xref: /freebsd/sys/geom/geom_io.c (revision 2008043f386721d58158e37e0d7e50df8095942d)
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 #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/sbuf.h>
51 #include <sys/stack.h>
52 #include <sys/sysctl.h>
53 #include <sys/vmem.h>
54 #include <machine/stack.h>
55 #include <machine/stdarg.h>
56 
57 #include <sys/errno.h>
58 #include <geom/geom.h>
59 #include <geom/geom_int.h>
60 #include <sys/devicestat.h>
61 
62 #include <vm/uma.h>
63 #include <vm/vm.h>
64 #include <vm/vm_param.h>
65 #include <vm/vm_kern.h>
66 #include <vm/vm_page.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_extern.h>
69 #include <vm/vm_map.h>
70 
71 #define KTR_GEOM_ENABLED \
72     ((KTR_COMPILE & KTR_GEOM) != 0 && (ktr_mask & KTR_GEOM) != 0)
73 
74 static int	g_io_transient_map_bio(struct bio *bp);
75 
76 static struct g_bioq g_bio_run_down;
77 static struct g_bioq g_bio_run_up;
78 
79 /*
80  * Pace is a hint that we've had some trouble recently allocating
81  * bios, so we should back off trying to send I/O down the stack
82  * a bit to let the problem resolve. When pacing, we also turn
83  * off direct dispatch to also reduce memory pressure from I/Os
84  * there, at the expxense of some added latency while the memory
85  * pressures exist. See g_io_schedule_down() for more details
86  * and limitations.
87  */
88 static volatile u_int __read_mostly pace;
89 
90 static uma_zone_t __read_mostly biozone;
91 
92 #include <machine/atomic.h>
93 
94 static void
95 g_bioq_lock(struct g_bioq *bq)
96 {
97 
98 	mtx_lock(&bq->bio_queue_lock);
99 }
100 
101 static void
102 g_bioq_unlock(struct g_bioq *bq)
103 {
104 
105 	mtx_unlock(&bq->bio_queue_lock);
106 }
107 
108 #if 0
109 static void
110 g_bioq_destroy(struct g_bioq *bq)
111 {
112 
113 	mtx_destroy(&bq->bio_queue_lock);
114 }
115 #endif
116 
117 static void
118 g_bioq_init(struct g_bioq *bq)
119 {
120 
121 	TAILQ_INIT(&bq->bio_queue);
122 	mtx_init(&bq->bio_queue_lock, "bio queue", NULL, MTX_DEF);
123 }
124 
125 static struct bio *
126 g_bioq_first(struct g_bioq *bq)
127 {
128 	struct bio *bp;
129 
130 	bp = TAILQ_FIRST(&bq->bio_queue);
131 	if (bp != NULL) {
132 		KASSERT((bp->bio_flags & BIO_ONQUEUE),
133 		    ("Bio not on queue bp=%p target %p", bp, bq));
134 		bp->bio_flags &= ~BIO_ONQUEUE;
135 		TAILQ_REMOVE(&bq->bio_queue, bp, bio_queue);
136 		bq->bio_queue_length--;
137 	}
138 	return (bp);
139 }
140 
141 struct bio *
142 g_new_bio(void)
143 {
144 	struct bio *bp;
145 
146 	bp = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
147 #ifdef KTR
148 	if (KTR_GEOM_ENABLED) {
149 		struct stack st;
150 
151 		CTR1(KTR_GEOM, "g_new_bio(): %p", bp);
152 		stack_save(&st);
153 		CTRSTACK(KTR_GEOM, &st, 3);
154 	}
155 #endif
156 	return (bp);
157 }
158 
159 struct bio *
160 g_alloc_bio(void)
161 {
162 	struct bio *bp;
163 
164 	bp = uma_zalloc(biozone, M_WAITOK | M_ZERO);
165 #ifdef KTR
166 	if (KTR_GEOM_ENABLED) {
167 		struct stack st;
168 
169 		CTR1(KTR_GEOM, "g_alloc_bio(): %p", bp);
170 		stack_save(&st);
171 		CTRSTACK(KTR_GEOM, &st, 3);
172 	}
173 #endif
174 	return (bp);
175 }
176 
177 void
178 g_destroy_bio(struct bio *bp)
179 {
180 #ifdef KTR
181 	if (KTR_GEOM_ENABLED) {
182 		struct stack st;
183 
184 		CTR1(KTR_GEOM, "g_destroy_bio(): %p", bp);
185 		stack_save(&st);
186 		CTRSTACK(KTR_GEOM, &st, 3);
187 	}
188 #endif
189 	uma_zfree(biozone, bp);
190 }
191 
192 struct bio *
193 g_clone_bio(struct bio *bp)
194 {
195 	struct bio *bp2;
196 
197 	bp2 = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
198 	if (bp2 != NULL) {
199 		bp2->bio_parent = bp;
200 		bp2->bio_cmd = bp->bio_cmd;
201 		/*
202 		 *  BIO_ORDERED flag may be used by disk drivers to enforce
203 		 *  ordering restrictions, so this flag needs to be cloned.
204 		 *  BIO_UNMAPPED, BIO_VLIST, and BIO_SWAP should be inherited,
205 		 *  to properly indicate which way the buffer is passed.
206 		 *  Other bio flags are not suitable for cloning.
207 		 */
208 		bp2->bio_flags = bp->bio_flags &
209 		    (BIO_ORDERED | BIO_UNMAPPED | BIO_VLIST | BIO_SWAP);
210 		bp2->bio_length = bp->bio_length;
211 		bp2->bio_offset = bp->bio_offset;
212 		bp2->bio_data = bp->bio_data;
213 		bp2->bio_ma = bp->bio_ma;
214 		bp2->bio_ma_n = bp->bio_ma_n;
215 		bp2->bio_ma_offset = bp->bio_ma_offset;
216 		bp2->bio_attribute = bp->bio_attribute;
217 		if (bp->bio_cmd == BIO_ZONE)
218 			bcopy(&bp->bio_zone, &bp2->bio_zone,
219 			    sizeof(bp->bio_zone));
220 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
221 		bp2->bio_track_bp = bp->bio_track_bp;
222 #endif
223 		bp->bio_children++;
224 	}
225 #ifdef KTR
226 	if (KTR_GEOM_ENABLED) {
227 		struct stack st;
228 
229 		CTR2(KTR_GEOM, "g_clone_bio(%p): %p", bp, bp2);
230 		stack_save(&st);
231 		CTRSTACK(KTR_GEOM, &st, 3);
232 	}
233 #endif
234 	return(bp2);
235 }
236 
237 struct bio *
238 g_duplicate_bio(struct bio *bp)
239 {
240 	struct bio *bp2;
241 
242 	bp2 = uma_zalloc(biozone, M_WAITOK | M_ZERO);
243 	bp2->bio_flags = bp->bio_flags & (BIO_UNMAPPED | BIO_VLIST | BIO_SWAP);
244 	bp2->bio_parent = bp;
245 	bp2->bio_cmd = bp->bio_cmd;
246 	bp2->bio_length = bp->bio_length;
247 	bp2->bio_offset = bp->bio_offset;
248 	bp2->bio_data = bp->bio_data;
249 	bp2->bio_ma = bp->bio_ma;
250 	bp2->bio_ma_n = bp->bio_ma_n;
251 	bp2->bio_ma_offset = bp->bio_ma_offset;
252 	bp2->bio_attribute = bp->bio_attribute;
253 	bp->bio_children++;
254 #ifdef KTR
255 	if (KTR_GEOM_ENABLED) {
256 		struct stack st;
257 
258 		CTR2(KTR_GEOM, "g_duplicate_bio(%p): %p", bp, bp2);
259 		stack_save(&st);
260 		CTRSTACK(KTR_GEOM, &st, 3);
261 	}
262 #endif
263 	return(bp2);
264 }
265 
266 void
267 g_reset_bio(struct bio *bp)
268 {
269 
270 	bzero(bp, sizeof(*bp));
271 }
272 
273 void
274 g_io_init(void)
275 {
276 
277 	g_bioq_init(&g_bio_run_down);
278 	g_bioq_init(&g_bio_run_up);
279 	biozone = uma_zcreate("g_bio", sizeof (struct bio),
280 	    NULL, NULL,
281 	    NULL, NULL,
282 	    0, 0);
283 }
284 
285 int
286 g_io_getattr(const char *attr, struct g_consumer *cp, int *len, void *ptr)
287 {
288 	struct bio *bp;
289 	int error;
290 
291 	g_trace(G_T_BIO, "bio_getattr(%s)", attr);
292 	bp = g_alloc_bio();
293 	bp->bio_cmd = BIO_GETATTR;
294 	bp->bio_done = NULL;
295 	bp->bio_attribute = attr;
296 	bp->bio_length = *len;
297 	bp->bio_data = ptr;
298 	g_io_request(bp, cp);
299 	error = biowait(bp, "ggetattr");
300 	*len = bp->bio_completed;
301 	g_destroy_bio(bp);
302 	return (error);
303 }
304 
305 int
306 g_io_zonecmd(struct disk_zone_args *zone_args, struct g_consumer *cp)
307 {
308 	struct bio *bp;
309 	int error;
310 
311 	g_trace(G_T_BIO, "bio_zone(%d)", zone_args->zone_cmd);
312 	bp = g_alloc_bio();
313 	bp->bio_cmd = BIO_ZONE;
314 	bp->bio_done = NULL;
315 	/*
316 	 * XXX KDM need to handle report zone data.
317 	 */
318 	bcopy(zone_args, &bp->bio_zone, sizeof(*zone_args));
319 	if (zone_args->zone_cmd == DISK_ZONE_REPORT_ZONES)
320 		bp->bio_length =
321 		    zone_args->zone_params.report.entries_allocated *
322 		    sizeof(struct disk_zone_rep_entry);
323 	else
324 		bp->bio_length = 0;
325 
326 	g_io_request(bp, cp);
327 	error = biowait(bp, "gzone");
328 	bcopy(&bp->bio_zone, zone_args, sizeof(*zone_args));
329 	g_destroy_bio(bp);
330 	return (error);
331 }
332 
333 /*
334  * Send a BIO_SPEEDUP down the stack. This is used to tell the lower layers that
335  * the upper layers have detected a resource shortage. The lower layers are
336  * advised to stop delaying I/O that they might be holding for performance
337  * reasons and to schedule it (non-trims) or complete it successfully (trims) as
338  * quickly as it can. bio_length is the amount of the shortage.  This call
339  * should be non-blocking. bio_resid is used to communicate back if the lower
340  * layers couldn't find bio_length worth of I/O to schedule or discard. A length
341  * of 0 means to do as much as you can (schedule the h/w queues full, discard
342  * all trims). flags are a hint from the upper layers to the lower layers what
343  * operation should be done.
344  */
345 int
346 g_io_speedup(off_t shortage, u_int flags, size_t *resid, struct g_consumer *cp)
347 {
348 	struct bio *bp;
349 	int error;
350 
351 	KASSERT((flags & (BIO_SPEEDUP_TRIM | BIO_SPEEDUP_WRITE)) != 0,
352 	    ("Invalid flags passed to g_io_speedup: %#x", flags));
353 	g_trace(G_T_BIO, "bio_speedup(%s, %jd, %#x)", cp->provider->name,
354 	    (intmax_t)shortage, flags);
355 	bp = g_new_bio();
356 	if (bp == NULL)
357 		return (ENOMEM);
358 	bp->bio_cmd = BIO_SPEEDUP;
359 	bp->bio_length = shortage;
360 	bp->bio_done = NULL;
361 	bp->bio_flags |= flags;
362 	g_io_request(bp, cp);
363 	error = biowait(bp, "gflush");
364 	*resid = bp->bio_resid;
365 	g_destroy_bio(bp);
366 	return (error);
367 }
368 
369 int
370 g_io_flush(struct g_consumer *cp)
371 {
372 	struct bio *bp;
373 	int error;
374 
375 	g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name);
376 	bp = g_alloc_bio();
377 	bp->bio_cmd = BIO_FLUSH;
378 	bp->bio_flags |= BIO_ORDERED;
379 	bp->bio_done = NULL;
380 	bp->bio_attribute = NULL;
381 	bp->bio_offset = cp->provider->mediasize;
382 	bp->bio_length = 0;
383 	bp->bio_data = NULL;
384 	g_io_request(bp, cp);
385 	error = biowait(bp, "gflush");
386 	g_destroy_bio(bp);
387 	return (error);
388 }
389 
390 static int
391 g_io_check(struct bio *bp)
392 {
393 	struct g_consumer *cp;
394 	struct g_provider *pp;
395 	off_t excess;
396 	int error;
397 
398 	biotrack(bp, __func__);
399 
400 	cp = bp->bio_from;
401 	pp = bp->bio_to;
402 
403 	/* Fail if access counters dont allow the operation */
404 	switch(bp->bio_cmd) {
405 	case BIO_READ:
406 	case BIO_GETATTR:
407 		if (cp->acr == 0)
408 			return (EPERM);
409 		break;
410 	case BIO_WRITE:
411 	case BIO_DELETE:
412 	case BIO_SPEEDUP:
413 	case BIO_FLUSH:
414 		if (cp->acw == 0)
415 			return (EPERM);
416 		break;
417 	case BIO_ZONE:
418 		if ((bp->bio_zone.zone_cmd == DISK_ZONE_REPORT_ZONES) ||
419 		    (bp->bio_zone.zone_cmd == DISK_ZONE_GET_PARAMS)) {
420 			if (cp->acr == 0)
421 				return (EPERM);
422 		} else if (cp->acw == 0)
423 			return (EPERM);
424 		break;
425 	default:
426 		return (EPERM);
427 	}
428 	/* if provider is marked for error, don't disturb. */
429 	if (pp->error)
430 		return (pp->error);
431 	if (cp->flags & G_CF_ORPHAN)
432 		return (ENXIO);
433 
434 	switch(bp->bio_cmd) {
435 	case BIO_READ:
436 	case BIO_WRITE:
437 	case BIO_DELETE:
438 		/* Zero sectorsize or mediasize is probably a lack of media. */
439 		if (pp->sectorsize == 0 || pp->mediasize == 0)
440 			return (ENXIO);
441 		/* Reject I/O not on sector boundary */
442 		if (bp->bio_offset % pp->sectorsize)
443 			return (EINVAL);
444 		/* Reject I/O not integral sector long */
445 		if (bp->bio_length % pp->sectorsize)
446 			return (EINVAL);
447 		/* Reject requests before or past the end of media. */
448 		if (bp->bio_offset < 0)
449 			return (EIO);
450 		if (bp->bio_offset > pp->mediasize)
451 			return (EIO);
452 
453 		/* Truncate requests to the end of providers media. */
454 		excess = bp->bio_offset + bp->bio_length;
455 		if (excess > bp->bio_to->mediasize) {
456 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
457 			    round_page(bp->bio_ma_offset +
458 			    bp->bio_length) / PAGE_SIZE == bp->bio_ma_n,
459 			    ("excess bio %p too short", bp));
460 			excess -= bp->bio_to->mediasize;
461 			bp->bio_length -= excess;
462 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
463 				bp->bio_ma_n = round_page(bp->bio_ma_offset +
464 				    bp->bio_length) / PAGE_SIZE;
465 			}
466 			if (excess > 0)
467 				CTR3(KTR_GEOM, "g_down truncated bio "
468 				    "%p provider %s by %d", bp,
469 				    bp->bio_to->name, excess);
470 		}
471 
472 		/* Deliver zero length transfers right here. */
473 		if (bp->bio_length == 0) {
474 			CTR2(KTR_GEOM, "g_down terminated 0-length "
475 			    "bp %p provider %s", bp, bp->bio_to->name);
476 			return (0);
477 		}
478 
479 		if ((bp->bio_flags & BIO_UNMAPPED) != 0 &&
480 		    (bp->bio_to->flags & G_PF_ACCEPT_UNMAPPED) == 0 &&
481 		    (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) {
482 			if ((error = g_io_transient_map_bio(bp)) >= 0)
483 				return (error);
484 		}
485 		break;
486 	default:
487 		break;
488 	}
489 	return (EJUSTRETURN);
490 }
491 
492 void
493 g_io_request(struct bio *bp, struct g_consumer *cp)
494 {
495 	struct g_provider *pp;
496 	int direct, error, first;
497 	uint8_t cmd;
498 
499 	biotrack(bp, __func__);
500 
501 	KASSERT(cp != NULL, ("NULL cp in g_io_request"));
502 	KASSERT(bp != NULL, ("NULL bp in g_io_request"));
503 	pp = cp->provider;
504 	KASSERT(pp != NULL, ("consumer not attached in g_io_request"));
505 #ifdef DIAGNOSTIC
506 	KASSERT(bp->bio_driver1 == NULL,
507 	    ("bio_driver1 used by the consumer (geom %s)", cp->geom->name));
508 	KASSERT(bp->bio_driver2 == NULL,
509 	    ("bio_driver2 used by the consumer (geom %s)", cp->geom->name));
510 	KASSERT(bp->bio_pflags == 0,
511 	    ("bio_pflags used by the consumer (geom %s)", cp->geom->name));
512 	/*
513 	 * Remember consumer's private fields, so we can detect if they were
514 	 * modified by the provider.
515 	 */
516 	bp->_bio_caller1 = bp->bio_caller1;
517 	bp->_bio_caller2 = bp->bio_caller2;
518 	bp->_bio_cflags = bp->bio_cflags;
519 #endif
520 
521 	cmd = bp->bio_cmd;
522 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_GETATTR) {
523 		KASSERT(bp->bio_data != NULL,
524 		    ("NULL bp->data in g_io_request(cmd=%hu)", bp->bio_cmd));
525 	}
526 	if (cmd == BIO_DELETE || cmd == BIO_FLUSH || cmd == BIO_SPEEDUP) {
527 		KASSERT(bp->bio_data == NULL,
528 		    ("non-NULL bp->data in g_io_request(cmd=%hu)",
529 		    bp->bio_cmd));
530 	}
531 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_DELETE) {
532 		KASSERT(bp->bio_offset % cp->provider->sectorsize == 0,
533 		    ("wrong offset %jd for sectorsize %u",
534 		    bp->bio_offset, cp->provider->sectorsize));
535 		KASSERT(bp->bio_length % cp->provider->sectorsize == 0,
536 		    ("wrong length %jd for sectorsize %u",
537 		    bp->bio_length, cp->provider->sectorsize));
538 	}
539 
540 	g_trace(G_T_BIO, "bio_request(%p) from %p(%s) to %p(%s) cmd %d",
541 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd);
542 
543 	bp->bio_from = cp;
544 	bp->bio_to = pp;
545 	bp->bio_error = 0;
546 	bp->bio_completed = 0;
547 
548 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
549 	    ("Bio already on queue bp=%p", bp));
550 
551 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
552 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
553 		binuptime(&bp->bio_t0);
554 	else
555 		getbinuptime(&bp->bio_t0);
556 	if (g_collectstats & G_STATS_CONSUMERS)
557 		devstat_start_transaction_bio_t0(cp->stat, bp);
558 	if (g_collectstats & G_STATS_PROVIDERS)
559 		devstat_start_transaction_bio_t0(pp->stat, bp);
560 #ifdef INVARIANTS
561 	atomic_add_int(&cp->nstart, 1);
562 #endif
563 
564 	direct = (cp->flags & G_CF_DIRECT_SEND) != 0 &&
565 	    (pp->flags & G_PF_DIRECT_RECEIVE) != 0 &&
566 	    curthread != g_down_td &&
567 	    ((pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ||
568 	    (bp->bio_flags & BIO_UNMAPPED) == 0 || THREAD_CAN_SLEEP()) &&
569 	    pace == 0;
570 	if (direct) {
571 		/* Block direct execution if less then half of stack left. */
572 		size_t	st, su;
573 		GET_STACK_USAGE(st, su);
574 		if (su * 2 > st)
575 			direct = 0;
576 	}
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 	biotrack(bp, __func__);
611 
612 	KASSERT(bp != NULL, ("NULL bp in g_io_deliver"));
613 	pp = bp->bio_to;
614 	KASSERT(pp != NULL, ("NULL bio_to in g_io_deliver"));
615 	cp = bp->bio_from;
616 	if (cp == NULL) {
617 		bp->bio_error = error;
618 		bp->bio_done(bp);
619 		return;
620 	}
621 	KASSERT(cp != NULL, ("NULL bio_from in g_io_deliver"));
622 	KASSERT(cp->geom != NULL, ("NULL bio_from->geom in g_io_deliver"));
623 #ifdef DIAGNOSTIC
624 	/*
625 	 * Some classes - GJournal in particular - can modify bio's
626 	 * private fields while the bio is in transit; G_GEOM_VOLATILE_BIO
627 	 * flag means it's an expected behaviour for that particular geom.
628 	 */
629 	if ((cp->geom->flags & G_GEOM_VOLATILE_BIO) == 0) {
630 		KASSERT(bp->bio_caller1 == bp->_bio_caller1,
631 		    ("bio_caller1 used by the provider %s", pp->name));
632 		KASSERT(bp->bio_caller2 == bp->_bio_caller2,
633 		    ("bio_caller2 used by the provider %s", pp->name));
634 		KASSERT(bp->bio_cflags == bp->_bio_cflags,
635 		    ("bio_cflags used by the provider %s", pp->name));
636 	}
637 #endif
638 	KASSERT(bp->bio_completed >= 0, ("bio_completed can't be less than 0"));
639 	KASSERT(bp->bio_completed <= bp->bio_length,
640 	    ("bio_completed can't be greater than bio_length"));
641 
642 	g_trace(G_T_BIO,
643 "g_io_deliver(%p) from %p(%s) to %p(%s) cmd %d error %d off %jd len %jd",
644 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd, error,
645 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
646 
647 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
648 	    ("Bio already on queue bp=%p", bp));
649 
650 	/*
651 	 * XXX: next two doesn't belong here
652 	 */
653 	bp->bio_bcount = bp->bio_length;
654 	bp->bio_resid = bp->bio_bcount - bp->bio_completed;
655 
656 	direct = (pp->flags & G_PF_DIRECT_SEND) &&
657 		 (cp->flags & G_CF_DIRECT_RECEIVE) &&
658 		 curthread != g_up_td;
659 	if (direct) {
660 		/* Block direct execution if less then half of stack left. */
661 		size_t	st, su;
662 		GET_STACK_USAGE(st, su);
663 		if (su * 2 > st)
664 			direct = 0;
665 	}
666 
667 	/*
668 	 * The statistics collection is lockless, as such, but we
669 	 * can not update one instance of the statistics from more
670 	 * than one thread at a time, so grab the lock first.
671 	 */
672 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
673 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
674 		binuptime(&now);
675 	mtxp = mtx_pool_find(mtxpool_sleep, pp);
676 	mtx_lock(mtxp);
677 	if (g_collectstats & G_STATS_PROVIDERS)
678 		devstat_end_transaction_bio_bt(pp->stat, bp, &now);
679 	if (g_collectstats & G_STATS_CONSUMERS)
680 		devstat_end_transaction_bio_bt(cp->stat, bp, &now);
681 #ifdef INVARIANTS
682 	cp->nend++;
683 #endif
684 	mtx_unlock(mtxp);
685 
686 	if (error != ENOMEM) {
687 		bp->bio_error = error;
688 		if (direct) {
689 			biodone(bp);
690 		} else {
691 			g_bioq_lock(&g_bio_run_up);
692 			first = TAILQ_EMPTY(&g_bio_run_up.bio_queue);
693 			TAILQ_INSERT_TAIL(&g_bio_run_up.bio_queue, bp, bio_queue);
694 			bp->bio_flags |= BIO_ONQUEUE;
695 			g_bio_run_up.bio_queue_length++;
696 			g_bioq_unlock(&g_bio_run_up);
697 			if (first)
698 				wakeup(&g_wait_up);
699 		}
700 		return;
701 	}
702 
703 	if (bootverbose)
704 		printf("ENOMEM %p on %p(%s)\n", bp, pp, pp->name);
705 	bp->bio_children = 0;
706 	bp->bio_inbed = 0;
707 	bp->bio_driver1 = NULL;
708 	bp->bio_driver2 = NULL;
709 	bp->bio_pflags = 0;
710 	g_io_request(bp, cp);
711 	pace = 1;
712 	return;
713 }
714 
715 SYSCTL_DECL(_kern_geom);
716 
717 static long transient_maps;
718 SYSCTL_LONG(_kern_geom, OID_AUTO, transient_maps, CTLFLAG_RD,
719     &transient_maps, 0,
720     "Total count of the transient mapping requests");
721 u_int transient_map_retries = 10;
722 SYSCTL_UINT(_kern_geom, OID_AUTO, transient_map_retries, CTLFLAG_RW,
723     &transient_map_retries, 0,
724     "Max count of retries used before giving up on creating transient map");
725 int transient_map_hard_failures;
726 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_hard_failures, CTLFLAG_RD,
727     &transient_map_hard_failures, 0,
728     "Failures to establish the transient mapping due to retry attempts "
729     "exhausted");
730 int transient_map_soft_failures;
731 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_soft_failures, CTLFLAG_RD,
732     &transient_map_soft_failures, 0,
733     "Count of retried failures to establish the transient mapping");
734 int inflight_transient_maps;
735 SYSCTL_INT(_kern_geom, OID_AUTO, inflight_transient_maps, CTLFLAG_RD,
736     &inflight_transient_maps, 0,
737     "Current count of the active transient maps");
738 
739 static int
740 g_io_transient_map_bio(struct bio *bp)
741 {
742 	vm_offset_t addr;
743 	long size;
744 	u_int retried;
745 
746 	KASSERT(unmapped_buf_allowed, ("unmapped disabled"));
747 
748 	size = round_page(bp->bio_ma_offset + bp->bio_length);
749 	KASSERT(size / PAGE_SIZE == bp->bio_ma_n, ("Bio too short %p", bp));
750 	addr = 0;
751 	retried = 0;
752 	atomic_add_long(&transient_maps, 1);
753 retry:
754 	if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) {
755 		if (transient_map_retries != 0 &&
756 		    retried >= transient_map_retries) {
757 			CTR2(KTR_GEOM, "g_down cannot map bp %p provider %s",
758 			    bp, bp->bio_to->name);
759 			atomic_add_int(&transient_map_hard_failures, 1);
760 			return (EDEADLK/* XXXKIB */);
761 		} else {
762 			/*
763 			 * Naive attempt to quisce the I/O to get more
764 			 * in-flight requests completed and defragment
765 			 * the transient_arena.
766 			 */
767 			CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d",
768 			    bp, bp->bio_to->name, retried);
769 			pause("g_d_tra", hz / 10);
770 			retried++;
771 			atomic_add_int(&transient_map_soft_failures, 1);
772 			goto retry;
773 		}
774 	}
775 	atomic_add_int(&inflight_transient_maps, 1);
776 	pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size));
777 	bp->bio_data = (caddr_t)addr + bp->bio_ma_offset;
778 	bp->bio_flags |= BIO_TRANSIENT_MAPPING;
779 	bp->bio_flags &= ~BIO_UNMAPPED;
780 	return (EJUSTRETURN);
781 }
782 
783 void
784 g_io_schedule_down(struct thread *tp __unused)
785 {
786 	struct bio *bp;
787 	int error;
788 
789 	for(;;) {
790 		g_bioq_lock(&g_bio_run_down);
791 		bp = g_bioq_first(&g_bio_run_down);
792 		if (bp == NULL) {
793 			CTR0(KTR_GEOM, "g_down going to sleep");
794 			msleep(&g_wait_down, &g_bio_run_down.bio_queue_lock,
795 			    PRIBIO | PDROP, "-", 0);
796 			continue;
797 		}
798 		CTR0(KTR_GEOM, "g_down has work to do");
799 		g_bioq_unlock(&g_bio_run_down);
800 		biotrack(bp, __func__);
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 g_io_schedule_up(struct thread *tp __unused)
849 {
850 	struct bio *bp;
851 
852 	for(;;) {
853 		g_bioq_lock(&g_bio_run_up);
854 		bp = g_bioq_first(&g_bio_run_up);
855 		if (bp == NULL) {
856 			CTR0(KTR_GEOM, "g_up going to sleep");
857 			msleep(&g_wait_up, &g_bio_run_up.bio_queue_lock,
858 			    PRIBIO | PDROP, "-", 0);
859 			continue;
860 		}
861 		g_bioq_unlock(&g_bio_run_up);
862 		THREAD_NO_SLEEPING();
863 		CTR4(KTR_GEOM, "g_up biodone bp %p provider %s off "
864 		    "%jd len %ld", bp, bp->bio_to->name,
865 		    bp->bio_offset, bp->bio_length);
866 		biodone(bp);
867 		THREAD_SLEEPING_OK();
868 	}
869 }
870 
871 void *
872 g_read_data(struct g_consumer *cp, off_t offset, off_t length, int *error)
873 {
874 	struct bio *bp;
875 	void *ptr;
876 	int errorc;
877 
878 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
879 	    length <= maxphys, ("g_read_data(): invalid length %jd",
880 	    (intmax_t)length));
881 
882 	bp = g_alloc_bio();
883 	bp->bio_cmd = BIO_READ;
884 	bp->bio_done = NULL;
885 	bp->bio_offset = offset;
886 	bp->bio_length = length;
887 	ptr = g_malloc(length, M_WAITOK);
888 	bp->bio_data = ptr;
889 	g_io_request(bp, cp);
890 	errorc = biowait(bp, "gread");
891 	if (errorc == 0 && bp->bio_completed != length)
892 		errorc = EIO;
893 	if (error != NULL)
894 		*error = errorc;
895 	g_destroy_bio(bp);
896 	if (errorc) {
897 		g_free(ptr);
898 		ptr = NULL;
899 	}
900 	return (ptr);
901 }
902 
903 /*
904  * A read function for use by ffs_sbget when used by GEOM-layer routines.
905  */
906 int
907 g_use_g_read_data(void *devfd, off_t loc, void **bufp, int size)
908 {
909 	struct g_consumer *cp;
910 
911 	KASSERT(*bufp == NULL,
912 	    ("g_use_g_read_data: non-NULL *bufp %p\n", *bufp));
913 
914 	cp = (struct g_consumer *)devfd;
915 	/*
916 	 * Take care not to issue an invalid I/O request. The offset of
917 	 * the superblock candidate must be multiples of the provider's
918 	 * sector size, otherwise an FFS can't exist on the provider
919 	 * anyway.
920 	 */
921 	if (loc % cp->provider->sectorsize != 0)
922 		return (ENOENT);
923 	*bufp = g_read_data(cp, loc, size, NULL);
924 	if (*bufp == NULL)
925 		return (ENOENT);
926 	return (0);
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 	if (error == 0 && bp->bio_completed != length)
948 		error = EIO;
949 	g_destroy_bio(bp);
950 	return (error);
951 }
952 
953 /*
954  * A write function for use by ffs_sbput when used by GEOM-layer routines.
955  */
956 int
957 g_use_g_write_data(void *devfd, off_t loc, void *buf, int size)
958 {
959 
960 	return (g_write_data((struct g_consumer *)devfd, loc, buf, size));
961 }
962 
963 int
964 g_delete_data(struct g_consumer *cp, off_t offset, off_t length)
965 {
966 	struct bio *bp;
967 	int error;
968 
969 	KASSERT(length > 0 && length >= cp->provider->sectorsize,
970 	    ("g_delete_data(): invalid length %jd", (intmax_t)length));
971 
972 	bp = g_alloc_bio();
973 	bp->bio_cmd = BIO_DELETE;
974 	bp->bio_done = NULL;
975 	bp->bio_offset = offset;
976 	bp->bio_length = length;
977 	bp->bio_data = NULL;
978 	g_io_request(bp, cp);
979 	error = biowait(bp, "gdelete");
980 	if (error == 0 && bp->bio_completed != length)
981 		error = EIO;
982 	g_destroy_bio(bp);
983 	return (error);
984 }
985 
986 void
987 g_print_bio(const char *prefix, const struct bio *bp, const char *fmtsuffix,
988     ...)
989 {
990 #ifndef PRINTF_BUFR_SIZE
991 #define PRINTF_BUFR_SIZE 64
992 #endif
993 	char bufr[PRINTF_BUFR_SIZE];
994 	struct sbuf sb, *sbp __unused;
995 	va_list ap;
996 
997 	sbp = sbuf_new(&sb, bufr, sizeof(bufr), SBUF_FIXEDLEN);
998 	KASSERT(sbp != NULL, ("sbuf_new misused?"));
999 
1000 	sbuf_set_drain(&sb, sbuf_printf_drain, NULL);
1001 
1002 	sbuf_cat(&sb, prefix);
1003 	g_format_bio(&sb, bp);
1004 
1005 	va_start(ap, fmtsuffix);
1006 	sbuf_vprintf(&sb, fmtsuffix, ap);
1007 	va_end(ap);
1008 
1009 	sbuf_nl_terminate(&sb);
1010 
1011 	sbuf_finish(&sb);
1012 	sbuf_delete(&sb);
1013 }
1014 
1015 void
1016 g_format_bio(struct sbuf *sb, const struct bio *bp)
1017 {
1018 	const char *pname, *cmd = NULL;
1019 
1020 	if (bp->bio_to != NULL)
1021 		pname = bp->bio_to->name;
1022 	else if (bp->bio_parent != NULL && bp->bio_parent->bio_to != NULL)
1023 		pname = bp->bio_parent->bio_to->name;
1024 	else
1025 		pname = "[unknown]";
1026 
1027 	switch (bp->bio_cmd) {
1028 	case BIO_GETATTR:
1029 		cmd = "GETATTR";
1030 		sbuf_printf(sb, "%s[%s(attr=%s)]", pname, cmd,
1031 		    bp->bio_attribute);
1032 		return;
1033 	case BIO_FLUSH:
1034 		cmd = "FLUSH";
1035 		sbuf_printf(sb, "%s[%s]", pname, cmd);
1036 		return;
1037 	case BIO_ZONE: {
1038 		char *subcmd = NULL;
1039 		cmd = "ZONE";
1040 		switch (bp->bio_zone.zone_cmd) {
1041 		case DISK_ZONE_OPEN:
1042 			subcmd = "OPEN";
1043 			break;
1044 		case DISK_ZONE_CLOSE:
1045 			subcmd = "CLOSE";
1046 			break;
1047 		case DISK_ZONE_FINISH:
1048 			subcmd = "FINISH";
1049 			break;
1050 		case DISK_ZONE_RWP:
1051 			subcmd = "RWP";
1052 			break;
1053 		case DISK_ZONE_REPORT_ZONES:
1054 			subcmd = "REPORT ZONES";
1055 			break;
1056 		case DISK_ZONE_GET_PARAMS:
1057 			subcmd = "GET PARAMS";
1058 			break;
1059 		default:
1060 			subcmd = "UNKNOWN";
1061 			break;
1062 		}
1063 		sbuf_printf(sb, "%s[%s,%s]", pname, cmd, subcmd);
1064 		return;
1065 	}
1066 	case BIO_READ:
1067 		cmd = "READ";
1068 		break;
1069 	case BIO_WRITE:
1070 		cmd = "WRITE";
1071 		break;
1072 	case BIO_DELETE:
1073 		cmd = "DELETE";
1074 		break;
1075 	default:
1076 		cmd = "UNKNOWN";
1077 		sbuf_printf(sb, "%s[%s()]", pname, cmd);
1078 		return;
1079 	}
1080 	sbuf_printf(sb, "%s[%s(offset=%jd, length=%jd)]", pname, cmd,
1081 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
1082 }
1083