1*4882a593Smuzhiyun // SPDX-License-Identifier: GPL-2.0
2*4882a593Smuzhiyun /*
3*4882a593Smuzhiyun * Compare and figure out the top N hottest streams
4*4882a593Smuzhiyun * Copyright (c) 2020, Intel Corporation.
5*4882a593Smuzhiyun * Author: Jin Yao
6*4882a593Smuzhiyun */
7*4882a593Smuzhiyun
8*4882a593Smuzhiyun #include <inttypes.h>
9*4882a593Smuzhiyun #include <stdlib.h>
10*4882a593Smuzhiyun #include <linux/zalloc.h>
11*4882a593Smuzhiyun #include "debug.h"
12*4882a593Smuzhiyun #include "hist.h"
13*4882a593Smuzhiyun #include "sort.h"
14*4882a593Smuzhiyun #include "stream.h"
15*4882a593Smuzhiyun #include "evlist.h"
16*4882a593Smuzhiyun
evsel_streams__delete(struct evsel_streams * es,int nr_evsel)17*4882a593Smuzhiyun static void evsel_streams__delete(struct evsel_streams *es, int nr_evsel)
18*4882a593Smuzhiyun {
19*4882a593Smuzhiyun for (int i = 0; i < nr_evsel; i++)
20*4882a593Smuzhiyun zfree(&es[i].streams);
21*4882a593Smuzhiyun
22*4882a593Smuzhiyun free(es);
23*4882a593Smuzhiyun }
24*4882a593Smuzhiyun
evlist_streams__delete(struct evlist_streams * els)25*4882a593Smuzhiyun void evlist_streams__delete(struct evlist_streams *els)
26*4882a593Smuzhiyun {
27*4882a593Smuzhiyun evsel_streams__delete(els->ev_streams, els->nr_evsel);
28*4882a593Smuzhiyun free(els);
29*4882a593Smuzhiyun }
30*4882a593Smuzhiyun
evlist_streams__new(int nr_evsel,int nr_streams_max)31*4882a593Smuzhiyun static struct evlist_streams *evlist_streams__new(int nr_evsel,
32*4882a593Smuzhiyun int nr_streams_max)
33*4882a593Smuzhiyun {
34*4882a593Smuzhiyun struct evlist_streams *els;
35*4882a593Smuzhiyun struct evsel_streams *es;
36*4882a593Smuzhiyun
37*4882a593Smuzhiyun els = zalloc(sizeof(*els));
38*4882a593Smuzhiyun if (!els)
39*4882a593Smuzhiyun return NULL;
40*4882a593Smuzhiyun
41*4882a593Smuzhiyun es = calloc(nr_evsel, sizeof(struct evsel_streams));
42*4882a593Smuzhiyun if (!es) {
43*4882a593Smuzhiyun free(els);
44*4882a593Smuzhiyun return NULL;
45*4882a593Smuzhiyun }
46*4882a593Smuzhiyun
47*4882a593Smuzhiyun for (int i = 0; i < nr_evsel; i++) {
48*4882a593Smuzhiyun struct evsel_streams *s = &es[i];
49*4882a593Smuzhiyun
50*4882a593Smuzhiyun s->streams = calloc(nr_streams_max, sizeof(struct stream));
51*4882a593Smuzhiyun if (!s->streams)
52*4882a593Smuzhiyun goto err;
53*4882a593Smuzhiyun
54*4882a593Smuzhiyun s->nr_streams_max = nr_streams_max;
55*4882a593Smuzhiyun s->evsel_idx = -1;
56*4882a593Smuzhiyun }
57*4882a593Smuzhiyun
58*4882a593Smuzhiyun els->ev_streams = es;
59*4882a593Smuzhiyun els->nr_evsel = nr_evsel;
60*4882a593Smuzhiyun return els;
61*4882a593Smuzhiyun
62*4882a593Smuzhiyun err:
63*4882a593Smuzhiyun evsel_streams__delete(es, nr_evsel);
64*4882a593Smuzhiyun return NULL;
65*4882a593Smuzhiyun }
66*4882a593Smuzhiyun
67*4882a593Smuzhiyun /*
68*4882a593Smuzhiyun * The cnodes with high hit number are hot callchains.
69*4882a593Smuzhiyun */
evsel_streams__set_hot_cnode(struct evsel_streams * es,struct callchain_node * cnode)70*4882a593Smuzhiyun static void evsel_streams__set_hot_cnode(struct evsel_streams *es,
71*4882a593Smuzhiyun struct callchain_node *cnode)
72*4882a593Smuzhiyun {
73*4882a593Smuzhiyun int i, idx = 0;
74*4882a593Smuzhiyun u64 hit;
75*4882a593Smuzhiyun
76*4882a593Smuzhiyun if (es->nr_streams < es->nr_streams_max) {
77*4882a593Smuzhiyun i = es->nr_streams;
78*4882a593Smuzhiyun es->streams[i].cnode = cnode;
79*4882a593Smuzhiyun es->nr_streams++;
80*4882a593Smuzhiyun return;
81*4882a593Smuzhiyun }
82*4882a593Smuzhiyun
83*4882a593Smuzhiyun /*
84*4882a593Smuzhiyun * Considering a few number of hot streams, only use simple
85*4882a593Smuzhiyun * way to find the cnode with smallest hit number and replace.
86*4882a593Smuzhiyun */
87*4882a593Smuzhiyun hit = (es->streams[0].cnode)->hit;
88*4882a593Smuzhiyun for (i = 1; i < es->nr_streams; i++) {
89*4882a593Smuzhiyun if ((es->streams[i].cnode)->hit < hit) {
90*4882a593Smuzhiyun hit = (es->streams[i].cnode)->hit;
91*4882a593Smuzhiyun idx = i;
92*4882a593Smuzhiyun }
93*4882a593Smuzhiyun }
94*4882a593Smuzhiyun
95*4882a593Smuzhiyun if (cnode->hit > hit)
96*4882a593Smuzhiyun es->streams[idx].cnode = cnode;
97*4882a593Smuzhiyun }
98*4882a593Smuzhiyun
update_hot_callchain(struct hist_entry * he,struct evsel_streams * es)99*4882a593Smuzhiyun static void update_hot_callchain(struct hist_entry *he,
100*4882a593Smuzhiyun struct evsel_streams *es)
101*4882a593Smuzhiyun {
102*4882a593Smuzhiyun struct rb_root *root = &he->sorted_chain;
103*4882a593Smuzhiyun struct rb_node *rb_node = rb_first(root);
104*4882a593Smuzhiyun struct callchain_node *cnode;
105*4882a593Smuzhiyun
106*4882a593Smuzhiyun while (rb_node) {
107*4882a593Smuzhiyun cnode = rb_entry(rb_node, struct callchain_node, rb_node);
108*4882a593Smuzhiyun evsel_streams__set_hot_cnode(es, cnode);
109*4882a593Smuzhiyun rb_node = rb_next(rb_node);
110*4882a593Smuzhiyun }
111*4882a593Smuzhiyun }
112*4882a593Smuzhiyun
init_hot_callchain(struct hists * hists,struct evsel_streams * es)113*4882a593Smuzhiyun static void init_hot_callchain(struct hists *hists, struct evsel_streams *es)
114*4882a593Smuzhiyun {
115*4882a593Smuzhiyun struct rb_node *next = rb_first_cached(&hists->entries);
116*4882a593Smuzhiyun
117*4882a593Smuzhiyun while (next) {
118*4882a593Smuzhiyun struct hist_entry *he;
119*4882a593Smuzhiyun
120*4882a593Smuzhiyun he = rb_entry(next, struct hist_entry, rb_node);
121*4882a593Smuzhiyun update_hot_callchain(he, es);
122*4882a593Smuzhiyun next = rb_next(&he->rb_node);
123*4882a593Smuzhiyun }
124*4882a593Smuzhiyun
125*4882a593Smuzhiyun es->streams_hits = callchain_total_hits(hists);
126*4882a593Smuzhiyun }
127*4882a593Smuzhiyun
evlist__init_callchain_streams(struct evlist * evlist,struct evlist_streams * els)128*4882a593Smuzhiyun static int evlist__init_callchain_streams(struct evlist *evlist,
129*4882a593Smuzhiyun struct evlist_streams *els)
130*4882a593Smuzhiyun {
131*4882a593Smuzhiyun struct evsel_streams *es = els->ev_streams;
132*4882a593Smuzhiyun struct evsel *pos;
133*4882a593Smuzhiyun int i = 0;
134*4882a593Smuzhiyun
135*4882a593Smuzhiyun BUG_ON(els->nr_evsel < evlist->core.nr_entries);
136*4882a593Smuzhiyun
137*4882a593Smuzhiyun evlist__for_each_entry(evlist, pos) {
138*4882a593Smuzhiyun struct hists *hists = evsel__hists(pos);
139*4882a593Smuzhiyun
140*4882a593Smuzhiyun hists__output_resort(hists, NULL);
141*4882a593Smuzhiyun init_hot_callchain(hists, &es[i]);
142*4882a593Smuzhiyun es[i].evsel_idx = pos->idx;
143*4882a593Smuzhiyun i++;
144*4882a593Smuzhiyun }
145*4882a593Smuzhiyun
146*4882a593Smuzhiyun return 0;
147*4882a593Smuzhiyun }
148*4882a593Smuzhiyun
evlist__create_streams(struct evlist * evlist,int nr_streams_max)149*4882a593Smuzhiyun struct evlist_streams *evlist__create_streams(struct evlist *evlist,
150*4882a593Smuzhiyun int nr_streams_max)
151*4882a593Smuzhiyun {
152*4882a593Smuzhiyun int nr_evsel = evlist->core.nr_entries, ret = -1;
153*4882a593Smuzhiyun struct evlist_streams *els = evlist_streams__new(nr_evsel,
154*4882a593Smuzhiyun nr_streams_max);
155*4882a593Smuzhiyun
156*4882a593Smuzhiyun if (!els)
157*4882a593Smuzhiyun return NULL;
158*4882a593Smuzhiyun
159*4882a593Smuzhiyun ret = evlist__init_callchain_streams(evlist, els);
160*4882a593Smuzhiyun if (ret) {
161*4882a593Smuzhiyun evlist_streams__delete(els);
162*4882a593Smuzhiyun return NULL;
163*4882a593Smuzhiyun }
164*4882a593Smuzhiyun
165*4882a593Smuzhiyun return els;
166*4882a593Smuzhiyun }
167*4882a593Smuzhiyun
evsel_streams__entry(struct evlist_streams * els,int evsel_idx)168*4882a593Smuzhiyun struct evsel_streams *evsel_streams__entry(struct evlist_streams *els,
169*4882a593Smuzhiyun int evsel_idx)
170*4882a593Smuzhiyun {
171*4882a593Smuzhiyun struct evsel_streams *es = els->ev_streams;
172*4882a593Smuzhiyun
173*4882a593Smuzhiyun for (int i = 0; i < els->nr_evsel; i++) {
174*4882a593Smuzhiyun if (es[i].evsel_idx == evsel_idx)
175*4882a593Smuzhiyun return &es[i];
176*4882a593Smuzhiyun }
177*4882a593Smuzhiyun
178*4882a593Smuzhiyun return NULL;
179*4882a593Smuzhiyun }
180*4882a593Smuzhiyun
stream__callchain_match(struct stream * base_stream,struct evsel_streams * es_pair)181*4882a593Smuzhiyun static struct stream *stream__callchain_match(struct stream *base_stream,
182*4882a593Smuzhiyun struct evsel_streams *es_pair)
183*4882a593Smuzhiyun {
184*4882a593Smuzhiyun for (int i = 0; i < es_pair->nr_streams; i++) {
185*4882a593Smuzhiyun struct stream *pair_stream = &es_pair->streams[i];
186*4882a593Smuzhiyun
187*4882a593Smuzhiyun if (callchain_cnode_matched(base_stream->cnode,
188*4882a593Smuzhiyun pair_stream->cnode)) {
189*4882a593Smuzhiyun return pair_stream;
190*4882a593Smuzhiyun }
191*4882a593Smuzhiyun }
192*4882a593Smuzhiyun
193*4882a593Smuzhiyun return NULL;
194*4882a593Smuzhiyun }
195*4882a593Smuzhiyun
stream__match(struct stream * base_stream,struct evsel_streams * es_pair)196*4882a593Smuzhiyun static struct stream *stream__match(struct stream *base_stream,
197*4882a593Smuzhiyun struct evsel_streams *es_pair)
198*4882a593Smuzhiyun {
199*4882a593Smuzhiyun return stream__callchain_match(base_stream, es_pair);
200*4882a593Smuzhiyun }
201*4882a593Smuzhiyun
stream__link(struct stream * base_stream,struct stream * pair_stream)202*4882a593Smuzhiyun static void stream__link(struct stream *base_stream, struct stream *pair_stream)
203*4882a593Smuzhiyun {
204*4882a593Smuzhiyun base_stream->pair_cnode = pair_stream->cnode;
205*4882a593Smuzhiyun pair_stream->pair_cnode = base_stream->cnode;
206*4882a593Smuzhiyun }
207*4882a593Smuzhiyun
evsel_streams__match(struct evsel_streams * es_base,struct evsel_streams * es_pair)208*4882a593Smuzhiyun void evsel_streams__match(struct evsel_streams *es_base,
209*4882a593Smuzhiyun struct evsel_streams *es_pair)
210*4882a593Smuzhiyun {
211*4882a593Smuzhiyun for (int i = 0; i < es_base->nr_streams; i++) {
212*4882a593Smuzhiyun struct stream *base_stream = &es_base->streams[i];
213*4882a593Smuzhiyun struct stream *pair_stream;
214*4882a593Smuzhiyun
215*4882a593Smuzhiyun pair_stream = stream__match(base_stream, es_pair);
216*4882a593Smuzhiyun if (pair_stream)
217*4882a593Smuzhiyun stream__link(base_stream, pair_stream);
218*4882a593Smuzhiyun }
219*4882a593Smuzhiyun }
220*4882a593Smuzhiyun
print_callchain_pair(struct stream * base_stream,int idx,struct evsel_streams * es_base,struct evsel_streams * es_pair)221*4882a593Smuzhiyun static void print_callchain_pair(struct stream *base_stream, int idx,
222*4882a593Smuzhiyun struct evsel_streams *es_base,
223*4882a593Smuzhiyun struct evsel_streams *es_pair)
224*4882a593Smuzhiyun {
225*4882a593Smuzhiyun struct callchain_node *base_cnode = base_stream->cnode;
226*4882a593Smuzhiyun struct callchain_node *pair_cnode = base_stream->pair_cnode;
227*4882a593Smuzhiyun struct callchain_list *base_chain, *pair_chain;
228*4882a593Smuzhiyun char buf1[512], buf2[512], cbuf1[256], cbuf2[256];
229*4882a593Smuzhiyun char *s1, *s2;
230*4882a593Smuzhiyun double pct;
231*4882a593Smuzhiyun
232*4882a593Smuzhiyun printf("\nhot chain pair %d:\n", idx);
233*4882a593Smuzhiyun
234*4882a593Smuzhiyun pct = (double)base_cnode->hit / (double)es_base->streams_hits;
235*4882a593Smuzhiyun scnprintf(buf1, sizeof(buf1), "cycles: %ld, hits: %.2f%%",
236*4882a593Smuzhiyun callchain_avg_cycles(base_cnode), pct * 100.0);
237*4882a593Smuzhiyun
238*4882a593Smuzhiyun pct = (double)pair_cnode->hit / (double)es_pair->streams_hits;
239*4882a593Smuzhiyun scnprintf(buf2, sizeof(buf2), "cycles: %ld, hits: %.2f%%",
240*4882a593Smuzhiyun callchain_avg_cycles(pair_cnode), pct * 100.0);
241*4882a593Smuzhiyun
242*4882a593Smuzhiyun printf("%35s\t%35s\n", buf1, buf2);
243*4882a593Smuzhiyun
244*4882a593Smuzhiyun printf("%35s\t%35s\n",
245*4882a593Smuzhiyun "---------------------------",
246*4882a593Smuzhiyun "--------------------------");
247*4882a593Smuzhiyun
248*4882a593Smuzhiyun pair_chain = list_first_entry(&pair_cnode->val,
249*4882a593Smuzhiyun struct callchain_list,
250*4882a593Smuzhiyun list);
251*4882a593Smuzhiyun
252*4882a593Smuzhiyun list_for_each_entry(base_chain, &base_cnode->val, list) {
253*4882a593Smuzhiyun if (&pair_chain->list == &pair_cnode->val)
254*4882a593Smuzhiyun return;
255*4882a593Smuzhiyun
256*4882a593Smuzhiyun s1 = callchain_list__sym_name(base_chain, cbuf1, sizeof(cbuf1),
257*4882a593Smuzhiyun false);
258*4882a593Smuzhiyun s2 = callchain_list__sym_name(pair_chain, cbuf2, sizeof(cbuf2),
259*4882a593Smuzhiyun false);
260*4882a593Smuzhiyun
261*4882a593Smuzhiyun scnprintf(buf1, sizeof(buf1), "%35s\t%35s", s1, s2);
262*4882a593Smuzhiyun printf("%s\n", buf1);
263*4882a593Smuzhiyun pair_chain = list_next_entry(pair_chain, list);
264*4882a593Smuzhiyun }
265*4882a593Smuzhiyun }
266*4882a593Smuzhiyun
print_stream_callchain(struct stream * stream,int idx,struct evsel_streams * es,bool pair)267*4882a593Smuzhiyun static void print_stream_callchain(struct stream *stream, int idx,
268*4882a593Smuzhiyun struct evsel_streams *es, bool pair)
269*4882a593Smuzhiyun {
270*4882a593Smuzhiyun struct callchain_node *cnode = stream->cnode;
271*4882a593Smuzhiyun struct callchain_list *chain;
272*4882a593Smuzhiyun char buf[512], cbuf[256], *s;
273*4882a593Smuzhiyun double pct;
274*4882a593Smuzhiyun
275*4882a593Smuzhiyun printf("\nhot chain %d:\n", idx);
276*4882a593Smuzhiyun
277*4882a593Smuzhiyun pct = (double)cnode->hit / (double)es->streams_hits;
278*4882a593Smuzhiyun scnprintf(buf, sizeof(buf), "cycles: %ld, hits: %.2f%%",
279*4882a593Smuzhiyun callchain_avg_cycles(cnode), pct * 100.0);
280*4882a593Smuzhiyun
281*4882a593Smuzhiyun if (pair) {
282*4882a593Smuzhiyun printf("%35s\t%35s\n", "", buf);
283*4882a593Smuzhiyun printf("%35s\t%35s\n",
284*4882a593Smuzhiyun "", "--------------------------");
285*4882a593Smuzhiyun } else {
286*4882a593Smuzhiyun printf("%35s\n", buf);
287*4882a593Smuzhiyun printf("%35s\n", "--------------------------");
288*4882a593Smuzhiyun }
289*4882a593Smuzhiyun
290*4882a593Smuzhiyun list_for_each_entry(chain, &cnode->val, list) {
291*4882a593Smuzhiyun s = callchain_list__sym_name(chain, cbuf, sizeof(cbuf), false);
292*4882a593Smuzhiyun
293*4882a593Smuzhiyun if (pair)
294*4882a593Smuzhiyun scnprintf(buf, sizeof(buf), "%35s\t%35s", "", s);
295*4882a593Smuzhiyun else
296*4882a593Smuzhiyun scnprintf(buf, sizeof(buf), "%35s", s);
297*4882a593Smuzhiyun
298*4882a593Smuzhiyun printf("%s\n", buf);
299*4882a593Smuzhiyun }
300*4882a593Smuzhiyun }
301*4882a593Smuzhiyun
callchain_streams_report(struct evsel_streams * es_base,struct evsel_streams * es_pair)302*4882a593Smuzhiyun static void callchain_streams_report(struct evsel_streams *es_base,
303*4882a593Smuzhiyun struct evsel_streams *es_pair)
304*4882a593Smuzhiyun {
305*4882a593Smuzhiyun struct stream *base_stream;
306*4882a593Smuzhiyun int i, idx = 0;
307*4882a593Smuzhiyun
308*4882a593Smuzhiyun printf("[ Matched hot streams ]\n");
309*4882a593Smuzhiyun for (i = 0; i < es_base->nr_streams; i++) {
310*4882a593Smuzhiyun base_stream = &es_base->streams[i];
311*4882a593Smuzhiyun if (base_stream->pair_cnode) {
312*4882a593Smuzhiyun print_callchain_pair(base_stream, ++idx,
313*4882a593Smuzhiyun es_base, es_pair);
314*4882a593Smuzhiyun }
315*4882a593Smuzhiyun }
316*4882a593Smuzhiyun
317*4882a593Smuzhiyun idx = 0;
318*4882a593Smuzhiyun printf("\n[ Hot streams in old perf data only ]\n");
319*4882a593Smuzhiyun for (i = 0; i < es_base->nr_streams; i++) {
320*4882a593Smuzhiyun base_stream = &es_base->streams[i];
321*4882a593Smuzhiyun if (!base_stream->pair_cnode) {
322*4882a593Smuzhiyun print_stream_callchain(base_stream, ++idx,
323*4882a593Smuzhiyun es_base, false);
324*4882a593Smuzhiyun }
325*4882a593Smuzhiyun }
326*4882a593Smuzhiyun
327*4882a593Smuzhiyun idx = 0;
328*4882a593Smuzhiyun printf("\n[ Hot streams in new perf data only ]\n");
329*4882a593Smuzhiyun for (i = 0; i < es_pair->nr_streams; i++) {
330*4882a593Smuzhiyun base_stream = &es_pair->streams[i];
331*4882a593Smuzhiyun if (!base_stream->pair_cnode) {
332*4882a593Smuzhiyun print_stream_callchain(base_stream, ++idx,
333*4882a593Smuzhiyun es_pair, true);
334*4882a593Smuzhiyun }
335*4882a593Smuzhiyun }
336*4882a593Smuzhiyun }
337*4882a593Smuzhiyun
evsel_streams__report(struct evsel_streams * es_base,struct evsel_streams * es_pair)338*4882a593Smuzhiyun void evsel_streams__report(struct evsel_streams *es_base,
339*4882a593Smuzhiyun struct evsel_streams *es_pair)
340*4882a593Smuzhiyun {
341*4882a593Smuzhiyun return callchain_streams_report(es_base, es_pair);
342*4882a593Smuzhiyun }
343