xref: /freebsd/sys/geom/geom_io.c (revision 09a53ad8f1318c5daae6cfb19d97f4f6459f0013)
1 /*-
2  * Copyright (c) 2002 Poul-Henning Kamp
3  * Copyright (c) 2002 Networks Associates Technology, Inc.
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * This software was developed for the FreeBSD Project by Poul-Henning Kamp
8  * and NAI Labs, the Security Research Division of Network Associates, Inc.
9  * under DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
10  * DARPA CHATS research program.
11  *
12  * Portions of this software were developed by Konstantin Belousov
13  * under sponsorship from the FreeBSD Foundation.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. The names of the authors may not be used to endorse or promote
24  *    products derived from this software without specific prior written
25  *    permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
28  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
31  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37  * SUCH DAMAGE.
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/kernel.h>
46 #include <sys/malloc.h>
47 #include <sys/bio.h>
48 #include <sys/ktr.h>
49 #include <sys/proc.h>
50 #include <sys/stack.h>
51 #include <sys/sysctl.h>
52 #include <sys/vmem.h>
53 
54 #include <sys/errno.h>
55 #include <geom/geom.h>
56 #include <geom/geom_int.h>
57 #include <sys/devicestat.h>
58 
59 #include <vm/uma.h>
60 #include <vm/vm.h>
61 #include <vm/vm_param.h>
62 #include <vm/vm_kern.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_object.h>
65 #include <vm/vm_extern.h>
66 #include <vm/vm_map.h>
67 
68 static int	g_io_transient_map_bio(struct bio *bp);
69 
70 static struct g_bioq g_bio_run_down;
71 static struct g_bioq g_bio_run_up;
72 static struct g_bioq g_bio_run_task;
73 
74 /*
75  * Pace is a hint that we've had some trouble recently allocating
76  * bios, so we should back off trying to send I/O down the stack
77  * a bit to let the problem resolve. When pacing, we also turn
78  * off direct dispatch to also reduce memory pressure from I/Os
79  * there, at the expxense of some added latency while the memory
80  * pressures exist. See g_io_schedule_down() for more details
81  * and limitations.
82  */
83 static volatile u_int pace;
84 
85 static uma_zone_t	biozone;
86 
87 /*
88  * The head of the list of classifiers used in g_io_request.
89  * Use g_register_classifier() and g_unregister_classifier()
90  * to add/remove entries to the list.
91  * Classifiers are invoked in registration order.
92  */
93 static TAILQ_HEAD(g_classifier_tailq, g_classifier_hook)
94     g_classifier_tailq = TAILQ_HEAD_INITIALIZER(g_classifier_tailq);
95 
96 #include <machine/atomic.h>
97 
98 static void
99 g_bioq_lock(struct g_bioq *bq)
100 {
101 
102 	mtx_lock(&bq->bio_queue_lock);
103 }
104 
105 static void
106 g_bioq_unlock(struct g_bioq *bq)
107 {
108 
109 	mtx_unlock(&bq->bio_queue_lock);
110 }
111 
112 #if 0
113 static void
114 g_bioq_destroy(struct g_bioq *bq)
115 {
116 
117 	mtx_destroy(&bq->bio_queue_lock);
118 }
119 #endif
120 
121 static void
122 g_bioq_init(struct g_bioq *bq)
123 {
124 
125 	TAILQ_INIT(&bq->bio_queue);
126 	mtx_init(&bq->bio_queue_lock, "bio queue", NULL, MTX_DEF);
127 }
128 
129 static struct bio *
130 g_bioq_first(struct g_bioq *bq)
131 {
132 	struct bio *bp;
133 
134 	bp = TAILQ_FIRST(&bq->bio_queue);
135 	if (bp != NULL) {
136 		KASSERT((bp->bio_flags & BIO_ONQUEUE),
137 		    ("Bio not on queue bp=%p target %p", bp, bq));
138 		bp->bio_flags &= ~BIO_ONQUEUE;
139 		TAILQ_REMOVE(&bq->bio_queue, bp, bio_queue);
140 		bq->bio_queue_length--;
141 	}
142 	return (bp);
143 }
144 
145 struct bio *
146 g_new_bio(void)
147 {
148 	struct bio *bp;
149 
150 	bp = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
151 #ifdef KTR
152 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
153 		struct stack st;
154 
155 		CTR1(KTR_GEOM, "g_new_bio(): %p", bp);
156 		stack_save(&st);
157 		CTRSTACK(KTR_GEOM, &st, 3, 0);
158 	}
159 #endif
160 	return (bp);
161 }
162 
163 struct bio *
164 g_alloc_bio(void)
165 {
166 	struct bio *bp;
167 
168 	bp = uma_zalloc(biozone, M_WAITOK | M_ZERO);
169 #ifdef KTR
170 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
171 		struct stack st;
172 
173 		CTR1(KTR_GEOM, "g_alloc_bio(): %p", bp);
174 		stack_save(&st);
175 		CTRSTACK(KTR_GEOM, &st, 3, 0);
176 	}
177 #endif
178 	return (bp);
179 }
180 
181 void
182 g_destroy_bio(struct bio *bp)
183 {
184 #ifdef KTR
185 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
186 		struct stack st;
187 
188 		CTR1(KTR_GEOM, "g_destroy_bio(): %p", bp);
189 		stack_save(&st);
190 		CTRSTACK(KTR_GEOM, &st, 3, 0);
191 	}
192 #endif
193 	uma_zfree(biozone, bp);
194 }
195 
196 struct bio *
197 g_clone_bio(struct bio *bp)
198 {
199 	struct bio *bp2;
200 
201 	bp2 = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
202 	if (bp2 != NULL) {
203 		bp2->bio_parent = bp;
204 		bp2->bio_cmd = bp->bio_cmd;
205 		/*
206 		 *  BIO_ORDERED flag may be used by disk drivers to enforce
207 		 *  ordering restrictions, so this flag needs to be cloned.
208 		 *  BIO_UNMAPPED and BIO_VLIST should be inherited, to properly
209 		 *  indicate which way the buffer is passed.
210 		 *  Other bio flags are not suitable for cloning.
211 		 */
212 		bp2->bio_flags = bp->bio_flags &
213 		    (BIO_ORDERED | BIO_UNMAPPED | BIO_VLIST);
214 		bp2->bio_length = bp->bio_length;
215 		bp2->bio_offset = bp->bio_offset;
216 		bp2->bio_data = bp->bio_data;
217 		bp2->bio_ma = bp->bio_ma;
218 		bp2->bio_ma_n = bp->bio_ma_n;
219 		bp2->bio_ma_offset = bp->bio_ma_offset;
220 		bp2->bio_attribute = bp->bio_attribute;
221 		if (bp->bio_cmd == BIO_ZONE)
222 			bcopy(&bp->bio_zone, &bp2->bio_zone,
223 			    sizeof(bp->bio_zone));
224 		/* Inherit classification info from the parent */
225 		bp2->bio_classifier1 = bp->bio_classifier1;
226 		bp2->bio_classifier2 = bp->bio_classifier2;
227 		bp->bio_children++;
228 	}
229 #ifdef KTR
230 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
231 		struct stack st;
232 
233 		CTR2(KTR_GEOM, "g_clone_bio(%p): %p", bp, bp2);
234 		stack_save(&st);
235 		CTRSTACK(KTR_GEOM, &st, 3, 0);
236 	}
237 #endif
238 	return(bp2);
239 }
240 
241 struct bio *
242 g_duplicate_bio(struct bio *bp)
243 {
244 	struct bio *bp2;
245 
246 	bp2 = uma_zalloc(biozone, M_WAITOK | M_ZERO);
247 	bp2->bio_flags = bp->bio_flags & (BIO_UNMAPPED | BIO_VLIST);
248 	bp2->bio_parent = bp;
249 	bp2->bio_cmd = bp->bio_cmd;
250 	bp2->bio_length = bp->bio_length;
251 	bp2->bio_offset = bp->bio_offset;
252 	bp2->bio_data = bp->bio_data;
253 	bp2->bio_ma = bp->bio_ma;
254 	bp2->bio_ma_n = bp->bio_ma_n;
255 	bp2->bio_ma_offset = bp->bio_ma_offset;
256 	bp2->bio_attribute = bp->bio_attribute;
257 	bp->bio_children++;
258 #ifdef KTR
259 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
260 		struct stack st;
261 
262 		CTR2(KTR_GEOM, "g_duplicate_bio(%p): %p", bp, bp2);
263 		stack_save(&st);
264 		CTRSTACK(KTR_GEOM, &st, 3, 0);
265 	}
266 #endif
267 	return(bp2);
268 }
269 
270 void
271 g_reset_bio(struct bio *bp)
272 {
273 
274 	bzero(bp, sizeof(*bp));
275 }
276 
277 void
278 g_io_init()
279 {
280 
281 	g_bioq_init(&g_bio_run_down);
282 	g_bioq_init(&g_bio_run_up);
283 	g_bioq_init(&g_bio_run_task);
284 	biozone = uma_zcreate("g_bio", sizeof (struct bio),
285 	    NULL, NULL,
286 	    NULL, NULL,
287 	    0, 0);
288 }
289 
290 int
291 g_io_getattr(const char *attr, struct g_consumer *cp, int *len, void *ptr)
292 {
293 	struct bio *bp;
294 	int error;
295 
296 	g_trace(G_T_BIO, "bio_getattr(%s)", attr);
297 	bp = g_alloc_bio();
298 	bp->bio_cmd = BIO_GETATTR;
299 	bp->bio_done = NULL;
300 	bp->bio_attribute = attr;
301 	bp->bio_length = *len;
302 	bp->bio_data = ptr;
303 	g_io_request(bp, cp);
304 	error = biowait(bp, "ggetattr");
305 	*len = bp->bio_completed;
306 	g_destroy_bio(bp);
307 	return (error);
308 }
309 
310 int
311 g_io_zonecmd(struct disk_zone_args *zone_args, struct g_consumer *cp)
312 {
313 	struct bio *bp;
314 	int error;
315 
316 	g_trace(G_T_BIO, "bio_zone(%d)", zone_args->zone_cmd);
317 	bp = g_alloc_bio();
318 	bp->bio_cmd = BIO_ZONE;
319 	bp->bio_done = NULL;
320 	/*
321 	 * XXX KDM need to handle report zone data.
322 	 */
323 	bcopy(zone_args, &bp->bio_zone, sizeof(*zone_args));
324 	if (zone_args->zone_cmd == DISK_ZONE_REPORT_ZONES)
325 		bp->bio_length =
326 		    zone_args->zone_params.report.entries_allocated *
327 		    sizeof(struct disk_zone_rep_entry);
328 	else
329 		bp->bio_length = 0;
330 
331 	g_io_request(bp, cp);
332 	error = biowait(bp, "gzone");
333 	bcopy(&bp->bio_zone, zone_args, sizeof(*zone_args));
334 	g_destroy_bio(bp);
335 	return (error);
336 }
337 
338 int
339 g_io_flush(struct g_consumer *cp)
340 {
341 	struct bio *bp;
342 	int error;
343 
344 	g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name);
345 	bp = g_alloc_bio();
346 	bp->bio_cmd = BIO_FLUSH;
347 	bp->bio_flags |= BIO_ORDERED;
348 	bp->bio_done = NULL;
349 	bp->bio_attribute = NULL;
350 	bp->bio_offset = cp->provider->mediasize;
351 	bp->bio_length = 0;
352 	bp->bio_data = NULL;
353 	g_io_request(bp, cp);
354 	error = biowait(bp, "gflush");
355 	g_destroy_bio(bp);
356 	return (error);
357 }
358 
359 static int
360 g_io_check(struct bio *bp)
361 {
362 	struct g_consumer *cp;
363 	struct g_provider *pp;
364 	off_t excess;
365 	int error;
366 
367 	cp = bp->bio_from;
368 	pp = bp->bio_to;
369 
370 	/* Fail if access counters dont allow the operation */
371 	switch(bp->bio_cmd) {
372 	case BIO_READ:
373 	case BIO_GETATTR:
374 		if (cp->acr == 0)
375 			return (EPERM);
376 		break;
377 	case BIO_WRITE:
378 	case BIO_DELETE:
379 	case BIO_FLUSH:
380 		if (cp->acw == 0)
381 			return (EPERM);
382 		break;
383 	case BIO_ZONE:
384 		if ((bp->bio_zone.zone_cmd == DISK_ZONE_REPORT_ZONES) ||
385 		    (bp->bio_zone.zone_cmd == DISK_ZONE_GET_PARAMS)) {
386 			if (cp->acr == 0)
387 				return (EPERM);
388 		} else if (cp->acw == 0)
389 			return (EPERM);
390 		break;
391 	default:
392 		return (EPERM);
393 	}
394 	/* if provider is marked for error, don't disturb. */
395 	if (pp->error)
396 		return (pp->error);
397 	if (cp->flags & G_CF_ORPHAN)
398 		return (ENXIO);
399 
400 	switch(bp->bio_cmd) {
401 	case BIO_READ:
402 	case BIO_WRITE:
403 	case BIO_DELETE:
404 		/* Zero sectorsize or mediasize is probably a lack of media. */
405 		if (pp->sectorsize == 0 || pp->mediasize == 0)
406 			return (ENXIO);
407 		/* Reject I/O not on sector boundary */
408 		if (bp->bio_offset % pp->sectorsize)
409 			return (EINVAL);
410 		/* Reject I/O not integral sector long */
411 		if (bp->bio_length % pp->sectorsize)
412 			return (EINVAL);
413 		/* Reject requests before or past the end of media. */
414 		if (bp->bio_offset < 0)
415 			return (EIO);
416 		if (bp->bio_offset > pp->mediasize)
417 			return (EIO);
418 
419 		/* Truncate requests to the end of providers media. */
420 		excess = bp->bio_offset + bp->bio_length;
421 		if (excess > bp->bio_to->mediasize) {
422 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
423 			    round_page(bp->bio_ma_offset +
424 			    bp->bio_length) / PAGE_SIZE == bp->bio_ma_n,
425 			    ("excess bio %p too short", bp));
426 			excess -= bp->bio_to->mediasize;
427 			bp->bio_length -= excess;
428 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
429 				bp->bio_ma_n = round_page(bp->bio_ma_offset +
430 				    bp->bio_length) / PAGE_SIZE;
431 			}
432 			if (excess > 0)
433 				CTR3(KTR_GEOM, "g_down truncated bio "
434 				    "%p provider %s by %d", bp,
435 				    bp->bio_to->name, excess);
436 		}
437 
438 		/* Deliver zero length transfers right here. */
439 		if (bp->bio_length == 0) {
440 			CTR2(KTR_GEOM, "g_down terminated 0-length "
441 			    "bp %p provider %s", bp, bp->bio_to->name);
442 			return (0);
443 		}
444 
445 		if ((bp->bio_flags & BIO_UNMAPPED) != 0 &&
446 		    (bp->bio_to->flags & G_PF_ACCEPT_UNMAPPED) == 0 &&
447 		    (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) {
448 			if ((error = g_io_transient_map_bio(bp)) >= 0)
449 				return (error);
450 		}
451 		break;
452 	default:
453 		break;
454 	}
455 	return (EJUSTRETURN);
456 }
457 
458 /*
459  * bio classification support.
460  *
461  * g_register_classifier() and g_unregister_classifier()
462  * are used to add/remove a classifier from the list.
463  * The list is protected using the g_bio_run_down lock,
464  * because the classifiers are called in this path.
465  *
466  * g_io_request() passes bio's that are not already classified
467  * (i.e. those with bio_classifier1 == NULL) to g_run_classifiers().
468  * Classifiers can store their result in the two fields
469  * bio_classifier1 and bio_classifier2.
470  * A classifier that updates one of the fields should
471  * return a non-zero value.
472  * If no classifier updates the field, g_run_classifiers() sets
473  * bio_classifier1 = BIO_NOTCLASSIFIED to avoid further calls.
474  */
475 
476 int
477 g_register_classifier(struct g_classifier_hook *hook)
478 {
479 
480 	g_bioq_lock(&g_bio_run_down);
481 	TAILQ_INSERT_TAIL(&g_classifier_tailq, hook, link);
482 	g_bioq_unlock(&g_bio_run_down);
483 
484 	return (0);
485 }
486 
487 void
488 g_unregister_classifier(struct g_classifier_hook *hook)
489 {
490 	struct g_classifier_hook *entry;
491 
492 	g_bioq_lock(&g_bio_run_down);
493 	TAILQ_FOREACH(entry, &g_classifier_tailq, link) {
494 		if (entry == hook) {
495 			TAILQ_REMOVE(&g_classifier_tailq, hook, link);
496 			break;
497 		}
498 	}
499 	g_bioq_unlock(&g_bio_run_down);
500 }
501 
502 static void
503 g_run_classifiers(struct bio *bp)
504 {
505 	struct g_classifier_hook *hook;
506 	int classified = 0;
507 
508 	TAILQ_FOREACH(hook, &g_classifier_tailq, link)
509 		classified |= hook->func(hook->arg, bp);
510 
511 	if (!classified)
512 		bp->bio_classifier1 = BIO_NOTCLASSIFIED;
513 }
514 
515 void
516 g_io_request(struct bio *bp, struct g_consumer *cp)
517 {
518 	struct g_provider *pp;
519 	struct mtx *mtxp;
520 	int direct, error, first;
521 	uint8_t cmd;
522 
523 	KASSERT(cp != NULL, ("NULL cp in g_io_request"));
524 	KASSERT(bp != NULL, ("NULL bp in g_io_request"));
525 	pp = cp->provider;
526 	KASSERT(pp != NULL, ("consumer not attached in g_io_request"));
527 #ifdef DIAGNOSTIC
528 	KASSERT(bp->bio_driver1 == NULL,
529 	    ("bio_driver1 used by the consumer (geom %s)", cp->geom->name));
530 	KASSERT(bp->bio_driver2 == NULL,
531 	    ("bio_driver2 used by the consumer (geom %s)", cp->geom->name));
532 	KASSERT(bp->bio_pflags == 0,
533 	    ("bio_pflags used by the consumer (geom %s)", cp->geom->name));
534 	/*
535 	 * Remember consumer's private fields, so we can detect if they were
536 	 * modified by the provider.
537 	 */
538 	bp->_bio_caller1 = bp->bio_caller1;
539 	bp->_bio_caller2 = bp->bio_caller2;
540 	bp->_bio_cflags = bp->bio_cflags;
541 #endif
542 
543 	cmd = bp->bio_cmd;
544 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_GETATTR) {
545 		KASSERT(bp->bio_data != NULL,
546 		    ("NULL bp->data in g_io_request(cmd=%hu)", bp->bio_cmd));
547 	}
548 	if (cmd == BIO_DELETE || cmd == BIO_FLUSH) {
549 		KASSERT(bp->bio_data == NULL,
550 		    ("non-NULL bp->data in g_io_request(cmd=%hu)",
551 		    bp->bio_cmd));
552 	}
553 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_DELETE) {
554 		KASSERT(bp->bio_offset % cp->provider->sectorsize == 0,
555 		    ("wrong offset %jd for sectorsize %u",
556 		    bp->bio_offset, cp->provider->sectorsize));
557 		KASSERT(bp->bio_length % cp->provider->sectorsize == 0,
558 		    ("wrong length %jd for sectorsize %u",
559 		    bp->bio_length, cp->provider->sectorsize));
560 	}
561 
562 	g_trace(G_T_BIO, "bio_request(%p) from %p(%s) to %p(%s) cmd %d",
563 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd);
564 
565 	bp->bio_from = cp;
566 	bp->bio_to = pp;
567 	bp->bio_error = 0;
568 	bp->bio_completed = 0;
569 
570 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
571 	    ("Bio already on queue bp=%p", bp));
572 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
573 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
574 		binuptime(&bp->bio_t0);
575 	else
576 		getbinuptime(&bp->bio_t0);
577 
578 #ifdef GET_STACK_USAGE
579 	direct = (cp->flags & G_CF_DIRECT_SEND) != 0 &&
580 	    (pp->flags & G_PF_DIRECT_RECEIVE) != 0 &&
581 	    !g_is_geom_thread(curthread) &&
582 	    ((pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ||
583 	    (bp->bio_flags & BIO_UNMAPPED) == 0 || THREAD_CAN_SLEEP()) &&
584 	    pace == 0;
585 	if (direct) {
586 		/* Block direct execution if less then half of stack left. */
587 		size_t	st, su;
588 		GET_STACK_USAGE(st, su);
589 		if (su * 2 > st)
590 			direct = 0;
591 	}
592 #else
593 	direct = 0;
594 #endif
595 
596 	if (!TAILQ_EMPTY(&g_classifier_tailq) && !bp->bio_classifier1) {
597 		g_bioq_lock(&g_bio_run_down);
598 		g_run_classifiers(bp);
599 		g_bioq_unlock(&g_bio_run_down);
600 	}
601 
602 	/*
603 	 * The statistics collection is lockless, as such, but we
604 	 * can not update one instance of the statistics from more
605 	 * than one thread at a time, so grab the lock first.
606 	 */
607 	mtxp = mtx_pool_find(mtxpool_sleep, pp);
608 	mtx_lock(mtxp);
609 	if (g_collectstats & G_STATS_PROVIDERS)
610 		devstat_start_transaction(pp->stat, &bp->bio_t0);
611 	if (g_collectstats & G_STATS_CONSUMERS)
612 		devstat_start_transaction(cp->stat, &bp->bio_t0);
613 	pp->nstart++;
614 	cp->nstart++;
615 	mtx_unlock(mtxp);
616 
617 	if (direct) {
618 		error = g_io_check(bp);
619 		if (error >= 0) {
620 			CTR3(KTR_GEOM, "g_io_request g_io_check on bp %p "
621 			    "provider %s returned %d", bp, bp->bio_to->name,
622 			    error);
623 			g_io_deliver(bp, error);
624 			return;
625 		}
626 		bp->bio_to->geom->start(bp);
627 	} else {
628 		g_bioq_lock(&g_bio_run_down);
629 		first = TAILQ_EMPTY(&g_bio_run_down.bio_queue);
630 		TAILQ_INSERT_TAIL(&g_bio_run_down.bio_queue, bp, bio_queue);
631 		bp->bio_flags |= BIO_ONQUEUE;
632 		g_bio_run_down.bio_queue_length++;
633 		g_bioq_unlock(&g_bio_run_down);
634 		/* Pass it on down. */
635 		if (first)
636 			wakeup(&g_wait_down);
637 	}
638 }
639 
640 void
641 g_io_deliver(struct bio *bp, int error)
642 {
643 	struct bintime now;
644 	struct g_consumer *cp;
645 	struct g_provider *pp;
646 	struct mtx *mtxp;
647 	int direct, first;
648 
649 	KASSERT(bp != NULL, ("NULL bp in g_io_deliver"));
650 	pp = bp->bio_to;
651 	KASSERT(pp != NULL, ("NULL bio_to in g_io_deliver"));
652 	cp = bp->bio_from;
653 	if (cp == NULL) {
654 		bp->bio_error = error;
655 		bp->bio_done(bp);
656 		return;
657 	}
658 	KASSERT(cp != NULL, ("NULL bio_from in g_io_deliver"));
659 	KASSERT(cp->geom != NULL, ("NULL bio_from->geom in g_io_deliver"));
660 #ifdef DIAGNOSTIC
661 	/*
662 	 * Some classes - GJournal in particular - can modify bio's
663 	 * private fields while the bio is in transit; G_GEOM_VOLATILE_BIO
664 	 * flag means it's an expected behaviour for that particular geom.
665 	 */
666 	if ((cp->geom->flags & G_GEOM_VOLATILE_BIO) == 0) {
667 		KASSERT(bp->bio_caller1 == bp->_bio_caller1,
668 		    ("bio_caller1 used by the provider %s", pp->name));
669 		KASSERT(bp->bio_caller2 == bp->_bio_caller2,
670 		    ("bio_caller2 used by the provider %s", pp->name));
671 		KASSERT(bp->bio_cflags == bp->_bio_cflags,
672 		    ("bio_cflags used by the provider %s", pp->name));
673 	}
674 #endif
675 	KASSERT(bp->bio_completed >= 0, ("bio_completed can't be less than 0"));
676 	KASSERT(bp->bio_completed <= bp->bio_length,
677 	    ("bio_completed can't be greater than bio_length"));
678 
679 	g_trace(G_T_BIO,
680 "g_io_deliver(%p) from %p(%s) to %p(%s) cmd %d error %d off %jd len %jd",
681 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd, error,
682 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
683 
684 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
685 	    ("Bio already on queue bp=%p", bp));
686 
687 	/*
688 	 * XXX: next two doesn't belong here
689 	 */
690 	bp->bio_bcount = bp->bio_length;
691 	bp->bio_resid = bp->bio_bcount - bp->bio_completed;
692 
693 #ifdef GET_STACK_USAGE
694 	direct = (pp->flags & G_PF_DIRECT_SEND) &&
695 		 (cp->flags & G_CF_DIRECT_RECEIVE) &&
696 		 !g_is_geom_thread(curthread);
697 	if (direct) {
698 		/* Block direct execution if less then half of stack left. */
699 		size_t	st, su;
700 		GET_STACK_USAGE(st, su);
701 		if (su * 2 > st)
702 			direct = 0;
703 	}
704 #else
705 	direct = 0;
706 #endif
707 
708 	/*
709 	 * The statistics collection is lockless, as such, but we
710 	 * can not update one instance of the statistics from more
711 	 * than one thread at a time, so grab the lock first.
712 	 */
713 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
714 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
715 		binuptime(&now);
716 	mtxp = mtx_pool_find(mtxpool_sleep, cp);
717 	mtx_lock(mtxp);
718 	if (g_collectstats & G_STATS_PROVIDERS)
719 		devstat_end_transaction_bio_bt(pp->stat, bp, &now);
720 	if (g_collectstats & G_STATS_CONSUMERS)
721 		devstat_end_transaction_bio_bt(cp->stat, bp, &now);
722 	cp->nend++;
723 	pp->nend++;
724 	mtx_unlock(mtxp);
725 
726 	if (error != ENOMEM) {
727 		bp->bio_error = error;
728 		if (direct) {
729 			biodone(bp);
730 		} else {
731 			g_bioq_lock(&g_bio_run_up);
732 			first = TAILQ_EMPTY(&g_bio_run_up.bio_queue);
733 			TAILQ_INSERT_TAIL(&g_bio_run_up.bio_queue, bp, bio_queue);
734 			bp->bio_flags |= BIO_ONQUEUE;
735 			g_bio_run_up.bio_queue_length++;
736 			g_bioq_unlock(&g_bio_run_up);
737 			if (first)
738 				wakeup(&g_wait_up);
739 		}
740 		return;
741 	}
742 
743 	if (bootverbose)
744 		printf("ENOMEM %p on %p(%s)\n", bp, pp, pp->name);
745 	bp->bio_children = 0;
746 	bp->bio_inbed = 0;
747 	bp->bio_driver1 = NULL;
748 	bp->bio_driver2 = NULL;
749 	bp->bio_pflags = 0;
750 	g_io_request(bp, cp);
751 	pace = 1;
752 	return;
753 }
754 
755 SYSCTL_DECL(_kern_geom);
756 
757 static long transient_maps;
758 SYSCTL_LONG(_kern_geom, OID_AUTO, transient_maps, CTLFLAG_RD,
759     &transient_maps, 0,
760     "Total count of the transient mapping requests");
761 u_int transient_map_retries = 10;
762 SYSCTL_UINT(_kern_geom, OID_AUTO, transient_map_retries, CTLFLAG_RW,
763     &transient_map_retries, 0,
764     "Max count of retries used before giving up on creating transient map");
765 int transient_map_hard_failures;
766 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_hard_failures, CTLFLAG_RD,
767     &transient_map_hard_failures, 0,
768     "Failures to establish the transient mapping due to retry attempts "
769     "exhausted");
770 int transient_map_soft_failures;
771 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_soft_failures, CTLFLAG_RD,
772     &transient_map_soft_failures, 0,
773     "Count of retried failures to establish the transient mapping");
774 int inflight_transient_maps;
775 SYSCTL_INT(_kern_geom, OID_AUTO, inflight_transient_maps, CTLFLAG_RD,
776     &inflight_transient_maps, 0,
777     "Current count of the active transient maps");
778 
779 static int
780 g_io_transient_map_bio(struct bio *bp)
781 {
782 	vm_offset_t addr;
783 	long size;
784 	u_int retried;
785 
786 	KASSERT(unmapped_buf_allowed, ("unmapped disabled"));
787 
788 	size = round_page(bp->bio_ma_offset + bp->bio_length);
789 	KASSERT(size / PAGE_SIZE == bp->bio_ma_n, ("Bio too short %p", bp));
790 	addr = 0;
791 	retried = 0;
792 	atomic_add_long(&transient_maps, 1);
793 retry:
794 	if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) {
795 		if (transient_map_retries != 0 &&
796 		    retried >= transient_map_retries) {
797 			CTR2(KTR_GEOM, "g_down cannot map bp %p provider %s",
798 			    bp, bp->bio_to->name);
799 			atomic_add_int(&transient_map_hard_failures, 1);
800 			return (EDEADLK/* XXXKIB */);
801 		} else {
802 			/*
803 			 * Naive attempt to quisce the I/O to get more
804 			 * in-flight requests completed and defragment
805 			 * the transient_arena.
806 			 */
807 			CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d",
808 			    bp, bp->bio_to->name, retried);
809 			pause("g_d_tra", hz / 10);
810 			retried++;
811 			atomic_add_int(&transient_map_soft_failures, 1);
812 			goto retry;
813 		}
814 	}
815 	atomic_add_int(&inflight_transient_maps, 1);
816 	pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size));
817 	bp->bio_data = (caddr_t)addr + bp->bio_ma_offset;
818 	bp->bio_flags |= BIO_TRANSIENT_MAPPING;
819 	bp->bio_flags &= ~BIO_UNMAPPED;
820 	return (EJUSTRETURN);
821 }
822 
823 void
824 g_io_schedule_down(struct thread *tp __unused)
825 {
826 	struct bio *bp;
827 	int error;
828 
829 	for(;;) {
830 		g_bioq_lock(&g_bio_run_down);
831 		bp = g_bioq_first(&g_bio_run_down);
832 		if (bp == NULL) {
833 			CTR0(KTR_GEOM, "g_down going to sleep");
834 			msleep(&g_wait_down, &g_bio_run_down.bio_queue_lock,
835 			    PRIBIO | PDROP, "-", 0);
836 			continue;
837 		}
838 		CTR0(KTR_GEOM, "g_down has work to do");
839 		g_bioq_unlock(&g_bio_run_down);
840 		if (pace != 0) {
841 			/*
842 			 * There has been at least one memory allocation
843 			 * failure since the last I/O completed. Pause 1ms to
844 			 * give the system a chance to free up memory. We only
845 			 * do this once because a large number of allocations
846 			 * can fail in the direct dispatch case and there's no
847 			 * relationship between the number of these failures and
848 			 * the length of the outage. If there's still an outage,
849 			 * we'll pause again and again until it's
850 			 * resolved. Older versions paused longer and once per
851 			 * allocation failure. This was OK for a single threaded
852 			 * g_down, but with direct dispatch would lead to max of
853 			 * 10 IOPs for minutes at a time when transient memory
854 			 * issues prevented allocation for a batch of requests
855 			 * from the upper layers.
856 			 *
857 			 * XXX This pacing is really lame. It needs to be solved
858 			 * by other methods. This is OK only because the worst
859 			 * case scenario is so rare. In the worst case scenario
860 			 * all memory is tied up waiting for I/O to complete
861 			 * which can never happen since we can't allocate bios
862 			 * for that I/O.
863 			 */
864 			CTR0(KTR_GEOM, "g_down pacing self");
865 			pause("g_down", min(hz/1000, 1));
866 			pace = 0;
867 		}
868 		CTR2(KTR_GEOM, "g_down processing bp %p provider %s", bp,
869 		    bp->bio_to->name);
870 		error = g_io_check(bp);
871 		if (error >= 0) {
872 			CTR3(KTR_GEOM, "g_down g_io_check on bp %p provider "
873 			    "%s returned %d", bp, bp->bio_to->name, error);
874 			g_io_deliver(bp, error);
875 			continue;
876 		}
877 		THREAD_NO_SLEEPING();
878 		CTR4(KTR_GEOM, "g_down starting bp %p provider %s off %ld "
879 		    "len %ld", bp, bp->bio_to->name, bp->bio_offset,
880 		    bp->bio_length);
881 		bp->bio_to->geom->start(bp);
882 		THREAD_SLEEPING_OK();
883 	}
884 }
885 
886 void
887 bio_taskqueue(struct bio *bp, bio_task_t *func, void *arg)
888 {
889 	bp->bio_task = func;
890 	bp->bio_task_arg = arg;
891 	/*
892 	 * The taskqueue is actually just a second queue off the "up"
893 	 * queue, so we use the same lock.
894 	 */
895 	g_bioq_lock(&g_bio_run_up);
896 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
897 	    ("Bio already on queue bp=%p target taskq", bp));
898 	bp->bio_flags |= BIO_ONQUEUE;
899 	TAILQ_INSERT_TAIL(&g_bio_run_task.bio_queue, bp, bio_queue);
900 	g_bio_run_task.bio_queue_length++;
901 	wakeup(&g_wait_up);
902 	g_bioq_unlock(&g_bio_run_up);
903 }
904 
905 
906 void
907 g_io_schedule_up(struct thread *tp __unused)
908 {
909 	struct bio *bp;
910 	for(;;) {
911 		g_bioq_lock(&g_bio_run_up);
912 		bp = g_bioq_first(&g_bio_run_task);
913 		if (bp != NULL) {
914 			g_bioq_unlock(&g_bio_run_up);
915 			THREAD_NO_SLEEPING();
916 			CTR1(KTR_GEOM, "g_up processing task bp %p", bp);
917 			bp->bio_task(bp->bio_task_arg);
918 			THREAD_SLEEPING_OK();
919 			continue;
920 		}
921 		bp = g_bioq_first(&g_bio_run_up);
922 		if (bp != NULL) {
923 			g_bioq_unlock(&g_bio_run_up);
924 			THREAD_NO_SLEEPING();
925 			CTR4(KTR_GEOM, "g_up biodone bp %p provider %s off "
926 			    "%jd len %ld", bp, bp->bio_to->name,
927 			    bp->bio_offset, bp->bio_length);
928 			biodone(bp);
929 			THREAD_SLEEPING_OK();
930 			continue;
931 		}
932 		CTR0(KTR_GEOM, "g_up going to sleep");
933 		msleep(&g_wait_up, &g_bio_run_up.bio_queue_lock,
934 		    PRIBIO | PDROP, "-", 0);
935 	}
936 }
937 
938 void *
939 g_read_data(struct g_consumer *cp, off_t offset, off_t length, int *error)
940 {
941 	struct bio *bp;
942 	void *ptr;
943 	int errorc;
944 
945 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
946 	    length <= MAXPHYS, ("g_read_data(): invalid length %jd",
947 	    (intmax_t)length));
948 
949 	bp = g_alloc_bio();
950 	bp->bio_cmd = BIO_READ;
951 	bp->bio_done = NULL;
952 	bp->bio_offset = offset;
953 	bp->bio_length = length;
954 	ptr = g_malloc(length, M_WAITOK);
955 	bp->bio_data = ptr;
956 	g_io_request(bp, cp);
957 	errorc = biowait(bp, "gread");
958 	if (error != NULL)
959 		*error = errorc;
960 	g_destroy_bio(bp);
961 	if (errorc) {
962 		g_free(ptr);
963 		ptr = NULL;
964 	}
965 	return (ptr);
966 }
967 
968 int
969 g_write_data(struct g_consumer *cp, off_t offset, void *ptr, off_t length)
970 {
971 	struct bio *bp;
972 	int error;
973 
974 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
975 	    length <= MAXPHYS, ("g_write_data(): invalid length %jd",
976 	    (intmax_t)length));
977 
978 	bp = g_alloc_bio();
979 	bp->bio_cmd = BIO_WRITE;
980 	bp->bio_done = NULL;
981 	bp->bio_offset = offset;
982 	bp->bio_length = length;
983 	bp->bio_data = ptr;
984 	g_io_request(bp, cp);
985 	error = biowait(bp, "gwrite");
986 	g_destroy_bio(bp);
987 	return (error);
988 }
989 
990 int
991 g_delete_data(struct g_consumer *cp, off_t offset, off_t length)
992 {
993 	struct bio *bp;
994 	int error;
995 
996 	KASSERT(length > 0 && length >= cp->provider->sectorsize,
997 	    ("g_delete_data(): invalid length %jd", (intmax_t)length));
998 
999 	bp = g_alloc_bio();
1000 	bp->bio_cmd = BIO_DELETE;
1001 	bp->bio_done = NULL;
1002 	bp->bio_offset = offset;
1003 	bp->bio_length = length;
1004 	bp->bio_data = NULL;
1005 	g_io_request(bp, cp);
1006 	error = biowait(bp, "gdelete");
1007 	g_destroy_bio(bp);
1008 	return (error);
1009 }
1010 
1011 void
1012 g_print_bio(struct bio *bp)
1013 {
1014 	const char *pname, *cmd = NULL;
1015 
1016 	if (bp->bio_to != NULL)
1017 		pname = bp->bio_to->name;
1018 	else
1019 		pname = "[unknown]";
1020 
1021 	switch (bp->bio_cmd) {
1022 	case BIO_GETATTR:
1023 		cmd = "GETATTR";
1024 		printf("%s[%s(attr=%s)]", pname, cmd, bp->bio_attribute);
1025 		return;
1026 	case BIO_FLUSH:
1027 		cmd = "FLUSH";
1028 		printf("%s[%s]", pname, cmd);
1029 		return;
1030 	case BIO_ZONE: {
1031 		char *subcmd = NULL;
1032 		cmd = "ZONE";
1033 		switch (bp->bio_zone.zone_cmd) {
1034 		case DISK_ZONE_OPEN:
1035 			subcmd = "OPEN";
1036 			break;
1037 		case DISK_ZONE_CLOSE:
1038 			subcmd = "CLOSE";
1039 			break;
1040 		case DISK_ZONE_FINISH:
1041 			subcmd = "FINISH";
1042 			break;
1043 		case DISK_ZONE_RWP:
1044 			subcmd = "RWP";
1045 			break;
1046 		case DISK_ZONE_REPORT_ZONES:
1047 			subcmd = "REPORT ZONES";
1048 			break;
1049 		case DISK_ZONE_GET_PARAMS:
1050 			subcmd = "GET PARAMS";
1051 			break;
1052 		default:
1053 			subcmd = "UNKNOWN";
1054 			break;
1055 		}
1056 		printf("%s[%s,%s]", pname, cmd, subcmd);
1057 		return;
1058 	}
1059 	case BIO_READ:
1060 		cmd = "READ";
1061 		break;
1062 	case BIO_WRITE:
1063 		cmd = "WRITE";
1064 		break;
1065 	case BIO_DELETE:
1066 		cmd = "DELETE";
1067 		break;
1068 	default:
1069 		cmd = "UNKNOWN";
1070 		printf("%s[%s()]", pname, cmd);
1071 		return;
1072 	}
1073 	printf("%s[%s(offset=%jd, length=%jd)]", pname, cmd,
1074 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
1075 }
1076