xref: /freebsd/sys/geom/geom_io.c (revision 6d83b381860bd8a1ea2043e26bdf2cf2ed640a38)
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/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/malloc.h>
46 #include <sys/bio.h>
47 #include <sys/ktr.h>
48 #include <sys/proc.h>
49 #include <sys/sbuf.h>
50 #include <sys/stack.h>
51 #include <sys/sysctl.h>
52 #include <sys/vmem.h>
53 #include <machine/stack.h>
54 #include <machine/stdarg.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 #define KTR_GEOM_ENABLED \
71     ((KTR_COMPILE & KTR_GEOM) != 0 && (ktr_mask & KTR_GEOM) != 0)
72 
73 static int	g_io_transient_map_bio(struct bio *bp);
74 
75 static struct g_bioq g_bio_run_down;
76 static struct g_bioq g_bio_run_up;
77 
78 static u_long nomem_count;
79 static u_long pause_count;
80 
81 /*
82  * Pace is a hint that we've had some trouble recently allocating
83  * bios, so we should back off trying to send I/O down the stack
84  * a bit to let the problem resolve. When pacing, we also turn
85  * off direct dispatch to also reduce memory pressure from I/Os
86  * there, at the expxense of some added latency while the memory
87  * pressures exist. See g_io_schedule_down() for more details
88  * and limitations.
89  */
90 static volatile u_int __read_mostly pace;
91 
92 static uma_zone_t __read_mostly biozone;
93 
94 #include <machine/atomic.h>
95 
96 static void
g_bioq_lock(struct g_bioq * bq)97 g_bioq_lock(struct g_bioq *bq)
98 {
99 
100 	mtx_lock(&bq->bio_queue_lock);
101 }
102 
103 static void
g_bioq_unlock(struct g_bioq * bq)104 g_bioq_unlock(struct g_bioq *bq)
105 {
106 
107 	mtx_unlock(&bq->bio_queue_lock);
108 }
109 
110 #if 0
111 static void
112 g_bioq_destroy(struct g_bioq *bq)
113 {
114 
115 	mtx_destroy(&bq->bio_queue_lock);
116 }
117 #endif
118 
119 static void
g_bioq_init(struct g_bioq * bq)120 g_bioq_init(struct g_bioq *bq)
121 {
122 
123 	TAILQ_INIT(&bq->bio_queue);
124 	mtx_init(&bq->bio_queue_lock, "bio queue", NULL, MTX_DEF);
125 }
126 
127 static struct bio *
g_bioq_first(struct g_bioq * bq)128 g_bioq_first(struct g_bioq *bq)
129 {
130 	struct bio *bp;
131 
132 	bp = TAILQ_FIRST(&bq->bio_queue);
133 	if (bp != NULL) {
134 		KASSERT((bp->bio_flags & BIO_ONQUEUE),
135 		    ("Bio not on queue bp=%p target %p", bp, bq));
136 		bp->bio_flags &= ~BIO_ONQUEUE;
137 		TAILQ_REMOVE(&bq->bio_queue, bp, bio_queue);
138 		bq->bio_queue_length--;
139 	}
140 	return (bp);
141 }
142 
143 struct bio *
g_new_bio(void)144 g_new_bio(void)
145 {
146 	struct bio *bp;
147 
148 	bp = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
149 #ifdef KTR
150 	if (KTR_GEOM_ENABLED) {
151 		struct stack st;
152 
153 		CTR1(KTR_GEOM, "g_new_bio(): %p", bp);
154 		stack_save(&st);
155 		CTRSTACK(KTR_GEOM, &st, 3);
156 	}
157 #endif
158 	return (bp);
159 }
160 
161 struct bio *
g_alloc_bio(void)162 g_alloc_bio(void)
163 {
164 	struct bio *bp;
165 
166 	bp = uma_zalloc(biozone, M_WAITOK | M_ZERO);
167 #ifdef KTR
168 	if (KTR_GEOM_ENABLED) {
169 		struct stack st;
170 
171 		CTR1(KTR_GEOM, "g_alloc_bio(): %p", bp);
172 		stack_save(&st);
173 		CTRSTACK(KTR_GEOM, &st, 3);
174 	}
175 #endif
176 	return (bp);
177 }
178 
179 void
g_destroy_bio(struct bio * bp)180 g_destroy_bio(struct bio *bp)
181 {
182 #ifdef KTR
183 	if (KTR_GEOM_ENABLED) {
184 		struct stack st;
185 
186 		CTR1(KTR_GEOM, "g_destroy_bio(): %p", bp);
187 		stack_save(&st);
188 		CTRSTACK(KTR_GEOM, &st, 3);
189 	}
190 #endif
191 	uma_zfree(biozone, bp);
192 }
193 
194 struct bio *
g_clone_bio(struct bio * bp)195 g_clone_bio(struct bio *bp)
196 {
197 	struct bio *bp2;
198 
199 	bp2 = uma_zalloc(biozone, M_NOWAIT | M_ZERO);
200 	if (bp2 != NULL) {
201 		bp2->bio_parent = bp;
202 		bp2->bio_cmd = bp->bio_cmd;
203 		/*
204 		 *  BIO_ORDERED flag may be used by disk drivers to enforce
205 		 *  ordering restrictions, so this flag needs to be cloned.
206 		 *  BIO_UNMAPPED, BIO_VLIST, and BIO_SWAP should be inherited,
207 		 *  to properly indicate which way the buffer is passed.
208 		 *  Other bio flags are not suitable for cloning.
209 		 */
210 		bp2->bio_flags = bp->bio_flags &
211 		    (BIO_ORDERED | BIO_UNMAPPED | BIO_VLIST | BIO_SWAP);
212 		bp2->bio_length = bp->bio_length;
213 		bp2->bio_offset = bp->bio_offset;
214 		bp2->bio_data = bp->bio_data;
215 		bp2->bio_ma = bp->bio_ma;
216 		bp2->bio_ma_n = bp->bio_ma_n;
217 		bp2->bio_ma_offset = bp->bio_ma_offset;
218 		bp2->bio_attribute = bp->bio_attribute;
219 		if (bp->bio_cmd == BIO_ZONE)
220 			bcopy(&bp->bio_zone, &bp2->bio_zone,
221 			    sizeof(bp->bio_zone));
222 #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING)
223 		bp2->bio_track_bp = bp->bio_track_bp;
224 #endif
225 		bp->bio_children++;
226 	}
227 #ifdef KTR
228 	if (KTR_GEOM_ENABLED) {
229 		struct stack st;
230 
231 		CTR2(KTR_GEOM, "g_clone_bio(%p): %p", bp, bp2);
232 		stack_save(&st);
233 		CTRSTACK(KTR_GEOM, &st, 3);
234 	}
235 #endif
236 	return(bp2);
237 }
238 
239 struct bio *
g_duplicate_bio(struct bio * bp)240 g_duplicate_bio(struct bio *bp)
241 {
242 	struct bio *bp2;
243 
244 	bp2 = uma_zalloc(biozone, M_WAITOK | M_ZERO);
245 	bp2->bio_flags = bp->bio_flags & (BIO_UNMAPPED | BIO_VLIST | BIO_SWAP);
246 	bp2->bio_parent = bp;
247 	bp2->bio_cmd = bp->bio_cmd;
248 	bp2->bio_length = bp->bio_length;
249 	bp2->bio_offset = bp->bio_offset;
250 	bp2->bio_data = bp->bio_data;
251 	bp2->bio_ma = bp->bio_ma;
252 	bp2->bio_ma_n = bp->bio_ma_n;
253 	bp2->bio_ma_offset = bp->bio_ma_offset;
254 	bp2->bio_attribute = bp->bio_attribute;
255 	bp->bio_children++;
256 #ifdef KTR
257 	if (KTR_GEOM_ENABLED) {
258 		struct stack st;
259 
260 		CTR2(KTR_GEOM, "g_duplicate_bio(%p): %p", bp, bp2);
261 		stack_save(&st);
262 		CTRSTACK(KTR_GEOM, &st, 3);
263 	}
264 #endif
265 	return(bp2);
266 }
267 
268 void
g_reset_bio(struct bio * bp)269 g_reset_bio(struct bio *bp)
270 {
271 
272 	bzero(bp, sizeof(*bp));
273 }
274 
275 void
g_io_init(void)276 g_io_init(void)
277 {
278 
279 	g_bioq_init(&g_bio_run_down);
280 	g_bioq_init(&g_bio_run_up);
281 	biozone = uma_zcreate("g_bio", sizeof (struct bio),
282 	    NULL, NULL,
283 	    NULL, NULL,
284 	    0, 0);
285 }
286 
287 int
g_io_getattr(const char * attr,struct g_consumer * cp,int * len,void * ptr)288 g_io_getattr(const char *attr, struct g_consumer *cp, int *len, void *ptr)
289 {
290 	struct bio *bp;
291 	int error;
292 
293 	g_trace(G_T_BIO, "bio_getattr(%s)", attr);
294 	bp = g_alloc_bio();
295 	bp->bio_cmd = BIO_GETATTR;
296 	bp->bio_done = NULL;
297 	bp->bio_attribute = attr;
298 	bp->bio_length = *len;
299 	bp->bio_data = ptr;
300 	g_io_request(bp, cp);
301 	error = biowait(bp, "ggetattr");
302 	*len = bp->bio_completed;
303 	g_destroy_bio(bp);
304 	return (error);
305 }
306 
307 int
g_io_zonecmd(struct disk_zone_args * zone_args,struct g_consumer * cp)308 g_io_zonecmd(struct disk_zone_args *zone_args, struct g_consumer *cp)
309 {
310 	struct bio *bp;
311 	int error;
312 
313 	g_trace(G_T_BIO, "bio_zone(%d)", zone_args->zone_cmd);
314 	bp = g_alloc_bio();
315 	bp->bio_cmd = BIO_ZONE;
316 	bp->bio_done = NULL;
317 	/*
318 	 * XXX KDM need to handle report zone data.
319 	 */
320 	bcopy(zone_args, &bp->bio_zone, sizeof(*zone_args));
321 	if (zone_args->zone_cmd == DISK_ZONE_REPORT_ZONES)
322 		bp->bio_length =
323 		    zone_args->zone_params.report.entries_allocated *
324 		    sizeof(struct disk_zone_rep_entry);
325 	else
326 		bp->bio_length = 0;
327 
328 	g_io_request(bp, cp);
329 	error = biowait(bp, "gzone");
330 	bcopy(&bp->bio_zone, zone_args, sizeof(*zone_args));
331 	g_destroy_bio(bp);
332 	return (error);
333 }
334 
335 /*
336  * Send a BIO_SPEEDUP down the stack. This is used to tell the lower layers that
337  * the upper layers have detected a resource shortage. The lower layers are
338  * advised to stop delaying I/O that they might be holding for performance
339  * reasons and to schedule it (non-trims) or complete it successfully (trims) as
340  * quickly as it can. bio_length is the amount of the shortage.  This call
341  * should be non-blocking. bio_resid is used to communicate back if the lower
342  * layers couldn't find bio_length worth of I/O to schedule or discard. A length
343  * of 0 means to do as much as you can (schedule the h/w queues full, discard
344  * all trims). flags are a hint from the upper layers to the lower layers what
345  * operation should be done.
346  */
347 int
g_io_speedup(off_t shortage,u_int flags,size_t * resid,struct g_consumer * cp)348 g_io_speedup(off_t shortage, u_int flags, size_t *resid, struct g_consumer *cp)
349 {
350 	struct bio *bp;
351 	int error;
352 
353 	KASSERT((flags & (BIO_SPEEDUP_TRIM | BIO_SPEEDUP_WRITE)) != 0,
354 	    ("Invalid flags passed to g_io_speedup: %#x", flags));
355 	g_trace(G_T_BIO, "bio_speedup(%s, %jd, %#x)", cp->provider->name,
356 	    (intmax_t)shortage, flags);
357 	bp = g_new_bio();
358 	if (bp == NULL)
359 		return (ENOMEM);
360 	bp->bio_cmd = BIO_SPEEDUP;
361 	bp->bio_length = shortage;
362 	bp->bio_done = NULL;
363 	bp->bio_flags |= flags;
364 	g_io_request(bp, cp);
365 	error = biowait(bp, "gflush");
366 	*resid = bp->bio_resid;
367 	g_destroy_bio(bp);
368 	return (error);
369 }
370 
371 int
g_io_flush(struct g_consumer * cp)372 g_io_flush(struct g_consumer *cp)
373 {
374 	struct bio *bp;
375 	int error;
376 
377 	g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name);
378 	bp = g_alloc_bio();
379 	bp->bio_cmd = BIO_FLUSH;
380 	bp->bio_flags |= BIO_ORDERED;
381 	bp->bio_done = NULL;
382 	bp->bio_attribute = NULL;
383 	bp->bio_offset = cp->provider->mediasize;
384 	bp->bio_length = 0;
385 	bp->bio_data = NULL;
386 	g_io_request(bp, cp);
387 	error = biowait(bp, "gflush");
388 	g_destroy_bio(bp);
389 	return (error);
390 }
391 
392 static int
g_io_check(struct bio * bp)393 g_io_check(struct bio *bp)
394 {
395 	struct g_consumer *cp;
396 	struct g_provider *pp;
397 	off_t excess;
398 	int error;
399 
400 	biotrack(bp, __func__);
401 
402 	cp = bp->bio_from;
403 	pp = bp->bio_to;
404 
405 	/* Fail if access counters dont allow the operation */
406 	switch(bp->bio_cmd) {
407 	case BIO_READ:
408 	case BIO_GETATTR:
409 		if (cp->acr == 0)
410 			return (EPERM);
411 		break;
412 	case BIO_WRITE:
413 	case BIO_DELETE:
414 	case BIO_SPEEDUP:
415 	case BIO_FLUSH:
416 		if (cp->acw == 0)
417 			return (EPERM);
418 		break;
419 	case BIO_ZONE:
420 		if ((bp->bio_zone.zone_cmd == DISK_ZONE_REPORT_ZONES) ||
421 		    (bp->bio_zone.zone_cmd == DISK_ZONE_GET_PARAMS)) {
422 			if (cp->acr == 0)
423 				return (EPERM);
424 		} else if (cp->acw == 0)
425 			return (EPERM);
426 		break;
427 	default:
428 		return (EPERM);
429 	}
430 	/* if provider is marked for error, don't disturb. */
431 	if (pp->error)
432 		return (pp->error);
433 	if (cp->flags & G_CF_ORPHAN)
434 		return (ENXIO);
435 
436 	switch(bp->bio_cmd) {
437 	case BIO_READ:
438 	case BIO_WRITE:
439 	case BIO_DELETE:
440 		/* Zero sectorsize or mediasize is probably a lack of media. */
441 		if (pp->sectorsize == 0 || pp->mediasize == 0)
442 			return (ENXIO);
443 		/* Reject I/O not on sector boundary */
444 		if (bp->bio_offset % pp->sectorsize)
445 			return (EINVAL);
446 		/* Reject I/O not integral sector long */
447 		if (bp->bio_length % pp->sectorsize)
448 			return (EINVAL);
449 		/* Reject requests before or past the end of media. */
450 		if (bp->bio_offset < 0)
451 			return (EIO);
452 		if (bp->bio_offset > pp->mediasize)
453 			return (EIO);
454 
455 		/* Truncate requests to the end of providers media. */
456 		excess = bp->bio_offset + bp->bio_length;
457 		if (excess > bp->bio_to->mediasize) {
458 			KASSERT((bp->bio_flags & BIO_UNMAPPED) == 0 ||
459 			    round_page(bp->bio_ma_offset +
460 			    bp->bio_length) / PAGE_SIZE == bp->bio_ma_n,
461 			    ("excess bio %p too short", bp));
462 			excess -= bp->bio_to->mediasize;
463 			bp->bio_length -= excess;
464 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
465 				bp->bio_ma_n = round_page(bp->bio_ma_offset +
466 				    bp->bio_length) / PAGE_SIZE;
467 			}
468 			if (excess > 0)
469 				CTR3(KTR_GEOM, "g_down truncated bio "
470 				    "%p provider %s by %d", bp,
471 				    bp->bio_to->name, excess);
472 		}
473 
474 		/* Deliver zero length transfers right here. */
475 		if (bp->bio_length == 0) {
476 			CTR2(KTR_GEOM, "g_down terminated 0-length "
477 			    "bp %p provider %s", bp, bp->bio_to->name);
478 			return (0);
479 		}
480 
481 		if ((bp->bio_flags & BIO_UNMAPPED) != 0 &&
482 		    (bp->bio_to->flags & G_PF_ACCEPT_UNMAPPED) == 0 &&
483 		    (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE)) {
484 			if ((error = g_io_transient_map_bio(bp)) >= 0)
485 				return (error);
486 		}
487 		break;
488 	default:
489 		break;
490 	}
491 	return (EJUSTRETURN);
492 }
493 
494 void
g_io_request(struct bio * bp,struct g_consumer * cp)495 g_io_request(struct bio *bp, struct g_consumer *cp)
496 {
497 	struct g_provider *pp;
498 	int direct, error, first;
499 	uint8_t cmd;
500 
501 	biotrack(bp, __func__);
502 
503 	KASSERT(cp != NULL, ("NULL cp in g_io_request"));
504 	KASSERT(bp != NULL, ("NULL bp in g_io_request"));
505 	pp = cp->provider;
506 	KASSERT(pp != NULL, ("consumer not attached in g_io_request"));
507 #ifdef DIAGNOSTIC
508 	KASSERT(bp->bio_driver1 == NULL,
509 	    ("bio_driver1 used by the consumer (geom %s)", cp->geom->name));
510 	KASSERT(bp->bio_driver2 == NULL,
511 	    ("bio_driver2 used by the consumer (geom %s)", cp->geom->name));
512 	KASSERT(bp->bio_pflags == 0,
513 	    ("bio_pflags used by the consumer (geom %s)", cp->geom->name));
514 	/*
515 	 * Remember consumer's private fields, so we can detect if they were
516 	 * modified by the provider.
517 	 */
518 	bp->_bio_caller1 = bp->bio_caller1;
519 	bp->_bio_caller2 = bp->bio_caller2;
520 	bp->_bio_cflags = bp->bio_cflags;
521 #endif
522 
523 	cmd = bp->bio_cmd;
524 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_GETATTR) {
525 		KASSERT(bp->bio_data != NULL,
526 		    ("NULL bp->data in g_io_request(cmd=%hu)", bp->bio_cmd));
527 	}
528 	if (cmd == BIO_DELETE || cmd == BIO_FLUSH || cmd == BIO_SPEEDUP) {
529 		KASSERT(bp->bio_data == NULL,
530 		    ("non-NULL bp->data in g_io_request(cmd=%hu)",
531 		    bp->bio_cmd));
532 	}
533 	if (cmd == BIO_READ || cmd == BIO_WRITE || cmd == BIO_DELETE) {
534 		KASSERT(bp->bio_offset % cp->provider->sectorsize == 0,
535 		    ("wrong offset %jd for sectorsize %u",
536 		    bp->bio_offset, cp->provider->sectorsize));
537 		KASSERT(bp->bio_length % cp->provider->sectorsize == 0,
538 		    ("wrong length %jd for sectorsize %u",
539 		    bp->bio_length, cp->provider->sectorsize));
540 	}
541 
542 	g_trace(G_T_BIO, "bio_request(%p) from %p(%s) to %p(%s) cmd %d",
543 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd);
544 
545 	bp->bio_from = cp;
546 	bp->bio_to = pp;
547 	bp->bio_error = 0;
548 	bp->bio_completed = 0;
549 
550 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
551 	    ("Bio already on queue bp=%p", bp));
552 
553 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
554 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
555 		binuptime(&bp->bio_t0);
556 	else
557 		getbinuptime(&bp->bio_t0);
558 	if (g_collectstats & G_STATS_CONSUMERS)
559 		devstat_start_transaction_bio_t0(cp->stat, bp);
560 	if (g_collectstats & G_STATS_PROVIDERS)
561 		devstat_start_transaction_bio_t0(pp->stat, bp);
562 #ifdef INVARIANTS
563 	atomic_add_int(&cp->nstart, 1);
564 #endif
565 
566 	direct = (cp->flags & G_CF_DIRECT_SEND) != 0 &&
567 	    (pp->flags & G_PF_DIRECT_RECEIVE) != 0 &&
568 	    curthread != g_down_td &&
569 	    ((pp->flags & G_PF_ACCEPT_UNMAPPED) != 0 ||
570 	    (bp->bio_flags & BIO_UNMAPPED) == 0 || THREAD_CAN_SLEEP()) &&
571 	    pace == 0;
572 	if (direct) {
573 		/* Block direct execution if less then half of stack left. */
574 		size_t	st, su;
575 		GET_STACK_USAGE(st, su);
576 		if (su * 2 > st)
577 			direct = 0;
578 	}
579 
580 	if (direct) {
581 		error = g_io_check(bp);
582 		if (error >= 0) {
583 			CTR3(KTR_GEOM, "g_io_request g_io_check on bp %p "
584 			    "provider %s returned %d", bp, bp->bio_to->name,
585 			    error);
586 			g_io_deliver(bp, error);
587 			return;
588 		}
589 		bp->bio_to->geom->start(bp);
590 	} else {
591 		g_bioq_lock(&g_bio_run_down);
592 		first = TAILQ_EMPTY(&g_bio_run_down.bio_queue);
593 		TAILQ_INSERT_TAIL(&g_bio_run_down.bio_queue, bp, bio_queue);
594 		bp->bio_flags |= BIO_ONQUEUE;
595 		g_bio_run_down.bio_queue_length++;
596 		g_bioq_unlock(&g_bio_run_down);
597 		/* Pass it on down. */
598 		if (first)
599 			wakeup(&g_wait_down);
600 	}
601 }
602 
603 void
g_io_deliver(struct bio * bp,int error)604 g_io_deliver(struct bio *bp, int error)
605 {
606 	struct bintime now;
607 	struct g_consumer *cp;
608 	struct g_provider *pp;
609 	struct mtx *mtxp;
610 	int direct, first;
611 
612 	biotrack(bp, __func__);
613 
614 	KASSERT(bp != NULL, ("NULL bp in g_io_deliver"));
615 	pp = bp->bio_to;
616 	KASSERT(pp != NULL, ("NULL bio_to in g_io_deliver"));
617 	cp = bp->bio_from;
618 	if (cp == NULL) {
619 		bp->bio_error = error;
620 		bp->bio_done(bp);
621 		return;
622 	}
623 	KASSERT(cp != NULL, ("NULL bio_from in g_io_deliver"));
624 	KASSERT(cp->geom != NULL, ("NULL bio_from->geom in g_io_deliver"));
625 #ifdef DIAGNOSTIC
626 	/*
627 	 * Some classes - GJournal in particular - can modify bio's
628 	 * private fields while the bio is in transit; G_GEOM_VOLATILE_BIO
629 	 * flag means it's an expected behaviour for that particular geom.
630 	 */
631 	if ((cp->geom->flags & G_GEOM_VOLATILE_BIO) == 0) {
632 		KASSERT(bp->bio_caller1 == bp->_bio_caller1,
633 		    ("bio_caller1 used by the provider %s", pp->name));
634 		KASSERT(bp->bio_caller2 == bp->_bio_caller2,
635 		    ("bio_caller2 used by the provider %s", pp->name));
636 		KASSERT(bp->bio_cflags == bp->_bio_cflags,
637 		    ("bio_cflags used by the provider %s", pp->name));
638 	}
639 #endif
640 	KASSERT(bp->bio_completed >= 0, ("bio_completed can't be less than 0"));
641 	KASSERT(bp->bio_completed <= bp->bio_length,
642 	    ("bio_completed can't be greater than bio_length"));
643 
644 	g_trace(G_T_BIO,
645 "g_io_deliver(%p) from %p(%s) to %p(%s) cmd %d error %d off %jd len %jd",
646 	    bp, cp, cp->geom->name, pp, pp->name, bp->bio_cmd, error,
647 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
648 
649 	KASSERT(!(bp->bio_flags & BIO_ONQUEUE),
650 	    ("Bio already on queue bp=%p", bp));
651 
652 	/*
653 	 * XXX: next two doesn't belong here
654 	 */
655 	bp->bio_bcount = bp->bio_length;
656 	bp->bio_resid = bp->bio_bcount - bp->bio_completed;
657 
658 	direct = (pp->flags & G_PF_DIRECT_SEND) &&
659 		 (cp->flags & G_CF_DIRECT_RECEIVE) &&
660 		 curthread != g_up_td;
661 	if (direct) {
662 		/* Block direct execution if less then half of stack left. */
663 		size_t	st, su;
664 		GET_STACK_USAGE(st, su);
665 		if (su * 2 > st)
666 			direct = 0;
667 	}
668 
669 	/*
670 	 * The statistics collection is lockless, as such, but we
671 	 * can not update one instance of the statistics from more
672 	 * than one thread at a time, so grab the lock first.
673 	 */
674 	if ((g_collectstats & G_STATS_CONSUMERS) != 0 ||
675 	    ((g_collectstats & G_STATS_PROVIDERS) != 0 && pp->stat != NULL))
676 		binuptime(&now);
677 	mtxp = mtx_pool_find(mtxpool_sleep, pp);
678 	mtx_lock(mtxp);
679 	if (g_collectstats & G_STATS_PROVIDERS)
680 		devstat_end_transaction_bio_bt(pp->stat, bp, &now);
681 	if (g_collectstats & G_STATS_CONSUMERS)
682 		devstat_end_transaction_bio_bt(cp->stat, bp, &now);
683 #ifdef INVARIANTS
684 	cp->nend++;
685 #endif
686 	mtx_unlock(mtxp);
687 
688 	if (error != ENOMEM) {
689 		bp->bio_error = error;
690 		if (direct) {
691 			biodone(bp);
692 		} else {
693 			g_bioq_lock(&g_bio_run_up);
694 			first = TAILQ_EMPTY(&g_bio_run_up.bio_queue);
695 			TAILQ_INSERT_TAIL(&g_bio_run_up.bio_queue, bp, bio_queue);
696 			bp->bio_flags |= BIO_ONQUEUE;
697 			g_bio_run_up.bio_queue_length++;
698 			g_bioq_unlock(&g_bio_run_up);
699 			if (first)
700 				wakeup(&g_wait_up);
701 		}
702 		return;
703 	}
704 
705 	if (bootverbose)
706 		printf("ENOMEM %p on %p(%s)\n", bp, pp, pp->name);
707 	atomic_add_long(&nomem_count, 1);	/* Rare event, but no locks held */
708 	bp->bio_children = 0;
709 	bp->bio_inbed = 0;
710 	bp->bio_driver1 = NULL;
711 	bp->bio_driver2 = NULL;
712 	bp->bio_pflags = 0;
713 	g_io_request(bp, cp);
714 	pace = 1;
715 	return;
716 }
717 
718 SYSCTL_DECL(_kern_geom);
719 
720 static long transient_maps;
721 SYSCTL_LONG(_kern_geom, OID_AUTO, transient_maps, CTLFLAG_RD,
722     &transient_maps, 0,
723     "Total count of the transient mapping requests");
724 u_int transient_map_retries = 10;
725 SYSCTL_UINT(_kern_geom, OID_AUTO, transient_map_retries, CTLFLAG_RW,
726     &transient_map_retries, 0,
727     "Max count of retries used before giving up on creating transient map");
728 int transient_map_hard_failures;
729 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_hard_failures, CTLFLAG_RD,
730     &transient_map_hard_failures, 0,
731     "Failures to establish the transient mapping due to retry attempts "
732     "exhausted");
733 int transient_map_soft_failures;
734 SYSCTL_INT(_kern_geom, OID_AUTO, transient_map_soft_failures, CTLFLAG_RD,
735     &transient_map_soft_failures, 0,
736     "Count of retried failures to establish the transient mapping");
737 int inflight_transient_maps;
738 SYSCTL_INT(_kern_geom, OID_AUTO, inflight_transient_maps, CTLFLAG_RD,
739     &inflight_transient_maps, 0,
740     "Current count of the active transient maps");
741 SYSCTL_ULONG(_kern_geom, OID_AUTO, nomem_count, CTLFLAG_RD,
742     &nomem_count, 0,
743     "Total count of requests completed with status of ENOMEM");
744 SYSCTL_ULONG(_kern_geom, OID_AUTO, pause_count, CTLFLAG_RD,
745     &pause_count, 0,
746     "Total count of requests stalled due to low memory in g_down");
747 
748 static int
g_io_transient_map_bio(struct bio * bp)749 g_io_transient_map_bio(struct bio *bp)
750 {
751 	vm_offset_t addr;
752 	long size;
753 	u_int retried;
754 
755 	KASSERT(unmapped_buf_allowed, ("unmapped disabled"));
756 
757 	size = round_page(bp->bio_ma_offset + bp->bio_length);
758 	KASSERT(size / PAGE_SIZE == bp->bio_ma_n, ("Bio too short %p", bp));
759 	addr = 0;
760 	retried = 0;
761 	atomic_add_long(&transient_maps, 1);
762 retry:
763 	if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) {
764 		if (transient_map_retries != 0 &&
765 		    retried >= transient_map_retries) {
766 			CTR2(KTR_GEOM, "g_down cannot map bp %p provider %s",
767 			    bp, bp->bio_to->name);
768 			atomic_add_int(&transient_map_hard_failures, 1);
769 			return (EDEADLK/* XXXKIB */);
770 		} else {
771 			/*
772 			 * Naive attempt to quisce the I/O to get more
773 			 * in-flight requests completed and defragment
774 			 * the transient_arena.
775 			 */
776 			CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d",
777 			    bp, bp->bio_to->name, retried);
778 			pause("g_d_tra", hz / 10);
779 			retried++;
780 			atomic_add_int(&transient_map_soft_failures, 1);
781 			goto retry;
782 		}
783 	}
784 	atomic_add_int(&inflight_transient_maps, 1);
785 	pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size));
786 	bp->bio_data = (caddr_t)addr + bp->bio_ma_offset;
787 	bp->bio_flags |= BIO_TRANSIENT_MAPPING;
788 	bp->bio_flags &= ~BIO_UNMAPPED;
789 	return (EJUSTRETURN);
790 }
791 
792 void
g_io_schedule_down(struct thread * tp __unused)793 g_io_schedule_down(struct thread *tp __unused)
794 {
795 	struct bio *bp;
796 	int error;
797 
798 	for(;;) {
799 		g_bioq_lock(&g_bio_run_down);
800 		bp = g_bioq_first(&g_bio_run_down);
801 		if (bp == NULL) {
802 			CTR0(KTR_GEOM, "g_down going to sleep");
803 			msleep(&g_wait_down, &g_bio_run_down.bio_queue_lock,
804 			    PRIBIO | PDROP, "-", 0);
805 			continue;
806 		}
807 		CTR0(KTR_GEOM, "g_down has work to do");
808 		g_bioq_unlock(&g_bio_run_down);
809 		biotrack(bp, __func__);
810 		if (pace != 0) {
811 			/*
812 			 * There has been at least one memory allocation failure
813 			 * since the last I/O completed. Pause 1ms to give the
814 			 * system a chance to free up memory. Pause time is not
815 			 * scaled to the number of I/O failures since they tend
816 			 * to cluster and the number is not predictive of how
817 			 * long a pause is needed.
818 			 *
819 			 * Older versions had a longer pause, which limited the
820 			 * IOPS to 10, which prolonged memory shortages that could
821 			 * be alleviated by I/O completing since it eliminated
822 			 * direct dispatch as well.
823 			 *
824 			 * XXX This pacing is really lame. It needs to be solved
825 			 * by other methods. This is OK only because the worst
826 			 * case scenario is so rare. In the worst case scenario
827 			 * all memory is tied up waiting for I/O to complete
828 			 * which can never happen since we can't allocate bios
829 			 * for that I/O.
830 			 */
831 			CTR0(KTR_GEOM, "g_down pacing self");
832 			pause_count++;		/* g_down has only one thread */
833 			pause_sbt("g_down", SBT_1MS, 0, 0);
834 			pace = 0;
835 		}
836 		CTR2(KTR_GEOM, "g_down processing bp %p provider %s", bp,
837 		    bp->bio_to->name);
838 		error = g_io_check(bp);
839 		if (error >= 0) {
840 			CTR3(KTR_GEOM, "g_down g_io_check on bp %p provider "
841 			    "%s returned %d", bp, bp->bio_to->name, error);
842 			g_io_deliver(bp, error);
843 			continue;
844 		}
845 		THREAD_NO_SLEEPING();
846 		CTR4(KTR_GEOM, "g_down starting bp %p provider %s off %ld "
847 		    "len %ld", bp, bp->bio_to->name, bp->bio_offset,
848 		    bp->bio_length);
849 		bp->bio_to->geom->start(bp);
850 		THREAD_SLEEPING_OK();
851 	}
852 }
853 
854 void
g_io_schedule_up(struct thread * tp __unused)855 g_io_schedule_up(struct thread *tp __unused)
856 {
857 	struct bio *bp;
858 
859 	for(;;) {
860 		g_bioq_lock(&g_bio_run_up);
861 		bp = g_bioq_first(&g_bio_run_up);
862 		if (bp == NULL) {
863 			CTR0(KTR_GEOM, "g_up going to sleep");
864 			msleep(&g_wait_up, &g_bio_run_up.bio_queue_lock,
865 			    PRIBIO | PDROP, "-", 0);
866 			continue;
867 		}
868 		g_bioq_unlock(&g_bio_run_up);
869 		THREAD_NO_SLEEPING();
870 		CTR4(KTR_GEOM, "g_up biodone bp %p provider %s off "
871 		    "%jd len %ld", bp, bp->bio_to->name,
872 		    bp->bio_offset, bp->bio_length);
873 		biodone(bp);
874 		THREAD_SLEEPING_OK();
875 	}
876 }
877 
878 void *
g_read_data(struct g_consumer * cp,off_t offset,off_t length,int * error)879 g_read_data(struct g_consumer *cp, off_t offset, off_t length, int *error)
880 {
881 	struct bio *bp;
882 	void *ptr;
883 	int errorc;
884 
885 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
886 	    length <= maxphys, ("g_read_data(): invalid length %jd",
887 	    (intmax_t)length));
888 
889 	bp = g_alloc_bio();
890 	bp->bio_cmd = BIO_READ;
891 	bp->bio_done = NULL;
892 	bp->bio_offset = offset;
893 	bp->bio_length = length;
894 	ptr = g_malloc(length, M_WAITOK);
895 	bp->bio_data = ptr;
896 	g_io_request(bp, cp);
897 	errorc = biowait(bp, "gread");
898 	if (errorc == 0 && bp->bio_completed != length)
899 		errorc = EIO;
900 	if (error != NULL)
901 		*error = errorc;
902 	g_destroy_bio(bp);
903 	if (errorc) {
904 		g_free(ptr);
905 		ptr = NULL;
906 	}
907 	return (ptr);
908 }
909 
910 /*
911  * A read function for use by ffs_sbget when used by GEOM-layer routines.
912  */
913 int
g_use_g_read_data(void * devfd,off_t loc,void ** bufp,int size)914 g_use_g_read_data(void *devfd, off_t loc, void **bufp, int size)
915 {
916 	struct g_consumer *cp;
917 
918 	KASSERT(*bufp == NULL,
919 	    ("g_use_g_read_data: non-NULL *bufp %p\n", *bufp));
920 
921 	cp = (struct g_consumer *)devfd;
922 	/*
923 	 * Take care not to issue an invalid I/O request. The offset of
924 	 * the superblock candidate must be multiples of the provider's
925 	 * sector size, otherwise an FFS can't exist on the provider
926 	 * anyway.
927 	 */
928 	if (loc % cp->provider->sectorsize != 0)
929 		return (ENOENT);
930 	*bufp = g_read_data(cp, loc, size, NULL);
931 	if (*bufp == NULL)
932 		return (ENOENT);
933 	return (0);
934 }
935 
936 int
g_write_data(struct g_consumer * cp,off_t offset,void * ptr,off_t length)937 g_write_data(struct g_consumer *cp, off_t offset, void *ptr, off_t length)
938 {
939 	struct bio *bp;
940 	int error;
941 
942 	KASSERT(length > 0 && length >= cp->provider->sectorsize &&
943 	    length <= maxphys, ("g_write_data(): invalid length %jd",
944 	    (intmax_t)length));
945 
946 	bp = g_alloc_bio();
947 	bp->bio_cmd = BIO_WRITE;
948 	bp->bio_done = NULL;
949 	bp->bio_offset = offset;
950 	bp->bio_length = length;
951 	bp->bio_data = ptr;
952 	g_io_request(bp, cp);
953 	error = biowait(bp, "gwrite");
954 	if (error == 0 && bp->bio_completed != length)
955 		error = EIO;
956 	g_destroy_bio(bp);
957 	return (error);
958 }
959 
960 /*
961  * A write function for use by ffs_sbput when used by GEOM-layer routines.
962  */
963 int
g_use_g_write_data(void * devfd,off_t loc,void * buf,int size)964 g_use_g_write_data(void *devfd, off_t loc, void *buf, int size)
965 {
966 
967 	return (g_write_data((struct g_consumer *)devfd, loc, buf, size));
968 }
969 
970 int
g_delete_data(struct g_consumer * cp,off_t offset,off_t length)971 g_delete_data(struct g_consumer *cp, off_t offset, off_t length)
972 {
973 	struct bio *bp;
974 	int error;
975 
976 	KASSERT(length > 0 && length >= cp->provider->sectorsize,
977 	    ("g_delete_data(): invalid length %jd", (intmax_t)length));
978 
979 	bp = g_alloc_bio();
980 	bp->bio_cmd = BIO_DELETE;
981 	bp->bio_done = NULL;
982 	bp->bio_offset = offset;
983 	bp->bio_length = length;
984 	bp->bio_data = NULL;
985 	g_io_request(bp, cp);
986 	error = biowait(bp, "gdelete");
987 	if (error == 0 && bp->bio_completed != length)
988 		error = EIO;
989 	g_destroy_bio(bp);
990 	return (error);
991 }
992 
993 void
g_print_bio(const char * prefix,const struct bio * bp,const char * fmtsuffix,...)994 g_print_bio(const char *prefix, const struct bio *bp, const char *fmtsuffix,
995     ...)
996 {
997 #ifndef PRINTF_BUFR_SIZE
998 #define PRINTF_BUFR_SIZE 64
999 #endif
1000 	char bufr[PRINTF_BUFR_SIZE];
1001 	struct sbuf sb, *sbp __unused;
1002 	va_list ap;
1003 
1004 	sbp = sbuf_new(&sb, bufr, sizeof(bufr), SBUF_FIXEDLEN);
1005 	KASSERT(sbp != NULL, ("sbuf_new misused?"));
1006 
1007 	sbuf_set_drain(&sb, sbuf_printf_drain, NULL);
1008 
1009 	sbuf_cat(&sb, prefix);
1010 	g_format_bio(&sb, bp);
1011 
1012 	va_start(ap, fmtsuffix);
1013 	sbuf_vprintf(&sb, fmtsuffix, ap);
1014 	va_end(ap);
1015 
1016 	sbuf_nl_terminate(&sb);
1017 
1018 	sbuf_finish(&sb);
1019 	sbuf_delete(&sb);
1020 }
1021 
1022 void
g_format_bio(struct sbuf * sb,const struct bio * bp)1023 g_format_bio(struct sbuf *sb, const struct bio *bp)
1024 {
1025 	const char *pname, *cmd = NULL;
1026 
1027 	if (bp->bio_to != NULL)
1028 		pname = bp->bio_to->name;
1029 	else if (bp->bio_parent != NULL && bp->bio_parent->bio_to != NULL)
1030 		pname = bp->bio_parent->bio_to->name;
1031 	else
1032 		pname = "[unknown]";
1033 
1034 	switch (bp->bio_cmd) {
1035 	case BIO_GETATTR:
1036 		cmd = "GETATTR";
1037 		sbuf_printf(sb, "%s[%s(attr=%s)]", pname, cmd,
1038 		    bp->bio_attribute);
1039 		return;
1040 	case BIO_FLUSH:
1041 		cmd = "FLUSH";
1042 		sbuf_printf(sb, "%s[%s]", pname, cmd);
1043 		return;
1044 	case BIO_ZONE: {
1045 		char *subcmd = NULL;
1046 		cmd = "ZONE";
1047 		switch (bp->bio_zone.zone_cmd) {
1048 		case DISK_ZONE_OPEN:
1049 			subcmd = "OPEN";
1050 			break;
1051 		case DISK_ZONE_CLOSE:
1052 			subcmd = "CLOSE";
1053 			break;
1054 		case DISK_ZONE_FINISH:
1055 			subcmd = "FINISH";
1056 			break;
1057 		case DISK_ZONE_RWP:
1058 			subcmd = "RWP";
1059 			break;
1060 		case DISK_ZONE_REPORT_ZONES:
1061 			subcmd = "REPORT ZONES";
1062 			break;
1063 		case DISK_ZONE_GET_PARAMS:
1064 			subcmd = "GET PARAMS";
1065 			break;
1066 		default:
1067 			subcmd = "UNKNOWN";
1068 			break;
1069 		}
1070 		sbuf_printf(sb, "%s[%s,%s]", pname, cmd, subcmd);
1071 		return;
1072 	}
1073 	case BIO_READ:
1074 		cmd = "READ";
1075 		break;
1076 	case BIO_WRITE:
1077 		cmd = "WRITE";
1078 		break;
1079 	case BIO_DELETE:
1080 		cmd = "DELETE";
1081 		break;
1082 	default:
1083 		cmd = "UNKNOWN";
1084 		sbuf_printf(sb, "%s[%s()]", pname, cmd);
1085 		return;
1086 	}
1087 	sbuf_printf(sb, "%s[%s(offset=%jd, length=%jd)]", pname, cmd,
1088 	    (intmax_t)bp->bio_offset, (intmax_t)bp->bio_length);
1089 }
1090