xref: /freebsd/sys/dev/md/md.c (revision d48a2d23039536ab9f92f88a72c27b9895796abf)
1 /*-
2  * ----------------------------------------------------------------------------
3  * "THE BEER-WARE LICENSE" (Revision 42):
4  * <phk@FreeBSD.ORG> wrote this file.  As long as you retain this notice you
5  * can do whatever you want with this stuff. If we meet some day, and you think
6  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
7  * ----------------------------------------------------------------------------
8  *
9  * $FreeBSD$
10  *
11  */
12 
13 /*-
14  * The following functions are based in the vn(4) driver: mdstart_swap(),
15  * mdstart_vnode(), mdcreate_swap(), mdcreate_vnode() and mddestroy(),
16  * and as such under the following copyright:
17  *
18  * Copyright (c) 1988 University of Utah.
19  * Copyright (c) 1990, 1993
20  *	The Regents of the University of California.  All rights reserved.
21  * Copyright (c) 2013 The FreeBSD Foundation
22  * All rights reserved.
23  *
24  * This code is derived from software contributed to Berkeley by
25  * the Systems Programming Group of the University of Utah Computer
26  * Science Department.
27  *
28  * Portions of this software were developed by Konstantin Belousov
29  * under sponsorship from the FreeBSD Foundation.
30  *
31  * Redistribution and use in source and binary forms, with or without
32  * modification, are permitted provided that the following conditions
33  * are met:
34  * 1. Redistributions of source code must retain the above copyright
35  *    notice, this list of conditions and the following disclaimer.
36  * 2. Redistributions in binary form must reproduce the above copyright
37  *    notice, this list of conditions and the following disclaimer in the
38  *    documentation and/or other materials provided with the distribution.
39  * 4. Neither the name of the University nor the names of its contributors
40  *    may be used to endorse or promote products derived from this software
41  *    without specific prior written permission.
42  *
43  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
44  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
45  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
46  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
47  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
48  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
49  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
50  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
51  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
52  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
53  * SUCH DAMAGE.
54  *
55  * from: Utah Hdr: vn.c 1.13 94/04/02
56  *
57  *	from: @(#)vn.c	8.6 (Berkeley) 4/1/94
58  * From: src/sys/dev/vn/vn.c,v 1.122 2000/12/16 16:06:03
59  */
60 
61 #include "opt_geom.h"
62 #include "opt_md.h"
63 
64 #include <sys/param.h>
65 #include <sys/systm.h>
66 #include <sys/bio.h>
67 #include <sys/buf.h>
68 #include <sys/conf.h>
69 #include <sys/devicestat.h>
70 #include <sys/fcntl.h>
71 #include <sys/kernel.h>
72 #include <sys/kthread.h>
73 #include <sys/limits.h>
74 #include <sys/linker.h>
75 #include <sys/lock.h>
76 #include <sys/malloc.h>
77 #include <sys/mdioctl.h>
78 #include <sys/mount.h>
79 #include <sys/mutex.h>
80 #include <sys/sx.h>
81 #include <sys/namei.h>
82 #include <sys/proc.h>
83 #include <sys/queue.h>
84 #include <sys/rwlock.h>
85 #include <sys/sbuf.h>
86 #include <sys/sched.h>
87 #include <sys/sf_buf.h>
88 #include <sys/sysctl.h>
89 #include <sys/vnode.h>
90 
91 #include <geom/geom.h>
92 #include <geom/geom_int.h>
93 
94 #include <vm/vm.h>
95 #include <vm/vm_param.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pager.h>
99 #include <vm/swap_pager.h>
100 #include <vm/uma.h>
101 
102 #include <machine/bus.h>
103 
104 #define MD_MODVER 1
105 
106 #define MD_SHUTDOWN	0x10000		/* Tell worker thread to terminate. */
107 #define	MD_EXITING	0x20000		/* Worker thread is exiting. */
108 
109 #ifndef MD_NSECT
110 #define MD_NSECT (10000 * 2)
111 #endif
112 
113 static MALLOC_DEFINE(M_MD, "md_disk", "Memory Disk");
114 static MALLOC_DEFINE(M_MDSECT, "md_sectors", "Memory Disk Sectors");
115 
116 static int md_debug;
117 SYSCTL_INT(_debug, OID_AUTO, mddebug, CTLFLAG_RW, &md_debug, 0,
118     "Enable md(4) debug messages");
119 static int md_malloc_wait;
120 SYSCTL_INT(_vm, OID_AUTO, md_malloc_wait, CTLFLAG_RW, &md_malloc_wait, 0,
121     "Allow malloc to wait for memory allocations");
122 
123 #if defined(MD_ROOT) && !defined(MD_ROOT_FSTYPE)
124 #define	MD_ROOT_FSTYPE	"ufs"
125 #endif
126 
127 #if defined(MD_ROOT)
128 /*
129  * Preloaded image gets put here.
130  */
131 #if defined(MD_ROOT_SIZE)
132 /*
133  * We put the mfs_root symbol into the oldmfs section of the kernel object file.
134  * Applications that patch the object with the image can determine
135  * the size looking at the oldmfs section size within the kernel.
136  */
137 u_char mfs_root[MD_ROOT_SIZE*1024] __attribute__ ((section ("oldmfs")));
138 const int mfs_root_size = sizeof(mfs_root);
139 #else
140 extern volatile u_char __weak_symbol mfs_root;
141 extern volatile u_char __weak_symbol mfs_root_end;
142 __GLOBL(mfs_root);
143 __GLOBL(mfs_root_end);
144 #define mfs_root_size ((uintptr_t)(&mfs_root_end - &mfs_root))
145 #endif
146 #endif
147 
148 static g_init_t g_md_init;
149 static g_fini_t g_md_fini;
150 static g_start_t g_md_start;
151 static g_access_t g_md_access;
152 static void g_md_dumpconf(struct sbuf *sb, const char *indent,
153     struct g_geom *gp, struct g_consumer *cp __unused, struct g_provider *pp);
154 
155 static struct cdev *status_dev = 0;
156 static struct sx md_sx;
157 static struct unrhdr *md_uh;
158 
159 static d_ioctl_t mdctlioctl;
160 
161 static struct cdevsw mdctl_cdevsw = {
162 	.d_version =	D_VERSION,
163 	.d_ioctl =	mdctlioctl,
164 	.d_name =	MD_NAME,
165 };
166 
167 struct g_class g_md_class = {
168 	.name = "MD",
169 	.version = G_VERSION,
170 	.init = g_md_init,
171 	.fini = g_md_fini,
172 	.start = g_md_start,
173 	.access = g_md_access,
174 	.dumpconf = g_md_dumpconf,
175 };
176 
177 DECLARE_GEOM_CLASS(g_md_class, g_md);
178 
179 
180 static LIST_HEAD(, md_s) md_softc_list = LIST_HEAD_INITIALIZER(md_softc_list);
181 
182 #define NINDIR	(PAGE_SIZE / sizeof(uintptr_t))
183 #define NMASK	(NINDIR-1)
184 static int nshift;
185 
186 static int md_vnode_pbuf_freecnt;
187 
188 struct indir {
189 	uintptr_t	*array;
190 	u_int		total;
191 	u_int		used;
192 	u_int		shift;
193 };
194 
195 struct md_s {
196 	int unit;
197 	LIST_ENTRY(md_s) list;
198 	struct bio_queue_head bio_queue;
199 	struct mtx queue_mtx;
200 	struct mtx stat_mtx;
201 	struct cdev *dev;
202 	enum md_types type;
203 	off_t mediasize;
204 	unsigned sectorsize;
205 	unsigned opencount;
206 	unsigned fwheads;
207 	unsigned fwsectors;
208 	unsigned flags;
209 	char name[20];
210 	struct proc *procp;
211 	struct g_geom *gp;
212 	struct g_provider *pp;
213 	int (*start)(struct md_s *sc, struct bio *bp);
214 	struct devstat *devstat;
215 
216 	/* MD_MALLOC related fields */
217 	struct indir *indir;
218 	uma_zone_t uma;
219 
220 	/* MD_PRELOAD related fields */
221 	u_char *pl_ptr;
222 	size_t pl_len;
223 
224 	/* MD_VNODE related fields */
225 	struct vnode *vnode;
226 	char file[PATH_MAX];
227 	struct ucred *cred;
228 
229 	/* MD_SWAP related fields */
230 	vm_object_t object;
231 };
232 
233 static struct indir *
234 new_indir(u_int shift)
235 {
236 	struct indir *ip;
237 
238 	ip = malloc(sizeof *ip, M_MD, (md_malloc_wait ? M_WAITOK : M_NOWAIT)
239 	    | M_ZERO);
240 	if (ip == NULL)
241 		return (NULL);
242 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
243 	    M_MDSECT, (md_malloc_wait ? M_WAITOK : M_NOWAIT) | M_ZERO);
244 	if (ip->array == NULL) {
245 		free(ip, M_MD);
246 		return (NULL);
247 	}
248 	ip->total = NINDIR;
249 	ip->shift = shift;
250 	return (ip);
251 }
252 
253 static void
254 del_indir(struct indir *ip)
255 {
256 
257 	free(ip->array, M_MDSECT);
258 	free(ip, M_MD);
259 }
260 
261 static void
262 destroy_indir(struct md_s *sc, struct indir *ip)
263 {
264 	int i;
265 
266 	for (i = 0; i < NINDIR; i++) {
267 		if (!ip->array[i])
268 			continue;
269 		if (ip->shift)
270 			destroy_indir(sc, (struct indir*)(ip->array[i]));
271 		else if (ip->array[i] > 255)
272 			uma_zfree(sc->uma, (void *)(ip->array[i]));
273 	}
274 	del_indir(ip);
275 }
276 
277 /*
278  * This function does the math and allocates the top level "indir" structure
279  * for a device of "size" sectors.
280  */
281 
282 static struct indir *
283 dimension(off_t size)
284 {
285 	off_t rcnt;
286 	struct indir *ip;
287 	int layer;
288 
289 	rcnt = size;
290 	layer = 0;
291 	while (rcnt > NINDIR) {
292 		rcnt /= NINDIR;
293 		layer++;
294 	}
295 
296 	/*
297 	 * XXX: the top layer is probably not fully populated, so we allocate
298 	 * too much space for ip->array in here.
299 	 */
300 	ip = malloc(sizeof *ip, M_MD, M_WAITOK | M_ZERO);
301 	ip->array = malloc(sizeof(uintptr_t) * NINDIR,
302 	    M_MDSECT, M_WAITOK | M_ZERO);
303 	ip->total = NINDIR;
304 	ip->shift = layer * nshift;
305 	return (ip);
306 }
307 
308 /*
309  * Read a given sector
310  */
311 
312 static uintptr_t
313 s_read(struct indir *ip, off_t offset)
314 {
315 	struct indir *cip;
316 	int idx;
317 	uintptr_t up;
318 
319 	if (md_debug > 1)
320 		printf("s_read(%jd)\n", (intmax_t)offset);
321 	up = 0;
322 	for (cip = ip; cip != NULL;) {
323 		if (cip->shift) {
324 			idx = (offset >> cip->shift) & NMASK;
325 			up = cip->array[idx];
326 			cip = (struct indir *)up;
327 			continue;
328 		}
329 		idx = offset & NMASK;
330 		return (cip->array[idx]);
331 	}
332 	return (0);
333 }
334 
335 /*
336  * Write a given sector, prune the tree if the value is 0
337  */
338 
339 static int
340 s_write(struct indir *ip, off_t offset, uintptr_t ptr)
341 {
342 	struct indir *cip, *lip[10];
343 	int idx, li;
344 	uintptr_t up;
345 
346 	if (md_debug > 1)
347 		printf("s_write(%jd, %p)\n", (intmax_t)offset, (void *)ptr);
348 	up = 0;
349 	li = 0;
350 	cip = ip;
351 	for (;;) {
352 		lip[li++] = cip;
353 		if (cip->shift) {
354 			idx = (offset >> cip->shift) & NMASK;
355 			up = cip->array[idx];
356 			if (up != 0) {
357 				cip = (struct indir *)up;
358 				continue;
359 			}
360 			/* Allocate branch */
361 			cip->array[idx] =
362 			    (uintptr_t)new_indir(cip->shift - nshift);
363 			if (cip->array[idx] == 0)
364 				return (ENOSPC);
365 			cip->used++;
366 			up = cip->array[idx];
367 			cip = (struct indir *)up;
368 			continue;
369 		}
370 		/* leafnode */
371 		idx = offset & NMASK;
372 		up = cip->array[idx];
373 		if (up != 0)
374 			cip->used--;
375 		cip->array[idx] = ptr;
376 		if (ptr != 0)
377 			cip->used++;
378 		break;
379 	}
380 	if (cip->used != 0 || li == 1)
381 		return (0);
382 	li--;
383 	while (cip->used == 0 && cip != ip) {
384 		li--;
385 		idx = (offset >> lip[li]->shift) & NMASK;
386 		up = lip[li]->array[idx];
387 		KASSERT(up == (uintptr_t)cip, ("md screwed up"));
388 		del_indir(cip);
389 		lip[li]->array[idx] = 0;
390 		lip[li]->used--;
391 		cip = lip[li];
392 	}
393 	return (0);
394 }
395 
396 
397 static int
398 g_md_access(struct g_provider *pp, int r, int w, int e)
399 {
400 	struct md_s *sc;
401 
402 	sc = pp->geom->softc;
403 	if (sc == NULL) {
404 		if (r <= 0 && w <= 0 && e <= 0)
405 			return (0);
406 		return (ENXIO);
407 	}
408 	r += pp->acr;
409 	w += pp->acw;
410 	e += pp->ace;
411 	if ((sc->flags & MD_READONLY) != 0 && w > 0)
412 		return (EROFS);
413 	if ((pp->acr + pp->acw + pp->ace) == 0 && (r + w + e) > 0) {
414 		sc->opencount = 1;
415 	} else if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0) {
416 		sc->opencount = 0;
417 	}
418 	return (0);
419 }
420 
421 static void
422 g_md_start(struct bio *bp)
423 {
424 	struct md_s *sc;
425 
426 	sc = bp->bio_to->geom->softc;
427 	if ((bp->bio_cmd == BIO_READ) || (bp->bio_cmd == BIO_WRITE)) {
428 		mtx_lock(&sc->stat_mtx);
429 		devstat_start_transaction_bio(sc->devstat, bp);
430 		mtx_unlock(&sc->stat_mtx);
431 	}
432 	mtx_lock(&sc->queue_mtx);
433 	bioq_disksort(&sc->bio_queue, bp);
434 	mtx_unlock(&sc->queue_mtx);
435 	wakeup(sc);
436 }
437 
438 #define	MD_MALLOC_MOVE_ZERO	1
439 #define	MD_MALLOC_MOVE_FILL	2
440 #define	MD_MALLOC_MOVE_READ	3
441 #define	MD_MALLOC_MOVE_WRITE	4
442 #define	MD_MALLOC_MOVE_CMP	5
443 
444 static int
445 md_malloc_move_ma(vm_page_t **mp, int *ma_offs, unsigned sectorsize,
446     void *ptr, u_char fill, int op)
447 {
448 	struct sf_buf *sf;
449 	vm_page_t m, *mp1;
450 	char *p, first;
451 	off_t *uc;
452 	unsigned n;
453 	int error, i, ma_offs1, sz, first_read;
454 
455 	m = NULL;
456 	error = 0;
457 	sf = NULL;
458 	/* if (op == MD_MALLOC_MOVE_CMP) { gcc */
459 		first = 0;
460 		first_read = 0;
461 		uc = ptr;
462 		mp1 = *mp;
463 		ma_offs1 = *ma_offs;
464 	/* } */
465 	sched_pin();
466 	for (n = sectorsize; n != 0; n -= sz) {
467 		sz = imin(PAGE_SIZE - *ma_offs, n);
468 		if (m != **mp) {
469 			if (sf != NULL)
470 				sf_buf_free(sf);
471 			m = **mp;
472 			sf = sf_buf_alloc(m, SFB_CPUPRIVATE |
473 			    (md_malloc_wait ? 0 : SFB_NOWAIT));
474 			if (sf == NULL) {
475 				error = ENOMEM;
476 				break;
477 			}
478 		}
479 		p = (char *)sf_buf_kva(sf) + *ma_offs;
480 		switch (op) {
481 		case MD_MALLOC_MOVE_ZERO:
482 			bzero(p, sz);
483 			break;
484 		case MD_MALLOC_MOVE_FILL:
485 			memset(p, fill, sz);
486 			break;
487 		case MD_MALLOC_MOVE_READ:
488 			bcopy(ptr, p, sz);
489 			cpu_flush_dcache(p, sz);
490 			break;
491 		case MD_MALLOC_MOVE_WRITE:
492 			bcopy(p, ptr, sz);
493 			break;
494 		case MD_MALLOC_MOVE_CMP:
495 			for (i = 0; i < sz; i++, p++) {
496 				if (!first_read) {
497 					*uc = (u_char)*p;
498 					first = *p;
499 					first_read = 1;
500 				} else if (*p != first) {
501 					error = EDOOFUS;
502 					break;
503 				}
504 			}
505 			break;
506 		default:
507 			KASSERT(0, ("md_malloc_move_ma unknown op %d\n", op));
508 			break;
509 		}
510 		if (error != 0)
511 			break;
512 		*ma_offs += sz;
513 		*ma_offs %= PAGE_SIZE;
514 		if (*ma_offs == 0)
515 			(*mp)++;
516 		ptr = (char *)ptr + sz;
517 	}
518 
519 	if (sf != NULL)
520 		sf_buf_free(sf);
521 	sched_unpin();
522 	if (op == MD_MALLOC_MOVE_CMP && error != 0) {
523 		*mp = mp1;
524 		*ma_offs = ma_offs1;
525 	}
526 	return (error);
527 }
528 
529 static int
530 md_malloc_move_vlist(bus_dma_segment_t **pvlist, int *pma_offs,
531     unsigned len, void *ptr, u_char fill, int op)
532 {
533 	bus_dma_segment_t *vlist;
534 	uint8_t *p, *end, first;
535 	off_t *uc;
536 	int ma_offs, seg_len;
537 
538 	vlist = *pvlist;
539 	ma_offs = *pma_offs;
540 	uc = ptr;
541 
542 	for (; len != 0; len -= seg_len) {
543 		seg_len = imin(vlist->ds_len - ma_offs, len);
544 		p = (uint8_t *)(uintptr_t)vlist->ds_addr + ma_offs;
545 		switch (op) {
546 		case MD_MALLOC_MOVE_ZERO:
547 			bzero(p, seg_len);
548 			break;
549 		case MD_MALLOC_MOVE_FILL:
550 			memset(p, fill, seg_len);
551 			break;
552 		case MD_MALLOC_MOVE_READ:
553 			bcopy(ptr, p, seg_len);
554 			cpu_flush_dcache(p, seg_len);
555 			break;
556 		case MD_MALLOC_MOVE_WRITE:
557 			bcopy(p, ptr, seg_len);
558 			break;
559 		case MD_MALLOC_MOVE_CMP:
560 			end = p + seg_len;
561 			first = *uc = *p;
562 			/* Confirm all following bytes match the first */
563 			while (++p < end) {
564 				if (*p != first)
565 					return (EDOOFUS);
566 			}
567 			break;
568 		default:
569 			KASSERT(0, ("md_malloc_move_vlist unknown op %d\n", op));
570 			break;
571 		}
572 
573 		ma_offs += seg_len;
574 		if (ma_offs == vlist->ds_len) {
575 			ma_offs = 0;
576 			vlist++;
577 		}
578 		ptr = (uint8_t *)ptr + seg_len;
579 	}
580 	*pvlist = vlist;
581 	*pma_offs = ma_offs;
582 
583 	return (0);
584 }
585 
586 static int
587 mdstart_malloc(struct md_s *sc, struct bio *bp)
588 {
589 	u_char *dst;
590 	vm_page_t *m;
591 	bus_dma_segment_t *vlist;
592 	int i, error, error1, ma_offs, notmapped;
593 	off_t secno, nsec, uc;
594 	uintptr_t sp, osp;
595 
596 	switch (bp->bio_cmd) {
597 	case BIO_READ:
598 	case BIO_WRITE:
599 	case BIO_DELETE:
600 		break;
601 	default:
602 		return (EOPNOTSUPP);
603 	}
604 
605 	notmapped = (bp->bio_flags & BIO_UNMAPPED) != 0;
606 	vlist = (bp->bio_flags & BIO_VLIST) != 0 ?
607 	    (bus_dma_segment_t *)bp->bio_data : NULL;
608 	if (notmapped) {
609 		m = bp->bio_ma;
610 		ma_offs = bp->bio_ma_offset;
611 		dst = NULL;
612 		KASSERT(vlist == NULL, ("vlists cannot be unmapped"));
613 	} else if (vlist != NULL) {
614 		ma_offs = bp->bio_ma_offset;
615 		dst = NULL;
616 	} else {
617 		dst = bp->bio_data;
618 	}
619 
620 	nsec = bp->bio_length / sc->sectorsize;
621 	secno = bp->bio_offset / sc->sectorsize;
622 	error = 0;
623 	while (nsec--) {
624 		osp = s_read(sc->indir, secno);
625 		if (bp->bio_cmd == BIO_DELETE) {
626 			if (osp != 0)
627 				error = s_write(sc->indir, secno, 0);
628 		} else if (bp->bio_cmd == BIO_READ) {
629 			if (osp == 0) {
630 				if (notmapped) {
631 					error = md_malloc_move_ma(&m, &ma_offs,
632 					    sc->sectorsize, NULL, 0,
633 					    MD_MALLOC_MOVE_ZERO);
634 				} else if (vlist != NULL) {
635 					error = md_malloc_move_vlist(&vlist,
636 					    &ma_offs, sc->sectorsize, NULL, 0,
637 					    MD_MALLOC_MOVE_ZERO);
638 				} else
639 					bzero(dst, sc->sectorsize);
640 			} else if (osp <= 255) {
641 				if (notmapped) {
642 					error = md_malloc_move_ma(&m, &ma_offs,
643 					    sc->sectorsize, NULL, osp,
644 					    MD_MALLOC_MOVE_FILL);
645 				} else if (vlist != NULL) {
646 					error = md_malloc_move_vlist(&vlist,
647 					    &ma_offs, sc->sectorsize, NULL, osp,
648 					    MD_MALLOC_MOVE_FILL);
649 				} else
650 					memset(dst, osp, sc->sectorsize);
651 			} else {
652 				if (notmapped) {
653 					error = md_malloc_move_ma(&m, &ma_offs,
654 					    sc->sectorsize, (void *)osp, 0,
655 					    MD_MALLOC_MOVE_READ);
656 				} else if (vlist != NULL) {
657 					error = md_malloc_move_vlist(&vlist,
658 					    &ma_offs, sc->sectorsize,
659 					    (void *)osp, 0,
660 					    MD_MALLOC_MOVE_READ);
661 				} else {
662 					bcopy((void *)osp, dst, sc->sectorsize);
663 					cpu_flush_dcache(dst, sc->sectorsize);
664 				}
665 			}
666 			osp = 0;
667 		} else if (bp->bio_cmd == BIO_WRITE) {
668 			if (sc->flags & MD_COMPRESS) {
669 				if (notmapped) {
670 					error1 = md_malloc_move_ma(&m, &ma_offs,
671 					    sc->sectorsize, &uc, 0,
672 					    MD_MALLOC_MOVE_CMP);
673 					i = error1 == 0 ? sc->sectorsize : 0;
674 				} else if (vlist != NULL) {
675 					error1 = md_malloc_move_vlist(&vlist,
676 					    &ma_offs, sc->sectorsize, &uc, 0,
677 					    MD_MALLOC_MOVE_CMP);
678 					i = error1 == 0 ? sc->sectorsize : 0;
679 				} else {
680 					uc = dst[0];
681 					for (i = 1; i < sc->sectorsize; i++) {
682 						if (dst[i] != uc)
683 							break;
684 					}
685 				}
686 			} else {
687 				i = 0;
688 				uc = 0;
689 			}
690 			if (i == sc->sectorsize) {
691 				if (osp != uc)
692 					error = s_write(sc->indir, secno, uc);
693 			} else {
694 				if (osp <= 255) {
695 					sp = (uintptr_t)uma_zalloc(sc->uma,
696 					    md_malloc_wait ? M_WAITOK :
697 					    M_NOWAIT);
698 					if (sp == 0) {
699 						error = ENOSPC;
700 						break;
701 					}
702 					if (notmapped) {
703 						error = md_malloc_move_ma(&m,
704 						    &ma_offs, sc->sectorsize,
705 						    (void *)sp, 0,
706 						    MD_MALLOC_MOVE_WRITE);
707 					} else if (vlist != NULL) {
708 						error = md_malloc_move_vlist(
709 						    &vlist, &ma_offs,
710 						    sc->sectorsize, (void *)sp,
711 						    0, MD_MALLOC_MOVE_WRITE);
712 					} else {
713 						bcopy(dst, (void *)sp,
714 						    sc->sectorsize);
715 					}
716 					error = s_write(sc->indir, secno, sp);
717 				} else {
718 					if (notmapped) {
719 						error = md_malloc_move_ma(&m,
720 						    &ma_offs, sc->sectorsize,
721 						    (void *)osp, 0,
722 						    MD_MALLOC_MOVE_WRITE);
723 					} else if (vlist != NULL) {
724 						error = md_malloc_move_vlist(
725 						    &vlist, &ma_offs,
726 						    sc->sectorsize, (void *)osp,
727 						    0, MD_MALLOC_MOVE_WRITE);
728 					} else {
729 						bcopy(dst, (void *)osp,
730 						    sc->sectorsize);
731 					}
732 					osp = 0;
733 				}
734 			}
735 		} else {
736 			error = EOPNOTSUPP;
737 		}
738 		if (osp > 255)
739 			uma_zfree(sc->uma, (void*)osp);
740 		if (error != 0)
741 			break;
742 		secno++;
743 		if (!notmapped && vlist == NULL)
744 			dst += sc->sectorsize;
745 	}
746 	bp->bio_resid = 0;
747 	return (error);
748 }
749 
750 static void
751 mdcopyto_vlist(void *src, bus_dma_segment_t *vlist, off_t offset, off_t len)
752 {
753 	off_t seg_len;
754 
755 	while (offset >= vlist->ds_len) {
756 		offset -= vlist->ds_len;
757 		vlist++;
758 	}
759 
760 	while (len != 0) {
761 		seg_len = omin(len, vlist->ds_len - offset);
762 		bcopy(src, (void *)(uintptr_t)(vlist->ds_addr + offset),
763 		    seg_len);
764 		offset = 0;
765 		src = (uint8_t *)src + seg_len;
766 		len -= seg_len;
767 		vlist++;
768 	}
769 }
770 
771 static void
772 mdcopyfrom_vlist(bus_dma_segment_t *vlist, off_t offset, void *dst, off_t len)
773 {
774 	off_t seg_len;
775 
776 	while (offset >= vlist->ds_len) {
777 		offset -= vlist->ds_len;
778 		vlist++;
779 	}
780 
781 	while (len != 0) {
782 		seg_len = omin(len, vlist->ds_len - offset);
783 		bcopy((void *)(uintptr_t)(vlist->ds_addr + offset), dst,
784 		    seg_len);
785 		offset = 0;
786 		dst = (uint8_t *)dst + seg_len;
787 		len -= seg_len;
788 		vlist++;
789 	}
790 }
791 
792 static int
793 mdstart_preload(struct md_s *sc, struct bio *bp)
794 {
795 	uint8_t *p;
796 
797 	p = sc->pl_ptr + bp->bio_offset;
798 	switch (bp->bio_cmd) {
799 	case BIO_READ:
800 		if ((bp->bio_flags & BIO_VLIST) != 0) {
801 			mdcopyto_vlist(p, (bus_dma_segment_t *)bp->bio_data,
802 			    bp->bio_ma_offset, bp->bio_length);
803 		} else {
804 			bcopy(p, bp->bio_data, bp->bio_length);
805 		}
806 		cpu_flush_dcache(bp->bio_data, bp->bio_length);
807 		break;
808 	case BIO_WRITE:
809 		if ((bp->bio_flags & BIO_VLIST) != 0) {
810 			mdcopyfrom_vlist((bus_dma_segment_t *)bp->bio_data,
811 			    bp->bio_ma_offset, p, bp->bio_length);
812 		} else {
813 			bcopy(bp->bio_data, p, bp->bio_length);
814 		}
815 		break;
816 	}
817 	bp->bio_resid = 0;
818 	return (0);
819 }
820 
821 static int
822 mdstart_vnode(struct md_s *sc, struct bio *bp)
823 {
824 	int error;
825 	struct uio auio;
826 	struct iovec aiov;
827 	struct iovec *piov;
828 	struct mount *mp;
829 	struct vnode *vp;
830 	struct buf *pb;
831 	bus_dma_segment_t *vlist;
832 	struct thread *td;
833 	off_t iolen, len, zerosize;
834 	int ma_offs, npages;
835 
836 	switch (bp->bio_cmd) {
837 	case BIO_READ:
838 		auio.uio_rw = UIO_READ;
839 		break;
840 	case BIO_WRITE:
841 	case BIO_DELETE:
842 		auio.uio_rw = UIO_WRITE;
843 		break;
844 	case BIO_FLUSH:
845 		break;
846 	default:
847 		return (EOPNOTSUPP);
848 	}
849 
850 	td = curthread;
851 	vp = sc->vnode;
852 	pb = NULL;
853 	piov = NULL;
854 	ma_offs = bp->bio_ma_offset;
855 	len = bp->bio_length;
856 
857 	/*
858 	 * VNODE I/O
859 	 *
860 	 * If an error occurs, we set BIO_ERROR but we do not set
861 	 * B_INVAL because (for a write anyway), the buffer is
862 	 * still valid.
863 	 */
864 
865 	if (bp->bio_cmd == BIO_FLUSH) {
866 		(void) vn_start_write(vp, &mp, V_WAIT);
867 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
868 		error = VOP_FSYNC(vp, MNT_WAIT, td);
869 		VOP_UNLOCK(vp, 0);
870 		vn_finished_write(mp);
871 		return (error);
872 	}
873 
874 	auio.uio_offset = (vm_ooffset_t)bp->bio_offset;
875 	auio.uio_resid = bp->bio_length;
876 	auio.uio_segflg = UIO_SYSSPACE;
877 	auio.uio_td = td;
878 
879 	if (bp->bio_cmd == BIO_DELETE) {
880 		/*
881 		 * Emulate BIO_DELETE by writing zeros.
882 		 */
883 		zerosize = ZERO_REGION_SIZE -
884 		    (ZERO_REGION_SIZE % sc->sectorsize);
885 		auio.uio_iovcnt = howmany(bp->bio_length, zerosize);
886 		piov = malloc(sizeof(*piov) * auio.uio_iovcnt, M_MD, M_WAITOK);
887 		auio.uio_iov = piov;
888 		while (len > 0) {
889 			piov->iov_base = __DECONST(void *, zero_region);
890 			piov->iov_len = len;
891 			if (len > zerosize)
892 				piov->iov_len = zerosize;
893 			len -= piov->iov_len;
894 			piov++;
895 		}
896 		piov = auio.uio_iov;
897 	} else if ((bp->bio_flags & BIO_VLIST) != 0) {
898 		piov = malloc(sizeof(*piov) * bp->bio_ma_n, M_MD, M_WAITOK);
899 		auio.uio_iov = piov;
900 		vlist = (bus_dma_segment_t *)bp->bio_data;
901 		while (len > 0) {
902 			piov->iov_base = (void *)(uintptr_t)(vlist->ds_addr +
903 			    ma_offs);
904 			piov->iov_len = vlist->ds_len - ma_offs;
905 			if (piov->iov_len > len)
906 				piov->iov_len = len;
907 			len -= piov->iov_len;
908 			ma_offs = 0;
909 			vlist++;
910 			piov++;
911 		}
912 		auio.uio_iovcnt = piov - auio.uio_iov;
913 		piov = auio.uio_iov;
914 	} else if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
915 		pb = getpbuf(&md_vnode_pbuf_freecnt);
916 		bp->bio_resid = len;
917 unmapped_step:
918 		npages = atop(min(MAXPHYS, round_page(len + (ma_offs &
919 		    PAGE_MASK))));
920 		iolen = min(ptoa(npages) - (ma_offs & PAGE_MASK), len);
921 		KASSERT(iolen > 0, ("zero iolen"));
922 		pmap_qenter((vm_offset_t)pb->b_data,
923 		    &bp->bio_ma[atop(ma_offs)], npages);
924 		aiov.iov_base = (void *)((vm_offset_t)pb->b_data +
925 		    (ma_offs & PAGE_MASK));
926 		aiov.iov_len = iolen;
927 		auio.uio_iov = &aiov;
928 		auio.uio_iovcnt = 1;
929 		auio.uio_resid = iolen;
930 	} else {
931 		aiov.iov_base = bp->bio_data;
932 		aiov.iov_len = bp->bio_length;
933 		auio.uio_iov = &aiov;
934 		auio.uio_iovcnt = 1;
935 	}
936 	/*
937 	 * When reading set IO_DIRECT to try to avoid double-caching
938 	 * the data.  When writing IO_DIRECT is not optimal.
939 	 */
940 	if (auio.uio_rw == UIO_READ) {
941 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
942 		error = VOP_READ(vp, &auio, IO_DIRECT, sc->cred);
943 		VOP_UNLOCK(vp, 0);
944 	} else {
945 		(void) vn_start_write(vp, &mp, V_WAIT);
946 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
947 		error = VOP_WRITE(vp, &auio, sc->flags & MD_ASYNC ? 0 : IO_SYNC,
948 		    sc->cred);
949 		VOP_UNLOCK(vp, 0);
950 		vn_finished_write(mp);
951 	}
952 
953 	if (pb != NULL) {
954 		pmap_qremove((vm_offset_t)pb->b_data, npages);
955 		if (error == 0) {
956 			len -= iolen;
957 			bp->bio_resid -= iolen;
958 			ma_offs += iolen;
959 			if (len > 0)
960 				goto unmapped_step;
961 		}
962 		relpbuf(pb, &md_vnode_pbuf_freecnt);
963 	}
964 
965 	free(piov, M_MD);
966 	if (pb == NULL)
967 		bp->bio_resid = auio.uio_resid;
968 	return (error);
969 }
970 
971 static int
972 mdstart_swap(struct md_s *sc, struct bio *bp)
973 {
974 	vm_page_t m;
975 	u_char *p;
976 	vm_pindex_t i, lastp;
977 	bus_dma_segment_t *vlist;
978 	int rv, ma_offs, offs, len, lastend;
979 
980 	switch (bp->bio_cmd) {
981 	case BIO_READ:
982 	case BIO_WRITE:
983 	case BIO_DELETE:
984 		break;
985 	default:
986 		return (EOPNOTSUPP);
987 	}
988 
989 	p = bp->bio_data;
990 	ma_offs = (bp->bio_flags & (BIO_UNMAPPED|BIO_VLIST)) != 0 ?
991 	    bp->bio_ma_offset : 0;
992 	vlist = (bp->bio_flags & BIO_VLIST) != 0 ?
993 	    (bus_dma_segment_t *)bp->bio_data : NULL;
994 
995 	/*
996 	 * offs is the offset at which to start operating on the
997 	 * next (ie, first) page.  lastp is the last page on
998 	 * which we're going to operate.  lastend is the ending
999 	 * position within that last page (ie, PAGE_SIZE if
1000 	 * we're operating on complete aligned pages).
1001 	 */
1002 	offs = bp->bio_offset % PAGE_SIZE;
1003 	lastp = (bp->bio_offset + bp->bio_length - 1) / PAGE_SIZE;
1004 	lastend = (bp->bio_offset + bp->bio_length - 1) % PAGE_SIZE + 1;
1005 
1006 	rv = VM_PAGER_OK;
1007 	VM_OBJECT_WLOCK(sc->object);
1008 	vm_object_pip_add(sc->object, 1);
1009 	for (i = bp->bio_offset / PAGE_SIZE; i <= lastp; i++) {
1010 		len = ((i == lastp) ? lastend : PAGE_SIZE) - offs;
1011 		m = vm_page_grab(sc->object, i, VM_ALLOC_SYSTEM);
1012 		if (bp->bio_cmd == BIO_READ) {
1013 			if (m->valid == VM_PAGE_BITS_ALL)
1014 				rv = VM_PAGER_OK;
1015 			else
1016 				rv = vm_pager_get_pages(sc->object, &m, 1,
1017 				    NULL, NULL);
1018 			if (rv == VM_PAGER_ERROR) {
1019 				vm_page_xunbusy(m);
1020 				break;
1021 			} else if (rv == VM_PAGER_FAIL) {
1022 				/*
1023 				 * Pager does not have the page.  Zero
1024 				 * the allocated page, and mark it as
1025 				 * valid. Do not set dirty, the page
1026 				 * can be recreated if thrown out.
1027 				 */
1028 				pmap_zero_page(m);
1029 				m->valid = VM_PAGE_BITS_ALL;
1030 			}
1031 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
1032 				pmap_copy_pages(&m, offs, bp->bio_ma,
1033 				    ma_offs, len);
1034 			} else if ((bp->bio_flags & BIO_VLIST) != 0) {
1035 				physcopyout_vlist(VM_PAGE_TO_PHYS(m) + offs,
1036 				    vlist, ma_offs, len);
1037 				cpu_flush_dcache(p, len);
1038 			} else {
1039 				physcopyout(VM_PAGE_TO_PHYS(m) + offs, p, len);
1040 				cpu_flush_dcache(p, len);
1041 			}
1042 		} else if (bp->bio_cmd == BIO_WRITE) {
1043 			if (len != PAGE_SIZE && m->valid != VM_PAGE_BITS_ALL)
1044 				rv = vm_pager_get_pages(sc->object, &m, 1,
1045 				    NULL, NULL);
1046 			else
1047 				rv = VM_PAGER_OK;
1048 			if (rv == VM_PAGER_ERROR) {
1049 				vm_page_xunbusy(m);
1050 				break;
1051 			}
1052 			if ((bp->bio_flags & BIO_UNMAPPED) != 0) {
1053 				pmap_copy_pages(bp->bio_ma, ma_offs, &m,
1054 				    offs, len);
1055 			} else if ((bp->bio_flags & BIO_VLIST) != 0) {
1056 				physcopyin_vlist(vlist, ma_offs,
1057 				    VM_PAGE_TO_PHYS(m) + offs, len);
1058 			} else {
1059 				physcopyin(p, VM_PAGE_TO_PHYS(m) + offs, len);
1060 			}
1061 			m->valid = VM_PAGE_BITS_ALL;
1062 		} else if (bp->bio_cmd == BIO_DELETE) {
1063 			if (len != PAGE_SIZE && m->valid != VM_PAGE_BITS_ALL)
1064 				rv = vm_pager_get_pages(sc->object, &m, 1,
1065 				    NULL, NULL);
1066 			else
1067 				rv = VM_PAGER_OK;
1068 			if (rv == VM_PAGER_ERROR) {
1069 				vm_page_xunbusy(m);
1070 				break;
1071 			}
1072 			if (len != PAGE_SIZE) {
1073 				pmap_zero_page_area(m, offs, len);
1074 				vm_page_clear_dirty(m, offs, len);
1075 				m->valid = VM_PAGE_BITS_ALL;
1076 			} else
1077 				vm_pager_page_unswapped(m);
1078 		}
1079 		vm_page_xunbusy(m);
1080 		vm_page_lock(m);
1081 		if (bp->bio_cmd == BIO_DELETE && len == PAGE_SIZE)
1082 			vm_page_free(m);
1083 		else
1084 			vm_page_activate(m);
1085 		vm_page_unlock(m);
1086 		if (bp->bio_cmd == BIO_WRITE) {
1087 			vm_page_dirty(m);
1088 			vm_pager_page_unswapped(m);
1089 		}
1090 
1091 		/* Actions on further pages start at offset 0 */
1092 		p += PAGE_SIZE - offs;
1093 		offs = 0;
1094 		ma_offs += len;
1095 	}
1096 	vm_object_pip_wakeup(sc->object);
1097 	VM_OBJECT_WUNLOCK(sc->object);
1098 	return (rv != VM_PAGER_ERROR ? 0 : ENOSPC);
1099 }
1100 
1101 static int
1102 mdstart_null(struct md_s *sc, struct bio *bp)
1103 {
1104 
1105 	switch (bp->bio_cmd) {
1106 	case BIO_READ:
1107 		bzero(bp->bio_data, bp->bio_length);
1108 		cpu_flush_dcache(bp->bio_data, bp->bio_length);
1109 		break;
1110 	case BIO_WRITE:
1111 		break;
1112 	}
1113 	bp->bio_resid = 0;
1114 	return (0);
1115 }
1116 
1117 static void
1118 md_kthread(void *arg)
1119 {
1120 	struct md_s *sc;
1121 	struct bio *bp;
1122 	int error;
1123 
1124 	sc = arg;
1125 	thread_lock(curthread);
1126 	sched_prio(curthread, PRIBIO);
1127 	thread_unlock(curthread);
1128 	if (sc->type == MD_VNODE)
1129 		curthread->td_pflags |= TDP_NORUNNINGBUF;
1130 
1131 	for (;;) {
1132 		mtx_lock(&sc->queue_mtx);
1133 		if (sc->flags & MD_SHUTDOWN) {
1134 			sc->flags |= MD_EXITING;
1135 			mtx_unlock(&sc->queue_mtx);
1136 			kproc_exit(0);
1137 		}
1138 		bp = bioq_takefirst(&sc->bio_queue);
1139 		if (!bp) {
1140 			msleep(sc, &sc->queue_mtx, PRIBIO | PDROP, "mdwait", 0);
1141 			continue;
1142 		}
1143 		mtx_unlock(&sc->queue_mtx);
1144 		if (bp->bio_cmd == BIO_GETATTR) {
1145 			if ((sc->fwsectors && sc->fwheads &&
1146 			    (g_handleattr_int(bp, "GEOM::fwsectors",
1147 			    sc->fwsectors) ||
1148 			    g_handleattr_int(bp, "GEOM::fwheads",
1149 			    sc->fwheads))) ||
1150 			    g_handleattr_int(bp, "GEOM::candelete", 1))
1151 				error = -1;
1152 			else
1153 				error = EOPNOTSUPP;
1154 		} else {
1155 			error = sc->start(sc, bp);
1156 		}
1157 
1158 		if (error != -1) {
1159 			bp->bio_completed = bp->bio_length;
1160 			if ((bp->bio_cmd == BIO_READ) || (bp->bio_cmd == BIO_WRITE))
1161 				devstat_end_transaction_bio(sc->devstat, bp);
1162 			g_io_deliver(bp, error);
1163 		}
1164 	}
1165 }
1166 
1167 static struct md_s *
1168 mdfind(int unit)
1169 {
1170 	struct md_s *sc;
1171 
1172 	LIST_FOREACH(sc, &md_softc_list, list) {
1173 		if (sc->unit == unit)
1174 			break;
1175 	}
1176 	return (sc);
1177 }
1178 
1179 static struct md_s *
1180 mdnew(int unit, int *errp, enum md_types type)
1181 {
1182 	struct md_s *sc;
1183 	int error;
1184 
1185 	*errp = 0;
1186 	if (unit == -1)
1187 		unit = alloc_unr(md_uh);
1188 	else
1189 		unit = alloc_unr_specific(md_uh, unit);
1190 
1191 	if (unit == -1) {
1192 		*errp = EBUSY;
1193 		return (NULL);
1194 	}
1195 
1196 	sc = (struct md_s *)malloc(sizeof *sc, M_MD, M_WAITOK | M_ZERO);
1197 	sc->type = type;
1198 	bioq_init(&sc->bio_queue);
1199 	mtx_init(&sc->queue_mtx, "md bio queue", NULL, MTX_DEF);
1200 	mtx_init(&sc->stat_mtx, "md stat", NULL, MTX_DEF);
1201 	sc->unit = unit;
1202 	sprintf(sc->name, "md%d", unit);
1203 	LIST_INSERT_HEAD(&md_softc_list, sc, list);
1204 	error = kproc_create(md_kthread, sc, &sc->procp, 0, 0,"%s", sc->name);
1205 	if (error == 0)
1206 		return (sc);
1207 	LIST_REMOVE(sc, list);
1208 	mtx_destroy(&sc->stat_mtx);
1209 	mtx_destroy(&sc->queue_mtx);
1210 	free_unr(md_uh, sc->unit);
1211 	free(sc, M_MD);
1212 	*errp = error;
1213 	return (NULL);
1214 }
1215 
1216 static void
1217 mdinit(struct md_s *sc)
1218 {
1219 	struct g_geom *gp;
1220 	struct g_provider *pp;
1221 
1222 	g_topology_lock();
1223 	gp = g_new_geomf(&g_md_class, "md%d", sc->unit);
1224 	gp->softc = sc;
1225 	pp = g_new_providerf(gp, "md%d", sc->unit);
1226 	pp->flags |= G_PF_DIRECT_SEND | G_PF_DIRECT_RECEIVE;
1227 	pp->mediasize = sc->mediasize;
1228 	pp->sectorsize = sc->sectorsize;
1229 	switch (sc->type) {
1230 	case MD_MALLOC:
1231 	case MD_VNODE:
1232 	case MD_SWAP:
1233 		pp->flags |= G_PF_ACCEPT_UNMAPPED;
1234 		break;
1235 	case MD_PRELOAD:
1236 	case MD_NULL:
1237 		break;
1238 	}
1239 	sc->gp = gp;
1240 	sc->pp = pp;
1241 	g_error_provider(pp, 0);
1242 	g_topology_unlock();
1243 	sc->devstat = devstat_new_entry("md", sc->unit, sc->sectorsize,
1244 	    DEVSTAT_ALL_SUPPORTED, DEVSTAT_TYPE_DIRECT, DEVSTAT_PRIORITY_MAX);
1245 }
1246 
1247 static int
1248 mdcreate_malloc(struct md_s *sc, struct md_ioctl *mdio)
1249 {
1250 	uintptr_t sp;
1251 	int error;
1252 	off_t u;
1253 
1254 	error = 0;
1255 	if (mdio->md_options & ~(MD_AUTOUNIT | MD_COMPRESS | MD_RESERVE))
1256 		return (EINVAL);
1257 	if (mdio->md_sectorsize != 0 && !powerof2(mdio->md_sectorsize))
1258 		return (EINVAL);
1259 	/* Compression doesn't make sense if we have reserved space */
1260 	if (mdio->md_options & MD_RESERVE)
1261 		mdio->md_options &= ~MD_COMPRESS;
1262 	if (mdio->md_fwsectors != 0)
1263 		sc->fwsectors = mdio->md_fwsectors;
1264 	if (mdio->md_fwheads != 0)
1265 		sc->fwheads = mdio->md_fwheads;
1266 	sc->flags = mdio->md_options & (MD_COMPRESS | MD_FORCE);
1267 	sc->indir = dimension(sc->mediasize / sc->sectorsize);
1268 	sc->uma = uma_zcreate(sc->name, sc->sectorsize, NULL, NULL, NULL, NULL,
1269 	    0x1ff, 0);
1270 	if (mdio->md_options & MD_RESERVE) {
1271 		off_t nsectors;
1272 
1273 		nsectors = sc->mediasize / sc->sectorsize;
1274 		for (u = 0; u < nsectors; u++) {
1275 			sp = (uintptr_t)uma_zalloc(sc->uma, (md_malloc_wait ?
1276 			    M_WAITOK : M_NOWAIT) | M_ZERO);
1277 			if (sp != 0)
1278 				error = s_write(sc->indir, u, sp);
1279 			else
1280 				error = ENOMEM;
1281 			if (error != 0)
1282 				break;
1283 		}
1284 	}
1285 	return (error);
1286 }
1287 
1288 
1289 static int
1290 mdsetcred(struct md_s *sc, struct ucred *cred)
1291 {
1292 	char *tmpbuf;
1293 	int error = 0;
1294 
1295 	/*
1296 	 * Set credits in our softc
1297 	 */
1298 
1299 	if (sc->cred)
1300 		crfree(sc->cred);
1301 	sc->cred = crhold(cred);
1302 
1303 	/*
1304 	 * Horrible kludge to establish credentials for NFS  XXX.
1305 	 */
1306 
1307 	if (sc->vnode) {
1308 		struct uio auio;
1309 		struct iovec aiov;
1310 
1311 		tmpbuf = malloc(sc->sectorsize, M_TEMP, M_WAITOK);
1312 		bzero(&auio, sizeof(auio));
1313 
1314 		aiov.iov_base = tmpbuf;
1315 		aiov.iov_len = sc->sectorsize;
1316 		auio.uio_iov = &aiov;
1317 		auio.uio_iovcnt = 1;
1318 		auio.uio_offset = 0;
1319 		auio.uio_rw = UIO_READ;
1320 		auio.uio_segflg = UIO_SYSSPACE;
1321 		auio.uio_resid = aiov.iov_len;
1322 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY);
1323 		error = VOP_READ(sc->vnode, &auio, 0, sc->cred);
1324 		VOP_UNLOCK(sc->vnode, 0);
1325 		free(tmpbuf, M_TEMP);
1326 	}
1327 	return (error);
1328 }
1329 
1330 static int
1331 mdcreate_vnode(struct md_s *sc, struct md_ioctl *mdio, struct thread *td)
1332 {
1333 	struct vattr vattr;
1334 	struct nameidata nd;
1335 	char *fname;
1336 	int error, flags;
1337 
1338 	/*
1339 	 * Kernel-originated requests must have the filename appended
1340 	 * to the mdio structure to protect against malicious software.
1341 	 */
1342 	fname = mdio->md_file;
1343 	if ((void *)fname != (void *)(mdio + 1)) {
1344 		error = copyinstr(fname, sc->file, sizeof(sc->file), NULL);
1345 		if (error != 0)
1346 			return (error);
1347 	} else
1348 		strlcpy(sc->file, fname, sizeof(sc->file));
1349 
1350 	/*
1351 	 * If the user specified that this is a read only device, don't
1352 	 * set the FWRITE mask before trying to open the backing store.
1353 	 */
1354 	flags = FREAD | ((mdio->md_options & MD_READONLY) ? 0 : FWRITE);
1355 	NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, sc->file, td);
1356 	error = vn_open(&nd, &flags, 0, NULL);
1357 	if (error != 0)
1358 		return (error);
1359 	NDFREE(&nd, NDF_ONLY_PNBUF);
1360 	if (nd.ni_vp->v_type != VREG) {
1361 		error = EINVAL;
1362 		goto bad;
1363 	}
1364 	error = VOP_GETATTR(nd.ni_vp, &vattr, td->td_ucred);
1365 	if (error != 0)
1366 		goto bad;
1367 	if (VOP_ISLOCKED(nd.ni_vp) != LK_EXCLUSIVE) {
1368 		vn_lock(nd.ni_vp, LK_UPGRADE | LK_RETRY);
1369 		if (nd.ni_vp->v_iflag & VI_DOOMED) {
1370 			/* Forced unmount. */
1371 			error = EBADF;
1372 			goto bad;
1373 		}
1374 	}
1375 	nd.ni_vp->v_vflag |= VV_MD;
1376 	VOP_UNLOCK(nd.ni_vp, 0);
1377 
1378 	if (mdio->md_fwsectors != 0)
1379 		sc->fwsectors = mdio->md_fwsectors;
1380 	if (mdio->md_fwheads != 0)
1381 		sc->fwheads = mdio->md_fwheads;
1382 	sc->flags = mdio->md_options & (MD_FORCE | MD_ASYNC);
1383 	if (!(flags & FWRITE))
1384 		sc->flags |= MD_READONLY;
1385 	sc->vnode = nd.ni_vp;
1386 
1387 	error = mdsetcred(sc, td->td_ucred);
1388 	if (error != 0) {
1389 		sc->vnode = NULL;
1390 		vn_lock(nd.ni_vp, LK_EXCLUSIVE | LK_RETRY);
1391 		nd.ni_vp->v_vflag &= ~VV_MD;
1392 		goto bad;
1393 	}
1394 	return (0);
1395 bad:
1396 	VOP_UNLOCK(nd.ni_vp, 0);
1397 	(void)vn_close(nd.ni_vp, flags, td->td_ucred, td);
1398 	return (error);
1399 }
1400 
1401 static int
1402 mddestroy(struct md_s *sc, struct thread *td)
1403 {
1404 
1405 	if (sc->gp) {
1406 		sc->gp->softc = NULL;
1407 		g_topology_lock();
1408 		g_wither_geom(sc->gp, ENXIO);
1409 		g_topology_unlock();
1410 		sc->gp = NULL;
1411 		sc->pp = NULL;
1412 	}
1413 	if (sc->devstat) {
1414 		devstat_remove_entry(sc->devstat);
1415 		sc->devstat = NULL;
1416 	}
1417 	mtx_lock(&sc->queue_mtx);
1418 	sc->flags |= MD_SHUTDOWN;
1419 	wakeup(sc);
1420 	while (!(sc->flags & MD_EXITING))
1421 		msleep(sc->procp, &sc->queue_mtx, PRIBIO, "mddestroy", hz / 10);
1422 	mtx_unlock(&sc->queue_mtx);
1423 	mtx_destroy(&sc->stat_mtx);
1424 	mtx_destroy(&sc->queue_mtx);
1425 	if (sc->vnode != NULL) {
1426 		vn_lock(sc->vnode, LK_EXCLUSIVE | LK_RETRY);
1427 		sc->vnode->v_vflag &= ~VV_MD;
1428 		VOP_UNLOCK(sc->vnode, 0);
1429 		(void)vn_close(sc->vnode, sc->flags & MD_READONLY ?
1430 		    FREAD : (FREAD|FWRITE), sc->cred, td);
1431 	}
1432 	if (sc->cred != NULL)
1433 		crfree(sc->cred);
1434 	if (sc->object != NULL)
1435 		vm_object_deallocate(sc->object);
1436 	if (sc->indir)
1437 		destroy_indir(sc, sc->indir);
1438 	if (sc->uma)
1439 		uma_zdestroy(sc->uma);
1440 
1441 	LIST_REMOVE(sc, list);
1442 	free_unr(md_uh, sc->unit);
1443 	free(sc, M_MD);
1444 	return (0);
1445 }
1446 
1447 static int
1448 mdresize(struct md_s *sc, struct md_ioctl *mdio)
1449 {
1450 	int error, res;
1451 	vm_pindex_t oldpages, newpages;
1452 
1453 	switch (sc->type) {
1454 	case MD_VNODE:
1455 	case MD_NULL:
1456 		break;
1457 	case MD_SWAP:
1458 		if (mdio->md_mediasize <= 0 ||
1459 		    (mdio->md_mediasize % PAGE_SIZE) != 0)
1460 			return (EDOM);
1461 		oldpages = OFF_TO_IDX(round_page(sc->mediasize));
1462 		newpages = OFF_TO_IDX(round_page(mdio->md_mediasize));
1463 		if (newpages < oldpages) {
1464 			VM_OBJECT_WLOCK(sc->object);
1465 			vm_object_page_remove(sc->object, newpages, 0, 0);
1466 			swap_pager_freespace(sc->object, newpages,
1467 			    oldpages - newpages);
1468 			swap_release_by_cred(IDX_TO_OFF(oldpages -
1469 			    newpages), sc->cred);
1470 			sc->object->charge = IDX_TO_OFF(newpages);
1471 			sc->object->size = newpages;
1472 			VM_OBJECT_WUNLOCK(sc->object);
1473 		} else if (newpages > oldpages) {
1474 			res = swap_reserve_by_cred(IDX_TO_OFF(newpages -
1475 			    oldpages), sc->cred);
1476 			if (!res)
1477 				return (ENOMEM);
1478 			if ((mdio->md_options & MD_RESERVE) ||
1479 			    (sc->flags & MD_RESERVE)) {
1480 				error = swap_pager_reserve(sc->object,
1481 				    oldpages, newpages - oldpages);
1482 				if (error < 0) {
1483 					swap_release_by_cred(
1484 					    IDX_TO_OFF(newpages - oldpages),
1485 					    sc->cred);
1486 					return (EDOM);
1487 				}
1488 			}
1489 			VM_OBJECT_WLOCK(sc->object);
1490 			sc->object->charge = IDX_TO_OFF(newpages);
1491 			sc->object->size = newpages;
1492 			VM_OBJECT_WUNLOCK(sc->object);
1493 		}
1494 		break;
1495 	default:
1496 		return (EOPNOTSUPP);
1497 	}
1498 
1499 	sc->mediasize = mdio->md_mediasize;
1500 	g_topology_lock();
1501 	g_resize_provider(sc->pp, sc->mediasize);
1502 	g_topology_unlock();
1503 	return (0);
1504 }
1505 
1506 static int
1507 mdcreate_swap(struct md_s *sc, struct md_ioctl *mdio, struct thread *td)
1508 {
1509 	vm_ooffset_t npage;
1510 	int error;
1511 
1512 	/*
1513 	 * Range check.  Disallow negative sizes and sizes not being
1514 	 * multiple of page size.
1515 	 */
1516 	if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0)
1517 		return (EDOM);
1518 
1519 	/*
1520 	 * Allocate an OBJT_SWAP object.
1521 	 *
1522 	 * Note the truncation.
1523 	 */
1524 
1525 	npage = mdio->md_mediasize / PAGE_SIZE;
1526 	if (mdio->md_fwsectors != 0)
1527 		sc->fwsectors = mdio->md_fwsectors;
1528 	if (mdio->md_fwheads != 0)
1529 		sc->fwheads = mdio->md_fwheads;
1530 	sc->object = vm_pager_allocate(OBJT_SWAP, NULL, PAGE_SIZE * npage,
1531 	    VM_PROT_DEFAULT, 0, td->td_ucred);
1532 	if (sc->object == NULL)
1533 		return (ENOMEM);
1534 	sc->flags = mdio->md_options & (MD_FORCE | MD_RESERVE);
1535 	if (mdio->md_options & MD_RESERVE) {
1536 		if (swap_pager_reserve(sc->object, 0, npage) < 0) {
1537 			error = EDOM;
1538 			goto finish;
1539 		}
1540 	}
1541 	error = mdsetcred(sc, td->td_ucred);
1542  finish:
1543 	if (error != 0) {
1544 		vm_object_deallocate(sc->object);
1545 		sc->object = NULL;
1546 	}
1547 	return (error);
1548 }
1549 
1550 static int
1551 mdcreate_null(struct md_s *sc, struct md_ioctl *mdio, struct thread *td)
1552 {
1553 
1554 	/*
1555 	 * Range check.  Disallow negative sizes and sizes not being
1556 	 * multiple of page size.
1557 	 */
1558 	if (sc->mediasize <= 0 || (sc->mediasize % PAGE_SIZE) != 0)
1559 		return (EDOM);
1560 
1561 	return (0);
1562 }
1563 
1564 static int
1565 xmdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td)
1566 {
1567 	struct md_ioctl *mdio;
1568 	struct md_s *sc;
1569 	int error, i;
1570 	unsigned sectsize;
1571 
1572 	if (md_debug)
1573 		printf("mdctlioctl(%s %lx %p %x %p)\n",
1574 			devtoname(dev), cmd, addr, flags, td);
1575 
1576 	mdio = (struct md_ioctl *)addr;
1577 	if (mdio->md_version != MDIOVERSION)
1578 		return (EINVAL);
1579 
1580 	/*
1581 	 * We assert the version number in the individual ioctl
1582 	 * handlers instead of out here because (a) it is possible we
1583 	 * may add another ioctl in the future which doesn't read an
1584 	 * mdio, and (b) the correct return value for an unknown ioctl
1585 	 * is ENOIOCTL, not EINVAL.
1586 	 */
1587 	error = 0;
1588 	switch (cmd) {
1589 	case MDIOCATTACH:
1590 		switch (mdio->md_type) {
1591 		case MD_MALLOC:
1592 		case MD_PRELOAD:
1593 		case MD_VNODE:
1594 		case MD_SWAP:
1595 		case MD_NULL:
1596 			break;
1597 		default:
1598 			return (EINVAL);
1599 		}
1600 		if (mdio->md_sectorsize == 0)
1601 			sectsize = DEV_BSIZE;
1602 		else
1603 			sectsize = mdio->md_sectorsize;
1604 		if (sectsize > MAXPHYS || mdio->md_mediasize < sectsize)
1605 			return (EINVAL);
1606 		if (mdio->md_options & MD_AUTOUNIT)
1607 			sc = mdnew(-1, &error, mdio->md_type);
1608 		else {
1609 			if (mdio->md_unit > INT_MAX)
1610 				return (EINVAL);
1611 			sc = mdnew(mdio->md_unit, &error, mdio->md_type);
1612 		}
1613 		if (sc == NULL)
1614 			return (error);
1615 		if (mdio->md_options & MD_AUTOUNIT)
1616 			mdio->md_unit = sc->unit;
1617 		sc->mediasize = mdio->md_mediasize;
1618 		sc->sectorsize = sectsize;
1619 		error = EDOOFUS;
1620 		switch (sc->type) {
1621 		case MD_MALLOC:
1622 			sc->start = mdstart_malloc;
1623 			error = mdcreate_malloc(sc, mdio);
1624 			break;
1625 		case MD_PRELOAD:
1626 			/*
1627 			 * We disallow attaching preloaded memory disks via
1628 			 * ioctl. Preloaded memory disks are automatically
1629 			 * attached in g_md_init().
1630 			 */
1631 			error = EOPNOTSUPP;
1632 			break;
1633 		case MD_VNODE:
1634 			sc->start = mdstart_vnode;
1635 			error = mdcreate_vnode(sc, mdio, td);
1636 			break;
1637 		case MD_SWAP:
1638 			sc->start = mdstart_swap;
1639 			error = mdcreate_swap(sc, mdio, td);
1640 			break;
1641 		case MD_NULL:
1642 			sc->start = mdstart_null;
1643 			error = mdcreate_null(sc, mdio, td);
1644 			break;
1645 		}
1646 		if (error != 0) {
1647 			mddestroy(sc, td);
1648 			return (error);
1649 		}
1650 
1651 		/* Prune off any residual fractional sector */
1652 		i = sc->mediasize % sc->sectorsize;
1653 		sc->mediasize -= i;
1654 
1655 		mdinit(sc);
1656 		return (0);
1657 	case MDIOCDETACH:
1658 		if (mdio->md_mediasize != 0 ||
1659 		    (mdio->md_options & ~MD_FORCE) != 0)
1660 			return (EINVAL);
1661 
1662 		sc = mdfind(mdio->md_unit);
1663 		if (sc == NULL)
1664 			return (ENOENT);
1665 		if (sc->opencount != 0 && !(sc->flags & MD_FORCE) &&
1666 		    !(mdio->md_options & MD_FORCE))
1667 			return (EBUSY);
1668 		return (mddestroy(sc, td));
1669 	case MDIOCRESIZE:
1670 		if ((mdio->md_options & ~(MD_FORCE | MD_RESERVE)) != 0)
1671 			return (EINVAL);
1672 
1673 		sc = mdfind(mdio->md_unit);
1674 		if (sc == NULL)
1675 			return (ENOENT);
1676 		if (mdio->md_mediasize < sc->sectorsize)
1677 			return (EINVAL);
1678 		if (mdio->md_mediasize < sc->mediasize &&
1679 		    !(sc->flags & MD_FORCE) &&
1680 		    !(mdio->md_options & MD_FORCE))
1681 			return (EBUSY);
1682 		return (mdresize(sc, mdio));
1683 	case MDIOCQUERY:
1684 		sc = mdfind(mdio->md_unit);
1685 		if (sc == NULL)
1686 			return (ENOENT);
1687 		mdio->md_type = sc->type;
1688 		mdio->md_options = sc->flags;
1689 		mdio->md_mediasize = sc->mediasize;
1690 		mdio->md_sectorsize = sc->sectorsize;
1691 		if (sc->type == MD_VNODE)
1692 			error = copyout(sc->file, mdio->md_file,
1693 			    strlen(sc->file) + 1);
1694 		return (error);
1695 	case MDIOCLIST:
1696 		i = 1;
1697 		LIST_FOREACH(sc, &md_softc_list, list) {
1698 			if (i == MDNPAD - 1)
1699 				mdio->md_pad[i] = -1;
1700 			else
1701 				mdio->md_pad[i++] = sc->unit;
1702 		}
1703 		mdio->md_pad[0] = i - 1;
1704 		return (0);
1705 	default:
1706 		return (ENOIOCTL);
1707 	};
1708 }
1709 
1710 static int
1711 mdctlioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td)
1712 {
1713 	int error;
1714 
1715 	sx_xlock(&md_sx);
1716 	error = xmdctlioctl(dev, cmd, addr, flags, td);
1717 	sx_xunlock(&md_sx);
1718 	return (error);
1719 }
1720 
1721 static void
1722 md_preloaded(u_char *image, size_t length, const char *name)
1723 {
1724 	struct md_s *sc;
1725 	int error;
1726 
1727 	sc = mdnew(-1, &error, MD_PRELOAD);
1728 	if (sc == NULL)
1729 		return;
1730 	sc->mediasize = length;
1731 	sc->sectorsize = DEV_BSIZE;
1732 	sc->pl_ptr = image;
1733 	sc->pl_len = length;
1734 	sc->start = mdstart_preload;
1735 #ifdef MD_ROOT
1736 	if (sc->unit == 0)
1737 		rootdevnames[0] = MD_ROOT_FSTYPE ":/dev/md0";
1738 #endif
1739 	mdinit(sc);
1740 	if (name != NULL) {
1741 		printf("%s%d: Preloaded image <%s> %zd bytes at %p\n",
1742 		    MD_NAME, sc->unit, name, length, image);
1743 	} else {
1744 		printf("%s%d: Embedded image %zd bytes at %p\n",
1745 		    MD_NAME, sc->unit, length, image);
1746 	}
1747 }
1748 
1749 static void
1750 g_md_init(struct g_class *mp __unused)
1751 {
1752 	caddr_t mod;
1753 	u_char *ptr, *name, *type;
1754 	unsigned len;
1755 	int i;
1756 
1757 	/* figure out log2(NINDIR) */
1758 	for (i = NINDIR, nshift = -1; i; nshift++)
1759 		i >>= 1;
1760 
1761 	mod = NULL;
1762 	sx_init(&md_sx, "MD config lock");
1763 	g_topology_unlock();
1764 	md_uh = new_unrhdr(0, INT_MAX, NULL);
1765 #ifdef MD_ROOT
1766 	if (mfs_root_size != 0) {
1767 		sx_xlock(&md_sx);
1768 		md_preloaded(__DEVOLATILE(u_char *, &mfs_root), mfs_root_size,
1769 		    NULL);
1770 		sx_xunlock(&md_sx);
1771 	}
1772 #endif
1773 	/* XXX: are preload_* static or do they need Giant ? */
1774 	while ((mod = preload_search_next_name(mod)) != NULL) {
1775 		name = (char *)preload_search_info(mod, MODINFO_NAME);
1776 		if (name == NULL)
1777 			continue;
1778 		type = (char *)preload_search_info(mod, MODINFO_TYPE);
1779 		if (type == NULL)
1780 			continue;
1781 		if (strcmp(type, "md_image") && strcmp(type, "mfs_root"))
1782 			continue;
1783 		ptr = preload_fetch_addr(mod);
1784 		len = preload_fetch_size(mod);
1785 		if (ptr != NULL && len != 0) {
1786 			sx_xlock(&md_sx);
1787 			md_preloaded(ptr, len, name);
1788 			sx_xunlock(&md_sx);
1789 		}
1790 	}
1791 	md_vnode_pbuf_freecnt = nswbuf / 10;
1792 	status_dev = make_dev(&mdctl_cdevsw, INT_MAX, UID_ROOT, GID_WHEEL,
1793 	    0600, MDCTL_NAME);
1794 	g_topology_lock();
1795 }
1796 
1797 static void
1798 g_md_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1799     struct g_consumer *cp __unused, struct g_provider *pp)
1800 {
1801 	struct md_s *mp;
1802 	char *type;
1803 
1804 	mp = gp->softc;
1805 	if (mp == NULL)
1806 		return;
1807 
1808 	switch (mp->type) {
1809 	case MD_MALLOC:
1810 		type = "malloc";
1811 		break;
1812 	case MD_PRELOAD:
1813 		type = "preload";
1814 		break;
1815 	case MD_VNODE:
1816 		type = "vnode";
1817 		break;
1818 	case MD_SWAP:
1819 		type = "swap";
1820 		break;
1821 	case MD_NULL:
1822 		type = "null";
1823 		break;
1824 	default:
1825 		type = "unknown";
1826 		break;
1827 	}
1828 
1829 	if (pp != NULL) {
1830 		if (indent == NULL) {
1831 			sbuf_printf(sb, " u %d", mp->unit);
1832 			sbuf_printf(sb, " s %ju", (uintmax_t) mp->sectorsize);
1833 			sbuf_printf(sb, " f %ju", (uintmax_t) mp->fwheads);
1834 			sbuf_printf(sb, " fs %ju", (uintmax_t) mp->fwsectors);
1835 			sbuf_printf(sb, " l %ju", (uintmax_t) mp->mediasize);
1836 			sbuf_printf(sb, " t %s", type);
1837 			if (mp->type == MD_VNODE && mp->vnode != NULL)
1838 				sbuf_printf(sb, " file %s", mp->file);
1839 		} else {
1840 			sbuf_printf(sb, "%s<unit>%d</unit>\n", indent,
1841 			    mp->unit);
1842 			sbuf_printf(sb, "%s<sectorsize>%ju</sectorsize>\n",
1843 			    indent, (uintmax_t) mp->sectorsize);
1844 			sbuf_printf(sb, "%s<fwheads>%ju</fwheads>\n",
1845 			    indent, (uintmax_t) mp->fwheads);
1846 			sbuf_printf(sb, "%s<fwsectors>%ju</fwsectors>\n",
1847 			    indent, (uintmax_t) mp->fwsectors);
1848 			sbuf_printf(sb, "%s<length>%ju</length>\n",
1849 			    indent, (uintmax_t) mp->mediasize);
1850 			sbuf_printf(sb, "%s<compression>%s</compression>\n", indent,
1851 			    (mp->flags & MD_COMPRESS) == 0 ? "off": "on");
1852 			sbuf_printf(sb, "%s<access>%s</access>\n", indent,
1853 			    (mp->flags & MD_READONLY) == 0 ? "read-write":
1854 			    "read-only");
1855 			sbuf_printf(sb, "%s<type>%s</type>\n", indent,
1856 			    type);
1857 			if (mp->type == MD_VNODE && mp->vnode != NULL) {
1858 				sbuf_printf(sb, "%s<file>", indent);
1859 				g_conf_printf_escaped(sb, "%s", mp->file);
1860 				sbuf_printf(sb, "</file>\n");
1861 			}
1862 		}
1863 	}
1864 }
1865 
1866 static void
1867 g_md_fini(struct g_class *mp __unused)
1868 {
1869 
1870 	sx_destroy(&md_sx);
1871 	if (status_dev != NULL)
1872 		destroy_dev(status_dev);
1873 	delete_unrhdr(md_uh);
1874 }
1875