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/param.h>
32#include <sys/disk.h>
33#include <sys/kernel.h>
34#include <sys/systm.h>
35#include <sys/bio.h>
36#include <sys/devicestat.h>
37#include <sys/sdt.h>
38#include <sys/sysctl.h>
39#include <sys/malloc.h>
40#include <sys/lock.h>
41#include <sys/mutex.h>
42#include <sys/conf.h>
43#include <vm/vm.h>
44#include <vm/pmap.h>
45
46#include <machine/atomic.h>
47
48SDT_PROVIDER_DEFINE(io);
49
50SDT_PROBE_DEFINE2(io, , , start, "struct bio *", "struct devstat *");
51SDT_PROBE_DEFINE2(io, , , done, "struct bio *", "struct devstat *");
52
53#define	DTRACE_DEVSTAT_BIO_START()	SDT_PROBE2(io, , , start, bp, ds)
54#define	DTRACE_DEVSTAT_BIO_DONE()	SDT_PROBE2(io, , , done, bp, ds)
55
56static int devstat_num_devs;
57static long devstat_generation = 1;
58static int devstat_version = DEVSTAT_VERSION;
59static int devstat_current_devnumber;
60static struct mtx devstat_mutex;
61MTX_SYSINIT(devstat_mutex, &devstat_mutex, "devstat", MTX_DEF);
62
63static struct devstatlist device_statq = STAILQ_HEAD_INITIALIZER(device_statq);
64static struct devstat *devstat_alloc(void);
65static void devstat_free(struct devstat *);
66static void devstat_add_entry(struct devstat *ds, const void *dev_name,
67		       int unit_number, uint32_t block_size,
68		       devstat_support_flags flags,
69		       devstat_type_flags device_type,
70		       devstat_priority priority);
71
72/*
73 * Allocate a devstat and initialize it
74 */
75struct devstat *
76devstat_new_entry(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	struct devstat *ds;
83
84	mtx_assert(&devstat_mutex, MA_NOTOWNED);
85
86	ds = devstat_alloc();
87	mtx_lock(&devstat_mutex);
88	if (unit_number == -1) {
89		ds->unit_number = unit_number;
90		ds->id = dev_name;
91		binuptime(&ds->creation_time);
92		devstat_generation++;
93	} else {
94		devstat_add_entry(ds, dev_name, unit_number, block_size,
95				  flags, device_type, priority);
96	}
97	mtx_unlock(&devstat_mutex);
98	return (ds);
99}
100
101/*
102 * Take a malloced and zeroed devstat structure given to us, fill it in
103 * and add it to the queue of devices.
104 */
105static void
106devstat_add_entry(struct devstat *ds, const void *dev_name,
107		  int unit_number, uint32_t block_size,
108		  devstat_support_flags flags,
109		  devstat_type_flags device_type,
110		  devstat_priority priority)
111{
112	struct devstatlist *devstat_head;
113	struct devstat *ds_tmp;
114
115	mtx_assert(&devstat_mutex, MA_OWNED);
116	devstat_num_devs++;
117
118	devstat_head = &device_statq;
119
120	/*
121	 * Priority sort.  Each driver passes in its priority when it adds
122	 * its devstat entry.  Drivers are sorted first by priority, and
123	 * then by probe order.
124	 *
125	 * For the first device, we just insert it, since the priority
126	 * doesn't really matter yet.  Subsequent devices are inserted into
127	 * the list using the order outlined above.
128	 */
129	if (devstat_num_devs == 1)
130		STAILQ_INSERT_TAIL(devstat_head, ds, dev_links);
131	else {
132		STAILQ_FOREACH(ds_tmp, devstat_head, dev_links) {
133			struct devstat *ds_next;
134
135			ds_next = STAILQ_NEXT(ds_tmp, dev_links);
136
137			/*
138			 * If we find a break between higher and lower
139			 * priority items, and if this item fits in the
140			 * break, insert it.  This also applies if the
141			 * "lower priority item" is the end of the list.
142			 */
143			if ((priority <= ds_tmp->priority)
144			 && ((ds_next == NULL)
145			   || (priority > ds_next->priority))) {
146				STAILQ_INSERT_AFTER(devstat_head, ds_tmp, ds,
147						    dev_links);
148				break;
149			} else if (priority > ds_tmp->priority) {
150				/*
151				 * If this is the case, we should be able
152				 * to insert ourselves at the head of the
153				 * list.  If we can't, something is wrong.
154				 */
155				if (ds_tmp == STAILQ_FIRST(devstat_head)) {
156					STAILQ_INSERT_HEAD(devstat_head,
157							   ds, dev_links);
158					break;
159				} else {
160					STAILQ_INSERT_TAIL(devstat_head,
161							   ds, dev_links);
162					printf("devstat_add_entry: HELP! "
163					       "sorting problem detected "
164					       "for name %p unit %d\n",
165					       dev_name, unit_number);
166					break;
167				}
168			}
169		}
170	}
171
172	ds->device_number = devstat_current_devnumber++;
173	ds->unit_number = unit_number;
174	strlcpy(ds->device_name, dev_name, DEVSTAT_NAME_LEN);
175	ds->block_size = block_size;
176	ds->flags = flags;
177	ds->device_type = device_type;
178	ds->priority = priority;
179	binuptime(&ds->creation_time);
180	devstat_generation++;
181}
182
183/*
184 * Remove a devstat structure from the list of devices.
185 */
186void
187devstat_remove_entry(struct devstat *ds)
188{
189	struct devstatlist *devstat_head;
190
191	mtx_assert(&devstat_mutex, MA_NOTOWNED);
192	if (ds == NULL)
193		return;
194
195	mtx_lock(&devstat_mutex);
196
197	devstat_head = &device_statq;
198
199	/* Remove this entry from the devstat queue */
200	atomic_add_acq_int(&ds->sequence1, 1);
201	if (ds->unit_number != -1) {
202		devstat_num_devs--;
203		STAILQ_REMOVE(devstat_head, ds, devstat, dev_links);
204	}
205	devstat_free(ds);
206	devstat_generation++;
207	mtx_unlock(&devstat_mutex);
208}
209
210/*
211 * Record a transaction start.
212 *
213 * See comments for devstat_end_transaction().  Ordering is very important
214 * here.
215 */
216void
217devstat_start_transaction(struct devstat *ds, const struct bintime *now)
218{
219
220	/* sanity check */
221	if (ds == NULL)
222		return;
223
224	atomic_add_acq_int(&ds->sequence1, 1);
225	/*
226	 * We only want to set the start time when we are going from idle
227	 * to busy.  The start time is really the start of the latest busy
228	 * period.
229	 */
230	if (atomic_fetchadd_int(&ds->start_count, 1) == ds->end_count) {
231		if (now != NULL)
232			ds->busy_from = *now;
233		else
234			binuptime(&ds->busy_from);
235	}
236	atomic_add_rel_int(&ds->sequence0, 1);
237}
238
239void
240devstat_start_transaction_bio(struct devstat *ds, struct bio *bp)
241{
242
243	/* sanity check */
244	if (ds == NULL)
245		return;
246
247	binuptime(&bp->bio_t0);
248	devstat_start_transaction_bio_t0(ds, bp);
249}
250
251void
252devstat_start_transaction_bio_t0(struct devstat *ds, struct bio *bp)
253{
254
255	/* sanity check */
256	if (ds == NULL)
257		return;
258
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 */
289void
290devstat_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}
333
334void
335devstat_end_transaction_bio(struct devstat *ds, const struct bio *bp)
336{
337
338	devstat_end_transaction_bio_bt(ds, bp, NULL);
339}
340
341void
342devstat_end_transaction_bio_bt(struct devstat *ds, const struct bio *bp,
343    const struct bintime *now)
344{
345	devstat_trans_flags flg;
346	devstat_tag_type tag;
347
348	/* sanity check */
349	if (ds == NULL)
350		return;
351
352	if (bp->bio_flags & BIO_ORDERED)
353		tag = DEVSTAT_TAG_ORDERED;
354	else
355		tag = DEVSTAT_TAG_SIMPLE;
356	if (bp->bio_cmd == BIO_DELETE)
357		flg = DEVSTAT_FREE;
358	else if ((bp->bio_cmd == BIO_READ)
359	      || ((bp->bio_cmd == BIO_ZONE)
360	       && (bp->bio_zone.zone_cmd == DISK_ZONE_REPORT_ZONES)))
361		flg = DEVSTAT_READ;
362	else if (bp->bio_cmd == BIO_WRITE)
363		flg = DEVSTAT_WRITE;
364	else
365		flg = DEVSTAT_NO_DATA;
366
367	devstat_end_transaction(ds, bp->bio_bcount - bp->bio_resid,
368				tag, flg, now, &bp->bio_t0);
369	DTRACE_DEVSTAT_BIO_DONE();
370}
371
372/*
373 * This is the sysctl handler for the devstat package.  The data pushed out
374 * on the kern.devstat.all sysctl variable consists of the current devstat
375 * generation number, and then an array of devstat structures, one for each
376 * device in the system.
377 *
378 * This is more cryptic that obvious, but basically we neither can nor
379 * want to hold the devstat_mutex for any amount of time, so we grab it
380 * only when we need to and keep an eye on devstat_generation all the time.
381 */
382static int
383sysctl_devstat(SYSCTL_HANDLER_ARGS)
384{
385	int error;
386	long mygen;
387	struct devstat *nds;
388
389	mtx_assert(&devstat_mutex, MA_NOTOWNED);
390
391	/*
392	 * XXX devstat_generation should really be "volatile" but that
393	 * XXX freaks out the sysctl macro below.  The places where we
394	 * XXX change it and inspect it are bracketed in the mutex which
395	 * XXX guarantees us proper write barriers.  I don't believe the
396	 * XXX compiler is allowed to optimize mygen away across calls
397	 * XXX to other functions, so the following is belived to be safe.
398	 */
399	mygen = devstat_generation;
400
401	error = SYSCTL_OUT(req, &mygen, sizeof(mygen));
402
403	if (devstat_num_devs == 0)
404		return(0);
405
406	if (error != 0)
407		return (error);
408
409	mtx_lock(&devstat_mutex);
410	nds = STAILQ_FIRST(&device_statq);
411	if (mygen != devstat_generation)
412		error = EBUSY;
413	mtx_unlock(&devstat_mutex);
414
415	if (error != 0)
416		return (error);
417
418	for (;nds != NULL;) {
419		error = SYSCTL_OUT(req, nds, sizeof(struct devstat));
420		if (error != 0)
421			return (error);
422		mtx_lock(&devstat_mutex);
423		if (mygen != devstat_generation)
424			error = EBUSY;
425		else
426			nds = STAILQ_NEXT(nds, dev_links);
427		mtx_unlock(&devstat_mutex);
428		if (error != 0)
429			return (error);
430	}
431	return(error);
432}
433
434/*
435 * Sysctl entries for devstat.  The first one is a node that all the rest
436 * hang off of.
437 */
438static SYSCTL_NODE(_kern, OID_AUTO, devstat, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
439    "Device Statistics");
440
441SYSCTL_PROC(_kern_devstat, OID_AUTO, all,
442    CTLFLAG_RD | CTLTYPE_OPAQUE | CTLFLAG_MPSAFE, NULL, 0,
443    sysctl_devstat, "S,devstat",
444    "All devices in the devstat list");
445/*
446 * Export the number of devices in the system so that userland utilities
447 * can determine how much memory to allocate to hold all the devices.
448 */
449SYSCTL_INT(_kern_devstat, OID_AUTO, numdevs, CTLFLAG_RD,
450    &devstat_num_devs, 0, "Number of devices in the devstat list");
451SYSCTL_LONG(_kern_devstat, OID_AUTO, generation, CTLFLAG_RD,
452    &devstat_generation, 0, "Devstat list generation");
453SYSCTL_INT(_kern_devstat, OID_AUTO, version, CTLFLAG_RD,
454    &devstat_version, 0, "Devstat list version number");
455
456/*
457 * Allocator for struct devstat structures.  We sub-allocate these from pages
458 * which we get from malloc.  These pages are exported for mmap(2)'ing through
459 * a miniature device driver
460 */
461
462#define statsperpage (PAGE_SIZE / sizeof(struct devstat))
463
464static d_ioctl_t devstat_ioctl;
465static d_mmap_t devstat_mmap;
466
467static struct cdevsw devstat_cdevsw = {
468	.d_version =	D_VERSION,
469	.d_ioctl =	devstat_ioctl,
470	.d_mmap =	devstat_mmap,
471	.d_name =	"devstat",
472};
473
474struct statspage {
475	TAILQ_ENTRY(statspage)	list;
476	struct devstat		*stat;
477	u_int			nfree;
478};
479
480static size_t pagelist_pages = 0;
481static TAILQ_HEAD(, statspage)	pagelist = TAILQ_HEAD_INITIALIZER(pagelist);
482static MALLOC_DEFINE(M_DEVSTAT, "devstat", "Device statistics");
483
484static int
485devstat_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int fflag,
486    struct thread *td)
487{
488	int error = ENOTTY;
489
490	switch (cmd) {
491	case DIOCGMEDIASIZE:
492		error = 0;
493		*(off_t *)data = pagelist_pages * PAGE_SIZE;
494		break;
495	}
496
497	return (error);
498}
499
500static int
501devstat_mmap(struct cdev *dev, vm_ooffset_t offset, vm_paddr_t *paddr,
502    int nprot, vm_memattr_t *memattr)
503{
504	struct statspage *spp;
505
506	if (nprot != VM_PROT_READ)
507		return (-1);
508	mtx_lock(&devstat_mutex);
509	TAILQ_FOREACH(spp, &pagelist, list) {
510		if (offset == 0) {
511			*paddr = vtophys(spp->stat);
512			mtx_unlock(&devstat_mutex);
513			return (0);
514		}
515		offset -= PAGE_SIZE;
516	}
517	mtx_unlock(&devstat_mutex);
518	return (-1);
519}
520
521static struct devstat *
522devstat_alloc(void)
523{
524	struct devstat *dsp;
525	struct statspage *spp, *spp2;
526	u_int u;
527	static int once;
528
529	mtx_assert(&devstat_mutex, MA_NOTOWNED);
530	if (!once) {
531		make_dev_credf(MAKEDEV_ETERNAL | MAKEDEV_CHECKNAME,
532		    &devstat_cdevsw, 0, NULL, UID_ROOT, GID_WHEEL, 0444,
533		    DEVSTAT_DEVICE_NAME);
534		once = 1;
535	}
536	spp2 = NULL;
537	mtx_lock(&devstat_mutex);
538	for (;;) {
539		TAILQ_FOREACH(spp, &pagelist, list) {
540			if (spp->nfree > 0)
541				break;
542		}
543		if (spp != NULL)
544			break;
545		mtx_unlock(&devstat_mutex);
546		spp2 = malloc(sizeof *spp, M_DEVSTAT, M_ZERO | M_WAITOK);
547		spp2->stat = malloc(PAGE_SIZE, M_DEVSTAT, M_ZERO | M_WAITOK);
548		spp2->nfree = statsperpage;
549
550		/*
551		 * If free statspages were added while the lock was released
552		 * just reuse them.
553		 */
554		mtx_lock(&devstat_mutex);
555		TAILQ_FOREACH(spp, &pagelist, list)
556			if (spp->nfree > 0)
557				break;
558		if (spp == NULL) {
559			spp = spp2;
560
561			/*
562			 * It would make more sense to add the new page at the
563			 * head but the order on the list determine the
564			 * sequence of the mapping so we can't do that.
565			 */
566			pagelist_pages++;
567			TAILQ_INSERT_TAIL(&pagelist, spp, list);
568		} else
569			break;
570	}
571	dsp = spp->stat;
572	for (u = 0; u < statsperpage; u++) {
573		if (dsp->allocated == 0)
574			break;
575		dsp++;
576	}
577	spp->nfree--;
578	dsp->allocated = 1;
579	mtx_unlock(&devstat_mutex);
580	if (spp2 != NULL && spp2 != spp) {
581		free(spp2->stat, M_DEVSTAT);
582		free(spp2, M_DEVSTAT);
583	}
584	return (dsp);
585}
586
587static void
588devstat_free(struct devstat *dsp)
589{
590	struct statspage *spp;
591
592	mtx_assert(&devstat_mutex, MA_OWNED);
593	bzero(dsp, sizeof *dsp);
594	TAILQ_FOREACH(spp, &pagelist, list) {
595		if (dsp >= spp->stat && dsp < (spp->stat + statsperpage)) {
596			spp->nfree++;
597			return;
598		}
599	}
600}
601
602SYSCTL_INT(_debug_sizeof, OID_AUTO, devstat, CTLFLAG_RD,
603    SYSCTL_NULL_INT_PTR, sizeof(struct devstat), "sizeof(struct devstat)");
604