xref: /freebsd/sys/geom/geom_io.c (revision 43faedc1339a9624c7acedb7f3e5624e64da5b99)
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_io_init()
269 {
270 
271 	g_bioq_init(&g_bio_run_down);
272 	g_bioq_init(&g_bio_run_up);
273 	g_bioq_init(&g_bio_run_task);
274 	biozone = uma_zcreate("g_bio", sizeof (struct bio),
275 	    NULL, NULL,
276 	    NULL, NULL,
277 	    0, 0);
278 }
279 
280 int
281 g_io_getattr(const char *attr, struct g_consumer *cp, int *len, void *ptr)
282 {
283 	struct bio *bp;
284 	int error;
285 
286 	g_trace(G_T_BIO, "bio_getattr(%s)", attr);
287 	bp = g_alloc_bio();
288 	bp->bio_cmd = BIO_GETATTR;
289 	bp->bio_done = NULL;
290 	bp->bio_attribute = attr;
291 	bp->bio_length = *len;
292 	bp->bio_data = ptr;
293 	g_io_request(bp, cp);
294 	error = biowait(bp, "ggetattr");
295 	*len = bp->bio_completed;
296 	g_destroy_bio(bp);
297 	return (error);
298 }
299 
300 int
301 g_io_flush(struct g_consumer *cp)
302 {
303 	struct bio *bp;
304 	int error;
305 
306 	g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name);
307 	bp = g_alloc_bio();
308 	bp->bio_cmd = BIO_FLUSH;
309 	bp->bio_flags |= BIO_ORDERED;
310 	bp->bio_done = NULL;
311 	bp->bio_attribute = NULL;
312 	bp->bio_offset = cp->provider->mediasize;
313 	bp->bio_length = 0;
314 	bp->bio_data = NULL;
315 	g_io_request(bp, cp);
316 	error = biowait(bp, "gflush");
317 	g_destroy_bio(bp);
318 	return (error);
319 }
320 
321 static int
322 g_io_check(struct bio *bp)
323 {
324 	struct g_consumer *cp;
325 	struct g_provider *pp;
326 	off_t excess;
327 	int error;
328 
329 	cp = bp->bio_from;
330 	pp = bp->bio_to;
331 
332 	/* Fail if access counters dont allow the operation */
333 	switch(bp->bio_cmd) {
334 	case BIO_READ:
335 	case BIO_GETATTR:
336 		if (cp->acr == 0)
337 			return (EPERM);
338 		break;
339 	case BIO_WRITE:
340 	case BIO_DELETE:
341 	case BIO_FLUSH:
342 		if (cp->acw == 0)
343 			return (EPERM);
344 		break;
345 	default:
346 		return (EPERM);
347 	}
348 	/* if provider is marked for error, don't disturb. */
349 	if (pp->error)
350 		return (pp->error);
351 	if (cp->flags & G_CF_ORPHAN)
352 		return (ENXIO);
353 
354 	switch(bp->bio_cmd) {
355 	case BIO_READ:
356 	case BIO_WRITE:
357 	case BIO_DELETE:
358 		/* Zero sectorsize or mediasize is probably a lack of media. */
359 		if (pp->sectorsize == 0 || pp->mediasize == 0)
360 			return (ENXIO);
361 		/* Reject I/O not on sector boundary */
362 		if (bp->bio_offset % pp->sectorsize)
363 			return (EINVAL);
364 		/* Reject I/O not integral sector long */
365 		if (bp->bio_length % pp->sectorsize)
366 			return (EINVAL);
367 		/* Reject requests before or past the end of media. */
368 		if (bp->bio_offset < 0)
369 			return (EIO);
370 		if (bp->bio_offset > pp->mediasize)
371 			return (EIO);
372 
373 		/* Truncate requests to the end of providers media. */
374 		excess = bp->bio_offset + bp->bio_length;
375 		if (excess > bp->bio_to->mediasize) {
376 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
377 			    round_page(bp->bio_ma_offset +
378 			    bp->bio_length) / PAGE_SIZE == bp->bio_ma_n,
379 			    ("excess bio %p too short", bp));
380 			excess -= bp->bio_to->mediasize;
381 			bp->bio_length -= excess;
382 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
383 				bp->bio_ma_n = round_page(bp->bio_ma_offset +
384 				    bp->bio_length) / PAGE_SIZE;
385 			}
386 			if (excess > 0)
387 				CTR3(KTR_GEOM, "g_down truncated bio "
388 				    "%p provider %s by %d", bp,
389 				    bp->bio_to->name, excess);
390 		}
391 
392 		/* Deliver zero length transfers right here. */
393 		if (bp->bio_length == 0) {
394 			CTR2(KTR_GEOM, "g_down terminated 0-length "
395 			    "bp %p provider %s", bp, bp->bio_to->name);
396 			return (0);
397 		}
398 
399 		if ((bp->bio_flags & BIO_UNMAPPED) != 0 &&
400 		    (bp->bio_to->flags & G_PF_ACCEPT_UNMAPPED) == 0 &&
401 		    (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) {
402 			if ((error = g_io_transient_map_bio(bp)) >= 0)
403 				return (error);
404 		}
405 		break;
406 	default:
407 		break;
408 	}
409 	return (EJUSTRETURN);
410 }
411 
412 /*
413  * bio classification support.
414  *
415  * g_register_classifier() and g_unregister_classifier()
416  * are used to add/remove a classifier from the list.
417  * The list is protected using the g_bio_run_down lock,
418  * because the classifiers are called in this path.
419  *
420  * g_io_request() passes bio's that are not already classified
421  * (i.e. those with bio_classifier1 == NULL) to g_run_classifiers().
422  * Classifiers can store their result in the two fields
423  * bio_classifier1 and bio_classifier2.
424  * A classifier that updates one of the fields should
425  * return a non-zero value.
426  * If no classifier updates the field, g_run_classifiers() sets
427  * bio_classifier1 = BIO_NOTCLASSIFIED to avoid further calls.
428  */
429 
430 int
431 g_register_classifier(struct g_classifier_hook *hook)
432 {
433 
434 	g_bioq_lock(&g_bio_run_down);
435 	TAILQ_INSERT_TAIL(&g_classifier_tailq, hook, link);
436 	g_bioq_unlock(&g_bio_run_down);
437 
438 	return (0);
439 }
440 
441 void
442 g_unregister_classifier(struct g_classifier_hook *hook)
443 {
444 	struct g_classifier_hook *entry;
445 
446 	g_bioq_lock(&g_bio_run_down);
447 	TAILQ_FOREACH(entry, &g_classifier_tailq, link) {
448 		if (entry == hook) {
449 			TAILQ_REMOVE(&g_classifier_tailq, hook, link);
450 			break;
451 		}
452 	}
453 	g_bioq_unlock(&g_bio_run_down);
454 }
455 
456 static void
457 g_run_classifiers(struct bio *bp)
458 {
459 	struct g_classifier_hook *hook;
460 	int classified = 0;
461 
462 	TAILQ_FOREACH(hook, &g_classifier_tailq, link)
463 		classified |= hook->func(hook->arg, bp);
464 
465 	if (!classified)
466 		bp->bio_classifier1 = BIO_NOTCLASSIFIED;
467 }
468 
469 void
470 g_io_request(struct bio *bp, struct g_consumer *cp)
471 {
472 	struct g_provider *pp;
473 	struct mtx *mtxp;
474 	int direct, error, first;
475 
476 	KASSERT(cp != NULL, ("NULL cp in g_io_request"));
477 	KASSERT(bp != NULL, ("NULL bp in g_io_request"));
478 	pp = cp->provider;
479 	KASSERT(pp != NULL, ("consumer not attached in g_io_request"));
480 #ifdef DIAGNOSTIC
481 	KASSERT(bp->bio_driver1 == NULL,
482 	    ("bio_driver1 used by the consumer (geom %s)", cp->geom->name));
483 	KASSERT(bp->bio_driver2 == NULL,
484 	    ("bio_driver2 used by the consumer (geom %s)", cp->geom->name));
485 	KASSERT(bp->bio_pflags == 0,
486 	    ("bio_pflags used by the consumer (geom %s)", cp->geom->name));
487 	/*
488 	 * Remember consumer's private fields, so we can detect if they were
489 	 * modified by the provider.
490 	 */
491 	bp->_bio_caller1 = bp->bio_caller1;
492 	bp->_bio_caller2 = bp->bio_caller2;
493 	bp->_bio_cflags = bp->bio_cflags;
494 #endif
495 
496 	if (bp->bio_cmd & (BIO_READ|BIO_WRITE|BIO_GETATTR)) {
497 		KASSERT(bp->bio_data != NULL,
498 		    ("NULL bp->data in g_io_request(cmd=%hhu)", bp->bio_cmd));
499 	}
500 	if (bp->bio_cmd & (BIO_DELETE|BIO_FLUSH)) {
501 		KASSERT(bp->bio_data == NULL,
502 		    ("non-NULL bp->data in g_io_request(cmd=%hhu)",
503 		    bp->bio_cmd));
504 	}
505 	if (bp->bio_cmd & (BIO_READ|BIO_WRITE|BIO_DELETE)) {
506 		KASSERT(bp->bio_offset % cp->provider->sectorsize == 0,
507 		    ("wrong offset %jd for sectorsize %u",
508 		    bp->bio_offset, cp->provider->sectorsize));
509 		KASSERT(bp->bio_length % cp->provider->sectorsize == 0,
510 		    ("wrong length %jd for sectorsize %u",
511 		    bp->bio_length, cp->provider->sectorsize));
512 	}
513 
514 	g_trace(G_T_BIO, "bio_request(%p) from %p(%s) to %p(%s) cmd %d",
515 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd);
516 
517 	bp->bio_from = cp;
518 	bp->bio_to = pp;
519 	bp->bio_error = 0;
520 	bp->bio_completed = 0;
521 
522 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
523 	    ("Bio already on queue bp=%p", bp));
524 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
525 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
526 		binuptime(&bp->bio_t0);
527 	else
528 		getbinuptime(&bp->bio_t0);
529 
530 #ifdef GET_STACK_USAGE
531 	direct = (cp->flags & G_CF_DIRECT_SEND) != 0 &&
532 	    (pp->flags & G_PF_DIRECT_RECEIVE) != 0 &&
533 	    !g_is_geom_thread(curthread) &&
534 	    ((pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ||
535 	    (bp->bio_flags & BIO_UNMAPPED) == 0 || THREAD_CAN_SLEEP()) &&
536 	    pace == 0;
537 	if (direct) {
538 		/* Block direct execution if less then half of stack left. */
539 		size_t	st, su;
540 		GET_STACK_USAGE(st, su);
541 		if (su * 2 > st)
542 			direct = 0;
543 	}
544 #else
545 	direct = 0;
546 #endif
547 
548 	if (!TAILQ_EMPTY(&g_classifier_tailq) && !bp->bio_classifier1) {
549 		g_bioq_lock(&g_bio_run_down);
550 		g_run_classifiers(bp);
551 		g_bioq_unlock(&g_bio_run_down);
552 	}
553 
554 	/*
555 	 * The statistics collection is lockless, as such, but we
556 	 * can not update one instance of the statistics from more
557 	 * than one thread at a time, so grab the lock first.
558 	 */
559 	mtxp = mtx_pool_find(mtxpool_sleep, pp);
560 	mtx_lock(mtxp);
561 	if (g_collectstats & G_STATS_PROVIDERS)
562 		devstat_start_transaction(pp->stat, &bp->bio_t0);
563 	if (g_collectstats & G_STATS_CONSUMERS)
564 		devstat_start_transaction(cp->stat, &bp->bio_t0);
565 	pp->nstart++;
566 	cp->nstart++;
567 	mtx_unlock(mtxp);
568 
569 	if (direct) {
570 		error = g_io_check(bp);
571 		if (error >= 0) {
572 			CTR3(KTR_GEOM, "g_io_request g_io_check on bp %p "
573 			    "provider %s returned %d", bp, bp->bio_to->name,
574 			    error);
575 			g_io_deliver(bp, error);
576 			return;
577 		}
578 		bp->bio_to->geom->start(bp);
579 	} else {
580 		g_bioq_lock(&g_bio_run_down);
581 		first = TAILQ_EMPTY(&g_bio_run_down.bio_queue);
582 		TAILQ_INSERT_TAIL(&g_bio_run_down.bio_queue, bp, bio_queue);
583 		bp->bio_flags |= BIO_ONQUEUE;
584 		g_bio_run_down.bio_queue_length++;
585 		g_bioq_unlock(&g_bio_run_down);
586 		/* Pass it on down. */
587 		if (first)
588 			wakeup(&g_wait_down);
589 	}
590 }
591 
592 void
593 g_io_deliver(struct bio *bp, int error)
594 {
595 	struct bintime now;
596 	struct g_consumer *cp;
597 	struct g_provider *pp;
598 	struct mtx *mtxp;
599 	int direct, first;
600 
601 	KASSERT(bp != NULL, ("NULL bp in g_io_deliver"));
602 	pp = bp->bio_to;
603 	KASSERT(pp != NULL, ("NULL bio_to in g_io_deliver"));
604 	cp = bp->bio_from;
605 	if (cp == NULL) {
606 		bp->bio_error = error;
607 		bp->bio_done(bp);
608 		return;
609 	}
610 	KASSERT(cp != NULL, ("NULL bio_from in g_io_deliver"));
611 	KASSERT(cp->geom != NULL, ("NULL bio_from->geom in g_io_deliver"));
612 #ifdef DIAGNOSTIC
613 	/*
614 	 * Some classes - GJournal in particular - can modify bio's
615 	 * private fields while the bio is in transit; G_GEOM_VOLATILE_BIO
616 	 * flag means it's an expected behaviour for that particular geom.
617 	 */
618 	if ((cp->geom->flags & G_GEOM_VOLATILE_BIO) == 0) {
619 		KASSERT(bp->bio_caller1 == bp->_bio_caller1,
620 		    ("bio_caller1 used by the provider %s", pp->name));
621 		KASSERT(bp->bio_caller2 == bp->_bio_caller2,
622 		    ("bio_caller2 used by the provider %s", pp->name));
623 		KASSERT(bp->bio_cflags == bp->_bio_cflags,
624 		    ("bio_cflags used by the provider %s", pp->name));
625 	}
626 #endif
627 	KASSERT(bp->bio_completed >= 0, ("bio_completed can't be less than 0"));
628 	KASSERT(bp->bio_completed <= bp->bio_length,
629 	    ("bio_completed can't be greater than bio_length"));
630 
631 	g_trace(G_T_BIO,
632 "g_io_deliver(%p) from %p(%s) to %p(%s) cmd %d error %d off %jd len %jd",
633 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd, error,
634 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
635 
636 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
637 	    ("Bio already on queue bp=%p", bp));
638 
639 	/*
640 	 * XXX: next two doesn't belong here
641 	 */
642 	bp->bio_bcount = bp->bio_length;
643 	bp->bio_resid = bp->bio_bcount - bp->bio_completed;
644 
645 #ifdef GET_STACK_USAGE
646 	direct = (pp->flags & G_PF_DIRECT_SEND) &&
647 		 (cp->flags & G_CF_DIRECT_RECEIVE) &&
648 		 !g_is_geom_thread(curthread);
649 	if (direct) {
650 		/* Block direct execution if less then half of stack left. */
651 		size_t	st, su;
652 		GET_STACK_USAGE(st, su);
653 		if (su * 2 > st)
654 			direct = 0;
655 	}
656 #else
657 	direct = 0;
658 #endif
659 
660 	/*
661 	 * The statistics collection is lockless, as such, but we
662 	 * can not update one instance of the statistics from more
663 	 * than one thread at a time, so grab the lock first.
664 	 */
665 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
666 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
667 		binuptime(&now);
668 	mtxp = mtx_pool_find(mtxpool_sleep, cp);
669 	mtx_lock(mtxp);
670 	if (g_collectstats & G_STATS_PROVIDERS)
671 		devstat_end_transaction_bio_bt(pp->stat, bp, &now);
672 	if (g_collectstats & G_STATS_CONSUMERS)
673 		devstat_end_transaction_bio_bt(cp->stat, bp, &now);
674 	cp->nend++;
675 	pp->nend++;
676 	mtx_unlock(mtxp);
677 
678 	if (error != ENOMEM) {
679 		bp->bio_error = error;
680 		if (direct) {
681 			biodone(bp);
682 		} else {
683 			g_bioq_lock(&g_bio_run_up);
684 			first = TAILQ_EMPTY(&g_bio_run_up.bio_queue);
685 			TAILQ_INSERT_TAIL(&g_bio_run_up.bio_queue, bp, bio_queue);
686 			bp->bio_flags |= BIO_ONQUEUE;
687 			g_bio_run_up.bio_queue_length++;
688 			g_bioq_unlock(&g_bio_run_up);
689 			if (first)
690 				wakeup(&g_wait_up);
691 		}
692 		return;
693 	}
694 
695 	if (bootverbose)
696 		printf("ENOMEM %p on %p(%s)\n", bp, pp, pp->name);
697 	bp->bio_children = 0;
698 	bp->bio_inbed = 0;
699 	bp->bio_driver1 = NULL;
700 	bp->bio_driver2 = NULL;
701 	bp->bio_pflags = 0;
702 	g_io_request(bp, cp);
703 	pace = 1;
704 	return;
705 }
706 
707 SYSCTL_DECL(_kern_geom);
708 
709 static long transient_maps;
710 SYSCTL_LONG(_kern_geom, OID_AUTO, transient_maps, CTLFLAG_RD,
711     &transient_maps, 0,
712     "Total count of the transient mapping requests");
713 u_int transient_map_retries = 10;
714 SYSCTL_UINT(_kern_geom, OID_AUTO, transient_map_retries, CTLFLAG_RW,
715     &transient_map_retries, 0,
716     "Max count of retries used before giving up on creating transient map");
717 int transient_map_hard_failures;
718 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_hard_failures, CTLFLAG_RD,
719     &transient_map_hard_failures, 0,
720     "Failures to establish the transient mapping due to retry attempts "
721     "exhausted");
722 int transient_map_soft_failures;
723 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_soft_failures, CTLFLAG_RD,
724     &transient_map_soft_failures, 0,
725     "Count of retried failures to establish the transient mapping");
726 int inflight_transient_maps;
727 SYSCTL_INT(_kern_geom, OID_AUTO, inflight_transient_maps, CTLFLAG_RD,
728     &inflight_transient_maps, 0,
729     "Current count of the active transient maps");
730 
731 static int
732 g_io_transient_map_bio(struct bio *bp)
733 {
734 	vm_offset_t addr;
735 	long size;
736 	u_int retried;
737 
738 	KASSERT(unmapped_buf_allowed, ("unmapped disabled"));
739 
740 	size = round_page(bp->bio_ma_offset + bp->bio_length);
741 	KASSERT(size / PAGE_SIZE == bp->bio_ma_n, ("Bio too short %p", bp));
742 	addr = 0;
743 	retried = 0;
744 	atomic_add_long(&transient_maps, 1);
745 retry:
746 	if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) {
747 		if (transient_map_retries != 0 &&
748 		    retried >= transient_map_retries) {
749 			CTR2(KTR_GEOM, "g_down cannot map bp %p provider %s",
750 			    bp, bp->bio_to->name);
751 			atomic_add_int(&transient_map_hard_failures, 1);
752 			return (EDEADLK/* XXXKIB */);
753 		} else {
754 			/*
755 			 * Naive attempt to quisce the I/O to get more
756 			 * in-flight requests completed and defragment
757 			 * the transient_arena.
758 			 */
759 			CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d",
760 			    bp, bp->bio_to->name, retried);
761 			pause("g_d_tra", hz / 10);
762 			retried++;
763 			atomic_add_int(&transient_map_soft_failures, 1);
764 			goto retry;
765 		}
766 	}
767 	atomic_add_int(&inflight_transient_maps, 1);
768 	pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size));
769 	bp->bio_data = (caddr_t)addr + bp->bio_ma_offset;
770 	bp->bio_flags |= BIO_TRANSIENT_MAPPING;
771 	bp->bio_flags &= ~BIO_UNMAPPED;
772 	return (EJUSTRETURN);
773 }
774 
775 void
776 g_io_schedule_down(struct thread *tp __unused)
777 {
778 	struct bio *bp;
779 	int error;
780 
781 	for(;;) {
782 		g_bioq_lock(&g_bio_run_down);
783 		bp = g_bioq_first(&g_bio_run_down);
784 		if (bp == NULL) {
785 			CTR0(KTR_GEOM, "g_down going to sleep");
786 			msleep(&g_wait_down, &g_bio_run_down.bio_queue_lock,
787 			    PRIBIO | PDROP, "-", 0);
788 			continue;
789 		}
790 		CTR0(KTR_GEOM, "g_down has work to do");
791 		g_bioq_unlock(&g_bio_run_down);
792 		if (pace != 0) {
793 			/*
794 			 * There has been at least one memory allocation
795 			 * failure since the last I/O completed. Pause 1ms to
796 			 * give the system a chance to free up memory. We only
797 			 * do this once because a large number of allocations
798 			 * can fail in the direct dispatch case and there's no
799 			 * relationship between the number of these failures and
800 			 * the length of the outage. If there's still an outage,
801 			 * we'll pause again and again until it's
802 			 * resolved. Older versions paused longer and once per
803 			 * allocation failure. This was OK for a single threaded
804 			 * g_down, but with direct dispatch would lead to max of
805 			 * 10 IOPs for minutes at a time when transient memory
806 			 * issues prevented allocation for a batch of requests
807 			 * from the upper layers.
808 			 *
809 			 * XXX This pacing is really lame. It needs to be solved
810 			 * by other methods. This is OK only because the worst
811 			 * case scenario is so rare. In the worst case scenario
812 			 * all memory is tied up waiting for I/O to complete
813 			 * which can never happen since we can't allocate bios
814 			 * for that I/O.
815 			 */
816 			CTR0(KTR_GEOM, "g_down pacing self");
817 			pause("g_down", min(hz/1000, 1));
818 			pace = 0;
819 		}
820 		CTR2(KTR_GEOM, "g_down processing bp %p provider %s", bp,
821 		    bp->bio_to->name);
822 		error = g_io_check(bp);
823 		if (error >= 0) {
824 			CTR3(KTR_GEOM, "g_down g_io_check on bp %p provider "
825 			    "%s returned %d", bp, bp->bio_to->name, error);
826 			g_io_deliver(bp, error);
827 			continue;
828 		}
829 		THREAD_NO_SLEEPING();
830 		CTR4(KTR_GEOM, "g_down starting bp %p provider %s off %ld "
831 		    "len %ld", bp, bp->bio_to->name, bp->bio_offset,
832 		    bp->bio_length);
833 		bp->bio_to->geom->start(bp);
834 		THREAD_SLEEPING_OK();
835 	}
836 }
837 
838 void
839 bio_taskqueue(struct bio *bp, bio_task_t *func, void *arg)
840 {
841 	bp->bio_task = func;
842 	bp->bio_task_arg = arg;
843 	/*
844 	 * The taskqueue is actually just a second queue off the "up"
845 	 * queue, so we use the same lock.
846 	 */
847 	g_bioq_lock(&g_bio_run_up);
848 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
849 	    ("Bio already on queue bp=%p target taskq", bp));
850 	bp->bio_flags |= BIO_ONQUEUE;
851 	TAILQ_INSERT_TAIL(&g_bio_run_task.bio_queue, bp, bio_queue);
852 	g_bio_run_task.bio_queue_length++;
853 	wakeup(&g_wait_up);
854 	g_bioq_unlock(&g_bio_run_up);
855 }
856 
857 
858 void
859 g_io_schedule_up(struct thread *tp __unused)
860 {
861 	struct bio *bp;
862 	for(;;) {
863 		g_bioq_lock(&g_bio_run_up);
864 		bp = g_bioq_first(&g_bio_run_task);
865 		if (bp != NULL) {
866 			g_bioq_unlock(&g_bio_run_up);
867 			THREAD_NO_SLEEPING();
868 			CTR1(KTR_GEOM, "g_up processing task bp %p", bp);
869 			bp->bio_task(bp->bio_task_arg);
870 			THREAD_SLEEPING_OK();
871 			continue;
872 		}
873 		bp = g_bioq_first(&g_bio_run_up);
874 		if (bp != NULL) {
875 			g_bioq_unlock(&g_bio_run_up);
876 			THREAD_NO_SLEEPING();
877 			CTR4(KTR_GEOM, "g_up biodone bp %p provider %s off "
878 			    "%jd len %ld", bp, bp->bio_to->name,
879 			    bp->bio_offset, bp->bio_length);
880 			biodone(bp);
881 			THREAD_SLEEPING_OK();
882 			continue;
883 		}
884 		CTR0(KTR_GEOM, "g_up going to sleep");
885 		msleep(&g_wait_up, &g_bio_run_up.bio_queue_lock,
886 		    PRIBIO | PDROP, "-", 0);
887 	}
888 }
889 
890 void *
891 g_read_data(struct g_consumer *cp, off_t offset, off_t length, int *error)
892 {
893 	struct bio *bp;
894 	void *ptr;
895 	int errorc;
896 
897 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
898 	    length <= MAXPHYS, ("g_read_data(): invalid length %jd",
899 	    (intmax_t)length));
900 
901 	bp = g_alloc_bio();
902 	bp->bio_cmd = BIO_READ;
903 	bp->bio_done = NULL;
904 	bp->bio_offset = offset;
905 	bp->bio_length = length;
906 	ptr = g_malloc(length, M_WAITOK);
907 	bp->bio_data = ptr;
908 	g_io_request(bp, cp);
909 	errorc = biowait(bp, "gread");
910 	if (error != NULL)
911 		*error = errorc;
912 	g_destroy_bio(bp);
913 	if (errorc) {
914 		g_free(ptr);
915 		ptr = NULL;
916 	}
917 	return (ptr);
918 }
919 
920 int
921 g_write_data(struct g_consumer *cp, off_t offset, void *ptr, off_t length)
922 {
923 	struct bio *bp;
924 	int error;
925 
926 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
927 	    length <= MAXPHYS, ("g_write_data(): invalid length %jd",
928 	    (intmax_t)length));
929 
930 	bp = g_alloc_bio();
931 	bp->bio_cmd = BIO_WRITE;
932 	bp->bio_done = NULL;
933 	bp->bio_offset = offset;
934 	bp->bio_length = length;
935 	bp->bio_data = ptr;
936 	g_io_request(bp, cp);
937 	error = biowait(bp, "gwrite");
938 	g_destroy_bio(bp);
939 	return (error);
940 }
941 
942 int
943 g_delete_data(struct g_consumer *cp, off_t offset, off_t length)
944 {
945 	struct bio *bp;
946 	int error;
947 
948 	KASSERT(length > 0 && length >= cp->provider->sectorsize,
949 	    ("g_delete_data(): invalid length %jd", (intmax_t)length));
950 
951 	bp = g_alloc_bio();
952 	bp->bio_cmd = BIO_DELETE;
953 	bp->bio_done = NULL;
954 	bp->bio_offset = offset;
955 	bp->bio_length = length;
956 	bp->bio_data = NULL;
957 	g_io_request(bp, cp);
958 	error = biowait(bp, "gdelete");
959 	g_destroy_bio(bp);
960 	return (error);
961 }
962 
963 void
964 g_print_bio(struct bio *bp)
965 {
966 	const char *pname, *cmd = NULL;
967 
968 	if (bp->bio_to != NULL)
969 		pname = bp->bio_to->name;
970 	else
971 		pname = "[unknown]";
972 
973 	switch (bp->bio_cmd) {
974 	case BIO_GETATTR:
975 		cmd = "GETATTR";
976 		printf("%s[%s(attr=%s)]", pname, cmd, bp->bio_attribute);
977 		return;
978 	case BIO_FLUSH:
979 		cmd = "FLUSH";
980 		printf("%s[%s]", pname, cmd);
981 		return;
982 	case BIO_READ:
983 		cmd = "READ";
984 		break;
985 	case BIO_WRITE:
986 		cmd = "WRITE";
987 		break;
988 	case BIO_DELETE:
989 		cmd = "DELETE";
990 		break;
991 	default:
992 		cmd = "UNKNOWN";
993 		printf("%s[%s()]", pname, cmd);
994 		return;
995 	}
996 	printf("%s[%s(offset=%jd, length=%jd)]", pname, cmd,
997 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
998 }
999