tprof.c revision 1.10.14.2 1 /* $NetBSD: tprof.c,v 1.10.14.2 2017/12/03 11:37:33 jdolecek Exp $ */
2
3 /*-
4 * Copyright (c)2008,2009,2010 YAMAMOTO Takashi,
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 */
28
29 #include <sys/cdefs.h>
30 __KERNEL_RCSID(0, "$NetBSD: tprof.c,v 1.10.14.2 2017/12/03 11:37:33 jdolecek Exp $");
31
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
35
36 #include <sys/cpu.h>
37 #include <sys/conf.h>
38 #include <sys/callout.h>
39 #include <sys/kmem.h>
40 #include <sys/module.h>
41 #include <sys/proc.h>
42 #include <sys/workqueue.h>
43 #include <sys/queue.h>
44
45 #include <dev/tprof/tprof.h>
46 #include <dev/tprof/tprof_ioctl.h>
47
48 #include "ioconf.h"
49
50 /*
51 * locking order:
52 * tprof_reader_lock -> tprof_lock
53 * tprof_startstop_lock -> tprof_lock
54 */
55
56 /*
57 * protected by:
58 * L: tprof_lock
59 * R: tprof_reader_lock
60 * S: tprof_startstop_lock
61 * s: writer should hold tprof_startstop_lock and tprof_lock
62 * reader should hold tprof_startstop_lock or tprof_lock
63 */
64
65 typedef struct tprof_buf {
66 u_int b_used;
67 u_int b_size;
68 u_int b_overflow;
69 u_int b_unused;
70 STAILQ_ENTRY(tprof_buf) b_list;
71 tprof_sample_t b_data[];
72 } tprof_buf_t;
73 #define TPROF_BUF_BYTESIZE(sz) \
74 (sizeof(tprof_buf_t) + (sz) * sizeof(tprof_sample_t))
75 #define TPROF_MAX_SAMPLES_PER_BUF 10000
76
77 #define TPROF_MAX_BUF 100
78
79 typedef struct {
80 tprof_buf_t *c_buf;
81 uint32_t c_cpuid;
82 struct work c_work;
83 callout_t c_callout;
84 } __aligned(CACHE_LINE_SIZE) tprof_cpu_t;
85
86 typedef struct tprof_backend {
87 const char *tb_name;
88 const tprof_backend_ops_t *tb_ops;
89 LIST_ENTRY(tprof_backend) tb_list;
90 int tb_usecount; /* S: */
91 } tprof_backend_t;
92
93 static kmutex_t tprof_lock;
94 static bool tprof_running; /* s: */
95 static u_int tprof_nworker; /* L: # of running worker LWPs */
96 static lwp_t *tprof_owner;
97 static STAILQ_HEAD(, tprof_buf) tprof_list; /* L: global buffer list */
98 static u_int tprof_nbuf_on_list; /* L: # of buffers on tprof_list */
99 static struct workqueue *tprof_wq;
100 static tprof_cpu_t tprof_cpus[MAXCPUS] __aligned(CACHE_LINE_SIZE);
101 static u_int tprof_samples_per_buf;
102
103 static tprof_backend_t *tprof_backend; /* S: */
104 static LIST_HEAD(, tprof_backend) tprof_backends =
105 LIST_HEAD_INITIALIZER(tprof_backend); /* S: */
106
107 static kmutex_t tprof_reader_lock;
108 static kcondvar_t tprof_reader_cv; /* L: */
109 static off_t tprof_reader_offset; /* R: */
110
111 static kmutex_t tprof_startstop_lock;
112 static kcondvar_t tprof_cv; /* L: */
113
114 static struct tprof_stat tprof_stat; /* L: */
115
116 static tprof_cpu_t *
117 tprof_cpu(struct cpu_info *ci)
118 {
119
120 return &tprof_cpus[cpu_index(ci)];
121 }
122
123 static tprof_cpu_t *
124 tprof_curcpu(void)
125 {
126
127 return tprof_cpu(curcpu());
128 }
129
130 static tprof_buf_t *
131 tprof_buf_alloc(void)
132 {
133 tprof_buf_t *new;
134 u_int size = tprof_samples_per_buf;
135
136 new = kmem_alloc(TPROF_BUF_BYTESIZE(size), KM_SLEEP);
137 new->b_used = 0;
138 new->b_size = size;
139 new->b_overflow = 0;
140 return new;
141 }
142
143 static void
144 tprof_buf_free(tprof_buf_t *buf)
145 {
146
147 kmem_free(buf, TPROF_BUF_BYTESIZE(buf->b_size));
148 }
149
150 static tprof_buf_t *
151 tprof_buf_switch(tprof_cpu_t *c, tprof_buf_t *new)
152 {
153 tprof_buf_t *old;
154
155 old = c->c_buf;
156 c->c_buf = new;
157 return old;
158 }
159
160 static tprof_buf_t *
161 tprof_buf_refresh(void)
162 {
163 tprof_cpu_t * const c = tprof_curcpu();
164 tprof_buf_t *new;
165
166 new = tprof_buf_alloc();
167 return tprof_buf_switch(c, new);
168 }
169
170 static void
171 tprof_worker(struct work *wk, void *dummy)
172 {
173 tprof_cpu_t * const c = tprof_curcpu();
174 tprof_buf_t *buf;
175 bool shouldstop;
176
177 KASSERT(wk == &c->c_work);
178 KASSERT(dummy == NULL);
179
180 /*
181 * get a per cpu buffer.
182 */
183 buf = tprof_buf_refresh();
184
185 /*
186 * and put it on the global list for read(2).
187 */
188 mutex_enter(&tprof_lock);
189 shouldstop = !tprof_running;
190 if (shouldstop) {
191 KASSERT(tprof_nworker > 0);
192 tprof_nworker--;
193 cv_broadcast(&tprof_cv);
194 cv_broadcast(&tprof_reader_cv);
195 }
196 if (buf->b_used == 0) {
197 tprof_stat.ts_emptybuf++;
198 } else if (tprof_nbuf_on_list < TPROF_MAX_BUF) {
199 tprof_stat.ts_sample += buf->b_used;
200 tprof_stat.ts_overflow += buf->b_overflow;
201 tprof_stat.ts_buf++;
202 STAILQ_INSERT_TAIL(&tprof_list, buf, b_list);
203 tprof_nbuf_on_list++;
204 buf = NULL;
205 cv_broadcast(&tprof_reader_cv);
206 } else {
207 tprof_stat.ts_dropbuf_sample += buf->b_used;
208 tprof_stat.ts_dropbuf++;
209 }
210 mutex_exit(&tprof_lock);
211 if (buf) {
212 tprof_buf_free(buf);
213 }
214 if (!shouldstop) {
215 callout_schedule(&c->c_callout, hz);
216 }
217 }
218
219 static void
220 tprof_kick(void *vp)
221 {
222 struct cpu_info * const ci = vp;
223 tprof_cpu_t * const c = tprof_cpu(ci);
224
225 workqueue_enqueue(tprof_wq, &c->c_work, ci);
226 }
227
228 static void
229 tprof_stop1(void)
230 {
231 CPU_INFO_ITERATOR cii;
232 struct cpu_info *ci;
233
234 KASSERT(mutex_owned(&tprof_startstop_lock));
235 KASSERT(tprof_nworker == 0);
236
237 for (CPU_INFO_FOREACH(cii, ci)) {
238 tprof_cpu_t * const c = tprof_cpu(ci);
239 tprof_buf_t *old;
240
241 old = tprof_buf_switch(c, NULL);
242 if (old != NULL) {
243 tprof_buf_free(old);
244 }
245 callout_destroy(&c->c_callout);
246 }
247 workqueue_destroy(tprof_wq);
248 }
249
250 static int
251 tprof_start(const struct tprof_param *param)
252 {
253 CPU_INFO_ITERATOR cii;
254 struct cpu_info *ci;
255 int error;
256 uint64_t freq;
257 tprof_backend_t *tb;
258
259 KASSERT(mutex_owned(&tprof_startstop_lock));
260 if (tprof_running) {
261 error = EBUSY;
262 goto done;
263 }
264
265 tb = tprof_backend;
266 if (tb == NULL) {
267 error = ENOENT;
268 goto done;
269 }
270 if (tb->tb_usecount > 0) {
271 error = EBUSY;
272 goto done;
273 }
274
275 tb->tb_usecount++;
276 freq = tb->tb_ops->tbo_estimate_freq();
277 tprof_samples_per_buf = MIN(freq * 2, TPROF_MAX_SAMPLES_PER_BUF);
278
279 error = workqueue_create(&tprof_wq, "tprofmv", tprof_worker, NULL,
280 PRI_NONE, IPL_SOFTCLOCK, WQ_MPSAFE | WQ_PERCPU);
281 if (error != 0) {
282 goto done;
283 }
284
285 for (CPU_INFO_FOREACH(cii, ci)) {
286 tprof_cpu_t * const c = tprof_cpu(ci);
287 tprof_buf_t *new;
288 tprof_buf_t *old;
289
290 new = tprof_buf_alloc();
291 old = tprof_buf_switch(c, new);
292 if (old != NULL) {
293 tprof_buf_free(old);
294 }
295 callout_init(&c->c_callout, CALLOUT_MPSAFE);
296 callout_setfunc(&c->c_callout, tprof_kick, ci);
297 }
298
299 error = tb->tb_ops->tbo_start(NULL);
300 if (error != 0) {
301 KASSERT(tb->tb_usecount > 0);
302 tb->tb_usecount--;
303 tprof_stop1();
304 goto done;
305 }
306
307 mutex_enter(&tprof_lock);
308 tprof_running = true;
309 mutex_exit(&tprof_lock);
310 for (CPU_INFO_FOREACH(cii, ci)) {
311 tprof_cpu_t * const c = tprof_cpu(ci);
312
313 mutex_enter(&tprof_lock);
314 tprof_nworker++;
315 mutex_exit(&tprof_lock);
316 workqueue_enqueue(tprof_wq, &c->c_work, ci);
317 }
318 done:
319 return error;
320 }
321
322 static void
323 tprof_stop(void)
324 {
325 tprof_backend_t *tb;
326
327 KASSERT(mutex_owned(&tprof_startstop_lock));
328 if (!tprof_running) {
329 goto done;
330 }
331
332 tb = tprof_backend;
333 KASSERT(tb->tb_usecount > 0);
334 tb->tb_ops->tbo_stop(NULL);
335 tb->tb_usecount--;
336
337 mutex_enter(&tprof_lock);
338 tprof_running = false;
339 cv_broadcast(&tprof_reader_cv);
340 while (tprof_nworker > 0) {
341 cv_wait(&tprof_cv, &tprof_lock);
342 }
343 mutex_exit(&tprof_lock);
344
345 tprof_stop1();
346 done:
347 ;
348 }
349
350 /*
351 * tprof_clear: drain unread samples.
352 */
353
354 static void
355 tprof_clear(void)
356 {
357 tprof_buf_t *buf;
358
359 mutex_enter(&tprof_reader_lock);
360 mutex_enter(&tprof_lock);
361 while ((buf = STAILQ_FIRST(&tprof_list)) != NULL) {
362 if (buf != NULL) {
363 STAILQ_REMOVE_HEAD(&tprof_list, b_list);
364 KASSERT(tprof_nbuf_on_list > 0);
365 tprof_nbuf_on_list--;
366 mutex_exit(&tprof_lock);
367 tprof_buf_free(buf);
368 mutex_enter(&tprof_lock);
369 }
370 }
371 KASSERT(tprof_nbuf_on_list == 0);
372 mutex_exit(&tprof_lock);
373 tprof_reader_offset = 0;
374 mutex_exit(&tprof_reader_lock);
375
376 memset(&tprof_stat, 0, sizeof(tprof_stat));
377 }
378
379 static tprof_backend_t *
380 tprof_backend_lookup(const char *name)
381 {
382 tprof_backend_t *tb;
383
384 KASSERT(mutex_owned(&tprof_startstop_lock));
385
386 LIST_FOREACH(tb, &tprof_backends, tb_list) {
387 if (!strcmp(tb->tb_name, name)) {
388 return tb;
389 }
390 }
391 return NULL;
392 }
393
394 /* -------------------- backend interfaces */
395
396 /*
397 * tprof_sample: record a sample on the per-cpu buffer.
398 *
399 * be careful; can be called in NMI context.
400 * we are bluntly assuming the followings are safe.
401 * curcpu()
402 * curlwp->l_lid
403 * curlwp->l_proc->p_pid
404 */
405
406 void
407 tprof_sample(tprof_backend_cookie_t *cookie, const tprof_frame_info_t *tfi)
408 {
409 tprof_cpu_t * const c = tprof_curcpu();
410 tprof_buf_t * const buf = c->c_buf;
411 tprof_sample_t *sp;
412 const uintptr_t pc = tfi->tfi_pc;
413 const lwp_t * const l = curlwp;
414 u_int idx;
415
416 idx = buf->b_used;
417 if (__predict_false(idx >= buf->b_size)) {
418 buf->b_overflow++;
419 return;
420 }
421 sp = &buf->b_data[idx];
422 sp->s_pid = l->l_proc->p_pid;
423 sp->s_lwpid = l->l_lid;
424 sp->s_cpuid = c->c_cpuid;
425 sp->s_flags = (tfi->tfi_inkernel) ? TPROF_SAMPLE_INKERNEL : 0;
426 sp->s_pc = pc;
427 buf->b_used = idx + 1;
428 }
429
430 /*
431 * tprof_backend_register:
432 */
433
434 int
435 tprof_backend_register(const char *name, const tprof_backend_ops_t *ops,
436 int vers)
437 {
438 tprof_backend_t *tb;
439
440 if (vers != TPROF_BACKEND_VERSION) {
441 return EINVAL;
442 }
443
444 mutex_enter(&tprof_startstop_lock);
445 tb = tprof_backend_lookup(name);
446 if (tb != NULL) {
447 mutex_exit(&tprof_startstop_lock);
448 return EEXIST;
449 }
450 #if 1 /* XXX for now */
451 if (!LIST_EMPTY(&tprof_backends)) {
452 mutex_exit(&tprof_startstop_lock);
453 return ENOTSUP;
454 }
455 #endif
456 tb = kmem_alloc(sizeof(*tb), KM_SLEEP);
457 tb->tb_name = name;
458 tb->tb_ops = ops;
459 tb->tb_usecount = 0;
460 LIST_INSERT_HEAD(&tprof_backends, tb, tb_list);
461 #if 1 /* XXX for now */
462 if (tprof_backend == NULL) {
463 tprof_backend = tb;
464 }
465 #endif
466 mutex_exit(&tprof_startstop_lock);
467
468 return 0;
469 }
470
471 /*
472 * tprof_backend_unregister:
473 */
474
475 int
476 tprof_backend_unregister(const char *name)
477 {
478 tprof_backend_t *tb;
479
480 mutex_enter(&tprof_startstop_lock);
481 tb = tprof_backend_lookup(name);
482 #if defined(DIAGNOSTIC)
483 if (tb == NULL) {
484 mutex_exit(&tprof_startstop_lock);
485 panic("%s: not found '%s'", __func__, name);
486 }
487 #endif /* defined(DIAGNOSTIC) */
488 if (tb->tb_usecount > 0) {
489 mutex_exit(&tprof_startstop_lock);
490 return EBUSY;
491 }
492 #if 1 /* XXX for now */
493 if (tprof_backend == tb) {
494 tprof_backend = NULL;
495 }
496 #endif
497 LIST_REMOVE(tb, tb_list);
498 mutex_exit(&tprof_startstop_lock);
499
500 kmem_free(tb, sizeof(*tb));
501
502 return 0;
503 }
504
505 /* -------------------- cdevsw interfaces */
506
507 static int
508 tprof_open(dev_t dev, int flags, int type, struct lwp *l)
509 {
510
511 if (minor(dev) != 0) {
512 return EXDEV;
513 }
514 mutex_enter(&tprof_lock);
515 if (tprof_owner != NULL) {
516 mutex_exit(&tprof_lock);
517 return EBUSY;
518 }
519 tprof_owner = curlwp;
520 mutex_exit(&tprof_lock);
521
522 return 0;
523 }
524
525 static int
526 tprof_close(dev_t dev, int flags, int type, struct lwp *l)
527 {
528
529 KASSERT(minor(dev) == 0);
530
531 mutex_enter(&tprof_startstop_lock);
532 mutex_enter(&tprof_lock);
533 tprof_owner = NULL;
534 mutex_exit(&tprof_lock);
535 tprof_stop();
536 tprof_clear();
537 mutex_exit(&tprof_startstop_lock);
538
539 return 0;
540 }
541
542 static int
543 tprof_read(dev_t dev, struct uio *uio, int flags)
544 {
545 tprof_buf_t *buf;
546 size_t bytes;
547 size_t resid;
548 size_t done;
549 int error = 0;
550
551 KASSERT(minor(dev) == 0);
552 mutex_enter(&tprof_reader_lock);
553 while (uio->uio_resid > 0 && error == 0) {
554 /*
555 * take the first buffer from the list.
556 */
557 mutex_enter(&tprof_lock);
558 buf = STAILQ_FIRST(&tprof_list);
559 if (buf == NULL) {
560 if (tprof_nworker == 0) {
561 mutex_exit(&tprof_lock);
562 error = 0;
563 break;
564 }
565 mutex_exit(&tprof_reader_lock);
566 error = cv_wait_sig(&tprof_reader_cv, &tprof_lock);
567 mutex_exit(&tprof_lock);
568 mutex_enter(&tprof_reader_lock);
569 continue;
570 }
571 STAILQ_REMOVE_HEAD(&tprof_list, b_list);
572 KASSERT(tprof_nbuf_on_list > 0);
573 tprof_nbuf_on_list--;
574 mutex_exit(&tprof_lock);
575
576 /*
577 * copy it out.
578 */
579 bytes = MIN(buf->b_used * sizeof(tprof_sample_t) -
580 tprof_reader_offset, uio->uio_resid);
581 resid = uio->uio_resid;
582 error = uiomove((char *)buf->b_data + tprof_reader_offset,
583 bytes, uio);
584 done = resid - uio->uio_resid;
585 tprof_reader_offset += done;
586
587 /*
588 * if we didn't consume the whole buffer,
589 * put it back to the list.
590 */
591 if (tprof_reader_offset <
592 buf->b_used * sizeof(tprof_sample_t)) {
593 mutex_enter(&tprof_lock);
594 STAILQ_INSERT_HEAD(&tprof_list, buf, b_list);
595 tprof_nbuf_on_list++;
596 cv_broadcast(&tprof_reader_cv);
597 mutex_exit(&tprof_lock);
598 } else {
599 tprof_buf_free(buf);
600 tprof_reader_offset = 0;
601 }
602 }
603 mutex_exit(&tprof_reader_lock);
604
605 return error;
606 }
607
608 static int
609 tprof_ioctl(dev_t dev, u_long cmd, void *data, int flags, struct lwp *l)
610 {
611 const struct tprof_param *param;
612 int error = 0;
613
614 KASSERT(minor(dev) == 0);
615
616 switch (cmd) {
617 case TPROF_IOC_GETVERSION:
618 *(int *)data = TPROF_VERSION;
619 break;
620 case TPROF_IOC_START:
621 param = data;
622 mutex_enter(&tprof_startstop_lock);
623 error = tprof_start(param);
624 mutex_exit(&tprof_startstop_lock);
625 break;
626 case TPROF_IOC_STOP:
627 mutex_enter(&tprof_startstop_lock);
628 tprof_stop();
629 mutex_exit(&tprof_startstop_lock);
630 break;
631 case TPROF_IOC_GETSTAT:
632 mutex_enter(&tprof_lock);
633 memcpy(data, &tprof_stat, sizeof(tprof_stat));
634 mutex_exit(&tprof_lock);
635 break;
636 default:
637 error = EINVAL;
638 break;
639 }
640
641 return error;
642 }
643
644 const struct cdevsw tprof_cdevsw = {
645 .d_open = tprof_open,
646 .d_close = tprof_close,
647 .d_read = tprof_read,
648 .d_write = nowrite,
649 .d_ioctl = tprof_ioctl,
650 .d_stop = nostop,
651 .d_tty = notty,
652 .d_poll = nopoll,
653 .d_mmap = nommap,
654 .d_kqfilter = nokqfilter,
655 .d_discard = nodiscard,
656 .d_flag = D_OTHER | D_MPSAFE
657 };
658
659 void
660 tprofattach(int nunits)
661 {
662
663 /* nothing */
664 }
665
666 MODULE(MODULE_CLASS_DRIVER, tprof, NULL);
667
668 static void
669 tprof_driver_init(void)
670 {
671 unsigned int i;
672
673 mutex_init(&tprof_lock, MUTEX_DEFAULT, IPL_NONE);
674 mutex_init(&tprof_reader_lock, MUTEX_DEFAULT, IPL_NONE);
675 mutex_init(&tprof_startstop_lock, MUTEX_DEFAULT, IPL_NONE);
676 cv_init(&tprof_cv, "tprof");
677 cv_init(&tprof_reader_cv, "tprof_rd");
678 STAILQ_INIT(&tprof_list);
679 for (i = 0; i < __arraycount(tprof_cpus); i++) {
680 tprof_cpu_t * const c = &tprof_cpus[i];
681
682 c->c_buf = NULL;
683 c->c_cpuid = i;
684 }
685 }
686
687 static void
688 tprof_driver_fini(void)
689 {
690
691 mutex_destroy(&tprof_lock);
692 mutex_destroy(&tprof_reader_lock);
693 mutex_destroy(&tprof_startstop_lock);
694 cv_destroy(&tprof_cv);
695 cv_destroy(&tprof_reader_cv);
696 }
697
698 static int
699 tprof_modcmd(modcmd_t cmd, void *arg)
700 {
701
702 switch (cmd) {
703 case MODULE_CMD_INIT:
704 tprof_driver_init();
705 #if defined(_MODULE)
706 {
707 devmajor_t bmajor = NODEVMAJOR;
708 devmajor_t cmajor = NODEVMAJOR;
709 int error;
710
711 error = devsw_attach("tprof", NULL, &bmajor,
712 &tprof_cdevsw, &cmajor);
713 if (error) {
714 tprof_driver_fini();
715 return error;
716 }
717 }
718 #endif /* defined(_MODULE) */
719 return 0;
720
721 case MODULE_CMD_FINI:
722 #if defined(_MODULE)
723 {
724 int error;
725 error = devsw_detach(NULL, &tprof_cdevsw);
726 if (error) {
727 return error;
728 }
729 }
730 #endif /* defined(_MODULE) */
731 tprof_driver_fini();
732 return 0;
733
734 default:
735 return ENOTTY;
736 }
737 }
738