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("$FreeBSD$");
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 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RWTUN, &g_virstor_debug,
88 0, "Debug level (2=production, 5=normal, 15=excessive)");
89
90 static u_int g_virstor_chunk_watermark = 100;
91 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RWTUN,
92 &g_virstor_chunk_watermark, 0,
93 "Minimum number of free chunks before issuing administrative warning");
94
95 static u_int g_virstor_component_watermark = 1;
96 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RWTUN,
97 &g_virstor_component_watermark, 0,
98 "Minimum number of free components before issuing administrative warning");
99
100 static int read_metadata(struct g_consumer *, struct g_virstor_metadata *);
101 static void write_metadata(struct g_consumer *, struct g_virstor_metadata *);
102 static int clear_metadata(struct g_virstor_component *);
103 static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *,
104 struct g_virstor_metadata *);
105 static struct g_geom *create_virstor_geom(struct g_class *,
106 struct g_virstor_metadata *);
107 static void virstor_check_and_run(struct g_virstor_softc *);
108 static u_int virstor_valid_components(struct g_virstor_softc *);
109 static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t,
110 boolean_t);
111 static void remove_component(struct g_virstor_softc *,
112 struct g_virstor_component *, boolean_t);
113 static void bioq_dismantle(struct bio_queue_head *);
114 static int allocate_chunk(struct g_virstor_softc *,
115 struct g_virstor_component **, u_int *, u_int *);
116 static void delay_destroy_consumer(void *, int);
117 static void dump_component(struct g_virstor_component *comp);
118 #if 0
119 static void dump_me(struct virstor_map_entry *me, unsigned int nr);
120 #endif
121
122 static void virstor_ctl_stop(struct gctl_req *, struct g_class *);
123 static void virstor_ctl_add(struct gctl_req *, struct g_class *);
124 static void virstor_ctl_remove(struct gctl_req *, struct g_class *);
125 static struct g_virstor_softc * virstor_find_geom(const struct g_class *,
126 const char *);
127 static void update_metadata(struct g_virstor_softc *);
128 static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *,
129 u_int, u_int);
130
131 static void g_virstor_orphan(struct g_consumer *);
132 static int g_virstor_access(struct g_provider *, int, int, int);
133 static void g_virstor_start(struct bio *);
134 static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *,
135 struct g_consumer *, struct g_provider *);
136 static void g_virstor_done(struct bio *);
137
138 static void invalid_call(void);
139 /*
140 * Initialise GEOM class (per-class callback)
141 */
142 static void
143 g_virstor_init(struct g_class *mp __unused)
144 {
145
146 /* Catch map struct size mismatch at compile time; Map entries must
147 * fit into MAXPHYS exactly, with no wasted space. */
148 CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS);
149
150 /* Init UMA zones, TAILQ's, other global vars */
151 }
152
153 /*
154 * Finalise GEOM class (per-class callback)
155 */
156 static void
157 g_virstor_fini(struct g_class *mp __unused)
158 {
159
160 /* Deinit UMA zones & global vars */
161 }
162
163 /*
164 * Config (per-class callback)
165 */
166 static void
167 g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb)
168 {
169 uint32_t *version;
170
171 g_topology_assert();
172
173 version = gctl_get_paraml(req, "version", sizeof(*version));
174 if (version == NULL) {
175 gctl_error(req, "Failed to get 'version' argument");
176 return;
177 }
178 if (*version != G_VIRSTOR_VERSION) {
179 gctl_error(req, "Userland and kernel versions out of sync");
180 return;
181 }
182
183 g_topology_unlock();
184 if (strcmp(verb, "add") == 0)
185 virstor_ctl_add(req, cp);
186 else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0)
187 virstor_ctl_stop(req, cp);
188 else if (strcmp(verb, "remove") == 0)
189 virstor_ctl_remove(req, cp);
190 else
191 gctl_error(req, "unknown verb: '%s'", verb);
192 g_topology_lock();
193 }
194
195 /*
196 * "stop" verb from userland
197 */
198 static void
199 virstor_ctl_stop(struct gctl_req *req, struct g_class *cp)
200 {
201 int *force, *nargs;
202 int i;
203
204 nargs = gctl_get_paraml(req, "nargs", sizeof *nargs);
205 if (nargs == NULL) {
206 gctl_error(req, "Error fetching argument '%s'", "nargs");
207 return;
208 }
209 if (*nargs < 1) {
210 gctl_error(req, "Invalid number of arguments");
211 return;
212 }
213 force = gctl_get_paraml(req, "force", sizeof *force);
214 if (force == NULL) {
215 gctl_error(req, "Error fetching argument '%s'", "force");
216 return;
217 }
218
219 g_topology_lock();
220 for (i = 0; i < *nargs; i++) {
221 char param[8];
222 const char *name;
223 struct g_virstor_softc *sc;
224 int error;
225
226 sprintf(param, "arg%d", i);
227 name = gctl_get_asciiparam(req, param);
228 if (name == NULL) {
229 gctl_error(req, "No 'arg%d' argument", i);
230 g_topology_unlock();
231 return;
232 }
233 sc = virstor_find_geom(cp, name);
234 if (sc == NULL) {
235 gctl_error(req, "Don't know anything about '%s'", name);
236 g_topology_unlock();
237 return;
238 }
239
240 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
241 sc->geom->name);
242 update_metadata(sc);
243 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
244 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
245 sc->geom->name, error);
246 }
247 }
248 g_topology_unlock();
249 }
250
251 /*
252 * "add" verb from userland - add new component(s) to the structure.
253 * This will be done all at once in here, without going through the
254 * .taste function for new components.
255 */
256 static void
257 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
258 {
259 /* Note: while this is going on, I/O is being done on
260 * the g_up and g_down threads. The idea is to make changes
261 * to softc members in a way that can atomically activate
262 * them all at once. */
263 struct g_virstor_softc *sc;
264 int *hardcode, *nargs;
265 const char *geom_name; /* geom to add a component to */
266 struct g_consumer *fcp;
267 struct g_virstor_bio_q *bq;
268 u_int added;
269 int error;
270 int i;
271
272 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
273 if (nargs == NULL) {
274 gctl_error(req, "Error fetching argument '%s'", "nargs");
275 return;
276 }
277 if (*nargs < 2) {
278 gctl_error(req, "Invalid number of arguments");
279 return;
280 }
281 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
282 if (hardcode == NULL) {
283 gctl_error(req, "Error fetching argument '%s'", "hardcode");
284 return;
285 }
286
287 /* Find "our" geom */
288 geom_name = gctl_get_asciiparam(req, "arg0");
289 if (geom_name == NULL) {
290 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
291 return;
292 }
293 sc = virstor_find_geom(cp, geom_name);
294 if (sc == NULL) {
295 gctl_error(req, "Don't know anything about '%s'", geom_name);
296 return;
297 }
298
299 if (virstor_valid_components(sc) != sc->n_components) {
300 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
301 "virstor %s", sc->geom->name);
302 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
303 return;
304 }
305
306 fcp = sc->components[0].gcons;
307 added = 0;
308 g_topology_lock();
309 for (i = 1; i < *nargs; i++) {
310 struct g_virstor_metadata md;
311 char aname[8];
312 const char *prov_name;
313 struct g_provider *pp;
314 struct g_consumer *cp;
315 u_int nc;
316 u_int j;
317
318 snprintf(aname, sizeof aname, "arg%d", i);
319 prov_name = gctl_get_asciiparam(req, aname);
320 if (prov_name == NULL) {
321 gctl_error(req, "Error fetching argument '%s'", aname);
322 g_topology_unlock();
323 return;
324 }
325 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
326 prov_name += sizeof(_PATH_DEV) - 1;
327
328 pp = g_provider_by_name(prov_name);
329 if (pp == NULL) {
330 /* This is the most common error so be verbose about it */
331 if (added != 0) {
332 gctl_error(req, "Invalid provider: '%s' (added"
333 " %u components)", prov_name, added);
334 update_metadata(sc);
335 } else {
336 gctl_error(req, "Invalid provider: '%s'",
337 prov_name);
338 }
339 g_topology_unlock();
340 return;
341 }
342 cp = g_new_consumer(sc->geom);
343 if (cp == NULL) {
344 gctl_error(req, "Cannot create consumer");
345 g_topology_unlock();
346 return;
347 }
348 error = g_attach(cp, pp);
349 if (error != 0) {
350 gctl_error(req, "Cannot attach a consumer to %s",
351 pp->name);
352 g_destroy_consumer(cp);
353 g_topology_unlock();
354 return;
355 }
356 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
357 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
358 if (error != 0) {
359 gctl_error(req, "Access request failed for %s",
360 pp->name);
361 g_destroy_consumer(cp);
362 g_topology_unlock();
363 return;
364 }
365 }
366 if (fcp->provider->sectorsize != pp->sectorsize) {
367 gctl_error(req, "Sector size doesn't fit for %s",
368 pp->name);
369 g_destroy_consumer(cp);
370 g_topology_unlock();
371 return;
372 }
373 for (j = 0; j < sc->n_components; j++) {
374 if (strcmp(sc->components[j].gcons->provider->name,
375 pp->name) == 0) {
376 gctl_error(req, "Component %s already in %s",
377 pp->name, sc->geom->name);
378 g_destroy_consumer(cp);
379 g_topology_unlock();
380 return;
381 }
382 }
383 sc->components = realloc(sc->components,
384 sizeof(*sc->components) * (sc->n_components + 1),
385 M_GVIRSTOR, M_WAITOK);
386
387 nc = sc->n_components;
388 sc->components[nc].gcons = cp;
389 sc->components[nc].sc = sc;
390 sc->components[nc].index = nc;
391 sc->components[nc].chunk_count = cp->provider->mediasize /
392 sc->chunk_size;
393 sc->components[nc].chunk_next = 0;
394 sc->components[nc].chunk_reserved = 0;
395
396 if (sc->components[nc].chunk_count < 4) {
397 gctl_error(req, "Provider too small: %s",
398 cp->provider->name);
399 g_destroy_consumer(cp);
400 g_topology_unlock();
401 return;
402 }
403 fill_metadata(sc, &md, nc, *hardcode);
404 write_metadata(cp, &md);
405 /* The new component becomes visible when n_components is
406 * incremented */
407 sc->n_components++;
408 added++;
409
410 }
411 /* This call to update_metadata() is critical. In case there's a
412 * power failure in the middle of it and some components are updated
413 * while others are not, there will be trouble on next .taste() iff
414 * a non-updated component is detected first */
415 update_metadata(sc);
416 g_topology_unlock();
417 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
418 sc->geom->name);
419 /* Fire off BIOs previously queued because there wasn't any
420 * physical space left. If the BIOs still can't be satisfied
421 * they will again be added to the end of the queue (during
422 * which the mutex will be recursed) */
423 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
424 bq->bio = NULL;
425 mtx_lock(&sc->delayed_bio_q_mtx);
426 /* First, insert a sentinel to the queue end, so we don't
427 * end up in an infinite loop if there's still no free
428 * space available. */
429 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
430 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
431 bq = STAILQ_FIRST(&sc->delayed_bio_q);
432 if (bq->bio != NULL) {
433 g_virstor_start(bq->bio);
434 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
435 free(bq, M_GVIRSTOR);
436 } else {
437 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
438 free(bq, M_GVIRSTOR);
439 break;
440 }
441 }
442 mtx_unlock(&sc->delayed_bio_q_mtx);
443
444 }
445
446 /*
447 * Find a geom handled by the class
448 */
449 static struct g_virstor_softc *
450 virstor_find_geom(const struct g_class *cp, const char *name)
451 {
452 struct g_geom *gp;
453
454 LIST_FOREACH(gp, &cp->geom, geom) {
455 if (strcmp(name, gp->name) == 0)
456 return (gp->softc);
457 }
458 return (NULL);
459 }
460
461 /*
462 * Update metadata on all components to reflect the current state
463 * of these fields:
464 * - chunk_next
465 * - flags
466 * - md_count
467 * Expects things to be set up so write_metadata() can work, i.e.
468 * the topology lock must be held.
469 */
470 static void
471 update_metadata(struct g_virstor_softc *sc)
472 {
473 struct g_virstor_metadata md;
474 u_int n;
475
476 if (virstor_valid_components(sc) != sc->n_components)
477 return; /* Incomplete device */
478 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
479 sc->geom->name);
480 /* Update metadata on components */
481 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
482 sc->geom->class->name, sc->geom->name);
483 g_topology_assert();
484 for (n = 0; n < sc->n_components; n++) {
485 read_metadata(sc->components[n].gcons, &md);
486 md.chunk_next = sc->components[n].chunk_next;
487 md.flags = sc->components[n].flags;
488 md.md_count = sc->n_components;
489 write_metadata(sc->components[n].gcons, &md);
490 }
491 }
492
493 /*
494 * Fills metadata (struct md) from information stored in softc and the nc'th
495 * component of virstor
496 */
497 static void
498 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
499 u_int nc, u_int hardcode)
500 {
501 struct g_virstor_component *c;
502
503 bzero(md, sizeof *md);
504 c = &sc->components[nc];
505
506 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
507 md->md_version = G_VIRSTOR_VERSION;
508 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
509 md->md_id = sc->id;
510 md->md_virsize = sc->virsize;
511 md->md_chunk_size = sc->chunk_size;
512 md->md_count = sc->n_components;
513
514 if (hardcode) {
515 strncpy(md->provider, c->gcons->provider->name,
516 sizeof md->provider);
517 }
518 md->no = nc;
519 md->provsize = c->gcons->provider->mediasize;
520 md->chunk_count = c->chunk_count;
521 md->chunk_next = c->chunk_next;
522 md->chunk_reserved = c->chunk_reserved;
523 md->flags = c->flags;
524 }
525
526 /*
527 * Remove a component from virstor device.
528 * Can only be done if the component is unallocated.
529 */
530 static void
531 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
532 {
533 /* As this is executed in parallel to I/O, operations on virstor
534 * structures must be as atomic as possible. */
535 struct g_virstor_softc *sc;
536 int *nargs;
537 const char *geom_name;
538 u_int removed;
539 int i;
540
541 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
542 if (nargs == NULL) {
543 gctl_error(req, "Error fetching argument '%s'", "nargs");
544 return;
545 }
546 if (*nargs < 2) {
547 gctl_error(req, "Invalid number of arguments");
548 return;
549 }
550 /* Find "our" geom */
551 geom_name = gctl_get_asciiparam(req, "arg0");
552 if (geom_name == NULL) {
553 gctl_error(req, "Error fetching argument '%s'",
554 "geom_name (arg0)");
555 return;
556 }
557 sc = virstor_find_geom(cp, geom_name);
558 if (sc == NULL) {
559 gctl_error(req, "Don't know anything about '%s'", geom_name);
560 return;
561 }
562
563 if (virstor_valid_components(sc) != sc->n_components) {
564 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
565 "virstor %s", sc->geom->name);
566 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
567 return;
568 }
569
570 removed = 0;
571 for (i = 1; i < *nargs; i++) {
572 char param[8];
573 const char *prov_name;
574 int j, found;
575 struct g_virstor_component *newcomp, *compbak;
576
577 sprintf(param, "arg%d", i);
578 prov_name = gctl_get_asciiparam(req, param);
579 if (prov_name == NULL) {
580 gctl_error(req, "Error fetching argument '%s'", param);
581 return;
582 }
583 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
584 prov_name += sizeof(_PATH_DEV) - 1;
585
586 found = -1;
587 for (j = 0; j < sc->n_components; j++) {
588 if (strcmp(sc->components[j].gcons->provider->name,
589 prov_name) == 0) {
590 found = j;
591 break;
592 }
593 }
594 if (found == -1) {
595 LOG_MSG(LVL_ERROR, "No %s component in %s",
596 prov_name, sc->geom->name);
597 continue;
598 }
599
600 compbak = sc->components;
601 newcomp = malloc(sc->n_components * sizeof(*sc->components),
602 M_GVIRSTOR, M_WAITOK | M_ZERO);
603 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
604 bcopy(&sc->components[found + 1], newcomp + found,
605 found * sizeof(*sc->components));
606 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
607 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
608 "removed from %s",
609 prov_name, sc->geom->name);
610 free(newcomp, M_GVIRSTOR);
611 /* We'll consider this non-fatal error */
612 continue;
613 }
614 /* Renumerate unallocated components */
615 for (j = 0; j < sc->n_components-1; j++) {
616 if ((sc->components[j].flags &
617 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
618 sc->components[j].index = j;
619 }
620 }
621 /* This is the critical section. If a component allocation
622 * event happens while both variables are not yet set,
623 * there will be trouble. Something will panic on encountering
624 * NULL sc->components[x].gcomp member.
625 * Luckily, component allocation happens very rarely and
626 * removing components is an abnormal action in any case. */
627 sc->components = newcomp;
628 sc->n_components--;
629 /* End critical section */
630
631 g_topology_lock();
632 if (clear_metadata(&compbak[found]) != 0) {
633 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
634 "metadata on %s", prov_name);
635 }
636 g_detach(compbak[found].gcons);
637 g_destroy_consumer(compbak[found].gcons);
638 g_topology_unlock();
639
640 free(compbak, M_GVIRSTOR);
641
642 removed++;
643 }
644
645 /* This call to update_metadata() is critical. In case there's a
646 * power failure in the middle of it and some components are updated
647 * while others are not, there will be trouble on next .taste() iff
648 * a non-updated component is detected first */
649 g_topology_lock();
650 update_metadata(sc);
651 g_topology_unlock();
652 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
653 sc->geom->name);
654 }
655
656 /*
657 * Clear metadata sector on component
658 */
659 static int
660 clear_metadata(struct g_virstor_component *comp)
661 {
662 char *buf;
663 int error;
664
665 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
666 comp->gcons->provider->name);
667 g_topology_assert();
668 error = g_access(comp->gcons, 0, 1, 0);
669 if (error != 0)
670 return (error);
671 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
672 M_WAITOK | M_ZERO);
673 error = g_write_data(comp->gcons,
674 comp->gcons->provider->mediasize -
675 comp->gcons->provider->sectorsize,
676 buf,
677 comp->gcons->provider->sectorsize);
678 free(buf, M_GVIRSTOR);
679 g_access(comp->gcons, 0, -1, 0);
680 return (error);
681 }
682
683 /*
684 * Destroy geom forcibly.
685 */
686 static int
687 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
688 struct g_geom *gp)
689 {
690 struct g_virstor_softc *sc;
691 int exitval;
692
693 sc = gp->softc;
694 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
695
696 exitval = 0;
697 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
698 gp->softc);
699
700 if (sc != NULL) {
701 #ifdef INVARIANTS
702 char *buf;
703 int error;
704 off_t off;
705 int isclean, count;
706 int n;
707
708 LOG_MSG(LVL_INFO, "INVARIANTS detected");
709 LOG_MSG(LVL_INFO, "Verifying allocation "
710 "table for %s", sc->geom->name);
711 count = 0;
712 for (n = 0; n < sc->chunk_count; n++) {
713 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
714 count++;
715 }
716 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
717 sc->geom->name, count);
718 n = off = count = 0;
719 isclean = 1;
720 if (virstor_valid_components(sc) != sc->n_components) {
721 /* This is a incomplete virstor device (not all
722 * components have been found) */
723 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
724 sc->geom->name);
725 goto bailout;
726 }
727 error = g_access(sc->components[0].gcons, 1, 0, 0);
728 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
729 error));
730 /* Compare the whole on-disk allocation table with what's
731 * currently in memory */
732 while (n < sc->chunk_count) {
733 buf = g_read_data(sc->components[0].gcons, off,
734 sc->sectorsize, &error);
735 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
736 "for read at %jd", error, off));
737 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
738 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
739 "entry %d, offset %jd", n, off);
740 isclean = 0;
741 count++;
742 }
743 n += sc->me_per_sector;
744 off += sc->sectorsize;
745 g_free(buf);
746 }
747 error = g_access(sc->components[0].gcons, -1, 0, 0);
748 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
749 __func__, error));
750 if (isclean != 1) {
751 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
752 "(%d sectors don't match, max %zu allocations)",
753 sc->geom->name, count,
754 count * sc->me_per_sector);
755 } else {
756 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
757 sc->geom->name);
758 }
759 bailout:
760 #endif
761 update_metadata(sc);
762 virstor_geom_destroy(sc, FALSE, FALSE);
763 exitval = EAGAIN;
764 } else
765 exitval = 0;
766 return (exitval);
767 }
768
769 /*
770 * Taste event (per-class callback)
771 * Examines a provider and creates geom instances if needed
772 */
773 static struct g_geom *
774 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
775 {
776 struct g_virstor_metadata md;
777 struct g_geom *gp;
778 struct g_consumer *cp;
779 struct g_virstor_softc *sc;
780 int error;
781
782 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
783 g_topology_assert();
784 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
785
786 /* We need a dummy geom to attach a consumer to the given provider */
787 gp = g_new_geomf(mp, "virstor:taste.helper");
788 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
789 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
790 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
791
792 cp = g_new_consumer(gp);
793 g_attach(cp, pp);
794 error = read_metadata(cp, &md);
795 g_detach(cp);
796 g_destroy_consumer(cp);
797 g_destroy_geom(gp);
798
799 if (error != 0)
800 return (NULL);
801
802 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
803 return (NULL);
804 if (md.md_version != G_VIRSTOR_VERSION) {
805 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
806 "to handle %s (%s) : %d should be %d",
807 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
808 return (NULL);
809 }
810 if (md.provsize != pp->mediasize)
811 return (NULL);
812
813 /* If the provider name is hardcoded, use the offered provider only
814 * if it's been offered with its proper name (the one used in
815 * the label command). */
816 if (md.provider[0] != '\0' &&
817 !g_compare_names(md.provider, pp->name))
818 return (NULL);
819
820 /* Iterate all geoms this class already knows about to see if a new
821 * geom instance of this class needs to be created (in case the provider
822 * is first from a (possibly) multi-consumer geom) or it just needs
823 * to be added to an existing instance. */
824 sc = NULL;
825 gp = NULL;
826 LIST_FOREACH(gp, &mp->geom, geom) {
827 sc = gp->softc;
828 if (sc == NULL)
829 continue;
830 if (strcmp(md.md_name, sc->geom->name) != 0)
831 continue;
832 if (md.md_id != sc->id)
833 continue;
834 break;
835 }
836 if (gp != NULL) { /* We found an existing geom instance; add to it */
837 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
838 error = add_provider_to_geom(sc, pp, &md);
839 if (error != 0) {
840 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
841 pp->name, md.md_name, error);
842 return (NULL);
843 }
844 } else { /* New geom instance needs to be created */
845 gp = create_virstor_geom(mp, &md);
846 if (gp == NULL) {
847 LOG_MSG(LVL_ERROR, "Error creating new instance of "
848 "class %s: %s", mp->name, md.md_name);
849 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
850 md.md_name, pp->name);
851 return (NULL);
852 }
853 sc = gp->softc;
854 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
855 md.md_name);
856 error = add_provider_to_geom(sc, pp, &md);
857 if (error != 0) {
858 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
859 pp->name, md.md_name, error);
860 virstor_geom_destroy(sc, TRUE, FALSE);
861 return (NULL);
862 }
863 }
864
865 return (gp);
866 }
867
868 /*
869 * Destroyes consumer passed to it in arguments. Used as a callback
870 * on g_event queue.
871 */
872 static void
873 delay_destroy_consumer(void *arg, int flags __unused)
874 {
875 struct g_consumer *c = arg;
876 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
877 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
878 c->provider->name);
879 g_detach(c);
880 g_destroy_consumer(c);
881 }
882
883 /*
884 * Remove a component (consumer) from geom instance; If it's the first
885 * component being removed, orphan the provider to announce geom's being
886 * dismantled
887 */
888 static void
889 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
890 boolean_t delay)
891 {
892 struct g_consumer *c;
893
894 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
895 sc->geom->name));
896 c = comp->gcons;
897
898 comp->gcons = NULL;
899 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
900 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
901 sc->geom->name);
902 if (sc->provider != NULL) {
903 LOG_MSG(LVL_INFO, "Removing provider %s", sc->provider->name);
904 g_wither_provider(sc->provider, ENXIO);
905 sc->provider = NULL;
906 }
907
908 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
909 return;
910 if (delay) {
911 /* Destroy consumer after it's tasted */
912 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
913 } else {
914 g_detach(c);
915 g_destroy_consumer(c);
916 }
917 }
918
919 /*
920 * Destroy geom - called internally
921 * See g_virstor_destroy_geom for the other one
922 */
923 static int
924 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
925 boolean_t delay)
926 {
927 struct g_provider *pp;
928 struct g_geom *gp;
929 u_int n;
930
931 g_topology_assert();
932
933 if (sc == NULL)
934 return (ENXIO);
935
936 pp = sc->provider;
937 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
938 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
939 "Device %s is still open.", pp->name);
940 if (!force)
941 return (EBUSY);
942 }
943
944 for (n = 0; n < sc->n_components; n++) {
945 if (sc->components[n].gcons != NULL)
946 remove_component(sc, &sc->components[n], delay);
947 }
948
949 gp = sc->geom;
950 gp->softc = NULL;
951
952 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
953 gp->name));
954
955 /* XXX: This might or might not work, since we're called with
956 * the topology lock held. Also, it might panic the kernel if
957 * the error'd BIO is in softupdates code. */
958 mtx_lock(&sc->delayed_bio_q_mtx);
959 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
960 struct g_virstor_bio_q *bq;
961 bq = STAILQ_FIRST(&sc->delayed_bio_q);
962 bq->bio->bio_error = ENOSPC;
963 g_io_deliver(bq->bio, EIO);
964 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
965 free(bq, M_GVIRSTOR);
966 }
967 mtx_unlock(&sc->delayed_bio_q_mtx);
968 mtx_destroy(&sc->delayed_bio_q_mtx);
969
970 free(sc->map, M_GVIRSTOR);
971 free(sc->components, M_GVIRSTOR);
972 bzero(sc, sizeof *sc);
973 free(sc, M_GVIRSTOR);
974
975 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
976 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
977 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
978
979 g_wither_geom(gp, ENXIO);
980
981 return (0);
982 }
983
984 /*
985 * Utility function: read metadata & decode. Wants topology lock to be
986 * held.
987 */
988 static int
989 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
990 {
991 struct g_provider *pp;
992 char *buf;
993 int error;
994
995 g_topology_assert();
996 error = g_access(cp, 1, 0, 0);
997 if (error != 0)
998 return (error);
999 pp = cp->provider;
1000 g_topology_unlock();
1001 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
1002 &error);
1003 g_topology_lock();
1004 g_access(cp, -1, 0, 0);
1005 if (buf == NULL)
1006 return (error);
1007
1008 virstor_metadata_decode(buf, md);
1009 g_free(buf);
1010
1011 return (0);
1012 }
1013
1014 /**
1015 * Utility function: encode & write metadata. Assumes topology lock is
1016 * held.
1017 *
1018 * There is no useful way of recovering from errors in this function,
1019 * not involving panicking the kernel. If the metadata cannot be written
1020 * the most we can do is notify the operator and hope he spots it and
1021 * replaces the broken drive.
1022 */
1023 static void
1024 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1025 {
1026 struct g_provider *pp;
1027 char *buf;
1028 int error;
1029
1030 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1031 ("Something's fishy in %s", __func__));
1032 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1033 g_topology_assert();
1034 error = g_access(cp, 0, 1, 0);
1035 if (error != 0) {
1036 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d",
1037 cp->provider->name, error);
1038 return;
1039 }
1040 pp = cp->provider;
1041
1042 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1043 bzero(buf, pp->sectorsize);
1044 virstor_metadata_encode(md, buf);
1045 g_topology_unlock();
1046 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1047 pp->sectorsize);
1048 g_topology_lock();
1049 g_access(cp, 0, -1, 0);
1050 free(buf, M_GVIRSTOR);
1051
1052 if (error != 0)
1053 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s",
1054 error, cp->provider->name);
1055 }
1056
1057 /*
1058 * Creates a new instance of this GEOM class, initialise softc
1059 */
1060 static struct g_geom *
1061 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1062 {
1063 struct g_geom *gp;
1064 struct g_virstor_softc *sc;
1065
1066 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1067 md->md_name, md->md_id);
1068
1069 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1070 md->md_virsize < md->md_chunk_size) {
1071 /* This is bogus configuration, and probably means data is
1072 * somehow corrupted. Panic, maybe? */
1073 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1074 md->md_name);
1075 return (NULL);
1076 }
1077
1078 /* Check if it's already created */
1079 LIST_FOREACH(gp, &mp->geom, geom) {
1080 sc = gp->softc;
1081 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1082 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1083 md->md_name);
1084 if (sc->id != md->md_id) {
1085 LOG_MSG(LVL_ERROR,
1086 "Some stale or invalid components "
1087 "exist for virstor device named %s. "
1088 "You will need to <CLEAR> all stale "
1089 "components and maybe reconfigure "
1090 "the virstor device. Tune "
1091 "kern.geom.virstor.debug sysctl up "
1092 "for more information.",
1093 sc->geom->name);
1094 }
1095 return (NULL);
1096 }
1097 }
1098 gp = g_new_geomf(mp, "%s", md->md_name);
1099 gp->softc = NULL; /* to circumevent races that test softc */
1100
1101 gp->start = g_virstor_start;
1102 gp->spoiled = g_virstor_orphan;
1103 gp->orphan = g_virstor_orphan;
1104 gp->access = g_virstor_access;
1105 gp->dumpconf = g_virstor_dumpconf;
1106
1107 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1108 sc->id = md->md_id;
1109 sc->n_components = md->md_count;
1110 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1111 M_GVIRSTOR, M_WAITOK | M_ZERO);
1112 sc->chunk_size = md->md_chunk_size;
1113 sc->virsize = md->md_virsize;
1114 STAILQ_INIT(&sc->delayed_bio_q);
1115 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1116 "gvirstor", MTX_DEF | MTX_RECURSE);
1117
1118 sc->geom = gp;
1119 sc->provider = NULL; /* virstor_check_and_run will create it */
1120 gp->softc = sc;
1121
1122 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1123
1124 return (gp);
1125 }
1126
1127 /*
1128 * Add provider to a GEOM class instance
1129 */
1130 static int
1131 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1132 struct g_virstor_metadata *md)
1133 {
1134 struct g_virstor_component *component;
1135 struct g_consumer *cp, *fcp;
1136 struct g_geom *gp;
1137 int error;
1138
1139 if (md->no >= sc->n_components)
1140 return (EINVAL);
1141
1142 /* "Current" compontent */
1143 component = &(sc->components[md->no]);
1144 if (component->gcons != NULL)
1145 return (EEXIST);
1146
1147 gp = sc->geom;
1148 fcp = LIST_FIRST(&gp->consumer);
1149
1150 cp = g_new_consumer(gp);
1151 error = g_attach(cp, pp);
1152
1153 if (error != 0) {
1154 g_destroy_consumer(cp);
1155 return (error);
1156 }
1157
1158 if (fcp != NULL) {
1159 if (fcp->provider->sectorsize != pp->sectorsize) {
1160 /* TODO: this can be made to work */
1161 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1162 "sector size (%d)", pp->name, sc->geom->name,
1163 pp->sectorsize);
1164 return (EINVAL);
1165 }
1166 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1167 /* Replicate access permissions from first "live" consumer
1168 * to the new one */
1169 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1170 if (error != 0) {
1171 g_detach(cp);
1172 g_destroy_consumer(cp);
1173 return (error);
1174 }
1175 }
1176 }
1177
1178 /* Bring up a new component */
1179 cp->private = component;
1180 component->gcons = cp;
1181 component->sc = sc;
1182 component->index = md->no;
1183 component->chunk_count = md->chunk_count;
1184 component->chunk_next = md->chunk_next;
1185 component->chunk_reserved = md->chunk_reserved;
1186 component->flags = md->flags;
1187
1188 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1189
1190 virstor_check_and_run(sc);
1191 return (0);
1192 }
1193
1194 /*
1195 * Check if everything's ready to create the geom provider & device entry,
1196 * create and start provider.
1197 * Called ultimately by .taste, from g_event thread
1198 */
1199 static void
1200 virstor_check_and_run(struct g_virstor_softc *sc)
1201 {
1202 off_t off;
1203 size_t n, count;
1204 int index;
1205 int error;
1206
1207 if (virstor_valid_components(sc) != sc->n_components)
1208 return;
1209
1210 if (virstor_valid_components(sc) == 0) {
1211 /* This is actually a candidate for panic() */
1212 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1213 sc->provider->name);
1214 return;
1215 }
1216
1217 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1218
1219 /* Initialise allocation map from the first consumer */
1220 sc->chunk_count = sc->virsize / sc->chunk_size;
1221 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1222 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1223 sc->provider->name,
1224 sc->chunk_count * (off_t)sc->chunk_size);
1225 }
1226 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1227 /* The following allocation is in order of 4MB - 8MB */
1228 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1229 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1230 __func__, sc->map_size, sc->provider->name));
1231 sc->map_sectors = sc->map_size / sc->sectorsize;
1232
1233 count = 0;
1234 for (n = 0; n < sc->n_components; n++)
1235 count += sc->components[n].chunk_count;
1236 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1237 "(%zu KB chunks)",
1238 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1239
1240 error = g_access(sc->components[0].gcons, 1, 0, 0);
1241 if (error != 0) {
1242 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1243 "read allocation map for %s",
1244 sc->components[0].gcons->provider->name,
1245 sc->geom->name);
1246 return;
1247 }
1248 /* Read in the allocation map */
1249 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1250 sc->components[0].gcons->provider->name);
1251 off = count = n = 0;
1252 while (count < sc->map_size) {
1253 struct g_virstor_map_entry *mapbuf;
1254 size_t bs;
1255
1256 bs = MIN(MAXPHYS, sc->map_size - count);
1257 if (bs % sc->sectorsize != 0) {
1258 /* Check for alignment errors */
1259 bs = rounddown(bs, sc->sectorsize);
1260 if (bs == 0)
1261 break;
1262 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1263 "for %s on %s", sc->geom->name,
1264 sc->components[0].gcons->provider->name);
1265 }
1266 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1267 if (mapbuf == NULL) {
1268 free(sc->map, M_GVIRSTOR);
1269 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1270 "for %s from %s (offset %ju) (error %d)",
1271 sc->geom->name,
1272 sc->components[0].gcons->provider->name,
1273 off, error);
1274 return;
1275 }
1276
1277 bcopy(mapbuf, &sc->map[n], bs);
1278 off += bs;
1279 count += bs;
1280 n += bs / sizeof *(sc->map);
1281 g_free(mapbuf);
1282 }
1283 g_access(sc->components[0].gcons, -1, 0, 0);
1284 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1285
1286 /* find first component with allocatable chunks */
1287 index = -1;
1288 for (n = 0; n < sc->n_components; n++) {
1289 if (sc->components[n].chunk_next <
1290 sc->components[n].chunk_count) {
1291 index = n;
1292 break;
1293 }
1294 }
1295 if (index == -1)
1296 /* not found? set it to the last component and handle it
1297 * later */
1298 index = sc->n_components - 1;
1299
1300 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1301 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1302 "(%d/%u: %s)", sc->geom->name,
1303 index+1,
1304 sc->n_components,
1305 sc->components[index].gcons->provider->name);
1306 }
1307 sc->curr_component = index;
1308
1309 if (sc->components[index].chunk_next >=
1310 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1311 LOG_MSG(LVL_WARNING,
1312 "Component %s of %s is running out of free space "
1313 "(%u chunks left)",
1314 sc->components[index].gcons->provider->name,
1315 sc->geom->name, sc->components[index].chunk_count -
1316 sc->components[index].chunk_next);
1317 }
1318
1319 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1320 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1321 LOG_MSG(LVL_ERROR,
1322 "%s: Map entries don't fit exactly in a sector (%s)",
1323 __func__, sc->geom->name);
1324 return;
1325 }
1326
1327 /* Recalculate allocated chunks in components & at the same time
1328 * verify map data is sane. We could trust metadata on this, but
1329 * we want to make sure. */
1330 for (n = 0; n < sc->n_components; n++)
1331 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1332
1333 for (n = 0; n < sc->chunk_count; n++) {
1334 if (sc->map[n].provider_no >= sc->n_components ||
1335 sc->map[n].provider_chunk >=
1336 sc->components[sc->map[n].provider_no].chunk_count) {
1337 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1338 __func__, (u_int)n, sc->geom->name);
1339 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1340 " provider_chunk: %u, chunk_count: %u", __func__,
1341 sc->map[n].provider_no, sc->n_components,
1342 sc->map[n].provider_chunk,
1343 sc->components[sc->map[n].provider_no].chunk_count);
1344 return;
1345 }
1346 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1347 sc->components[sc->map[n].provider_no].chunk_next++;
1348 }
1349
1350 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1351 sc->geom->name);
1352
1353 sc->provider->sectorsize = sc->sectorsize;
1354 sc->provider->mediasize = sc->virsize;
1355 g_error_provider(sc->provider, 0);
1356
1357 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1358 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1359 "chunk %u", sc->provider->name, sc->curr_component,
1360 sc->components[sc->curr_component].chunk_next);
1361 }
1362
1363 /*
1364 * Returns count of active providers in this geom instance
1365 */
1366 static u_int
1367 virstor_valid_components(struct g_virstor_softc *sc)
1368 {
1369 unsigned int nc, i;
1370
1371 nc = 0;
1372 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1373 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1374 for (i = 0; i < sc->n_components; i++)
1375 if (sc->components[i].gcons != NULL)
1376 nc++;
1377 return (nc);
1378 }
1379
1380 /*
1381 * Called when the consumer gets orphaned (?)
1382 */
1383 static void
1384 g_virstor_orphan(struct g_consumer *cp)
1385 {
1386 struct g_virstor_softc *sc;
1387 struct g_virstor_component *comp;
1388 struct g_geom *gp;
1389
1390 g_topology_assert();
1391 gp = cp->geom;
1392 sc = gp->softc;
1393 if (sc == NULL)
1394 return;
1395
1396 comp = cp->private;
1397 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1398 __func__));
1399 remove_component(sc, comp, FALSE);
1400 if (LIST_EMPTY(&gp->consumer))
1401 virstor_geom_destroy(sc, TRUE, FALSE);
1402 }
1403
1404 /*
1405 * Called to notify geom when it's been opened, and for what intent
1406 */
1407 static int
1408 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1409 {
1410 struct g_consumer *c, *c2, *tmp;
1411 struct g_virstor_softc *sc;
1412 struct g_geom *gp;
1413 int error;
1414
1415 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1416 gp = pp->geom;
1417 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1418 sc = gp->softc;
1419
1420 /* Grab an exclusive bit to propagate on our consumers on first open */
1421 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1422 de++;
1423 /* ... drop it on close */
1424 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1425 de--;
1426 if (sc != NULL)
1427 update_metadata(sc);
1428 }
1429
1430 error = ENXIO;
1431 LIST_FOREACH_SAFE(c, &gp->consumer, consumer, tmp) {
1432 error = g_access(c, dr, dw, de);
1433 if (error != 0)
1434 goto fail;
1435 if (c->acr == 0 && c->acw == 0 && c->ace == 0 &&
1436 c->flags & G_CF_ORPHAN) {
1437 g_detach(c);
1438 g_destroy_consumer(c);
1439 }
1440 }
1441
1442 if (sc != NULL && LIST_EMPTY(&gp->consumer))
1443 virstor_geom_destroy(sc, TRUE, FALSE);
1444
1445 return (error);
1446
1447 fail:
1448 /* Backout earlier changes */
1449 LIST_FOREACH(c2, &gp->consumer, consumer) {
1450 if (c2 == c)
1451 break;
1452 g_access(c2, -dr, -dw, -de);
1453 }
1454 return (error);
1455 }
1456
1457 /*
1458 * Generate XML dump of current state
1459 */
1460 static void
1461 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1462 struct g_consumer *cp, struct g_provider *pp)
1463 {
1464 struct g_virstor_softc *sc;
1465
1466 g_topology_assert();
1467 sc = gp->softc;
1468
1469 if (sc == NULL || pp != NULL)
1470 return;
1471
1472 if (cp != NULL) {
1473 /* For each component */
1474 struct g_virstor_component *comp;
1475
1476 comp = cp->private;
1477 if (comp == NULL)
1478 return;
1479 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1480 indent, comp->index);
1481 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1482 indent, comp->chunk_count);
1483 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1484 indent, comp->chunk_next);
1485 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1486 indent, comp->chunk_reserved);
1487 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1488 indent,
1489 comp->chunk_next > 0 ? 100 -
1490 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1491 comp->chunk_count : 100);
1492 } else {
1493 /* For the whole thing */
1494 u_int count, used, i;
1495 off_t size;
1496
1497 count = used = size = 0;
1498 for (i = 0; i < sc->n_components; i++) {
1499 if (sc->components[i].gcons != NULL) {
1500 count += sc->components[i].chunk_count;
1501 used += sc->components[i].chunk_next +
1502 sc->components[i].chunk_reserved;
1503 size += sc->components[i].gcons->
1504 provider->mediasize;
1505 }
1506 }
1507
1508 sbuf_printf(sb, "%s<Status>"
1509 "Components=%u, Online=%u</Status>\n", indent,
1510 sc->n_components, virstor_valid_components(sc));
1511 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1512 indent, 100-(used * 100) / count);
1513 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1514 sc->chunk_size);
1515 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1516 indent, used > 0 ? 100 - (used * 100) / count : 100);
1517 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1518 indent, count);
1519 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1520 indent, sc->chunk_count);
1521 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1522 indent,
1523 (count * 100) / sc->chunk_count);
1524 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1525 indent, size);
1526 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1527 sc->virsize);
1528 }
1529 }
1530
1531 /*
1532 * GEOM .done handler
1533 * Can't use standard handler because one requested IO may
1534 * fork into additional data IOs
1535 */
1536 static void
1537 g_virstor_done(struct bio *b)
1538 {
1539 struct g_virstor_softc *sc;
1540 struct bio *parent_b;
1541
1542 parent_b = b->bio_parent;
1543 sc = parent_b->bio_to->geom->softc;
1544
1545 if (b->bio_error != 0) {
1546 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1547 b->bio_error, b->bio_offset, b->bio_length,
1548 b->bio_to->name);
1549 if (parent_b->bio_error == 0)
1550 parent_b->bio_error = b->bio_error;
1551 }
1552
1553 parent_b->bio_inbed++;
1554 parent_b->bio_completed += b->bio_completed;
1555
1556 if (parent_b->bio_children == parent_b->bio_inbed) {
1557 parent_b->bio_completed = parent_b->bio_length;
1558 g_io_deliver(parent_b, parent_b->bio_error);
1559 }
1560 g_destroy_bio(b);
1561 }
1562
1563 /*
1564 * I/O starts here
1565 * Called in g_down thread
1566 */
1567 static void
1568 g_virstor_start(struct bio *b)
1569 {
1570 struct g_virstor_softc *sc;
1571 struct g_virstor_component *comp;
1572 struct bio *cb;
1573 struct g_provider *pp;
1574 char *addr;
1575 off_t offset, length;
1576 struct bio_queue_head bq;
1577 size_t chunk_size; /* cached for convenience */
1578 u_int count;
1579
1580 pp = b->bio_to;
1581 sc = pp->geom->softc;
1582 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1583 b->bio_to->error, b->bio_to->name));
1584
1585 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1586
1587 switch (b->bio_cmd) {
1588 case BIO_READ:
1589 case BIO_WRITE:
1590 case BIO_DELETE:
1591 break;
1592 default:
1593 g_io_deliver(b, EOPNOTSUPP);
1594 return;
1595 }
1596
1597 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1598 bioq_init(&bq);
1599
1600 chunk_size = sc->chunk_size;
1601 addr = b->bio_data;
1602 offset = b->bio_offset; /* virtual offset and length */
1603 length = b->bio_length;
1604
1605 while (length > 0) {
1606 size_t chunk_index, in_chunk_offset, in_chunk_length;
1607 struct virstor_map_entry *me;
1608
1609 chunk_index = offset / chunk_size; /* round downwards */
1610 in_chunk_offset = offset % chunk_size;
1611 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1612 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1613 b->bio_cmd == BIO_READ ? "R" : "W",
1614 offset, length,
1615 chunk_index, in_chunk_offset, in_chunk_length);
1616 me = &sc->map[chunk_index];
1617
1618 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1619 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1620 /* Reads from unallocated chunks return zeroed
1621 * buffers */
1622 if (b->bio_cmd == BIO_READ)
1623 bzero(addr, in_chunk_length);
1624 } else {
1625 comp = &sc->components[me->provider_no];
1626
1627 cb = g_clone_bio(b);
1628 if (cb == NULL) {
1629 bioq_dismantle(&bq);
1630 if (b->bio_error == 0)
1631 b->bio_error = ENOMEM;
1632 g_io_deliver(b, b->bio_error);
1633 return;
1634 }
1635 cb->bio_to = comp->gcons->provider;
1636 cb->bio_done = g_virstor_done;
1637 cb->bio_offset =
1638 (off_t)me->provider_chunk * (off_t)chunk_size
1639 + in_chunk_offset;
1640 cb->bio_length = in_chunk_length;
1641 cb->bio_data = addr;
1642 cb->bio_caller1 = comp;
1643 bioq_disksort(&bq, cb);
1644 }
1645 } else { /* handle BIO_WRITE */
1646 KASSERT(b->bio_cmd == BIO_WRITE,
1647 ("%s: Unknown command %d", __func__,
1648 b->bio_cmd));
1649
1650 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1651 /* We have a virtual chunk, represented by
1652 * the "me" entry, but it's not yet allocated
1653 * (tied to) a physical chunk. So do it now. */
1654 struct virstor_map_entry *data_me;
1655 u_int phys_chunk, comp_no;
1656 off_t s_offset;
1657 int error;
1658
1659 error = allocate_chunk(sc, &comp, &comp_no,
1660 &phys_chunk);
1661 if (error != 0) {
1662 /* We cannot allocate a physical chunk
1663 * to satisfy this request, so we'll
1664 * delay it to when we can...
1665 * XXX: this will prevent the fs from
1666 * being umounted! */
1667 struct g_virstor_bio_q *biq;
1668 biq = malloc(sizeof *biq, M_GVIRSTOR,
1669 M_NOWAIT);
1670 if (biq == NULL) {
1671 bioq_dismantle(&bq);
1672 if (b->bio_error == 0)
1673 b->bio_error = ENOMEM;
1674 g_io_deliver(b, b->bio_error);
1675 return;
1676 }
1677 biq->bio = b;
1678 mtx_lock(&sc->delayed_bio_q_mtx);
1679 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1680 biq, linkage);
1681 mtx_unlock(&sc->delayed_bio_q_mtx);
1682 LOG_MSG(LVL_WARNING, "Delaying BIO "
1683 "(size=%ju) until free physical "
1684 "space can be found on %s",
1685 b->bio_length,
1686 sc->provider->name);
1687 return;
1688 }
1689 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1690 "for %s",
1691 phys_chunk,
1692 comp->gcons->provider->name,
1693 sc->provider->name);
1694
1695 me->provider_no = comp_no;
1696 me->provider_chunk = phys_chunk;
1697 me->flags |= VIRSTOR_MAP_ALLOCATED;
1698
1699 cb = g_clone_bio(b);
1700 if (cb == NULL) {
1701 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1702 me->provider_no = 0;
1703 me->provider_chunk = 0;
1704 bioq_dismantle(&bq);
1705 if (b->bio_error == 0)
1706 b->bio_error = ENOMEM;
1707 g_io_deliver(b, b->bio_error);
1708 return;
1709 }
1710
1711 /* The allocation table is stored continuously
1712 * at the start of the drive. We need to
1713 * calculate the offset of the sector that holds
1714 * this map entry both on the drive and in the
1715 * map array.
1716 * sc_offset will end up pointing to the drive
1717 * sector. */
1718 s_offset = chunk_index * sizeof *me;
1719 s_offset = rounddown(s_offset, sc->sectorsize);
1720
1721 /* data_me points to map entry sector
1722 * in memory (analogous to offset) */
1723 data_me = &sc->map[rounddown(chunk_index,
1724 sc->me_per_sector)];
1725
1726 /* Commit sector with map entry to storage */
1727 cb->bio_to = sc->components[0].gcons->provider;
1728 cb->bio_done = g_virstor_done;
1729 cb->bio_offset = s_offset;
1730 cb->bio_data = (char *)data_me;
1731 cb->bio_length = sc->sectorsize;
1732 cb->bio_caller1 = &sc->components[0];
1733 bioq_disksort(&bq, cb);
1734 }
1735
1736 comp = &sc->components[me->provider_no];
1737 cb = g_clone_bio(b);
1738 if (cb == NULL) {
1739 bioq_dismantle(&bq);
1740 if (b->bio_error == 0)
1741 b->bio_error = ENOMEM;
1742 g_io_deliver(b, b->bio_error);
1743 return;
1744 }
1745 /* Finally, handle the data */
1746 cb->bio_to = comp->gcons->provider;
1747 cb->bio_done = g_virstor_done;
1748 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1749 in_chunk_offset;
1750 cb->bio_length = in_chunk_length;
1751 cb->bio_data = addr;
1752 cb->bio_caller1 = comp;
1753 bioq_disksort(&bq, cb);
1754 }
1755 addr += in_chunk_length;
1756 length -= in_chunk_length;
1757 offset += in_chunk_length;
1758 }
1759
1760 /* Fire off bio's here */
1761 count = 0;
1762 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1763 bioq_remove(&bq, cb);
1764 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1765 comp = cb->bio_caller1;
1766 cb->bio_caller1 = NULL;
1767 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1768 cb->bio_offset, cb->bio_length);
1769 g_io_request(cb, comp->gcons);
1770 count++;
1771 }
1772 if (count == 0) { /* We handled everything locally */
1773 b->bio_completed = b->bio_length;
1774 g_io_deliver(b, 0);
1775 }
1776
1777 }
1778
1779 /*
1780 * Allocate a chunk from a physical provider. Returns physical component,
1781 * chunk index relative to the component and the component's index.
1782 */
1783 static int
1784 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1785 u_int *comp_no_p, u_int *chunk)
1786 {
1787 u_int comp_no;
1788
1789 KASSERT(sc->curr_component < sc->n_components,
1790 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1791
1792 comp_no = sc->curr_component;
1793 *comp = &sc->components[comp_no];
1794 dump_component(*comp);
1795 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1796 /* This component is full. Allocate next component */
1797 if (comp_no >= sc->n_components-1) {
1798 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1799 sc->geom->name);
1800 return (-1);
1801 }
1802 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1803 sc->curr_component = ++comp_no;
1804
1805 *comp = &sc->components[comp_no];
1806 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1807 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1808 "(switching to %u/%u: %s)", sc->geom->name,
1809 comp_no+1, sc->n_components,
1810 (*comp)->gcons->provider->name);
1811 /* Take care not to overwrite reserved chunks */
1812 if ( (*comp)->chunk_reserved > 0 &&
1813 (*comp)->chunk_next < (*comp)->chunk_reserved)
1814 (*comp)->chunk_next = (*comp)->chunk_reserved;
1815
1816 (*comp)->flags |=
1817 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1818 dump_component(*comp);
1819 *comp_no_p = comp_no;
1820 *chunk = (*comp)->chunk_next++;
1821 } else {
1822 *comp_no_p = comp_no;
1823 *chunk = (*comp)->chunk_next++;
1824 }
1825 return (0);
1826 }
1827
1828 /* Dump a component */
1829 static void
1830 dump_component(struct g_virstor_component *comp)
1831 {
1832
1833 if (g_virstor_debug < LVL_DEBUG2)
1834 return;
1835 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1836 printf(" chunk_count: %u\n", comp->chunk_count);
1837 printf(" chunk_next: %u\n", comp->chunk_next);
1838 printf(" flags: %u\n", comp->flags);
1839 }
1840
1841 #if 0
1842 /* Dump a map entry */
1843 static void
1844 dump_me(struct virstor_map_entry *me, unsigned int nr)
1845 {
1846 if (g_virstor_debug < LVL_DEBUG)
1847 return;
1848 printf("VIRT. CHUNK #%d: ", nr);
1849 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1850 printf("(unallocated)\n");
1851 else
1852 printf("allocated at provider %u, provider_chunk %u\n",
1853 me->provider_no, me->provider_chunk);
1854 }
1855 #endif
1856
1857 /*
1858 * Dismantle bio_queue and destroy its components
1859 */
1860 static void
1861 bioq_dismantle(struct bio_queue_head *bq)
1862 {
1863 struct bio *b;
1864
1865 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1866 bioq_remove(bq, b);
1867 g_destroy_bio(b);
1868 }
1869 }
1870
1871 /*
1872 * The function that shouldn't be called.
1873 * When this is called, the stack is already garbled because of
1874 * argument mismatch. There's nothing to do now but panic, which is
1875 * accidentally the whole purpose of this function.
1876 * Motivation: to guard from accidentally calling geom methods when
1877 * they shouldn't be called. (see g_..._taste)
1878 */
1879 static void
1880 invalid_call(void)
1881 {
1882 panic("invalid_call() has just been called. Something's fishy here.");
1883 }
1884
1885 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */
1886 MODULE_VERSION(geom_virstor, 0);
Cache object: 29f43fb68dcc99a8a452ea2b35a619e2
|