1 /*-
2 * Copyright (c) 2006-2007 Ivan Voras <ivoras@freebsd.org>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27 /* Implementation notes:
28 * - "Components" are wrappers around providers that make up the
29 * virtual storage (i.e. a virstor has "physical" components)
30 */
31
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD: releng/10.0/sys/geom/virstor/g_virstor.c 239021 2012-08-03 20:24:16Z jimharris $");
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/sx.h>
42 #include <sys/bio.h>
43 #include <sys/sbuf.h>
44 #include <sys/sysctl.h>
45 #include <sys/malloc.h>
46 #include <sys/time.h>
47 #include <sys/proc.h>
48 #include <sys/kthread.h>
49 #include <sys/mutex.h>
50 #include <vm/uma.h>
51 #include <geom/geom.h>
52
53 #include <geom/virstor/g_virstor.h>
54 #include <geom/virstor/g_virstor_md.h>
55
56 FEATURE(g_virstor, "GEOM virtual storage support");
57
58 /* Declare malloc(9) label */
59 static MALLOC_DEFINE(M_GVIRSTOR, "gvirstor", "GEOM_VIRSTOR Data");
60
61 /* GEOM class methods */
62 static g_init_t g_virstor_init;
63 static g_fini_t g_virstor_fini;
64 static g_taste_t g_virstor_taste;
65 static g_ctl_req_t g_virstor_config;
66 static g_ctl_destroy_geom_t g_virstor_destroy_geom;
67
68 /* Declare & initialize class structure ("geom class") */
69 struct g_class g_virstor_class = {
70 .name = G_VIRSTOR_CLASS_NAME,
71 .version = G_VERSION,
72 .init = g_virstor_init,
73 .fini = g_virstor_fini,
74 .taste = g_virstor_taste,
75 .ctlreq = g_virstor_config,
76 .destroy_geom = g_virstor_destroy_geom
77 /* The .dumpconf and the rest are only usable for a geom instance, so
78 * they will be set when such instance is created. */
79 };
80
81 /* Declare sysctl's and loader tunables */
82 SYSCTL_DECL(_kern_geom);
83 static SYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0,
84 "GEOM_GVIRSTOR information");
85
86 static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */
87 TUNABLE_INT("kern.geom.virstor.debug", &g_virstor_debug);
88 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RW, &g_virstor_debug,
89 0, "Debug level (2=production, 5=normal, 15=excessive)");
90
91 static u_int g_virstor_chunk_watermark = 100;
92 TUNABLE_INT("kern.geom.virstor.chunk_watermark", &g_virstor_chunk_watermark);
93 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RW,
94 &g_virstor_chunk_watermark, 0,
95 "Minimum number of free chunks before issuing administrative warning");
96
97 static u_int g_virstor_component_watermark = 1;
98 TUNABLE_INT("kern.geom.virstor.component_watermark",
99 &g_virstor_component_watermark);
100 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RW,
101 &g_virstor_component_watermark, 0,
102 "Minimum number of free components before issuing administrative warning");
103
104 static int read_metadata(struct g_consumer *, struct g_virstor_metadata *);
105 static void write_metadata(struct g_consumer *, struct g_virstor_metadata *);
106 static int clear_metadata(struct g_virstor_component *);
107 static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *,
108 struct g_virstor_metadata *);
109 static struct g_geom *create_virstor_geom(struct g_class *,
110 struct g_virstor_metadata *);
111 static void virstor_check_and_run(struct g_virstor_softc *);
112 static u_int virstor_valid_components(struct g_virstor_softc *);
113 static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t,
114 boolean_t);
115 static void remove_component(struct g_virstor_softc *,
116 struct g_virstor_component *, boolean_t);
117 static void bioq_dismantle(struct bio_queue_head *);
118 static int allocate_chunk(struct g_virstor_softc *,
119 struct g_virstor_component **, u_int *, u_int *);
120 static void delay_destroy_consumer(void *, int);
121 static void dump_component(struct g_virstor_component *comp);
122 #if 0
123 static void dump_me(struct virstor_map_entry *me, unsigned int nr);
124 #endif
125
126 static void virstor_ctl_stop(struct gctl_req *, struct g_class *);
127 static void virstor_ctl_add(struct gctl_req *, struct g_class *);
128 static void virstor_ctl_remove(struct gctl_req *, struct g_class *);
129 static struct g_virstor_softc * virstor_find_geom(const struct g_class *,
130 const char *);
131 static void update_metadata(struct g_virstor_softc *);
132 static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *,
133 u_int, u_int);
134
135 static void g_virstor_orphan(struct g_consumer *);
136 static int g_virstor_access(struct g_provider *, int, int, int);
137 static void g_virstor_start(struct bio *);
138 static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *,
139 struct g_consumer *, struct g_provider *);
140 static void g_virstor_done(struct bio *);
141
142 static void invalid_call(void);
143 /*
144 * Initialise GEOM class (per-class callback)
145 */
146 static void
147 g_virstor_init(struct g_class *mp __unused)
148 {
149
150 /* Catch map struct size mismatch at compile time; Map entries must
151 * fit into MAXPHYS exactly, with no wasted space. */
152 CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS);
153
154 /* Init UMA zones, TAILQ's, other global vars */
155 }
156
157 /*
158 * Finalise GEOM class (per-class callback)
159 */
160 static void
161 g_virstor_fini(struct g_class *mp __unused)
162 {
163
164 /* Deinit UMA zones & global vars */
165 }
166
167 /*
168 * Config (per-class callback)
169 */
170 static void
171 g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb)
172 {
173 uint32_t *version;
174
175 g_topology_assert();
176
177 version = gctl_get_paraml(req, "version", sizeof(*version));
178 if (version == NULL) {
179 gctl_error(req, "Failed to get 'version' argument");
180 return;
181 }
182 if (*version != G_VIRSTOR_VERSION) {
183 gctl_error(req, "Userland and kernel versions out of sync");
184 return;
185 }
186
187 g_topology_unlock();
188 if (strcmp(verb, "add") == 0)
189 virstor_ctl_add(req, cp);
190 else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0)
191 virstor_ctl_stop(req, cp);
192 else if (strcmp(verb, "remove") == 0)
193 virstor_ctl_remove(req, cp);
194 else
195 gctl_error(req, "unknown verb: '%s'", verb);
196 g_topology_lock();
197 }
198
199 /*
200 * "stop" verb from userland
201 */
202 static void
203 virstor_ctl_stop(struct gctl_req *req, struct g_class *cp)
204 {
205 int *force, *nargs;
206 int i;
207
208 nargs = gctl_get_paraml(req, "nargs", sizeof *nargs);
209 if (nargs == NULL) {
210 gctl_error(req, "Error fetching argument '%s'", "nargs");
211 return;
212 }
213 if (*nargs < 1) {
214 gctl_error(req, "Invalid number of arguments");
215 return;
216 }
217 force = gctl_get_paraml(req, "force", sizeof *force);
218 if (force == NULL) {
219 gctl_error(req, "Error fetching argument '%s'", "force");
220 return;
221 }
222
223 g_topology_lock();
224 for (i = 0; i < *nargs; i++) {
225 char param[8];
226 const char *name;
227 struct g_virstor_softc *sc;
228 int error;
229
230 sprintf(param, "arg%d", i);
231 name = gctl_get_asciiparam(req, param);
232 if (name == NULL) {
233 gctl_error(req, "No 'arg%d' argument", i);
234 g_topology_unlock();
235 return;
236 }
237 sc = virstor_find_geom(cp, name);
238 if (sc == NULL) {
239 gctl_error(req, "Don't know anything about '%s'", name);
240 g_topology_unlock();
241 return;
242 }
243
244 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
245 sc->geom->name);
246 update_metadata(sc);
247 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
248 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
249 sc->geom->name, error);
250 }
251 }
252 g_topology_unlock();
253 }
254
255 /*
256 * "add" verb from userland - add new component(s) to the structure.
257 * This will be done all at once in here, without going through the
258 * .taste function for new components.
259 */
260 static void
261 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
262 {
263 /* Note: while this is going on, I/O is being done on
264 * the g_up and g_down threads. The idea is to make changes
265 * to softc members in a way that can atomically activate
266 * them all at once. */
267 struct g_virstor_softc *sc;
268 int *hardcode, *nargs;
269 const char *geom_name; /* geom to add a component to */
270 struct g_consumer *fcp;
271 struct g_virstor_bio_q *bq;
272 u_int added;
273 int error;
274 int i;
275
276 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
277 if (nargs == NULL) {
278 gctl_error(req, "Error fetching argument '%s'", "nargs");
279 return;
280 }
281 if (*nargs < 2) {
282 gctl_error(req, "Invalid number of arguments");
283 return;
284 }
285 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
286 if (hardcode == NULL) {
287 gctl_error(req, "Error fetching argument '%s'", "hardcode");
288 return;
289 }
290
291 /* Find "our" geom */
292 geom_name = gctl_get_asciiparam(req, "arg0");
293 if (geom_name == NULL) {
294 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
295 return;
296 }
297 sc = virstor_find_geom(cp, geom_name);
298 if (sc == NULL) {
299 gctl_error(req, "Don't know anything about '%s'", geom_name);
300 return;
301 }
302
303 if (virstor_valid_components(sc) != sc->n_components) {
304 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
305 "virstor %s", sc->geom->name);
306 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
307 return;
308 }
309
310 fcp = sc->components[0].gcons;
311 added = 0;
312 g_topology_lock();
313 for (i = 1; i < *nargs; i++) {
314 struct g_virstor_metadata md;
315 char aname[8];
316 const char *prov_name;
317 struct g_provider *pp;
318 struct g_consumer *cp;
319 u_int nc;
320 u_int j;
321
322 snprintf(aname, sizeof aname, "arg%d", i);
323 prov_name = gctl_get_asciiparam(req, aname);
324 if (prov_name == NULL) {
325 gctl_error(req, "Error fetching argument '%s'", aname);
326 g_topology_unlock();
327 return;
328 }
329 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
330 prov_name += sizeof(_PATH_DEV) - 1;
331
332 pp = g_provider_by_name(prov_name);
333 if (pp == NULL) {
334 /* This is the most common error so be verbose about it */
335 if (added != 0) {
336 gctl_error(req, "Invalid provider: '%s' (added"
337 " %u components)", prov_name, added);
338 update_metadata(sc);
339 } else {
340 gctl_error(req, "Invalid provider: '%s'",
341 prov_name);
342 }
343 g_topology_unlock();
344 return;
345 }
346 cp = g_new_consumer(sc->geom);
347 if (cp == NULL) {
348 gctl_error(req, "Cannot create consumer");
349 g_topology_unlock();
350 return;
351 }
352 error = g_attach(cp, pp);
353 if (error != 0) {
354 gctl_error(req, "Cannot attach a consumer to %s",
355 pp->name);
356 g_destroy_consumer(cp);
357 g_topology_unlock();
358 return;
359 }
360 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
361 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
362 if (error != 0) {
363 gctl_error(req, "Access request failed for %s",
364 pp->name);
365 g_destroy_consumer(cp);
366 g_topology_unlock();
367 return;
368 }
369 }
370 if (fcp->provider->sectorsize != pp->sectorsize) {
371 gctl_error(req, "Sector size doesn't fit for %s",
372 pp->name);
373 g_destroy_consumer(cp);
374 g_topology_unlock();
375 return;
376 }
377 for (j = 0; j < sc->n_components; j++) {
378 if (strcmp(sc->components[j].gcons->provider->name,
379 pp->name) == 0) {
380 gctl_error(req, "Component %s already in %s",
381 pp->name, sc->geom->name);
382 g_destroy_consumer(cp);
383 g_topology_unlock();
384 return;
385 }
386 }
387 sc->components = realloc(sc->components,
388 sizeof(*sc->components) * (sc->n_components + 1),
389 M_GVIRSTOR, M_WAITOK);
390
391 nc = sc->n_components;
392 sc->components[nc].gcons = cp;
393 sc->components[nc].sc = sc;
394 sc->components[nc].index = nc;
395 sc->components[nc].chunk_count = cp->provider->mediasize /
396 sc->chunk_size;
397 sc->components[nc].chunk_next = 0;
398 sc->components[nc].chunk_reserved = 0;
399
400 if (sc->components[nc].chunk_count < 4) {
401 gctl_error(req, "Provider too small: %s",
402 cp->provider->name);
403 g_destroy_consumer(cp);
404 g_topology_unlock();
405 return;
406 }
407 fill_metadata(sc, &md, nc, *hardcode);
408 write_metadata(cp, &md);
409 /* The new component becomes visible when n_components is
410 * incremented */
411 sc->n_components++;
412 added++;
413
414 }
415 /* This call to update_metadata() is critical. In case there's a
416 * power failure in the middle of it and some components are updated
417 * while others are not, there will be trouble on next .taste() iff
418 * a non-updated component is detected first */
419 update_metadata(sc);
420 g_topology_unlock();
421 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
422 sc->geom->name);
423 /* Fire off BIOs previously queued because there wasn't any
424 * physical space left. If the BIOs still can't be satisfied
425 * they will again be added to the end of the queue (during
426 * which the mutex will be recursed) */
427 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
428 bq->bio = NULL;
429 mtx_lock(&sc->delayed_bio_q_mtx);
430 /* First, insert a sentinel to the queue end, so we don't
431 * end up in an infinite loop if there's still no free
432 * space available. */
433 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
434 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
435 bq = STAILQ_FIRST(&sc->delayed_bio_q);
436 if (bq->bio != NULL) {
437 g_virstor_start(bq->bio);
438 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
439 free(bq, M_GVIRSTOR);
440 } else {
441 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
442 free(bq, M_GVIRSTOR);
443 break;
444 }
445 }
446 mtx_unlock(&sc->delayed_bio_q_mtx);
447
448 }
449
450 /*
451 * Find a geom handled by the class
452 */
453 static struct g_virstor_softc *
454 virstor_find_geom(const struct g_class *cp, const char *name)
455 {
456 struct g_geom *gp;
457
458 LIST_FOREACH(gp, &cp->geom, geom) {
459 if (strcmp(name, gp->name) == 0)
460 return (gp->softc);
461 }
462 return (NULL);
463 }
464
465 /*
466 * Update metadata on all components to reflect the current state
467 * of these fields:
468 * - chunk_next
469 * - flags
470 * - md_count
471 * Expects things to be set up so write_metadata() can work, i.e.
472 * the topology lock must be held.
473 */
474 static void
475 update_metadata(struct g_virstor_softc *sc)
476 {
477 struct g_virstor_metadata md;
478 int n;
479
480 if (virstor_valid_components(sc) != sc->n_components)
481 return; /* Incomplete device */
482 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
483 sc->geom->name);
484 /* Update metadata on components */
485 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
486 sc->geom->class->name, sc->geom->name);
487 g_topology_assert();
488 for (n = 0; n < sc->n_components; n++) {
489 read_metadata(sc->components[n].gcons, &md);
490 md.chunk_next = sc->components[n].chunk_next;
491 md.flags = sc->components[n].flags;
492 md.md_count = sc->n_components;
493 write_metadata(sc->components[n].gcons, &md);
494 }
495 }
496
497 /*
498 * Fills metadata (struct md) from information stored in softc and the nc'th
499 * component of virstor
500 */
501 static void
502 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
503 u_int nc, u_int hardcode)
504 {
505 struct g_virstor_component *c;
506
507 bzero(md, sizeof *md);
508 c = &sc->components[nc];
509
510 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
511 md->md_version = G_VIRSTOR_VERSION;
512 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
513 md->md_id = sc->id;
514 md->md_virsize = sc->virsize;
515 md->md_chunk_size = sc->chunk_size;
516 md->md_count = sc->n_components;
517
518 if (hardcode) {
519 strncpy(md->provider, c->gcons->provider->name,
520 sizeof md->provider);
521 }
522 md->no = nc;
523 md->provsize = c->gcons->provider->mediasize;
524 md->chunk_count = c->chunk_count;
525 md->chunk_next = c->chunk_next;
526 md->chunk_reserved = c->chunk_reserved;
527 md->flags = c->flags;
528 }
529
530 /*
531 * Remove a component from virstor device.
532 * Can only be done if the component is unallocated.
533 */
534 static void
535 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
536 {
537 /* As this is executed in parallel to I/O, operations on virstor
538 * structures must be as atomic as possible. */
539 struct g_virstor_softc *sc;
540 int *nargs;
541 const char *geom_name;
542 u_int removed;
543 int i;
544
545 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
546 if (nargs == NULL) {
547 gctl_error(req, "Error fetching argument '%s'", "nargs");
548 return;
549 }
550 if (*nargs < 2) {
551 gctl_error(req, "Invalid number of arguments");
552 return;
553 }
554 /* Find "our" geom */
555 geom_name = gctl_get_asciiparam(req, "arg0");
556 if (geom_name == NULL) {
557 gctl_error(req, "Error fetching argument '%s'",
558 "geom_name (arg0)");
559 return;
560 }
561 sc = virstor_find_geom(cp, geom_name);
562 if (sc == NULL) {
563 gctl_error(req, "Don't know anything about '%s'", geom_name);
564 return;
565 }
566
567 if (virstor_valid_components(sc) != sc->n_components) {
568 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
569 "virstor %s", sc->geom->name);
570 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
571 return;
572 }
573
574 removed = 0;
575 for (i = 1; i < *nargs; i++) {
576 char param[8];
577 const char *prov_name;
578 int j, found;
579 struct g_virstor_component *newcomp, *compbak;
580
581 sprintf(param, "arg%d", i);
582 prov_name = gctl_get_asciiparam(req, param);
583 if (prov_name == NULL) {
584 gctl_error(req, "Error fetching argument '%s'", param);
585 return;
586 }
587 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
588 prov_name += sizeof(_PATH_DEV) - 1;
589
590 found = -1;
591 for (j = 0; j < sc->n_components; j++) {
592 if (strcmp(sc->components[j].gcons->provider->name,
593 prov_name) == 0) {
594 found = j;
595 break;
596 }
597 }
598 if (found == -1) {
599 LOG_MSG(LVL_ERROR, "No %s component in %s",
600 prov_name, sc->geom->name);
601 continue;
602 }
603
604 compbak = sc->components;
605 newcomp = malloc(sc->n_components * sizeof(*sc->components),
606 M_GVIRSTOR, M_WAITOK | M_ZERO);
607 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
608 bcopy(&sc->components[found + 1], newcomp + found,
609 found * sizeof(*sc->components));
610 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
611 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
612 "removed from %s",
613 prov_name, sc->geom->name);
614 free(newcomp, M_GVIRSTOR);
615 /* We'll consider this non-fatal error */
616 continue;
617 }
618 /* Renumerate unallocated components */
619 for (j = 0; j < sc->n_components-1; j++) {
620 if ((sc->components[j].flags &
621 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
622 sc->components[j].index = j;
623 }
624 }
625 /* This is the critical section. If a component allocation
626 * event happens while both variables are not yet set,
627 * there will be trouble. Something will panic on encountering
628 * NULL sc->components[x].gcomp member.
629 * Luckily, component allocation happens very rarely and
630 * removing components is an abnormal action in any case. */
631 sc->components = newcomp;
632 sc->n_components--;
633 /* End critical section */
634
635 g_topology_lock();
636 if (clear_metadata(&compbak[found]) != 0) {
637 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
638 "metadata on %s", prov_name);
639 }
640 g_detach(compbak[found].gcons);
641 g_destroy_consumer(compbak[found].gcons);
642 g_topology_unlock();
643
644 free(compbak, M_GVIRSTOR);
645
646 removed++;
647 }
648
649 /* This call to update_metadata() is critical. In case there's a
650 * power failure in the middle of it and some components are updated
651 * while others are not, there will be trouble on next .taste() iff
652 * a non-updated component is detected first */
653 g_topology_lock();
654 update_metadata(sc);
655 g_topology_unlock();
656 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
657 sc->geom->name);
658 }
659
660 /*
661 * Clear metadata sector on component
662 */
663 static int
664 clear_metadata(struct g_virstor_component *comp)
665 {
666 char *buf;
667 int error;
668
669 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
670 comp->gcons->provider->name);
671 g_topology_assert();
672 error = g_access(comp->gcons, 0, 1, 0);
673 if (error != 0)
674 return (error);
675 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
676 M_WAITOK | M_ZERO);
677 error = g_write_data(comp->gcons,
678 comp->gcons->provider->mediasize -
679 comp->gcons->provider->sectorsize,
680 buf,
681 comp->gcons->provider->sectorsize);
682 free(buf, M_GVIRSTOR);
683 g_access(comp->gcons, 0, -1, 0);
684 return (error);
685 }
686
687 /*
688 * Destroy geom forcibly.
689 */
690 static int
691 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
692 struct g_geom *gp)
693 {
694 struct g_virstor_softc *sc;
695 int exitval;
696
697 sc = gp->softc;
698 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
699
700 exitval = 0;
701 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
702 gp->softc);
703
704 if (sc != NULL) {
705 #ifdef INVARIANTS
706 char *buf;
707 int error;
708 off_t off;
709 int isclean, count;
710 int n;
711
712 LOG_MSG(LVL_INFO, "INVARIANTS detected");
713 LOG_MSG(LVL_INFO, "Verifying allocation "
714 "table for %s", sc->geom->name);
715 count = 0;
716 for (n = 0; n < sc->chunk_count; n++) {
717 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
718 count++;
719 }
720 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
721 sc->geom->name, count);
722 n = off = count = 0;
723 isclean = 1;
724 if (virstor_valid_components(sc) != sc->n_components) {
725 /* This is a incomplete virstor device (not all
726 * components have been found) */
727 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
728 sc->geom->name);
729 goto bailout;
730 }
731 error = g_access(sc->components[0].gcons, 1, 0, 0);
732 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
733 error));
734 /* Compare the whole on-disk allocation table with what's
735 * currently in memory */
736 while (n < sc->chunk_count) {
737 buf = g_read_data(sc->components[0].gcons, off,
738 sc->sectorsize, &error);
739 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
740 "for read at %jd", error, off));
741 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
742 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
743 "entry %d, offset %jd", n, off);
744 isclean = 0;
745 count++;
746 }
747 n += sc->me_per_sector;
748 off += sc->sectorsize;
749 g_free(buf);
750 }
751 error = g_access(sc->components[0].gcons, -1, 0, 0);
752 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
753 __func__, error));
754 if (isclean != 1) {
755 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
756 "(%d sectors don't match, max %zu allocations)",
757 sc->geom->name, count,
758 count * sc->me_per_sector);
759 } else {
760 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
761 sc->geom->name);
762 }
763 bailout:
764 #endif
765 update_metadata(sc);
766 virstor_geom_destroy(sc, FALSE, FALSE);
767 exitval = EAGAIN;
768 } else
769 exitval = 0;
770 return (exitval);
771 }
772
773 /*
774 * Taste event (per-class callback)
775 * Examines a provider and creates geom instances if needed
776 */
777 static struct g_geom *
778 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
779 {
780 struct g_virstor_metadata md;
781 struct g_geom *gp;
782 struct g_consumer *cp;
783 struct g_virstor_softc *sc;
784 int error;
785
786 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
787 g_topology_assert();
788 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
789
790 /* We need a dummy geom to attach a consumer to the given provider */
791 gp = g_new_geomf(mp, "virstor:taste.helper");
792 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
793 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
794 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
795
796 cp = g_new_consumer(gp);
797 g_attach(cp, pp);
798 error = read_metadata(cp, &md);
799 g_detach(cp);
800 g_destroy_consumer(cp);
801 g_destroy_geom(gp);
802
803 if (error != 0)
804 return (NULL);
805
806 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
807 return (NULL);
808 if (md.md_version != G_VIRSTOR_VERSION) {
809 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
810 "to handle %s (%s) : %d should be %d",
811 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
812 return (NULL);
813 }
814 if (md.provsize != pp->mediasize)
815 return (NULL);
816
817 /* If the provider name is hardcoded, use the offered provider only
818 * if it's been offered with its proper name (the one used in
819 * the label command). */
820 if (md.provider[0] != '\0' &&
821 !g_compare_names(md.provider, pp->name))
822 return (NULL);
823
824 /* Iterate all geoms this class already knows about to see if a new
825 * geom instance of this class needs to be created (in case the provider
826 * is first from a (possibly) multi-consumer geom) or it just needs
827 * to be added to an existing instance. */
828 sc = NULL;
829 gp = NULL;
830 LIST_FOREACH(gp, &mp->geom, geom) {
831 sc = gp->softc;
832 if (sc == NULL)
833 continue;
834 if (strcmp(md.md_name, sc->geom->name) != 0)
835 continue;
836 if (md.md_id != sc->id)
837 continue;
838 break;
839 }
840 if (gp != NULL) { /* We found an existing geom instance; add to it */
841 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
842 error = add_provider_to_geom(sc, pp, &md);
843 if (error != 0) {
844 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
845 pp->name, md.md_name, error);
846 return (NULL);
847 }
848 } else { /* New geom instance needs to be created */
849 gp = create_virstor_geom(mp, &md);
850 if (gp == NULL) {
851 LOG_MSG(LVL_ERROR, "Error creating new instance of "
852 "class %s: %s", mp->name, md.md_name);
853 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
854 md.md_name, pp->name);
855 return (NULL);
856 }
857 sc = gp->softc;
858 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
859 md.md_name);
860 error = add_provider_to_geom(sc, pp, &md);
861 if (error != 0) {
862 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
863 pp->name, md.md_name, error);
864 virstor_geom_destroy(sc, TRUE, FALSE);
865 return (NULL);
866 }
867 }
868
869 return (gp);
870 }
871
872 /*
873 * Destroyes consumer passed to it in arguments. Used as a callback
874 * on g_event queue.
875 */
876 static void
877 delay_destroy_consumer(void *arg, int flags __unused)
878 {
879 struct g_consumer *c = arg;
880 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
881 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
882 c->provider->name);
883 g_detach(c);
884 g_destroy_consumer(c);
885 }
886
887 /*
888 * Remove a component (consumer) from geom instance; If it's the first
889 * component being removed, orphan the provider to announce geom's being
890 * dismantled
891 */
892 static void
893 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
894 boolean_t delay)
895 {
896 struct g_consumer *c;
897
898 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
899 sc->geom->name));
900 c = comp->gcons;
901
902 comp->gcons = NULL;
903 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
904 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
905 sc->geom->name);
906 if (sc->provider != NULL) {
907 /* Whither, GEOM? */
908 sc->provider->flags |= G_PF_WITHER;
909 g_orphan_provider(sc->provider, ENXIO);
910 sc->provider = NULL;
911 LOG_MSG(LVL_INFO, "Removing provider %s", sc->geom->name);
912 }
913
914 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
915 g_access(c, -c->acr, -c->acw, -c->ace);
916 if (delay) {
917 /* Destroy consumer after it's tasted */
918 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
919 } else {
920 g_detach(c);
921 g_destroy_consumer(c);
922 }
923 }
924
925 /*
926 * Destroy geom - called internally
927 * See g_virstor_destroy_geom for the other one
928 */
929 static int
930 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
931 boolean_t delay)
932 {
933 struct g_provider *pp;
934 struct g_geom *gp;
935 int n;
936
937 g_topology_assert();
938
939 if (sc == NULL)
940 return (ENXIO);
941
942 pp = sc->provider;
943 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
944 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
945 "Device %s is still open.", pp->name);
946 if (!force)
947 return (EBUSY);
948 }
949
950 for (n = 0; n < sc->n_components; n++) {
951 if (sc->components[n].gcons != NULL)
952 remove_component(sc, &sc->components[n], delay);
953 }
954
955 gp = sc->geom;
956 gp->softc = NULL;
957
958 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
959 gp->name));
960
961 /* XXX: This might or might not work, since we're called with
962 * the topology lock held. Also, it might panic the kernel if
963 * the error'd BIO is in softupdates code. */
964 mtx_lock(&sc->delayed_bio_q_mtx);
965 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
966 struct g_virstor_bio_q *bq;
967 bq = STAILQ_FIRST(&sc->delayed_bio_q);
968 bq->bio->bio_error = ENOSPC;
969 g_io_deliver(bq->bio, EIO);
970 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
971 free(bq, M_GVIRSTOR);
972 }
973 mtx_unlock(&sc->delayed_bio_q_mtx);
974 mtx_destroy(&sc->delayed_bio_q_mtx);
975
976 free(sc->map, M_GVIRSTOR);
977 free(sc->components, M_GVIRSTOR);
978 bzero(sc, sizeof *sc);
979 free(sc, M_GVIRSTOR);
980
981 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
982 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
983 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
984
985 g_wither_geom(gp, ENXIO);
986
987 return (0);
988 }
989
990 /*
991 * Utility function: read metadata & decode. Wants topology lock to be
992 * held.
993 */
994 static int
995 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
996 {
997 struct g_provider *pp;
998 char *buf;
999 int error;
1000
1001 g_topology_assert();
1002 error = g_access(cp, 1, 0, 0);
1003 if (error != 0)
1004 return (error);
1005 pp = cp->provider;
1006 g_topology_unlock();
1007 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
1008 &error);
1009 g_topology_lock();
1010 g_access(cp, -1, 0, 0);
1011 if (buf == NULL)
1012 return (error);
1013
1014 virstor_metadata_decode(buf, md);
1015 g_free(buf);
1016
1017 return (0);
1018 }
1019
1020 /**
1021 * Utility function: encode & write metadata. Assumes topology lock is
1022 * held.
1023 *
1024 * There is no useful way of recovering from errors in this function,
1025 * not involving panicking the kernel. If the metadata cannot be written
1026 * the most we can do is notify the operator and hope he spots it and
1027 * replaces the broken drive.
1028 */
1029 static void
1030 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1031 {
1032 struct g_provider *pp;
1033 char *buf;
1034 int error;
1035
1036 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1037 ("Something's fishy in %s", __func__));
1038 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1039 g_topology_assert();
1040 error = g_access(cp, 0, 1, 0);
1041 if (error != 0) {
1042 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d",
1043 cp->provider->name, error);
1044 return;
1045 }
1046 pp = cp->provider;
1047
1048 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1049 virstor_metadata_encode(md, buf);
1050 g_topology_unlock();
1051 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1052 pp->sectorsize);
1053 g_topology_lock();
1054 g_access(cp, 0, -1, 0);
1055 free(buf, M_GVIRSTOR);
1056
1057 if (error != 0)
1058 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s",
1059 error, cp->provider->name);
1060 }
1061
1062 /*
1063 * Creates a new instance of this GEOM class, initialise softc
1064 */
1065 static struct g_geom *
1066 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1067 {
1068 struct g_geom *gp;
1069 struct g_virstor_softc *sc;
1070
1071 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1072 md->md_name, md->md_id);
1073
1074 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1075 md->md_virsize < md->md_chunk_size) {
1076 /* This is bogus configuration, and probably means data is
1077 * somehow corrupted. Panic, maybe? */
1078 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1079 md->md_name);
1080 return (NULL);
1081 }
1082
1083 /* Check if it's already created */
1084 LIST_FOREACH(gp, &mp->geom, geom) {
1085 sc = gp->softc;
1086 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1087 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1088 md->md_name);
1089 if (sc->id != md->md_id) {
1090 LOG_MSG(LVL_ERROR,
1091 "Some stale or invalid components "
1092 "exist for virstor device named %s. "
1093 "You will need to <CLEAR> all stale "
1094 "components and maybe reconfigure "
1095 "the virstor device. Tune "
1096 "kern.geom.virstor.debug sysctl up "
1097 "for more information.",
1098 sc->geom->name);
1099 }
1100 return (NULL);
1101 }
1102 }
1103 gp = g_new_geomf(mp, "%s", md->md_name);
1104 gp->softc = NULL; /* to circumevent races that test softc */
1105
1106 gp->start = g_virstor_start;
1107 gp->spoiled = g_virstor_orphan;
1108 gp->orphan = g_virstor_orphan;
1109 gp->access = g_virstor_access;
1110 gp->dumpconf = g_virstor_dumpconf;
1111
1112 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1113 sc->id = md->md_id;
1114 sc->n_components = md->md_count;
1115 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1116 M_GVIRSTOR, M_WAITOK | M_ZERO);
1117 sc->chunk_size = md->md_chunk_size;
1118 sc->virsize = md->md_virsize;
1119 STAILQ_INIT(&sc->delayed_bio_q);
1120 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1121 "gvirstor", MTX_DEF | MTX_RECURSE);
1122
1123 sc->geom = gp;
1124 sc->provider = NULL; /* virstor_check_and_run will create it */
1125 gp->softc = sc;
1126
1127 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1128
1129 return (gp);
1130 }
1131
1132 /*
1133 * Add provider to a GEOM class instance
1134 */
1135 static int
1136 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1137 struct g_virstor_metadata *md)
1138 {
1139 struct g_virstor_component *component;
1140 struct g_consumer *cp, *fcp;
1141 struct g_geom *gp;
1142 int error;
1143
1144 if (md->no >= sc->n_components)
1145 return (EINVAL);
1146
1147 /* "Current" compontent */
1148 component = &(sc->components[md->no]);
1149 if (component->gcons != NULL)
1150 return (EEXIST);
1151
1152 gp = sc->geom;
1153 fcp = LIST_FIRST(&gp->consumer);
1154
1155 cp = g_new_consumer(gp);
1156 error = g_attach(cp, pp);
1157
1158 if (error != 0) {
1159 g_destroy_consumer(cp);
1160 return (error);
1161 }
1162
1163 if (fcp != NULL) {
1164 if (fcp->provider->sectorsize != pp->sectorsize) {
1165 /* TODO: this can be made to work */
1166 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1167 "sector size (%d)", pp->name, sc->geom->name,
1168 pp->sectorsize);
1169 return (EINVAL);
1170 }
1171 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1172 /* Replicate access permissions from first "live" consumer
1173 * to the new one */
1174 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1175 if (error != 0) {
1176 g_detach(cp);
1177 g_destroy_consumer(cp);
1178 return (error);
1179 }
1180 }
1181 }
1182
1183 /* Bring up a new component */
1184 cp->private = component;
1185 component->gcons = cp;
1186 component->sc = sc;
1187 component->index = md->no;
1188 component->chunk_count = md->chunk_count;
1189 component->chunk_next = md->chunk_next;
1190 component->chunk_reserved = md->chunk_reserved;
1191 component->flags = md->flags;
1192
1193 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1194
1195 virstor_check_and_run(sc);
1196 return (0);
1197 }
1198
1199 /*
1200 * Check if everything's ready to create the geom provider & device entry,
1201 * create and start provider.
1202 * Called ultimately by .taste, from g_event thread
1203 */
1204 static void
1205 virstor_check_and_run(struct g_virstor_softc *sc)
1206 {
1207 off_t off;
1208 size_t n, count;
1209 int index;
1210 int error;
1211
1212 if (virstor_valid_components(sc) != sc->n_components)
1213 return;
1214
1215 if (virstor_valid_components(sc) == 0) {
1216 /* This is actually a candidate for panic() */
1217 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1218 sc->provider->name);
1219 return;
1220 }
1221
1222 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1223
1224 /* Initialise allocation map from the first consumer */
1225 sc->chunk_count = sc->virsize / sc->chunk_size;
1226 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1227 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1228 sc->provider->name,
1229 sc->chunk_count * (off_t)sc->chunk_size);
1230 }
1231 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1232 /* The following allocation is in order of 4MB - 8MB */
1233 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1234 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1235 __func__, sc->map_size, sc->provider->name));
1236 sc->map_sectors = sc->map_size / sc->sectorsize;
1237
1238 count = 0;
1239 for (n = 0; n < sc->n_components; n++)
1240 count += sc->components[n].chunk_count;
1241 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1242 "(%zu KB chunks)",
1243 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1244
1245 error = g_access(sc->components[0].gcons, 1, 0, 0);
1246 if (error != 0) {
1247 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1248 "read allocation map for %s",
1249 sc->components[0].gcons->provider->name,
1250 sc->geom->name);
1251 return;
1252 }
1253 /* Read in the allocation map */
1254 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1255 sc->components[0].gcons->provider->name);
1256 off = count = n = 0;
1257 while (count < sc->map_size) {
1258 struct g_virstor_map_entry *mapbuf;
1259 size_t bs;
1260
1261 bs = MIN(MAXPHYS, sc->map_size - count);
1262 if (bs % sc->sectorsize != 0) {
1263 /* Check for alignment errors */
1264 bs = (bs / sc->sectorsize) * sc->sectorsize;
1265 if (bs == 0)
1266 break;
1267 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1268 "for %s on %s", sc->geom->name,
1269 sc->components[0].gcons->provider->name);
1270 }
1271 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1272 if (mapbuf == NULL) {
1273 free(sc->map, M_GVIRSTOR);
1274 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1275 "for %s from %s (offset %ju) (error %d)",
1276 sc->geom->name,
1277 sc->components[0].gcons->provider->name,
1278 off, error);
1279 return;
1280 }
1281
1282 bcopy(mapbuf, &sc->map[n], bs);
1283 off += bs;
1284 count += bs;
1285 n += bs / sizeof *(sc->map);
1286 g_free(mapbuf);
1287 }
1288 g_access(sc->components[0].gcons, -1, 0, 0);
1289 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1290
1291 /* find first component with allocatable chunks */
1292 index = -1;
1293 for (n = 0; n < sc->n_components; n++) {
1294 if (sc->components[n].chunk_next <
1295 sc->components[n].chunk_count) {
1296 index = n;
1297 break;
1298 }
1299 }
1300 if (index == -1)
1301 /* not found? set it to the last component and handle it
1302 * later */
1303 index = sc->n_components - 1;
1304
1305 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1306 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1307 "(%d/%u: %s)", sc->geom->name,
1308 index+1,
1309 sc->n_components,
1310 sc->components[index].gcons->provider->name);
1311 }
1312 sc->curr_component = index;
1313
1314 if (sc->components[index].chunk_next >=
1315 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1316 LOG_MSG(LVL_WARNING,
1317 "Component %s of %s is running out of free space "
1318 "(%u chunks left)",
1319 sc->components[index].gcons->provider->name,
1320 sc->geom->name, sc->components[index].chunk_count -
1321 sc->components[index].chunk_next);
1322 }
1323
1324 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1325 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1326 LOG_MSG(LVL_ERROR,
1327 "%s: Map entries don't fit exactly in a sector (%s)",
1328 __func__, sc->geom->name);
1329 return;
1330 }
1331
1332 /* Recalculate allocated chunks in components & at the same time
1333 * verify map data is sane. We could trust metadata on this, but
1334 * we want to make sure. */
1335 for (n = 0; n < sc->n_components; n++)
1336 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1337
1338 for (n = 0; n < sc->chunk_count; n++) {
1339 if (sc->map[n].provider_no >= sc->n_components ||
1340 sc->map[n].provider_chunk >=
1341 sc->components[sc->map[n].provider_no].chunk_count) {
1342 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1343 __func__, (u_int)n, sc->geom->name);
1344 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1345 " provider_chunk: %u, chunk_count: %u", __func__,
1346 sc->map[n].provider_no, sc->n_components,
1347 sc->map[n].provider_chunk,
1348 sc->components[sc->map[n].provider_no].chunk_count);
1349 return;
1350 }
1351 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1352 sc->components[sc->map[n].provider_no].chunk_next++;
1353 }
1354
1355 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1356 sc->geom->name);
1357
1358 sc->provider->sectorsize = sc->sectorsize;
1359 sc->provider->mediasize = sc->virsize;
1360 g_error_provider(sc->provider, 0);
1361
1362 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1363 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1364 "chunk %u", sc->provider->name, sc->curr_component,
1365 sc->components[sc->curr_component].chunk_next);
1366 }
1367
1368 /*
1369 * Returns count of active providers in this geom instance
1370 */
1371 static u_int
1372 virstor_valid_components(struct g_virstor_softc *sc)
1373 {
1374 unsigned int nc, i;
1375
1376 nc = 0;
1377 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1378 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1379 for (i = 0; i < sc->n_components; i++)
1380 if (sc->components[i].gcons != NULL)
1381 nc++;
1382 return (nc);
1383 }
1384
1385 /*
1386 * Called when the consumer gets orphaned (?)
1387 */
1388 static void
1389 g_virstor_orphan(struct g_consumer *cp)
1390 {
1391 struct g_virstor_softc *sc;
1392 struct g_virstor_component *comp;
1393 struct g_geom *gp;
1394
1395 g_topology_assert();
1396 gp = cp->geom;
1397 sc = gp->softc;
1398 if (sc == NULL)
1399 return;
1400
1401 comp = cp->private;
1402 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1403 __func__));
1404 remove_component(sc, comp, FALSE);
1405 if (virstor_valid_components(sc) == 0)
1406 virstor_geom_destroy(sc, TRUE, FALSE);
1407 }
1408
1409 /*
1410 * Called to notify geom when it's been opened, and for what intent
1411 */
1412 static int
1413 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1414 {
1415 struct g_consumer *c;
1416 struct g_virstor_softc *sc;
1417 struct g_geom *gp;
1418 int error;
1419
1420 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1421 gp = pp->geom;
1422 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1423 sc = gp->softc;
1424
1425 if (sc == NULL) {
1426 /* It seems that .access can be called with negative dr,dw,dx
1427 * in this case but I want to check for myself */
1428 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s",
1429 dr, dw, de, pp->name);
1430 /* This should only happen when geom is withered so
1431 * allow only negative requests */
1432 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
1433 ("%s: Positive access for %s", __func__, pp->name));
1434 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
1435 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed",
1436 pp->name);
1437 return (0);
1438 }
1439
1440 /* Grab an exclusive bit to propagate on our consumers on first open */
1441 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1442 de++;
1443 /* ... drop it on close */
1444 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1445 de--;
1446 update_metadata(sc); /* Writes statistical information */
1447 }
1448
1449 error = ENXIO;
1450 LIST_FOREACH(c, &gp->consumer, consumer) {
1451 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
1452 error = g_access(c, dr, dw, de);
1453 if (error != 0) {
1454 struct g_consumer *c2;
1455
1456 /* Backout earlier changes */
1457 LIST_FOREACH(c2, &gp->consumer, consumer) {
1458 if (c2 == c) /* all eariler components fixed */
1459 return (error);
1460 g_access(c2, -dr, -dw, -de);
1461 }
1462 }
1463 }
1464
1465 return (error);
1466 }
1467
1468 /*
1469 * Generate XML dump of current state
1470 */
1471 static void
1472 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1473 struct g_consumer *cp, struct g_provider *pp)
1474 {
1475 struct g_virstor_softc *sc;
1476
1477 g_topology_assert();
1478 sc = gp->softc;
1479
1480 if (sc == NULL || pp != NULL)
1481 return;
1482
1483 if (cp != NULL) {
1484 /* For each component */
1485 struct g_virstor_component *comp;
1486
1487 comp = cp->private;
1488 if (comp == NULL)
1489 return;
1490 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1491 indent, comp->index);
1492 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1493 indent, comp->chunk_count);
1494 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1495 indent, comp->chunk_next);
1496 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1497 indent, comp->chunk_reserved);
1498 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1499 indent,
1500 comp->chunk_next > 0 ? 100 -
1501 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1502 comp->chunk_count : 100);
1503 } else {
1504 /* For the whole thing */
1505 u_int count, used, i;
1506 off_t size;
1507
1508 count = used = size = 0;
1509 for (i = 0; i < sc->n_components; i++) {
1510 if (sc->components[i].gcons != NULL) {
1511 count += sc->components[i].chunk_count;
1512 used += sc->components[i].chunk_next +
1513 sc->components[i].chunk_reserved;
1514 size += sc->components[i].gcons->
1515 provider->mediasize;
1516 }
1517 }
1518
1519 sbuf_printf(sb, "%s<Status>"
1520 "Components=%u, Online=%u</Status>\n", indent,
1521 sc->n_components, virstor_valid_components(sc));
1522 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1523 indent, 100-(used * 100) / count);
1524 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1525 sc->chunk_size);
1526 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1527 indent, used > 0 ? 100 - (used * 100) / count : 100);
1528 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1529 indent, count);
1530 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1531 indent, sc->chunk_count);
1532 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1533 indent,
1534 (count * 100) / sc->chunk_count);
1535 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1536 indent, size);
1537 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1538 sc->virsize);
1539 }
1540 }
1541
1542 /*
1543 * GEOM .done handler
1544 * Can't use standard handler because one requested IO may
1545 * fork into additional data IOs
1546 */
1547 static void
1548 g_virstor_done(struct bio *b)
1549 {
1550 struct g_virstor_softc *sc;
1551 struct bio *parent_b;
1552
1553 parent_b = b->bio_parent;
1554 sc = parent_b->bio_to->geom->softc;
1555
1556 if (b->bio_error != 0) {
1557 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1558 b->bio_error, b->bio_offset, b->bio_length,
1559 b->bio_to->name);
1560 if (parent_b->bio_error == 0)
1561 parent_b->bio_error = b->bio_error;
1562 }
1563
1564 parent_b->bio_inbed++;
1565 parent_b->bio_completed += b->bio_completed;
1566
1567 if (parent_b->bio_children == parent_b->bio_inbed) {
1568 parent_b->bio_completed = parent_b->bio_length;
1569 g_io_deliver(parent_b, parent_b->bio_error);
1570 }
1571 g_destroy_bio(b);
1572 }
1573
1574 /*
1575 * I/O starts here
1576 * Called in g_down thread
1577 */
1578 static void
1579 g_virstor_start(struct bio *b)
1580 {
1581 struct g_virstor_softc *sc;
1582 struct g_virstor_component *comp;
1583 struct bio *cb;
1584 struct g_provider *pp;
1585 char *addr;
1586 off_t offset, length;
1587 struct bio_queue_head bq;
1588 size_t chunk_size; /* cached for convenience */
1589 u_int count;
1590
1591 pp = b->bio_to;
1592 sc = pp->geom->softc;
1593 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1594 b->bio_to->error, b->bio_to->name));
1595
1596 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1597
1598 switch (b->bio_cmd) {
1599 case BIO_READ:
1600 case BIO_WRITE:
1601 case BIO_DELETE:
1602 break;
1603 default:
1604 g_io_deliver(b, EOPNOTSUPP);
1605 return;
1606 }
1607
1608 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1609 bioq_init(&bq);
1610
1611 chunk_size = sc->chunk_size;
1612 addr = b->bio_data;
1613 offset = b->bio_offset; /* virtual offset and length */
1614 length = b->bio_length;
1615
1616 while (length > 0) {
1617 size_t chunk_index, in_chunk_offset, in_chunk_length;
1618 struct virstor_map_entry *me;
1619
1620 chunk_index = offset / chunk_size; /* round downwards */
1621 in_chunk_offset = offset % chunk_size;
1622 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1623 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1624 b->bio_cmd == BIO_READ ? "R" : "W",
1625 offset, length,
1626 chunk_index, in_chunk_offset, in_chunk_length);
1627 me = &sc->map[chunk_index];
1628
1629 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1630 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1631 /* Reads from unallocated chunks return zeroed
1632 * buffers */
1633 if (b->bio_cmd == BIO_READ)
1634 bzero(addr, in_chunk_length);
1635 } else {
1636 comp = &sc->components[me->provider_no];
1637
1638 cb = g_clone_bio(b);
1639 if (cb == NULL) {
1640 bioq_dismantle(&bq);
1641 if (b->bio_error == 0)
1642 b->bio_error = ENOMEM;
1643 g_io_deliver(b, b->bio_error);
1644 return;
1645 }
1646 cb->bio_to = comp->gcons->provider;
1647 cb->bio_done = g_virstor_done;
1648 cb->bio_offset =
1649 (off_t)me->provider_chunk * (off_t)chunk_size
1650 + in_chunk_offset;
1651 cb->bio_length = in_chunk_length;
1652 cb->bio_data = addr;
1653 cb->bio_caller1 = comp;
1654 bioq_disksort(&bq, cb);
1655 }
1656 } else { /* handle BIO_WRITE */
1657 KASSERT(b->bio_cmd == BIO_WRITE,
1658 ("%s: Unknown command %d", __func__,
1659 b->bio_cmd));
1660
1661 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1662 /* We have a virtual chunk, represented by
1663 * the "me" entry, but it's not yet allocated
1664 * (tied to) a physical chunk. So do it now. */
1665 struct virstor_map_entry *data_me;
1666 u_int phys_chunk, comp_no;
1667 off_t s_offset;
1668 int error;
1669
1670 error = allocate_chunk(sc, &comp, &comp_no,
1671 &phys_chunk);
1672 if (error != 0) {
1673 /* We cannot allocate a physical chunk
1674 * to satisfy this request, so we'll
1675 * delay it to when we can...
1676 * XXX: this will prevent the fs from
1677 * being umounted! */
1678 struct g_virstor_bio_q *biq;
1679 biq = malloc(sizeof *biq, M_GVIRSTOR,
1680 M_NOWAIT);
1681 if (biq == NULL) {
1682 bioq_dismantle(&bq);
1683 if (b->bio_error == 0)
1684 b->bio_error = ENOMEM;
1685 g_io_deliver(b, b->bio_error);
1686 return;
1687 }
1688 biq->bio = b;
1689 mtx_lock(&sc->delayed_bio_q_mtx);
1690 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1691 biq, linkage);
1692 mtx_unlock(&sc->delayed_bio_q_mtx);
1693 LOG_MSG(LVL_WARNING, "Delaying BIO "
1694 "(size=%ju) until free physical "
1695 "space can be found on %s",
1696 b->bio_length,
1697 sc->provider->name);
1698 return;
1699 }
1700 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1701 "for %s",
1702 phys_chunk,
1703 comp->gcons->provider->name,
1704 sc->provider->name);
1705
1706 me->provider_no = comp_no;
1707 me->provider_chunk = phys_chunk;
1708 me->flags |= VIRSTOR_MAP_ALLOCATED;
1709
1710 cb = g_clone_bio(b);
1711 if (cb == NULL) {
1712 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1713 me->provider_no = 0;
1714 me->provider_chunk = 0;
1715 bioq_dismantle(&bq);
1716 if (b->bio_error == 0)
1717 b->bio_error = ENOMEM;
1718 g_io_deliver(b, b->bio_error);
1719 return;
1720 }
1721
1722 /* The allocation table is stored continuously
1723 * at the start of the drive. We need to
1724 * calculate the offset of the sector that holds
1725 * this map entry both on the drive and in the
1726 * map array.
1727 * sc_offset will end up pointing to the drive
1728 * sector. */
1729 s_offset = chunk_index * sizeof *me;
1730 s_offset = (s_offset / sc->sectorsize) *
1731 sc->sectorsize;
1732
1733 /* data_me points to map entry sector
1734 * in memory (analoguos to offset) */
1735 data_me = &sc->map[(chunk_index /
1736 sc->me_per_sector) * sc->me_per_sector];
1737
1738 /* Commit sector with map entry to storage */
1739 cb->bio_to = sc->components[0].gcons->provider;
1740 cb->bio_done = g_virstor_done;
1741 cb->bio_offset = s_offset;
1742 cb->bio_data = (char *)data_me;
1743 cb->bio_length = sc->sectorsize;
1744 cb->bio_caller1 = &sc->components[0];
1745 bioq_disksort(&bq, cb);
1746 }
1747
1748 comp = &sc->components[me->provider_no];
1749 cb = g_clone_bio(b);
1750 if (cb == NULL) {
1751 bioq_dismantle(&bq);
1752 if (b->bio_error == 0)
1753 b->bio_error = ENOMEM;
1754 g_io_deliver(b, b->bio_error);
1755 return;
1756 }
1757 /* Finally, handle the data */
1758 cb->bio_to = comp->gcons->provider;
1759 cb->bio_done = g_virstor_done;
1760 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1761 in_chunk_offset;
1762 cb->bio_length = in_chunk_length;
1763 cb->bio_data = addr;
1764 cb->bio_caller1 = comp;
1765 bioq_disksort(&bq, cb);
1766 }
1767 addr += in_chunk_length;
1768 length -= in_chunk_length;
1769 offset += in_chunk_length;
1770 }
1771
1772 /* Fire off bio's here */
1773 count = 0;
1774 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1775 bioq_remove(&bq, cb);
1776 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1777 comp = cb->bio_caller1;
1778 cb->bio_caller1 = NULL;
1779 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1780 cb->bio_offset, cb->bio_length);
1781 g_io_request(cb, comp->gcons);
1782 count++;
1783 }
1784 if (count == 0) { /* We handled everything locally */
1785 b->bio_completed = b->bio_length;
1786 g_io_deliver(b, 0);
1787 }
1788
1789 }
1790
1791 /*
1792 * Allocate a chunk from a physical provider. Returns physical component,
1793 * chunk index relative to the component and the component's index.
1794 */
1795 static int
1796 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1797 u_int *comp_no_p, u_int *chunk)
1798 {
1799 u_int comp_no;
1800
1801 KASSERT(sc->curr_component < sc->n_components,
1802 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1803
1804 comp_no = sc->curr_component;
1805 *comp = &sc->components[comp_no];
1806 dump_component(*comp);
1807 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1808 /* This component is full. Allocate next component */
1809 if (comp_no >= sc->n_components-1) {
1810 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1811 sc->geom->name);
1812 return (-1);
1813 }
1814 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1815 sc->curr_component = ++comp_no;
1816
1817 *comp = &sc->components[comp_no];
1818 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1819 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1820 "(switching to %u/%u: %s)", sc->geom->name,
1821 comp_no+1, sc->n_components,
1822 (*comp)->gcons->provider->name);
1823 /* Take care not to overwrite reserved chunks */
1824 if ( (*comp)->chunk_reserved > 0 &&
1825 (*comp)->chunk_next < (*comp)->chunk_reserved)
1826 (*comp)->chunk_next = (*comp)->chunk_reserved;
1827
1828 (*comp)->flags |=
1829 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1830 dump_component(*comp);
1831 *comp_no_p = comp_no;
1832 *chunk = (*comp)->chunk_next++;
1833 } else {
1834 *comp_no_p = comp_no;
1835 *chunk = (*comp)->chunk_next++;
1836 }
1837 return (0);
1838 }
1839
1840 /* Dump a component */
1841 static void
1842 dump_component(struct g_virstor_component *comp)
1843 {
1844
1845 if (g_virstor_debug < LVL_DEBUG2)
1846 return;
1847 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1848 printf(" chunk_count: %u\n", comp->chunk_count);
1849 printf(" chunk_next: %u\n", comp->chunk_next);
1850 printf(" flags: %u\n", comp->flags);
1851 }
1852
1853 #if 0
1854 /* Dump a map entry */
1855 static void
1856 dump_me(struct virstor_map_entry *me, unsigned int nr)
1857 {
1858 if (g_virstor_debug < LVL_DEBUG)
1859 return;
1860 printf("VIRT. CHUNK #%d: ", nr);
1861 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1862 printf("(unallocated)\n");
1863 else
1864 printf("allocated at provider %u, provider_chunk %u\n",
1865 me->provider_no, me->provider_chunk);
1866 }
1867 #endif
1868
1869 /*
1870 * Dismantle bio_queue and destroy its components
1871 */
1872 static void
1873 bioq_dismantle(struct bio_queue_head *bq)
1874 {
1875 struct bio *b;
1876
1877 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1878 bioq_remove(bq, b);
1879 g_destroy_bio(b);
1880 }
1881 }
1882
1883 /*
1884 * The function that shouldn't be called.
1885 * When this is called, the stack is already garbled because of
1886 * argument mismatch. There's nothing to do now but panic, which is
1887 * accidentally the whole purpose of this function.
1888 * Motivation: to guard from accidentally calling geom methods when
1889 * they shouldn't be called. (see g_..._taste)
1890 */
1891 static void
1892 invalid_call(void)
1893 {
1894 panic("invalid_call() has just been called. Something's fishy here.");
1895 }
1896
1897 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */
Cache object: 5ba1f71b99765121690b7dcb7396c0f8
|