xref: /freebsd/sys/geom/geom_io.c (revision 74ca7bf1d4c7173d5575ba168bc4b5f6d181ff5a)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2002 Poul-Henning Kamp
5  * Copyright (c) 2002 Networks Associates Technology, Inc.
6  * Copyright (c) 2013 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * This software was developed for the FreeBSD Project by Poul-Henning Kamp
10  * and NAI Labs, the Security Research Division of Network Associates, Inc.
11  * under DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the
12  * DARPA CHATS research program.
13  *
14  * Portions of this software were developed by Konstantin Belousov
15  * under sponsorship from the FreeBSD Foundation.
16  *
17  * Redistribution and use in source and binary forms, with or without
18  * modification, are permitted provided that the following conditions
19  * are met:
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  * 3. The names of the authors may not be used to endorse or promote
26  *    products derived from this software without specific prior written
27  *    permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  */
41 
42 #include <sys/cdefs.h>
43 __FBSDID("$FreeBSD$");
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/malloc.h>
49 #include <sys/bio.h>
50 #include <sys/ktr.h>
51 #include <sys/proc.h>
52 #include <sys/stack.h>
53 #include <sys/sysctl.h>
54 #include <sys/vmem.h>
55 
56 #include <sys/errno.h>
57 #include <geom/geom.h>
58 #include <geom/geom_int.h>
59 #include <sys/devicestat.h>
60 
61 #include <vm/uma.h>
62 #include <vm/vm.h>
63 #include <vm/vm_param.h>
64 #include <vm/vm_kern.h>
65 #include <vm/vm_page.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_extern.h>
68 #include <vm/vm_map.h>
69 
70 static int	g_io_transient_map_bio(struct bio *bp);
71 
72 static struct g_bioq g_bio_run_down;
73 static struct g_bioq g_bio_run_up;
74 
75 /*
76  * Pace is a hint that we've had some trouble recently allocating
77  * bios, so we should back off trying to send I/O down the stack
78  * a bit to let the problem resolve. When pacing, we also turn
79  * off direct dispatch to also reduce memory pressure from I/Os
80  * there, at the expxense of some added latency while the memory
81  * pressures exist. See g_io_schedule_down() for more details
82  * and limitations.
83  */
84 static volatile u_int pace;
85 
86 static uma_zone_t	biozone;
87 
88 /*
89  * The head of the list of classifiers used in g_io_request.
90  * Use g_register_classifier() and g_unregister_classifier()
91  * to add/remove entries to the list.
92  * Classifiers are invoked in registration order.
93  */
94 static TAILQ_HEAD(g_classifier_tailq, g_classifier_hook)
95     g_classifier_tailq = TAILQ_HEAD_INITIALIZER(g_classifier_tailq);
96 
97 #include <machine/atomic.h>
98 
99 static void
100 g_bioq_lock(struct g_bioq *bq)
101 {
102 
103 	mtx_lock(&bq->bio_queue_lock);
104 }
105 
106 static void
107 g_bioq_unlock(struct g_bioq *bq)
108 {
109 
110 	mtx_unlock(&bq->bio_queue_lock);
111 }
112 
113 #if 0
114 static void
115 g_bioq_destroy(struct g_bioq *bq)
116 {
117 
118 	mtx_destroy(&bq->bio_queue_lock);
119 }
120 #endif
121 
122 static void
123 g_bioq_init(struct g_bioq *bq)
124 {
125 
126 	TAILQ_INIT(&bq->bio_queue);
127 	mtx_init(&bq->bio_queue_lock, "bio queue", NULL, MTX_DEF);
128 }
129 
130 static struct bio *
131 g_bioq_first(struct g_bioq *bq)
132 {
133 	struct bio *bp;
134 
135 	bp = TAILQ_FIRST(&bq->bio_queue);
136 	if (bp != NULL) {
137 		KASSERT((bp->bio_flags & BIO_ONQUEUE),
138 		    ("Bio not on queue bp=%p target %p", bp, bq));
139 		bp->bio_flags &= ~BIO_ONQUEUE;
140 		TAILQ_REMOVE(&bq->bio_queue, bp, bio_queue);
141 		bq->bio_queue_length--;
142 	}
143 	return (bp);
144 }
145 
146 struct bio *
147 g_new_bio(void)
148 {
149 	struct bio *bp;
150 
151 	bp = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
152 #ifdef KTR
153 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
154 		struct stack st;
155 
156 		CTR1(KTR_GEOM, "g_new_bio(): %p", bp);
157 		stack_save(&st);
158 		CTRSTACK(KTR_GEOM, &st, 3, 0);
159 	}
160 #endif
161 	return (bp);
162 }
163 
164 struct bio *
165 g_alloc_bio(void)
166 {
167 	struct bio *bp;
168 
169 	bp = uma_zalloc(biozone, M_WAITOK | M_ZERO);
170 #ifdef KTR
171 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
172 		struct stack st;
173 
174 		CTR1(KTR_GEOM, "g_alloc_bio(): %p", bp);
175 		stack_save(&st);
176 		CTRSTACK(KTR_GEOM, &st, 3, 0);
177 	}
178 #endif
179 	return (bp);
180 }
181 
182 void
183 g_destroy_bio(struct bio *bp)
184 {
185 #ifdef KTR
186 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
187 		struct stack st;
188 
189 		CTR1(KTR_GEOM, "g_destroy_bio(): %p", bp);
190 		stack_save(&st);
191 		CTRSTACK(KTR_GEOM, &st, 3, 0);
192 	}
193 #endif
194 	uma_zfree(biozone, bp);
195 }
196 
197 struct bio *
198 g_clone_bio(struct bio *bp)
199 {
200 	struct bio *bp2;
201 
202 	bp2 = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
203 	if (bp2 != NULL) {
204 		bp2->bio_parent = bp;
205 		bp2->bio_cmd = bp->bio_cmd;
206 		/*
207 		 *  BIO_ORDERED flag may be used by disk drivers to enforce
208 		 *  ordering restrictions, so this flag needs to be cloned.
209 		 *  BIO_UNMAPPED and BIO_VLIST should be inherited, to properly
210 		 *  indicate which way the buffer is passed.
211 		 *  Other bio flags are not suitable for cloning.
212 		 */
213 		bp2->bio_flags = bp->bio_flags &
214 		    (BIO_ORDERED | BIO_UNMAPPED | BIO_VLIST);
215 		bp2->bio_length = bp->bio_length;
216 		bp2->bio_offset = bp->bio_offset;
217 		bp2->bio_data = bp->bio_data;
218 		bp2->bio_ma = bp->bio_ma;
219 		bp2->bio_ma_n = bp->bio_ma_n;
220 		bp2->bio_ma_offset = bp->bio_ma_offset;
221 		bp2->bio_attribute = bp->bio_attribute;
222 		if (bp->bio_cmd == BIO_ZONE)
223 			bcopy(&bp->bio_zone, &bp2->bio_zone,
224 			    sizeof(bp->bio_zone));
225 		/* Inherit classification info from the parent */
226 		bp2->bio_classifier1 = bp->bio_classifier1;
227 		bp2->bio_classifier2 = bp->bio_classifier2;
228 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
229 		bp2->bio_track_bp = bp->bio_track_bp;
230 #endif
231 		bp->bio_children++;
232 	}
233 #ifdef KTR
234 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
235 		struct stack st;
236 
237 		CTR2(KTR_GEOM, "g_clone_bio(%p): %p", bp, bp2);
238 		stack_save(&st);
239 		CTRSTACK(KTR_GEOM, &st, 3, 0);
240 	}
241 #endif
242 	return(bp2);
243 }
244 
245 struct bio *
246 g_duplicate_bio(struct bio *bp)
247 {
248 	struct bio *bp2;
249 
250 	bp2 = uma_zalloc(biozone, M_WAITOK | M_ZERO);
251 	bp2->bio_flags = bp->bio_flags & (BIO_UNMAPPED | BIO_VLIST);
252 	bp2->bio_parent = bp;
253 	bp2->bio_cmd = bp->bio_cmd;
254 	bp2->bio_length = bp->bio_length;
255 	bp2->bio_offset = bp->bio_offset;
256 	bp2->bio_data = bp->bio_data;
257 	bp2->bio_ma = bp->bio_ma;
258 	bp2->bio_ma_n = bp->bio_ma_n;
259 	bp2->bio_ma_offset = bp->bio_ma_offset;
260 	bp2->bio_attribute = bp->bio_attribute;
261 	bp->bio_children++;
262 #ifdef KTR
263 	if ((KTR_COMPILE & KTR_GEOM) && (ktr_mask & KTR_GEOM)) {
264 		struct stack st;
265 
266 		CTR2(KTR_GEOM, "g_duplicate_bio(%p): %p", bp, bp2);
267 		stack_save(&st);
268 		CTRSTACK(KTR_GEOM, &st, 3, 0);
269 	}
270 #endif
271 	return(bp2);
272 }
273 
274 void
275 g_reset_bio(struct bio *bp)
276 {
277 
278 	bzero(bp, sizeof(*bp));
279 }
280 
281 void
282 g_io_init()
283 {
284 
285 	g_bioq_init(&g_bio_run_down);
286 	g_bioq_init(&g_bio_run_up);
287 	biozone = uma_zcreate("g_bio", sizeof (struct bio),
288 	    NULL, NULL,
289 	    NULL, NULL,
290 	    0, 0);
291 }
292 
293 int
294 g_io_getattr(const char *attr, struct g_consumer *cp, int *len, void *ptr)
295 {
296 	struct bio *bp;
297 	int error;
298 
299 	g_trace(G_T_BIO, "bio_getattr(%s)", attr);
300 	bp = g_alloc_bio();
301 	bp->bio_cmd = BIO_GETATTR;
302 	bp->bio_done = NULL;
303 	bp->bio_attribute = attr;
304 	bp->bio_length = *len;
305 	bp->bio_data = ptr;
306 	g_io_request(bp, cp);
307 	error = biowait(bp, "ggetattr");
308 	*len = bp->bio_completed;
309 	g_destroy_bio(bp);
310 	return (error);
311 }
312 
313 int
314 g_io_zonecmd(struct disk_zone_args *zone_args, struct g_consumer *cp)
315 {
316 	struct bio *bp;
317 	int error;
318 
319 	g_trace(G_T_BIO, "bio_zone(%d)", zone_args->zone_cmd);
320 	bp = g_alloc_bio();
321 	bp->bio_cmd = BIO_ZONE;
322 	bp->bio_done = NULL;
323 	/*
324 	 * XXX KDM need to handle report zone data.
325 	 */
326 	bcopy(zone_args, &bp->bio_zone, sizeof(*zone_args));
327 	if (zone_args->zone_cmd == DISK_ZONE_REPORT_ZONES)
328 		bp->bio_length =
329 		    zone_args->zone_params.report.entries_allocated *
330 		    sizeof(struct disk_zone_rep_entry);
331 	else
332 		bp->bio_length = 0;
333 
334 	g_io_request(bp, cp);
335 	error = biowait(bp, "gzone");
336 	bcopy(&bp->bio_zone, zone_args, sizeof(*zone_args));
337 	g_destroy_bio(bp);
338 	return (error);
339 }
340 
341 int
342 g_io_flush(struct g_consumer *cp)
343 {
344 	struct bio *bp;
345 	int error;
346 
347 	g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name);
348 	bp = g_alloc_bio();
349 	bp->bio_cmd = BIO_FLUSH;
350 	bp->bio_flags |= BIO_ORDERED;
351 	bp->bio_done = NULL;
352 	bp->bio_attribute = NULL;
353 	bp->bio_offset = cp->provider->mediasize;
354 	bp->bio_length = 0;
355 	bp->bio_data = NULL;
356 	g_io_request(bp, cp);
357 	error = biowait(bp, "gflush");
358 	g_destroy_bio(bp);
359 	return (error);
360 }
361 
362 static int
363 g_io_check(struct bio *bp)
364 {
365 	struct g_consumer *cp;
366 	struct g_provider *pp;
367 	off_t excess;
368 	int error;
369 
370 	biotrack(bp, __func__);
371 
372 	cp = bp->bio_from;
373 	pp = bp->bio_to;
374 
375 	/* Fail if access counters dont allow the operation */
376 	switch(bp->bio_cmd) {
377 	case BIO_READ:
378 	case BIO_GETATTR:
379 		if (cp->acr == 0)
380 			return (EPERM);
381 		break;
382 	case BIO_WRITE:
383 	case BIO_DELETE:
384 	case BIO_FLUSH:
385 		if (cp->acw == 0)
386 			return (EPERM);
387 		break;
388 	case BIO_ZONE:
389 		if ((bp->bio_zone.zone_cmd == DISK_ZONE_REPORT_ZONES) ||
390 		    (bp->bio_zone.zone_cmd == DISK_ZONE_GET_PARAMS)) {
391 			if (cp->acr == 0)
392 				return (EPERM);
393 		} else if (cp->acw == 0)
394 			return (EPERM);
395 		break;
396 	default:
397 		return (EPERM);
398 	}
399 	/* if provider is marked for error, don't disturb. */
400 	if (pp->error)
401 		return (pp->error);
402 	if (cp->flags & G_CF_ORPHAN)
403 		return (ENXIO);
404 
405 	switch(bp->bio_cmd) {
406 	case BIO_READ:
407 	case BIO_WRITE:
408 	case BIO_DELETE:
409 		/* Zero sectorsize or mediasize is probably a lack of media. */
410 		if (pp->sectorsize == 0 || pp->mediasize == 0)
411 			return (ENXIO);
412 		/* Reject I/O not on sector boundary */
413 		if (bp->bio_offset % pp->sectorsize)
414 			return (EINVAL);
415 		/* Reject I/O not integral sector long */
416 		if (bp->bio_length % pp->sectorsize)
417 			return (EINVAL);
418 		/* Reject requests before or past the end of media. */
419 		if (bp->bio_offset < 0)
420 			return (EIO);
421 		if (bp->bio_offset > pp->mediasize)
422 			return (EIO);
423 
424 		/* Truncate requests to the end of providers media. */
425 		excess = bp->bio_offset + bp->bio_length;
426 		if (excess > bp->bio_to->mediasize) {
427 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
428 			    round_page(bp->bio_ma_offset +
429 			    bp->bio_length) / PAGE_SIZE == bp->bio_ma_n,
430 			    ("excess bio %p too short", bp));
431 			excess -= bp->bio_to->mediasize;
432 			bp->bio_length -= excess;
433 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
434 				bp->bio_ma_n = round_page(bp->bio_ma_offset +
435 				    bp->bio_length) / PAGE_SIZE;
436 			}
437 			if (excess > 0)
438 				CTR3(KTR_GEOM, "g_down truncated bio "
439 				    "%p provider %s by %d", bp,
440 				    bp->bio_to->name, excess);
441 		}
442 
443 		/* Deliver zero length transfers right here. */
444 		if (bp->bio_length == 0) {
445 			CTR2(KTR_GEOM, "g_down terminated 0-length "
446 			    "bp %p provider %s", bp, bp->bio_to->name);
447 			return (0);
448 		}
449 
450 		if ((bp->bio_flags & BIO_UNMAPPED) != 0 &&
451 		    (bp->bio_to->flags & G_PF_ACCEPT_UNMAPPED) == 0 &&
452 		    (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) {
453 			if ((error = g_io_transient_map_bio(bp)) >= 0)
454 				return (error);
455 		}
456 		break;
457 	default:
458 		break;
459 	}
460 	return (EJUSTRETURN);
461 }
462 
463 /*
464  * bio classification support.
465  *
466  * g_register_classifier() and g_unregister_classifier()
467  * are used to add/remove a classifier from the list.
468  * The list is protected using the g_bio_run_down lock,
469  * because the classifiers are called in this path.
470  *
471  * g_io_request() passes bio's that are not already classified
472  * (i.e. those with bio_classifier1 == NULL) to g_run_classifiers().
473  * Classifiers can store their result in the two fields
474  * bio_classifier1 and bio_classifier2.
475  * A classifier that updates one of the fields should
476  * return a non-zero value.
477  * If no classifier updates the field, g_run_classifiers() sets
478  * bio_classifier1 = BIO_NOTCLASSIFIED to avoid further calls.
479  */
480 
481 int
482 g_register_classifier(struct g_classifier_hook *hook)
483 {
484 
485 	g_bioq_lock(&g_bio_run_down);
486 	TAILQ_INSERT_TAIL(&g_classifier_tailq, hook, link);
487 	g_bioq_unlock(&g_bio_run_down);
488 
489 	return (0);
490 }
491 
492 void
493 g_unregister_classifier(struct g_classifier_hook *hook)
494 {
495 	struct g_classifier_hook *entry;
496 
497 	g_bioq_lock(&g_bio_run_down);
498 	TAILQ_FOREACH(entry, &g_classifier_tailq, link) {
499 		if (entry == hook) {
500 			TAILQ_REMOVE(&g_classifier_tailq, hook, link);
501 			break;
502 		}
503 	}
504 	g_bioq_unlock(&g_bio_run_down);
505 }
506 
507 static void
508 g_run_classifiers(struct bio *bp)
509 {
510 	struct g_classifier_hook *hook;
511 	int classified = 0;
512 
513 	biotrack(bp, __func__);
514 
515 	TAILQ_FOREACH(hook, &g_classifier_tailq, link)
516 		classified |= hook->func(hook->arg, bp);
517 
518 	if (!classified)
519 		bp->bio_classifier1 = BIO_NOTCLASSIFIED;
520 }
521 
522 void
523 g_io_request(struct bio *bp, struct g_consumer *cp)
524 {
525 	struct g_provider *pp;
526 	struct mtx *mtxp;
527 	int direct, error, first;
528 	uint8_t cmd;
529 
530 	biotrack(bp, __func__);
531 
532 	KASSERT(cp != NULL, ("NULL cp in g_io_request"));
533 	KASSERT(bp != NULL, ("NULL bp in g_io_request"));
534 	pp = cp->provider;
535 	KASSERT(pp != NULL, ("consumer not attached in g_io_request"));
536 #ifdef DIAGNOSTIC
537 	KASSERT(bp->bio_driver1 == NULL,
538 	    ("bio_driver1 used by the consumer (geom %s)", cp->geom->name));
539 	KASSERT(bp->bio_driver2 == NULL,
540 	    ("bio_driver2 used by the consumer (geom %s)", cp->geom->name));
541 	KASSERT(bp->bio_pflags == 0,
542 	    ("bio_pflags used by the consumer (geom %s)", cp->geom->name));
543 	/*
544 	 * Remember consumer's private fields, so we can detect if they were
545 	 * modified by the provider.
546 	 */
547 	bp->_bio_caller1 = bp->bio_caller1;
548 	bp->_bio_caller2 = bp->bio_caller2;
549 	bp->_bio_cflags = bp->bio_cflags;
550 #endif
551 
552 	cmd = bp->bio_cmd;
553 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_GETATTR) {
554 		KASSERT(bp->bio_data != NULL,
555 		    ("NULL bp->data in g_io_request(cmd=%hu)", bp->bio_cmd));
556 	}
557 	if (cmd == BIO_DELETE || cmd == BIO_FLUSH) {
558 		KASSERT(bp->bio_data == NULL,
559 		    ("non-NULL bp->data in g_io_request(cmd=%hu)",
560 		    bp->bio_cmd));
561 	}
562 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_DELETE) {
563 		KASSERT(bp->bio_offset % cp->provider->sectorsize == 0,
564 		    ("wrong offset %jd for sectorsize %u",
565 		    bp->bio_offset, cp->provider->sectorsize));
566 		KASSERT(bp->bio_length % cp->provider->sectorsize == 0,
567 		    ("wrong length %jd for sectorsize %u",
568 		    bp->bio_length, cp->provider->sectorsize));
569 	}
570 
571 	g_trace(G_T_BIO, "bio_request(%p) from %p(%s) to %p(%s) cmd %d",
572 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd);
573 
574 	bp->bio_from = cp;
575 	bp->bio_to = pp;
576 	bp->bio_error = 0;
577 	bp->bio_completed = 0;
578 
579 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
580 	    ("Bio already on queue bp=%p", bp));
581 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
582 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
583 		binuptime(&bp->bio_t0);
584 	else
585 		getbinuptime(&bp->bio_t0);
586 
587 #ifdef GET_STACK_USAGE
588 	direct = (cp->flags & G_CF_DIRECT_SEND) != 0 &&
589 	    (pp->flags & G_PF_DIRECT_RECEIVE) != 0 &&
590 	    !g_is_geom_thread(curthread) &&
591 	    ((pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ||
592 	    (bp->bio_flags & BIO_UNMAPPED) == 0 || THREAD_CAN_SLEEP()) &&
593 	    pace == 0;
594 	if (direct) {
595 		/* Block direct execution if less then half of stack left. */
596 		size_t	st, su;
597 		GET_STACK_USAGE(st, su);
598 		if (su * 2 > st)
599 			direct = 0;
600 	}
601 #else
602 	direct = 0;
603 #endif
604 
605 	if (!TAILQ_EMPTY(&g_classifier_tailq) && !bp->bio_classifier1) {
606 		g_bioq_lock(&g_bio_run_down);
607 		g_run_classifiers(bp);
608 		g_bioq_unlock(&g_bio_run_down);
609 	}
610 
611 	/*
612 	 * The statistics collection is lockless, as such, but we
613 	 * can not update one instance of the statistics from more
614 	 * than one thread at a time, so grab the lock first.
615 	 */
616 	mtxp = mtx_pool_find(mtxpool_sleep, pp);
617 	mtx_lock(mtxp);
618 	if (g_collectstats & G_STATS_PROVIDERS)
619 		devstat_start_transaction(pp->stat, &bp->bio_t0);
620 	if (g_collectstats & G_STATS_CONSUMERS)
621 		devstat_start_transaction(cp->stat, &bp->bio_t0);
622 	pp->nstart++;
623 	cp->nstart++;
624 	mtx_unlock(mtxp);
625 
626 	if (direct) {
627 		error = g_io_check(bp);
628 		if (error >= 0) {
629 			CTR3(KTR_GEOM, "g_io_request g_io_check on bp %p "
630 			    "provider %s returned %d", bp, bp->bio_to->name,
631 			    error);
632 			g_io_deliver(bp, error);
633 			return;
634 		}
635 		bp->bio_to->geom->start(bp);
636 	} else {
637 		g_bioq_lock(&g_bio_run_down);
638 		first = TAILQ_EMPTY(&g_bio_run_down.bio_queue);
639 		TAILQ_INSERT_TAIL(&g_bio_run_down.bio_queue, bp, bio_queue);
640 		bp->bio_flags |= BIO_ONQUEUE;
641 		g_bio_run_down.bio_queue_length++;
642 		g_bioq_unlock(&g_bio_run_down);
643 		/* Pass it on down. */
644 		if (first)
645 			wakeup(&g_wait_down);
646 	}
647 }
648 
649 void
650 g_io_deliver(struct bio *bp, int error)
651 {
652 	struct bintime now;
653 	struct g_consumer *cp;
654 	struct g_provider *pp;
655 	struct mtx *mtxp;
656 	int direct, first;
657 
658 	biotrack(bp, __func__);
659 
660 	KASSERT(bp != NULL, ("NULL bp in g_io_deliver"));
661 	pp = bp->bio_to;
662 	KASSERT(pp != NULL, ("NULL bio_to in g_io_deliver"));
663 	cp = bp->bio_from;
664 	if (cp == NULL) {
665 		bp->bio_error = error;
666 		bp->bio_done(bp);
667 		return;
668 	}
669 	KASSERT(cp != NULL, ("NULL bio_from in g_io_deliver"));
670 	KASSERT(cp->geom != NULL, ("NULL bio_from->geom in g_io_deliver"));
671 #ifdef DIAGNOSTIC
672 	/*
673 	 * Some classes - GJournal in particular - can modify bio's
674 	 * private fields while the bio is in transit; G_GEOM_VOLATILE_BIO
675 	 * flag means it's an expected behaviour for that particular geom.
676 	 */
677 	if ((cp->geom->flags & G_GEOM_VOLATILE_BIO) == 0) {
678 		KASSERT(bp->bio_caller1 == bp->_bio_caller1,
679 		    ("bio_caller1 used by the provider %s", pp->name));
680 		KASSERT(bp->bio_caller2 == bp->_bio_caller2,
681 		    ("bio_caller2 used by the provider %s", pp->name));
682 		KASSERT(bp->bio_cflags == bp->_bio_cflags,
683 		    ("bio_cflags used by the provider %s", pp->name));
684 	}
685 #endif
686 	KASSERT(bp->bio_completed >= 0, ("bio_completed can't be less than 0"));
687 	KASSERT(bp->bio_completed <= bp->bio_length,
688 	    ("bio_completed can't be greater than bio_length"));
689 
690 	g_trace(G_T_BIO,
691 "g_io_deliver(%p) from %p(%s) to %p(%s) cmd %d error %d off %jd len %jd",
692 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd, error,
693 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
694 
695 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
696 	    ("Bio already on queue bp=%p", bp));
697 
698 	/*
699 	 * XXX: next two doesn't belong here
700 	 */
701 	bp->bio_bcount = bp->bio_length;
702 	bp->bio_resid = bp->bio_bcount - bp->bio_completed;
703 
704 #ifdef GET_STACK_USAGE
705 	direct = (pp->flags & G_PF_DIRECT_SEND) &&
706 		 (cp->flags & G_CF_DIRECT_RECEIVE) &&
707 		 !g_is_geom_thread(curthread);
708 	if (direct) {
709 		/* Block direct execution if less then half of stack left. */
710 		size_t	st, su;
711 		GET_STACK_USAGE(st, su);
712 		if (su * 2 > st)
713 			direct = 0;
714 	}
715 #else
716 	direct = 0;
717 #endif
718 
719 	/*
720 	 * The statistics collection is lockless, as such, but we
721 	 * can not update one instance of the statistics from more
722 	 * than one thread at a time, so grab the lock first.
723 	 */
724 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
725 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
726 		binuptime(&now);
727 	mtxp = mtx_pool_find(mtxpool_sleep, cp);
728 	mtx_lock(mtxp);
729 	if (g_collectstats & G_STATS_PROVIDERS)
730 		devstat_end_transaction_bio_bt(pp->stat, bp, &now);
731 	if (g_collectstats & G_STATS_CONSUMERS)
732 		devstat_end_transaction_bio_bt(cp->stat, bp, &now);
733 	cp->nend++;
734 	pp->nend++;
735 	mtx_unlock(mtxp);
736 
737 	if (error != ENOMEM) {
738 		bp->bio_error = error;
739 		if (direct) {
740 			biodone(bp);
741 		} else {
742 			g_bioq_lock(&g_bio_run_up);
743 			first = TAILQ_EMPTY(&g_bio_run_up.bio_queue);
744 			TAILQ_INSERT_TAIL(&g_bio_run_up.bio_queue, bp, bio_queue);
745 			bp->bio_flags |= BIO_ONQUEUE;
746 			g_bio_run_up.bio_queue_length++;
747 			g_bioq_unlock(&g_bio_run_up);
748 			if (first)
749 				wakeup(&g_wait_up);
750 		}
751 		return;
752 	}
753 
754 	if (bootverbose)
755 		printf("ENOMEM %p on %p(%s)\n", bp, pp, pp->name);
756 	bp->bio_children = 0;
757 	bp->bio_inbed = 0;
758 	bp->bio_driver1 = NULL;
759 	bp->bio_driver2 = NULL;
760 	bp->bio_pflags = 0;
761 	g_io_request(bp, cp);
762 	pace = 1;
763 	return;
764 }
765 
766 SYSCTL_DECL(_kern_geom);
767 
768 static long transient_maps;
769 SYSCTL_LONG(_kern_geom, OID_AUTO, transient_maps, CTLFLAG_RD,
770     &transient_maps, 0,
771     "Total count of the transient mapping requests");
772 u_int transient_map_retries = 10;
773 SYSCTL_UINT(_kern_geom, OID_AUTO, transient_map_retries, CTLFLAG_RW,
774     &transient_map_retries, 0,
775     "Max count of retries used before giving up on creating transient map");
776 int transient_map_hard_failures;
777 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_hard_failures, CTLFLAG_RD,
778     &transient_map_hard_failures, 0,
779     "Failures to establish the transient mapping due to retry attempts "
780     "exhausted");
781 int transient_map_soft_failures;
782 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_soft_failures, CTLFLAG_RD,
783     &transient_map_soft_failures, 0,
784     "Count of retried failures to establish the transient mapping");
785 int inflight_transient_maps;
786 SYSCTL_INT(_kern_geom, OID_AUTO, inflight_transient_maps, CTLFLAG_RD,
787     &inflight_transient_maps, 0,
788     "Current count of the active transient maps");
789 
790 static int
791 g_io_transient_map_bio(struct bio *bp)
792 {
793 	vm_offset_t addr;
794 	long size;
795 	u_int retried;
796 
797 	KASSERT(unmapped_buf_allowed, ("unmapped disabled"));
798 
799 	size = round_page(bp->bio_ma_offset + bp->bio_length);
800 	KASSERT(size / PAGE_SIZE == bp->bio_ma_n, ("Bio too short %p", bp));
801 	addr = 0;
802 	retried = 0;
803 	atomic_add_long(&transient_maps, 1);
804 retry:
805 	if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) {
806 		if (transient_map_retries != 0 &&
807 		    retried >= transient_map_retries) {
808 			CTR2(KTR_GEOM, "g_down cannot map bp %p provider %s",
809 			    bp, bp->bio_to->name);
810 			atomic_add_int(&transient_map_hard_failures, 1);
811 			return (EDEADLK/* XXXKIB */);
812 		} else {
813 			/*
814 			 * Naive attempt to quisce the I/O to get more
815 			 * in-flight requests completed and defragment
816 			 * the transient_arena.
817 			 */
818 			CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d",
819 			    bp, bp->bio_to->name, retried);
820 			pause("g_d_tra", hz / 10);
821 			retried++;
822 			atomic_add_int(&transient_map_soft_failures, 1);
823 			goto retry;
824 		}
825 	}
826 	atomic_add_int(&inflight_transient_maps, 1);
827 	pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size));
828 	bp->bio_data = (caddr_t)addr + bp->bio_ma_offset;
829 	bp->bio_flags |= BIO_TRANSIENT_MAPPING;
830 	bp->bio_flags &= ~BIO_UNMAPPED;
831 	return (EJUSTRETURN);
832 }
833 
834 void
835 g_io_schedule_down(struct thread *tp __unused)
836 {
837 	struct bio *bp;
838 	int error;
839 
840 	for(;;) {
841 		g_bioq_lock(&g_bio_run_down);
842 		bp = g_bioq_first(&g_bio_run_down);
843 		if (bp == NULL) {
844 			CTR0(KTR_GEOM, "g_down going to sleep");
845 			msleep(&g_wait_down, &g_bio_run_down.bio_queue_lock,
846 			    PRIBIO | PDROP, "-", 0);
847 			continue;
848 		}
849 		CTR0(KTR_GEOM, "g_down has work to do");
850 		g_bioq_unlock(&g_bio_run_down);
851 		biotrack(bp, __func__);
852 		if (pace != 0) {
853 			/*
854 			 * There has been at least one memory allocation
855 			 * failure since the last I/O completed. Pause 1ms to
856 			 * give the system a chance to free up memory. We only
857 			 * do this once because a large number of allocations
858 			 * can fail in the direct dispatch case and there's no
859 			 * relationship between the number of these failures and
860 			 * the length of the outage. If there's still an outage,
861 			 * we'll pause again and again until it's
862 			 * resolved. Older versions paused longer and once per
863 			 * allocation failure. This was OK for a single threaded
864 			 * g_down, but with direct dispatch would lead to max of
865 			 * 10 IOPs for minutes at a time when transient memory
866 			 * issues prevented allocation for a batch of requests
867 			 * from the upper layers.
868 			 *
869 			 * XXX This pacing is really lame. It needs to be solved
870 			 * by other methods. This is OK only because the worst
871 			 * case scenario is so rare. In the worst case scenario
872 			 * all memory is tied up waiting for I/O to complete
873 			 * which can never happen since we can't allocate bios
874 			 * for that I/O.
875 			 */
876 			CTR0(KTR_GEOM, "g_down pacing self");
877 			pause("g_down", min(hz/1000, 1));
878 			pace = 0;
879 		}
880 		CTR2(KTR_GEOM, "g_down processing bp %p provider %s", bp,
881 		    bp->bio_to->name);
882 		error = g_io_check(bp);
883 		if (error >= 0) {
884 			CTR3(KTR_GEOM, "g_down g_io_check on bp %p provider "
885 			    "%s returned %d", bp, bp->bio_to->name, error);
886 			g_io_deliver(bp, error);
887 			continue;
888 		}
889 		THREAD_NO_SLEEPING();
890 		CTR4(KTR_GEOM, "g_down starting bp %p provider %s off %ld "
891 		    "len %ld", bp, bp->bio_to->name, bp->bio_offset,
892 		    bp->bio_length);
893 		bp->bio_to->geom->start(bp);
894 		THREAD_SLEEPING_OK();
895 	}
896 }
897 
898 void
899 g_io_schedule_up(struct thread *tp __unused)
900 {
901 	struct bio *bp;
902 
903 	for(;;) {
904 		g_bioq_lock(&g_bio_run_up);
905 		bp = g_bioq_first(&g_bio_run_up);
906 		if (bp == NULL) {
907 			CTR0(KTR_GEOM, "g_up going to sleep");
908 			msleep(&g_wait_up, &g_bio_run_up.bio_queue_lock,
909 			    PRIBIO | PDROP, "-", 0);
910 			continue;
911 		}
912 		g_bioq_unlock(&g_bio_run_up);
913 		THREAD_NO_SLEEPING();
914 		CTR4(KTR_GEOM, "g_up biodone bp %p provider %s off "
915 		    "%jd len %ld", bp, bp->bio_to->name,
916 		    bp->bio_offset, bp->bio_length);
917 		biodone(bp);
918 		THREAD_SLEEPING_OK();
919 	}
920 }
921 
922 void *
923 g_read_data(struct g_consumer *cp, off_t offset, off_t length, int *error)
924 {
925 	struct bio *bp;
926 	void *ptr;
927 	int errorc;
928 
929 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
930 	    length <= MAXPHYS, ("g_read_data(): invalid length %jd",
931 	    (intmax_t)length));
932 
933 	bp = g_alloc_bio();
934 	bp->bio_cmd = BIO_READ;
935 	bp->bio_done = NULL;
936 	bp->bio_offset = offset;
937 	bp->bio_length = length;
938 	ptr = g_malloc(length, M_WAITOK);
939 	bp->bio_data = ptr;
940 	g_io_request(bp, cp);
941 	errorc = biowait(bp, "gread");
942 	if (error != NULL)
943 		*error = errorc;
944 	g_destroy_bio(bp);
945 	if (errorc) {
946 		g_free(ptr);
947 		ptr = NULL;
948 	}
949 	return (ptr);
950 }
951 
952 /*
953  * A read function for use by ffs_sbget when used by GEOM-layer routines.
954  */
955 int
956 g_use_g_read_data(void *devfd, off_t loc, void **bufp, int size)
957 {
958 	struct g_consumer *cp;
959 
960 	KASSERT(*bufp == NULL,
961 	    ("g_use_g_read_data: non-NULL *bufp %p\n", *bufp));
962 
963 	cp = (struct g_consumer *)devfd;
964 	/*
965 	 * Take care not to issue an invalid I/O request. The offset of
966 	 * the superblock candidate must be multiples of the provider's
967 	 * sector size, otherwise an FFS can't exist on the provider
968 	 * anyway.
969 	 */
970 	if (loc % cp->provider->sectorsize != 0)
971 		return (ENOENT);
972 	*bufp = g_read_data(cp, loc, size, NULL);
973 	if (*bufp == NULL)
974 		return (ENOENT);
975 	return (0);
976 }
977 
978 int
979 g_write_data(struct g_consumer *cp, off_t offset, void *ptr, off_t length)
980 {
981 	struct bio *bp;
982 	int error;
983 
984 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
985 	    length <= MAXPHYS, ("g_write_data(): invalid length %jd",
986 	    (intmax_t)length));
987 
988 	bp = g_alloc_bio();
989 	bp->bio_cmd = BIO_WRITE;
990 	bp->bio_done = NULL;
991 	bp->bio_offset = offset;
992 	bp->bio_length = length;
993 	bp->bio_data = ptr;
994 	g_io_request(bp, cp);
995 	error = biowait(bp, "gwrite");
996 	g_destroy_bio(bp);
997 	return (error);
998 }
999 
1000 /*
1001  * A write function for use by ffs_sbput when used by GEOM-layer routines.
1002  */
1003 int
1004 g_use_g_write_data(void *devfd, off_t loc, void *buf, int size)
1005 {
1006 
1007 	return (g_write_data((struct g_consumer *)devfd, loc, buf, size));
1008 }
1009 
1010 int
1011 g_delete_data(struct g_consumer *cp, off_t offset, off_t length)
1012 {
1013 	struct bio *bp;
1014 	int error;
1015 
1016 	KASSERT(length > 0 && length >= cp->provider->sectorsize,
1017 	    ("g_delete_data(): invalid length %jd", (intmax_t)length));
1018 
1019 	bp = g_alloc_bio();
1020 	bp->bio_cmd = BIO_DELETE;
1021 	bp->bio_done = NULL;
1022 	bp->bio_offset = offset;
1023 	bp->bio_length = length;
1024 	bp->bio_data = NULL;
1025 	g_io_request(bp, cp);
1026 	error = biowait(bp, "gdelete");
1027 	g_destroy_bio(bp);
1028 	return (error);
1029 }
1030 
1031 void
1032 g_print_bio(struct bio *bp)
1033 {
1034 	const char *pname, *cmd = NULL;
1035 
1036 	if (bp->bio_to != NULL)
1037 		pname = bp->bio_to->name;
1038 	else
1039 		pname = "[unknown]";
1040 
1041 	switch (bp->bio_cmd) {
1042 	case BIO_GETATTR:
1043 		cmd = "GETATTR";
1044 		printf("%s[%s(attr=%s)]", pname, cmd, bp->bio_attribute);
1045 		return;
1046 	case BIO_FLUSH:
1047 		cmd = "FLUSH";
1048 		printf("%s[%s]", pname, cmd);
1049 		return;
1050 	case BIO_ZONE: {
1051 		char *subcmd = NULL;
1052 		cmd = "ZONE";
1053 		switch (bp->bio_zone.zone_cmd) {
1054 		case DISK_ZONE_OPEN:
1055 			subcmd = "OPEN";
1056 			break;
1057 		case DISK_ZONE_CLOSE:
1058 			subcmd = "CLOSE";
1059 			break;
1060 		case DISK_ZONE_FINISH:
1061 			subcmd = "FINISH";
1062 			break;
1063 		case DISK_ZONE_RWP:
1064 			subcmd = "RWP";
1065 			break;
1066 		case DISK_ZONE_REPORT_ZONES:
1067 			subcmd = "REPORT ZONES";
1068 			break;
1069 		case DISK_ZONE_GET_PARAMS:
1070 			subcmd = "GET PARAMS";
1071 			break;
1072 		default:
1073 			subcmd = "UNKNOWN";
1074 			break;
1075 		}
1076 		printf("%s[%s,%s]", pname, cmd, subcmd);
1077 		return;
1078 	}
1079 	case BIO_READ:
1080 		cmd = "READ";
1081 		break;
1082 	case BIO_WRITE:
1083 		cmd = "WRITE";
1084 		break;
1085 	case BIO_DELETE:
1086 		cmd = "DELETE";
1087 		break;
1088 	default:
1089 		cmd = "UNKNOWN";
1090 		printf("%s[%s()]", pname, cmd);
1091 		return;
1092 	}
1093 	printf("%s[%s(offset=%jd, length=%jd)]", pname, cmd,
1094 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
1095 }
1096