xref: /freebsd/sys/dev/md/md.c (revision 6af83ee0d2941d18880b6aaa2b4facd1d30c6106)
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  *
22  * This code is derived from software contributed to Berkeley by
23  * the Systems Programming Group of the University of Utah Computer
24  * Science Department.
25  *
26  * Redistribution and use in source and binary forms, with or without
27  * modification, are permitted provided that the following conditions
28  * are met:
29  * 1. Redistributions of source code must retain the above copyright
30  *    notice, this list of conditions and the following disclaimer.
31  * 2. Redistributions in binary form must reproduce the above copyright
32  *    notice, this list of conditions and the following disclaimer in the
33  *    documentation and/or other materials provided with the distribution.
34  * 4. Neither the name of the University nor the names of its contributors
35  *    may be used to endorse or promote products derived from this software
36  *    without specific prior written permission.
37  *
38  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
39  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
40  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
41  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
42  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
43  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
44  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
45  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
46  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
47  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
48  * SUCH DAMAGE.
49  *
50  * from: Utah Hdr: vn.c 1.13 94/04/02
51  *
52  *	from: @(#)vn.c	8.6 (Berkeley) 4/1/94
53  * From: src/sys/dev/vn/vn.c,v 1.122 2000/12/16 16:06:03
54  */
55 
56 #include "opt_geom.h"
57 #include "opt_md.h"
58 
59 #include <sys/param.h>
60 #include <sys/systm.h>
61 #include <sys/bio.h>
62 #include <sys/conf.h>
63 #include <sys/fcntl.h>
64 #include <sys/kernel.h>
65 #include <sys/kthread.h>
66 #include <sys/linker.h>
67 #include <sys/lock.h>
68 #include <sys/malloc.h>
69 #include <sys/mdioctl.h>
70 #include <sys/mutex.h>
71 #include <sys/namei.h>
72 #include <sys/proc.h>
73 #include <sys/queue.h>
74 #include <sys/sched.h>
75 #include <sys/sf_buf.h>
76 #include <sys/sysctl.h>
77 #include <sys/vnode.h>
78 
79 #include <geom/geom.h>
80 
81 #include <vm/vm.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_page.h>
84 #include <vm/vm_pager.h>
85 #include <vm/swap_pager.h>
86 #include <vm/uma.h>
87 
88 #define MD_MODVER 1
89 
90 #define MD_SHUTDOWN 0x10000	/* Tell worker thread to terminate. */
91 
92 #ifndef MD_NSECT
93 #define MD_NSECT (10000 * 2)
94 #endif
95 
96 static MALLOC_DEFINE(M_MD, "MD disk", "Memory Disk");
97 static MALLOC_DEFINE(M_MDSECT, "MD sectors", "Memory Disk Sectors");
98 
99 static int md_debug;
100 SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0, "");
101 
102 #if defined(MD_ROOT) && defined(MD_ROOT_SIZE)
103 /* Image gets put here: */
104 static u_char mfs_root[MD_ROOT_SIZE*1024] = "MFS Filesystem goes here";
105 static u_char end_mfs_root[] __unused = "MFS Filesystem had better STOP here";
106 #endif
107 
108 static g_init_t g_md_init;
109 static g_fini_t g_md_fini;
110 static g_start_t g_md_start;
111 static g_access_t g_md_access;
112 
113 static int	mdunits;
114 static struct cdev *status_dev = 0;
115 
116 static d_ioctl_t mdctlioctl;
117 
118 static struct cdevsw mdctl_cdevsw = {
119 	.d_version =	D_VERSION,
120 	.d_flags =	D_NEEDGIANT,
121 	.d_ioctl =	mdctlioctl,
122 	.d_name =	MD_NAME,
123 };
124 
125 struct g_class g_md_class = {
126 	.name = "MD",
127 	.version = G_VERSION,
128 	.init = g_md_init,
129 	.fini = g_md_fini,
130 	.start = g_md_start,
131 	.access = g_md_access,
132 };
133 
134 DECLARE_GEOM_CLASS(g_md_class, g_md);
135 
136 
137 static LIST_HEAD(, md_s) md_softc_list = LIST_HEAD_INITIALIZER(&md_softc_list);
138 
139 #define NINDIR	(PAGE_SIZE / sizeof(uintptr_t))
140 #define NMASK	(NINDIR-1)
141 static int nshift;
142 
143 struct indir {
144 	uintptr_t	*array;
145 	u_int		total;
146 	u_int		used;
147 	u_int		shift;
148 };
149 
150 struct md_s {
151 	int unit;
152 	LIST_ENTRY(md_s) list;
153 	struct bio_queue_head bio_queue;
154 	struct mtx queue_mtx;
155 	struct cdev *dev;
156 	enum md_types type;
157 	off_t mediasize;
158 	unsigned sectorsize;
159 	unsigned opencount;
160 	unsigned fwheads;
161 	unsigned fwsectors;
162 	unsigned flags;
163 	char name[20];
164 	struct proc *procp;
165 	struct g_geom *gp;
166 	struct g_provider *pp;
167 
168 	/* MD_MALLOC related fields */
169 	struct indir *indir;
170 	uma_zone_t uma;
171 
172 	/* MD_PRELOAD related fields */
173 	u_char *pl_ptr;
174 	size_t pl_len;
175 
176 	/* MD_VNODE related fields */
177 	struct vnode *vnode;
178 	char file[PATH_MAX];
179 	struct ucred *cred;
180 
181 	/* MD_SWAP related fields */
182 	vm_object_t object;
183 };
184 
185 static int mddestroy(struct md_s *sc, struct thread *td);
186 
187 static struct indir *
188 new_indir(u_int shift)
189 {
190 	struct indir *ip;
191 
192 	ip = malloc(sizeof *ip, M_MD, M_NOWAIT | M_ZERO);
193 	if (ip == NULL)
194 		return (NULL);
195 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
196 	    M_MDSECT, M_NOWAIT | M_ZERO);
197 	if (ip->array == NULL) {
198 		free(ip, M_MD);
199 		return (NULL);
200 	}
201 	ip->total = NINDIR;
202 	ip->shift = shift;
203 	return (ip);
204 }
205 
206 static void
207 del_indir(struct indir *ip)
208 {
209 
210 	free(ip->array, M_MDSECT);
211 	free(ip, M_MD);
212 }
213 
214 static void
215 destroy_indir(struct md_s *sc, struct indir *ip)
216 {
217 	int i;
218 
219 	for (i = 0; i < NINDIR; i++) {
220 		if (!ip->array[i])
221 			continue;
222 		if (ip->shift)
223 			destroy_indir(sc, (struct indir*)(ip->array[i]));
224 		else if (ip->array[i] > 255)
225 			uma_zfree(sc->uma, (void *)(ip->array[i]));
226 	}
227 	del_indir(ip);
228 }
229 
230 /*
231  * This function does the math and alloctes the top level "indir" structure
232  * for a device of "size" sectors.
233  */
234 
235 static struct indir *
236 dimension(off_t size)
237 {
238 	off_t rcnt;
239 	struct indir *ip;
240 	int i, layer;
241 
242 	rcnt = size;
243 	layer = 0;
244 	while (rcnt > NINDIR) {
245 		rcnt /= NINDIR;
246 		layer++;
247 	}
248 	/* figure out log2(NINDIR) */
249 	for (i = NINDIR, nshift = -1; i; nshift++)
250 		i >>= 1;
251 
252 	/*
253 	 * XXX: the top layer is probably not fully populated, so we allocate
254 	 * too much space for ip->array in here.
255 	 */
256 	ip = malloc(sizeof *ip, M_MD, M_WAITOK | M_ZERO);
257 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
258 	    M_MDSECT, M_WAITOK | M_ZERO);
259 	ip->total = NINDIR;
260 	ip->shift = layer * nshift;
261 	return (ip);
262 }
263 
264 /*
265  * Read a given sector
266  */
267 
268 static uintptr_t
269 s_read(struct indir *ip, off_t offset)
270 {
271 	struct indir *cip;
272 	int idx;
273 	uintptr_t up;
274 
275 	if (md_debug > 1)
276 		printf("s_read(%jd)\n", (intmax_t)offset);
277 	up = 0;
278 	for (cip = ip; cip != NULL;) {
279 		if (cip->shift) {
280 			idx = (offset >> cip->shift) & NMASK;
281 			up = cip->array[idx];
282 			cip = (struct indir *)up;
283 			continue;
284 		}
285 		idx = offset & NMASK;
286 		return (cip->array[idx]);
287 	}
288 	return (0);
289 }
290 
291 /*
292  * Write a given sector, prune the tree if the value is 0
293  */
294 
295 static int
296 s_write(struct indir *ip, off_t offset, uintptr_t ptr)
297 {
298 	struct indir *cip, *lip[10];
299 	int idx, li;
300 	uintptr_t up;
301 
302 	if (md_debug > 1)
303 		printf("s_write(%jd, %p)\n", (intmax_t)offset, (void *)ptr);
304 	up = 0;
305 	li = 0;
306 	cip = ip;
307 	for (;;) {
308 		lip[li++] = cip;
309 		if (cip->shift) {
310 			idx = (offset >> cip->shift) & NMASK;
311 			up = cip->array[idx];
312 			if (up != 0) {
313 				cip = (struct indir *)up;
314 				continue;
315 			}
316 			/* Allocate branch */
317 			cip->array[idx] =
318 			    (uintptr_t)new_indir(cip->shift - nshift);
319 			if (cip->array[idx] == 0)
320 				return (ENOSPC);
321 			cip->used++;
322 			up = cip->array[idx];
323 			cip = (struct indir *)up;
324 			continue;
325 		}
326 		/* leafnode */
327 		idx = offset & NMASK;
328 		up = cip->array[idx];
329 		if (up != 0)
330 			cip->used--;
331 		cip->array[idx] = ptr;
332 		if (ptr != 0)
333 			cip->used++;
334 		break;
335 	}
336 	if (cip->used != 0 || li == 1)
337 		return (0);
338 	li--;
339 	while (cip->used == 0 && cip != ip) {
340 		li--;
341 		idx = (offset >> lip[li]->shift) & NMASK;
342 		up = lip[li]->array[idx];
343 		KASSERT(up == (uintptr_t)cip, ("md screwed up"));
344 		del_indir(cip);
345 		lip[li]->array[idx] = 0;
346 		lip[li]->used--;
347 		cip = lip[li];
348 	}
349 	return (0);
350 }
351 
352 
353 static int
354 g_md_access(struct g_provider *pp, int r, int w, int e)
355 {
356 	struct md_s *sc;
357 
358 	sc = pp->geom->softc;
359 	if (sc == NULL)
360 		return (ENXIO);
361 	r += pp->acr;
362 	w += pp->acw;
363 	e += pp->ace;
364 	if ((pp->acr + pp->acw + pp->ace) == 0 && (r + w + e) > 0) {
365 		sc->opencount = 1;
366 	} else if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0) {
367 		sc->opencount = 0;
368 	}
369 	return (0);
370 }
371 
372 static void
373 g_md_start(struct bio *bp)
374 {
375 	struct md_s *sc;
376 
377 	sc = bp->bio_to->geom->softc;
378 	mtx_lock(&sc->queue_mtx);
379 	bioq_disksort(&sc->bio_queue, bp);
380 	mtx_unlock(&sc->queue_mtx);
381 	wakeup(sc);
382 }
383 
384 
385 
386 static int
387 mdstart_malloc(struct md_s *sc, struct bio *bp)
388 {
389 	int i, error;
390 	u_char *dst;
391 	off_t secno, nsec, uc;
392 	uintptr_t sp, osp;
393 
394 	nsec = bp->bio_length / sc->sectorsize;
395 	secno = bp->bio_offset / sc->sectorsize;
396 	dst = bp->bio_data;
397 	error = 0;
398 	while (nsec--) {
399 		osp = s_read(sc->indir, secno);
400 		if (bp->bio_cmd == BIO_DELETE) {
401 			if (osp != 0)
402 				error = s_write(sc->indir, secno, 0);
403 		} else if (bp->bio_cmd == BIO_READ) {
404 			if (osp == 0)
405 				bzero(dst, sc->sectorsize);
406 			else if (osp <= 255)
407 				for (i = 0; i < sc->sectorsize; i++)
408 					dst[i] = osp;
409 			else
410 				bcopy((void *)osp, dst, sc->sectorsize);
411 			osp = 0;
412 		} else if (bp->bio_cmd == BIO_WRITE) {
413 			if (sc->flags & MD_COMPRESS) {
414 				uc = dst[0];
415 				for (i = 1; i < sc->sectorsize; i++)
416 					if (dst[i] != uc)
417 						break;
418 			} else {
419 				i = 0;
420 				uc = 0;
421 			}
422 			if (i == sc->sectorsize) {
423 				if (osp != uc)
424 					error = s_write(sc->indir, secno, uc);
425 			} else {
426 				if (osp <= 255) {
427 					sp = (uintptr_t)uma_zalloc(sc->uma,
428 					    M_NOWAIT);
429 					if (sp == 0) {
430 						error = ENOSPC;
431 						break;
432 					}
433 					bcopy(dst, (void *)sp, sc->sectorsize);
434 					error = s_write(sc->indir, secno, sp);
435 				} else {
436 					bcopy(dst, (void *)osp, sc->sectorsize);
437 					osp = 0;
438 				}
439 			}
440 		} else {
441 			error = EOPNOTSUPP;
442 		}
443 		if (osp > 255)
444 			uma_zfree(sc->uma, (void*)osp);
445 		if (error != 0)
446 			break;
447 		secno++;
448 		dst += sc->sectorsize;
449 	}
450 	bp->bio_resid = 0;
451 	return (error);
452 }
453 
454 static int
455 mdstart_preload(struct md_s *sc, struct bio *bp)
456 {
457 
458 	switch (bp->bio_cmd) {
459 	case BIO_READ:
460 		bcopy(sc->pl_ptr + bp->bio_offset, bp->bio_data,
461 		    bp->bio_length);
462 		break;
463 	case BIO_WRITE:
464 		bcopy(bp->bio_data, sc->pl_ptr + bp->bio_offset,
465 		    bp->bio_length);
466 		break;
467 	}
468 	bp->bio_resid = 0;
469 	return (0);
470 }
471 
472 static int
473 mdstart_vnode(struct md_s *sc, struct bio *bp)
474 {
475 	int error;
476 	struct uio auio;
477 	struct iovec aiov;
478 	struct mount *mp;
479 
480 	/*
481 	 * VNODE I/O
482 	 *
483 	 * If an error occurs, we set BIO_ERROR but we do not set
484 	 * B_INVAL because (for a write anyway), the buffer is
485 	 * still valid.
486 	 */
487 
488 	bzero(&auio, sizeof(auio));
489 
490 	aiov.iov_base = bp->bio_data;
491 	aiov.iov_len = bp->bio_length;
492 	auio.uio_iov = &aiov;
493 	auio.uio_iovcnt = 1;
494 	auio.uio_offset = (vm_ooffset_t)bp->bio_offset;
495 	auio.uio_segflg = UIO_SYSSPACE;
496 	if(bp->bio_cmd == BIO_READ)
497 		auio.uio_rw = UIO_READ;
498 	else if(bp->bio_cmd == BIO_WRITE)
499 		auio.uio_rw = UIO_WRITE;
500 	else
501 		panic("wrong BIO_OP in mdstart_vnode");
502 	auio.uio_resid = bp->bio_length;
503 	auio.uio_td = curthread;
504 	/*
505 	 * When reading set IO_DIRECT to try to avoid double-caching
506 	 * the data.  When writing IO_DIRECT is not optimal.
507 	 */
508 	if (bp->bio_cmd == BIO_READ) {
509 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY, curthread);
510 		error = VOP_READ(sc->vnode, &auio, IO_DIRECT, sc->cred);
511 		VOP_UNLOCK(sc->vnode, 0, curthread);
512 	} else {
513 		(void) vn_start_write(sc->vnode, &mp, V_WAIT);
514 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY, curthread);
515 		error = VOP_WRITE(sc->vnode, &auio,
516 		    sc->flags & MD_ASYNC ? 0 : IO_SYNC, sc->cred);
517 		VOP_UNLOCK(sc->vnode, 0, curthread);
518 		vn_finished_write(mp);
519 	}
520 	bp->bio_resid = auio.uio_resid;
521 	return (error);
522 }
523 
524 static int
525 mdstart_swap(struct md_s *sc, struct bio *bp)
526 {
527 	struct sf_buf *sf;
528 	int rv, offs, len, lastend;
529 	vm_pindex_t i, lastp;
530 	vm_page_t m;
531 	u_char *p;
532 
533 	p = bp->bio_data;
534 
535 	/*
536 	 * offs is the ofset at whih to start operating on the
537 	 * next (ie, first) page.  lastp is the last page on
538 	 * which we're going to operate.  lastend is the ending
539 	 * position within that last page (ie, PAGE_SIZE if
540 	 * we're operating on complete aligned pages).
541 	 */
542 	offs = bp->bio_offset % PAGE_SIZE;
543 	lastp = (bp->bio_offset + bp->bio_length - 1) / PAGE_SIZE;
544 	lastend = (bp->bio_offset + bp->bio_length - 1) % PAGE_SIZE + 1;
545 
546 	rv = VM_PAGER_OK;
547 	VM_OBJECT_LOCK(sc->object);
548 	vm_object_pip_add(sc->object, 1);
549 	for (i = bp->bio_offset / PAGE_SIZE; i <= lastp; i++) {
550 		len = ((i == lastp) ? lastend : PAGE_SIZE) - offs;
551 
552 		m = vm_page_grab(sc->object, i,
553 		    VM_ALLOC_NORMAL|VM_ALLOC_RETRY);
554 		VM_OBJECT_UNLOCK(sc->object);
555 		sched_pin();
556 		sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
557 		VM_OBJECT_LOCK(sc->object);
558 		if (bp->bio_cmd == BIO_READ) {
559 			if (m->valid != VM_PAGE_BITS_ALL)
560 				rv = vm_pager_get_pages(sc->object, &m, 1, 0);
561 			if (rv == VM_PAGER_ERROR) {
562 				sf_buf_free(sf);
563 				sched_unpin();
564 				vm_page_lock_queues();
565 				vm_page_wakeup(m);
566 				vm_page_unlock_queues();
567 				break;
568 			}
569 			bcopy((void *)(sf_buf_kva(sf) + offs), p, len);
570 		} else if (bp->bio_cmd == BIO_WRITE) {
571 			if (len != PAGE_SIZE && m->valid != VM_PAGE_BITS_ALL)
572 				rv = vm_pager_get_pages(sc->object, &m, 1, 0);
573 			if (rv == VM_PAGER_ERROR) {
574 				sf_buf_free(sf);
575 				sched_unpin();
576 				vm_page_lock_queues();
577 				vm_page_wakeup(m);
578 				vm_page_unlock_queues();
579 				break;
580 			}
581 			bcopy(p, (void *)(sf_buf_kva(sf) + offs), len);
582 			m->valid = VM_PAGE_BITS_ALL;
583 #if 0
584 		} else if (bp->bio_cmd == BIO_DELETE) {
585 			if (len != PAGE_SIZE && m->valid != VM_PAGE_BITS_ALL)
586 				rv = vm_pager_get_pages(sc->object, &m, 1, 0);
587 			if (rv == VM_PAGER_ERROR) {
588 				sf_buf_free(sf);
589 				sched_unpin();
590 				vm_page_lock_queues();
591 				vm_page_wakeup(m);
592 				vm_page_unlock_queues();
593 				break;
594 			}
595 			bzero((void *)(sf_buf_kva(sf) + offs), len);
596 			vm_page_dirty(m);
597 			m->valid = VM_PAGE_BITS_ALL;
598 #endif
599 		}
600 		sf_buf_free(sf);
601 		sched_unpin();
602 		vm_page_lock_queues();
603 		vm_page_wakeup(m);
604 		vm_page_activate(m);
605 		if (bp->bio_cmd == BIO_WRITE)
606 			vm_page_dirty(m);
607 		vm_page_unlock_queues();
608 
609 		/* Actions on further pages start at offset 0 */
610 		p += PAGE_SIZE - offs;
611 		offs = 0;
612 #if 0
613 if (bootverbose || bp->bio_offset / PAGE_SIZE < 17)
614 printf("wire_count %d busy %d flags %x hold_count %d act_count %d queue %d valid %d dirty %d @ %d\n",
615     m->wire_count, m->busy,
616     m->flags, m->hold_count, m->act_count, m->queue, m->valid, m->dirty, i);
617 #endif
618 	}
619 	vm_object_pip_subtract(sc->object, 1);
620 	vm_object_set_writeable_dirty(sc->object);
621 	VM_OBJECT_UNLOCK(sc->object);
622 	return (rv != VM_PAGER_ERROR ? 0 : ENOSPC);
623 }
624 
625 static void
626 md_kthread(void *arg)
627 {
628 	struct md_s *sc;
629 	struct bio *bp;
630 	int error, hasgiant;
631 
632 	sc = arg;
633 	mtx_lock_spin(&sched_lock);
634 	sched_prio(curthread, PRIBIO);
635 	mtx_unlock_spin(&sched_lock);
636 
637 	switch (sc->type) {
638 	case MD_VNODE:
639 		mtx_lock(&Giant);
640 		hasgiant = 1;
641 		break;
642 	case MD_MALLOC:
643 	case MD_PRELOAD:
644 	case MD_SWAP:
645 	default:
646 		hasgiant = 0;
647 		break;
648 	}
649 
650 	for (;;) {
651 		mtx_lock(&sc->queue_mtx);
652 		bp = bioq_takefirst(&sc->bio_queue);
653 		if (!bp) {
654 			if (sc->flags & MD_SHUTDOWN) {
655 				mtx_unlock(&sc->queue_mtx);
656 				sc->procp = NULL;
657 				wakeup(&sc->procp);
658 				if (hasgiant)
659 					mtx_unlock(&Giant);
660 				kthread_exit(0);
661 			}
662 			msleep(sc, &sc->queue_mtx, PRIBIO | PDROP, "mdwait", 0);
663 			continue;
664 		}
665 		mtx_unlock(&sc->queue_mtx);
666 		if (bp->bio_cmd == BIO_GETATTR) {
667 			if (sc->fwsectors && sc->fwheads &&
668 			    (g_handleattr_int(bp, "GEOM::fwsectors",
669 			    sc->fwsectors) ||
670 			    g_handleattr_int(bp, "GEOM::fwheads",
671 			    sc->fwheads)))
672 				error = -1;
673 			else
674 				error = EOPNOTSUPP;
675 		} else {
676 			switch (sc->type) {
677 			case MD_MALLOC:
678 				error = mdstart_malloc(sc, bp);
679 				break;
680 			case MD_PRELOAD:
681 				error = mdstart_preload(sc, bp);
682 				break;
683 			case MD_VNODE:
684 				mtx_assert(&Giant, MA_OWNED);
685 				error = mdstart_vnode(sc, bp);
686 				mtx_assert(&Giant, MA_OWNED);
687 				break;
688 			case MD_SWAP:
689 				error = mdstart_swap(sc, bp);
690 				break;
691 			default:
692 				panic("Impossible md(type)");
693 				break;
694 			}
695 		}
696 
697 		if (error != -1) {
698 			bp->bio_completed = bp->bio_length;
699 			g_io_deliver(bp, error);
700 		}
701 	}
702 }
703 
704 static struct md_s *
705 mdfind(int unit)
706 {
707 	struct md_s *sc;
708 
709 	/* XXX: LOCK(unique unit numbers) */
710 	LIST_FOREACH(sc, &md_softc_list, list) {
711 		if (sc->unit == unit)
712 			break;
713 	}
714 	/* XXX: UNLOCK(unique unit numbers) */
715 	return (sc);
716 }
717 
718 static struct md_s *
719 mdnew(int unit)
720 {
721 	struct md_s *sc;
722 	int error, max = -1;
723 
724 	/* XXX: LOCK(unique unit numbers) */
725 	LIST_FOREACH(sc, &md_softc_list, list) {
726 		if (sc->unit == unit) {
727 			/* XXX: UNLOCK(unique unit numbers) */
728 			return (NULL);
729 		}
730 		if (sc->unit > max)
731 			max = sc->unit;
732 	}
733 	if (unit == -1)
734 		unit = max + 1;
735 	sc = (struct md_s *)malloc(sizeof *sc, M_MD, M_WAITOK | M_ZERO);
736 	sc->unit = unit;
737 	bioq_init(&sc->bio_queue);
738 	mtx_init(&sc->queue_mtx, "md bio queue", NULL, MTX_DEF);
739 	sprintf(sc->name, "md%d", unit);
740 	error = kthread_create(md_kthread, sc, &sc->procp, 0, 0,"%s", sc->name);
741 	if (error != 0) {
742 		free(sc, M_MD);
743 		return (NULL);
744 	}
745 	LIST_INSERT_HEAD(&md_softc_list, sc, list);
746 	/* XXX: UNLOCK(unique unit numbers) */
747 	return (sc);
748 }
749 
750 static void
751 mdinit(struct md_s *sc)
752 {
753 
754 	struct g_geom *gp;
755 	struct g_provider *pp;
756 
757 	DROP_GIANT();
758 	g_topology_lock();
759 	gp = g_new_geomf(&g_md_class, "md%d", sc->unit);
760 	gp->softc = sc;
761 	pp = g_new_providerf(gp, "md%d", sc->unit);
762 	pp->mediasize = sc->mediasize;
763 	pp->sectorsize = sc->sectorsize;
764 	sc->gp = gp;
765 	sc->pp = pp;
766 	g_error_provider(pp, 0);
767 	g_topology_unlock();
768 	PICKUP_GIANT();
769 }
770 
771 /*
772  * XXX: we should check that the range they feed us is mapped.
773  * XXX: we should implement read-only.
774  */
775 
776 static int
777 mdcreate_preload(struct md_s *sc, struct md_ioctl *mdio)
778 {
779 
780 	if (mdio->md_options & ~(MD_AUTOUNIT | MD_FORCE))
781 		return (EINVAL);
782 	sc->flags = mdio->md_options & MD_FORCE;
783 	/* Cast to pointer size, then to pointer to avoid warning */
784 	sc->pl_ptr = (u_char *)(uintptr_t)mdio->md_base;
785 	sc->pl_len = (size_t)sc->mediasize;
786 	return (0);
787 }
788 
789 
790 static int
791 mdcreate_malloc(struct md_s *sc, struct md_ioctl *mdio)
792 {
793 	uintptr_t sp;
794 	int error;
795 	off_t u;
796 
797 	error = 0;
798 	if (mdio->md_options & ~(MD_AUTOUNIT | MD_COMPRESS | MD_RESERVE))
799 		return (EINVAL);
800 	if (mdio->md_sectorsize != 0 && !powerof2(mdio->md_sectorsize))
801 		return (EINVAL);
802 	/* Compression doesn't make sense if we have reserved space */
803 	if (mdio->md_options & MD_RESERVE)
804 		mdio->md_options &= ~MD_COMPRESS;
805 	if (mdio->md_fwsectors != 0)
806 		sc->fwsectors = mdio->md_fwsectors;
807 	if (mdio->md_fwheads != 0)
808 		sc->fwheads = mdio->md_fwheads;
809 	sc->flags = mdio->md_options & (MD_COMPRESS | MD_FORCE);
810 	sc->indir = dimension(sc->mediasize / sc->sectorsize);
811 	sc->uma = uma_zcreate(sc->name, sc->sectorsize, NULL, NULL, NULL, NULL,
812 	    0x1ff, 0);
813 	if (mdio->md_options & MD_RESERVE) {
814 		off_t nsectors;
815 
816 		nsectors = sc->mediasize / sc->sectorsize;
817 		for (u = 0; u < nsectors; u++) {
818 			sp = (uintptr_t)uma_zalloc(sc->uma, M_NOWAIT | M_ZERO);
819 			if (sp != 0)
820 				error = s_write(sc->indir, u, sp);
821 			else
822 				error = ENOMEM;
823 			if (error != 0)
824 				break;
825 		}
826 	}
827 	return (error);
828 }
829 
830 
831 static int
832 mdsetcred(struct md_s *sc, struct ucred *cred)
833 {
834 	char *tmpbuf;
835 	int error = 0;
836 
837 	/*
838 	 * Set credits in our softc
839 	 */
840 
841 	if (sc->cred)
842 		crfree(sc->cred);
843 	sc->cred = crhold(cred);
844 
845 	/*
846 	 * Horrible kludge to establish credentials for NFS  XXX.
847 	 */
848 
849 	if (sc->vnode) {
850 		struct uio auio;
851 		struct iovec aiov;
852 
853 		tmpbuf = malloc(sc->sectorsize, M_TEMP, M_WAITOK);
854 		bzero(&auio, sizeof(auio));
855 
856 		aiov.iov_base = tmpbuf;
857 		aiov.iov_len = sc->sectorsize;
858 		auio.uio_iov = &aiov;
859 		auio.uio_iovcnt = 1;
860 		auio.uio_offset = 0;
861 		auio.uio_rw = UIO_READ;
862 		auio.uio_segflg = UIO_SYSSPACE;
863 		auio.uio_resid = aiov.iov_len;
864 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY, curthread);
865 		error = VOP_READ(sc->vnode, &auio, 0, sc->cred);
866 		VOP_UNLOCK(sc->vnode, 0, curthread);
867 		free(tmpbuf, M_TEMP);
868 	}
869 	return (error);
870 }
871 
872 static int
873 mdcreate_vnode(struct md_s *sc, struct md_ioctl *mdio, struct thread *td)
874 {
875 	struct vattr vattr;
876 	struct nameidata nd;
877 	int error, flags;
878 
879 	error = copyinstr(mdio->md_file, sc->file, sizeof(sc->file), NULL);
880 	if (error != 0)
881 		return (error);
882 	flags = FREAD|FWRITE;
883 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, sc->file, td);
884 	error = vn_open(&nd, &flags, 0, -1);
885 	if (error != 0) {
886 		NDFREE(&nd, NDF_ONLY_PNBUF);
887 		if (error != EACCES && error != EPERM && error != EROFS)
888 			return (error);
889 		flags &= ~FWRITE;
890 		NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, sc->file, td);
891 		error = vn_open(&nd, &flags, 0, -1);
892 	}
893 	NDFREE(&nd, NDF_ONLY_PNBUF);
894 	if (error != 0)
895 		return (error);
896 	if (nd.ni_vp->v_type != VREG ||
897 	    (error = VOP_GETATTR(nd.ni_vp, &vattr, td->td_ucred, td))) {
898 		VOP_UNLOCK(nd.ni_vp, 0, td);
899 		(void)vn_close(nd.ni_vp, flags, td->td_ucred, td);
900 		return (error ? error : EINVAL);
901 	}
902 	VOP_UNLOCK(nd.ni_vp, 0, td);
903 
904 	if (mdio->md_fwsectors != 0)
905 		sc->fwsectors = mdio->md_fwsectors;
906 	if (mdio->md_fwheads != 0)
907 		sc->fwheads = mdio->md_fwheads;
908 	sc->flags = mdio->md_options & (MD_FORCE | MD_ASYNC);
909 	if (!(flags & FWRITE))
910 		sc->flags |= MD_READONLY;
911 	sc->vnode = nd.ni_vp;
912 
913 	error = mdsetcred(sc, td->td_ucred);
914 	if (error != 0) {
915 		(void)vn_close(nd.ni_vp, flags, td->td_ucred, td);
916 		return (error);
917 	}
918 	return (0);
919 }
920 
921 static void
922 md_zapit(void *p, int cancel)
923 {
924 	if (cancel)
925 		return;
926 	g_wither_geom(p, ENXIO);
927 }
928 
929 static int
930 mddestroy(struct md_s *sc, struct thread *td)
931 {
932 
933 	GIANT_REQUIRED;
934 
935 	mtx_destroy(&sc->queue_mtx);
936 	if (sc->gp) {
937 		sc->gp->softc = NULL;
938 		g_waitfor_event(md_zapit, sc->gp, M_WAITOK, sc->gp, NULL);
939 		sc->gp = NULL;
940 		sc->pp = NULL;
941 	}
942 	sc->flags |= MD_SHUTDOWN;
943 	wakeup(sc);
944 	while (sc->procp != NULL)
945 		tsleep(&sc->procp, PRIBIO, "mddestroy", hz / 10);
946 	if (sc->vnode != NULL)
947 		(void)vn_close(sc->vnode, sc->flags & MD_READONLY ?
948 		    FREAD : (FREAD|FWRITE), sc->cred, td);
949 	if (sc->cred != NULL)
950 		crfree(sc->cred);
951 	if (sc->object != NULL)
952 		vm_object_deallocate(sc->object);
953 	if (sc->indir)
954 		destroy_indir(sc, sc->indir);
955 	if (sc->uma)
956 		uma_zdestroy(sc->uma);
957 
958 	/* XXX: LOCK(unique unit numbers) */
959 	LIST_REMOVE(sc, list);
960 	/* XXX: UNLOCK(unique unit numbers) */
961 	free(sc, M_MD);
962 	return (0);
963 }
964 
965 static int
966 mdcreate_swap(struct md_s *sc, struct md_ioctl *mdio, struct thread *td)
967 {
968 	vm_ooffset_t npage;
969 	int error;
970 
971 	GIANT_REQUIRED;
972 
973 	/*
974 	 * Range check.  Disallow negative sizes or any size less then the
975 	 * size of a page.  Then round to a page.
976 	 */
977 	if (sc->mediasize == 0 || (sc->mediasize % PAGE_SIZE) != 0)
978 		return (EDOM);
979 
980 	/*
981 	 * Allocate an OBJT_SWAP object.
982 	 *
983 	 * Note the truncation.
984 	 */
985 
986 	npage = mdio->md_mediasize / PAGE_SIZE;
987 	if (mdio->md_fwsectors != 0)
988 		sc->fwsectors = mdio->md_fwsectors;
989 	if (mdio->md_fwheads != 0)
990 		sc->fwheads = mdio->md_fwheads;
991 	sc->object = vm_pager_allocate(OBJT_SWAP, NULL, PAGE_SIZE * npage,
992 	    VM_PROT_DEFAULT, 0);
993 	if (sc->object == NULL)
994 		return (ENOMEM);
995 	sc->flags = mdio->md_options & MD_FORCE;
996 	if (mdio->md_options & MD_RESERVE) {
997 		if (swap_pager_reserve(sc->object, 0, npage) < 0) {
998 			vm_object_deallocate(sc->object);
999 			sc->object = NULL;
1000 			return (EDOM);
1001 		}
1002 	}
1003 	error = mdsetcred(sc, td->td_ucred);
1004 	if (error != 0) {
1005 		vm_object_deallocate(sc->object);
1006 		sc->object = NULL;
1007 	}
1008 	return (error);
1009 }
1010 
1011 static int
1012 mddetach(int unit, struct thread *td)
1013 {
1014 	struct md_s *sc;
1015 
1016 	sc = mdfind(unit);
1017 	if (sc == NULL)
1018 		return (ENOENT);
1019 	if (sc->opencount != 0 && !(sc->flags & MD_FORCE))
1020 		return (EBUSY);
1021 	switch(sc->type) {
1022 	case MD_VNODE:
1023 	case MD_SWAP:
1024 	case MD_MALLOC:
1025 	case MD_PRELOAD:
1026 		return (mddestroy(sc, td));
1027 	default:
1028 		return (EOPNOTSUPP);
1029 	}
1030 }
1031 
1032 static int
1033 mdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td)
1034 {
1035 	struct md_ioctl *mdio;
1036 	struct md_s *sc;
1037 	int error, i;
1038 
1039 	if (md_debug)
1040 		printf("mdctlioctl(%s %lx %p %x %p)\n",
1041 			devtoname(dev), cmd, addr, flags, td);
1042 
1043 	/*
1044 	 * We assert the version number in the individual ioctl
1045 	 * handlers instead of out here because (a) it is possible we
1046 	 * may add another ioctl in the future which doesn't read an
1047 	 * mdio, and (b) the correct return value for an unknown ioctl
1048 	 * is ENOIOCTL, not EINVAL.
1049 	 */
1050 	mdio = (struct md_ioctl *)addr;
1051 	switch (cmd) {
1052 	case MDIOCATTACH:
1053 		if (mdio->md_version != MDIOVERSION)
1054 			return (EINVAL);
1055 		switch (mdio->md_type) {
1056 		case MD_MALLOC:
1057 		case MD_PRELOAD:
1058 		case MD_VNODE:
1059 		case MD_SWAP:
1060 			break;
1061 		default:
1062 			return (EINVAL);
1063 		}
1064 		if (mdio->md_options & MD_AUTOUNIT) {
1065 			sc = mdnew(-1);
1066 			mdio->md_unit = sc->unit;
1067 		} else {
1068 			sc = mdnew(mdio->md_unit);
1069 			if (sc == NULL)
1070 				return (EBUSY);
1071 		}
1072 		sc->type = mdio->md_type;
1073 		sc->mediasize = mdio->md_mediasize;
1074 		if (mdio->md_sectorsize == 0)
1075 			sc->sectorsize = DEV_BSIZE;
1076 		else
1077 			sc->sectorsize = mdio->md_sectorsize;
1078 		error = EDOOFUS;
1079 		switch (sc->type) {
1080 		case MD_MALLOC:
1081 			error = mdcreate_malloc(sc, mdio);
1082 			break;
1083 		case MD_PRELOAD:
1084 			error = mdcreate_preload(sc, mdio);
1085 			break;
1086 		case MD_VNODE:
1087 			error = mdcreate_vnode(sc, mdio, td);
1088 			break;
1089 		case MD_SWAP:
1090 			error = mdcreate_swap(sc, mdio, td);
1091 			break;
1092 		}
1093 		if (error != 0) {
1094 			mddestroy(sc, td);
1095 			return (error);
1096 		}
1097 		mdinit(sc);
1098 		return (0);
1099 	case MDIOCDETACH:
1100 		if (mdio->md_version != MDIOVERSION)
1101 			return (EINVAL);
1102 		if (mdio->md_mediasize != 0 || mdio->md_options != 0)
1103 			return (EINVAL);
1104 		return (mddetach(mdio->md_unit, td));
1105 	case MDIOCQUERY:
1106 		if (mdio->md_version != MDIOVERSION)
1107 			return (EINVAL);
1108 		sc = mdfind(mdio->md_unit);
1109 		if (sc == NULL)
1110 			return (ENOENT);
1111 		mdio->md_type = sc->type;
1112 		mdio->md_options = sc->flags;
1113 		mdio->md_mediasize = sc->mediasize;
1114 		mdio->md_sectorsize = sc->sectorsize;
1115 		if (sc->type == MD_VNODE) {
1116 			error = copyout(sc->file, mdio->md_file,
1117 			    strlen(sc->file) + 1);
1118 			if (error != 0)
1119 				return (error);
1120 		}
1121 		return (0);
1122 	case MDIOCLIST:
1123 		if (mdio->md_version != MDIOVERSION)
1124 			return (EINVAL);
1125 		i = 1;
1126 		LIST_FOREACH(sc, &md_softc_list, list) {
1127 			if (i == MDNPAD - 1)
1128 				mdio->md_pad[i] = -1;
1129 			else
1130 				mdio->md_pad[i++] = sc->unit;
1131 		}
1132 		mdio->md_pad[0] = i - 1;
1133 		return (0);
1134 	default:
1135 		return (ENOIOCTL);
1136 	};
1137 	return (ENOIOCTL);
1138 }
1139 
1140 static void
1141 md_preloaded(u_char *image, size_t length)
1142 {
1143 	struct md_s *sc;
1144 
1145 	sc = mdnew(-1);
1146 	if (sc == NULL)
1147 		return;
1148 	sc->type = MD_PRELOAD;
1149 	sc->mediasize = length;
1150 	sc->sectorsize = DEV_BSIZE;
1151 	sc->pl_ptr = image;
1152 	sc->pl_len = length;
1153 #ifdef MD_ROOT
1154 	if (sc->unit == 0)
1155 		rootdevnames[0] = "ufs:/dev/md0";
1156 #endif
1157 	mdinit(sc);
1158 }
1159 
1160 static void
1161 g_md_init(struct g_class *mp __unused)
1162 {
1163 
1164 	caddr_t mod;
1165 	caddr_t c;
1166 	u_char *ptr, *name, *type;
1167 	unsigned len;
1168 
1169 	mod = NULL;
1170 	g_topology_unlock();
1171 #ifdef MD_ROOT_SIZE
1172 	md_preloaded(mfs_root, MD_ROOT_SIZE*1024);
1173 #endif
1174 	while ((mod = preload_search_next_name(mod)) != NULL) {
1175 		name = (char *)preload_search_info(mod, MODINFO_NAME);
1176 		type = (char *)preload_search_info(mod, MODINFO_TYPE);
1177 		if (name == NULL)
1178 			continue;
1179 		if (type == NULL)
1180 			continue;
1181 		if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
1182 			continue;
1183 		c = preload_search_info(mod, MODINFO_ADDR);
1184 		ptr = *(u_char **)c;
1185 		c = preload_search_info(mod, MODINFO_SIZE);
1186 		len = *(size_t *)c;
1187 		printf("%s%d: Preloaded image <%s> %d bytes at %p\n",
1188 		    MD_NAME, mdunits, name, len, ptr);
1189 		md_preloaded(ptr, len);
1190 	}
1191 	status_dev = make_dev(&mdctl_cdevsw, MAXMINOR, UID_ROOT, GID_WHEEL,
1192 	    0600, MDCTL_NAME);
1193 	g_topology_lock();
1194 }
1195 
1196 static void
1197 g_md_fini(struct g_class *mp __unused)
1198 {
1199 
1200 	if (status_dev != NULL)
1201 		destroy_dev(status_dev);
1202 }
1203