xref: /freebsd/sys/geom/union/g_union.c (revision abd87254)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2022 Marshall Kirk McKusick <mckusick@mckusick.com>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/param.h>
29 #include <sys/bio.h>
30 #include <sys/buf.h>
31 #include <sys/ctype.h>
32 #include <sys/kernel.h>
33 #include <sys/lock.h>
34 #include <sys/malloc.h>
35 #include <sys/module.h>
36 #include <sys/reboot.h>
37 #include <sys/rwlock.h>
38 #include <sys/sbuf.h>
39 #include <sys/sysctl.h>
40 
41 #include <geom/geom.h>
42 #include <geom/geom_dbg.h>
43 #include <geom/union/g_union.h>
44 
45 SYSCTL_DECL(_kern_geom);
46 static SYSCTL_NODE(_kern_geom, OID_AUTO, union, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
47     "GEOM_UNION stuff");
48 static u_int g_union_debug = 0;
49 SYSCTL_UINT(_kern_geom_union, OID_AUTO, debug, CTLFLAG_RW, &g_union_debug, 0,
50     "Debug level");
51 
52 static void g_union_config(struct gctl_req *req, struct g_class *mp,
53     const char *verb);
54 static g_access_t g_union_access;
55 static g_start_t g_union_start;
56 static g_dumpconf_t g_union_dumpconf;
57 static g_orphan_t g_union_orphan;
58 static int g_union_destroy_geom(struct gctl_req *req, struct g_class *mp,
59     struct g_geom *gp);
60 static g_provgone_t g_union_providergone;
61 static g_resize_t g_union_resize;
62 
63 struct g_class g_union_class = {
64 	.name = G_UNION_CLASS_NAME,
65 	.version = G_VERSION,
66 	.ctlreq = g_union_config,
67 	.access = g_union_access,
68 	.start = g_union_start,
69 	.dumpconf = g_union_dumpconf,
70 	.orphan = g_union_orphan,
71 	.destroy_geom = g_union_destroy_geom,
72 	.providergone = g_union_providergone,
73 	.resize = g_union_resize,
74 };
75 
76 static void g_union_ctl_create(struct gctl_req *req, struct g_class *mp, bool);
77 static intmax_t g_union_fetcharg(struct gctl_req *req, const char *name);
78 static bool g_union_verify_nprefix(const char *name);
79 static void g_union_ctl_destroy(struct gctl_req *req, struct g_class *mp, bool);
80 static struct g_geom *g_union_find_geom(struct g_class *mp, const char *name);
81 static void g_union_ctl_reset(struct gctl_req *req, struct g_class *mp, bool);
82 static void g_union_ctl_revert(struct gctl_req *req, struct g_class *mp, bool);
83 static void g_union_revert(struct g_union_softc *sc);
84 static void g_union_doio(struct g_union_wip *wip);
85 static void g_union_ctl_commit(struct gctl_req *req, struct g_class *mp, bool);
86 static void g_union_setmap(struct bio *bp, struct g_union_softc *sc);
87 static bool g_union_getmap(struct bio *bp, struct g_union_softc *sc,
88 	off_t *len2read);
89 static void g_union_done(struct bio *bp);
90 static void g_union_kerneldump(struct bio *bp, struct g_union_softc *sc);
91 static int g_union_dumper(void *, void *, off_t, size_t);
92 static int g_union_destroy(struct gctl_req *req, struct g_geom *gp, bool force);
93 
94 /*
95  * Operate on union-specific configuration commands.
96  */
97 static void
98 g_union_config(struct gctl_req *req, struct g_class *mp, const char *verb)
99 {
100 	uint32_t *version, *verbose;
101 
102 	g_topology_assert();
103 
104 	version = gctl_get_paraml(req, "version", sizeof(*version));
105 	if (version == NULL) {
106 		gctl_error(req, "No '%s' argument.", "version");
107 		return;
108 	}
109 	if (*version != G_UNION_VERSION) {
110 		gctl_error(req, "Userland and kernel parts are out of sync.");
111 		return;
112 	}
113 	verbose = gctl_get_paraml(req, "verbose", sizeof(*verbose));
114 	if (verbose == NULL) {
115 		gctl_error(req, "No '%s' argument.", "verbose");
116 		return;
117 	}
118 	if (strcmp(verb, "create") == 0) {
119 		g_union_ctl_create(req, mp, *verbose);
120 		return;
121 	} else if (strcmp(verb, "destroy") == 0) {
122 		g_union_ctl_destroy(req, mp, *verbose);
123 		return;
124 	} else if (strcmp(verb, "reset") == 0) {
125 		g_union_ctl_reset(req, mp, *verbose);
126 		return;
127 	} else if (strcmp(verb, "revert") == 0) {
128 		g_union_ctl_revert(req, mp, *verbose);
129 		return;
130 	} else if (strcmp(verb, "commit") == 0) {
131 		g_union_ctl_commit(req, mp, *verbose);
132 		return;
133 	}
134 
135 	gctl_error(req, "Unknown verb.");
136 }
137 
138 /*
139  * Create a union device.
140  */
141 static void
142 g_union_ctl_create(struct gctl_req *req, struct g_class *mp, bool verbose)
143 {
144 	struct g_provider *upperpp, *lowerpp, *newpp;
145 	struct g_consumer *uppercp, *lowercp;
146 	struct g_union_softc *sc;
147 	struct g_geom_alias *gap;
148 	struct g_geom *gp;
149 	intmax_t offset, secsize, size, needed;
150 	const char *gunionname;
151 	int *nargs, error, i, n;
152 	char name[64];
153 
154 	g_topology_assert();
155 
156 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
157 	if (nargs == NULL) {
158 		gctl_error(req, "No '%s' argument.", "nargs");
159 		return;
160 	}
161 	if (*nargs < 2) {
162 		gctl_error(req, "Missing device(s).");
163 		return;
164 	}
165 	if (*nargs > 2) {
166 		gctl_error(req, "Extra device(s).");
167 		return;
168 	}
169 
170 	offset = g_union_fetcharg(req, "offset");
171 	size = g_union_fetcharg(req, "size");
172 	secsize = g_union_fetcharg(req, "secsize");
173 	gunionname = gctl_get_asciiparam(req, "gunionname");
174 
175 	upperpp = gctl_get_provider(req, "arg0");
176 	lowerpp = gctl_get_provider(req, "arg1");
177 	if (upperpp == NULL || lowerpp == NULL)
178 		/* error message provided by gctl_get_provider() */
179 		return;
180 	/* Create the union */
181 	if (secsize == 0)
182 		secsize = lowerpp->sectorsize;
183 	else if ((secsize % lowerpp->sectorsize) != 0) {
184 		gctl_error(req, "Sector size %jd is not a multiple of lower "
185 		    "provider %s's %jd sector size.", (intmax_t)secsize,
186 		    lowerpp->name, (intmax_t)lowerpp->sectorsize);
187 		return;
188 	}
189 	if (secsize > maxphys) {
190 		gctl_error(req, "Too big secsize %jd for lower provider %s.",
191 		    (intmax_t)secsize, lowerpp->name);
192 		return;
193 	}
194 	if (secsize % upperpp->sectorsize != 0) {
195 		gctl_error(req, "Sector size %jd is not a multiple of upper "
196 		    "provider %s's %jd sector size.", (intmax_t)secsize,
197 		    upperpp->name, (intmax_t)upperpp->sectorsize);
198 		return;
199 	}
200 	if ((offset % secsize) != 0) {
201 		gctl_error(req, "Offset %jd is not a multiple of lower "
202 		    "provider %s's %jd sector size.", (intmax_t)offset,
203 		    lowerpp->name, (intmax_t)lowerpp->sectorsize);
204 		return;
205 	}
206 	if (size == 0)
207 		size = lowerpp->mediasize - offset;
208 	else
209 		size -= offset;
210 	if ((size % secsize) != 0) {
211 		gctl_error(req, "Size %jd is not a multiple of sector size "
212 		    "%jd.", (intmax_t)size, (intmax_t)secsize);
213 		return;
214 	}
215 	if (offset + size < lowerpp->mediasize) {
216 		gctl_error(req, "Size %jd is too small for lower provider %s, "
217 		    "needs %jd.", (intmax_t)(offset + size), lowerpp->name,
218 		    lowerpp->mediasize);
219 		return;
220 	}
221 	if (size > upperpp->mediasize) {
222 		gctl_error(req, "Upper provider %s size (%jd) is too small, "
223 		    "needs %jd.", upperpp->name, (intmax_t)upperpp->mediasize,
224 		    (intmax_t)size);
225 		return;
226 	}
227 	if (gunionname != NULL && !g_union_verify_nprefix(gunionname)) {
228 		gctl_error(req, "Gunion name %s must be alphanumeric.",
229 		    gunionname);
230 		return;
231 	}
232 	if (gunionname != NULL) {
233 		n = snprintf(name, sizeof(name), "%s%s", gunionname,
234 		    G_UNION_SUFFIX);
235 	} else {
236 		n = snprintf(name, sizeof(name), "%s-%s%s", upperpp->name,
237 		    lowerpp->name, G_UNION_SUFFIX);
238 	}
239 	if (n <= 0 || n >= sizeof(name)) {
240 		gctl_error(req, "Invalid provider name.");
241 		return;
242 	}
243 	LIST_FOREACH(gp, &mp->geom, geom) {
244 		if (strcmp(gp->name, name) == 0) {
245 			gctl_error(req, "Provider %s already exists.", name);
246 			return;
247 		}
248 	}
249 	gp = g_new_geomf(mp, "%s", name);
250 	sc = g_malloc(sizeof(*sc), M_WAITOK | M_ZERO);
251 	rw_init(&sc->sc_rwlock, "gunion");
252 	TAILQ_INIT(&sc->sc_wiplist);
253 	sc->sc_offset = offset;
254 	sc->sc_size = size;
255 	sc->sc_sectorsize = secsize;
256 	sc->sc_reads = 0;
257 	sc->sc_writes = 0;
258 	sc->sc_deletes = 0;
259 	sc->sc_getattrs = 0;
260 	sc->sc_flushes = 0;
261 	sc->sc_speedups = 0;
262 	sc->sc_cmd0s = 0;
263 	sc->sc_cmd1s = 0;
264 	sc->sc_cmd2s = 0;
265 	sc->sc_readbytes = 0;
266 	sc->sc_wrotebytes = 0;
267 	sc->sc_writemap_memory = 0;
268 	gp->softc = sc;
269 
270 	newpp = g_new_providerf(gp, "%s", gp->name);
271 	newpp->flags |= G_PF_DIRECT_SEND | G_PF_DIRECT_RECEIVE;
272 	newpp->mediasize = size;
273 	newpp->sectorsize = secsize;
274 	LIST_FOREACH(gap, &upperpp->aliases, ga_next)
275 		g_provider_add_alias(newpp, "%s%s", gap->ga_alias,
276 		    G_UNION_SUFFIX);
277 	LIST_FOREACH(gap, &lowerpp->aliases, ga_next)
278 		g_provider_add_alias(newpp, "%s%s", gap->ga_alias,
279 		    G_UNION_SUFFIX);
280 	lowercp = g_new_consumer(gp);
281 	lowercp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
282 	if ((error = g_attach(lowercp, lowerpp)) != 0) {
283 		gctl_error(req, "Error %d: cannot attach to provider %s.",
284 		    error, lowerpp->name);
285 		goto fail1;
286 	}
287 	/* request read and exclusive access for lower */
288 	if ((error = g_access(lowercp, 1, 0, 1)) != 0) {
289 		gctl_error(req, "Error %d: cannot obtain exclusive access to "
290 		    "%s.\n\tMust be unmounted or mounted read-only.", error,
291 		    lowerpp->name);
292 		goto fail2;
293 	}
294 	uppercp = g_new_consumer(gp);
295 	uppercp->flags |= G_CF_DIRECT_SEND | G_CF_DIRECT_RECEIVE;
296 	if ((error = g_attach(uppercp, upperpp)) != 0) {
297 		gctl_error(req, "Error %d: cannot attach to provider %s.",
298 		    error, upperpp->name);
299 		goto fail3;
300 	}
301 	/* request read, write, and exclusive access for upper */
302 	if ((error = g_access(uppercp, 1, 1, 1)) != 0) {
303 		gctl_error(req, "Error %d: cannot obtain write access to %s.",
304 		    error, upperpp->name);
305 		goto fail4;
306 	}
307 	sc->sc_uppercp = uppercp;
308 	sc->sc_lowercp = lowercp;
309 
310 	newpp->flags |= (upperpp->flags & G_PF_ACCEPT_UNMAPPED) &
311 	    (lowerpp->flags & G_PF_ACCEPT_UNMAPPED);
312 	g_error_provider(newpp, 0);
313 	/*
314 	 * Allocate the map that tracks the sectors that have been written
315 	 * to the top layer. We use a 2-level hierarchy as that lets us
316 	 * map up to 1 petabyte using allocations of less than 33 Mb
317 	 * when using 4K byte sectors (or 268 Mb with 512 byte sectors).
318 	 *
319 	 * We totally populate the leaf nodes rather than allocating them
320 	 * as they are first used because their usage occurs in the
321 	 * g_union_start() routine that may be running in the g_down
322 	 * thread which cannot sleep.
323 	 */
324 	sc->sc_map_size = roundup(size / secsize, BITS_PER_ENTRY);
325 	needed = sc->sc_map_size / BITS_PER_ENTRY;
326 	for (sc->sc_root_size = 1;
327 	     sc->sc_root_size * sc->sc_root_size < needed;
328 	     sc->sc_root_size++)
329 		continue;
330 	sc->sc_writemap_root = g_malloc(sc->sc_root_size * sizeof(uint64_t *),
331 	    M_WAITOK | M_ZERO);
332 	sc->sc_leaf_size = sc->sc_root_size;
333 	sc->sc_bits_per_leaf = sc->sc_leaf_size * BITS_PER_ENTRY;
334 	sc->sc_leafused = g_malloc(roundup(sc->sc_root_size, BITS_PER_ENTRY),
335 	    M_WAITOK | M_ZERO);
336 	for (i = 0; i < sc->sc_root_size; i++)
337 		sc->sc_writemap_root[i] =
338 		    g_malloc(sc->sc_leaf_size * sizeof(uint64_t),
339 		    M_WAITOK | M_ZERO);
340 	sc->sc_writemap_memory =
341 	    (sc->sc_root_size + sc->sc_root_size * sc->sc_leaf_size) *
342 	    sizeof(uint64_t) + roundup(sc->sc_root_size, BITS_PER_ENTRY);
343 	if (verbose)
344 		gctl_msg(req, 0, "Device %s created with memory map size %jd.",
345 		    gp->name, (intmax_t)sc->sc_writemap_memory);
346 	gctl_post_messages(req);
347 	G_UNION_DEBUG(1, "Device %s created with memory map size %jd.",
348 	    gp->name, (intmax_t)sc->sc_writemap_memory);
349 	return;
350 
351 fail4:
352 	g_detach(uppercp);
353 fail3:
354 	g_destroy_consumer(uppercp);
355 	g_access(lowercp, -1, 0, -1);
356 fail2:
357 	g_detach(lowercp);
358 fail1:
359 	g_destroy_consumer(lowercp);
360 	g_destroy_provider(newpp);
361 	g_destroy_geom(gp);
362 }
363 
364 /*
365  * Fetch named option and verify that it is positive.
366  */
367 static intmax_t
368 g_union_fetcharg(struct gctl_req *req, const char *name)
369 {
370 	intmax_t *val;
371 
372 	val = gctl_get_paraml_opt(req, name, sizeof(*val));
373 	if (val == NULL)
374 		return (0);
375 	if (*val >= 0)
376 		return (*val);
377 	gctl_msg(req, EINVAL, "Invalid '%s' (%jd): negative value, "
378 	    "using default.", name, *val);
379 	return (0);
380 }
381 
382 /*
383  * Verify that a name is alphanumeric.
384  */
385 static bool
386 g_union_verify_nprefix(const char *name)
387 {
388 	int i;
389 
390 	for (i = 0; i < strlen(name); i++) {
391 		if (isalpha(name[i]) == 0 && isdigit(name[i]) == 0) {
392 			return (false);
393 		}
394 	}
395 	return (true);
396 }
397 
398 /*
399  * Destroy a union device.
400  */
401 static void
402 g_union_ctl_destroy(struct gctl_req *req, struct g_class *mp, bool verbose)
403 {
404 	int *nargs, *force, error, i;
405 	struct g_geom *gp;
406 	const char *name;
407 	char param[16];
408 
409 	g_topology_assert();
410 
411 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
412 	if (nargs == NULL) {
413 		gctl_error(req, "No '%s' argument.", "nargs");
414 		return;
415 	}
416 	if (*nargs <= 0) {
417 		gctl_error(req, "Missing device(s).");
418 		return;
419 	}
420 	force = gctl_get_paraml(req, "force", sizeof(*force));
421 	if (force == NULL) {
422 		gctl_error(req, "No 'force' argument.");
423 		return;
424 	}
425 
426 	for (i = 0; i < *nargs; i++) {
427 		snprintf(param, sizeof(param), "arg%d", i);
428 		name = gctl_get_asciiparam(req, param);
429 		if (name == NULL) {
430 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
431 			continue;
432 		}
433 		if (strncmp(name, _PATH_DEV, strlen(_PATH_DEV)) == 0)
434 			name += strlen(_PATH_DEV);
435 		gp = g_union_find_geom(mp, name);
436 		if (gp == NULL) {
437 			gctl_msg(req, EINVAL, "Device %s is invalid.", name);
438 			continue;
439 		}
440 		error = g_union_destroy(verbose ? req : NULL, gp, *force);
441 		if (error != 0)
442 			gctl_msg(req, error, "Error %d: "
443 			    "cannot destroy device %s.", error, gp->name);
444 	}
445 	gctl_post_messages(req);
446 }
447 
448 /*
449  * Find a union geom.
450  */
451 static struct g_geom *
452 g_union_find_geom(struct g_class *mp, const char *name)
453 {
454 	struct g_geom *gp;
455 
456 	LIST_FOREACH(gp, &mp->geom, geom) {
457 		if (strcmp(gp->name, name) == 0)
458 			return (gp);
459 	}
460 	return (NULL);
461 }
462 
463 /*
464  * Zero out all the statistics associated with a union device.
465  */
466 static void
467 g_union_ctl_reset(struct gctl_req *req, struct g_class *mp, bool verbose)
468 {
469 	struct g_union_softc *sc;
470 	struct g_provider *pp;
471 	struct g_geom *gp;
472 	char param[16];
473 	int i, *nargs;
474 
475 	g_topology_assert();
476 
477 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
478 	if (nargs == NULL) {
479 		gctl_error(req, "No '%s' argument.", "nargs");
480 		return;
481 	}
482 	if (*nargs <= 0) {
483 		gctl_error(req, "Missing device(s).");
484 		return;
485 	}
486 
487 	for (i = 0; i < *nargs; i++) {
488 		snprintf(param, sizeof(param), "arg%d", i);
489 		pp = gctl_get_provider(req, param);
490 		if (pp == NULL) {
491 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
492 			continue;
493 		}
494 		gp = pp->geom;
495 		if (gp->class != mp) {
496 			gctl_msg(req, EINVAL, "Provider %s is invalid.",
497 			    pp->name);
498 			continue;
499 		}
500 		sc = gp->softc;
501 		sc->sc_reads = 0;
502 		sc->sc_writes = 0;
503 		sc->sc_deletes = 0;
504 		sc->sc_getattrs = 0;
505 		sc->sc_flushes = 0;
506 		sc->sc_speedups = 0;
507 		sc->sc_cmd0s = 0;
508 		sc->sc_cmd1s = 0;
509 		sc->sc_cmd2s = 0;
510 		sc->sc_readbytes = 0;
511 		sc->sc_wrotebytes = 0;
512 		if (verbose)
513 			gctl_msg(req, 0, "Device %s has been reset.", pp->name);
514 		G_UNION_DEBUG(1, "Device %s has been reset.", pp->name);
515 	}
516 	gctl_post_messages(req);
517 }
518 
519 /*
520  * Revert all write requests made to the top layer of the union.
521  */
522 static void
523 g_union_ctl_revert(struct gctl_req *req, struct g_class *mp, bool verbose)
524 {
525 	struct g_union_softc *sc;
526 	struct g_provider *pp;
527 	struct g_geom *gp;
528 	char param[16];
529 	int i, *nargs;
530 
531 	g_topology_assert();
532 
533 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
534 	if (nargs == NULL) {
535 		gctl_error(req, "No '%s' argument.", "nargs");
536 		return;
537 	}
538 	if (*nargs <= 0) {
539 		gctl_error(req, "Missing device(s).");
540 		return;
541 	}
542 
543 	for (i = 0; i < *nargs; i++) {
544 		snprintf(param, sizeof(param), "arg%d", i);
545 		pp = gctl_get_provider(req, param);
546 		if (pp == NULL) {
547 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
548 			continue;
549 		}
550 		gp = pp->geom;
551 		if (gp->class != mp) {
552 			gctl_msg(req, EINVAL, "Provider %s is invalid.",
553 			    pp->name);
554 			continue;
555 		}
556 		sc = gp->softc;
557 		if (g_union_get_writelock(sc) != 0) {
558 			gctl_msg(req, EINVAL, "Revert already in progress for "
559 			    "provider %s.", pp->name);
560 			continue;
561 		}
562 		/*
563 		 * No mount or other use of union is allowed.
564 		 */
565 		if (pp->acr > 0 || pp->acw > 0 || pp->ace > 0) {
566 			gctl_msg(req, EPERM, "Unable to get exclusive access "
567 			    "for reverting of %s;\n\t%s cannot be mounted or "
568 			    "otherwise open during a revert.",
569 			     pp->name, pp->name);
570 			g_union_rel_writelock(sc);
571 			continue;
572 		}
573 		g_union_revert(sc);
574 		g_union_rel_writelock(sc);
575 		if (verbose)
576 			gctl_msg(req, 0, "Device %s has been reverted.",
577 			    pp->name);
578 		G_UNION_DEBUG(1, "Device %s has been reverted.", pp->name);
579 	}
580 	gctl_post_messages(req);
581 }
582 
583 /*
584  * Revert union writes by zero'ing out the writemap.
585  */
586 static void
587 g_union_revert(struct g_union_softc *sc)
588 {
589 	int i;
590 
591 	G_WLOCK(sc);
592 	for (i = 0; i < sc->sc_root_size; i++)
593 		memset(sc->sc_writemap_root[i], 0,
594 		    sc->sc_leaf_size * sizeof(uint64_t));
595 	memset(sc->sc_leafused, 0, roundup(sc->sc_root_size, BITS_PER_ENTRY));
596 	G_WUNLOCK(sc);
597 }
598 
599 /*
600  * Commit all the writes made in the top layer to the lower layer.
601  */
602 static void
603 g_union_ctl_commit(struct gctl_req *req, struct g_class *mp, bool verbose)
604 {
605 	struct g_union_softc *sc;
606 	struct g_provider *pp, *lowerpp;
607 	struct g_consumer *lowercp;
608 	struct g_geom *gp;
609 	struct bio *bp;
610 	char param[16];
611 	off_t len2rd, len2wt, savelen;
612 	int i, error, error1, *nargs, *force, *reboot;
613 
614 	g_topology_assert();
615 
616 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
617 	if (nargs == NULL) {
618 		gctl_error(req, "No '%s' argument.", "nargs");
619 		return;
620 	}
621 	if (*nargs <= 0) {
622 		gctl_error(req, "Missing device(s).");
623 		return;
624 	}
625 	force = gctl_get_paraml(req, "force", sizeof(*force));
626 	if (force == NULL) {
627 		gctl_error(req, "No 'force' argument.");
628 		return;
629 	}
630 	reboot = gctl_get_paraml(req, "reboot", sizeof(*reboot));
631 	if (reboot == NULL) {
632 		gctl_error(req, "No 'reboot' argument.");
633 		return;
634 	}
635 
636 	/* Get a bio buffer to do our I/O */
637 	bp = g_alloc_bio();
638 	bp->bio_data = g_malloc(MAXBSIZE, M_WAITOK);
639 	bp->bio_done = biodone;
640 	for (i = 0; i < *nargs; i++) {
641 		snprintf(param, sizeof(param), "arg%d", i);
642 		pp = gctl_get_provider(req, param);
643 		if (pp == NULL) {
644 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
645 			continue;
646 		}
647 		gp = pp->geom;
648 		if (gp->class != mp) {
649 			gctl_msg(req, EINVAL, "Provider %s is invalid.",
650 			    pp->name);
651 			continue;
652 		}
653 		sc = gp->softc;
654 		if (g_union_get_writelock(sc) != 0) {
655 			gctl_msg(req, EINVAL, "Commit already in progress for "
656 			    "provider %s.", pp->name);
657 			continue;
658 		}
659 
660 		/* upgrade to write access for lower */
661 		lowercp = sc->sc_lowercp;
662 		lowerpp = lowercp->provider;
663 		/*
664 		 * No mount or other use of union is allowed, unless the
665 		 * -f flag is given which allows read-only mount or usage.
666 		 */
667 		if ((*force == false && pp->acr > 0) || pp->acw > 0 ||
668 		     pp->ace > 0) {
669 			gctl_msg(req, EPERM, "Unable to get exclusive access "
670 			    "for writing of %s.\n\tNote that %s cannot be "
671 			    "mounted or otherwise\n\topen during a commit "
672 			    "unless the -f flag is used.", pp->name, pp->name);
673 			g_union_rel_writelock(sc);
674 			continue;
675 		}
676 		/*
677 		 * No mount or other use of lower media is allowed, unless the
678 		 * -f flag is given which allows read-only mount or usage.
679 		 */
680 		if ((*force == false && lowerpp->acr > lowercp->acr) ||
681 		     lowerpp->acw > lowercp->acw ||
682 		     lowerpp->ace > lowercp->ace) {
683 			gctl_msg(req, EPERM, "provider %s is unable to get "
684 			    "exclusive access to %s\n\tfor writing. Note that "
685 			    "%s cannot be mounted or otherwise open\n\tduring "
686 			    "a commit unless the -f flag is used.", pp->name,
687 			    lowerpp->name, lowerpp->name);
688 			g_union_rel_writelock(sc);
689 			continue;
690 		}
691 		if ((error = g_access(lowercp, 0, 1, 0)) != 0) {
692 			gctl_msg(req, error, "Error %d: provider %s is unable "
693 			    "to access %s for writing.", error, pp->name,
694 			    lowerpp->name);
695 			g_union_rel_writelock(sc);
696 			continue;
697 		}
698 		g_topology_unlock();
699 		/* Loop over write map copying across written blocks */
700 		bp->bio_offset = 0;
701 		bp->bio_length = sc->sc_map_size * sc->sc_sectorsize;
702 		G_RLOCK(sc);
703 		error = 0;
704 		while (bp->bio_length > 0) {
705 			if (!g_union_getmap(bp, sc, &len2rd)) {
706 				/* not written, so skip */
707 				bp->bio_offset += len2rd;
708 				bp->bio_length -= len2rd;
709 				continue;
710 			}
711 			G_RUNLOCK(sc);
712 			/* need to read then write len2rd sectors */
713 			for ( ; len2rd > 0; len2rd -= len2wt) {
714 				/* limit ourselves to MAXBSIZE size I/Os */
715 				len2wt = len2rd;
716 				if (len2wt > MAXBSIZE)
717 					len2wt = MAXBSIZE;
718 				savelen = bp->bio_length;
719 				bp->bio_length = len2wt;
720 				bp->bio_cmd = BIO_READ;
721 				g_io_request(bp, sc->sc_uppercp);
722 				if ((error = biowait(bp, "rdunion")) != 0) {
723 					gctl_msg(req, error, "Commit read "
724 					    "error %d in provider %s, commit "
725 					    "aborted.", error, pp->name);
726 					goto cleanup;
727 				}
728 				bp->bio_flags &= ~BIO_DONE;
729 				bp->bio_cmd = BIO_WRITE;
730 				g_io_request(bp, lowercp);
731 				if ((error = biowait(bp, "wtunion")) != 0) {
732 					gctl_msg(req, error, "Commit write "
733 					    "error %d in provider %s, commit "
734 					    "aborted.", error, pp->name);
735 					goto cleanup;
736 				}
737 				bp->bio_flags &= ~BIO_DONE;
738 				bp->bio_offset += len2wt;
739 				bp->bio_length = savelen - len2wt;
740 			}
741 			G_RLOCK(sc);
742 		}
743 		G_RUNLOCK(sc);
744 		/* clear the write map */
745 		g_union_revert(sc);
746 cleanup:
747 		g_topology_lock();
748 		/* return lower to previous access */
749 		if ((error1 = g_access(lowercp, 0, -1, 0)) != 0) {
750 			G_UNION_DEBUG(2, "Error %d: device %s could not reset "
751 			    "access to %s (r=0 w=-1 e=0).", error1, pp->name,
752 			    lowerpp->name);
753 		}
754 		g_union_rel_writelock(sc);
755 		if (error == 0 && verbose)
756 			gctl_msg(req, 0, "Device %s has been committed.",
757 			    pp->name);
758 		G_UNION_DEBUG(1, "Device %s has been committed.", pp->name);
759 	}
760 	gctl_post_messages(req);
761 	g_free(bp->bio_data);
762 	g_destroy_bio(bp);
763 	if (*reboot)
764 		kern_reboot(RB_AUTOBOOT);
765 }
766 
767 /*
768  * Generally allow access unless a commit is in progress.
769  */
770 static int
771 g_union_access(struct g_provider *pp, int r, int w, int e)
772 {
773 	struct g_union_softc *sc;
774 
775 	sc = pp->geom->softc;
776 	if (sc == NULL) {
777 		if (r <= 0 && w <= 0 && e <= 0)
778 			return (0);
779 		return (ENXIO);
780 	}
781 	r += pp->acr;
782 	w += pp->acw;
783 	e += pp->ace;
784 	if (g_union_get_writelock(sc) != 0) {
785 		if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0)
786 			return (0);
787 		return (EBUSY);
788 	}
789 	g_union_rel_writelock(sc);
790 	return (0);
791 }
792 
793 /*
794  * Initiate an I/O operation on the union device.
795  */
796 static void
797 g_union_start(struct bio *bp)
798 {
799 	struct g_union_softc *sc;
800 	struct g_union_wip *wip;
801 	struct bio *cbp;
802 
803 	sc = bp->bio_to->geom->softc;
804 	if (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE) {
805 		wip = g_malloc(sizeof(*wip), M_NOWAIT);
806 		if (wip == NULL) {
807 			g_io_deliver(bp, ENOMEM);
808 			return;
809 		}
810 		TAILQ_INIT(&wip->wip_waiting);
811 		wip->wip_bp = bp;
812 		wip->wip_sc = sc;
813 		wip->wip_start = bp->bio_offset + sc->sc_offset;
814 		wip->wip_end = wip->wip_start + bp->bio_length - 1;
815 		wip->wip_numios = 1;
816 		wip->wip_error = 0;
817 		g_union_doio(wip);
818 		return;
819 	}
820 
821 	/*
822 	 * All commands other than read and write are passed through to
823 	 * the upper-level device since it is writable and thus able to
824 	 * respond to delete, flush, and speedup requests.
825 	 */
826 	cbp = g_clone_bio(bp);
827 	if (cbp == NULL) {
828 		g_io_deliver(bp, ENOMEM);
829 		return;
830 	}
831 	cbp->bio_offset = bp->bio_offset + sc->sc_offset;
832 	cbp->bio_done = g_std_done;
833 
834 	switch (cbp->bio_cmd) {
835 	case BIO_DELETE:
836 		G_UNION_LOGREQ(cbp, "Delete request received.");
837 		atomic_add_long(&sc->sc_deletes, 1);
838 		break;
839 	case BIO_GETATTR:
840 		G_UNION_LOGREQ(cbp, "Getattr request received.");
841 		atomic_add_long(&sc->sc_getattrs, 1);
842 		if (strcmp(cbp->bio_attribute, "GEOM::kerneldump") != 0)
843 			/* forward the GETATTR to the lower-level device */
844 			break;
845 		g_union_kerneldump(bp, sc);
846 		return;
847 	case BIO_FLUSH:
848 		G_UNION_LOGREQ(cbp, "Flush request received.");
849 		atomic_add_long(&sc->sc_flushes, 1);
850 		break;
851 	case BIO_SPEEDUP:
852 		G_UNION_LOGREQ(cbp, "Speedup request received.");
853 		atomic_add_long(&sc->sc_speedups, 1);
854 		break;
855 	case BIO_CMD0:
856 		G_UNION_LOGREQ(cbp, "Cmd0 request received.");
857 		atomic_add_long(&sc->sc_cmd0s, 1);
858 		break;
859 	case BIO_CMD1:
860 		G_UNION_LOGREQ(cbp, "Cmd1 request received.");
861 		atomic_add_long(&sc->sc_cmd1s, 1);
862 		break;
863 	case BIO_CMD2:
864 		G_UNION_LOGREQ(cbp, "Cmd2 request received.");
865 		atomic_add_long(&sc->sc_cmd2s, 1);
866 		break;
867 	default:
868 		G_UNION_LOGREQ(cbp, "Unknown (%d) request received.",
869 		    cbp->bio_cmd);
870 		break;
871 	}
872 	g_io_request(cbp, sc->sc_uppercp);
873 }
874 
875 /*
876  * Initiate a read or write operation on the union device.
877  */
878 static void
879 g_union_doio(struct g_union_wip *wip)
880 {
881 	struct g_union_softc *sc;
882 	struct g_consumer *cp, *firstcp;
883 	struct g_union_wip *activewip;
884 	struct bio *cbp, *firstbp;
885 	off_t rdlen, len2rd, offset;
886 	int iocnt, needstoblock;
887 	char *level;
888 
889 	/*
890 	 * To maintain consistency, we cannot allow concurrent reads
891 	 * or writes to the same block.
892 	 *
893 	 * A work-in-progress (wip) structure is allocated for each
894 	 * read or write request. All active requests are kept on the
895 	 * softc sc_wiplist. As each request arrives, it is checked to
896 	 * see if it overlaps any of the active entries. If it does not
897 	 * overlap, then it is added to the active list and initiated.
898 	 * If it does overlap an active entry, it is added to the
899 	 * wip_waiting list for the active entry that it overlaps.
900 	 * When an active entry completes, it restarts all the requests
901 	 * on its wip_waiting list.
902 	 */
903 	sc = wip->wip_sc;
904 	G_WLOCK(sc);
905 	TAILQ_FOREACH(activewip, &sc->sc_wiplist, wip_next) {
906 		if (wip->wip_end < activewip->wip_start ||
907 		    wip->wip_start > activewip->wip_end)
908 			continue;
909 		needstoblock = 1;
910 		if (wip->wip_bp->bio_cmd == BIO_WRITE)
911 			if (activewip->wip_bp->bio_cmd == BIO_WRITE)
912 				sc->sc_writeblockwrite += 1;
913 			else
914 				sc->sc_readblockwrite += 1;
915 		else
916 			if (activewip->wip_bp->bio_cmd == BIO_WRITE)
917 				sc->sc_writeblockread += 1;
918 			else {
919 				sc->sc_readcurrentread += 1;
920 				needstoblock = 0;
921 			}
922 		/* Put request on a waiting list if necessary */
923 		if (needstoblock) {
924 			TAILQ_INSERT_TAIL(&activewip->wip_waiting, wip,
925 			    wip_next);
926 			G_WUNLOCK(sc);
927 			return;
928 		}
929 	}
930 	/* Put request on the active list */
931 	TAILQ_INSERT_TAIL(&sc->sc_wiplist, wip, wip_next);
932 
933 	/*
934 	 * Process I/O requests that have been cleared to go.
935 	 */
936 	cbp = g_clone_bio(wip->wip_bp);
937 	if (cbp == NULL) {
938 		TAILQ_REMOVE(&sc->sc_wiplist, wip, wip_next);
939 		G_WUNLOCK(sc);
940 		KASSERT(TAILQ_FIRST(&wip->wip_waiting) == NULL,
941 		    ("g_union_doio: non-empty work-in-progress waiting queue"));
942 		g_io_deliver(wip->wip_bp, ENOMEM);
943 		g_free(wip);
944 		return;
945 	}
946 	G_WUNLOCK(sc);
947 	cbp->bio_caller1 = wip;
948 	cbp->bio_done = g_union_done;
949 	cbp->bio_offset = wip->wip_start;
950 
951 	/*
952 	 * Writes are always done to the top level. The blocks that
953 	 * are written are recorded in the bitmap when the I/O completes.
954 	 */
955 	if (cbp->bio_cmd == BIO_WRITE) {
956 		G_UNION_LOGREQ(cbp, "Sending %jd byte write request to upper "
957 		    "level.", cbp->bio_length);
958 		atomic_add_long(&sc->sc_writes, 1);
959 		atomic_add_long(&sc->sc_wrotebytes, cbp->bio_length);
960 		g_io_request(cbp, sc->sc_uppercp);
961 		return;
962 	}
963 	/*
964 	 * The usual read case is that we either read the top layer
965 	 * if the block has been previously written or the bottom layer
966 	 * if it has not been written. However, it is possible that
967 	 * only part of the block has been written, For example we may
968 	 * have written a UFS/FFS file fragment comprising several
969 	 * sectors out of an 8-sector block.  Here, if the entire
970 	 * 8-sector block is read for example by a snapshot needing
971 	 * to copy the full block, then we need to read the written
972 	 * sectors from the upper level and the unwritten sectors from
973 	 * the lower level. We do this by alternately reading from the
974 	 * top and bottom layers until we complete the read. We
975 	 * simplify for the common case to just do the I/O and return.
976 	 */
977 	atomic_add_long(&sc->sc_reads, 1);
978 	atomic_add_long(&sc->sc_readbytes, cbp->bio_length);
979 	rdlen = cbp->bio_length;
980 	offset = 0;
981 	for (iocnt = 0; ; iocnt++) {
982 		if (g_union_getmap(cbp, sc, &len2rd)) {
983 			/* read top */
984 			cp = sc->sc_uppercp;
985 			level = "upper";
986 		} else {
987 			/* read bottom */
988 			cp = sc->sc_lowercp;
989 			level = "lower";
990 		}
991 		/* Check if only a single read is required */
992 		if (iocnt == 0 && rdlen == len2rd) {
993 			G_UNION_LOGREQLVL((cp == sc->sc_uppercp) ?
994 			    3 : 4, cbp, "Sending %jd byte read "
995 			    "request to %s level.", len2rd, level);
996 			g_io_request(cbp, cp);
997 			return;
998 		}
999 		cbp->bio_length = len2rd;
1000 		if ((cbp->bio_flags & BIO_UNMAPPED) != 0)
1001 			cbp->bio_ma_offset += offset;
1002 		else
1003 			cbp->bio_data += offset;
1004 		offset += len2rd;
1005 		rdlen -= len2rd;
1006 		G_UNION_LOGREQLVL(3, cbp, "Sending %jd byte read "
1007 		    "request to %s level.", len2rd, level);
1008 		/*
1009 		 * To avoid prematurely notifying our consumer
1010 		 * that their I/O has completed, we have to delay
1011 		 * issuing our first I/O request until we have
1012 		 * issued all the additional I/O requests.
1013 		 */
1014 		if (iocnt > 0) {
1015 			atomic_add_long(&wip->wip_numios, 1);
1016 			g_io_request(cbp, cp);
1017 		} else {
1018 			firstbp = cbp;
1019 			firstcp = cp;
1020 		}
1021 		if (rdlen == 0)
1022 			break;
1023 		/* set up for next read */
1024 		cbp = g_clone_bio(wip->wip_bp);
1025 		if (cbp == NULL) {
1026 			wip->wip_error = ENOMEM;
1027 			atomic_add_long(&wip->wip_numios, -1);
1028 			break;
1029 		}
1030 		cbp->bio_caller1 = wip;
1031 		cbp->bio_done = g_union_done;
1032 		cbp->bio_offset += offset;
1033 		cbp->bio_length = rdlen;
1034 		atomic_add_long(&sc->sc_reads, 1);
1035 	}
1036 	/* We have issued all our I/O, so start the first one */
1037 	g_io_request(firstbp, firstcp);
1038 	return;
1039 }
1040 
1041 /*
1042  * Used when completing a union I/O operation.
1043  */
1044 static void
1045 g_union_done(struct bio *bp)
1046 {
1047 	struct g_union_wip *wip, *waitingwip;
1048 	struct g_union_softc *sc;
1049 
1050 	wip = bp->bio_caller1;
1051 	if (wip->wip_error != 0 && bp->bio_error == 0)
1052 		bp->bio_error = wip->wip_error;
1053 	wip->wip_error = 0;
1054 	if (atomic_fetchadd_long(&wip->wip_numios, -1) == 1) {
1055 		sc = wip->wip_sc;
1056 		G_WLOCK(sc);
1057 		if (bp->bio_cmd == BIO_WRITE)
1058 			g_union_setmap(bp, sc);
1059 		TAILQ_REMOVE(&sc->sc_wiplist, wip, wip_next);
1060 		G_WUNLOCK(sc);
1061 		while ((waitingwip = TAILQ_FIRST(&wip->wip_waiting)) != NULL) {
1062 			TAILQ_REMOVE(&wip->wip_waiting, waitingwip, wip_next);
1063 			g_union_doio(waitingwip);
1064 		}
1065 		g_free(wip);
1066 	}
1067 	g_std_done(bp);
1068 }
1069 
1070 /*
1071  * Record blocks that have been written in the map.
1072  */
1073 static void
1074 g_union_setmap(struct bio *bp, struct g_union_softc *sc)
1075 {
1076 	size_t root_idx;
1077 	uint64_t **leaf;
1078 	uint64_t *wordp;
1079 	off_t start, numsec;
1080 
1081 	G_WLOCKOWNED(sc);
1082 	KASSERT(bp->bio_offset % sc->sc_sectorsize == 0,
1083 	    ("g_union_setmap: offset not on sector boundry"));
1084 	KASSERT(bp->bio_length % sc->sc_sectorsize == 0,
1085 	    ("g_union_setmap: length not a multiple of sectors"));
1086 	start = bp->bio_offset / sc->sc_sectorsize;
1087 	numsec = bp->bio_length / sc->sc_sectorsize;
1088 	KASSERT(start + numsec <= sc->sc_map_size,
1089 	    ("g_union_setmap: block %jd is out of range", start + numsec));
1090 	for ( ; numsec > 0; numsec--, start++) {
1091 		root_idx = start / sc->sc_bits_per_leaf;
1092 		leaf = &sc->sc_writemap_root[root_idx];
1093 		wordp = &(*leaf)
1094 		    [(start % sc->sc_bits_per_leaf) / BITS_PER_ENTRY];
1095 		*wordp |= 1ULL << (start % BITS_PER_ENTRY);
1096 		sc->sc_leafused[root_idx / BITS_PER_ENTRY] |=
1097 		    1ULL << (root_idx % BITS_PER_ENTRY);
1098 	}
1099 }
1100 
1101 /*
1102  * Check map to determine whether blocks have been written.
1103  *
1104  * Return true if they have been written so should be read from the top
1105  * layer. Return false if they have not been written so should be read
1106  * from the bottom layer. Return in len2read the bytes to be read. See
1107  * the comment above the BIO_READ implementation in g_union_start() for
1108  * an explantion of why len2read may be shorter than the buffer length.
1109  */
1110 static bool
1111 g_union_getmap(struct bio *bp, struct g_union_softc *sc, off_t *len2read)
1112 {
1113 	off_t start, numsec, leafresid, bitloc;
1114 	bool first, maptype, retval;
1115 	uint64_t *leaf, word;
1116 	size_t root_idx;
1117 
1118 	KASSERT(bp->bio_offset % sc->sc_sectorsize == 0,
1119 	    ("g_union_getmap: offset not on sector boundry"));
1120 	KASSERT(bp->bio_length % sc->sc_sectorsize == 0,
1121 	    ("g_union_getmap: length not a multiple of sectors"));
1122 	start = bp->bio_offset / sc->sc_sectorsize;
1123 	numsec = bp->bio_length / sc->sc_sectorsize;
1124 	G_UNION_DEBUG(4, "g_union_getmap: check %jd sectors starting at %jd\n",
1125 	    numsec, start);
1126 	KASSERT(start + numsec <= sc->sc_map_size,
1127 	    ("g_union_getmap: block %jd is out of range", start + numsec));
1128 		root_idx = start / sc->sc_bits_per_leaf;
1129 	first = true;
1130 	maptype = false;
1131 	while (numsec > 0) {
1132 		/* Check first if the leaf records any written sectors */
1133 		root_idx = start / sc->sc_bits_per_leaf;
1134 		leafresid = sc->sc_bits_per_leaf -
1135 		    (start % sc->sc_bits_per_leaf);
1136 		if (((sc->sc_leafused[root_idx / BITS_PER_ENTRY]) &
1137 		    (1ULL << (root_idx % BITS_PER_ENTRY))) == 0) {
1138 			if (first) {
1139 				maptype = false;
1140 				first = false;
1141 			}
1142 			if (maptype)
1143 				break;
1144 			numsec -= leafresid;
1145 			start += leafresid;
1146 			continue;
1147 		}
1148 		/* Check up to a word boundry, then check word by word */
1149 		leaf = sc->sc_writemap_root[root_idx];
1150 		word = leaf[(start % sc->sc_bits_per_leaf) / BITS_PER_ENTRY];
1151 		bitloc = start % BITS_PER_ENTRY;
1152 		if (bitloc == 0 && (word == 0 || word == ~0)) {
1153 			if (first) {
1154 				if (word == 0)
1155 					maptype = false;
1156 				else
1157 					maptype = true;
1158 				first = false;
1159 			}
1160 			if ((word == 0 && maptype) ||
1161 			    (word == ~0 && !maptype))
1162 				break;
1163 			numsec -= BITS_PER_ENTRY;
1164 			start += BITS_PER_ENTRY;
1165 			continue;
1166 		}
1167 		for ( ; bitloc < BITS_PER_ENTRY; bitloc ++) {
1168 			retval = (word & (1ULL << bitloc)) != 0;
1169 			if (first) {
1170 				maptype = retval;
1171 				first = false;
1172 			}
1173 			if (maptype == retval) {
1174 				numsec--;
1175 				start++;
1176 				continue;
1177 			}
1178 			goto out;
1179 		}
1180 	}
1181 out:
1182 	if (numsec < 0) {
1183 		start += numsec;
1184 		numsec = 0;
1185 	}
1186 	*len2read = bp->bio_length - (numsec * sc->sc_sectorsize);
1187 	G_UNION_DEBUG(maptype ? 3 : 4,
1188 	    "g_union_getmap: return maptype %swritten for %jd "
1189 	    "sectors ending at %jd\n", maptype ? "" : "NOT ",
1190 	    *len2read / sc->sc_sectorsize, start - 1);
1191 	return (maptype);
1192 }
1193 
1194 /*
1195  * Fill in details for a BIO_GETATTR request.
1196  */
1197 static void
1198 g_union_kerneldump(struct bio *bp, struct g_union_softc *sc)
1199 {
1200 	struct g_kerneldump *gkd;
1201 	struct g_geom *gp;
1202 	struct g_provider *pp;
1203 
1204 	gkd = (struct g_kerneldump *)bp->bio_data;
1205 	gp = bp->bio_to->geom;
1206 	g_trace(G_T_TOPOLOGY, "%s(%s, %jd, %jd)", __func__, gp->name,
1207 	    (intmax_t)gkd->offset, (intmax_t)gkd->length);
1208 
1209 	pp = LIST_FIRST(&gp->provider);
1210 
1211 	gkd->di.dumper = g_union_dumper;
1212 	gkd->di.priv = sc;
1213 	gkd->di.blocksize = pp->sectorsize;
1214 	gkd->di.maxiosize = DFLTPHYS;
1215 	gkd->di.mediaoffset = sc->sc_offset + gkd->offset;
1216 	if (gkd->offset > sc->sc_size) {
1217 		g_io_deliver(bp, ENODEV);
1218 		return;
1219 	}
1220 	if (gkd->offset + gkd->length > sc->sc_size)
1221 		gkd->length = sc->sc_size - gkd->offset;
1222 	gkd->di.mediasize = gkd->length;
1223 	g_io_deliver(bp, 0);
1224 }
1225 
1226 /*
1227  * Handler for g_union_kerneldump().
1228  */
1229 static int
1230 g_union_dumper(void *priv, void *virtual, off_t offset, size_t length)
1231 {
1232 
1233 	return (0);
1234 }
1235 
1236 /*
1237  * List union statistics.
1238  */
1239 static void
1240 g_union_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1241     struct g_consumer *cp, struct g_provider *pp)
1242 {
1243 	struct g_union_softc *sc;
1244 
1245 	if (pp != NULL || cp != NULL || gp->softc == NULL)
1246 		return;
1247 	sc = gp->softc;
1248 	sbuf_printf(sb, "%s<Reads>%ju</Reads>\n", indent,
1249 	    (uintmax_t)sc->sc_reads);
1250 	sbuf_printf(sb, "%s<Writes>%ju</Writes>\n", indent,
1251 	    (uintmax_t)sc->sc_writes);
1252 	sbuf_printf(sb, "%s<Deletes>%ju</Deletes>\n", indent,
1253 	    (uintmax_t)sc->sc_deletes);
1254 	sbuf_printf(sb, "%s<Getattrs>%ju</Getattrs>\n", indent,
1255 	    (uintmax_t)sc->sc_getattrs);
1256 	sbuf_printf(sb, "%s<Flushes>%ju</Flushes>\n", indent,
1257 	    (uintmax_t)sc->sc_flushes);
1258 	sbuf_printf(sb, "%s<Speedups>%ju</Speedups>\n", indent,
1259 	    (uintmax_t)sc->sc_speedups);
1260 	sbuf_printf(sb, "%s<Cmd0s>%ju</Cmd0s>\n", indent,
1261 	    (uintmax_t)sc->sc_cmd0s);
1262 	sbuf_printf(sb, "%s<Cmd1s>%ju</Cmd1s>\n", indent,
1263 	    (uintmax_t)sc->sc_cmd1s);
1264 	sbuf_printf(sb, "%s<Cmd2s>%ju</Cmd2s>\n", indent,
1265 	    (uintmax_t)sc->sc_cmd2s);
1266 	sbuf_printf(sb, "%s<ReadCurrentRead>%ju</ReadCurrentRead>\n", indent,
1267 	    (uintmax_t)sc->sc_readcurrentread);
1268 	sbuf_printf(sb, "%s<ReadBlockWrite>%ju</ReadBlockWrite>\n", indent,
1269 	    (uintmax_t)sc->sc_readblockwrite);
1270 	sbuf_printf(sb, "%s<WriteBlockRead>%ju</WriteBlockRead>\n", indent,
1271 	    (uintmax_t)sc->sc_writeblockread);
1272 	sbuf_printf(sb, "%s<WriteBlockWrite>%ju</WriteBlockWrite>\n", indent,
1273 	    (uintmax_t)sc->sc_writeblockwrite);
1274 	sbuf_printf(sb, "%s<ReadBytes>%ju</ReadBytes>\n", indent,
1275 	    (uintmax_t)sc->sc_readbytes);
1276 	sbuf_printf(sb, "%s<WroteBytes>%ju</WroteBytes>\n", indent,
1277 	    (uintmax_t)sc->sc_wrotebytes);
1278 	sbuf_printf(sb, "%s<Offset>%jd</Offset>\n", indent,
1279 	    (intmax_t)sc->sc_offset);
1280 }
1281 
1282 /*
1283  * Clean up an orphaned geom.
1284  */
1285 static void
1286 g_union_orphan(struct g_consumer *cp)
1287 {
1288 
1289 	g_topology_assert();
1290 	g_union_destroy(NULL, cp->geom, true);
1291 }
1292 
1293 /*
1294  * Clean up a union geom.
1295  */
1296 static int
1297 g_union_destroy_geom(struct gctl_req *req, struct g_class *mp,
1298     struct g_geom *gp)
1299 {
1300 
1301 	return (g_union_destroy(NULL, gp, false));
1302 }
1303 
1304 /*
1305  * Clean up a union device.
1306  */
1307 static int
1308 g_union_destroy(struct gctl_req *req, struct g_geom *gp, bool force)
1309 {
1310 	struct g_union_softc *sc;
1311 	struct g_provider *pp;
1312 	int error;
1313 
1314 	g_topology_assert();
1315 	sc = gp->softc;
1316 	if (sc == NULL)
1317 		return (ENXIO);
1318 	pp = LIST_FIRST(&gp->provider);
1319 	if ((sc->sc_flags & DOING_COMMIT) != 0 ||
1320 	    (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0))) {
1321 		if (force) {
1322 			if (req != NULL)
1323 				gctl_msg(req, 0, "Device %s is still in use, "
1324 				    "so is being forcibly removed.", gp->name);
1325 			G_UNION_DEBUG(1, "Device %s is still in use, so "
1326 			    "is being forcibly removed.", gp->name);
1327 		} else {
1328 			if (req != NULL)
1329 				gctl_msg(req, EBUSY, "Device %s is still open "
1330 				    "(r=%d w=%d e=%d).", gp->name, pp->acr,
1331 				    pp->acw, pp->ace);
1332 			G_UNION_DEBUG(1, "Device %s is still open "
1333 			    "(r=%d w=%d e=%d).", gp->name, pp->acr,
1334 			    pp->acw, pp->ace);
1335 			return (EBUSY);
1336 		}
1337 	} else {
1338 		if (req != NULL)
1339 			gctl_msg(req, 0, "Device %s removed.", gp->name);
1340 		G_UNION_DEBUG(1, "Device %s removed.", gp->name);
1341 	}
1342 	/* Close consumers */
1343 	if ((error = g_access(sc->sc_lowercp, -1, 0, -1)) != 0)
1344 		G_UNION_DEBUG(2, "Error %d: device %s could not reset access "
1345 		    "to %s.", error, gp->name, sc->sc_lowercp->provider->name);
1346 	if ((error = g_access(sc->sc_uppercp, -1, -1, -1)) != 0)
1347 		G_UNION_DEBUG(2, "Error %d: device %s could not reset access "
1348 		    "to %s.", error, gp->name, sc->sc_uppercp->provider->name);
1349 
1350 	g_wither_geom(gp, ENXIO);
1351 
1352 	return (0);
1353 }
1354 
1355 /*
1356  * Clean up a union provider.
1357  */
1358 static void
1359 g_union_providergone(struct g_provider *pp)
1360 {
1361 	struct g_geom *gp;
1362 	struct g_union_softc *sc;
1363 	size_t i;
1364 
1365 	gp = pp->geom;
1366 	sc = gp->softc;
1367 	gp->softc = NULL;
1368 	for (i = 0; i < sc->sc_root_size; i++)
1369 		g_free(sc->sc_writemap_root[i]);
1370 	g_free(sc->sc_writemap_root);
1371 	g_free(sc->sc_leafused);
1372 	rw_destroy(&sc->sc_rwlock);
1373 	g_free(sc);
1374 }
1375 
1376 /*
1377  * Respond to a resized provider.
1378  */
1379 static void
1380 g_union_resize(struct g_consumer *cp)
1381 {
1382 	struct g_union_softc *sc;
1383 	struct g_geom *gp;
1384 
1385 	g_topology_assert();
1386 
1387 	gp = cp->geom;
1388 	sc = gp->softc;
1389 
1390 	/*
1391 	 * If size has gotten bigger, ignore it and just keep using
1392 	 * the space we already had. Otherwise we are done.
1393 	 */
1394 	if (sc->sc_size < cp->provider->mediasize - sc->sc_offset)
1395 		return;
1396 	g_union_destroy(NULL, gp, true);
1397 }
1398 
1399 DECLARE_GEOM_CLASS(g_union_class, g_union);
1400 MODULE_VERSION(geom_union, 0);
1401