xref: /freebsd/sys/geom/union/g_union.c (revision ecca34fa969579e53e4940ef9d7a7e170e872e77)
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
g_union_config(struct gctl_req * req,struct g_class * mp,const char * verb)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
g_union_ctl_create(struct gctl_req * req,struct g_class * mp,bool verbose)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_geom(mp, 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 	rw_destroy(&sc->sc_rwlock);
362 	g_free(sc);
363 	g_destroy_geom(gp);
364 }
365 
366 /*
367  * Fetch named option and verify that it is positive.
368  */
369 static intmax_t
g_union_fetcharg(struct gctl_req * req,const char * name)370 g_union_fetcharg(struct gctl_req *req, const char *name)
371 {
372 	intmax_t *val;
373 
374 	val = gctl_get_paraml_opt(req, name, sizeof(*val));
375 	if (val == NULL)
376 		return (0);
377 	if (*val >= 0)
378 		return (*val);
379 	gctl_msg(req, EINVAL, "Invalid '%s' (%jd): negative value, "
380 	    "using default.", name, *val);
381 	return (0);
382 }
383 
384 /*
385  * Verify that a name is alphanumeric.
386  */
387 static bool
g_union_verify_nprefix(const char * name)388 g_union_verify_nprefix(const char *name)
389 {
390 	int i;
391 
392 	for (i = 0; i < strlen(name); i++) {
393 		if (isalpha(name[i]) == 0 && isdigit(name[i]) == 0) {
394 			return (false);
395 		}
396 	}
397 	return (true);
398 }
399 
400 /*
401  * Destroy a union device.
402  */
403 static void
g_union_ctl_destroy(struct gctl_req * req,struct g_class * mp,bool verbose)404 g_union_ctl_destroy(struct gctl_req *req, struct g_class *mp, bool verbose)
405 {
406 	int *nargs, *force, error, i;
407 	struct g_geom *gp;
408 	const char *name;
409 	char param[16];
410 
411 	g_topology_assert();
412 
413 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
414 	if (nargs == NULL) {
415 		gctl_error(req, "No '%s' argument.", "nargs");
416 		return;
417 	}
418 	if (*nargs <= 0) {
419 		gctl_error(req, "Missing device(s).");
420 		return;
421 	}
422 	force = gctl_get_paraml(req, "force", sizeof(*force));
423 	if (force == NULL) {
424 		gctl_error(req, "No 'force' argument.");
425 		return;
426 	}
427 
428 	for (i = 0; i < *nargs; i++) {
429 		snprintf(param, sizeof(param), "arg%d", i);
430 		name = gctl_get_asciiparam(req, param);
431 		if (name == NULL) {
432 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
433 			continue;
434 		}
435 		if (strncmp(name, _PATH_DEV, strlen(_PATH_DEV)) == 0)
436 			name += strlen(_PATH_DEV);
437 		gp = g_union_find_geom(mp, name);
438 		if (gp == NULL) {
439 			gctl_msg(req, EINVAL, "Device %s is invalid.", name);
440 			continue;
441 		}
442 		error = g_union_destroy(verbose ? req : NULL, gp, *force);
443 		if (error != 0)
444 			gctl_msg(req, error, "Error %d: "
445 			    "cannot destroy device %s.", error, gp->name);
446 	}
447 	gctl_post_messages(req);
448 }
449 
450 /*
451  * Find a union geom.
452  */
453 static struct g_geom *
g_union_find_geom(struct g_class * mp,const char * name)454 g_union_find_geom(struct g_class *mp, const char *name)
455 {
456 	struct g_geom *gp;
457 
458 	LIST_FOREACH(gp, &mp->geom, geom) {
459 		if (strcmp(gp->name, name) == 0)
460 			return (gp);
461 	}
462 	return (NULL);
463 }
464 
465 /*
466  * Zero out all the statistics associated with a union device.
467  */
468 static void
g_union_ctl_reset(struct gctl_req * req,struct g_class * mp,bool verbose)469 g_union_ctl_reset(struct gctl_req *req, struct g_class *mp, bool verbose)
470 {
471 	struct g_union_softc *sc;
472 	struct g_provider *pp;
473 	struct g_geom *gp;
474 	char param[16];
475 	int i, *nargs;
476 
477 	g_topology_assert();
478 
479 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
480 	if (nargs == NULL) {
481 		gctl_error(req, "No '%s' argument.", "nargs");
482 		return;
483 	}
484 	if (*nargs <= 0) {
485 		gctl_error(req, "Missing device(s).");
486 		return;
487 	}
488 
489 	for (i = 0; i < *nargs; i++) {
490 		snprintf(param, sizeof(param), "arg%d", i);
491 		pp = gctl_get_provider(req, param);
492 		if (pp == NULL) {
493 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
494 			continue;
495 		}
496 		gp = pp->geom;
497 		if (gp->class != mp) {
498 			gctl_msg(req, EINVAL, "Provider %s is invalid.",
499 			    pp->name);
500 			continue;
501 		}
502 		sc = gp->softc;
503 		sc->sc_reads = 0;
504 		sc->sc_writes = 0;
505 		sc->sc_deletes = 0;
506 		sc->sc_getattrs = 0;
507 		sc->sc_flushes = 0;
508 		sc->sc_speedups = 0;
509 		sc->sc_cmd0s = 0;
510 		sc->sc_cmd1s = 0;
511 		sc->sc_cmd2s = 0;
512 		sc->sc_readbytes = 0;
513 		sc->sc_wrotebytes = 0;
514 		if (verbose)
515 			gctl_msg(req, 0, "Device %s has been reset.", pp->name);
516 		G_UNION_DEBUG(1, "Device %s has been reset.", pp->name);
517 	}
518 	gctl_post_messages(req);
519 }
520 
521 /*
522  * Revert all write requests made to the top layer of the union.
523  */
524 static void
g_union_ctl_revert(struct gctl_req * req,struct g_class * mp,bool verbose)525 g_union_ctl_revert(struct gctl_req *req, struct g_class *mp, bool verbose)
526 {
527 	struct g_union_softc *sc;
528 	struct g_provider *pp;
529 	struct g_geom *gp;
530 	char param[16];
531 	int i, *nargs;
532 
533 	g_topology_assert();
534 
535 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
536 	if (nargs == NULL) {
537 		gctl_error(req, "No '%s' argument.", "nargs");
538 		return;
539 	}
540 	if (*nargs <= 0) {
541 		gctl_error(req, "Missing device(s).");
542 		return;
543 	}
544 
545 	for (i = 0; i < *nargs; i++) {
546 		snprintf(param, sizeof(param), "arg%d", i);
547 		pp = gctl_get_provider(req, param);
548 		if (pp == NULL) {
549 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
550 			continue;
551 		}
552 		gp = pp->geom;
553 		if (gp->class != mp) {
554 			gctl_msg(req, EINVAL, "Provider %s is invalid.",
555 			    pp->name);
556 			continue;
557 		}
558 		sc = gp->softc;
559 		if (g_union_get_writelock(sc) != 0) {
560 			gctl_msg(req, EINVAL, "Revert already in progress for "
561 			    "provider %s.", pp->name);
562 			continue;
563 		}
564 		/*
565 		 * No mount or other use of union is allowed.
566 		 */
567 		if (pp->acr > 0 || pp->acw > 0 || pp->ace > 0) {
568 			gctl_msg(req, EPERM, "Unable to get exclusive access "
569 			    "for reverting of %s;\n\t%s cannot be mounted or "
570 			    "otherwise open during a revert.",
571 			     pp->name, pp->name);
572 			g_union_rel_writelock(sc);
573 			continue;
574 		}
575 		g_union_revert(sc);
576 		g_union_rel_writelock(sc);
577 		if (verbose)
578 			gctl_msg(req, 0, "Device %s has been reverted.",
579 			    pp->name);
580 		G_UNION_DEBUG(1, "Device %s has been reverted.", pp->name);
581 	}
582 	gctl_post_messages(req);
583 }
584 
585 /*
586  * Revert union writes by zero'ing out the writemap.
587  */
588 static void
g_union_revert(struct g_union_softc * sc)589 g_union_revert(struct g_union_softc *sc)
590 {
591 	int i;
592 
593 	G_WLOCK(sc);
594 	for (i = 0; i < sc->sc_root_size; i++)
595 		memset(sc->sc_writemap_root[i], 0,
596 		    sc->sc_leaf_size * sizeof(uint64_t));
597 	memset(sc->sc_leafused, 0, roundup(sc->sc_root_size, BITS_PER_ENTRY));
598 	G_WUNLOCK(sc);
599 }
600 
601 /*
602  * Commit all the writes made in the top layer to the lower layer.
603  */
604 static void
g_union_ctl_commit(struct gctl_req * req,struct g_class * mp,bool verbose)605 g_union_ctl_commit(struct gctl_req *req, struct g_class *mp, bool verbose)
606 {
607 	struct g_union_softc *sc;
608 	struct g_provider *pp, *lowerpp;
609 	struct g_consumer *lowercp;
610 	struct g_geom *gp;
611 	struct bio *bp;
612 	char param[16];
613 	off_t len2rd, len2wt, savelen;
614 	int i, error, error1, *nargs, *force, *reboot;
615 
616 	g_topology_assert();
617 
618 	nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
619 	if (nargs == NULL) {
620 		gctl_error(req, "No '%s' argument.", "nargs");
621 		return;
622 	}
623 	if (*nargs <= 0) {
624 		gctl_error(req, "Missing device(s).");
625 		return;
626 	}
627 	force = gctl_get_paraml(req, "force", sizeof(*force));
628 	if (force == NULL) {
629 		gctl_error(req, "No 'force' argument.");
630 		return;
631 	}
632 	reboot = gctl_get_paraml(req, "reboot", sizeof(*reboot));
633 	if (reboot == NULL) {
634 		gctl_error(req, "No 'reboot' argument.");
635 		return;
636 	}
637 
638 	/* Get a bio buffer to do our I/O */
639 	bp = g_alloc_bio();
640 	bp->bio_data = g_malloc(MAXBSIZE, M_WAITOK);
641 	bp->bio_done = biodone;
642 	for (i = 0; i < *nargs; i++) {
643 		snprintf(param, sizeof(param), "arg%d", i);
644 		pp = gctl_get_provider(req, param);
645 		if (pp == NULL) {
646 			gctl_msg(req, EINVAL, "No '%s' argument.", param);
647 			continue;
648 		}
649 		gp = pp->geom;
650 		if (gp->class != mp) {
651 			gctl_msg(req, EINVAL, "Provider %s is invalid.",
652 			    pp->name);
653 			continue;
654 		}
655 		sc = gp->softc;
656 		if (g_union_get_writelock(sc) != 0) {
657 			gctl_msg(req, EINVAL, "Commit already in progress for "
658 			    "provider %s.", pp->name);
659 			continue;
660 		}
661 
662 		/* upgrade to write access for lower */
663 		lowercp = sc->sc_lowercp;
664 		lowerpp = lowercp->provider;
665 		/*
666 		 * No mount or other use of union is allowed, unless the
667 		 * -f flag is given which allows read-only mount or usage.
668 		 */
669 		if ((*force == false && pp->acr > 0) || pp->acw > 0 ||
670 		     pp->ace > 0) {
671 			gctl_msg(req, EPERM, "Unable to get exclusive access "
672 			    "for writing of %s.\n\tNote that %s cannot be "
673 			    "mounted or otherwise\n\topen during a commit "
674 			    "unless the -f flag is used.", pp->name, pp->name);
675 			g_union_rel_writelock(sc);
676 			continue;
677 		}
678 		/*
679 		 * No mount or other use of lower media is allowed, unless the
680 		 * -f flag is given which allows read-only mount or usage.
681 		 */
682 		if ((*force == false && lowerpp->acr > lowercp->acr) ||
683 		     lowerpp->acw > lowercp->acw ||
684 		     lowerpp->ace > lowercp->ace) {
685 			gctl_msg(req, EPERM, "provider %s is unable to get "
686 			    "exclusive access to %s\n\tfor writing. Note that "
687 			    "%s cannot be mounted or otherwise open\n\tduring "
688 			    "a commit unless the -f flag is used.", pp->name,
689 			    lowerpp->name, lowerpp->name);
690 			g_union_rel_writelock(sc);
691 			continue;
692 		}
693 		if ((error = g_access(lowercp, 0, 1, 0)) != 0) {
694 			gctl_msg(req, error, "Error %d: provider %s is unable "
695 			    "to access %s for writing.", error, pp->name,
696 			    lowerpp->name);
697 			g_union_rel_writelock(sc);
698 			continue;
699 		}
700 		g_topology_unlock();
701 		/* Loop over write map copying across written blocks */
702 		bp->bio_offset = 0;
703 		bp->bio_length = sc->sc_map_size * sc->sc_sectorsize;
704 		G_RLOCK(sc);
705 		error = 0;
706 		while (bp->bio_length > 0) {
707 			if (!g_union_getmap(bp, sc, &len2rd)) {
708 				/* not written, so skip */
709 				bp->bio_offset += len2rd;
710 				bp->bio_length -= len2rd;
711 				continue;
712 			}
713 			G_RUNLOCK(sc);
714 			/* need to read then write len2rd sectors */
715 			for ( ; len2rd > 0; len2rd -= len2wt) {
716 				/* limit ourselves to MAXBSIZE size I/Os */
717 				len2wt = len2rd;
718 				if (len2wt > MAXBSIZE)
719 					len2wt = MAXBSIZE;
720 				savelen = bp->bio_length;
721 				bp->bio_length = len2wt;
722 				bp->bio_cmd = BIO_READ;
723 				g_io_request(bp, sc->sc_uppercp);
724 				if ((error = biowait(bp, "rdunion")) != 0) {
725 					gctl_msg(req, error, "Commit read "
726 					    "error %d in provider %s, commit "
727 					    "aborted.", error, pp->name);
728 					goto cleanup;
729 				}
730 				bp->bio_flags &= ~BIO_DONE;
731 				bp->bio_cmd = BIO_WRITE;
732 				g_io_request(bp, lowercp);
733 				if ((error = biowait(bp, "wtunion")) != 0) {
734 					gctl_msg(req, error, "Commit write "
735 					    "error %d in provider %s, commit "
736 					    "aborted.", error, pp->name);
737 					goto cleanup;
738 				}
739 				bp->bio_flags &= ~BIO_DONE;
740 				bp->bio_offset += len2wt;
741 				bp->bio_length = savelen - len2wt;
742 			}
743 			G_RLOCK(sc);
744 		}
745 		G_RUNLOCK(sc);
746 		/* clear the write map */
747 		g_union_revert(sc);
748 cleanup:
749 		g_topology_lock();
750 		/* return lower to previous access */
751 		if ((error1 = g_access(lowercp, 0, -1, 0)) != 0) {
752 			G_UNION_DEBUG(2, "Error %d: device %s could not reset "
753 			    "access to %s (r=0 w=-1 e=0).", error1, pp->name,
754 			    lowerpp->name);
755 		}
756 		g_union_rel_writelock(sc);
757 		if (error == 0 && verbose)
758 			gctl_msg(req, 0, "Device %s has been committed.",
759 			    pp->name);
760 		G_UNION_DEBUG(1, "Device %s has been committed.", pp->name);
761 	}
762 	gctl_post_messages(req);
763 	g_free(bp->bio_data);
764 	g_destroy_bio(bp);
765 	if (*reboot)
766 		kern_reboot(RB_AUTOBOOT);
767 }
768 
769 /*
770  * Generally allow access unless a commit is in progress.
771  */
772 static int
g_union_access(struct g_provider * pp,int r,int w,int e)773 g_union_access(struct g_provider *pp, int r, int w, int e)
774 {
775 	struct g_union_softc *sc;
776 
777 	sc = pp->geom->softc;
778 	if (sc == NULL) {
779 		if (r <= 0 && w <= 0 && e <= 0)
780 			return (0);
781 		return (ENXIO);
782 	}
783 	r += pp->acr;
784 	w += pp->acw;
785 	e += pp->ace;
786 	if (g_union_get_writelock(sc) != 0) {
787 		if ((pp->acr + pp->acw + pp->ace) > 0 && (r + w + e) == 0)
788 			return (0);
789 		return (EBUSY);
790 	}
791 	g_union_rel_writelock(sc);
792 	return (0);
793 }
794 
795 /*
796  * Initiate an I/O operation on the union device.
797  */
798 static void
g_union_start(struct bio * bp)799 g_union_start(struct bio *bp)
800 {
801 	struct g_union_softc *sc;
802 	struct g_union_wip *wip;
803 	struct bio *cbp;
804 
805 	sc = bp->bio_to->geom->softc;
806 	if (bp->bio_cmd == BIO_READ || bp->bio_cmd == BIO_WRITE) {
807 		wip = g_malloc(sizeof(*wip), M_NOWAIT);
808 		if (wip == NULL) {
809 			g_io_deliver(bp, ENOMEM);
810 			return;
811 		}
812 		TAILQ_INIT(&wip->wip_waiting);
813 		wip->wip_bp = bp;
814 		wip->wip_sc = sc;
815 		wip->wip_start = bp->bio_offset + sc->sc_offset;
816 		wip->wip_end = wip->wip_start + bp->bio_length - 1;
817 		wip->wip_numios = 1;
818 		wip->wip_error = 0;
819 		g_union_doio(wip);
820 		return;
821 	}
822 
823 	/*
824 	 * All commands other than read and write are passed through to
825 	 * the upper-level device since it is writable and thus able to
826 	 * respond to delete, flush, and speedup requests.
827 	 */
828 	cbp = g_clone_bio(bp);
829 	if (cbp == NULL) {
830 		g_io_deliver(bp, ENOMEM);
831 		return;
832 	}
833 	cbp->bio_offset = bp->bio_offset + sc->sc_offset;
834 	cbp->bio_done = g_std_done;
835 
836 	switch (cbp->bio_cmd) {
837 	case BIO_DELETE:
838 		G_UNION_LOGREQ(cbp, "Delete request received.");
839 		atomic_add_long(&sc->sc_deletes, 1);
840 		break;
841 	case BIO_GETATTR:
842 		G_UNION_LOGREQ(cbp, "Getattr request received.");
843 		atomic_add_long(&sc->sc_getattrs, 1);
844 		if (strcmp(cbp->bio_attribute, "GEOM::kerneldump") != 0)
845 			/* forward the GETATTR to the lower-level device */
846 			break;
847 		g_union_kerneldump(bp, sc);
848 		return;
849 	case BIO_FLUSH:
850 		G_UNION_LOGREQ(cbp, "Flush request received.");
851 		atomic_add_long(&sc->sc_flushes, 1);
852 		break;
853 	case BIO_SPEEDUP:
854 		G_UNION_LOGREQ(cbp, "Speedup request received.");
855 		atomic_add_long(&sc->sc_speedups, 1);
856 		break;
857 	case BIO_CMD0:
858 		G_UNION_LOGREQ(cbp, "Cmd0 request received.");
859 		atomic_add_long(&sc->sc_cmd0s, 1);
860 		break;
861 	case BIO_CMD1:
862 		G_UNION_LOGREQ(cbp, "Cmd1 request received.");
863 		atomic_add_long(&sc->sc_cmd1s, 1);
864 		break;
865 	case BIO_CMD2:
866 		G_UNION_LOGREQ(cbp, "Cmd2 request received.");
867 		atomic_add_long(&sc->sc_cmd2s, 1);
868 		break;
869 	default:
870 		G_UNION_LOGREQ(cbp, "Unknown (%d) request received.",
871 		    cbp->bio_cmd);
872 		break;
873 	}
874 	g_io_request(cbp, sc->sc_uppercp);
875 }
876 
877 /*
878  * Initiate a read or write operation on the union device.
879  */
880 static void
g_union_doio(struct g_union_wip * wip)881 g_union_doio(struct g_union_wip *wip)
882 {
883 	struct g_union_softc *sc;
884 	struct g_consumer *cp, *firstcp;
885 	struct g_union_wip *activewip;
886 	struct bio *cbp, *firstbp;
887 	off_t rdlen, len2rd, offset;
888 	int iocnt, needstoblock;
889 	char *level;
890 
891 	/*
892 	 * To maintain consistency, we cannot allow concurrent reads
893 	 * or writes to the same block.
894 	 *
895 	 * A work-in-progress (wip) structure is allocated for each
896 	 * read or write request. All active requests are kept on the
897 	 * softc sc_wiplist. As each request arrives, it is checked to
898 	 * see if it overlaps any of the active entries. If it does not
899 	 * overlap, then it is added to the active list and initiated.
900 	 * If it does overlap an active entry, it is added to the
901 	 * wip_waiting list for the active entry that it overlaps.
902 	 * When an active entry completes, it restarts all the requests
903 	 * on its wip_waiting list.
904 	 */
905 	sc = wip->wip_sc;
906 	G_WLOCK(sc);
907 	TAILQ_FOREACH(activewip, &sc->sc_wiplist, wip_next) {
908 		if (wip->wip_end < activewip->wip_start ||
909 		    wip->wip_start > activewip->wip_end)
910 			continue;
911 		needstoblock = 1;
912 		if (wip->wip_bp->bio_cmd == BIO_WRITE)
913 			if (activewip->wip_bp->bio_cmd == BIO_WRITE)
914 				sc->sc_writeblockwrite += 1;
915 			else
916 				sc->sc_readblockwrite += 1;
917 		else
918 			if (activewip->wip_bp->bio_cmd == BIO_WRITE)
919 				sc->sc_writeblockread += 1;
920 			else {
921 				sc->sc_readcurrentread += 1;
922 				needstoblock = 0;
923 			}
924 		/* Put request on a waiting list if necessary */
925 		if (needstoblock) {
926 			TAILQ_INSERT_TAIL(&activewip->wip_waiting, wip,
927 			    wip_next);
928 			G_WUNLOCK(sc);
929 			return;
930 		}
931 	}
932 	/* Put request on the active list */
933 	TAILQ_INSERT_TAIL(&sc->sc_wiplist, wip, wip_next);
934 
935 	/*
936 	 * Process I/O requests that have been cleared to go.
937 	 */
938 	cbp = g_clone_bio(wip->wip_bp);
939 	if (cbp == NULL) {
940 		TAILQ_REMOVE(&sc->sc_wiplist, wip, wip_next);
941 		G_WUNLOCK(sc);
942 		KASSERT(TAILQ_FIRST(&wip->wip_waiting) == NULL,
943 		    ("g_union_doio: non-empty work-in-progress waiting queue"));
944 		g_io_deliver(wip->wip_bp, ENOMEM);
945 		g_free(wip);
946 		return;
947 	}
948 	G_WUNLOCK(sc);
949 	cbp->bio_caller1 = wip;
950 	cbp->bio_done = g_union_done;
951 	cbp->bio_offset = wip->wip_start;
952 
953 	/*
954 	 * Writes are always done to the top level. The blocks that
955 	 * are written are recorded in the bitmap when the I/O completes.
956 	 */
957 	if (cbp->bio_cmd == BIO_WRITE) {
958 		G_UNION_LOGREQ(cbp, "Sending %jd byte write request to upper "
959 		    "level.", cbp->bio_length);
960 		atomic_add_long(&sc->sc_writes, 1);
961 		atomic_add_long(&sc->sc_wrotebytes, cbp->bio_length);
962 		g_io_request(cbp, sc->sc_uppercp);
963 		return;
964 	}
965 	/*
966 	 * The usual read case is that we either read the top layer
967 	 * if the block has been previously written or the bottom layer
968 	 * if it has not been written. However, it is possible that
969 	 * only part of the block has been written, For example we may
970 	 * have written a UFS/FFS file fragment comprising several
971 	 * sectors out of an 8-sector block.  Here, if the entire
972 	 * 8-sector block is read for example by a snapshot needing
973 	 * to copy the full block, then we need to read the written
974 	 * sectors from the upper level and the unwritten sectors from
975 	 * the lower level. We do this by alternately reading from the
976 	 * top and bottom layers until we complete the read. We
977 	 * simplify for the common case to just do the I/O and return.
978 	 */
979 	atomic_add_long(&sc->sc_reads, 1);
980 	atomic_add_long(&sc->sc_readbytes, cbp->bio_length);
981 	rdlen = cbp->bio_length;
982 	offset = 0;
983 	for (iocnt = 0; ; iocnt++) {
984 		if (g_union_getmap(cbp, sc, &len2rd)) {
985 			/* read top */
986 			cp = sc->sc_uppercp;
987 			level = "upper";
988 		} else {
989 			/* read bottom */
990 			cp = sc->sc_lowercp;
991 			level = "lower";
992 		}
993 		/* Check if only a single read is required */
994 		if (iocnt == 0 && rdlen == len2rd) {
995 			G_UNION_LOGREQLVL((cp == sc->sc_uppercp) ?
996 			    3 : 4, cbp, "Sending %jd byte read "
997 			    "request to %s level.", len2rd, level);
998 			g_io_request(cbp, cp);
999 			return;
1000 		}
1001 		cbp->bio_length = len2rd;
1002 		if ((cbp->bio_flags & BIO_UNMAPPED) != 0)
1003 			cbp->bio_ma_offset += offset;
1004 		else
1005 			cbp->bio_data += offset;
1006 		offset += len2rd;
1007 		rdlen -= len2rd;
1008 		G_UNION_LOGREQLVL(3, cbp, "Sending %jd byte read "
1009 		    "request to %s level.", len2rd, level);
1010 		/*
1011 		 * To avoid prematurely notifying our consumer
1012 		 * that their I/O has completed, we have to delay
1013 		 * issuing our first I/O request until we have
1014 		 * issued all the additional I/O requests.
1015 		 */
1016 		if (iocnt > 0) {
1017 			atomic_add_long(&wip->wip_numios, 1);
1018 			g_io_request(cbp, cp);
1019 		} else {
1020 			firstbp = cbp;
1021 			firstcp = cp;
1022 		}
1023 		if (rdlen == 0)
1024 			break;
1025 		/* set up for next read */
1026 		cbp = g_clone_bio(wip->wip_bp);
1027 		if (cbp == NULL) {
1028 			wip->wip_error = ENOMEM;
1029 			atomic_add_long(&wip->wip_numios, -1);
1030 			break;
1031 		}
1032 		cbp->bio_caller1 = wip;
1033 		cbp->bio_done = g_union_done;
1034 		cbp->bio_offset += offset;
1035 		cbp->bio_length = rdlen;
1036 		atomic_add_long(&sc->sc_reads, 1);
1037 	}
1038 	/* We have issued all our I/O, so start the first one */
1039 	g_io_request(firstbp, firstcp);
1040 	return;
1041 }
1042 
1043 /*
1044  * Used when completing a union I/O operation.
1045  */
1046 static void
g_union_done(struct bio * bp)1047 g_union_done(struct bio *bp)
1048 {
1049 	struct g_union_wip *wip, *waitingwip;
1050 	struct g_union_softc *sc;
1051 
1052 	wip = bp->bio_caller1;
1053 	if (wip->wip_error != 0 && bp->bio_error == 0)
1054 		bp->bio_error = wip->wip_error;
1055 	wip->wip_error = 0;
1056 	if (atomic_fetchadd_long(&wip->wip_numios, -1) == 1) {
1057 		sc = wip->wip_sc;
1058 		G_WLOCK(sc);
1059 		if (bp->bio_cmd == BIO_WRITE)
1060 			g_union_setmap(bp, sc);
1061 		TAILQ_REMOVE(&sc->sc_wiplist, wip, wip_next);
1062 		G_WUNLOCK(sc);
1063 		while ((waitingwip = TAILQ_FIRST(&wip->wip_waiting)) != NULL) {
1064 			TAILQ_REMOVE(&wip->wip_waiting, waitingwip, wip_next);
1065 			g_union_doio(waitingwip);
1066 		}
1067 		g_free(wip);
1068 	}
1069 	g_std_done(bp);
1070 }
1071 
1072 /*
1073  * Record blocks that have been written in the map.
1074  */
1075 static void
g_union_setmap(struct bio * bp,struct g_union_softc * sc)1076 g_union_setmap(struct bio *bp, struct g_union_softc *sc)
1077 {
1078 	size_t root_idx;
1079 	uint64_t **leaf;
1080 	uint64_t *wordp;
1081 	off_t start, numsec;
1082 
1083 	G_WLOCKOWNED(sc);
1084 	KASSERT(bp->bio_offset % sc->sc_sectorsize == 0,
1085 	    ("g_union_setmap: offset not on sector boundry"));
1086 	KASSERT(bp->bio_length % sc->sc_sectorsize == 0,
1087 	    ("g_union_setmap: length not a multiple of sectors"));
1088 	start = bp->bio_offset / sc->sc_sectorsize;
1089 	numsec = bp->bio_length / sc->sc_sectorsize;
1090 	KASSERT(start + numsec <= sc->sc_map_size,
1091 	    ("g_union_setmap: block %jd is out of range", start + numsec));
1092 	for ( ; numsec > 0; numsec--, start++) {
1093 		root_idx = start / sc->sc_bits_per_leaf;
1094 		leaf = &sc->sc_writemap_root[root_idx];
1095 		wordp = &(*leaf)
1096 		    [(start % sc->sc_bits_per_leaf) / BITS_PER_ENTRY];
1097 		*wordp |= 1ULL << (start % BITS_PER_ENTRY);
1098 		sc->sc_leafused[root_idx / BITS_PER_ENTRY] |=
1099 		    1ULL << (root_idx % BITS_PER_ENTRY);
1100 	}
1101 }
1102 
1103 /*
1104  * Check map to determine whether blocks have been written.
1105  *
1106  * Return true if they have been written so should be read from the top
1107  * layer. Return false if they have not been written so should be read
1108  * from the bottom layer. Return in len2read the bytes to be read. See
1109  * the comment above the BIO_READ implementation in g_union_start() for
1110  * an explantion of why len2read may be shorter than the buffer length.
1111  */
1112 static bool
g_union_getmap(struct bio * bp,struct g_union_softc * sc,off_t * len2read)1113 g_union_getmap(struct bio *bp, struct g_union_softc *sc, off_t *len2read)
1114 {
1115 	off_t start, numsec, leafresid, bitloc;
1116 	bool first, maptype, retval;
1117 	uint64_t *leaf, word;
1118 	size_t root_idx;
1119 
1120 	KASSERT(bp->bio_offset % sc->sc_sectorsize == 0,
1121 	    ("g_union_getmap: offset not on sector boundry"));
1122 	KASSERT(bp->bio_length % sc->sc_sectorsize == 0,
1123 	    ("g_union_getmap: length not a multiple of sectors"));
1124 	start = bp->bio_offset / sc->sc_sectorsize;
1125 	numsec = bp->bio_length / sc->sc_sectorsize;
1126 	G_UNION_DEBUG(4, "g_union_getmap: check %jd sectors starting at %jd\n",
1127 	    numsec, start);
1128 	KASSERT(start + numsec <= sc->sc_map_size,
1129 	    ("g_union_getmap: block %jd is out of range", start + numsec));
1130 		root_idx = start / sc->sc_bits_per_leaf;
1131 	first = true;
1132 	maptype = false;
1133 	while (numsec > 0) {
1134 		/* Check first if the leaf records any written sectors */
1135 		root_idx = start / sc->sc_bits_per_leaf;
1136 		leafresid = sc->sc_bits_per_leaf -
1137 		    (start % sc->sc_bits_per_leaf);
1138 		if (((sc->sc_leafused[root_idx / BITS_PER_ENTRY]) &
1139 		    (1ULL << (root_idx % BITS_PER_ENTRY))) == 0) {
1140 			if (first) {
1141 				maptype = false;
1142 				first = false;
1143 			}
1144 			if (maptype)
1145 				break;
1146 			numsec -= leafresid;
1147 			start += leafresid;
1148 			continue;
1149 		}
1150 		/* Check up to a word boundry, then check word by word */
1151 		leaf = sc->sc_writemap_root[root_idx];
1152 		word = leaf[(start % sc->sc_bits_per_leaf) / BITS_PER_ENTRY];
1153 		bitloc = start % BITS_PER_ENTRY;
1154 		if (bitloc == 0 && (word == 0 || word == ~0)) {
1155 			if (first) {
1156 				if (word == 0)
1157 					maptype = false;
1158 				else
1159 					maptype = true;
1160 				first = false;
1161 			}
1162 			if ((word == 0 && maptype) ||
1163 			    (word == ~0 && !maptype))
1164 				break;
1165 			numsec -= BITS_PER_ENTRY;
1166 			start += BITS_PER_ENTRY;
1167 			continue;
1168 		}
1169 		for ( ; bitloc < BITS_PER_ENTRY; bitloc ++) {
1170 			retval = (word & (1ULL << bitloc)) != 0;
1171 			if (first) {
1172 				maptype = retval;
1173 				first = false;
1174 			}
1175 			if (maptype == retval) {
1176 				numsec--;
1177 				start++;
1178 				continue;
1179 			}
1180 			goto out;
1181 		}
1182 	}
1183 out:
1184 	if (numsec < 0) {
1185 		start += numsec;
1186 		numsec = 0;
1187 	}
1188 	*len2read = bp->bio_length - (numsec * sc->sc_sectorsize);
1189 	G_UNION_DEBUG(maptype ? 3 : 4,
1190 	    "g_union_getmap: return maptype %swritten for %jd "
1191 	    "sectors ending at %jd\n", maptype ? "" : "NOT ",
1192 	    *len2read / sc->sc_sectorsize, start - 1);
1193 	return (maptype);
1194 }
1195 
1196 /*
1197  * Fill in details for a BIO_GETATTR request.
1198  */
1199 static void
g_union_kerneldump(struct bio * bp,struct g_union_softc * sc)1200 g_union_kerneldump(struct bio *bp, struct g_union_softc *sc)
1201 {
1202 	struct g_kerneldump *gkd;
1203 	struct g_geom *gp;
1204 	struct g_provider *pp;
1205 
1206 	gkd = (struct g_kerneldump *)bp->bio_data;
1207 	gp = bp->bio_to->geom;
1208 	g_trace(G_T_TOPOLOGY, "%s(%s, %jd, %jd)", __func__, gp->name,
1209 	    (intmax_t)gkd->offset, (intmax_t)gkd->length);
1210 
1211 	pp = LIST_FIRST(&gp->provider);
1212 
1213 	gkd->di.dumper = g_union_dumper;
1214 	gkd->di.priv = sc;
1215 	gkd->di.blocksize = pp->sectorsize;
1216 	gkd->di.maxiosize = DFLTPHYS;
1217 	gkd->di.mediaoffset = sc->sc_offset + gkd->offset;
1218 	if (gkd->offset > sc->sc_size) {
1219 		g_io_deliver(bp, ENODEV);
1220 		return;
1221 	}
1222 	if (gkd->offset + gkd->length > sc->sc_size)
1223 		gkd->length = sc->sc_size - gkd->offset;
1224 	gkd->di.mediasize = gkd->length;
1225 	g_io_deliver(bp, 0);
1226 }
1227 
1228 /*
1229  * Handler for g_union_kerneldump().
1230  */
1231 static int
g_union_dumper(void * priv,void * virtual,off_t offset,size_t length)1232 g_union_dumper(void *priv, void *virtual, off_t offset, size_t length)
1233 {
1234 
1235 	return (0);
1236 }
1237 
1238 /*
1239  * List union statistics.
1240  */
1241 static void
g_union_dumpconf(struct sbuf * sb,const char * indent,struct g_geom * gp,struct g_consumer * cp,struct g_provider * pp)1242 g_union_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1243     struct g_consumer *cp, struct g_provider *pp)
1244 {
1245 	struct g_union_softc *sc;
1246 
1247 	if (pp != NULL || cp != NULL || gp->softc == NULL)
1248 		return;
1249 	sc = gp->softc;
1250 	sbuf_printf(sb, "%s<Reads>%ju</Reads>\n", indent,
1251 	    (uintmax_t)sc->sc_reads);
1252 	sbuf_printf(sb, "%s<Writes>%ju</Writes>\n", indent,
1253 	    (uintmax_t)sc->sc_writes);
1254 	sbuf_printf(sb, "%s<Deletes>%ju</Deletes>\n", indent,
1255 	    (uintmax_t)sc->sc_deletes);
1256 	sbuf_printf(sb, "%s<Getattrs>%ju</Getattrs>\n", indent,
1257 	    (uintmax_t)sc->sc_getattrs);
1258 	sbuf_printf(sb, "%s<Flushes>%ju</Flushes>\n", indent,
1259 	    (uintmax_t)sc->sc_flushes);
1260 	sbuf_printf(sb, "%s<Speedups>%ju</Speedups>\n", indent,
1261 	    (uintmax_t)sc->sc_speedups);
1262 	sbuf_printf(sb, "%s<Cmd0s>%ju</Cmd0s>\n", indent,
1263 	    (uintmax_t)sc->sc_cmd0s);
1264 	sbuf_printf(sb, "%s<Cmd1s>%ju</Cmd1s>\n", indent,
1265 	    (uintmax_t)sc->sc_cmd1s);
1266 	sbuf_printf(sb, "%s<Cmd2s>%ju</Cmd2s>\n", indent,
1267 	    (uintmax_t)sc->sc_cmd2s);
1268 	sbuf_printf(sb, "%s<ReadCurrentRead>%ju</ReadCurrentRead>\n", indent,
1269 	    (uintmax_t)sc->sc_readcurrentread);
1270 	sbuf_printf(sb, "%s<ReadBlockWrite>%ju</ReadBlockWrite>\n", indent,
1271 	    (uintmax_t)sc->sc_readblockwrite);
1272 	sbuf_printf(sb, "%s<WriteBlockRead>%ju</WriteBlockRead>\n", indent,
1273 	    (uintmax_t)sc->sc_writeblockread);
1274 	sbuf_printf(sb, "%s<WriteBlockWrite>%ju</WriteBlockWrite>\n", indent,
1275 	    (uintmax_t)sc->sc_writeblockwrite);
1276 	sbuf_printf(sb, "%s<ReadBytes>%ju</ReadBytes>\n", indent,
1277 	    (uintmax_t)sc->sc_readbytes);
1278 	sbuf_printf(sb, "%s<WroteBytes>%ju</WroteBytes>\n", indent,
1279 	    (uintmax_t)sc->sc_wrotebytes);
1280 	sbuf_printf(sb, "%s<Offset>%jd</Offset>\n", indent,
1281 	    (intmax_t)sc->sc_offset);
1282 }
1283 
1284 /*
1285  * Clean up an orphaned geom.
1286  */
1287 static void
g_union_orphan(struct g_consumer * cp)1288 g_union_orphan(struct g_consumer *cp)
1289 {
1290 
1291 	g_topology_assert();
1292 	g_union_destroy(NULL, cp->geom, true);
1293 }
1294 
1295 /*
1296  * Clean up a union geom.
1297  */
1298 static int
g_union_destroy_geom(struct gctl_req * req,struct g_class * mp,struct g_geom * gp)1299 g_union_destroy_geom(struct gctl_req *req, struct g_class *mp,
1300     struct g_geom *gp)
1301 {
1302 
1303 	return (g_union_destroy(NULL, gp, false));
1304 }
1305 
1306 /*
1307  * Clean up a union device.
1308  */
1309 static int
g_union_destroy(struct gctl_req * req,struct g_geom * gp,bool force)1310 g_union_destroy(struct gctl_req *req, struct g_geom *gp, bool force)
1311 {
1312 	struct g_union_softc *sc;
1313 	struct g_provider *pp;
1314 	int error;
1315 
1316 	g_topology_assert();
1317 	sc = gp->softc;
1318 	if (sc == NULL)
1319 		return (ENXIO);
1320 	pp = LIST_FIRST(&gp->provider);
1321 	if ((sc->sc_flags & DOING_COMMIT) != 0 ||
1322 	    (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0))) {
1323 		if (force) {
1324 			if (req != NULL)
1325 				gctl_msg(req, 0, "Device %s is still in use, "
1326 				    "so is being forcibly removed.", gp->name);
1327 			G_UNION_DEBUG(1, "Device %s is still in use, so "
1328 			    "is being forcibly removed.", gp->name);
1329 		} else {
1330 			if (req != NULL)
1331 				gctl_msg(req, EBUSY, "Device %s is still open "
1332 				    "(r=%d w=%d e=%d).", gp->name, pp->acr,
1333 				    pp->acw, pp->ace);
1334 			G_UNION_DEBUG(1, "Device %s is still open "
1335 			    "(r=%d w=%d e=%d).", gp->name, pp->acr,
1336 			    pp->acw, pp->ace);
1337 			return (EBUSY);
1338 		}
1339 	} else {
1340 		if (req != NULL)
1341 			gctl_msg(req, 0, "Device %s removed.", gp->name);
1342 		G_UNION_DEBUG(1, "Device %s removed.", gp->name);
1343 	}
1344 	/* Close consumers */
1345 	if ((error = g_access(sc->sc_lowercp, -1, 0, -1)) != 0)
1346 		G_UNION_DEBUG(2, "Error %d: device %s could not reset access "
1347 		    "to %s.", error, gp->name, sc->sc_lowercp->provider->name);
1348 	if ((error = g_access(sc->sc_uppercp, -1, -1, -1)) != 0)
1349 		G_UNION_DEBUG(2, "Error %d: device %s could not reset access "
1350 		    "to %s.", error, gp->name, sc->sc_uppercp->provider->name);
1351 
1352 	g_wither_geom(gp, ENXIO);
1353 
1354 	return (0);
1355 }
1356 
1357 /*
1358  * Clean up a union provider.
1359  */
1360 static void
g_union_providergone(struct g_provider * pp)1361 g_union_providergone(struct g_provider *pp)
1362 {
1363 	struct g_geom *gp;
1364 	struct g_union_softc *sc;
1365 	size_t i;
1366 
1367 	gp = pp->geom;
1368 	sc = gp->softc;
1369 	gp->softc = NULL;
1370 	for (i = 0; i < sc->sc_root_size; i++)
1371 		g_free(sc->sc_writemap_root[i]);
1372 	g_free(sc->sc_writemap_root);
1373 	g_free(sc->sc_leafused);
1374 	rw_destroy(&sc->sc_rwlock);
1375 	g_free(sc);
1376 }
1377 
1378 /*
1379  * Respond to a resized provider.
1380  */
1381 static void
g_union_resize(struct g_consumer * cp)1382 g_union_resize(struct g_consumer *cp)
1383 {
1384 	struct g_union_softc *sc;
1385 	struct g_geom *gp;
1386 
1387 	g_topology_assert();
1388 
1389 	gp = cp->geom;
1390 	sc = gp->softc;
1391 
1392 	/*
1393 	 * If size has gotten bigger, ignore it and just keep using
1394 	 * the space we already had. Otherwise we are done.
1395 	 */
1396 	if (sc->sc_size < cp->provider->mediasize - sc->sc_offset)
1397 		return;
1398 	g_union_destroy(NULL, gp, true);
1399 }
1400 
1401 DECLARE_GEOM_CLASS(g_union_class, g_union);
1402 MODULE_VERSION(geom_union, 0);
1403