xref: /freebsd/sys/geom/geom_io.c (revision 8ef24a0d4b28fe230e20637f56869cc4148cd2ca)
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 		/* Inherit classification info from the parent */
222 		bp2->bio_classifier1 = bp->bio_classifier1;
223 		bp2->bio_classifier2 = bp->bio_classifier2;
224 		bp->bio_children++;
225 	}
226 #ifdef KTR
227 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
228 		struct stack st;
229 
230 		CTR2(KTR_GEOM, "g_clone_bio(%p): %p", bp, bp2);
231 		stack_save(&st);
232 		CTRSTACK(KTR_GEOM, &st, 3, 0);
233 	}
234 #endif
235 	return(bp2);
236 }
237 
238 struct bio *
239 g_duplicate_bio(struct bio *bp)
240 {
241 	struct bio *bp2;
242 
243 	bp2 = uma_zalloc(biozone, M_WAITOK | M_ZERO);
244 	bp2->bio_flags = bp->bio_flags & (BIO_UNMAPPED | BIO_VLIST);
245 	bp2->bio_parent = bp;
246 	bp2->bio_cmd = bp->bio_cmd;
247 	bp2->bio_length = bp->bio_length;
248 	bp2->bio_offset = bp->bio_offset;
249 	bp2->bio_data = bp->bio_data;
250 	bp2->bio_ma = bp->bio_ma;
251 	bp2->bio_ma_n = bp->bio_ma_n;
252 	bp2->bio_ma_offset = bp->bio_ma_offset;
253 	bp2->bio_attribute = bp->bio_attribute;
254 	bp->bio_children++;
255 #ifdef KTR
256 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
257 		struct stack st;
258 
259 		CTR2(KTR_GEOM, "g_duplicate_bio(%p): %p", bp, bp2);
260 		stack_save(&st);
261 		CTRSTACK(KTR_GEOM, &st, 3, 0);
262 	}
263 #endif
264 	return(bp2);
265 }
266 
267 void
268 g_reset_bio(struct bio *bp)
269 {
270 
271 	bzero(bp, sizeof(*bp));
272 }
273 
274 void
275 g_io_init()
276 {
277 
278 	g_bioq_init(&g_bio_run_down);
279 	g_bioq_init(&g_bio_run_up);
280 	g_bioq_init(&g_bio_run_task);
281 	biozone = uma_zcreate("g_bio", sizeof (struct bio),
282 	    NULL, NULL,
283 	    NULL, NULL,
284 	    0, 0);
285 }
286 
287 int
288 g_io_getattr(const char *attr, struct g_consumer *cp, int *len, void *ptr)
289 {
290 	struct bio *bp;
291 	int error;
292 
293 	g_trace(G_T_BIO, "bio_getattr(%s)", attr);
294 	bp = g_alloc_bio();
295 	bp->bio_cmd = BIO_GETATTR;
296 	bp->bio_done = NULL;
297 	bp->bio_attribute = attr;
298 	bp->bio_length = *len;
299 	bp->bio_data = ptr;
300 	g_io_request(bp, cp);
301 	error = biowait(bp, "ggetattr");
302 	*len = bp->bio_completed;
303 	g_destroy_bio(bp);
304 	return (error);
305 }
306 
307 int
308 g_io_flush(struct g_consumer *cp)
309 {
310 	struct bio *bp;
311 	int error;
312 
313 	g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name);
314 	bp = g_alloc_bio();
315 	bp->bio_cmd = BIO_FLUSH;
316 	bp->bio_flags |= BIO_ORDERED;
317 	bp->bio_done = NULL;
318 	bp->bio_attribute = NULL;
319 	bp->bio_offset = cp->provider->mediasize;
320 	bp->bio_length = 0;
321 	bp->bio_data = NULL;
322 	g_io_request(bp, cp);
323 	error = biowait(bp, "gflush");
324 	g_destroy_bio(bp);
325 	return (error);
326 }
327 
328 static int
329 g_io_check(struct bio *bp)
330 {
331 	struct g_consumer *cp;
332 	struct g_provider *pp;
333 	off_t excess;
334 	int error;
335 
336 	cp = bp->bio_from;
337 	pp = bp->bio_to;
338 
339 	/* Fail if access counters dont allow the operation */
340 	switch(bp->bio_cmd) {
341 	case BIO_READ:
342 	case BIO_GETATTR:
343 		if (cp->acr == 0)
344 			return (EPERM);
345 		break;
346 	case BIO_WRITE:
347 	case BIO_DELETE:
348 	case BIO_FLUSH:
349 		if (cp->acw == 0)
350 			return (EPERM);
351 		break;
352 	default:
353 		return (EPERM);
354 	}
355 	/* if provider is marked for error, don't disturb. */
356 	if (pp->error)
357 		return (pp->error);
358 	if (cp->flags & G_CF_ORPHAN)
359 		return (ENXIO);
360 
361 	switch(bp->bio_cmd) {
362 	case BIO_READ:
363 	case BIO_WRITE:
364 	case BIO_DELETE:
365 		/* Zero sectorsize or mediasize is probably a lack of media. */
366 		if (pp->sectorsize == 0 || pp->mediasize == 0)
367 			return (ENXIO);
368 		/* Reject I/O not on sector boundary */
369 		if (bp->bio_offset % pp->sectorsize)
370 			return (EINVAL);
371 		/* Reject I/O not integral sector long */
372 		if (bp->bio_length % pp->sectorsize)
373 			return (EINVAL);
374 		/* Reject requests before or past the end of media. */
375 		if (bp->bio_offset < 0)
376 			return (EIO);
377 		if (bp->bio_offset > pp->mediasize)
378 			return (EIO);
379 
380 		/* Truncate requests to the end of providers media. */
381 		excess = bp->bio_offset + bp->bio_length;
382 		if (excess > bp->bio_to->mediasize) {
383 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
384 			    round_page(bp->bio_ma_offset +
385 			    bp->bio_length) / PAGE_SIZE == bp->bio_ma_n,
386 			    ("excess bio %p too short", bp));
387 			excess -= bp->bio_to->mediasize;
388 			bp->bio_length -= excess;
389 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
390 				bp->bio_ma_n = round_page(bp->bio_ma_offset +
391 				    bp->bio_length) / PAGE_SIZE;
392 			}
393 			if (excess > 0)
394 				CTR3(KTR_GEOM, "g_down truncated bio "
395 				    "%p provider %s by %d", bp,
396 				    bp->bio_to->name, excess);
397 		}
398 
399 		/* Deliver zero length transfers right here. */
400 		if (bp->bio_length == 0) {
401 			CTR2(KTR_GEOM, "g_down terminated 0-length "
402 			    "bp %p provider %s", bp, bp->bio_to->name);
403 			return (0);
404 		}
405 
406 		if ((bp->bio_flags & BIO_UNMAPPED) != 0 &&
407 		    (bp->bio_to->flags & G_PF_ACCEPT_UNMAPPED) == 0 &&
408 		    (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) {
409 			if ((error = g_io_transient_map_bio(bp)) >= 0)
410 				return (error);
411 		}
412 		break;
413 	default:
414 		break;
415 	}
416 	return (EJUSTRETURN);
417 }
418 
419 /*
420  * bio classification support.
421  *
422  * g_register_classifier() and g_unregister_classifier()
423  * are used to add/remove a classifier from the list.
424  * The list is protected using the g_bio_run_down lock,
425  * because the classifiers are called in this path.
426  *
427  * g_io_request() passes bio's that are not already classified
428  * (i.e. those with bio_classifier1 == NULL) to g_run_classifiers().
429  * Classifiers can store their result in the two fields
430  * bio_classifier1 and bio_classifier2.
431  * A classifier that updates one of the fields should
432  * return a non-zero value.
433  * If no classifier updates the field, g_run_classifiers() sets
434  * bio_classifier1 = BIO_NOTCLASSIFIED to avoid further calls.
435  */
436 
437 int
438 g_register_classifier(struct g_classifier_hook *hook)
439 {
440 
441 	g_bioq_lock(&g_bio_run_down);
442 	TAILQ_INSERT_TAIL(&g_classifier_tailq, hook, link);
443 	g_bioq_unlock(&g_bio_run_down);
444 
445 	return (0);
446 }
447 
448 void
449 g_unregister_classifier(struct g_classifier_hook *hook)
450 {
451 	struct g_classifier_hook *entry;
452 
453 	g_bioq_lock(&g_bio_run_down);
454 	TAILQ_FOREACH(entry, &g_classifier_tailq, link) {
455 		if (entry == hook) {
456 			TAILQ_REMOVE(&g_classifier_tailq, hook, link);
457 			break;
458 		}
459 	}
460 	g_bioq_unlock(&g_bio_run_down);
461 }
462 
463 static void
464 g_run_classifiers(struct bio *bp)
465 {
466 	struct g_classifier_hook *hook;
467 	int classified = 0;
468 
469 	TAILQ_FOREACH(hook, &g_classifier_tailq, link)
470 		classified |= hook->func(hook->arg, bp);
471 
472 	if (!classified)
473 		bp->bio_classifier1 = BIO_NOTCLASSIFIED;
474 }
475 
476 void
477 g_io_request(struct bio *bp, struct g_consumer *cp)
478 {
479 	struct g_provider *pp;
480 	struct mtx *mtxp;
481 	int direct, error, first;
482 	uint8_t cmd;
483 
484 	KASSERT(cp != NULL, ("NULL cp in g_io_request"));
485 	KASSERT(bp != NULL, ("NULL bp in g_io_request"));
486 	pp = cp->provider;
487 	KASSERT(pp != NULL, ("consumer not attached in g_io_request"));
488 #ifdef DIAGNOSTIC
489 	KASSERT(bp->bio_driver1 == NULL,
490 	    ("bio_driver1 used by the consumer (geom %s)", cp->geom->name));
491 	KASSERT(bp->bio_driver2 == NULL,
492 	    ("bio_driver2 used by the consumer (geom %s)", cp->geom->name));
493 	KASSERT(bp->bio_pflags == 0,
494 	    ("bio_pflags used by the consumer (geom %s)", cp->geom->name));
495 	/*
496 	 * Remember consumer's private fields, so we can detect if they were
497 	 * modified by the provider.
498 	 */
499 	bp->_bio_caller1 = bp->bio_caller1;
500 	bp->_bio_caller2 = bp->bio_caller2;
501 	bp->_bio_cflags = bp->bio_cflags;
502 #endif
503 
504 	cmd = bp->bio_cmd;
505 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_GETATTR) {
506 		KASSERT(bp->bio_data != NULL,
507 		    ("NULL bp->data in g_io_request(cmd=%hu)", bp->bio_cmd));
508 	}
509 	if (cmd == BIO_DELETE || cmd == BIO_FLUSH) {
510 		KASSERT(bp->bio_data == NULL,
511 		    ("non-NULL bp->data in g_io_request(cmd=%hu)",
512 		    bp->bio_cmd));
513 	}
514 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_DELETE) {
515 		KASSERT(bp->bio_offset % cp->provider->sectorsize == 0,
516 		    ("wrong offset %jd for sectorsize %u",
517 		    bp->bio_offset, cp->provider->sectorsize));
518 		KASSERT(bp->bio_length % cp->provider->sectorsize == 0,
519 		    ("wrong length %jd for sectorsize %u",
520 		    bp->bio_length, cp->provider->sectorsize));
521 	}
522 
523 	g_trace(G_T_BIO, "bio_request(%p) from %p(%s) to %p(%s) cmd %d",
524 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd);
525 
526 	bp->bio_from = cp;
527 	bp->bio_to = pp;
528 	bp->bio_error = 0;
529 	bp->bio_completed = 0;
530 
531 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
532 	    ("Bio already on queue bp=%p", bp));
533 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
534 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
535 		binuptime(&bp->bio_t0);
536 	else
537 		getbinuptime(&bp->bio_t0);
538 
539 #ifdef GET_STACK_USAGE
540 	direct = (cp->flags & G_CF_DIRECT_SEND) != 0 &&
541 	    (pp->flags & G_PF_DIRECT_RECEIVE) != 0 &&
542 	    !g_is_geom_thread(curthread) &&
543 	    ((pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ||
544 	    (bp->bio_flags & BIO_UNMAPPED) == 0 || THREAD_CAN_SLEEP()) &&
545 	    pace == 0;
546 	if (direct) {
547 		/* Block direct execution if less then half of stack left. */
548 		size_t	st, su;
549 		GET_STACK_USAGE(st, su);
550 		if (su * 2 > st)
551 			direct = 0;
552 	}
553 #else
554 	direct = 0;
555 #endif
556 
557 	if (!TAILQ_EMPTY(&g_classifier_tailq) && !bp->bio_classifier1) {
558 		g_bioq_lock(&g_bio_run_down);
559 		g_run_classifiers(bp);
560 		g_bioq_unlock(&g_bio_run_down);
561 	}
562 
563 	/*
564 	 * The statistics collection is lockless, as such, but we
565 	 * can not update one instance of the statistics from more
566 	 * than one thread at a time, so grab the lock first.
567 	 */
568 	mtxp = mtx_pool_find(mtxpool_sleep, pp);
569 	mtx_lock(mtxp);
570 	if (g_collectstats & G_STATS_PROVIDERS)
571 		devstat_start_transaction(pp->stat, &bp->bio_t0);
572 	if (g_collectstats & G_STATS_CONSUMERS)
573 		devstat_start_transaction(cp->stat, &bp->bio_t0);
574 	pp->nstart++;
575 	cp->nstart++;
576 	mtx_unlock(mtxp);
577 
578 	if (direct) {
579 		error = g_io_check(bp);
580 		if (error >= 0) {
581 			CTR3(KTR_GEOM, "g_io_request g_io_check on bp %p "
582 			    "provider %s returned %d", bp, bp->bio_to->name,
583 			    error);
584 			g_io_deliver(bp, error);
585 			return;
586 		}
587 		bp->bio_to->geom->start(bp);
588 	} else {
589 		g_bioq_lock(&g_bio_run_down);
590 		first = TAILQ_EMPTY(&g_bio_run_down.bio_queue);
591 		TAILQ_INSERT_TAIL(&g_bio_run_down.bio_queue, bp, bio_queue);
592 		bp->bio_flags |= BIO_ONQUEUE;
593 		g_bio_run_down.bio_queue_length++;
594 		g_bioq_unlock(&g_bio_run_down);
595 		/* Pass it on down. */
596 		if (first)
597 			wakeup(&g_wait_down);
598 	}
599 }
600 
601 void
602 g_io_deliver(struct bio *bp, int error)
603 {
604 	struct bintime now;
605 	struct g_consumer *cp;
606 	struct g_provider *pp;
607 	struct mtx *mtxp;
608 	int direct, first;
609 
610 	KASSERT(bp != NULL, ("NULL bp in g_io_deliver"));
611 	pp = bp->bio_to;
612 	KASSERT(pp != NULL, ("NULL bio_to in g_io_deliver"));
613 	cp = bp->bio_from;
614 	if (cp == NULL) {
615 		bp->bio_error = error;
616 		bp->bio_done(bp);
617 		return;
618 	}
619 	KASSERT(cp != NULL, ("NULL bio_from in g_io_deliver"));
620 	KASSERT(cp->geom != NULL, ("NULL bio_from->geom in g_io_deliver"));
621 #ifdef DIAGNOSTIC
622 	/*
623 	 * Some classes - GJournal in particular - can modify bio's
624 	 * private fields while the bio is in transit; G_GEOM_VOLATILE_BIO
625 	 * flag means it's an expected behaviour for that particular geom.
626 	 */
627 	if ((cp->geom->flags & G_GEOM_VOLATILE_BIO) == 0) {
628 		KASSERT(bp->bio_caller1 == bp->_bio_caller1,
629 		    ("bio_caller1 used by the provider %s", pp->name));
630 		KASSERT(bp->bio_caller2 == bp->_bio_caller2,
631 		    ("bio_caller2 used by the provider %s", pp->name));
632 		KASSERT(bp->bio_cflags == bp->_bio_cflags,
633 		    ("bio_cflags used by the provider %s", pp->name));
634 	}
635 #endif
636 	KASSERT(bp->bio_completed >= 0, ("bio_completed can't be less than 0"));
637 	KASSERT(bp->bio_completed <= bp->bio_length,
638 	    ("bio_completed can't be greater than bio_length"));
639 
640 	g_trace(G_T_BIO,
641 "g_io_deliver(%p) from %p(%s) to %p(%s) cmd %d error %d off %jd len %jd",
642 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd, error,
643 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
644 
645 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
646 	    ("Bio already on queue bp=%p", bp));
647 
648 	/*
649 	 * XXX: next two doesn't belong here
650 	 */
651 	bp->bio_bcount = bp->bio_length;
652 	bp->bio_resid = bp->bio_bcount - bp->bio_completed;
653 
654 #ifdef GET_STACK_USAGE
655 	direct = (pp->flags & G_PF_DIRECT_SEND) &&
656 		 (cp->flags & G_CF_DIRECT_RECEIVE) &&
657 		 !g_is_geom_thread(curthread);
658 	if (direct) {
659 		/* Block direct execution if less then half of stack left. */
660 		size_t	st, su;
661 		GET_STACK_USAGE(st, su);
662 		if (su * 2 > st)
663 			direct = 0;
664 	}
665 #else
666 	direct = 0;
667 #endif
668 
669 	/*
670 	 * The statistics collection is lockless, as such, but we
671 	 * can not update one instance of the statistics from more
672 	 * than one thread at a time, so grab the lock first.
673 	 */
674 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
675 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
676 		binuptime(&now);
677 	mtxp = mtx_pool_find(mtxpool_sleep, cp);
678 	mtx_lock(mtxp);
679 	if (g_collectstats & G_STATS_PROVIDERS)
680 		devstat_end_transaction_bio_bt(pp->stat, bp, &now);
681 	if (g_collectstats & G_STATS_CONSUMERS)
682 		devstat_end_transaction_bio_bt(cp->stat, bp, &now);
683 	cp->nend++;
684 	pp->nend++;
685 	mtx_unlock(mtxp);
686 
687 	if (error != ENOMEM) {
688 		bp->bio_error = error;
689 		if (direct) {
690 			biodone(bp);
691 		} else {
692 			g_bioq_lock(&g_bio_run_up);
693 			first = TAILQ_EMPTY(&g_bio_run_up.bio_queue);
694 			TAILQ_INSERT_TAIL(&g_bio_run_up.bio_queue, bp, bio_queue);
695 			bp->bio_flags |= BIO_ONQUEUE;
696 			g_bio_run_up.bio_queue_length++;
697 			g_bioq_unlock(&g_bio_run_up);
698 			if (first)
699 				wakeup(&g_wait_up);
700 		}
701 		return;
702 	}
703 
704 	if (bootverbose)
705 		printf("ENOMEM %p on %p(%s)\n", bp, pp, pp->name);
706 	bp->bio_children = 0;
707 	bp->bio_inbed = 0;
708 	bp->bio_driver1 = NULL;
709 	bp->bio_driver2 = NULL;
710 	bp->bio_pflags = 0;
711 	g_io_request(bp, cp);
712 	pace = 1;
713 	return;
714 }
715 
716 SYSCTL_DECL(_kern_geom);
717 
718 static long transient_maps;
719 SYSCTL_LONG(_kern_geom, OID_AUTO, transient_maps, CTLFLAG_RD,
720     &transient_maps, 0,
721     "Total count of the transient mapping requests");
722 u_int transient_map_retries = 10;
723 SYSCTL_UINT(_kern_geom, OID_AUTO, transient_map_retries, CTLFLAG_RW,
724     &transient_map_retries, 0,
725     "Max count of retries used before giving up on creating transient map");
726 int transient_map_hard_failures;
727 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_hard_failures, CTLFLAG_RD,
728     &transient_map_hard_failures, 0,
729     "Failures to establish the transient mapping due to retry attempts "
730     "exhausted");
731 int transient_map_soft_failures;
732 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_soft_failures, CTLFLAG_RD,
733     &transient_map_soft_failures, 0,
734     "Count of retried failures to establish the transient mapping");
735 int inflight_transient_maps;
736 SYSCTL_INT(_kern_geom, OID_AUTO, inflight_transient_maps, CTLFLAG_RD,
737     &inflight_transient_maps, 0,
738     "Current count of the active transient maps");
739 
740 static int
741 g_io_transient_map_bio(struct bio *bp)
742 {
743 	vm_offset_t addr;
744 	long size;
745 	u_int retried;
746 
747 	KASSERT(unmapped_buf_allowed, ("unmapped disabled"));
748 
749 	size = round_page(bp->bio_ma_offset + bp->bio_length);
750 	KASSERT(size / PAGE_SIZE == bp->bio_ma_n, ("Bio too short %p", bp));
751 	addr = 0;
752 	retried = 0;
753 	atomic_add_long(&transient_maps, 1);
754 retry:
755 	if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) {
756 		if (transient_map_retries != 0 &&
757 		    retried >= transient_map_retries) {
758 			CTR2(KTR_GEOM, "g_down cannot map bp %p provider %s",
759 			    bp, bp->bio_to->name);
760 			atomic_add_int(&transient_map_hard_failures, 1);
761 			return (EDEADLK/* XXXKIB */);
762 		} else {
763 			/*
764 			 * Naive attempt to quisce the I/O to get more
765 			 * in-flight requests completed and defragment
766 			 * the transient_arena.
767 			 */
768 			CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d",
769 			    bp, bp->bio_to->name, retried);
770 			pause("g_d_tra", hz / 10);
771 			retried++;
772 			atomic_add_int(&transient_map_soft_failures, 1);
773 			goto retry;
774 		}
775 	}
776 	atomic_add_int(&inflight_transient_maps, 1);
777 	pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size));
778 	bp->bio_data = (caddr_t)addr + bp->bio_ma_offset;
779 	bp->bio_flags |= BIO_TRANSIENT_MAPPING;
780 	bp->bio_flags &= ~BIO_UNMAPPED;
781 	return (EJUSTRETURN);
782 }
783 
784 void
785 g_io_schedule_down(struct thread *tp __unused)
786 {
787 	struct bio *bp;
788 	int error;
789 
790 	for(;;) {
791 		g_bioq_lock(&g_bio_run_down);
792 		bp = g_bioq_first(&g_bio_run_down);
793 		if (bp == NULL) {
794 			CTR0(KTR_GEOM, "g_down going to sleep");
795 			msleep(&g_wait_down, &g_bio_run_down.bio_queue_lock,
796 			    PRIBIO | PDROP, "-", 0);
797 			continue;
798 		}
799 		CTR0(KTR_GEOM, "g_down has work to do");
800 		g_bioq_unlock(&g_bio_run_down);
801 		if (pace != 0) {
802 			/*
803 			 * There has been at least one memory allocation
804 			 * failure since the last I/O completed. Pause 1ms to
805 			 * give the system a chance to free up memory. We only
806 			 * do this once because a large number of allocations
807 			 * can fail in the direct dispatch case and there's no
808 			 * relationship between the number of these failures and
809 			 * the length of the outage. If there's still an outage,
810 			 * we'll pause again and again until it's
811 			 * resolved. Older versions paused longer and once per
812 			 * allocation failure. This was OK for a single threaded
813 			 * g_down, but with direct dispatch would lead to max of
814 			 * 10 IOPs for minutes at a time when transient memory
815 			 * issues prevented allocation for a batch of requests
816 			 * from the upper layers.
817 			 *
818 			 * XXX This pacing is really lame. It needs to be solved
819 			 * by other methods. This is OK only because the worst
820 			 * case scenario is so rare. In the worst case scenario
821 			 * all memory is tied up waiting for I/O to complete
822 			 * which can never happen since we can't allocate bios
823 			 * for that I/O.
824 			 */
825 			CTR0(KTR_GEOM, "g_down pacing self");
826 			pause("g_down", min(hz/1000, 1));
827 			pace = 0;
828 		}
829 		CTR2(KTR_GEOM, "g_down processing bp %p provider %s", bp,
830 		    bp->bio_to->name);
831 		error = g_io_check(bp);
832 		if (error >= 0) {
833 			CTR3(KTR_GEOM, "g_down g_io_check on bp %p provider "
834 			    "%s returned %d", bp, bp->bio_to->name, error);
835 			g_io_deliver(bp, error);
836 			continue;
837 		}
838 		THREAD_NO_SLEEPING();
839 		CTR4(KTR_GEOM, "g_down starting bp %p provider %s off %ld "
840 		    "len %ld", bp, bp->bio_to->name, bp->bio_offset,
841 		    bp->bio_length);
842 		bp->bio_to->geom->start(bp);
843 		THREAD_SLEEPING_OK();
844 	}
845 }
846 
847 void
848 bio_taskqueue(struct bio *bp, bio_task_t *func, void *arg)
849 {
850 	bp->bio_task = func;
851 	bp->bio_task_arg = arg;
852 	/*
853 	 * The taskqueue is actually just a second queue off the "up"
854 	 * queue, so we use the same lock.
855 	 */
856 	g_bioq_lock(&g_bio_run_up);
857 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
858 	    ("Bio already on queue bp=%p target taskq", bp));
859 	bp->bio_flags |= BIO_ONQUEUE;
860 	TAILQ_INSERT_TAIL(&g_bio_run_task.bio_queue, bp, bio_queue);
861 	g_bio_run_task.bio_queue_length++;
862 	wakeup(&g_wait_up);
863 	g_bioq_unlock(&g_bio_run_up);
864 }
865 
866 
867 void
868 g_io_schedule_up(struct thread *tp __unused)
869 {
870 	struct bio *bp;
871 	for(;;) {
872 		g_bioq_lock(&g_bio_run_up);
873 		bp = g_bioq_first(&g_bio_run_task);
874 		if (bp != NULL) {
875 			g_bioq_unlock(&g_bio_run_up);
876 			THREAD_NO_SLEEPING();
877 			CTR1(KTR_GEOM, "g_up processing task bp %p", bp);
878 			bp->bio_task(bp->bio_task_arg);
879 			THREAD_SLEEPING_OK();
880 			continue;
881 		}
882 		bp = g_bioq_first(&g_bio_run_up);
883 		if (bp != NULL) {
884 			g_bioq_unlock(&g_bio_run_up);
885 			THREAD_NO_SLEEPING();
886 			CTR4(KTR_GEOM, "g_up biodone bp %p provider %s off "
887 			    "%jd len %ld", bp, bp->bio_to->name,
888 			    bp->bio_offset, bp->bio_length);
889 			biodone(bp);
890 			THREAD_SLEEPING_OK();
891 			continue;
892 		}
893 		CTR0(KTR_GEOM, "g_up going to sleep");
894 		msleep(&g_wait_up, &g_bio_run_up.bio_queue_lock,
895 		    PRIBIO | PDROP, "-", 0);
896 	}
897 }
898 
899 void *
900 g_read_data(struct g_consumer *cp, off_t offset, off_t length, int *error)
901 {
902 	struct bio *bp;
903 	void *ptr;
904 	int errorc;
905 
906 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
907 	    length <= MAXPHYS, ("g_read_data(): invalid length %jd",
908 	    (intmax_t)length));
909 
910 	bp = g_alloc_bio();
911 	bp->bio_cmd = BIO_READ;
912 	bp->bio_done = NULL;
913 	bp->bio_offset = offset;
914 	bp->bio_length = length;
915 	ptr = g_malloc(length, M_WAITOK);
916 	bp->bio_data = ptr;
917 	g_io_request(bp, cp);
918 	errorc = biowait(bp, "gread");
919 	if (error != NULL)
920 		*error = errorc;
921 	g_destroy_bio(bp);
922 	if (errorc) {
923 		g_free(ptr);
924 		ptr = NULL;
925 	}
926 	return (ptr);
927 }
928 
929 int
930 g_write_data(struct g_consumer *cp, off_t offset, void *ptr, off_t length)
931 {
932 	struct bio *bp;
933 	int error;
934 
935 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
936 	    length <= MAXPHYS, ("g_write_data(): invalid length %jd",
937 	    (intmax_t)length));
938 
939 	bp = g_alloc_bio();
940 	bp->bio_cmd = BIO_WRITE;
941 	bp->bio_done = NULL;
942 	bp->bio_offset = offset;
943 	bp->bio_length = length;
944 	bp->bio_data = ptr;
945 	g_io_request(bp, cp);
946 	error = biowait(bp, "gwrite");
947 	g_destroy_bio(bp);
948 	return (error);
949 }
950 
951 int
952 g_delete_data(struct g_consumer *cp, off_t offset, off_t length)
953 {
954 	struct bio *bp;
955 	int error;
956 
957 	KASSERT(length > 0 && length >= cp->provider->sectorsize,
958 	    ("g_delete_data(): invalid length %jd", (intmax_t)length));
959 
960 	bp = g_alloc_bio();
961 	bp->bio_cmd = BIO_DELETE;
962 	bp->bio_done = NULL;
963 	bp->bio_offset = offset;
964 	bp->bio_length = length;
965 	bp->bio_data = NULL;
966 	g_io_request(bp, cp);
967 	error = biowait(bp, "gdelete");
968 	g_destroy_bio(bp);
969 	return (error);
970 }
971 
972 void
973 g_print_bio(struct bio *bp)
974 {
975 	const char *pname, *cmd = NULL;
976 
977 	if (bp->bio_to != NULL)
978 		pname = bp->bio_to->name;
979 	else
980 		pname = "[unknown]";
981 
982 	switch (bp->bio_cmd) {
983 	case BIO_GETATTR:
984 		cmd = "GETATTR";
985 		printf("%s[%s(attr=%s)]", pname, cmd, bp->bio_attribute);
986 		return;
987 	case BIO_FLUSH:
988 		cmd = "FLUSH";
989 		printf("%s[%s]", pname, cmd);
990 		return;
991 	case BIO_READ:
992 		cmd = "READ";
993 		break;
994 	case BIO_WRITE:
995 		cmd = "WRITE";
996 		break;
997 	case BIO_DELETE:
998 		cmd = "DELETE";
999 		break;
1000 	default:
1001 		cmd = "UNKNOWN";
1002 		printf("%s[%s()]", pname, cmd);
1003 		return;
1004 	}
1005 	printf("%s[%s(offset=%jd, length=%jd)]", pname, cmd,
1006 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
1007 }
1008