xref: /freebsd/sys/kern/subr_devstat.c (revision 480093f4440d54b30b3025afeac24b48f2ba7a2e)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1997, 1998, 1999 Kenneth D. Merry.
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  * 3. The name of the author may not be used to endorse or promote products
16  *    derived from this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/param.h>
35 #include <sys/kernel.h>
36 #include <sys/systm.h>
37 #include <sys/bio.h>
38 #include <sys/devicestat.h>
39 #include <sys/sdt.h>
40 #include <sys/sysctl.h>
41 #include <sys/malloc.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/conf.h>
45 #include <vm/vm.h>
46 #include <vm/pmap.h>
47 
48 #include <machine/atomic.h>
49 
50 SDT_PROVIDER_DEFINE(io);
51 
52 SDT_PROBE_DEFINE2(io, , , start, "struct bio *", "struct devstat *");
53 SDT_PROBE_DEFINE2(io, , , done, "struct bio *", "struct devstat *");
54 SDT_PROBE_DEFINE2(io, , , wait__start, "struct bio *",
55     "struct devstat *");
56 SDT_PROBE_DEFINE2(io, , , wait__done, "struct bio *",
57     "struct devstat *");
58 
59 #define	DTRACE_DEVSTAT_START()		SDT_PROBE2(io, , , start, NULL, ds)
60 #define	DTRACE_DEVSTAT_BIO_START()	SDT_PROBE2(io, , , start, bp, ds)
61 #define	DTRACE_DEVSTAT_DONE()		SDT_PROBE2(io, , , done, NULL, ds)
62 #define	DTRACE_DEVSTAT_BIO_DONE()	SDT_PROBE2(io, , , done, bp, ds)
63 #define	DTRACE_DEVSTAT_WAIT_START()	SDT_PROBE2(io, , , wait__start, NULL, ds)
64 #define	DTRACE_DEVSTAT_WAIT_DONE()	SDT_PROBE2(io, , , wait__done, NULL, ds)
65 
66 static int devstat_num_devs;
67 static long devstat_generation = 1;
68 static int devstat_version = DEVSTAT_VERSION;
69 static int devstat_current_devnumber;
70 static struct mtx devstat_mutex;
71 MTX_SYSINIT(devstat_mutex, &devstat_mutex, "devstat", MTX_DEF);
72 
73 static struct devstatlist device_statq = STAILQ_HEAD_INITIALIZER(device_statq);
74 static struct devstat *devstat_alloc(void);
75 static void devstat_free(struct devstat *);
76 static void devstat_add_entry(struct devstat *ds, const void *dev_name,
77 		       int unit_number, uint32_t block_size,
78 		       devstat_support_flags flags,
79 		       devstat_type_flags device_type,
80 		       devstat_priority priority);
81 
82 /*
83  * Allocate a devstat and initialize it
84  */
85 struct devstat *
86 devstat_new_entry(const void *dev_name,
87 		  int unit_number, uint32_t block_size,
88 		  devstat_support_flags flags,
89 		  devstat_type_flags device_type,
90 		  devstat_priority priority)
91 {
92 	struct devstat *ds;
93 
94 	mtx_assert(&devstat_mutex, MA_NOTOWNED);
95 
96 	ds = devstat_alloc();
97 	mtx_lock(&devstat_mutex);
98 	if (unit_number == -1) {
99 		ds->unit_number = unit_number;
100 		ds->id = dev_name;
101 		binuptime(&ds->creation_time);
102 		devstat_generation++;
103 	} else {
104 		devstat_add_entry(ds, dev_name, unit_number, block_size,
105 				  flags, device_type, priority);
106 	}
107 	mtx_unlock(&devstat_mutex);
108 	return (ds);
109 }
110 
111 /*
112  * Take a malloced and zeroed devstat structure given to us, fill it in
113  * and add it to the queue of devices.
114  */
115 static void
116 devstat_add_entry(struct devstat *ds, const void *dev_name,
117 		  int unit_number, uint32_t block_size,
118 		  devstat_support_flags flags,
119 		  devstat_type_flags device_type,
120 		  devstat_priority priority)
121 {
122 	struct devstatlist *devstat_head;
123 	struct devstat *ds_tmp;
124 
125 	mtx_assert(&devstat_mutex, MA_OWNED);
126 	devstat_num_devs++;
127 
128 	devstat_head = &device_statq;
129 
130 	/*
131 	 * Priority sort.  Each driver passes in its priority when it adds
132 	 * its devstat entry.  Drivers are sorted first by priority, and
133 	 * then by probe order.
134 	 *
135 	 * For the first device, we just insert it, since the priority
136 	 * doesn't really matter yet.  Subsequent devices are inserted into
137 	 * the list using the order outlined above.
138 	 */
139 	if (devstat_num_devs == 1)
140 		STAILQ_INSERT_TAIL(devstat_head, ds, dev_links);
141 	else {
142 		STAILQ_FOREACH(ds_tmp, devstat_head, dev_links) {
143 			struct devstat *ds_next;
144 
145 			ds_next = STAILQ_NEXT(ds_tmp, dev_links);
146 
147 			/*
148 			 * If we find a break between higher and lower
149 			 * priority items, and if this item fits in the
150 			 * break, insert it.  This also applies if the
151 			 * "lower priority item" is the end of the list.
152 			 */
153 			if ((priority <= ds_tmp->priority)
154 			 && ((ds_next == NULL)
155 			   || (priority > ds_next->priority))) {
156 				STAILQ_INSERT_AFTER(devstat_head, ds_tmp, ds,
157 						    dev_links);
158 				break;
159 			} else if (priority > ds_tmp->priority) {
160 				/*
161 				 * If this is the case, we should be able
162 				 * to insert ourselves at the head of the
163 				 * list.  If we can't, something is wrong.
164 				 */
165 				if (ds_tmp == STAILQ_FIRST(devstat_head)) {
166 					STAILQ_INSERT_HEAD(devstat_head,
167 							   ds, dev_links);
168 					break;
169 				} else {
170 					STAILQ_INSERT_TAIL(devstat_head,
171 							   ds, dev_links);
172 					printf("devstat_add_entry: HELP! "
173 					       "sorting problem detected "
174 					       "for name %p unit %d\n",
175 					       dev_name, unit_number);
176 					break;
177 				}
178 			}
179 		}
180 	}
181 
182 	ds->device_number = devstat_current_devnumber++;
183 	ds->unit_number = unit_number;
184 	strlcpy(ds->device_name, dev_name, DEVSTAT_NAME_LEN);
185 	ds->block_size = block_size;
186 	ds->flags = flags;
187 	ds->device_type = device_type;
188 	ds->priority = priority;
189 	binuptime(&ds->creation_time);
190 	devstat_generation++;
191 }
192 
193 /*
194  * Remove a devstat structure from the list of devices.
195  */
196 void
197 devstat_remove_entry(struct devstat *ds)
198 {
199 	struct devstatlist *devstat_head;
200 
201 	mtx_assert(&devstat_mutex, MA_NOTOWNED);
202 	if (ds == NULL)
203 		return;
204 
205 	mtx_lock(&devstat_mutex);
206 
207 	devstat_head = &device_statq;
208 
209 	/* Remove this entry from the devstat queue */
210 	atomic_add_acq_int(&ds->sequence1, 1);
211 	if (ds->unit_number != -1) {
212 		devstat_num_devs--;
213 		STAILQ_REMOVE(devstat_head, ds, devstat, dev_links);
214 	}
215 	devstat_free(ds);
216 	devstat_generation++;
217 	mtx_unlock(&devstat_mutex);
218 }
219 
220 /*
221  * Record a transaction start.
222  *
223  * See comments for devstat_end_transaction().  Ordering is very important
224  * here.
225  */
226 void
227 devstat_start_transaction(struct devstat *ds, const struct bintime *now)
228 {
229 
230 	/* sanity check */
231 	if (ds == NULL)
232 		return;
233 
234 	atomic_add_acq_int(&ds->sequence1, 1);
235 	/*
236 	 * We only want to set the start time when we are going from idle
237 	 * to busy.  The start time is really the start of the latest busy
238 	 * period.
239 	 */
240 	if (atomic_fetchadd_int(&ds->start_count, 1) == ds->end_count) {
241 		if (now != NULL)
242 			ds->busy_from = *now;
243 		else
244 			binuptime(&ds->busy_from);
245 	}
246 	atomic_add_rel_int(&ds->sequence0, 1);
247 	DTRACE_DEVSTAT_START();
248 }
249 
250 void
251 devstat_start_transaction_bio(struct devstat *ds, struct bio *bp)
252 {
253 
254 	/* sanity check */
255 	if (ds == NULL)
256 		return;
257 
258 	binuptime(&bp->bio_t0);
259 	devstat_start_transaction(ds, &bp->bio_t0);
260 	DTRACE_DEVSTAT_BIO_START();
261 }
262 
263 /*
264  * Record the ending of a transaction, and incrment the various counters.
265  *
266  * Ordering in this function, and in devstat_start_transaction() is VERY
267  * important.  The idea here is to run without locks, so we are very
268  * careful to only modify some fields on the way "down" (i.e. at
269  * transaction start) and some fields on the way "up" (i.e. at transaction
270  * completion).  One exception is busy_from, which we only modify in
271  * devstat_start_transaction() when there are no outstanding transactions,
272  * and thus it can't be modified in devstat_end_transaction()
273  * simultaneously.
274  *
275  * The sequence0 and sequence1 fields are provided to enable an application
276  * spying on the structures with mmap(2) to tell when a structure is in a
277  * consistent state or not.
278  *
279  * For this to work 100% reliably, it is important that the two fields
280  * are at opposite ends of the structure and that they are incremented
281  * in the opposite order of how a memcpy(3) in userland would copy them.
282  * We assume that the copying happens front to back, but there is actually
283  * no way short of writing your own memcpy(3) replacement to guarantee
284  * this will be the case.
285  *
286  * In addition to this, being a kind of locks, they must be updated with
287  * atomic instructions using appropriate memory barriers.
288  */
289 void
290 devstat_end_transaction(struct devstat *ds, uint32_t bytes,
291 			devstat_tag_type tag_type, devstat_trans_flags flags,
292 			const struct bintime *now, const struct bintime *then)
293 {
294 	struct bintime dt, lnow;
295 
296 	/* sanity check */
297 	if (ds == NULL)
298 		return;
299 
300 	if (now == NULL) {
301 		binuptime(&lnow);
302 		now = &lnow;
303 	}
304 
305 	atomic_add_acq_int(&ds->sequence1, 1);
306 	/* Update byte and operations counts */
307 	ds->bytes[flags] += bytes;
308 	ds->operations[flags]++;
309 
310 	/*
311 	 * Keep a count of the various tag types sent.
312 	 */
313 	if ((ds->flags & DEVSTAT_NO_ORDERED_TAGS) == 0 &&
314 	    tag_type != DEVSTAT_TAG_NONE)
315 		ds->tag_types[tag_type]++;
316 
317 	if (then != NULL) {
318 		/* Update duration of operations */
319 		dt = *now;
320 		bintime_sub(&dt, then);
321 		bintime_add(&ds->duration[flags], &dt);
322 	}
323 
324 	/* Accumulate busy time */
325 	dt = *now;
326 	bintime_sub(&dt, &ds->busy_from);
327 	bintime_add(&ds->busy_time, &dt);
328 	ds->busy_from = *now;
329 
330 	ds->end_count++;
331 	atomic_add_rel_int(&ds->sequence0, 1);
332 	DTRACE_DEVSTAT_DONE();
333 }
334 
335 void
336 devstat_end_transaction_bio(struct devstat *ds, const struct bio *bp)
337 {
338 
339 	devstat_end_transaction_bio_bt(ds, bp, NULL);
340 }
341 
342 void
343 devstat_end_transaction_bio_bt(struct devstat *ds, const struct bio *bp,
344     const struct bintime *now)
345 {
346 	devstat_trans_flags flg;
347 	devstat_tag_type tag;
348 
349 	/* sanity check */
350 	if (ds == NULL)
351 		return;
352 
353 	if (bp->bio_flags & BIO_ORDERED)
354 		tag = DEVSTAT_TAG_ORDERED;
355 	else
356 		tag = DEVSTAT_TAG_SIMPLE;
357 	if (bp->bio_cmd == BIO_DELETE)
358 		flg = DEVSTAT_FREE;
359 	else if ((bp->bio_cmd == BIO_READ)
360 	      || ((bp->bio_cmd == BIO_ZONE)
361 	       && (bp->bio_zone.zone_cmd == DISK_ZONE_REPORT_ZONES)))
362 		flg = DEVSTAT_READ;
363 	else if (bp->bio_cmd == BIO_WRITE)
364 		flg = DEVSTAT_WRITE;
365 	else
366 		flg = DEVSTAT_NO_DATA;
367 
368 	devstat_end_transaction(ds, bp->bio_bcount - bp->bio_resid,
369 				tag, flg, now, &bp->bio_t0);
370 	DTRACE_DEVSTAT_BIO_DONE();
371 }
372 
373 /*
374  * This is the sysctl handler for the devstat package.  The data pushed out
375  * on the kern.devstat.all sysctl variable consists of the current devstat
376  * generation number, and then an array of devstat structures, one for each
377  * device in the system.
378  *
379  * This is more cryptic that obvious, but basically we neither can nor
380  * want to hold the devstat_mutex for any amount of time, so we grab it
381  * only when we need to and keep an eye on devstat_generation all the time.
382  */
383 static int
384 sysctl_devstat(SYSCTL_HANDLER_ARGS)
385 {
386 	int error;
387 	long mygen;
388 	struct devstat *nds;
389 
390 	mtx_assert(&devstat_mutex, MA_NOTOWNED);
391 
392 	/*
393 	 * XXX devstat_generation should really be "volatile" but that
394 	 * XXX freaks out the sysctl macro below.  The places where we
395 	 * XXX change it and inspect it are bracketed in the mutex which
396 	 * XXX guarantees us proper write barriers.  I don't believe the
397 	 * XXX compiler is allowed to optimize mygen away across calls
398 	 * XXX to other functions, so the following is belived to be safe.
399 	 */
400 	mygen = devstat_generation;
401 
402 	error = SYSCTL_OUT(req, &mygen, sizeof(mygen));
403 
404 	if (devstat_num_devs == 0)
405 		return(0);
406 
407 	if (error != 0)
408 		return (error);
409 
410 	mtx_lock(&devstat_mutex);
411 	nds = STAILQ_FIRST(&device_statq);
412 	if (mygen != devstat_generation)
413 		error = EBUSY;
414 	mtx_unlock(&devstat_mutex);
415 
416 	if (error != 0)
417 		return (error);
418 
419 	for (;nds != NULL;) {
420 		error = SYSCTL_OUT(req, nds, sizeof(struct devstat));
421 		if (error != 0)
422 			return (error);
423 		mtx_lock(&devstat_mutex);
424 		if (mygen != devstat_generation)
425 			error = EBUSY;
426 		else
427 			nds = STAILQ_NEXT(nds, dev_links);
428 		mtx_unlock(&devstat_mutex);
429 		if (error != 0)
430 			return (error);
431 	}
432 	return(error);
433 }
434 
435 /*
436  * Sysctl entries for devstat.  The first one is a node that all the rest
437  * hang off of.
438  */
439 static SYSCTL_NODE(_kern, OID_AUTO, devstat, CTLFLAG_RD, NULL,
440     "Device Statistics");
441 
442 SYSCTL_PROC(_kern_devstat, OID_AUTO, all, CTLFLAG_RD|CTLTYPE_OPAQUE,
443     NULL, 0, sysctl_devstat, "S,devstat", "All devices in the devstat list");
444 /*
445  * Export the number of devices in the system so that userland utilities
446  * can determine how much memory to allocate to hold all the devices.
447  */
448 SYSCTL_INT(_kern_devstat, OID_AUTO, numdevs, CTLFLAG_RD,
449     &devstat_num_devs, 0, "Number of devices in the devstat list");
450 SYSCTL_LONG(_kern_devstat, OID_AUTO, generation, CTLFLAG_RD,
451     &devstat_generation, 0, "Devstat list generation");
452 SYSCTL_INT(_kern_devstat, OID_AUTO, version, CTLFLAG_RD,
453     &devstat_version, 0, "Devstat list version number");
454 
455 /*
456  * Allocator for struct devstat structures.  We sub-allocate these from pages
457  * which we get from malloc.  These pages are exported for mmap(2)'ing through
458  * a miniature device driver
459  */
460 
461 #define statsperpage (PAGE_SIZE / sizeof(struct devstat))
462 
463 static d_mmap_t devstat_mmap;
464 
465 static struct cdevsw devstat_cdevsw = {
466 	.d_version =	D_VERSION,
467 	.d_mmap =	devstat_mmap,
468 	.d_name =	"devstat",
469 };
470 
471 struct statspage {
472 	TAILQ_ENTRY(statspage)	list;
473 	struct devstat		*stat;
474 	u_int			nfree;
475 };
476 
477 static TAILQ_HEAD(, statspage)	pagelist = TAILQ_HEAD_INITIALIZER(pagelist);
478 static MALLOC_DEFINE(M_DEVSTAT, "devstat", "Device statistics");
479 
480 static int
481 devstat_mmap(struct cdev *dev, vm_ooffset_t offset, vm_paddr_t *paddr,
482     int nprot, vm_memattr_t *memattr)
483 {
484 	struct statspage *spp;
485 
486 	if (nprot != VM_PROT_READ)
487 		return (-1);
488 	mtx_lock(&devstat_mutex);
489 	TAILQ_FOREACH(spp, &pagelist, list) {
490 		if (offset == 0) {
491 			*paddr = vtophys(spp->stat);
492 			mtx_unlock(&devstat_mutex);
493 			return (0);
494 		}
495 		offset -= PAGE_SIZE;
496 	}
497 	mtx_unlock(&devstat_mutex);
498 	return (-1);
499 }
500 
501 static struct devstat *
502 devstat_alloc(void)
503 {
504 	struct devstat *dsp;
505 	struct statspage *spp, *spp2;
506 	u_int u;
507 	static int once;
508 
509 	mtx_assert(&devstat_mutex, MA_NOTOWNED);
510 	if (!once) {
511 		make_dev_credf(MAKEDEV_ETERNAL | MAKEDEV_CHECKNAME,
512 		    &devstat_cdevsw, 0, NULL, UID_ROOT, GID_WHEEL, 0444,
513 		    DEVSTAT_DEVICE_NAME);
514 		once = 1;
515 	}
516 	spp2 = NULL;
517 	mtx_lock(&devstat_mutex);
518 	for (;;) {
519 		TAILQ_FOREACH(spp, &pagelist, list) {
520 			if (spp->nfree > 0)
521 				break;
522 		}
523 		if (spp != NULL)
524 			break;
525 		mtx_unlock(&devstat_mutex);
526 		spp2 = malloc(sizeof *spp, M_DEVSTAT, M_ZERO | M_WAITOK);
527 		spp2->stat = malloc(PAGE_SIZE, M_DEVSTAT, M_ZERO | M_WAITOK);
528 		spp2->nfree = statsperpage;
529 
530 		/*
531 		 * If free statspages were added while the lock was released
532 		 * just reuse them.
533 		 */
534 		mtx_lock(&devstat_mutex);
535 		TAILQ_FOREACH(spp, &pagelist, list)
536 			if (spp->nfree > 0)
537 				break;
538 		if (spp == NULL) {
539 			spp = spp2;
540 
541 			/*
542 			 * It would make more sense to add the new page at the
543 			 * head but the order on the list determine the
544 			 * sequence of the mapping so we can't do that.
545 			 */
546 			TAILQ_INSERT_TAIL(&pagelist, spp, list);
547 		} else
548 			break;
549 	}
550 	dsp = spp->stat;
551 	for (u = 0; u < statsperpage; u++) {
552 		if (dsp->allocated == 0)
553 			break;
554 		dsp++;
555 	}
556 	spp->nfree--;
557 	dsp->allocated = 1;
558 	mtx_unlock(&devstat_mutex);
559 	if (spp2 != NULL && spp2 != spp) {
560 		free(spp2->stat, M_DEVSTAT);
561 		free(spp2, M_DEVSTAT);
562 	}
563 	return (dsp);
564 }
565 
566 static void
567 devstat_free(struct devstat *dsp)
568 {
569 	struct statspage *spp;
570 
571 	mtx_assert(&devstat_mutex, MA_OWNED);
572 	bzero(dsp, sizeof *dsp);
573 	TAILQ_FOREACH(spp, &pagelist, list) {
574 		if (dsp >= spp->stat && dsp < (spp->stat + statsperpage)) {
575 			spp->nfree++;
576 			return;
577 		}
578 	}
579 }
580 
581 SYSCTL_INT(_debug_sizeof, OID_AUTO, devstat, CTLFLAG_RD,
582     SYSCTL_NULL_INT_PTR, sizeof(struct devstat), "sizeof(struct devstat)");
583