xref: /freebsd/sys/dev/md/md.c (revision f677a9e2672665f4eb3dd4111c07ee8f1f954262)
1 /*-
2  * ----------------------------------------------------------------------------
3  * "THE BEER-WARE LICENSE" (Revision 42):
4  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
5  * can do whatever you want with this stuff. If we meet some day, and you think
6  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
7  * ----------------------------------------------------------------------------
8  *
9  * $FreeBSD$
10  *
11  */
12 
13 /*-
14  * The following functions are based in the vn(4) driver: mdstart_swap(),
15  * mdstart_vnode(), mdcreate_swap(), mdcreate_vnode() and mddestroy(),
16  * and as such under the following copyright:
17  *
18  * Copyright (c) 1988 University of Utah.
19  * Copyright (c) 1990, 1993
20  *	The Regents of the University of California.  All rights reserved.
21  * Copyright (c) 2013 The FreeBSD Foundation
22  * All rights reserved.
23  *
24  * This code is derived from software contributed to Berkeley by
25  * the Systems Programming Group of the University of Utah Computer
26  * Science Department.
27  *
28  * Portions of this software were developed by Konstantin Belousov
29  * under sponsorship from the FreeBSD Foundation.
30  *
31  * Redistribution and use in source and binary forms, with or without
32  * modification, are permitted provided that the following conditions
33  * are met:
34  * 1. Redistributions of source code must retain the above copyright
35  *    notice, this list of conditions and the following disclaimer.
36  * 2. Redistributions in binary form must reproduce the above copyright
37  *    notice, this list of conditions and the following disclaimer in the
38  *    documentation and/or other materials provided with the distribution.
39  * 4. Neither the name of the University nor the names of its contributors
40  *    may be used to endorse or promote products derived from this software
41  *    without specific prior written permission.
42  *
43  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
44  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
45  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
46  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
47  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
48  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
49  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
50  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
51  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
52  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
53  * SUCH DAMAGE.
54  *
55  * from: Utah Hdr: vn.c 1.13 94/04/02
56  *
57  *	from: @(#)vn.c	8.6 (Berkeley) 4/1/94
58  * From: src/sys/dev/vn/vn.c,v 1.122 2000/12/16 16:06:03
59  */
60 
61 #include "opt_geom.h"
62 #include "opt_md.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bio.h>
67 #include <sys/buf.h>
68 #include <sys/conf.h>
69 #include <sys/devicestat.h>
70 #include <sys/fcntl.h>
71 #include <sys/kernel.h>
72 #include <sys/kthread.h>
73 #include <sys/limits.h>
74 #include <sys/linker.h>
75 #include <sys/lock.h>
76 #include <sys/malloc.h>
77 #include <sys/mdioctl.h>
78 #include <sys/mount.h>
79 #include <sys/mutex.h>
80 #include <sys/sx.h>
81 #include <sys/namei.h>
82 #include <sys/proc.h>
83 #include <sys/queue.h>
84 #include <sys/rwlock.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/sf_buf.h>
88 #include <sys/sysctl.h>
89 #include <sys/vnode.h>
90 
91 #include <geom/geom.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_param.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pager.h>
98 #include <vm/swap_pager.h>
99 #include <vm/uma.h>
100 
101 #define MD_MODVER 1
102 
103 #define MD_SHUTDOWN	0x10000		/* Tell worker thread to terminate. */
104 #define	MD_EXITING	0x20000		/* Worker thread is exiting. */
105 
106 #ifndef MD_NSECT
107 #define MD_NSECT (10000 * 2)
108 #endif
109 
110 static MALLOC_DEFINE(M_MD, "md_disk", "Memory Disk");
111 static MALLOC_DEFINE(M_MDSECT, "md_sectors", "Memory Disk Sectors");
112 
113 static int md_debug;
114 SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0,
115     "Enable md(4) debug messages");
116 static int md_malloc_wait;
117 SYSCTL_INT(_vm, OID_AUTO, md_malloc_wait, CTLFLAG_RW, &md_malloc_wait, 0,
118     "Allow malloc to wait for memory allocations");
119 
120 #if defined(MD_ROOT) && !defined(MD_ROOT_FSTYPE)
121 #define	MD_ROOT_FSTYPE	"ufs"
122 #endif
123 
124 #if defined(MD_ROOT) && defined(MD_ROOT_SIZE)
125 /*
126  * Preloaded image gets put here.
127  * Applications that patch the object with the image can determine
128  * the size looking at the start and end markers (strings),
129  * so we want them contiguous.
130  */
131 static struct {
132 	u_char start[MD_ROOT_SIZE*1024];
133 	u_char end[128];
134 } mfs_root = {
135 	.start = "MFS Filesystem goes here",
136 	.end = "MFS Filesystem had better STOP here",
137 };
138 #endif
139 
140 static g_init_t g_md_init;
141 static g_fini_t g_md_fini;
142 static g_start_t g_md_start;
143 static g_access_t g_md_access;
144 static void g_md_dumpconf(struct sbuf *sb, const char *indent,
145     struct g_geom *gp, struct g_consumer *cp __unused, struct g_provider *pp);
146 
147 static struct cdev *status_dev = 0;
148 static struct sx md_sx;
149 static struct unrhdr *md_uh;
150 
151 static d_ioctl_t mdctlioctl;
152 
153 static struct cdevsw mdctl_cdevsw = {
154 	.d_version =	D_VERSION,
155 	.d_ioctl =	mdctlioctl,
156 	.d_name =	MD_NAME,
157 };
158 
159 struct g_class g_md_class = {
160 	.name = "MD",
161 	.version = G_VERSION,
162 	.init = g_md_init,
163 	.fini = g_md_fini,
164 	.start = g_md_start,
165 	.access = g_md_access,
166 	.dumpconf = g_md_dumpconf,
167 };
168 
169 DECLARE_GEOM_CLASS(g_md_class, g_md);
170 
171 
172 static LIST_HEAD(, md_s) md_softc_list = LIST_HEAD_INITIALIZER(md_softc_list);
173 
174 #define NINDIR	(PAGE_SIZE / sizeof(uintptr_t))
175 #define NMASK	(NINDIR-1)
176 static int nshift;
177 
178 static int md_vnode_pbuf_freecnt;
179 
180 struct indir {
181 	uintptr_t	*array;
182 	u_int		total;
183 	u_int		used;
184 	u_int		shift;
185 };
186 
187 struct md_s {
188 	int unit;
189 	LIST_ENTRY(md_s) list;
190 	struct bio_queue_head bio_queue;
191 	struct mtx queue_mtx;
192 	struct cdev *dev;
193 	enum md_types type;
194 	off_t mediasize;
195 	unsigned sectorsize;
196 	unsigned opencount;
197 	unsigned fwheads;
198 	unsigned fwsectors;
199 	unsigned flags;
200 	char name[20];
201 	struct proc *procp;
202 	struct g_geom *gp;
203 	struct g_provider *pp;
204 	int (*start)(struct md_s *sc, struct bio *bp);
205 	struct devstat *devstat;
206 
207 	/* MD_MALLOC related fields */
208 	struct indir *indir;
209 	uma_zone_t uma;
210 
211 	/* MD_PRELOAD related fields */
212 	u_char *pl_ptr;
213 	size_t pl_len;
214 
215 	/* MD_VNODE related fields */
216 	struct vnode *vnode;
217 	char file[PATH_MAX];
218 	struct ucred *cred;
219 
220 	/* MD_SWAP related fields */
221 	vm_object_t object;
222 };
223 
224 static struct indir *
225 new_indir(u_int shift)
226 {
227 	struct indir *ip;
228 
229 	ip = malloc(sizeof *ip, M_MD, (md_malloc_wait ? M_WAITOK : M_NOWAIT)
230 	    | M_ZERO);
231 	if (ip == NULL)
232 		return (NULL);
233 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
234 	    M_MDSECT, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO);
235 	if (ip->array == NULL) {
236 		free(ip, M_MD);
237 		return (NULL);
238 	}
239 	ip->total = NINDIR;
240 	ip->shift = shift;
241 	return (ip);
242 }
243 
244 static void
245 del_indir(struct indir *ip)
246 {
247 
248 	free(ip->array, M_MDSECT);
249 	free(ip, M_MD);
250 }
251 
252 static void
253 destroy_indir(struct md_s *sc, struct indir *ip)
254 {
255 	int i;
256 
257 	for (i = 0; i < NINDIR; i++) {
258 		if (!ip->array[i])
259 			continue;
260 		if (ip->shift)
261 			destroy_indir(sc, (struct indir*)(ip->array[i]));
262 		else if (ip->array[i] > 255)
263 			uma_zfree(sc->uma, (void *)(ip->array[i]));
264 	}
265 	del_indir(ip);
266 }
267 
268 /*
269  * This function does the math and allocates the top level "indir" structure
270  * for a device of "size" sectors.
271  */
272 
273 static struct indir *
274 dimension(off_t size)
275 {
276 	off_t rcnt;
277 	struct indir *ip;
278 	int layer;
279 
280 	rcnt = size;
281 	layer = 0;
282 	while (rcnt > NINDIR) {
283 		rcnt /= NINDIR;
284 		layer++;
285 	}
286 
287 	/*
288 	 * XXX: the top layer is probably not fully populated, so we allocate
289 	 * too much space for ip->array in here.
290 	 */
291 	ip = malloc(sizeof *ip, M_MD, M_WAITOK | M_ZERO);
292 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
293 	    M_MDSECT, M_WAITOK | M_ZERO);
294 	ip->total = NINDIR;
295 	ip->shift = layer * nshift;
296 	return (ip);
297 }
298 
299 /*
300  * Read a given sector
301  */
302 
303 static uintptr_t
304 s_read(struct indir *ip, off_t offset)
305 {
306 	struct indir *cip;
307 	int idx;
308 	uintptr_t up;
309 
310 	if (md_debug > 1)
311 		printf("s_read(%jd)\n", (intmax_t)offset);
312 	up = 0;
313 	for (cip = ip; cip != NULL;) {
314 		if (cip->shift) {
315 			idx = (offset >> cip->shift) & NMASK;
316 			up = cip->array[idx];
317 			cip = (struct indir *)up;
318 			continue;
319 		}
320 		idx = offset & NMASK;
321 		return (cip->array[idx]);
322 	}
323 	return (0);
324 }
325 
326 /*
327  * Write a given sector, prune the tree if the value is 0
328  */
329 
330 static int
331 s_write(struct indir *ip, off_t offset, uintptr_t ptr)
332 {
333 	struct indir *cip, *lip[10];
334 	int idx, li;
335 	uintptr_t up;
336 
337 	if (md_debug > 1)
338 		printf("s_write(%jd, %p)\n", (intmax_t)offset, (void *)ptr);
339 	up = 0;
340 	li = 0;
341 	cip = ip;
342 	for (;;) {
343 		lip[li++] = cip;
344 		if (cip->shift) {
345 			idx = (offset >> cip->shift) & NMASK;
346 			up = cip->array[idx];
347 			if (up != 0) {
348 				cip = (struct indir *)up;
349 				continue;
350 			}
351 			/* Allocate branch */
352 			cip->array[idx] =
353 			    (uintptr_t)new_indir(cip->shift - nshift);
354 			if (cip->array[idx] == 0)
355 				return (ENOSPC);
356 			cip->used++;
357 			up = cip->array[idx];
358 			cip = (struct indir *)up;
359 			continue;
360 		}
361 		/* leafnode */
362 		idx = offset & NMASK;
363 		up = cip->array[idx];
364 		if (up != 0)
365 			cip->used--;
366 		cip->array[idx] = ptr;
367 		if (ptr != 0)
368 			cip->used++;
369 		break;
370 	}
371 	if (cip->used != 0 || li == 1)
372 		return (0);
373 	li--;
374 	while (cip->used == 0 && cip != ip) {
375 		li--;
376 		idx = (offset >> lip[li]->shift) & NMASK;
377 		up = lip[li]->array[idx];
378 		KASSERT(up == (uintptr_t)cip, ("md screwed up"));
379 		del_indir(cip);
380 		lip[li]->array[idx] = 0;
381 		lip[li]->used--;
382 		cip = lip[li];
383 	}
384 	return (0);
385 }
386 
387 
388 static int
389 g_md_access(struct g_provider *pp, int r, int w, int e)
390 {
391 	struct md_s *sc;
392 
393 	sc = pp->geom->softc;
394 	if (sc == NULL) {
395 		if (r <= 0 && w <= 0 && e <= 0)
396 			return (0);
397 		return (ENXIO);
398 	}
399 	r += pp->acr;
400 	w += pp->acw;
401 	e += pp->ace;
402 	if ((sc->flags & MD_READONLY) != 0 && w > 0)
403 		return (EROFS);
404 	if ((pp->acr + pp->acw + pp->ace) == 0 && (r + w + e) > 0) {
405 		sc->opencount = 1;
406 	} else if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0) {
407 		sc->opencount = 0;
408 	}
409 	return (0);
410 }
411 
412 static void
413 g_md_start(struct bio *bp)
414 {
415 	struct md_s *sc;
416 
417 	sc = bp->bio_to->geom->softc;
418 	if ((bp->bio_cmd == BIO_READ) || (bp->bio_cmd == BIO_WRITE))
419 		devstat_start_transaction_bio(sc->devstat, bp);
420 	mtx_lock(&sc->queue_mtx);
421 	bioq_disksort(&sc->bio_queue, bp);
422 	mtx_unlock(&sc->queue_mtx);
423 	wakeup(sc);
424 }
425 
426 #define	MD_MALLOC_MOVE_ZERO	1
427 #define	MD_MALLOC_MOVE_FILL	2
428 #define	MD_MALLOC_MOVE_READ	3
429 #define	MD_MALLOC_MOVE_WRITE	4
430 #define	MD_MALLOC_MOVE_CMP	5
431 
432 static int
433 md_malloc_move(vm_page_t **mp, int *ma_offs, unsigned sectorsize,
434     void *ptr, u_char fill, int op)
435 {
436 	struct sf_buf *sf;
437 	vm_page_t m, *mp1;
438 	char *p, first;
439 	off_t *uc;
440 	unsigned n;
441 	int error, i, ma_offs1, sz, first_read;
442 
443 	m = NULL;
444 	error = 0;
445 	sf = NULL;
446 	/* if (op == MD_MALLOC_MOVE_CMP) { gcc */
447 		first = 0;
448 		first_read = 0;
449 		uc = ptr;
450 		mp1 = *mp;
451 		ma_offs1 = *ma_offs;
452 	/* } */
453 	sched_pin();
454 	for (n = sectorsize; n != 0; n -= sz) {
455 		sz = imin(PAGE_SIZE - *ma_offs, n);
456 		if (m != **mp) {
457 			if (sf != NULL)
458 				sf_buf_free(sf);
459 			m = **mp;
460 			sf = sf_buf_alloc(m, SFB_CPUPRIVATE |
461 			    (md_malloc_wait ? 0 : SFB_NOWAIT));
462 			if (sf == NULL) {
463 				error = ENOMEM;
464 				break;
465 			}
466 		}
467 		p = (char *)sf_buf_kva(sf) + *ma_offs;
468 		switch (op) {
469 		case MD_MALLOC_MOVE_ZERO:
470 			bzero(p, sz);
471 			break;
472 		case MD_MALLOC_MOVE_FILL:
473 			memset(p, fill, sz);
474 			break;
475 		case MD_MALLOC_MOVE_READ:
476 			bcopy(ptr, p, sz);
477 			cpu_flush_dcache(p, sz);
478 			break;
479 		case MD_MALLOC_MOVE_WRITE:
480 			bcopy(p, ptr, sz);
481 			break;
482 		case MD_MALLOC_MOVE_CMP:
483 			for (i = 0; i < sz; i++, p++) {
484 				if (!first_read) {
485 					*uc = (u_char)*p;
486 					first = *p;
487 					first_read = 1;
488 				} else if (*p != first) {
489 					error = EDOOFUS;
490 					break;
491 				}
492 			}
493 			break;
494 		default:
495 			KASSERT(0, ("md_malloc_move unknown op %d\n", op));
496 			break;
497 		}
498 		if (error != 0)
499 			break;
500 		*ma_offs += sz;
501 		*ma_offs %= PAGE_SIZE;
502 		if (*ma_offs == 0)
503 			(*mp)++;
504 		ptr = (char *)ptr + sz;
505 	}
506 
507 	if (sf != NULL)
508 		sf_buf_free(sf);
509 	sched_unpin();
510 	if (op == MD_MALLOC_MOVE_CMP && error != 0) {
511 		*mp = mp1;
512 		*ma_offs = ma_offs1;
513 	}
514 	return (error);
515 }
516 
517 static int
518 mdstart_malloc(struct md_s *sc, struct bio *bp)
519 {
520 	u_char *dst;
521 	vm_page_t *m;
522 	int i, error, error1, ma_offs, notmapped;
523 	off_t secno, nsec, uc;
524 	uintptr_t sp, osp;
525 
526 	switch (bp->bio_cmd) {
527 	case BIO_READ:
528 	case BIO_WRITE:
529 	case BIO_DELETE:
530 		break;
531 	default:
532 		return (EOPNOTSUPP);
533 	}
534 
535 	notmapped = (bp->bio_flags & BIO_UNMAPPED) != 0;
536 	if (notmapped) {
537 		m = bp->bio_ma;
538 		ma_offs = bp->bio_ma_offset;
539 		dst = NULL;
540 	} else {
541 		dst = bp->bio_data;
542 	}
543 
544 	nsec = bp->bio_length / sc->sectorsize;
545 	secno = bp->bio_offset / sc->sectorsize;
546 	error = 0;
547 	while (nsec--) {
548 		osp = s_read(sc->indir, secno);
549 		if (bp->bio_cmd == BIO_DELETE) {
550 			if (osp != 0)
551 				error = s_write(sc->indir, secno, 0);
552 		} else if (bp->bio_cmd == BIO_READ) {
553 			if (osp == 0) {
554 				if (notmapped) {
555 					error = md_malloc_move(&m, &ma_offs,
556 					    sc->sectorsize, NULL, 0,
557 					    MD_MALLOC_MOVE_ZERO);
558 				} else
559 					bzero(dst, sc->sectorsize);
560 			} else if (osp <= 255) {
561 				if (notmapped) {
562 					error = md_malloc_move(&m, &ma_offs,
563 					    sc->sectorsize, NULL, osp,
564 					    MD_MALLOC_MOVE_FILL);
565 				} else
566 					memset(dst, osp, sc->sectorsize);
567 			} else {
568 				if (notmapped) {
569 					error = md_malloc_move(&m, &ma_offs,
570 					    sc->sectorsize, (void *)osp, 0,
571 					    MD_MALLOC_MOVE_READ);
572 				} else {
573 					bcopy((void *)osp, dst, sc->sectorsize);
574 					cpu_flush_dcache(dst, sc->sectorsize);
575 				}
576 			}
577 			osp = 0;
578 		} else if (bp->bio_cmd == BIO_WRITE) {
579 			if (sc->flags & MD_COMPRESS) {
580 				if (notmapped) {
581 					error1 = md_malloc_move(&m, &ma_offs,
582 					    sc->sectorsize, &uc, 0,
583 					    MD_MALLOC_MOVE_CMP);
584 					i = error1 == 0 ? sc->sectorsize : 0;
585 				} else {
586 					uc = dst[0];
587 					for (i = 1; i < sc->sectorsize; i++) {
588 						if (dst[i] != uc)
589 							break;
590 					}
591 				}
592 			} else {
593 				i = 0;
594 				uc = 0;
595 			}
596 			if (i == sc->sectorsize) {
597 				if (osp != uc)
598 					error = s_write(sc->indir, secno, uc);
599 			} else {
600 				if (osp <= 255) {
601 					sp = (uintptr_t)uma_zalloc(sc->uma,
602 					    md_malloc_wait ? M_WAITOK :
603 					    M_NOWAIT);
604 					if (sp == 0) {
605 						error = ENOSPC;
606 						break;
607 					}
608 					if (notmapped) {
609 						error = md_malloc_move(&m,
610 						    &ma_offs, sc->sectorsize,
611 						    (void *)sp, 0,
612 						    MD_MALLOC_MOVE_WRITE);
613 					} else {
614 						bcopy(dst, (void *)sp,
615 						    sc->sectorsize);
616 					}
617 					error = s_write(sc->indir, secno, sp);
618 				} else {
619 					if (notmapped) {
620 						error = md_malloc_move(&m,
621 						    &ma_offs, sc->sectorsize,
622 						    (void *)osp, 0,
623 						    MD_MALLOC_MOVE_WRITE);
624 					} else {
625 						bcopy(dst, (void *)osp,
626 						    sc->sectorsize);
627 					}
628 					osp = 0;
629 				}
630 			}
631 		} else {
632 			error = EOPNOTSUPP;
633 		}
634 		if (osp > 255)
635 			uma_zfree(sc->uma, (void*)osp);
636 		if (error != 0)
637 			break;
638 		secno++;
639 		if (!notmapped)
640 			dst += sc->sectorsize;
641 	}
642 	bp->bio_resid = 0;
643 	return (error);
644 }
645 
646 static int
647 mdstart_preload(struct md_s *sc, struct bio *bp)
648 {
649 
650 	switch (bp->bio_cmd) {
651 	case BIO_READ:
652 		bcopy(sc->pl_ptr + bp->bio_offset, bp->bio_data,
653 		    bp->bio_length);
654 		cpu_flush_dcache(bp->bio_data, bp->bio_length);
655 		break;
656 	case BIO_WRITE:
657 		bcopy(bp->bio_data, sc->pl_ptr + bp->bio_offset,
658 		    bp->bio_length);
659 		break;
660 	}
661 	bp->bio_resid = 0;
662 	return (0);
663 }
664 
665 static int
666 mdstart_vnode(struct md_s *sc, struct bio *bp)
667 {
668 	int error;
669 	struct uio auio;
670 	struct iovec aiov;
671 	struct mount *mp;
672 	struct vnode *vp;
673 	struct buf *pb;
674 	struct thread *td;
675 	off_t end, zerosize;
676 
677 	switch (bp->bio_cmd) {
678 	case BIO_READ:
679 	case BIO_WRITE:
680 	case BIO_DELETE:
681 	case BIO_FLUSH:
682 		break;
683 	default:
684 		return (EOPNOTSUPP);
685 	}
686 
687 	td = curthread;
688 	vp = sc->vnode;
689 
690 	/*
691 	 * VNODE I/O
692 	 *
693 	 * If an error occurs, we set BIO_ERROR but we do not set
694 	 * B_INVAL because (for a write anyway), the buffer is
695 	 * still valid.
696 	 */
697 
698 	if (bp->bio_cmd == BIO_FLUSH) {
699 		(void) vn_start_write(vp, &mp, V_WAIT);
700 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
701 		error = VOP_FSYNC(vp, MNT_WAIT, td);
702 		VOP_UNLOCK(vp, 0);
703 		vn_finished_write(mp);
704 		return (error);
705 	}
706 
707 	bzero(&auio, sizeof(auio));
708 
709 	/*
710 	 * Special case for BIO_DELETE.  On the surface, this is very
711 	 * similar to BIO_WRITE, except that we write from our own
712 	 * fixed-length buffer, so we have to loop.  The net result is
713 	 * that the two cases end up having very little in common.
714 	 */
715 	if (bp->bio_cmd == BIO_DELETE) {
716 		zerosize = ZERO_REGION_SIZE -
717 		    (ZERO_REGION_SIZE % sc->sectorsize);
718 		auio.uio_iov = &aiov;
719 		auio.uio_iovcnt = 1;
720 		auio.uio_offset = (vm_ooffset_t)bp->bio_offset;
721 		auio.uio_segflg = UIO_SYSSPACE;
722 		auio.uio_rw = UIO_WRITE;
723 		auio.uio_td = td;
724 		end = bp->bio_offset + bp->bio_length;
725 		(void) vn_start_write(vp, &mp, V_WAIT);
726 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
727 		error = 0;
728 		while (auio.uio_offset < end) {
729 			aiov.iov_base = __DECONST(void *, zero_region);
730 			aiov.iov_len = end - auio.uio_offset;
731 			if (aiov.iov_len > zerosize)
732 				aiov.iov_len = zerosize;
733 			auio.uio_resid = aiov.iov_len;
734 			error = VOP_WRITE(vp, &auio,
735 			    sc->flags & MD_ASYNC ? 0 : IO_SYNC, sc->cred);
736 			if (error != 0)
737 				break;
738 		}
739 		VOP_UNLOCK(vp, 0);
740 		vn_finished_write(mp);
741 		bp->bio_resid = end - auio.uio_offset;
742 		return (error);
743 	}
744 
745 	KASSERT(bp->bio_length <= MAXPHYS, ("bio_length %jd",
746 	    (uintmax_t)bp->bio_length));
747 	if ((bp->bio_flags & BIO_UNMAPPED) == 0) {
748 		pb = NULL;
749 		aiov.iov_base = bp->bio_data;
750 	} else {
751 		pb = getpbuf(&md_vnode_pbuf_freecnt);
752 		pmap_qenter((vm_offset_t)pb->b_data, bp->bio_ma, bp->bio_ma_n);
753 		aiov.iov_base = (void *)((vm_offset_t)pb->b_data +
754 		    bp->bio_ma_offset);
755 	}
756 	aiov.iov_len = bp->bio_length;
757 	auio.uio_iov = &aiov;
758 	auio.uio_iovcnt = 1;
759 	auio.uio_offset = (vm_ooffset_t)bp->bio_offset;
760 	auio.uio_segflg = UIO_SYSSPACE;
761 	if (bp->bio_cmd == BIO_READ)
762 		auio.uio_rw = UIO_READ;
763 	else if (bp->bio_cmd == BIO_WRITE)
764 		auio.uio_rw = UIO_WRITE;
765 	else
766 		panic("wrong BIO_OP in mdstart_vnode");
767 	auio.uio_resid = bp->bio_length;
768 	auio.uio_td = td;
769 	/*
770 	 * When reading set IO_DIRECT to try to avoid double-caching
771 	 * the data.  When writing IO_DIRECT is not optimal.
772 	 */
773 	if (bp->bio_cmd == BIO_READ) {
774 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
775 		error = VOP_READ(vp, &auio, IO_DIRECT, sc->cred);
776 		VOP_UNLOCK(vp, 0);
777 	} else {
778 		(void) vn_start_write(vp, &mp, V_WAIT);
779 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
780 		error = VOP_WRITE(vp, &auio, sc->flags & MD_ASYNC ? 0 : IO_SYNC,
781 		    sc->cred);
782 		VOP_UNLOCK(vp, 0);
783 		vn_finished_write(mp);
784 	}
785 	if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
786 		pmap_qremove((vm_offset_t)pb->b_data, bp->bio_ma_n);
787 		relpbuf(pb, &md_vnode_pbuf_freecnt);
788 	}
789 	bp->bio_resid = auio.uio_resid;
790 	return (error);
791 }
792 
793 static int
794 mdstart_swap(struct md_s *sc, struct bio *bp)
795 {
796 	vm_page_t m;
797 	u_char *p;
798 	vm_pindex_t i, lastp;
799 	int rv, ma_offs, offs, len, lastend;
800 
801 	switch (bp->bio_cmd) {
802 	case BIO_READ:
803 	case BIO_WRITE:
804 	case BIO_DELETE:
805 		break;
806 	default:
807 		return (EOPNOTSUPP);
808 	}
809 
810 	p = bp->bio_data;
811 	ma_offs = (bp->bio_flags & BIO_UNMAPPED) == 0 ? 0 : bp->bio_ma_offset;
812 
813 	/*
814 	 * offs is the offset at which to start operating on the
815 	 * next (ie, first) page.  lastp is the last page on
816 	 * which we're going to operate.  lastend is the ending
817 	 * position within that last page (ie, PAGE_SIZE if
818 	 * we're operating on complete aligned pages).
819 	 */
820 	offs = bp->bio_offset % PAGE_SIZE;
821 	lastp = (bp->bio_offset + bp->bio_length - 1) / PAGE_SIZE;
822 	lastend = (bp->bio_offset + bp->bio_length - 1) % PAGE_SIZE + 1;
823 
824 	rv = VM_PAGER_OK;
825 	VM_OBJECT_WLOCK(sc->object);
826 	vm_object_pip_add(sc->object, 1);
827 	for (i = bp->bio_offset / PAGE_SIZE; i <= lastp; i++) {
828 		len = ((i == lastp) ? lastend : PAGE_SIZE) - offs;
829 		m = vm_page_grab(sc->object, i, VM_ALLOC_SYSTEM);
830 		if (bp->bio_cmd == BIO_READ) {
831 			if (m->valid == VM_PAGE_BITS_ALL)
832 				rv = VM_PAGER_OK;
833 			else
834 				rv = vm_pager_get_pages(sc->object, &m, 1, 0);
835 			if (rv == VM_PAGER_ERROR) {
836 				vm_page_xunbusy(m);
837 				break;
838 			} else if (rv == VM_PAGER_FAIL) {
839 				/*
840 				 * Pager does not have the page.  Zero
841 				 * the allocated page, and mark it as
842 				 * valid. Do not set dirty, the page
843 				 * can be recreated if thrown out.
844 				 */
845 				pmap_zero_page(m);
846 				m->valid = VM_PAGE_BITS_ALL;
847 			}
848 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
849 				pmap_copy_pages(&m, offs, bp->bio_ma,
850 				    ma_offs, len);
851 			} else {
852 				physcopyout(VM_PAGE_TO_PHYS(m) + offs, p, len);
853 				cpu_flush_dcache(p, len);
854 			}
855 		} else if (bp->bio_cmd == BIO_WRITE) {
856 			if (len != PAGE_SIZE && m->valid != VM_PAGE_BITS_ALL)
857 				rv = vm_pager_get_pages(sc->object, &m, 1, 0);
858 			else
859 				rv = VM_PAGER_OK;
860 			if (rv == VM_PAGER_ERROR) {
861 				vm_page_xunbusy(m);
862 				break;
863 			}
864 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
865 				pmap_copy_pages(bp->bio_ma, ma_offs, &m,
866 				    offs, len);
867 			} else {
868 				physcopyin(p, VM_PAGE_TO_PHYS(m) + offs, len);
869 			}
870 			m->valid = VM_PAGE_BITS_ALL;
871 		} else if (bp->bio_cmd == BIO_DELETE) {
872 			if (len != PAGE_SIZE && m->valid != VM_PAGE_BITS_ALL)
873 				rv = vm_pager_get_pages(sc->object, &m, 1, 0);
874 			else
875 				rv = VM_PAGER_OK;
876 			if (rv == VM_PAGER_ERROR) {
877 				vm_page_xunbusy(m);
878 				break;
879 			}
880 			if (len != PAGE_SIZE) {
881 				pmap_zero_page_area(m, offs, len);
882 				vm_page_clear_dirty(m, offs, len);
883 				m->valid = VM_PAGE_BITS_ALL;
884 			} else
885 				vm_pager_page_unswapped(m);
886 		}
887 		vm_page_xunbusy(m);
888 		vm_page_lock(m);
889 		if (bp->bio_cmd == BIO_DELETE && len == PAGE_SIZE)
890 			vm_page_free(m);
891 		else
892 			vm_page_activate(m);
893 		vm_page_unlock(m);
894 		if (bp->bio_cmd == BIO_WRITE)
895 			vm_page_dirty(m);
896 
897 		/* Actions on further pages start at offset 0 */
898 		p += PAGE_SIZE - offs;
899 		offs = 0;
900 		ma_offs += len;
901 	}
902 	vm_object_pip_subtract(sc->object, 1);
903 	VM_OBJECT_WUNLOCK(sc->object);
904 	return (rv != VM_PAGER_ERROR ? 0 : ENOSPC);
905 }
906 
907 static void
908 md_kthread(void *arg)
909 {
910 	struct md_s *sc;
911 	struct bio *bp;
912 	int error;
913 
914 	sc = arg;
915 	thread_lock(curthread);
916 	sched_prio(curthread, PRIBIO);
917 	thread_unlock(curthread);
918 	if (sc->type == MD_VNODE)
919 		curthread->td_pflags |= TDP_NORUNNINGBUF;
920 
921 	for (;;) {
922 		mtx_lock(&sc->queue_mtx);
923 		if (sc->flags & MD_SHUTDOWN) {
924 			sc->flags |= MD_EXITING;
925 			mtx_unlock(&sc->queue_mtx);
926 			kproc_exit(0);
927 		}
928 		bp = bioq_takefirst(&sc->bio_queue);
929 		if (!bp) {
930 			msleep(sc, &sc->queue_mtx, PRIBIO | PDROP, "mdwait", 0);
931 			continue;
932 		}
933 		mtx_unlock(&sc->queue_mtx);
934 		if (bp->bio_cmd == BIO_GETATTR) {
935 			if ((sc->fwsectors && sc->fwheads &&
936 			    (g_handleattr_int(bp, "GEOM::fwsectors",
937 			    sc->fwsectors) ||
938 			    g_handleattr_int(bp, "GEOM::fwheads",
939 			    sc->fwheads))) ||
940 			    g_handleattr_int(bp, "GEOM::candelete", 1))
941 				error = -1;
942 			else
943 				error = EOPNOTSUPP;
944 		} else {
945 			error = sc->start(sc, bp);
946 		}
947 
948 		if (error != -1) {
949 			bp->bio_completed = bp->bio_length;
950 			if ((bp->bio_cmd == BIO_READ) || (bp->bio_cmd == BIO_WRITE))
951 				devstat_end_transaction_bio(sc->devstat, bp);
952 			g_io_deliver(bp, error);
953 		}
954 	}
955 }
956 
957 static struct md_s *
958 mdfind(int unit)
959 {
960 	struct md_s *sc;
961 
962 	LIST_FOREACH(sc, &md_softc_list, list) {
963 		if (sc->unit == unit)
964 			break;
965 	}
966 	return (sc);
967 }
968 
969 static struct md_s *
970 mdnew(int unit, int *errp, enum md_types type)
971 {
972 	struct md_s *sc;
973 	int error;
974 
975 	*errp = 0;
976 	if (unit == -1)
977 		unit = alloc_unr(md_uh);
978 	else
979 		unit = alloc_unr_specific(md_uh, unit);
980 
981 	if (unit == -1) {
982 		*errp = EBUSY;
983 		return (NULL);
984 	}
985 
986 	sc = (struct md_s *)malloc(sizeof *sc, M_MD, M_WAITOK | M_ZERO);
987 	sc->type = type;
988 	bioq_init(&sc->bio_queue);
989 	mtx_init(&sc->queue_mtx, "md bio queue", NULL, MTX_DEF);
990 	sc->unit = unit;
991 	sprintf(sc->name, "md%d", unit);
992 	LIST_INSERT_HEAD(&md_softc_list, sc, list);
993 	error = kproc_create(md_kthread, sc, &sc->procp, 0, 0,"%s", sc->name);
994 	if (error == 0)
995 		return (sc);
996 	LIST_REMOVE(sc, list);
997 	mtx_destroy(&sc->queue_mtx);
998 	free_unr(md_uh, sc->unit);
999 	free(sc, M_MD);
1000 	*errp = error;
1001 	return (NULL);
1002 }
1003 
1004 static void
1005 mdinit(struct md_s *sc)
1006 {
1007 	struct g_geom *gp;
1008 	struct g_provider *pp;
1009 
1010 	g_topology_lock();
1011 	gp = g_new_geomf(&g_md_class, "md%d", sc->unit);
1012 	gp->softc = sc;
1013 	pp = g_new_providerf(gp, "md%d", sc->unit);
1014 	pp->mediasize = sc->mediasize;
1015 	pp->sectorsize = sc->sectorsize;
1016 	switch (sc->type) {
1017 	case MD_MALLOC:
1018 	case MD_VNODE:
1019 	case MD_SWAP:
1020 		pp->flags |= G_PF_ACCEPT_UNMAPPED;
1021 		break;
1022 	case MD_PRELOAD:
1023 		break;
1024 	}
1025 	sc->gp = gp;
1026 	sc->pp = pp;
1027 	g_error_provider(pp, 0);
1028 	g_topology_unlock();
1029 	sc->devstat = devstat_new_entry("md", sc->unit, sc->sectorsize,
1030 	    DEVSTAT_ALL_SUPPORTED, DEVSTAT_TYPE_DIRECT, DEVSTAT_PRIORITY_MAX);
1031 }
1032 
1033 static int
1034 mdcreate_malloc(struct md_s *sc, struct md_ioctl *mdio)
1035 {
1036 	uintptr_t sp;
1037 	int error;
1038 	off_t u;
1039 
1040 	error = 0;
1041 	if (mdio->md_options & ~(MD_AUTOUNIT | MD_COMPRESS | MD_RESERVE))
1042 		return (EINVAL);
1043 	if (mdio->md_sectorsize != 0 && !powerof2(mdio->md_sectorsize))
1044 		return (EINVAL);
1045 	/* Compression doesn't make sense if we have reserved space */
1046 	if (mdio->md_options & MD_RESERVE)
1047 		mdio->md_options &= ~MD_COMPRESS;
1048 	if (mdio->md_fwsectors != 0)
1049 		sc->fwsectors = mdio->md_fwsectors;
1050 	if (mdio->md_fwheads != 0)
1051 		sc->fwheads = mdio->md_fwheads;
1052 	sc->flags = mdio->md_options & (MD_COMPRESS | MD_FORCE);
1053 	sc->indir = dimension(sc->mediasize / sc->sectorsize);
1054 	sc->uma = uma_zcreate(sc->name, sc->sectorsize, NULL, NULL, NULL, NULL,
1055 	    0x1ff, 0);
1056 	if (mdio->md_options & MD_RESERVE) {
1057 		off_t nsectors;
1058 
1059 		nsectors = sc->mediasize / sc->sectorsize;
1060 		for (u = 0; u < nsectors; u++) {
1061 			sp = (uintptr_t)uma_zalloc(sc->uma, (md_malloc_wait ?
1062 			    M_WAITOK : M_NOWAIT) | M_ZERO);
1063 			if (sp != 0)
1064 				error = s_write(sc->indir, u, sp);
1065 			else
1066 				error = ENOMEM;
1067 			if (error != 0)
1068 				break;
1069 		}
1070 	}
1071 	return (error);
1072 }
1073 
1074 
1075 static int
1076 mdsetcred(struct md_s *sc, struct ucred *cred)
1077 {
1078 	char *tmpbuf;
1079 	int error = 0;
1080 
1081 	/*
1082 	 * Set credits in our softc
1083 	 */
1084 
1085 	if (sc->cred)
1086 		crfree(sc->cred);
1087 	sc->cred = crhold(cred);
1088 
1089 	/*
1090 	 * Horrible kludge to establish credentials for NFS  XXX.
1091 	 */
1092 
1093 	if (sc->vnode) {
1094 		struct uio auio;
1095 		struct iovec aiov;
1096 
1097 		tmpbuf = malloc(sc->sectorsize, M_TEMP, M_WAITOK);
1098 		bzero(&auio, sizeof(auio));
1099 
1100 		aiov.iov_base = tmpbuf;
1101 		aiov.iov_len = sc->sectorsize;
1102 		auio.uio_iov = &aiov;
1103 		auio.uio_iovcnt = 1;
1104 		auio.uio_offset = 0;
1105 		auio.uio_rw = UIO_READ;
1106 		auio.uio_segflg = UIO_SYSSPACE;
1107 		auio.uio_resid = aiov.iov_len;
1108 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY);
1109 		error = VOP_READ(sc->vnode, &auio, 0, sc->cred);
1110 		VOP_UNLOCK(sc->vnode, 0);
1111 		free(tmpbuf, M_TEMP);
1112 	}
1113 	return (error);
1114 }
1115 
1116 static int
1117 mdcreate_vnode(struct md_s *sc, struct md_ioctl *mdio, struct thread *td)
1118 {
1119 	struct vattr vattr;
1120 	struct nameidata nd;
1121 	char *fname;
1122 	int error, flags;
1123 
1124 	/*
1125 	 * Kernel-originated requests must have the filename appended
1126 	 * to the mdio structure to protect against malicious software.
1127 	 */
1128 	fname = mdio->md_file;
1129 	if ((void *)fname != (void *)(mdio + 1)) {
1130 		error = copyinstr(fname, sc->file, sizeof(sc->file), NULL);
1131 		if (error != 0)
1132 			return (error);
1133 	} else
1134 		strlcpy(sc->file, fname, sizeof(sc->file));
1135 
1136 	/*
1137 	 * If the user specified that this is a read only device, don't
1138 	 * set the FWRITE mask before trying to open the backing store.
1139 	 */
1140 	flags = FREAD | ((mdio->md_options & MD_READONLY) ? 0 : FWRITE);
1141 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, sc->file, td);
1142 	error = vn_open(&nd, &flags, 0, NULL);
1143 	if (error != 0)
1144 		return (error);
1145 	NDFREE(&nd, NDF_ONLY_PNBUF);
1146 	if (nd.ni_vp->v_type != VREG) {
1147 		error = EINVAL;
1148 		goto bad;
1149 	}
1150 	error = VOP_GETATTR(nd.ni_vp, &vattr, td->td_ucred);
1151 	if (error != 0)
1152 		goto bad;
1153 	if (VOP_ISLOCKED(nd.ni_vp) != LK_EXCLUSIVE) {
1154 		vn_lock(nd.ni_vp, LK_UPGRADE | LK_RETRY);
1155 		if (nd.ni_vp->v_iflag & VI_DOOMED) {
1156 			/* Forced unmount. */
1157 			error = EBADF;
1158 			goto bad;
1159 		}
1160 	}
1161 	nd.ni_vp->v_vflag |= VV_MD;
1162 	VOP_UNLOCK(nd.ni_vp, 0);
1163 
1164 	if (mdio->md_fwsectors != 0)
1165 		sc->fwsectors = mdio->md_fwsectors;
1166 	if (mdio->md_fwheads != 0)
1167 		sc->fwheads = mdio->md_fwheads;
1168 	sc->flags = mdio->md_options & (MD_FORCE | MD_ASYNC);
1169 	if (!(flags & FWRITE))
1170 		sc->flags |= MD_READONLY;
1171 	sc->vnode = nd.ni_vp;
1172 
1173 	error = mdsetcred(sc, td->td_ucred);
1174 	if (error != 0) {
1175 		sc->vnode = NULL;
1176 		vn_lock(nd.ni_vp, LK_EXCLUSIVE | LK_RETRY);
1177 		nd.ni_vp->v_vflag &= ~VV_MD;
1178 		goto bad;
1179 	}
1180 	return (0);
1181 bad:
1182 	VOP_UNLOCK(nd.ni_vp, 0);
1183 	(void)vn_close(nd.ni_vp, flags, td->td_ucred, td);
1184 	return (error);
1185 }
1186 
1187 static int
1188 mddestroy(struct md_s *sc, struct thread *td)
1189 {
1190 
1191 	if (sc->gp) {
1192 		sc->gp->softc = NULL;
1193 		g_topology_lock();
1194 		g_wither_geom(sc->gp, ENXIO);
1195 		g_topology_unlock();
1196 		sc->gp = NULL;
1197 		sc->pp = NULL;
1198 	}
1199 	if (sc->devstat) {
1200 		devstat_remove_entry(sc->devstat);
1201 		sc->devstat = NULL;
1202 	}
1203 	mtx_lock(&sc->queue_mtx);
1204 	sc->flags |= MD_SHUTDOWN;
1205 	wakeup(sc);
1206 	while (!(sc->flags & MD_EXITING))
1207 		msleep(sc->procp, &sc->queue_mtx, PRIBIO, "mddestroy", hz / 10);
1208 	mtx_unlock(&sc->queue_mtx);
1209 	mtx_destroy(&sc->queue_mtx);
1210 	if (sc->vnode != NULL) {
1211 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY);
1212 		sc->vnode->v_vflag &= ~VV_MD;
1213 		VOP_UNLOCK(sc->vnode, 0);
1214 		(void)vn_close(sc->vnode, sc->flags & MD_READONLY ?
1215 		    FREAD : (FREAD|FWRITE), sc->cred, td);
1216 	}
1217 	if (sc->cred != NULL)
1218 		crfree(sc->cred);
1219 	if (sc->object != NULL)
1220 		vm_object_deallocate(sc->object);
1221 	if (sc->indir)
1222 		destroy_indir(sc, sc->indir);
1223 	if (sc->uma)
1224 		uma_zdestroy(sc->uma);
1225 
1226 	LIST_REMOVE(sc, list);
1227 	free_unr(md_uh, sc->unit);
1228 	free(sc, M_MD);
1229 	return (0);
1230 }
1231 
1232 static int
1233 mdresize(struct md_s *sc, struct md_ioctl *mdio)
1234 {
1235 	int error, res;
1236 	vm_pindex_t oldpages, newpages;
1237 
1238 	switch (sc->type) {
1239 	case MD_VNODE:
1240 		break;
1241 	case MD_SWAP:
1242 		if (mdio->md_mediasize <= 0 ||
1243 		    (mdio->md_mediasize % PAGE_SIZE) != 0)
1244 			return (EDOM);
1245 		oldpages = OFF_TO_IDX(round_page(sc->mediasize));
1246 		newpages = OFF_TO_IDX(round_page(mdio->md_mediasize));
1247 		if (newpages < oldpages) {
1248 			VM_OBJECT_WLOCK(sc->object);
1249 			vm_object_page_remove(sc->object, newpages, 0, 0);
1250 			swap_pager_freespace(sc->object, newpages,
1251 			    oldpages - newpages);
1252 			swap_release_by_cred(IDX_TO_OFF(oldpages -
1253 			    newpages), sc->cred);
1254 			sc->object->charge = IDX_TO_OFF(newpages);
1255 			sc->object->size = newpages;
1256 			VM_OBJECT_WUNLOCK(sc->object);
1257 		} else if (newpages > oldpages) {
1258 			res = swap_reserve_by_cred(IDX_TO_OFF(newpages -
1259 			    oldpages), sc->cred);
1260 			if (!res)
1261 				return (ENOMEM);
1262 			if ((mdio->md_options & MD_RESERVE) ||
1263 			    (sc->flags & MD_RESERVE)) {
1264 				error = swap_pager_reserve(sc->object,
1265 				    oldpages, newpages - oldpages);
1266 				if (error < 0) {
1267 					swap_release_by_cred(
1268 					    IDX_TO_OFF(newpages - oldpages),
1269 					    sc->cred);
1270 					return (EDOM);
1271 				}
1272 			}
1273 			VM_OBJECT_WLOCK(sc->object);
1274 			sc->object->charge = IDX_TO_OFF(newpages);
1275 			sc->object->size = newpages;
1276 			VM_OBJECT_WUNLOCK(sc->object);
1277 		}
1278 		break;
1279 	default:
1280 		return (EOPNOTSUPP);
1281 	}
1282 
1283 	sc->mediasize = mdio->md_mediasize;
1284 	g_topology_lock();
1285 	g_resize_provider(sc->pp, sc->mediasize);
1286 	g_topology_unlock();
1287 	return (0);
1288 }
1289 
1290 static int
1291 mdcreate_swap(struct md_s *sc, struct md_ioctl *mdio, struct thread *td)
1292 {
1293 	vm_ooffset_t npage;
1294 	int error;
1295 
1296 	/*
1297 	 * Range check.  Disallow negative sizes or any size less then the
1298 	 * size of a page.  Then round to a page.
1299 	 */
1300 	if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0)
1301 		return (EDOM);
1302 
1303 	/*
1304 	 * Allocate an OBJT_SWAP object.
1305 	 *
1306 	 * Note the truncation.
1307 	 */
1308 
1309 	npage = mdio->md_mediasize / PAGE_SIZE;
1310 	if (mdio->md_fwsectors != 0)
1311 		sc->fwsectors = mdio->md_fwsectors;
1312 	if (mdio->md_fwheads != 0)
1313 		sc->fwheads = mdio->md_fwheads;
1314 	sc->object = vm_pager_allocate(OBJT_SWAP, NULL, PAGE_SIZE * npage,
1315 	    VM_PROT_DEFAULT, 0, td->td_ucred);
1316 	if (sc->object == NULL)
1317 		return (ENOMEM);
1318 	sc->flags = mdio->md_options & (MD_FORCE | MD_RESERVE);
1319 	if (mdio->md_options & MD_RESERVE) {
1320 		if (swap_pager_reserve(sc->object, 0, npage) < 0) {
1321 			error = EDOM;
1322 			goto finish;
1323 		}
1324 	}
1325 	error = mdsetcred(sc, td->td_ucred);
1326  finish:
1327 	if (error != 0) {
1328 		vm_object_deallocate(sc->object);
1329 		sc->object = NULL;
1330 	}
1331 	return (error);
1332 }
1333 
1334 
1335 static int
1336 xmdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td)
1337 {
1338 	struct md_ioctl *mdio;
1339 	struct md_s *sc;
1340 	int error, i;
1341 	unsigned sectsize;
1342 
1343 	if (md_debug)
1344 		printf("mdctlioctl(%s %lx %p %x %p)\n",
1345 			devtoname(dev), cmd, addr, flags, td);
1346 
1347 	mdio = (struct md_ioctl *)addr;
1348 	if (mdio->md_version != MDIOVERSION)
1349 		return (EINVAL);
1350 
1351 	/*
1352 	 * We assert the version number in the individual ioctl
1353 	 * handlers instead of out here because (a) it is possible we
1354 	 * may add another ioctl in the future which doesn't read an
1355 	 * mdio, and (b) the correct return value for an unknown ioctl
1356 	 * is ENOIOCTL, not EINVAL.
1357 	 */
1358 	error = 0;
1359 	switch (cmd) {
1360 	case MDIOCATTACH:
1361 		switch (mdio->md_type) {
1362 		case MD_MALLOC:
1363 		case MD_PRELOAD:
1364 		case MD_VNODE:
1365 		case MD_SWAP:
1366 			break;
1367 		default:
1368 			return (EINVAL);
1369 		}
1370 		if (mdio->md_sectorsize == 0)
1371 			sectsize = DEV_BSIZE;
1372 		else
1373 			sectsize = mdio->md_sectorsize;
1374 		if (sectsize > MAXPHYS || mdio->md_mediasize < sectsize)
1375 			return (EINVAL);
1376 		if (mdio->md_options & MD_AUTOUNIT)
1377 			sc = mdnew(-1, &error, mdio->md_type);
1378 		else {
1379 			if (mdio->md_unit > INT_MAX)
1380 				return (EINVAL);
1381 			sc = mdnew(mdio->md_unit, &error, mdio->md_type);
1382 		}
1383 		if (sc == NULL)
1384 			return (error);
1385 		if (mdio->md_options & MD_AUTOUNIT)
1386 			mdio->md_unit = sc->unit;
1387 		sc->mediasize = mdio->md_mediasize;
1388 		sc->sectorsize = sectsize;
1389 		error = EDOOFUS;
1390 		switch (sc->type) {
1391 		case MD_MALLOC:
1392 			sc->start = mdstart_malloc;
1393 			error = mdcreate_malloc(sc, mdio);
1394 			break;
1395 		case MD_PRELOAD:
1396 			/*
1397 			 * We disallow attaching preloaded memory disks via
1398 			 * ioctl. Preloaded memory disks are automatically
1399 			 * attached in g_md_init().
1400 			 */
1401 			error = EOPNOTSUPP;
1402 			break;
1403 		case MD_VNODE:
1404 			sc->start = mdstart_vnode;
1405 			error = mdcreate_vnode(sc, mdio, td);
1406 			break;
1407 		case MD_SWAP:
1408 			sc->start = mdstart_swap;
1409 			error = mdcreate_swap(sc, mdio, td);
1410 			break;
1411 		}
1412 		if (error != 0) {
1413 			mddestroy(sc, td);
1414 			return (error);
1415 		}
1416 
1417 		/* Prune off any residual fractional sector */
1418 		i = sc->mediasize % sc->sectorsize;
1419 		sc->mediasize -= i;
1420 
1421 		mdinit(sc);
1422 		return (0);
1423 	case MDIOCDETACH:
1424 		if (mdio->md_mediasize != 0 ||
1425 		    (mdio->md_options & ~MD_FORCE) != 0)
1426 			return (EINVAL);
1427 
1428 		sc = mdfind(mdio->md_unit);
1429 		if (sc == NULL)
1430 			return (ENOENT);
1431 		if (sc->opencount != 0 && !(sc->flags & MD_FORCE) &&
1432 		    !(mdio->md_options & MD_FORCE))
1433 			return (EBUSY);
1434 		return (mddestroy(sc, td));
1435 	case MDIOCRESIZE:
1436 		if ((mdio->md_options & ~(MD_FORCE | MD_RESERVE)) != 0)
1437 			return (EINVAL);
1438 
1439 		sc = mdfind(mdio->md_unit);
1440 		if (sc == NULL)
1441 			return (ENOENT);
1442 		if (mdio->md_mediasize < sc->sectorsize)
1443 			return (EINVAL);
1444 		if (mdio->md_mediasize < sc->mediasize &&
1445 		    !(sc->flags & MD_FORCE) &&
1446 		    !(mdio->md_options & MD_FORCE))
1447 			return (EBUSY);
1448 		return (mdresize(sc, mdio));
1449 	case MDIOCQUERY:
1450 		sc = mdfind(mdio->md_unit);
1451 		if (sc == NULL)
1452 			return (ENOENT);
1453 		mdio->md_type = sc->type;
1454 		mdio->md_options = sc->flags;
1455 		mdio->md_mediasize = sc->mediasize;
1456 		mdio->md_sectorsize = sc->sectorsize;
1457 		if (sc->type == MD_VNODE)
1458 			error = copyout(sc->file, mdio->md_file,
1459 			    strlen(sc->file) + 1);
1460 		return (error);
1461 	case MDIOCLIST:
1462 		i = 1;
1463 		LIST_FOREACH(sc, &md_softc_list, list) {
1464 			if (i == MDNPAD - 1)
1465 				mdio->md_pad[i] = -1;
1466 			else
1467 				mdio->md_pad[i++] = sc->unit;
1468 		}
1469 		mdio->md_pad[0] = i - 1;
1470 		return (0);
1471 	default:
1472 		return (ENOIOCTL);
1473 	};
1474 }
1475 
1476 static int
1477 mdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td)
1478 {
1479 	int error;
1480 
1481 	sx_xlock(&md_sx);
1482 	error = xmdctlioctl(dev, cmd, addr, flags, td);
1483 	sx_xunlock(&md_sx);
1484 	return (error);
1485 }
1486 
1487 static void
1488 md_preloaded(u_char *image, size_t length, const char *name)
1489 {
1490 	struct md_s *sc;
1491 	int error;
1492 
1493 	sc = mdnew(-1, &error, MD_PRELOAD);
1494 	if (sc == NULL)
1495 		return;
1496 	sc->mediasize = length;
1497 	sc->sectorsize = DEV_BSIZE;
1498 	sc->pl_ptr = image;
1499 	sc->pl_len = length;
1500 	sc->start = mdstart_preload;
1501 #ifdef MD_ROOT
1502 	if (sc->unit == 0)
1503 		rootdevnames[0] = MD_ROOT_FSTYPE ":/dev/md0";
1504 #endif
1505 	mdinit(sc);
1506 	if (name != NULL) {
1507 		printf("%s%d: Preloaded image <%s> %zd bytes at %p\n",
1508 		    MD_NAME, sc->unit, name, length, image);
1509 	}
1510 }
1511 
1512 static void
1513 g_md_init(struct g_class *mp __unused)
1514 {
1515 	caddr_t mod;
1516 	u_char *ptr, *name, *type;
1517 	unsigned len;
1518 	int i;
1519 
1520 	/* figure out log2(NINDIR) */
1521 	for (i = NINDIR, nshift = -1; i; nshift++)
1522 		i >>= 1;
1523 
1524 	mod = NULL;
1525 	sx_init(&md_sx, "MD config lock");
1526 	g_topology_unlock();
1527 	md_uh = new_unrhdr(0, INT_MAX, NULL);
1528 #ifdef MD_ROOT_SIZE
1529 	sx_xlock(&md_sx);
1530 	md_preloaded(mfs_root.start, sizeof(mfs_root.start), NULL);
1531 	sx_xunlock(&md_sx);
1532 #endif
1533 	/* XXX: are preload_* static or do they need Giant ? */
1534 	while ((mod = preload_search_next_name(mod)) != NULL) {
1535 		name = (char *)preload_search_info(mod, MODINFO_NAME);
1536 		if (name == NULL)
1537 			continue;
1538 		type = (char *)preload_search_info(mod, MODINFO_TYPE);
1539 		if (type == NULL)
1540 			continue;
1541 		if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
1542 			continue;
1543 		ptr = preload_fetch_addr(mod);
1544 		len = preload_fetch_size(mod);
1545 		if (ptr != NULL && len != 0) {
1546 			sx_xlock(&md_sx);
1547 			md_preloaded(ptr, len, name);
1548 			sx_xunlock(&md_sx);
1549 		}
1550 	}
1551 	md_vnode_pbuf_freecnt = nswbuf / 10;
1552 	status_dev = make_dev(&mdctl_cdevsw, INT_MAX, UID_ROOT, GID_WHEEL,
1553 	    0600, MDCTL_NAME);
1554 	g_topology_lock();
1555 }
1556 
1557 static void
1558 g_md_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1559     struct g_consumer *cp __unused, struct g_provider *pp)
1560 {
1561 	struct md_s *mp;
1562 	char *type;
1563 
1564 	mp = gp->softc;
1565 	if (mp == NULL)
1566 		return;
1567 
1568 	switch (mp->type) {
1569 	case MD_MALLOC:
1570 		type = "malloc";
1571 		break;
1572 	case MD_PRELOAD:
1573 		type = "preload";
1574 		break;
1575 	case MD_VNODE:
1576 		type = "vnode";
1577 		break;
1578 	case MD_SWAP:
1579 		type = "swap";
1580 		break;
1581 	default:
1582 		type = "unknown";
1583 		break;
1584 	}
1585 
1586 	if (pp != NULL) {
1587 		if (indent == NULL) {
1588 			sbuf_printf(sb, " u %d", mp->unit);
1589 			sbuf_printf(sb, " s %ju", (uintmax_t) mp->sectorsize);
1590 			sbuf_printf(sb, " f %ju", (uintmax_t) mp->fwheads);
1591 			sbuf_printf(sb, " fs %ju", (uintmax_t) mp->fwsectors);
1592 			sbuf_printf(sb, " l %ju", (uintmax_t) mp->mediasize);
1593 			sbuf_printf(sb, " t %s", type);
1594 			if (mp->type == MD_VNODE && mp->vnode != NULL)
1595 				sbuf_printf(sb, " file %s", mp->file);
1596 		} else {
1597 			sbuf_printf(sb, "%s<unit>%d</unit>\n", indent,
1598 			    mp->unit);
1599 			sbuf_printf(sb, "%s<sectorsize>%ju</sectorsize>\n",
1600 			    indent, (uintmax_t) mp->sectorsize);
1601 			sbuf_printf(sb, "%s<fwheads>%ju</fwheads>\n",
1602 			    indent, (uintmax_t) mp->fwheads);
1603 			sbuf_printf(sb, "%s<fwsectors>%ju</fwsectors>\n",
1604 			    indent, (uintmax_t) mp->fwsectors);
1605 			sbuf_printf(sb, "%s<length>%ju</length>\n",
1606 			    indent, (uintmax_t) mp->mediasize);
1607 			sbuf_printf(sb, "%s<compression>%s</compression>\n", indent,
1608 			    (mp->flags & MD_COMPRESS) == 0 ? "off": "on");
1609 			sbuf_printf(sb, "%s<access>%s</access>\n", indent,
1610 			    (mp->flags & MD_READONLY) == 0 ? "read-write":
1611 			    "read-only");
1612 			sbuf_printf(sb, "%s<type>%s</type>\n", indent,
1613 			    type);
1614 			if (mp->type == MD_VNODE && mp->vnode != NULL)
1615 				sbuf_printf(sb, "%s<file>%s</file>\n",
1616 				    indent, mp->file);
1617 		}
1618 	}
1619 }
1620 
1621 static void
1622 g_md_fini(struct g_class *mp __unused)
1623 {
1624 
1625 	sx_destroy(&md_sx);
1626 	if (status_dev != NULL)
1627 		destroy_dev(status_dev);
1628 	delete_unrhdr(md_uh);
1629 }
1630