
840*4882a593Smuzhiyun
841*4882a593Smuzhiyun if (spe->synth_opts.flc) {
842*4882a593Smuzhiyun spe->sample_flc = true;
843*4882a593Smuzhiyun
844*4882a593Smuzhiyun /* Level 1 data cache miss */
845*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
846*4882a593Smuzhiyun if (err)
847*4882a593Smuzhiyun return err;
848*4882a593Smuzhiyun spe->l1d_miss_id = id;
849*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "l1d-miss");
850*4882a593Smuzhiyun id += 1;
851*4882a593Smuzhiyun
852*4882a593Smuzhiyun /* Level 1 data cache access */
853*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
854*4882a593Smuzhiyun if (err)
855*4882a593Smuzhiyun return err;
856*4882a593Smuzhiyun spe->l1d_access_id = id;
857*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "l1d-access");
858*4882a593Smuzhiyun id += 1;
859*4882a593Smuzhiyun }
860*4882a593Smuzhiyun
861*4882a593Smuzhiyun if (spe->synth_opts.llc) {
862*4882a593Smuzhiyun spe->sample_llc = true;
863*4882a593Smuzhiyun
864*4882a593Smuzhiyun /* Last level cache miss */
865*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
866*4882a593Smuzhiyun if (err)
867*4882a593Smuzhiyun return err;
868*4882a593Smuzhiyun spe->llc_miss_id = id;
869*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "llc-miss");
870*4882a593Smuzhiyun id += 1;
871*4882a593Smuzhiyun
872*4882a593Smuzhiyun /* Last level cache access */
873*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
874*4882a593Smuzhiyun if (err)
875*4882a593Smuzhiyun return err;
876*4882a593Smuzhiyun spe->llc_access_id = id;
877*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "llc-access");
878*4882a593Smuzhiyun id += 1;
879*4882a593Smuzhiyun }
880*4882a593Smuzhiyun
881*4882a593Smuzhiyun if (spe->synth_opts.tlb) {
882*4882a593Smuzhiyun spe->sample_tlb = true;
883*4882a593Smuzhiyun
884*4882a593Smuzhiyun /* TLB miss */
885*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
886*4882a593Smuzhiyun if (err)
887*4882a593Smuzhiyun return err;
888*4882a593Smuzhiyun spe->tlb_miss_id = id;
889*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "tlb-miss");
890*4882a593Smuzhiyun id += 1;
891*4882a593Smuzhiyun
892*4882a593Smuzhiyun /* TLB access */
893*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
894*4882a593Smuzhiyun if (err)
895*4882a593Smuzhiyun return err;
896*4882a593Smuzhiyun spe->tlb_access_id = id;
897*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "tlb-access");
898*4882a593Smuzhiyun id += 1;
899*4882a593Smuzhiyun }
900*4882a593Smuzhiyun
901*4882a593Smuzhiyun if (spe->synth_opts.branches) {
902*4882a593Smuzhiyun spe->sample_branch = true;
903*4882a593Smuzhiyun
904*4882a593Smuzhiyun /* Branch miss */
905*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
906*4882a593Smuzhiyun if (err)
907*4882a593Smuzhiyun return err;
908*4882a593Smuzhiyun spe->branch_miss_id = id;
909*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "branch-miss");
910*4882a593Smuzhiyun id += 1;
911*4882a593Smuzhiyun }
912*4882a593Smuzhiyun
913*4882a593Smuzhiyun if (spe->synth_opts.remote_access) {
914*4882a593Smuzhiyun spe->sample_remote_access = true;
915*4882a593Smuzhiyun
916*4882a593Smuzhiyun /* Remote access */
917*4882a593Smuzhiyun err = arm_spe_synth_event(session, &attr, id);
918*4882a593Smuzhiyun if (err)
919*4882a593Smuzhiyun return err;
920*4882a593Smuzhiyun spe->remote_access_id = id;
921*4882a593Smuzhiyun arm_spe_set_event_name(evlist, id, "remote-access");
922*4882a593Smuzhiyun id += 1;
923*4882a593Smuzhiyun }
924*4882a593Smuzhiyun
925*4882a593Smuzhiyun return 0;
926*4882a593Smuzhiyun }
927*4882a593Smuzhiyun
arm_spe_process_auxtrace_info(union perf_event * event,struct perf_session * session)928*4882a593Smuzhiyun int arm_spe_process_auxtrace_info(union perf_event *event,
929*4882a593Smuzhiyun struct perf_session *session)
930*4882a593Smuzhiyun {
931*4882a593Smuzhiyun struct perf_record_auxtrace_info *auxtrace_info = &event->auxtrace_info;
932*4882a593Smuzhiyun size_t min_sz = sizeof(u64) * ARM_SPE_AUXTRACE_PRIV_MAX;
933*4882a593Smuzhiyun struct arm_spe *spe;
934*4882a593Smuzhiyun int err;
935*4882a593Smuzhiyun
936*4882a593Smuzhiyun if (auxtrace_info->header.size < sizeof(struct perf_record_auxtrace_info) +
937*4882a593Smuzhiyun min_sz)
938*4882a593Smuzhiyun return -EINVAL;
939*4882a593Smuzhiyun
940*4882a593Smuzhiyun spe = zalloc(sizeof(struct arm_spe));
941*4882a593Smuzhiyun if (!spe)
942*4882a593Smuzhiyun return -ENOMEM;
943*4882a593Smuzhiyun
944*4882a593Smuzhiyun err = auxtrace_queues__init(&spe->queues);
945*4882a593Smuzhiyun if (err)
946*4882a593Smuzhiyun goto err_free;
947*4882a593Smuzhiyun
948*4882a593Smuzhiyun spe->session = session;
949*4882a593Smuzhiyun spe->machine = &session->machines.host; /* No kvm support */
950*4882a593Smuzhiyun spe->auxtrace_type = auxtrace_info->type;
951*4882a593Smuzhiyun spe->pmu_type = auxtrace_info->priv[ARM_SPE_PMU_TYPE];
952*4882a593Smuzhiyun
953*4882a593Smuzhiyun spe->timeless_decoding = arm_spe__is_timeless_decoding(spe);
954*4882a593Smuzhiyun spe->auxtrace.process_event = arm_spe_process_event;
955*4882a593Smuzhiyun spe->auxtrace.process_auxtrace_event = arm_spe_process_auxtrace_event;
956*4882a593Smuzhiyun spe->auxtrace.flush_events = arm_spe_flush;
957*4882a593Smuzhiyun spe->auxtrace.free_events = arm_spe_free_events;
958*4882a593Smuzhiyun spe->auxtrace.free = arm_spe_free;
959*4882a593Smuzhiyun spe->auxtrace.evsel_is_auxtrace = arm_spe_evsel_is_auxtrace;
960*4882a593Smuzhiyun session->auxtrace = &spe->auxtrace;
961*4882a593Smuzhiyun
962*4882a593Smuzhiyun arm_spe_print_info(&auxtrace_info->priv[0]);
963*4882a593Smuzhiyun
964*4882a593Smuzhiyun if (dump_trace)
965*4882a593Smuzhiyun return 0;
966*4882a593Smuzhiyun
967*4882a593Smuzhiyun if (session->itrace_synth_opts && session->itrace_synth_opts->set)
968*4882a593Smuzhiyun spe->synth_opts = *session->itrace_synth_opts;
969*4882a593Smuzhiyun else
970*4882a593Smuzhiyun itrace_synth_opts__set_default(&spe->synth_opts, false);
971*4882a593Smuzhiyun
972*4882a593Smuzhiyun err = arm_spe_synth_events(spe, session);
973*4882a593Smuzhiyun if (err)
974*4882a593Smuzhiyun goto err_free_queues;
975*4882a593Smuzhiyun
976*4882a593Smuzhiyun err = auxtrace_queues__process_index(&spe->queues, session);
977*4882a593Smuzhiyun if (err)
978*4882a593Smuzhiyun goto err_free_queues;
979*4882a593Smuzhiyun
980*4882a593Smuzhiyun if (spe->queues.populated)
981*4882a593Smuzhiyun spe->data_queued = true;
982*4882a593Smuzhiyun
983*4882a593Smuzhiyun return 0;
984*4882a593Smuzhiyun
985*4882a593Smuzhiyun err_free_queues:
986*4882a593Smuzhiyun auxtrace_queues__free(&spe->queues);
987*4882a593Smuzhiyun session->auxtrace = NULL;
988*4882a593Smuzhiyun err_free:
989*4882a593Smuzhiyun free(spe);
990*4882a593Smuzhiyun return err;
991*4882a593Smuzhiyun }
992