xref: /freebsd/sys/dev/md/md.c (revision ca53e5aedfebcc1b4091b68e01b2d5cae923f85e)
1 /*-
2  * SPDX-License-Identifier: (Beerware AND BSD-3-Clause)
3  *
4  * ----------------------------------------------------------------------------
5  * "THE BEER-WARE LICENSE" (Revision 42):
6  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
7  * can do whatever you want with this stuff. If we meet some day, and you think
8  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
9  * ----------------------------------------------------------------------------
10  *
11  * $FreeBSD$
12  *
13  */
14 
15 /*-
16  * The following functions are based on the vn(4) driver: mdstart_swap(),
17  * mdstart_vnode(), mdcreate_swap(), mdcreate_vnode() and mddestroy(),
18  * and as such under the following copyright:
19  *
20  * Copyright (c) 1988 University of Utah.
21  * Copyright (c) 1990, 1993
22  *	The Regents of the University of California.  All rights reserved.
23  * Copyright (c) 2013 The FreeBSD Foundation
24  * All rights reserved.
25  *
26  * This code is derived from software contributed to Berkeley by
27  * the Systems Programming Group of the University of Utah Computer
28  * Science Department.
29  *
30  * Portions of this software were developed by Konstantin Belousov
31  * under sponsorship from the FreeBSD Foundation.
32  *
33  * Redistribution and use in source and binary forms, with or without
34  * modification, are permitted provided that the following conditions
35  * are met:
36  * 1. Redistributions of source code must retain the above copyright
37  *    notice, this list of conditions and the following disclaimer.
38  * 2. Redistributions in binary form must reproduce the above copyright
39  *    notice, this list of conditions and the following disclaimer in the
40  *    documentation and/or other materials provided with the distribution.
41  * 3. Neither the name of the University nor the names of its contributors
42  *    may be used to endorse or promote products derived from this software
43  *    without specific prior written permission.
44  *
45  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
46  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
47  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
48  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
49  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
50  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
51  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
52  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
53  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
54  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
55  * SUCH DAMAGE.
56  *
57  * from: Utah Hdr: vn.c 1.13 94/04/02
58  *
59  *	from: @(#)vn.c	8.6 (Berkeley) 4/1/94
60  * From: src/sys/dev/vn/vn.c,v 1.122 2000/12/16 16:06:03
61  */
62 
63 #include "opt_rootdevname.h"
64 #include "opt_geom.h"
65 #include "opt_md.h"
66 
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/bio.h>
70 #include <sys/buf.h>
71 #include <sys/conf.h>
72 #include <sys/devicestat.h>
73 #include <sys/fcntl.h>
74 #include <sys/kernel.h>
75 #include <sys/kthread.h>
76 #include <sys/limits.h>
77 #include <sys/linker.h>
78 #include <sys/lock.h>
79 #include <sys/malloc.h>
80 #include <sys/mdioctl.h>
81 #include <sys/mount.h>
82 #include <sys/mutex.h>
83 #include <sys/sx.h>
84 #include <sys/namei.h>
85 #include <sys/proc.h>
86 #include <sys/queue.h>
87 #include <sys/rwlock.h>
88 #include <sys/sbuf.h>
89 #include <sys/sched.h>
90 #include <sys/sf_buf.h>
91 #include <sys/sysctl.h>
92 #include <sys/uio.h>
93 #include <sys/vnode.h>
94 #include <sys/disk.h>
95 
96 #include <geom/geom.h>
97 #include <geom/geom_int.h>
98 
99 #include <vm/vm.h>
100 #include <vm/vm_param.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/uma.h>
106 
107 #include <machine/bus.h>
108 
109 #define MD_MODVER 1
110 
111 #define MD_SHUTDOWN	0x10000		/* Tell worker thread to terminate. */
112 #define	MD_EXITING	0x20000		/* Worker thread is exiting. */
113 #define MD_PROVIDERGONE	0x40000		/* Safe to free the softc */
114 
115 #ifndef MD_NSECT
116 #define MD_NSECT (10000 * 2)
117 #endif
118 
119 struct md_req {
120 	unsigned	md_unit;	/* unit number */
121 	enum md_types	md_type;	/* type of disk */
122 	off_t		md_mediasize;	/* size of disk in bytes */
123 	unsigned	md_sectorsize;	/* sectorsize */
124 	unsigned	md_options;	/* options */
125 	int		md_fwheads;	/* firmware heads */
126 	int		md_fwsectors;	/* firmware sectors */
127 	char		*md_file;	/* pathname of file to mount */
128 	enum uio_seg	md_file_seg;	/* location of md_file */
129 	char		*md_label;	/* label of the device (userspace) */
130 	int		*md_units;	/* pointer to units array (kernel) */
131 	size_t		md_units_nitems; /* items in md_units array */
132 };
133 
134 #ifdef COMPAT_FREEBSD32
135 struct md_ioctl32 {
136 	unsigned	md_version;
137 	unsigned	md_unit;
138 	enum md_types	md_type;
139 	uint32_t	md_file;
140 	off_t		md_mediasize;
141 	unsigned	md_sectorsize;
142 	unsigned	md_options;
143 	uint64_t	md_base;
144 	int		md_fwheads;
145 	int		md_fwsectors;
146 	uint32_t	md_label;
147 	int		md_pad[MDNPAD];
148 } __attribute__((__packed__));
149 CTASSERT((sizeof(struct md_ioctl32)) == 436);
150 
151 #define	MDIOCATTACH_32	_IOC_NEWTYPE(MDIOCATTACH, struct md_ioctl32)
152 #define	MDIOCDETACH_32	_IOC_NEWTYPE(MDIOCDETACH, struct md_ioctl32)
153 #define	MDIOCQUERY_32	_IOC_NEWTYPE(MDIOCQUERY, struct md_ioctl32)
154 #define	MDIOCRESIZE_32	_IOC_NEWTYPE(MDIOCRESIZE, struct md_ioctl32)
155 #endif /* COMPAT_FREEBSD32 */
156 
157 static MALLOC_DEFINE(M_MD, "md_disk", "Memory Disk");
158 static MALLOC_DEFINE(M_MDSECT, "md_sectors", "Memory Disk Sectors");
159 
160 static int md_debug;
161 SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0,
162     "Enable md(4) debug messages");
163 static int md_malloc_wait;
164 SYSCTL_INT(_vm, OID_AUTO, md_malloc_wait, CTLFLAG_RW, &md_malloc_wait, 0,
165     "Allow malloc to wait for memory allocations");
166 
167 #if defined(MD_ROOT) && !defined(MD_ROOT_FSTYPE)
168 #define	MD_ROOT_FSTYPE	"ufs"
169 #endif
170 
171 #if defined(MD_ROOT)
172 /*
173  * Preloaded image gets put here.
174  */
175 #if defined(MD_ROOT_SIZE)
176 /*
177  * We put the mfs_root symbol into the oldmfs section of the kernel object file.
178  * Applications that patch the object with the image can determine
179  * the size looking at the oldmfs section size within the kernel.
180  */
181 u_char mfs_root[MD_ROOT_SIZE*1024] __attribute__ ((section ("oldmfs")));
182 const int mfs_root_size = sizeof(mfs_root);
183 #elif defined(MD_ROOT_MEM)
184 /* MD region already mapped in the memory */
185 u_char *mfs_root;
186 int mfs_root_size;
187 #else
188 extern volatile u_char __weak_symbol mfs_root;
189 extern volatile u_char __weak_symbol mfs_root_end;
190 __GLOBL(mfs_root);
191 __GLOBL(mfs_root_end);
192 #define mfs_root_size ((uintptr_t)(&mfs_root_end - &mfs_root))
193 #endif
194 #endif
195 
196 static g_init_t g_md_init;
197 static g_fini_t g_md_fini;
198 static g_start_t g_md_start;
199 static g_access_t g_md_access;
200 static void g_md_dumpconf(struct sbuf *sb, const char *indent,
201     struct g_geom *gp, struct g_consumer *cp __unused, struct g_provider *pp);
202 static g_provgone_t g_md_providergone;
203 
204 static struct cdev *status_dev = NULL;
205 static struct sx md_sx;
206 static struct unrhdr *md_uh;
207 
208 static d_ioctl_t mdctlioctl;
209 
210 static struct cdevsw mdctl_cdevsw = {
211 	.d_version =	D_VERSION,
212 	.d_ioctl =	mdctlioctl,
213 	.d_name =	MD_NAME,
214 };
215 
216 struct g_class g_md_class = {
217 	.name = "MD",
218 	.version = G_VERSION,
219 	.init = g_md_init,
220 	.fini = g_md_fini,
221 	.start = g_md_start,
222 	.access = g_md_access,
223 	.dumpconf = g_md_dumpconf,
224 	.providergone = g_md_providergone,
225 };
226 
227 DECLARE_GEOM_CLASS(g_md_class, g_md);
228 
229 static LIST_HEAD(, md_s) md_softc_list = LIST_HEAD_INITIALIZER(md_softc_list);
230 
231 #define NINDIR	(PAGE_SIZE / sizeof(uintptr_t))
232 #define NMASK	(NINDIR-1)
233 static int nshift;
234 
235 static uma_zone_t md_pbuf_zone;
236 
237 struct indir {
238 	uintptr_t	*array;
239 	u_int		total;
240 	u_int		used;
241 	u_int		shift;
242 };
243 
244 struct md_s {
245 	int unit;
246 	LIST_ENTRY(md_s) list;
247 	struct bio_queue_head bio_queue;
248 	struct mtx queue_mtx;
249 	struct cdev *dev;
250 	enum md_types type;
251 	off_t mediasize;
252 	unsigned sectorsize;
253 	unsigned opencount;
254 	unsigned fwheads;
255 	unsigned fwsectors;
256 	char ident[32];
257 	unsigned flags;
258 	char name[20];
259 	struct proc *procp;
260 	struct g_geom *gp;
261 	struct g_provider *pp;
262 	int (*start)(struct md_s *sc, struct bio *bp);
263 	struct devstat *devstat;
264 
265 	/* MD_MALLOC related fields */
266 	struct indir *indir;
267 	uma_zone_t uma;
268 
269 	/* MD_PRELOAD related fields */
270 	u_char *pl_ptr;
271 	size_t pl_len;
272 
273 	/* MD_VNODE related fields */
274 	struct vnode *vnode;
275 	char file[PATH_MAX];
276 	char label[PATH_MAX];
277 	struct ucred *cred;
278 
279 	/* MD_SWAP related fields */
280 	vm_object_t object;
281 };
282 
283 static struct indir *
284 new_indir(u_int shift)
285 {
286 	struct indir *ip;
287 
288 	ip = malloc(sizeof *ip, M_MD, (md_malloc_wait ? M_WAITOK : M_NOWAIT)
289 	    | M_ZERO);
290 	if (ip == NULL)
291 		return (NULL);
292 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
293 	    M_MDSECT, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO);
294 	if (ip->array == NULL) {
295 		free(ip, M_MD);
296 		return (NULL);
297 	}
298 	ip->total = NINDIR;
299 	ip->shift = shift;
300 	return (ip);
301 }
302 
303 static void
304 del_indir(struct indir *ip)
305 {
306 
307 	free(ip->array, M_MDSECT);
308 	free(ip, M_MD);
309 }
310 
311 static void
312 destroy_indir(struct md_s *sc, struct indir *ip)
313 {
314 	int i;
315 
316 	for (i = 0; i < NINDIR; i++) {
317 		if (!ip->array[i])
318 			continue;
319 		if (ip->shift)
320 			destroy_indir(sc, (struct indir*)(ip->array[i]));
321 		else if (ip->array[i] > 255)
322 			uma_zfree(sc->uma, (void *)(ip->array[i]));
323 	}
324 	del_indir(ip);
325 }
326 
327 /*
328  * This function does the math and allocates the top level "indir" structure
329  * for a device of "size" sectors.
330  */
331 
332 static struct indir *
333 dimension(off_t size)
334 {
335 	off_t rcnt;
336 	struct indir *ip;
337 	int layer;
338 
339 	rcnt = size;
340 	layer = 0;
341 	while (rcnt > NINDIR) {
342 		rcnt /= NINDIR;
343 		layer++;
344 	}
345 
346 	/*
347 	 * XXX: the top layer is probably not fully populated, so we allocate
348 	 * too much space for ip->array in here.
349 	 */
350 	ip = malloc(sizeof *ip, M_MD, M_WAITOK | M_ZERO);
351 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
352 	    M_MDSECT, M_WAITOK | M_ZERO);
353 	ip->total = NINDIR;
354 	ip->shift = layer * nshift;
355 	return (ip);
356 }
357 
358 /*
359  * Read a given sector
360  */
361 
362 static uintptr_t
363 s_read(struct indir *ip, off_t offset)
364 {
365 	struct indir *cip;
366 	int idx;
367 	uintptr_t up;
368 
369 	if (md_debug > 1)
370 		printf("s_read(%jd)\n", (intmax_t)offset);
371 	up = 0;
372 	for (cip = ip; cip != NULL;) {
373 		if (cip->shift) {
374 			idx = (offset >> cip->shift) & NMASK;
375 			up = cip->array[idx];
376 			cip = (struct indir *)up;
377 			continue;
378 		}
379 		idx = offset & NMASK;
380 		return (cip->array[idx]);
381 	}
382 	return (0);
383 }
384 
385 /*
386  * Write a given sector, prune the tree if the value is 0
387  */
388 
389 static int
390 s_write(struct indir *ip, off_t offset, uintptr_t ptr)
391 {
392 	struct indir *cip, *lip[10];
393 	int idx, li;
394 	uintptr_t up;
395 
396 	if (md_debug > 1)
397 		printf("s_write(%jd, %p)\n", (intmax_t)offset, (void *)ptr);
398 	up = 0;
399 	li = 0;
400 	cip = ip;
401 	for (;;) {
402 		lip[li++] = cip;
403 		if (cip->shift) {
404 			idx = (offset >> cip->shift) & NMASK;
405 			up = cip->array[idx];
406 			if (up != 0) {
407 				cip = (struct indir *)up;
408 				continue;
409 			}
410 			/* Allocate branch */
411 			cip->array[idx] =
412 			    (uintptr_t)new_indir(cip->shift - nshift);
413 			if (cip->array[idx] == 0)
414 				return (ENOSPC);
415 			cip->used++;
416 			up = cip->array[idx];
417 			cip = (struct indir *)up;
418 			continue;
419 		}
420 		/* leafnode */
421 		idx = offset & NMASK;
422 		up = cip->array[idx];
423 		if (up != 0)
424 			cip->used--;
425 		cip->array[idx] = ptr;
426 		if (ptr != 0)
427 			cip->used++;
428 		break;
429 	}
430 	if (cip->used != 0 || li == 1)
431 		return (0);
432 	li--;
433 	while (cip->used == 0 && cip != ip) {
434 		li--;
435 		idx = (offset >> lip[li]->shift) & NMASK;
436 		up = lip[li]->array[idx];
437 		KASSERT(up == (uintptr_t)cip, ("md screwed up"));
438 		del_indir(cip);
439 		lip[li]->array[idx] = 0;
440 		lip[li]->used--;
441 		cip = lip[li];
442 	}
443 	return (0);
444 }
445 
446 static int
447 g_md_access(struct g_provider *pp, int r, int w, int e)
448 {
449 	struct md_s *sc;
450 
451 	sc = pp->geom->softc;
452 	if (sc == NULL) {
453 		if (r <= 0 && w <= 0 && e <= 0)
454 			return (0);
455 		return (ENXIO);
456 	}
457 	r += pp->acr;
458 	w += pp->acw;
459 	e += pp->ace;
460 	if ((sc->flags & MD_READONLY) != 0 && w > 0)
461 		return (EROFS);
462 	if ((pp->acr + pp->acw + pp->ace) == 0 && (r + w + e) > 0) {
463 		sc->opencount = 1;
464 	} else if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0) {
465 		sc->opencount = 0;
466 	}
467 	return (0);
468 }
469 
470 static void
471 g_md_start(struct bio *bp)
472 {
473 	struct md_s *sc;
474 
475 	sc = bp->bio_to->geom->softc;
476 	if ((bp->bio_cmd == BIO_READ) || (bp->bio_cmd == BIO_WRITE)) {
477 		devstat_start_transaction_bio(sc->devstat, bp);
478 	}
479 	mtx_lock(&sc->queue_mtx);
480 	bioq_disksort(&sc->bio_queue, bp);
481 	wakeup(sc);
482 	mtx_unlock(&sc->queue_mtx);
483 }
484 
485 #define	MD_MALLOC_MOVE_ZERO	1
486 #define	MD_MALLOC_MOVE_FILL	2
487 #define	MD_MALLOC_MOVE_READ	3
488 #define	MD_MALLOC_MOVE_WRITE	4
489 #define	MD_MALLOC_MOVE_CMP	5
490 
491 static int
492 md_malloc_move_ma(vm_page_t **mp, int *ma_offs, unsigned sectorsize,
493     void *ptr, u_char fill, int op)
494 {
495 	struct sf_buf *sf;
496 	vm_page_t m, *mp1;
497 	char *p, first;
498 	off_t *uc;
499 	unsigned n;
500 	int error, i, ma_offs1, sz, first_read;
501 
502 	m = NULL;
503 	error = 0;
504 	sf = NULL;
505 	/* if (op == MD_MALLOC_MOVE_CMP) { gcc */
506 		first = 0;
507 		first_read = 0;
508 		uc = ptr;
509 		mp1 = *mp;
510 		ma_offs1 = *ma_offs;
511 	/* } */
512 	sched_pin();
513 	for (n = sectorsize; n != 0; n -= sz) {
514 		sz = imin(PAGE_SIZE - *ma_offs, n);
515 		if (m != **mp) {
516 			if (sf != NULL)
517 				sf_buf_free(sf);
518 			m = **mp;
519 			sf = sf_buf_alloc(m, SFB_CPUPRIVATE |
520 			    (md_malloc_wait ? 0 : SFB_NOWAIT));
521 			if (sf == NULL) {
522 				error = ENOMEM;
523 				break;
524 			}
525 		}
526 		p = (char *)sf_buf_kva(sf) + *ma_offs;
527 		switch (op) {
528 		case MD_MALLOC_MOVE_ZERO:
529 			bzero(p, sz);
530 			break;
531 		case MD_MALLOC_MOVE_FILL:
532 			memset(p, fill, sz);
533 			break;
534 		case MD_MALLOC_MOVE_READ:
535 			bcopy(ptr, p, sz);
536 			cpu_flush_dcache(p, sz);
537 			break;
538 		case MD_MALLOC_MOVE_WRITE:
539 			bcopy(p, ptr, sz);
540 			break;
541 		case MD_MALLOC_MOVE_CMP:
542 			for (i = 0; i < sz; i++, p++) {
543 				if (!first_read) {
544 					*uc = (u_char)*p;
545 					first = *p;
546 					first_read = 1;
547 				} else if (*p != first) {
548 					error = EDOOFUS;
549 					break;
550 				}
551 			}
552 			break;
553 		default:
554 			KASSERT(0, ("md_malloc_move_ma unknown op %d\n", op));
555 			break;
556 		}
557 		if (error != 0)
558 			break;
559 		*ma_offs += sz;
560 		*ma_offs %= PAGE_SIZE;
561 		if (*ma_offs == 0)
562 			(*mp)++;
563 		ptr = (char *)ptr + sz;
564 	}
565 
566 	if (sf != NULL)
567 		sf_buf_free(sf);
568 	sched_unpin();
569 	if (op == MD_MALLOC_MOVE_CMP && error != 0) {
570 		*mp = mp1;
571 		*ma_offs = ma_offs1;
572 	}
573 	return (error);
574 }
575 
576 static int
577 md_malloc_move_vlist(bus_dma_segment_t **pvlist, int *pma_offs,
578     unsigned len, void *ptr, u_char fill, int op)
579 {
580 	bus_dma_segment_t *vlist;
581 	uint8_t *p, *end, first;
582 	off_t *uc;
583 	int ma_offs, seg_len;
584 
585 	vlist = *pvlist;
586 	ma_offs = *pma_offs;
587 	uc = ptr;
588 
589 	for (; len != 0; len -= seg_len) {
590 		seg_len = imin(vlist->ds_len - ma_offs, len);
591 		p = (uint8_t *)(uintptr_t)vlist->ds_addr + ma_offs;
592 		switch (op) {
593 		case MD_MALLOC_MOVE_ZERO:
594 			bzero(p, seg_len);
595 			break;
596 		case MD_MALLOC_MOVE_FILL:
597 			memset(p, fill, seg_len);
598 			break;
599 		case MD_MALLOC_MOVE_READ:
600 			bcopy(ptr, p, seg_len);
601 			cpu_flush_dcache(p, seg_len);
602 			break;
603 		case MD_MALLOC_MOVE_WRITE:
604 			bcopy(p, ptr, seg_len);
605 			break;
606 		case MD_MALLOC_MOVE_CMP:
607 			end = p + seg_len;
608 			first = *uc = *p;
609 			/* Confirm all following bytes match the first */
610 			while (++p < end) {
611 				if (*p != first)
612 					return (EDOOFUS);
613 			}
614 			break;
615 		default:
616 			KASSERT(0, ("md_malloc_move_vlist unknown op %d\n", op));
617 			break;
618 		}
619 
620 		ma_offs += seg_len;
621 		if (ma_offs == vlist->ds_len) {
622 			ma_offs = 0;
623 			vlist++;
624 		}
625 		ptr = (uint8_t *)ptr + seg_len;
626 	}
627 	*pvlist = vlist;
628 	*pma_offs = ma_offs;
629 
630 	return (0);
631 }
632 
633 static int
634 mdstart_malloc(struct md_s *sc, struct bio *bp)
635 {
636 	u_char *dst;
637 	vm_page_t *m;
638 	bus_dma_segment_t *vlist;
639 	int i, error, error1, ma_offs, notmapped;
640 	off_t secno, nsec, uc;
641 	uintptr_t sp, osp;
642 
643 	switch (bp->bio_cmd) {
644 	case BIO_READ:
645 	case BIO_WRITE:
646 	case BIO_DELETE:
647 		break;
648 	default:
649 		return (EOPNOTSUPP);
650 	}
651 
652 	notmapped = (bp->bio_flags & BIO_UNMAPPED) != 0;
653 	vlist = (bp->bio_flags & BIO_VLIST) != 0 ?
654 	    (bus_dma_segment_t *)bp->bio_data : NULL;
655 	if (notmapped) {
656 		m = bp->bio_ma;
657 		ma_offs = bp->bio_ma_offset;
658 		dst = NULL;
659 		KASSERT(vlist == NULL, ("vlists cannot be unmapped"));
660 	} else if (vlist != NULL) {
661 		ma_offs = bp->bio_ma_offset;
662 		dst = NULL;
663 	} else {
664 		dst = bp->bio_data;
665 	}
666 
667 	nsec = bp->bio_length / sc->sectorsize;
668 	secno = bp->bio_offset / sc->sectorsize;
669 	error = 0;
670 	while (nsec--) {
671 		osp = s_read(sc->indir, secno);
672 		if (bp->bio_cmd == BIO_DELETE) {
673 			if (osp != 0)
674 				error = s_write(sc->indir, secno, 0);
675 		} else if (bp->bio_cmd == BIO_READ) {
676 			if (osp == 0) {
677 				if (notmapped) {
678 					error = md_malloc_move_ma(&m, &ma_offs,
679 					    sc->sectorsize, NULL, 0,
680 					    MD_MALLOC_MOVE_ZERO);
681 				} else if (vlist != NULL) {
682 					error = md_malloc_move_vlist(&vlist,
683 					    &ma_offs, sc->sectorsize, NULL, 0,
684 					    MD_MALLOC_MOVE_ZERO);
685 				} else
686 					bzero(dst, sc->sectorsize);
687 			} else if (osp <= 255) {
688 				if (notmapped) {
689 					error = md_malloc_move_ma(&m, &ma_offs,
690 					    sc->sectorsize, NULL, osp,
691 					    MD_MALLOC_MOVE_FILL);
692 				} else if (vlist != NULL) {
693 					error = md_malloc_move_vlist(&vlist,
694 					    &ma_offs, sc->sectorsize, NULL, osp,
695 					    MD_MALLOC_MOVE_FILL);
696 				} else
697 					memset(dst, osp, sc->sectorsize);
698 			} else {
699 				if (notmapped) {
700 					error = md_malloc_move_ma(&m, &ma_offs,
701 					    sc->sectorsize, (void *)osp, 0,
702 					    MD_MALLOC_MOVE_READ);
703 				} else if (vlist != NULL) {
704 					error = md_malloc_move_vlist(&vlist,
705 					    &ma_offs, sc->sectorsize,
706 					    (void *)osp, 0,
707 					    MD_MALLOC_MOVE_READ);
708 				} else {
709 					bcopy((void *)osp, dst, sc->sectorsize);
710 					cpu_flush_dcache(dst, sc->sectorsize);
711 				}
712 			}
713 			osp = 0;
714 		} else if (bp->bio_cmd == BIO_WRITE) {
715 			if (sc->flags & MD_COMPRESS) {
716 				if (notmapped) {
717 					error1 = md_malloc_move_ma(&m, &ma_offs,
718 					    sc->sectorsize, &uc, 0,
719 					    MD_MALLOC_MOVE_CMP);
720 					i = error1 == 0 ? sc->sectorsize : 0;
721 				} else if (vlist != NULL) {
722 					error1 = md_malloc_move_vlist(&vlist,
723 					    &ma_offs, sc->sectorsize, &uc, 0,
724 					    MD_MALLOC_MOVE_CMP);
725 					i = error1 == 0 ? sc->sectorsize : 0;
726 				} else {
727 					uc = dst[0];
728 					for (i = 1; i < sc->sectorsize; i++) {
729 						if (dst[i] != uc)
730 							break;
731 					}
732 				}
733 			} else {
734 				i = 0;
735 				uc = 0;
736 			}
737 			if (i == sc->sectorsize) {
738 				if (osp != uc)
739 					error = s_write(sc->indir, secno, uc);
740 			} else {
741 				if (osp <= 255) {
742 					sp = (uintptr_t)uma_zalloc(sc->uma,
743 					    md_malloc_wait ? M_WAITOK :
744 					    M_NOWAIT);
745 					if (sp == 0) {
746 						error = ENOSPC;
747 						break;
748 					}
749 					if (notmapped) {
750 						error = md_malloc_move_ma(&m,
751 						    &ma_offs, sc->sectorsize,
752 						    (void *)sp, 0,
753 						    MD_MALLOC_MOVE_WRITE);
754 					} else if (vlist != NULL) {
755 						error = md_malloc_move_vlist(
756 						    &vlist, &ma_offs,
757 						    sc->sectorsize, (void *)sp,
758 						    0, MD_MALLOC_MOVE_WRITE);
759 					} else {
760 						bcopy(dst, (void *)sp,
761 						    sc->sectorsize);
762 					}
763 					error = s_write(sc->indir, secno, sp);
764 				} else {
765 					if (notmapped) {
766 						error = md_malloc_move_ma(&m,
767 						    &ma_offs, sc->sectorsize,
768 						    (void *)osp, 0,
769 						    MD_MALLOC_MOVE_WRITE);
770 					} else if (vlist != NULL) {
771 						error = md_malloc_move_vlist(
772 						    &vlist, &ma_offs,
773 						    sc->sectorsize, (void *)osp,
774 						    0, MD_MALLOC_MOVE_WRITE);
775 					} else {
776 						bcopy(dst, (void *)osp,
777 						    sc->sectorsize);
778 					}
779 					osp = 0;
780 				}
781 			}
782 		} else {
783 			error = EOPNOTSUPP;
784 		}
785 		if (osp > 255)
786 			uma_zfree(sc->uma, (void*)osp);
787 		if (error != 0)
788 			break;
789 		secno++;
790 		if (!notmapped && vlist == NULL)
791 			dst += sc->sectorsize;
792 	}
793 	bp->bio_resid = 0;
794 	return (error);
795 }
796 
797 static void
798 mdcopyto_vlist(void *src, bus_dma_segment_t *vlist, off_t offset, off_t len)
799 {
800 	off_t seg_len;
801 
802 	while (offset >= vlist->ds_len) {
803 		offset -= vlist->ds_len;
804 		vlist++;
805 	}
806 
807 	while (len != 0) {
808 		seg_len = omin(len, vlist->ds_len - offset);
809 		bcopy(src, (void *)(uintptr_t)(vlist->ds_addr + offset),
810 		    seg_len);
811 		offset = 0;
812 		src = (uint8_t *)src + seg_len;
813 		len -= seg_len;
814 		vlist++;
815 	}
816 }
817 
818 static void
819 mdcopyfrom_vlist(bus_dma_segment_t *vlist, off_t offset, void *dst, off_t len)
820 {
821 	off_t seg_len;
822 
823 	while (offset >= vlist->ds_len) {
824 		offset -= vlist->ds_len;
825 		vlist++;
826 	}
827 
828 	while (len != 0) {
829 		seg_len = omin(len, vlist->ds_len - offset);
830 		bcopy((void *)(uintptr_t)(vlist->ds_addr + offset), dst,
831 		    seg_len);
832 		offset = 0;
833 		dst = (uint8_t *)dst + seg_len;
834 		len -= seg_len;
835 		vlist++;
836 	}
837 }
838 
839 static int
840 mdstart_preload(struct md_s *sc, struct bio *bp)
841 {
842 	uint8_t *p;
843 
844 	p = sc->pl_ptr + bp->bio_offset;
845 	switch (bp->bio_cmd) {
846 	case BIO_READ:
847 		if ((bp->bio_flags & BIO_VLIST) != 0) {
848 			mdcopyto_vlist(p, (bus_dma_segment_t *)bp->bio_data,
849 			    bp->bio_ma_offset, bp->bio_length);
850 		} else {
851 			bcopy(p, bp->bio_data, bp->bio_length);
852 		}
853 		cpu_flush_dcache(bp->bio_data, bp->bio_length);
854 		break;
855 	case BIO_WRITE:
856 		if ((bp->bio_flags & BIO_VLIST) != 0) {
857 			mdcopyfrom_vlist((bus_dma_segment_t *)bp->bio_data,
858 			    bp->bio_ma_offset, p, bp->bio_length);
859 		} else {
860 			bcopy(bp->bio_data, p, bp->bio_length);
861 		}
862 		break;
863 	}
864 	bp->bio_resid = 0;
865 	return (0);
866 }
867 
868 static int
869 mdstart_vnode(struct md_s *sc, struct bio *bp)
870 {
871 	int error;
872 	struct uio auio;
873 	struct iovec aiov;
874 	struct iovec *piov;
875 	struct mount *mp;
876 	struct vnode *vp;
877 	struct buf *pb;
878 	bus_dma_segment_t *vlist;
879 	struct thread *td;
880 	off_t iolen, iostart, len, zerosize;
881 	int ma_offs, npages;
882 
883 	switch (bp->bio_cmd) {
884 	case BIO_READ:
885 		auio.uio_rw = UIO_READ;
886 		break;
887 	case BIO_WRITE:
888 	case BIO_DELETE:
889 		auio.uio_rw = UIO_WRITE;
890 		break;
891 	case BIO_FLUSH:
892 		break;
893 	default:
894 		return (EOPNOTSUPP);
895 	}
896 
897 	td = curthread;
898 	vp = sc->vnode;
899 	pb = NULL;
900 	piov = NULL;
901 	ma_offs = bp->bio_ma_offset;
902 	len = bp->bio_length;
903 
904 	/*
905 	 * VNODE I/O
906 	 *
907 	 * If an error occurs, we set BIO_ERROR but we do not set
908 	 * B_INVAL because (for a write anyway), the buffer is
909 	 * still valid.
910 	 */
911 
912 	if (bp->bio_cmd == BIO_FLUSH) {
913 		(void) vn_start_write(vp, &mp, V_WAIT);
914 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
915 		error = VOP_FSYNC(vp, MNT_WAIT, td);
916 		VOP_UNLOCK(vp);
917 		vn_finished_write(mp);
918 		return (error);
919 	}
920 
921 	auio.uio_offset = (vm_ooffset_t)bp->bio_offset;
922 	auio.uio_resid = bp->bio_length;
923 	auio.uio_segflg = UIO_SYSSPACE;
924 	auio.uio_td = td;
925 
926 	if (bp->bio_cmd == BIO_DELETE) {
927 		/*
928 		 * Emulate BIO_DELETE by writing zeros.
929 		 */
930 		zerosize = ZERO_REGION_SIZE -
931 		    (ZERO_REGION_SIZE % sc->sectorsize);
932 		auio.uio_iovcnt = howmany(bp->bio_length, zerosize);
933 		piov = malloc(sizeof(*piov) * auio.uio_iovcnt, M_MD, M_WAITOK);
934 		auio.uio_iov = piov;
935 		while (len > 0) {
936 			piov->iov_base = __DECONST(void *, zero_region);
937 			piov->iov_len = len;
938 			if (len > zerosize)
939 				piov->iov_len = zerosize;
940 			len -= piov->iov_len;
941 			piov++;
942 		}
943 		piov = auio.uio_iov;
944 	} else if ((bp->bio_flags & BIO_VLIST) != 0) {
945 		piov = malloc(sizeof(*piov) * bp->bio_ma_n, M_MD, M_WAITOK);
946 		auio.uio_iov = piov;
947 		vlist = (bus_dma_segment_t *)bp->bio_data;
948 		while (len > 0) {
949 			piov->iov_base = (void *)(uintptr_t)(vlist->ds_addr +
950 			    ma_offs);
951 			piov->iov_len = vlist->ds_len - ma_offs;
952 			if (piov->iov_len > len)
953 				piov->iov_len = len;
954 			len -= piov->iov_len;
955 			ma_offs = 0;
956 			vlist++;
957 			piov++;
958 		}
959 		auio.uio_iovcnt = piov - auio.uio_iov;
960 		piov = auio.uio_iov;
961 	} else if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
962 		pb = uma_zalloc(md_pbuf_zone, M_WAITOK);
963 		MPASS((pb->b_flags & B_MAXPHYS) != 0);
964 		bp->bio_resid = len;
965 unmapped_step:
966 		npages = atop(min(maxphys, round_page(len + (ma_offs &
967 		    PAGE_MASK))));
968 		iolen = min(ptoa(npages) - (ma_offs & PAGE_MASK), len);
969 		KASSERT(iolen > 0, ("zero iolen"));
970 		pmap_qenter((vm_offset_t)pb->b_data,
971 		    &bp->bio_ma[atop(ma_offs)], npages);
972 		aiov.iov_base = (void *)((vm_offset_t)pb->b_data +
973 		    (ma_offs & PAGE_MASK));
974 		aiov.iov_len = iolen;
975 		auio.uio_iov = &aiov;
976 		auio.uio_iovcnt = 1;
977 		auio.uio_resid = iolen;
978 	} else {
979 		aiov.iov_base = bp->bio_data;
980 		aiov.iov_len = bp->bio_length;
981 		auio.uio_iov = &aiov;
982 		auio.uio_iovcnt = 1;
983 	}
984 	iostart = auio.uio_offset;
985 	if (auio.uio_rw == UIO_READ) {
986 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
987 		error = VOP_READ(vp, &auio, 0, sc->cred);
988 		VOP_UNLOCK(vp);
989 	} else {
990 		(void) vn_start_write(vp, &mp, V_WAIT);
991 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
992 		error = VOP_WRITE(vp, &auio, sc->flags & MD_ASYNC ? 0 : IO_SYNC,
993 		    sc->cred);
994 		VOP_UNLOCK(vp);
995 		vn_finished_write(mp);
996 		if (error == 0)
997 			sc->flags &= ~MD_VERIFY;
998 	}
999 
1000 	/* When MD_CACHE is set, try to avoid double-caching the data. */
1001 	if (error == 0 && (sc->flags & MD_CACHE) == 0)
1002 		VOP_ADVISE(vp, iostart, auio.uio_offset - 1,
1003 		    POSIX_FADV_DONTNEED);
1004 
1005 	if (pb != NULL) {
1006 		pmap_qremove((vm_offset_t)pb->b_data, npages);
1007 		if (error == 0) {
1008 			len -= iolen;
1009 			bp->bio_resid -= iolen;
1010 			ma_offs += iolen;
1011 			if (len > 0)
1012 				goto unmapped_step;
1013 		}
1014 		uma_zfree(md_pbuf_zone, pb);
1015 	}
1016 
1017 	free(piov, M_MD);
1018 	if (pb == NULL)
1019 		bp->bio_resid = auio.uio_resid;
1020 	return (error);
1021 }
1022 
1023 static int
1024 mdstart_swap(struct md_s *sc, struct bio *bp)
1025 {
1026 	vm_page_t m;
1027 	u_char *p;
1028 	vm_pindex_t i, lastp;
1029 	bus_dma_segment_t *vlist;
1030 	int rv, ma_offs, offs, len, lastend;
1031 
1032 	switch (bp->bio_cmd) {
1033 	case BIO_READ:
1034 	case BIO_WRITE:
1035 	case BIO_DELETE:
1036 		break;
1037 	default:
1038 		return (EOPNOTSUPP);
1039 	}
1040 
1041 	p = bp->bio_data;
1042 	ma_offs = (bp->bio_flags & (BIO_UNMAPPED|BIO_VLIST)) != 0 ?
1043 	    bp->bio_ma_offset : 0;
1044 	vlist = (bp->bio_flags & BIO_VLIST) != 0 ?
1045 	    (bus_dma_segment_t *)bp->bio_data : NULL;
1046 
1047 	/*
1048 	 * offs is the offset at which to start operating on the
1049 	 * next (ie, first) page.  lastp is the last page on
1050 	 * which we're going to operate.  lastend is the ending
1051 	 * position within that last page (ie, PAGE_SIZE if
1052 	 * we're operating on complete aligned pages).
1053 	 */
1054 	offs = bp->bio_offset % PAGE_SIZE;
1055 	lastp = (bp->bio_offset + bp->bio_length - 1) / PAGE_SIZE;
1056 	lastend = (bp->bio_offset + bp->bio_length - 1) % PAGE_SIZE + 1;
1057 
1058 	rv = VM_PAGER_OK;
1059 	vm_object_pip_add(sc->object, 1);
1060 	for (i = bp->bio_offset / PAGE_SIZE; i <= lastp; i++) {
1061 		len = ((i == lastp) ? lastend : PAGE_SIZE) - offs;
1062 		m = vm_page_grab_unlocked(sc->object, i, VM_ALLOC_SYSTEM);
1063 		if (bp->bio_cmd == BIO_READ) {
1064 			if (vm_page_all_valid(m))
1065 				rv = VM_PAGER_OK;
1066 			else
1067 				rv = vm_pager_get_pages(sc->object, &m, 1,
1068 				    NULL, NULL);
1069 			if (rv == VM_PAGER_ERROR) {
1070 				VM_OBJECT_WLOCK(sc->object);
1071 				vm_page_free(m);
1072 				VM_OBJECT_WUNLOCK(sc->object);
1073 				break;
1074 			} else if (rv == VM_PAGER_FAIL) {
1075 				/*
1076 				 * Pager does not have the page.  Zero
1077 				 * the allocated page, and mark it as
1078 				 * valid. Do not set dirty, the page
1079 				 * can be recreated if thrown out.
1080 				 */
1081 				pmap_zero_page(m);
1082 				vm_page_valid(m);
1083 			}
1084 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
1085 				pmap_copy_pages(&m, offs, bp->bio_ma,
1086 				    ma_offs, len);
1087 			} else if ((bp->bio_flags & BIO_VLIST) != 0) {
1088 				physcopyout_vlist(VM_PAGE_TO_PHYS(m) + offs,
1089 				    vlist, ma_offs, len);
1090 				cpu_flush_dcache(p, len);
1091 			} else {
1092 				physcopyout(VM_PAGE_TO_PHYS(m) + offs, p, len);
1093 				cpu_flush_dcache(p, len);
1094 			}
1095 		} else if (bp->bio_cmd == BIO_WRITE) {
1096 			if (len == PAGE_SIZE || vm_page_all_valid(m))
1097 				rv = VM_PAGER_OK;
1098 			else
1099 				rv = vm_pager_get_pages(sc->object, &m, 1,
1100 				    NULL, NULL);
1101 			if (rv == VM_PAGER_ERROR) {
1102 				VM_OBJECT_WLOCK(sc->object);
1103 				vm_page_free(m);
1104 				VM_OBJECT_WUNLOCK(sc->object);
1105 				break;
1106 			} else if (rv == VM_PAGER_FAIL)
1107 				pmap_zero_page(m);
1108 
1109 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
1110 				pmap_copy_pages(bp->bio_ma, ma_offs, &m,
1111 				    offs, len);
1112 			} else if ((bp->bio_flags & BIO_VLIST) != 0) {
1113 				physcopyin_vlist(vlist, ma_offs,
1114 				    VM_PAGE_TO_PHYS(m) + offs, len);
1115 			} else {
1116 				physcopyin(p, VM_PAGE_TO_PHYS(m) + offs, len);
1117 			}
1118 
1119 			vm_page_valid(m);
1120 			vm_page_set_dirty(m);
1121 		} else if (bp->bio_cmd == BIO_DELETE) {
1122 			if (len == PAGE_SIZE || vm_page_all_valid(m))
1123 				rv = VM_PAGER_OK;
1124 			else
1125 				rv = vm_pager_get_pages(sc->object, &m, 1,
1126 				    NULL, NULL);
1127 			VM_OBJECT_WLOCK(sc->object);
1128 			if (rv == VM_PAGER_ERROR) {
1129 				vm_page_free(m);
1130 				VM_OBJECT_WUNLOCK(sc->object);
1131 				break;
1132 			} else if (rv == VM_PAGER_FAIL) {
1133 				vm_page_free(m);
1134 				m = NULL;
1135 			} else {
1136 				/* Page is valid. */
1137 				if (len != PAGE_SIZE) {
1138 					pmap_zero_page_area(m, offs, len);
1139 					vm_page_set_dirty(m);
1140 				} else {
1141 					vm_pager_page_unswapped(m);
1142 					vm_page_free(m);
1143 					m = NULL;
1144 				}
1145 			}
1146 			VM_OBJECT_WUNLOCK(sc->object);
1147 		}
1148 		if (m != NULL) {
1149 			vm_page_xunbusy(m);
1150 
1151 			/*
1152 			 * The page may be deactivated prior to setting
1153 			 * PGA_REFERENCED, but in this case it will be
1154 			 * reactivated by the page daemon.
1155 			 */
1156 			if (vm_page_active(m))
1157 				vm_page_reference(m);
1158 			else
1159 				vm_page_activate(m);
1160 		}
1161 
1162 		/* Actions on further pages start at offset 0 */
1163 		p += PAGE_SIZE - offs;
1164 		offs = 0;
1165 		ma_offs += len;
1166 	}
1167 	vm_object_pip_wakeup(sc->object);
1168 	return (rv != VM_PAGER_ERROR ? 0 : ENOSPC);
1169 }
1170 
1171 static int
1172 mdstart_null(struct md_s *sc, struct bio *bp)
1173 {
1174 
1175 	switch (bp->bio_cmd) {
1176 	case BIO_READ:
1177 		bzero(bp->bio_data, bp->bio_length);
1178 		cpu_flush_dcache(bp->bio_data, bp->bio_length);
1179 		break;
1180 	case BIO_WRITE:
1181 		break;
1182 	}
1183 	bp->bio_resid = 0;
1184 	return (0);
1185 }
1186 
1187 static void
1188 md_kthread(void *arg)
1189 {
1190 	struct md_s *sc;
1191 	struct bio *bp;
1192 	int error;
1193 
1194 	sc = arg;
1195 	thread_lock(curthread);
1196 	sched_prio(curthread, PRIBIO);
1197 	thread_unlock(curthread);
1198 	if (sc->type == MD_VNODE)
1199 		curthread->td_pflags |= TDP_NORUNNINGBUF;
1200 
1201 	for (;;) {
1202 		mtx_lock(&sc->queue_mtx);
1203 		if (sc->flags & MD_SHUTDOWN) {
1204 			sc->flags |= MD_EXITING;
1205 			mtx_unlock(&sc->queue_mtx);
1206 			kproc_exit(0);
1207 		}
1208 		bp = bioq_takefirst(&sc->bio_queue);
1209 		if (!bp) {
1210 			msleep(sc, &sc->queue_mtx, PRIBIO | PDROP, "mdwait", 0);
1211 			continue;
1212 		}
1213 		mtx_unlock(&sc->queue_mtx);
1214 		if (bp->bio_cmd == BIO_GETATTR) {
1215 			int isv = ((sc->flags & MD_VERIFY) != 0);
1216 
1217 			if ((sc->fwsectors && sc->fwheads &&
1218 			    (g_handleattr_int(bp, "GEOM::fwsectors",
1219 			    sc->fwsectors) ||
1220 			    g_handleattr_int(bp, "GEOM::fwheads",
1221 			    sc->fwheads))) ||
1222 			    g_handleattr_int(bp, "GEOM::candelete", 1))
1223 				error = -1;
1224 			else if (sc->ident[0] != '\0' &&
1225 			    g_handleattr_str(bp, "GEOM::ident", sc->ident))
1226 				error = -1;
1227 			else if (g_handleattr_int(bp, "MNT::verified", isv))
1228 				error = -1;
1229 			else
1230 				error = EOPNOTSUPP;
1231 		} else {
1232 			error = sc->start(sc, bp);
1233 		}
1234 
1235 		if (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE) {
1236 			/*
1237 			 * Devstat uses (bio_bcount, bio_resid) for
1238 			 * determining the length of the completed part of
1239 			 * the i/o.  g_io_deliver() will translate from
1240 			 * bio_completed to that, but it also destroys the
1241 			 * bio so we must do our own translation.
1242 			 */
1243 			bp->bio_bcount = bp->bio_length;
1244 			bp->bio_resid = (error == -1 ? bp->bio_bcount : 0);
1245 			devstat_end_transaction_bio(sc->devstat, bp);
1246 		}
1247 		if (error != -1) {
1248 			bp->bio_completed = bp->bio_length;
1249 			g_io_deliver(bp, error);
1250 		}
1251 	}
1252 }
1253 
1254 static struct md_s *
1255 mdfind(int unit)
1256 {
1257 	struct md_s *sc;
1258 
1259 	LIST_FOREACH(sc, &md_softc_list, list) {
1260 		if (sc->unit == unit)
1261 			break;
1262 	}
1263 	return (sc);
1264 }
1265 
1266 static struct md_s *
1267 mdnew(int unit, int *errp, enum md_types type)
1268 {
1269 	struct md_s *sc;
1270 	int error;
1271 
1272 	*errp = 0;
1273 	if (unit == -1)
1274 		unit = alloc_unr(md_uh);
1275 	else
1276 		unit = alloc_unr_specific(md_uh, unit);
1277 
1278 	if (unit == -1) {
1279 		*errp = EBUSY;
1280 		return (NULL);
1281 	}
1282 
1283 	sc = (struct md_s *)malloc(sizeof *sc, M_MD, M_WAITOK | M_ZERO);
1284 	sc->type = type;
1285 	bioq_init(&sc->bio_queue);
1286 	mtx_init(&sc->queue_mtx, "md bio queue", NULL, MTX_DEF);
1287 	sc->unit = unit;
1288 	sprintf(sc->name, "md%d", unit);
1289 	LIST_INSERT_HEAD(&md_softc_list, sc, list);
1290 	error = kproc_create(md_kthread, sc, &sc->procp, 0, 0,"%s", sc->name);
1291 	if (error == 0)
1292 		return (sc);
1293 	LIST_REMOVE(sc, list);
1294 	mtx_destroy(&sc->queue_mtx);
1295 	free_unr(md_uh, sc->unit);
1296 	free(sc, M_MD);
1297 	*errp = error;
1298 	return (NULL);
1299 }
1300 
1301 static void
1302 mdinit(struct md_s *sc)
1303 {
1304 	struct g_geom *gp;
1305 	struct g_provider *pp;
1306 
1307 	g_topology_lock();
1308 	gp = g_new_geomf(&g_md_class, "md%d", sc->unit);
1309 	gp->softc = sc;
1310 	pp = g_new_providerf(gp, "md%d", sc->unit);
1311 	devstat_remove_entry(pp->stat);
1312 	pp->stat = NULL;
1313 	pp->flags |= G_PF_DIRECT_SEND | G_PF_DIRECT_RECEIVE;
1314 	pp->mediasize = sc->mediasize;
1315 	pp->sectorsize = sc->sectorsize;
1316 	switch (sc->type) {
1317 	case MD_MALLOC:
1318 	case MD_VNODE:
1319 	case MD_SWAP:
1320 		pp->flags |= G_PF_ACCEPT_UNMAPPED;
1321 		break;
1322 	case MD_PRELOAD:
1323 	case MD_NULL:
1324 		break;
1325 	}
1326 	sc->gp = gp;
1327 	sc->pp = pp;
1328 	sc->devstat = devstat_new_entry("md", sc->unit, sc->sectorsize,
1329 	    DEVSTAT_ALL_SUPPORTED, DEVSTAT_TYPE_DIRECT, DEVSTAT_PRIORITY_MAX);
1330 	sc->devstat->id = pp;
1331 	g_error_provider(pp, 0);
1332 	g_topology_unlock();
1333 }
1334 
1335 static int
1336 mdcreate_malloc(struct md_s *sc, struct md_req *mdr)
1337 {
1338 	uintptr_t sp;
1339 	int error;
1340 	off_t u;
1341 
1342 	error = 0;
1343 	if (mdr->md_options & ~(MD_AUTOUNIT | MD_COMPRESS | MD_RESERVE))
1344 		return (EINVAL);
1345 	if (mdr->md_sectorsize != 0 && !powerof2(mdr->md_sectorsize))
1346 		return (EINVAL);
1347 	/* Compression doesn't make sense if we have reserved space */
1348 	if (mdr->md_options & MD_RESERVE)
1349 		mdr->md_options &= ~MD_COMPRESS;
1350 	if (mdr->md_fwsectors != 0)
1351 		sc->fwsectors = mdr->md_fwsectors;
1352 	if (mdr->md_fwheads != 0)
1353 		sc->fwheads = mdr->md_fwheads;
1354 	sc->flags = mdr->md_options & (MD_COMPRESS | MD_FORCE);
1355 	sc->indir = dimension(sc->mediasize / sc->sectorsize);
1356 	sc->uma = uma_zcreate(sc->name, sc->sectorsize, NULL, NULL, NULL, NULL,
1357 	    0x1ff, 0);
1358 	if (mdr->md_options & MD_RESERVE) {
1359 		off_t nsectors;
1360 
1361 		nsectors = sc->mediasize / sc->sectorsize;
1362 		for (u = 0; u < nsectors; u++) {
1363 			sp = (uintptr_t)uma_zalloc(sc->uma, (md_malloc_wait ?
1364 			    M_WAITOK : M_NOWAIT) | M_ZERO);
1365 			if (sp != 0)
1366 				error = s_write(sc->indir, u, sp);
1367 			else
1368 				error = ENOMEM;
1369 			if (error != 0)
1370 				break;
1371 		}
1372 	}
1373 	return (error);
1374 }
1375 
1376 static int
1377 mdsetcred(struct md_s *sc, struct ucred *cred)
1378 {
1379 	char *tmpbuf;
1380 	int error = 0;
1381 
1382 	/*
1383 	 * Set credits in our softc
1384 	 */
1385 
1386 	if (sc->cred)
1387 		crfree(sc->cred);
1388 	sc->cred = crhold(cred);
1389 
1390 	/*
1391 	 * Horrible kludge to establish credentials for NFS  XXX.
1392 	 */
1393 
1394 	if (sc->vnode) {
1395 		struct uio auio;
1396 		struct iovec aiov;
1397 
1398 		tmpbuf = malloc(sc->sectorsize, M_TEMP, M_WAITOK);
1399 		bzero(&auio, sizeof(auio));
1400 
1401 		aiov.iov_base = tmpbuf;
1402 		aiov.iov_len = sc->sectorsize;
1403 		auio.uio_iov = &aiov;
1404 		auio.uio_iovcnt = 1;
1405 		auio.uio_offset = 0;
1406 		auio.uio_rw = UIO_READ;
1407 		auio.uio_segflg = UIO_SYSSPACE;
1408 		auio.uio_resid = aiov.iov_len;
1409 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY);
1410 		error = VOP_READ(sc->vnode, &auio, 0, sc->cred);
1411 		VOP_UNLOCK(sc->vnode);
1412 		free(tmpbuf, M_TEMP);
1413 	}
1414 	return (error);
1415 }
1416 
1417 static int
1418 mdcreate_vnode(struct md_s *sc, struct md_req *mdr, struct thread *td)
1419 {
1420 	struct vattr vattr;
1421 	struct nameidata nd;
1422 	char *fname;
1423 	int error, flags;
1424 
1425 	fname = mdr->md_file;
1426 	if (mdr->md_file_seg == UIO_USERSPACE) {
1427 		error = copyinstr(fname, sc->file, sizeof(sc->file), NULL);
1428 		if (error != 0)
1429 			return (error);
1430 	} else if (mdr->md_file_seg == UIO_SYSSPACE)
1431 		strlcpy(sc->file, fname, sizeof(sc->file));
1432 	else
1433 		return (EDOOFUS);
1434 
1435 	/*
1436 	 * If the user specified that this is a read only device, don't
1437 	 * set the FWRITE mask before trying to open the backing store.
1438 	 */
1439 	flags = FREAD | ((mdr->md_options & MD_READONLY) ? 0 : FWRITE) \
1440 	    | ((mdr->md_options & MD_VERIFY) ? O_VERIFY : 0);
1441 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, sc->file, td);
1442 	error = vn_open(&nd, &flags, 0, NULL);
1443 	if (error != 0)
1444 		return (error);
1445 	NDFREE(&nd, NDF_ONLY_PNBUF);
1446 	if (nd.ni_vp->v_type != VREG) {
1447 		error = EINVAL;
1448 		goto bad;
1449 	}
1450 	error = VOP_GETATTR(nd.ni_vp, &vattr, td->td_ucred);
1451 	if (error != 0)
1452 		goto bad;
1453 	if (VOP_ISLOCKED(nd.ni_vp) != LK_EXCLUSIVE) {
1454 		vn_lock(nd.ni_vp, LK_UPGRADE | LK_RETRY);
1455 		if (VN_IS_DOOMED(nd.ni_vp)) {
1456 			/* Forced unmount. */
1457 			error = EBADF;
1458 			goto bad;
1459 		}
1460 	}
1461 	nd.ni_vp->v_vflag |= VV_MD;
1462 	VOP_UNLOCK(nd.ni_vp);
1463 
1464 	if (mdr->md_fwsectors != 0)
1465 		sc->fwsectors = mdr->md_fwsectors;
1466 	if (mdr->md_fwheads != 0)
1467 		sc->fwheads = mdr->md_fwheads;
1468 	snprintf(sc->ident, sizeof(sc->ident), "MD-DEV%ju-INO%ju",
1469 	    (uintmax_t)vattr.va_fsid, (uintmax_t)vattr.va_fileid);
1470 	sc->flags = mdr->md_options & (MD_ASYNC | MD_CACHE | MD_FORCE |
1471 	    MD_VERIFY);
1472 	if (!(flags & FWRITE))
1473 		sc->flags |= MD_READONLY;
1474 	sc->vnode = nd.ni_vp;
1475 
1476 	error = mdsetcred(sc, td->td_ucred);
1477 	if (error != 0) {
1478 		sc->vnode = NULL;
1479 		vn_lock(nd.ni_vp, LK_EXCLUSIVE | LK_RETRY);
1480 		nd.ni_vp->v_vflag &= ~VV_MD;
1481 		goto bad;
1482 	}
1483 	return (0);
1484 bad:
1485 	VOP_UNLOCK(nd.ni_vp);
1486 	(void)vn_close(nd.ni_vp, flags, td->td_ucred, td);
1487 	return (error);
1488 }
1489 
1490 static void
1491 g_md_providergone(struct g_provider *pp)
1492 {
1493 	struct md_s *sc = pp->geom->softc;
1494 
1495 	mtx_lock(&sc->queue_mtx);
1496 	sc->flags |= MD_PROVIDERGONE;
1497 	wakeup(&sc->flags);
1498 	mtx_unlock(&sc->queue_mtx);
1499 }
1500 
1501 static int
1502 mddestroy(struct md_s *sc, struct thread *td)
1503 {
1504 
1505 	if (sc->gp) {
1506 		g_topology_lock();
1507 		g_wither_geom(sc->gp, ENXIO);
1508 		g_topology_unlock();
1509 
1510 		mtx_lock(&sc->queue_mtx);
1511 		while (!(sc->flags & MD_PROVIDERGONE))
1512 			msleep(&sc->flags, &sc->queue_mtx, PRIBIO, "mddestroy", 0);
1513 		mtx_unlock(&sc->queue_mtx);
1514 	}
1515 	if (sc->devstat) {
1516 		devstat_remove_entry(sc->devstat);
1517 		sc->devstat = NULL;
1518 	}
1519 	mtx_lock(&sc->queue_mtx);
1520 	sc->flags |= MD_SHUTDOWN;
1521 	wakeup(sc);
1522 	while (!(sc->flags & MD_EXITING))
1523 		msleep(sc->procp, &sc->queue_mtx, PRIBIO, "mddestroy", hz / 10);
1524 	mtx_unlock(&sc->queue_mtx);
1525 	mtx_destroy(&sc->queue_mtx);
1526 	if (sc->vnode != NULL) {
1527 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY);
1528 		sc->vnode->v_vflag &= ~VV_MD;
1529 		VOP_UNLOCK(sc->vnode);
1530 		(void)vn_close(sc->vnode, sc->flags & MD_READONLY ?
1531 		    FREAD : (FREAD|FWRITE), sc->cred, td);
1532 	}
1533 	if (sc->cred != NULL)
1534 		crfree(sc->cred);
1535 	if (sc->object != NULL)
1536 		vm_object_deallocate(sc->object);
1537 	if (sc->indir)
1538 		destroy_indir(sc, sc->indir);
1539 	if (sc->uma)
1540 		uma_zdestroy(sc->uma);
1541 
1542 	LIST_REMOVE(sc, list);
1543 	free_unr(md_uh, sc->unit);
1544 	free(sc, M_MD);
1545 	return (0);
1546 }
1547 
1548 static int
1549 mdresize(struct md_s *sc, struct md_req *mdr)
1550 {
1551 	int error, res;
1552 	vm_pindex_t oldpages, newpages;
1553 
1554 	switch (sc->type) {
1555 	case MD_VNODE:
1556 	case MD_NULL:
1557 		break;
1558 	case MD_SWAP:
1559 		if (mdr->md_mediasize <= 0 ||
1560 		    (mdr->md_mediasize % PAGE_SIZE) != 0)
1561 			return (EDOM);
1562 		oldpages = OFF_TO_IDX(sc->mediasize);
1563 		newpages = OFF_TO_IDX(mdr->md_mediasize);
1564 		if (newpages < oldpages) {
1565 			VM_OBJECT_WLOCK(sc->object);
1566 			vm_object_page_remove(sc->object, newpages, 0, 0);
1567 			swap_release_by_cred(IDX_TO_OFF(oldpages -
1568 			    newpages), sc->cred);
1569 			sc->object->charge = IDX_TO_OFF(newpages);
1570 			sc->object->size = newpages;
1571 			VM_OBJECT_WUNLOCK(sc->object);
1572 		} else if (newpages > oldpages) {
1573 			res = swap_reserve_by_cred(IDX_TO_OFF(newpages -
1574 			    oldpages), sc->cred);
1575 			if (!res)
1576 				return (ENOMEM);
1577 			if ((mdr->md_options & MD_RESERVE) ||
1578 			    (sc->flags & MD_RESERVE)) {
1579 				error = swap_pager_reserve(sc->object,
1580 				    oldpages, newpages - oldpages);
1581 				if (error < 0) {
1582 					swap_release_by_cred(
1583 					    IDX_TO_OFF(newpages - oldpages),
1584 					    sc->cred);
1585 					return (EDOM);
1586 				}
1587 			}
1588 			VM_OBJECT_WLOCK(sc->object);
1589 			sc->object->charge = IDX_TO_OFF(newpages);
1590 			sc->object->size = newpages;
1591 			VM_OBJECT_WUNLOCK(sc->object);
1592 		}
1593 		break;
1594 	default:
1595 		return (EOPNOTSUPP);
1596 	}
1597 
1598 	sc->mediasize = mdr->md_mediasize;
1599 	g_topology_lock();
1600 	g_resize_provider(sc->pp, sc->mediasize);
1601 	g_topology_unlock();
1602 	return (0);
1603 }
1604 
1605 static int
1606 mdcreate_swap(struct md_s *sc, struct md_req *mdr, struct thread *td)
1607 {
1608 	vm_ooffset_t npage;
1609 	int error;
1610 
1611 	/*
1612 	 * Range check.  Disallow negative sizes and sizes not being
1613 	 * multiple of page size.
1614 	 */
1615 	if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0)
1616 		return (EDOM);
1617 
1618 	/*
1619 	 * Allocate an OBJT_SWAP object.
1620 	 *
1621 	 * Note the truncation.
1622 	 */
1623 
1624 	if ((mdr->md_options & MD_VERIFY) != 0)
1625 		return (EINVAL);
1626 	npage = mdr->md_mediasize / PAGE_SIZE;
1627 	if (mdr->md_fwsectors != 0)
1628 		sc->fwsectors = mdr->md_fwsectors;
1629 	if (mdr->md_fwheads != 0)
1630 		sc->fwheads = mdr->md_fwheads;
1631 	sc->object = vm_pager_allocate(OBJT_SWAP, NULL, PAGE_SIZE * npage,
1632 	    VM_PROT_DEFAULT, 0, td->td_ucred);
1633 	if (sc->object == NULL)
1634 		return (ENOMEM);
1635 	sc->flags = mdr->md_options & (MD_FORCE | MD_RESERVE);
1636 	if (mdr->md_options & MD_RESERVE) {
1637 		if (swap_pager_reserve(sc->object, 0, npage) < 0) {
1638 			error = EDOM;
1639 			goto finish;
1640 		}
1641 	}
1642 	error = mdsetcred(sc, td->td_ucred);
1643  finish:
1644 	if (error != 0) {
1645 		vm_object_deallocate(sc->object);
1646 		sc->object = NULL;
1647 	}
1648 	return (error);
1649 }
1650 
1651 static int
1652 mdcreate_null(struct md_s *sc, struct md_req *mdr, struct thread *td)
1653 {
1654 
1655 	/*
1656 	 * Range check.  Disallow negative sizes and sizes not being
1657 	 * multiple of page size.
1658 	 */
1659 	if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0)
1660 		return (EDOM);
1661 
1662 	return (0);
1663 }
1664 
1665 static int
1666 kern_mdattach_locked(struct thread *td, struct md_req *mdr)
1667 {
1668 	struct md_s *sc;
1669 	unsigned sectsize;
1670 	int error, i;
1671 
1672 	sx_assert(&md_sx, SA_XLOCKED);
1673 
1674 	switch (mdr->md_type) {
1675 	case MD_MALLOC:
1676 	case MD_PRELOAD:
1677 	case MD_VNODE:
1678 	case MD_SWAP:
1679 	case MD_NULL:
1680 		break;
1681 	default:
1682 		return (EINVAL);
1683 	}
1684 	if (mdr->md_sectorsize == 0)
1685 		sectsize = DEV_BSIZE;
1686 	else
1687 		sectsize = mdr->md_sectorsize;
1688 	if (sectsize > maxphys || mdr->md_mediasize < sectsize)
1689 		return (EINVAL);
1690 	if (mdr->md_options & MD_AUTOUNIT)
1691 		sc = mdnew(-1, &error, mdr->md_type);
1692 	else {
1693 		if (mdr->md_unit > INT_MAX)
1694 			return (EINVAL);
1695 		sc = mdnew(mdr->md_unit, &error, mdr->md_type);
1696 	}
1697 	if (sc == NULL)
1698 		return (error);
1699 	if (mdr->md_label != NULL)
1700 		error = copyinstr(mdr->md_label, sc->label,
1701 		    sizeof(sc->label), NULL);
1702 	if (error != 0)
1703 		goto err_after_new;
1704 	if (mdr->md_options & MD_AUTOUNIT)
1705 		mdr->md_unit = sc->unit;
1706 	sc->mediasize = mdr->md_mediasize;
1707 	sc->sectorsize = sectsize;
1708 	error = EDOOFUS;
1709 	switch (sc->type) {
1710 	case MD_MALLOC:
1711 		sc->start = mdstart_malloc;
1712 		error = mdcreate_malloc(sc, mdr);
1713 		break;
1714 	case MD_PRELOAD:
1715 		/*
1716 		 * We disallow attaching preloaded memory disks via
1717 		 * ioctl. Preloaded memory disks are automatically
1718 		 * attached in g_md_init().
1719 		 */
1720 		error = EOPNOTSUPP;
1721 		break;
1722 	case MD_VNODE:
1723 		sc->start = mdstart_vnode;
1724 		error = mdcreate_vnode(sc, mdr, td);
1725 		break;
1726 	case MD_SWAP:
1727 		sc->start = mdstart_swap;
1728 		error = mdcreate_swap(sc, mdr, td);
1729 		break;
1730 	case MD_NULL:
1731 		sc->start = mdstart_null;
1732 		error = mdcreate_null(sc, mdr, td);
1733 		break;
1734 	}
1735 err_after_new:
1736 	if (error != 0) {
1737 		mddestroy(sc, td);
1738 		return (error);
1739 	}
1740 
1741 	/* Prune off any residual fractional sector */
1742 	i = sc->mediasize % sc->sectorsize;
1743 	sc->mediasize -= i;
1744 
1745 	mdinit(sc);
1746 	return (0);
1747 }
1748 
1749 static int
1750 kern_mdattach(struct thread *td, struct md_req *mdr)
1751 {
1752 	int error;
1753 
1754 	sx_xlock(&md_sx);
1755 	error = kern_mdattach_locked(td, mdr);
1756 	sx_xunlock(&md_sx);
1757 	return (error);
1758 }
1759 
1760 static int
1761 kern_mddetach_locked(struct thread *td, struct md_req *mdr)
1762 {
1763 	struct md_s *sc;
1764 
1765 	sx_assert(&md_sx, SA_XLOCKED);
1766 
1767 	if (mdr->md_mediasize != 0 ||
1768 	    (mdr->md_options & ~MD_FORCE) != 0)
1769 		return (EINVAL);
1770 
1771 	sc = mdfind(mdr->md_unit);
1772 	if (sc == NULL)
1773 		return (ENOENT);
1774 	if (sc->opencount != 0 && !(sc->flags & MD_FORCE) &&
1775 	    !(mdr->md_options & MD_FORCE))
1776 		return (EBUSY);
1777 	return (mddestroy(sc, td));
1778 }
1779 
1780 static int
1781 kern_mddetach(struct thread *td, struct md_req *mdr)
1782 {
1783 	int error;
1784 
1785 	sx_xlock(&md_sx);
1786 	error = kern_mddetach_locked(td, mdr);
1787 	sx_xunlock(&md_sx);
1788 	return (error);
1789 }
1790 
1791 static int
1792 kern_mdresize_locked(struct md_req *mdr)
1793 {
1794 	struct md_s *sc;
1795 
1796 	sx_assert(&md_sx, SA_XLOCKED);
1797 
1798 	if ((mdr->md_options & ~(MD_FORCE | MD_RESERVE)) != 0)
1799 		return (EINVAL);
1800 
1801 	sc = mdfind(mdr->md_unit);
1802 	if (sc == NULL)
1803 		return (ENOENT);
1804 	if (mdr->md_mediasize < sc->sectorsize)
1805 		return (EINVAL);
1806 	if (mdr->md_mediasize < sc->mediasize &&
1807 	    !(sc->flags & MD_FORCE) &&
1808 	    !(mdr->md_options & MD_FORCE))
1809 		return (EBUSY);
1810 	return (mdresize(sc, mdr));
1811 }
1812 
1813 static int
1814 kern_mdresize(struct md_req *mdr)
1815 {
1816 	int error;
1817 
1818 	sx_xlock(&md_sx);
1819 	error = kern_mdresize_locked(mdr);
1820 	sx_xunlock(&md_sx);
1821 	return (error);
1822 }
1823 
1824 static int
1825 kern_mdquery_locked(struct md_req *mdr)
1826 {
1827 	struct md_s *sc;
1828 	int error;
1829 
1830 	sx_assert(&md_sx, SA_XLOCKED);
1831 
1832 	sc = mdfind(mdr->md_unit);
1833 	if (sc == NULL)
1834 		return (ENOENT);
1835 	mdr->md_type = sc->type;
1836 	mdr->md_options = sc->flags;
1837 	mdr->md_mediasize = sc->mediasize;
1838 	mdr->md_sectorsize = sc->sectorsize;
1839 	error = 0;
1840 	if (mdr->md_label != NULL) {
1841 		error = copyout(sc->label, mdr->md_label,
1842 		    strlen(sc->label) + 1);
1843 		if (error != 0)
1844 			return (error);
1845 	}
1846 	if (sc->type == MD_VNODE ||
1847 	    (sc->type == MD_PRELOAD && mdr->md_file != NULL))
1848 		error = copyout(sc->file, mdr->md_file,
1849 		    strlen(sc->file) + 1);
1850 	return (error);
1851 }
1852 
1853 static int
1854 kern_mdquery(struct md_req *mdr)
1855 {
1856 	int error;
1857 
1858 	sx_xlock(&md_sx);
1859 	error = kern_mdquery_locked(mdr);
1860 	sx_xunlock(&md_sx);
1861 	return (error);
1862 }
1863 
1864 /* Copy members that are not userspace pointers. */
1865 #define	MD_IOCTL2REQ(mdio, mdr) do {					\
1866 	(mdr)->md_unit = (mdio)->md_unit;				\
1867 	(mdr)->md_type = (mdio)->md_type;				\
1868 	(mdr)->md_mediasize = (mdio)->md_mediasize;			\
1869 	(mdr)->md_sectorsize = (mdio)->md_sectorsize;			\
1870 	(mdr)->md_options = (mdio)->md_options;				\
1871 	(mdr)->md_fwheads = (mdio)->md_fwheads;				\
1872 	(mdr)->md_fwsectors = (mdio)->md_fwsectors;			\
1873 	(mdr)->md_units = &(mdio)->md_pad[0];				\
1874 	(mdr)->md_units_nitems = nitems((mdio)->md_pad);		\
1875 } while(0)
1876 
1877 /* Copy members that might have been updated */
1878 #define MD_REQ2IOCTL(mdr, mdio) do {					\
1879 	(mdio)->md_unit = (mdr)->md_unit;				\
1880 	(mdio)->md_type = (mdr)->md_type;				\
1881 	(mdio)->md_mediasize = (mdr)->md_mediasize;			\
1882 	(mdio)->md_sectorsize = (mdr)->md_sectorsize;			\
1883 	(mdio)->md_options = (mdr)->md_options;				\
1884 	(mdio)->md_fwheads = (mdr)->md_fwheads;				\
1885 	(mdio)->md_fwsectors = (mdr)->md_fwsectors;			\
1886 } while(0)
1887 
1888 static int
1889 mdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags,
1890     struct thread *td)
1891 {
1892 	struct md_req mdr;
1893 	int error;
1894 
1895 	if (md_debug)
1896 		printf("mdctlioctl(%s %lx %p %x %p)\n",
1897 			devtoname(dev), cmd, addr, flags, td);
1898 
1899 	bzero(&mdr, sizeof(mdr));
1900 	switch (cmd) {
1901 	case MDIOCATTACH:
1902 	case MDIOCDETACH:
1903 	case MDIOCRESIZE:
1904 	case MDIOCQUERY: {
1905 		struct md_ioctl *mdio = (struct md_ioctl *)addr;
1906 		if (mdio->md_version != MDIOVERSION)
1907 			return (EINVAL);
1908 		MD_IOCTL2REQ(mdio, &mdr);
1909 		mdr.md_file = mdio->md_file;
1910 		mdr.md_file_seg = UIO_USERSPACE;
1911 		/* If the file is adjacent to the md_ioctl it's in kernel. */
1912 		if ((void *)mdio->md_file == (void *)(mdio + 1))
1913 			mdr.md_file_seg = UIO_SYSSPACE;
1914 		mdr.md_label = mdio->md_label;
1915 		break;
1916 	}
1917 #ifdef COMPAT_FREEBSD32
1918 	case MDIOCATTACH_32:
1919 	case MDIOCDETACH_32:
1920 	case MDIOCRESIZE_32:
1921 	case MDIOCQUERY_32: {
1922 		struct md_ioctl32 *mdio = (struct md_ioctl32 *)addr;
1923 		if (mdio->md_version != MDIOVERSION)
1924 			return (EINVAL);
1925 		MD_IOCTL2REQ(mdio, &mdr);
1926 		mdr.md_file = (void *)(uintptr_t)mdio->md_file;
1927 		mdr.md_file_seg = UIO_USERSPACE;
1928 		mdr.md_label = (void *)(uintptr_t)mdio->md_label;
1929 		break;
1930 	}
1931 #endif
1932 	default:
1933 		/* Fall through to handler switch. */
1934 		break;
1935 	}
1936 
1937 	error = 0;
1938 	switch (cmd) {
1939 	case MDIOCATTACH:
1940 #ifdef COMPAT_FREEBSD32
1941 	case MDIOCATTACH_32:
1942 #endif
1943 		error = kern_mdattach(td, &mdr);
1944 		break;
1945 	case MDIOCDETACH:
1946 #ifdef COMPAT_FREEBSD32
1947 	case MDIOCDETACH_32:
1948 #endif
1949 		error = kern_mddetach(td, &mdr);
1950 		break;
1951 	case MDIOCRESIZE:
1952 #ifdef COMPAT_FREEBSD32
1953 	case MDIOCRESIZE_32:
1954 #endif
1955 		error = kern_mdresize(&mdr);
1956 		break;
1957 	case MDIOCQUERY:
1958 #ifdef COMPAT_FREEBSD32
1959 	case MDIOCQUERY_32:
1960 #endif
1961 		error = kern_mdquery(&mdr);
1962 		break;
1963 	default:
1964 		error = ENOIOCTL;
1965 	}
1966 
1967 	switch (cmd) {
1968 	case MDIOCATTACH:
1969 	case MDIOCQUERY: {
1970 		struct md_ioctl *mdio = (struct md_ioctl *)addr;
1971 		MD_REQ2IOCTL(&mdr, mdio);
1972 		break;
1973 	}
1974 #ifdef COMPAT_FREEBSD32
1975 	case MDIOCATTACH_32:
1976 	case MDIOCQUERY_32: {
1977 		struct md_ioctl32 *mdio = (struct md_ioctl32 *)addr;
1978 		MD_REQ2IOCTL(&mdr, mdio);
1979 		break;
1980 	}
1981 #endif
1982 	default:
1983 		/* Other commands to not alter mdr. */
1984 		break;
1985 	}
1986 
1987 	return (error);
1988 }
1989 
1990 static void
1991 md_preloaded(u_char *image, size_t length, const char *name)
1992 {
1993 	struct md_s *sc;
1994 	int error;
1995 
1996 	sc = mdnew(-1, &error, MD_PRELOAD);
1997 	if (sc == NULL)
1998 		return;
1999 	sc->mediasize = length;
2000 	sc->sectorsize = DEV_BSIZE;
2001 	sc->pl_ptr = image;
2002 	sc->pl_len = length;
2003 	sc->start = mdstart_preload;
2004 	if (name != NULL)
2005 		strlcpy(sc->file, name, sizeof(sc->file));
2006 #ifdef MD_ROOT
2007 	if (sc->unit == 0) {
2008 #ifndef ROOTDEVNAME
2009 		rootdevnames[0] = MD_ROOT_FSTYPE ":/dev/md0";
2010 #endif
2011 #ifdef MD_ROOT_READONLY
2012 		sc->flags |= MD_READONLY;
2013 #endif
2014 	}
2015 #endif
2016 	mdinit(sc);
2017 	if (name != NULL) {
2018 		printf("%s%d: Preloaded image <%s> %zd bytes at %p\n",
2019 		    MD_NAME, sc->unit, name, length, image);
2020 	} else {
2021 		printf("%s%d: Embedded image %zd bytes at %p\n",
2022 		    MD_NAME, sc->unit, length, image);
2023 	}
2024 }
2025 
2026 static void
2027 g_md_init(struct g_class *mp __unused)
2028 {
2029 	caddr_t mod;
2030 	u_char *ptr, *name, *type;
2031 	unsigned len;
2032 	int i;
2033 
2034 	/* figure out log2(NINDIR) */
2035 	for (i = NINDIR, nshift = -1; i; nshift++)
2036 		i >>= 1;
2037 
2038 	mod = NULL;
2039 	sx_init(&md_sx, "MD config lock");
2040 	g_topology_unlock();
2041 	md_uh = new_unrhdr(0, INT_MAX, NULL);
2042 #ifdef MD_ROOT
2043 	if (mfs_root_size != 0) {
2044 		sx_xlock(&md_sx);
2045 #ifdef MD_ROOT_MEM
2046 		md_preloaded(mfs_root, mfs_root_size, NULL);
2047 #else
2048 		md_preloaded(__DEVOLATILE(u_char *, &mfs_root), mfs_root_size,
2049 		    NULL);
2050 #endif
2051 		sx_xunlock(&md_sx);
2052 	}
2053 #endif
2054 	/* XXX: are preload_* static or do they need Giant ? */
2055 	while ((mod = preload_search_next_name(mod)) != NULL) {
2056 		name = (char *)preload_search_info(mod, MODINFO_NAME);
2057 		if (name == NULL)
2058 			continue;
2059 		type = (char *)preload_search_info(mod, MODINFO_TYPE);
2060 		if (type == NULL)
2061 			continue;
2062 		if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
2063 			continue;
2064 		ptr = preload_fetch_addr(mod);
2065 		len = preload_fetch_size(mod);
2066 		if (ptr != NULL && len != 0) {
2067 			sx_xlock(&md_sx);
2068 			md_preloaded(ptr, len, name);
2069 			sx_xunlock(&md_sx);
2070 		}
2071 	}
2072 	md_pbuf_zone = pbuf_zsecond_create("mdpbuf", nswbuf / 10);
2073 	status_dev = make_dev(&mdctl_cdevsw, INT_MAX, UID_ROOT, GID_WHEEL,
2074 	    0600, MDCTL_NAME);
2075 	g_topology_lock();
2076 }
2077 
2078 static void
2079 g_md_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
2080     struct g_consumer *cp __unused, struct g_provider *pp)
2081 {
2082 	struct md_s *mp;
2083 	char *type;
2084 
2085 	mp = gp->softc;
2086 	if (mp == NULL)
2087 		return;
2088 
2089 	switch (mp->type) {
2090 	case MD_MALLOC:
2091 		type = "malloc";
2092 		break;
2093 	case MD_PRELOAD:
2094 		type = "preload";
2095 		break;
2096 	case MD_VNODE:
2097 		type = "vnode";
2098 		break;
2099 	case MD_SWAP:
2100 		type = "swap";
2101 		break;
2102 	case MD_NULL:
2103 		type = "null";
2104 		break;
2105 	default:
2106 		type = "unknown";
2107 		break;
2108 	}
2109 
2110 	if (pp != NULL) {
2111 		if (indent == NULL) {
2112 			sbuf_printf(sb, " u %d", mp->unit);
2113 			sbuf_printf(sb, " s %ju", (uintmax_t) mp->sectorsize);
2114 			sbuf_printf(sb, " f %ju", (uintmax_t) mp->fwheads);
2115 			sbuf_printf(sb, " fs %ju", (uintmax_t) mp->fwsectors);
2116 			sbuf_printf(sb, " l %ju", (uintmax_t) mp->mediasize);
2117 			sbuf_printf(sb, " t %s", type);
2118 			if ((mp->type == MD_VNODE && mp->vnode != NULL) ||
2119 			    (mp->type == MD_PRELOAD && mp->file[0] != '\0'))
2120 				sbuf_printf(sb, " file %s", mp->file);
2121 			sbuf_printf(sb, " label %s", mp->label);
2122 		} else {
2123 			sbuf_printf(sb, "%s<unit>%d</unit>\n", indent,
2124 			    mp->unit);
2125 			sbuf_printf(sb, "%s<sectorsize>%ju</sectorsize>\n",
2126 			    indent, (uintmax_t) mp->sectorsize);
2127 			sbuf_printf(sb, "%s<fwheads>%ju</fwheads>\n",
2128 			    indent, (uintmax_t) mp->fwheads);
2129 			sbuf_printf(sb, "%s<fwsectors>%ju</fwsectors>\n",
2130 			    indent, (uintmax_t) mp->fwsectors);
2131 			if (mp->ident[0] != '\0') {
2132 				sbuf_printf(sb, "%s<ident>", indent);
2133 				g_conf_printf_escaped(sb, "%s", mp->ident);
2134 				sbuf_printf(sb, "</ident>\n");
2135 			}
2136 			sbuf_printf(sb, "%s<length>%ju</length>\n",
2137 			    indent, (uintmax_t) mp->mediasize);
2138 			sbuf_printf(sb, "%s<compression>%s</compression>\n", indent,
2139 			    (mp->flags & MD_COMPRESS) == 0 ? "off": "on");
2140 			sbuf_printf(sb, "%s<access>%s</access>\n", indent,
2141 			    (mp->flags & MD_READONLY) == 0 ? "read-write":
2142 			    "read-only");
2143 			sbuf_printf(sb, "%s<type>%s</type>\n", indent,
2144 			    type);
2145 			if ((mp->type == MD_VNODE && mp->vnode != NULL) ||
2146 			    (mp->type == MD_PRELOAD && mp->file[0] != '\0')) {
2147 				sbuf_printf(sb, "%s<file>", indent);
2148 				g_conf_printf_escaped(sb, "%s", mp->file);
2149 				sbuf_printf(sb, "</file>\n");
2150 			}
2151 			if (mp->type == MD_VNODE)
2152 				sbuf_printf(sb, "%s<cache>%s</cache>\n", indent,
2153 				    (mp->flags & MD_CACHE) == 0 ? "off": "on");
2154 			sbuf_printf(sb, "%s<label>", indent);
2155 			g_conf_printf_escaped(sb, "%s", mp->label);
2156 			sbuf_printf(sb, "</label>\n");
2157 		}
2158 	}
2159 }
2160 
2161 static void
2162 g_md_fini(struct g_class *mp __unused)
2163 {
2164 
2165 	sx_destroy(&md_sx);
2166 	if (status_dev != NULL)
2167 		destroy_dev(status_dev);
2168 	uma_zdestroy(md_pbuf_zone);
2169 	delete_unrhdr(md_uh);
2170 }
2171