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/9.0/sys/geom/virstor/g_virstor.c 223921 2011-07-11 05:22:31Z ae $");
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 SYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0, "GEOM_GVIRSTOR information");
84
85 static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */
86 TUNABLE_INT("kern.geom.virstor.debug", &g_virstor_debug);
87 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RW, &g_virstor_debug,
88 0, "Debug level (2=production, 5=normal, 15=excessive)");
89
90 static u_int g_virstor_chunk_watermark = 100;
91 TUNABLE_INT("kern.geom.virstor.chunk_watermark", &g_virstor_chunk_watermark);
92 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RW,
93 &g_virstor_chunk_watermark, 0,
94 "Minimum number of free chunks before issuing administrative warning");
95
96 static u_int g_virstor_component_watermark = 1;
97 TUNABLE_INT("kern.geom.virstor.component_watermark",
98 &g_virstor_component_watermark);
99 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RW,
100 &g_virstor_component_watermark, 0,
101 "Minimum number of free components before issuing administrative warning");
102
103 static int read_metadata(struct g_consumer *, struct g_virstor_metadata *);
104 static void write_metadata(struct g_consumer *, struct g_virstor_metadata *);
105 static int clear_metadata(struct g_virstor_component *);
106 static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *,
107 struct g_virstor_metadata *);
108 static struct g_geom *create_virstor_geom(struct g_class *,
109 struct g_virstor_metadata *);
110 static void virstor_check_and_run(struct g_virstor_softc *);
111 static u_int virstor_valid_components(struct g_virstor_softc *);
112 static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t,
113 boolean_t);
114 static void remove_component(struct g_virstor_softc *,
115 struct g_virstor_component *, boolean_t);
116 static void bioq_dismantle(struct bio_queue_head *);
117 static int allocate_chunk(struct g_virstor_softc *,
118 struct g_virstor_component **, u_int *, u_int *);
119 static void delay_destroy_consumer(void *, int);
120 static void dump_component(struct g_virstor_component *comp);
121 #if 0
122 static void dump_me(struct virstor_map_entry *me, unsigned int nr);
123 #endif
124
125 static void virstor_ctl_stop(struct gctl_req *, struct g_class *);
126 static void virstor_ctl_add(struct gctl_req *, struct g_class *);
127 static void virstor_ctl_remove(struct gctl_req *, struct g_class *);
128 static struct g_virstor_softc * virstor_find_geom(const struct g_class *,
129 const char *);
130 static void update_metadata(struct g_virstor_softc *);
131 static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *,
132 u_int, u_int);
133
134 static void g_virstor_orphan(struct g_consumer *);
135 static int g_virstor_access(struct g_provider *, int, int, int);
136 static void g_virstor_start(struct bio *);
137 static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *,
138 struct g_consumer *, struct g_provider *);
139 static void g_virstor_done(struct bio *);
140
141 static void invalid_call(void);
142 /*
143 * Initialise GEOM class (per-class callback)
144 */
145 static void
146 g_virstor_init(struct g_class *mp __unused)
147 {
148
149 /* Catch map struct size mismatch at compile time; Map entries must
150 * fit into MAXPHYS exactly, with no wasted space. */
151 CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS);
152
153 /* Init UMA zones, TAILQ's, other global vars */
154 }
155
156 /*
157 * Finalise GEOM class (per-class callback)
158 */
159 static void
160 g_virstor_fini(struct g_class *mp __unused)
161 {
162
163 /* Deinit UMA zones & global vars */
164 }
165
166 /*
167 * Config (per-class callback)
168 */
169 static void
170 g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb)
171 {
172 uint32_t *version;
173
174 g_topology_assert();
175
176 version = gctl_get_paraml(req, "version", sizeof(*version));
177 if (version == NULL) {
178 gctl_error(req, "Failed to get 'version' argument");
179 return;
180 }
181 if (*version != G_VIRSTOR_VERSION) {
182 gctl_error(req, "Userland and kernel versions out of sync");
183 return;
184 }
185
186 g_topology_unlock();
187 if (strcmp(verb, "add") == 0)
188 virstor_ctl_add(req, cp);
189 else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0)
190 virstor_ctl_stop(req, cp);
191 else if (strcmp(verb, "remove") == 0)
192 virstor_ctl_remove(req, cp);
193 else
194 gctl_error(req, "unknown verb: '%s'", verb);
195 g_topology_lock();
196 }
197
198 /*
199 * "stop" verb from userland
200 */
201 static void
202 virstor_ctl_stop(struct gctl_req *req, struct g_class *cp)
203 {
204 int *force, *nargs;
205 int i;
206
207 nargs = gctl_get_paraml(req, "nargs", sizeof *nargs);
208 if (nargs == NULL) {
209 gctl_error(req, "Error fetching argument '%s'", "nargs");
210 return;
211 }
212 if (*nargs < 1) {
213 gctl_error(req, "Invalid number of arguments");
214 return;
215 }
216 force = gctl_get_paraml(req, "force", sizeof *force);
217 if (force == NULL) {
218 gctl_error(req, "Error fetching argument '%s'", "force");
219 return;
220 }
221
222 g_topology_lock();
223 for (i = 0; i < *nargs; i++) {
224 char param[8];
225 const char *name;
226 struct g_virstor_softc *sc;
227 int error;
228
229 sprintf(param, "arg%d", i);
230 name = gctl_get_asciiparam(req, param);
231 if (name == NULL) {
232 gctl_error(req, "No 'arg%d' argument", i);
233 g_topology_unlock();
234 return;
235 }
236 sc = virstor_find_geom(cp, name);
237 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
238 sc->geom->name);
239 update_metadata(sc);
240 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
241 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
242 sc->geom->name, error);
243 }
244 }
245 g_topology_unlock();
246 }
247
248 /*
249 * "add" verb from userland - add new component(s) to the structure.
250 * This will be done all at once in here, without going through the
251 * .taste function for new components.
252 */
253 static void
254 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
255 {
256 /* Note: while this is going on, I/O is being done on
257 * the g_up and g_down threads. The idea is to make changes
258 * to softc members in a way that can atomically activate
259 * them all at once. */
260 struct g_virstor_softc *sc;
261 int *hardcode, *nargs;
262 const char *geom_name; /* geom to add a component to */
263 struct g_consumer *fcp;
264 struct g_virstor_bio_q *bq;
265 u_int added;
266 int error;
267 int i;
268
269 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
270 if (nargs == NULL) {
271 gctl_error(req, "Error fetching argument '%s'", "nargs");
272 return;
273 }
274 if (*nargs < 2) {
275 gctl_error(req, "Invalid number of arguments");
276 return;
277 }
278 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
279 if (hardcode == NULL) {
280 gctl_error(req, "Error fetching argument '%s'", "hardcode");
281 return;
282 }
283
284 /* Find "our" geom */
285 geom_name = gctl_get_asciiparam(req, "arg0");
286 if (geom_name == NULL) {
287 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
288 return;
289 }
290 sc = virstor_find_geom(cp, geom_name);
291 if (sc == NULL) {
292 gctl_error(req, "Don't know anything about '%s'", geom_name);
293 return;
294 }
295
296 if (virstor_valid_components(sc) != sc->n_components) {
297 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
298 "virstor %s", sc->geom->name);
299 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
300 return;
301 }
302
303 fcp = sc->components[0].gcons;
304 added = 0;
305 g_topology_lock();
306 for (i = 1; i < *nargs; i++) {
307 struct g_virstor_metadata md;
308 char aname[8];
309 const char *prov_name;
310 struct g_provider *pp;
311 struct g_consumer *cp;
312 u_int nc;
313 u_int j;
314
315 snprintf(aname, sizeof aname, "arg%d", i);
316 prov_name = gctl_get_asciiparam(req, aname);
317 if (prov_name == NULL) {
318 gctl_error(req, "Error fetching argument '%s'", aname);
319 g_topology_unlock();
320 return;
321 }
322 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
323 prov_name += sizeof(_PATH_DEV) - 1;
324
325 pp = g_provider_by_name(prov_name);
326 if (pp == NULL) {
327 /* This is the most common error so be verbose about it */
328 if (added != 0) {
329 gctl_error(req, "Invalid provider: '%s' (added"
330 " %u components)", prov_name, added);
331 update_metadata(sc);
332 } else {
333 gctl_error(req, "Invalid provider: '%s'",
334 prov_name);
335 }
336 g_topology_unlock();
337 return;
338 }
339 cp = g_new_consumer(sc->geom);
340 if (cp == NULL) {
341 gctl_error(req, "Cannot create consumer");
342 g_topology_unlock();
343 return;
344 }
345 error = g_attach(cp, pp);
346 if (error != 0) {
347 gctl_error(req, "Cannot attach a consumer to %s",
348 pp->name);
349 g_destroy_consumer(cp);
350 g_topology_unlock();
351 return;
352 }
353 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
354 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
355 if (error != 0) {
356 gctl_error(req, "Access request failed for %s",
357 pp->name);
358 g_destroy_consumer(cp);
359 g_topology_unlock();
360 return;
361 }
362 }
363 if (fcp->provider->sectorsize != pp->sectorsize) {
364 gctl_error(req, "Sector size doesn't fit for %s",
365 pp->name);
366 g_destroy_consumer(cp);
367 g_topology_unlock();
368 return;
369 }
370 for (j = 0; j < sc->n_components; j++) {
371 if (strcmp(sc->components[j].gcons->provider->name,
372 pp->name) == 0) {
373 gctl_error(req, "Component %s already in %s",
374 pp->name, sc->geom->name);
375 g_destroy_consumer(cp);
376 g_topology_unlock();
377 return;
378 }
379 }
380 sc->components = realloc(sc->components,
381 sizeof(*sc->components) * (sc->n_components + 1),
382 M_GVIRSTOR, M_WAITOK);
383
384 nc = sc->n_components;
385 sc->components[nc].gcons = cp;
386 sc->components[nc].sc = sc;
387 sc->components[nc].index = nc;
388 sc->components[nc].chunk_count = cp->provider->mediasize /
389 sc->chunk_size;
390 sc->components[nc].chunk_next = 0;
391 sc->components[nc].chunk_reserved = 0;
392
393 if (sc->components[nc].chunk_count < 4) {
394 gctl_error(req, "Provider too small: %s",
395 cp->provider->name);
396 g_destroy_consumer(cp);
397 g_topology_unlock();
398 return;
399 }
400 fill_metadata(sc, &md, nc, *hardcode);
401 write_metadata(cp, &md);
402 /* The new component becomes visible when n_components is
403 * incremented */
404 sc->n_components++;
405 added++;
406
407 }
408 /* This call to update_metadata() is critical. In case there's a
409 * power failure in the middle of it and some components are updated
410 * while others are not, there will be trouble on next .taste() iff
411 * a non-updated component is detected first */
412 update_metadata(sc);
413 g_topology_unlock();
414 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
415 sc->geom->name);
416 /* Fire off BIOs previously queued because there wasn't any
417 * physical space left. If the BIOs still can't be satisfied
418 * they will again be added to the end of the queue (during
419 * which the mutex will be recursed) */
420 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
421 bq->bio = NULL;
422 mtx_lock(&sc->delayed_bio_q_mtx);
423 /* First, insert a sentinel to the queue end, so we don't
424 * end up in an infinite loop if there's still no free
425 * space available. */
426 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
427 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
428 bq = STAILQ_FIRST(&sc->delayed_bio_q);
429 if (bq->bio != NULL) {
430 g_virstor_start(bq->bio);
431 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
432 free(bq, M_GVIRSTOR);
433 } else {
434 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
435 free(bq, M_GVIRSTOR);
436 break;
437 }
438 }
439 mtx_unlock(&sc->delayed_bio_q_mtx);
440
441 }
442
443 /*
444 * Find a geom handled by the class
445 */
446 static struct g_virstor_softc *
447 virstor_find_geom(const struct g_class *cp, const char *name)
448 {
449 struct g_geom *gp;
450
451 LIST_FOREACH(gp, &cp->geom, geom) {
452 if (strcmp(name, gp->name) == 0)
453 return (gp->softc);
454 }
455 return (NULL);
456 }
457
458 /*
459 * Update metadata on all components to reflect the current state
460 * of these fields:
461 * - chunk_next
462 * - flags
463 * - md_count
464 * Expects things to be set up so write_metadata() can work, i.e.
465 * the topology lock must be held.
466 */
467 static void
468 update_metadata(struct g_virstor_softc *sc)
469 {
470 struct g_virstor_metadata md;
471 int n;
472
473 if (virstor_valid_components(sc) != sc->n_components)
474 return; /* Incomplete device */
475 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
476 sc->geom->name);
477 /* Update metadata on components */
478 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
479 sc->geom->class->name, sc->geom->name);
480 g_topology_assert();
481 for (n = 0; n < sc->n_components; n++) {
482 read_metadata(sc->components[n].gcons, &md);
483 md.chunk_next = sc->components[n].chunk_next;
484 md.flags = sc->components[n].flags;
485 md.md_count = sc->n_components;
486 write_metadata(sc->components[n].gcons, &md);
487 }
488 }
489
490 /*
491 * Fills metadata (struct md) from information stored in softc and the nc'th
492 * component of virstor
493 */
494 static void
495 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
496 u_int nc, u_int hardcode)
497 {
498 struct g_virstor_component *c;
499
500 bzero(md, sizeof *md);
501 c = &sc->components[nc];
502
503 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
504 md->md_version = G_VIRSTOR_VERSION;
505 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
506 md->md_id = sc->id;
507 md->md_virsize = sc->virsize;
508 md->md_chunk_size = sc->chunk_size;
509 md->md_count = sc->n_components;
510
511 if (hardcode) {
512 strncpy(md->provider, c->gcons->provider->name,
513 sizeof md->provider);
514 }
515 md->no = nc;
516 md->provsize = c->gcons->provider->mediasize;
517 md->chunk_count = c->chunk_count;
518 md->chunk_next = c->chunk_next;
519 md->chunk_reserved = c->chunk_reserved;
520 md->flags = c->flags;
521 }
522
523 /*
524 * Remove a component from virstor device.
525 * Can only be done if the component is unallocated.
526 */
527 static void
528 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
529 {
530 /* As this is executed in parallel to I/O, operations on virstor
531 * structures must be as atomic as possible. */
532 struct g_virstor_softc *sc;
533 int *nargs;
534 const char *geom_name;
535 u_int removed;
536 int i;
537
538 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
539 if (nargs == NULL) {
540 gctl_error(req, "Error fetching argument '%s'", "nargs");
541 return;
542 }
543 if (*nargs < 2) {
544 gctl_error(req, "Invalid number of arguments");
545 return;
546 }
547 /* Find "our" geom */
548 geom_name = gctl_get_asciiparam(req, "arg0");
549 if (geom_name == NULL) {
550 gctl_error(req, "Error fetching argument '%s'",
551 "geom_name (arg0)");
552 return;
553 }
554 sc = virstor_find_geom(cp, geom_name);
555 if (sc == NULL) {
556 gctl_error(req, "Don't know anything about '%s'", geom_name);
557 return;
558 }
559
560 if (virstor_valid_components(sc) != sc->n_components) {
561 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
562 "virstor %s", sc->geom->name);
563 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
564 return;
565 }
566
567 removed = 0;
568 for (i = 1; i < *nargs; i++) {
569 char param[8];
570 const char *prov_name;
571 int j, found;
572 struct g_virstor_component *newcomp, *compbak;
573
574 sprintf(param, "arg%d", i);
575 prov_name = gctl_get_asciiparam(req, param);
576 if (prov_name == NULL) {
577 gctl_error(req, "Error fetching argument '%s'", param);
578 return;
579 }
580 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
581 prov_name += sizeof(_PATH_DEV) - 1;
582
583 found = -1;
584 for (j = 0; j < sc->n_components; j++) {
585 if (strcmp(sc->components[j].gcons->provider->name,
586 prov_name) == 0) {
587 found = j;
588 break;
589 }
590 }
591 if (found == -1) {
592 LOG_MSG(LVL_ERROR, "No %s component in %s",
593 prov_name, sc->geom->name);
594 continue;
595 }
596
597 compbak = sc->components;
598 newcomp = malloc(sc->n_components * sizeof(*sc->components),
599 M_GVIRSTOR, M_WAITOK | M_ZERO);
600 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
601 bcopy(&sc->components[found + 1], newcomp + found,
602 found * sizeof(*sc->components));
603 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
604 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
605 "removed from %s",
606 prov_name, sc->geom->name);
607 free(newcomp, M_GVIRSTOR);
608 /* We'll consider this non-fatal error */
609 continue;
610 }
611 /* Renumerate unallocated components */
612 for (j = 0; j < sc->n_components-1; j++) {
613 if ((sc->components[j].flags &
614 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
615 sc->components[j].index = j;
616 }
617 }
618 /* This is the critical section. If a component allocation
619 * event happens while both variables are not yet set,
620 * there will be trouble. Something will panic on encountering
621 * NULL sc->components[x].gcomp member.
622 * Luckily, component allocation happens very rarely and
623 * removing components is an abnormal action in any case. */
624 sc->components = newcomp;
625 sc->n_components--;
626 /* End critical section */
627
628 g_topology_lock();
629 if (clear_metadata(&compbak[found]) != 0) {
630 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
631 "metadata on %s", prov_name);
632 }
633 g_detach(compbak[found].gcons);
634 g_destroy_consumer(compbak[found].gcons);
635 g_topology_unlock();
636
637 free(compbak, M_GVIRSTOR);
638
639 removed++;
640 }
641
642 /* This call to update_metadata() is critical. In case there's a
643 * power failure in the middle of it and some components are updated
644 * while others are not, there will be trouble on next .taste() iff
645 * a non-updated component is detected first */
646 g_topology_lock();
647 update_metadata(sc);
648 g_topology_unlock();
649 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
650 sc->geom->name);
651 }
652
653 /*
654 * Clear metadata sector on component
655 */
656 static int
657 clear_metadata(struct g_virstor_component *comp)
658 {
659 char *buf;
660 int error;
661
662 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
663 comp->gcons->provider->name);
664 g_topology_assert();
665 error = g_access(comp->gcons, 0, 1, 0);
666 if (error != 0)
667 return (error);
668 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
669 M_WAITOK | M_ZERO);
670 error = g_write_data(comp->gcons,
671 comp->gcons->provider->mediasize -
672 comp->gcons->provider->sectorsize,
673 buf,
674 comp->gcons->provider->sectorsize);
675 free(buf, M_GVIRSTOR);
676 g_access(comp->gcons, 0, -1, 0);
677 return (error);
678 }
679
680 /*
681 * Destroy geom forcibly.
682 */
683 static int
684 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
685 struct g_geom *gp)
686 {
687 struct g_virstor_softc *sc;
688 int exitval;
689
690 sc = gp->softc;
691 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
692
693 exitval = 0;
694 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
695 gp->softc);
696
697 if (sc != NULL) {
698 #ifdef INVARIANTS
699 char *buf;
700 int error;
701 off_t off;
702 int isclean, count;
703 int n;
704
705 LOG_MSG(LVL_INFO, "INVARIANTS detected");
706 LOG_MSG(LVL_INFO, "Verifying allocation "
707 "table for %s", sc->geom->name);
708 count = 0;
709 for (n = 0; n < sc->chunk_count; n++) {
710 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
711 count++;
712 }
713 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
714 sc->geom->name, count);
715 n = off = count = 0;
716 isclean = 1;
717 if (virstor_valid_components(sc) != sc->n_components) {
718 /* This is a incomplete virstor device (not all
719 * components have been found) */
720 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
721 sc->geom->name);
722 goto bailout;
723 }
724 error = g_access(sc->components[0].gcons, 1, 0, 0);
725 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
726 error));
727 /* Compare the whole on-disk allocation table with what's
728 * currently in memory */
729 while (n < sc->chunk_count) {
730 buf = g_read_data(sc->components[0].gcons, off,
731 sc->sectorsize, &error);
732 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
733 "for read at %jd", error, off));
734 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
735 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
736 "entry %d, offset %jd", n, off);
737 isclean = 0;
738 count++;
739 }
740 n += sc->me_per_sector;
741 off += sc->sectorsize;
742 g_free(buf);
743 }
744 error = g_access(sc->components[0].gcons, -1, 0, 0);
745 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
746 __func__, error));
747 if (isclean != 1) {
748 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
749 "(%d sectors don't match, max %zu allocations)",
750 sc->geom->name, count,
751 count * sc->me_per_sector);
752 } else {
753 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
754 sc->geom->name);
755 }
756 bailout:
757 #endif
758 update_metadata(sc);
759 virstor_geom_destroy(sc, FALSE, FALSE);
760 exitval = EAGAIN;
761 } else
762 exitval = 0;
763 return (exitval);
764 }
765
766 /*
767 * Taste event (per-class callback)
768 * Examines a provider and creates geom instances if needed
769 */
770 static struct g_geom *
771 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
772 {
773 struct g_virstor_metadata md;
774 struct g_geom *gp;
775 struct g_consumer *cp;
776 struct g_virstor_softc *sc;
777 int error;
778
779 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
780 g_topology_assert();
781 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
782
783 /* We need a dummy geom to attach a consumer to the given provider */
784 gp = g_new_geomf(mp, "virstor:taste.helper");
785 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
786 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
787 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
788
789 cp = g_new_consumer(gp);
790 g_attach(cp, pp);
791 error = read_metadata(cp, &md);
792 g_detach(cp);
793 g_destroy_consumer(cp);
794 g_destroy_geom(gp);
795
796 if (error != 0)
797 return (NULL);
798
799 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
800 return (NULL);
801 if (md.md_version != G_VIRSTOR_VERSION) {
802 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
803 "to handle %s (%s) : %d should be %d",
804 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
805 return (NULL);
806 }
807 if (md.provsize != pp->mediasize)
808 return (NULL);
809
810 /* If the provider name is hardcoded, use the offered provider only
811 * if it's been offered with its proper name (the one used in
812 * the label command). */
813 if (md.provider[0] != '\0' &&
814 !g_compare_names(md.provider, pp->name))
815 return (NULL);
816
817 /* Iterate all geoms this class already knows about to see if a new
818 * geom instance of this class needs to be created (in case the provider
819 * is first from a (possibly) multi-consumer geom) or it just needs
820 * to be added to an existing instance. */
821 sc = NULL;
822 gp = NULL;
823 LIST_FOREACH(gp, &mp->geom, geom) {
824 sc = gp->softc;
825 if (sc == NULL)
826 continue;
827 if (strcmp(md.md_name, sc->geom->name) != 0)
828 continue;
829 if (md.md_id != sc->id)
830 continue;
831 break;
832 }
833 if (gp != NULL) { /* We found an existing geom instance; add to it */
834 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
835 error = add_provider_to_geom(sc, pp, &md);
836 if (error != 0) {
837 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
838 pp->name, md.md_name, error);
839 return (NULL);
840 }
841 } else { /* New geom instance needs to be created */
842 gp = create_virstor_geom(mp, &md);
843 if (gp == NULL) {
844 LOG_MSG(LVL_ERROR, "Error creating new instance of "
845 "class %s: %s", mp->name, md.md_name);
846 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
847 md.md_name, pp->name);
848 return (NULL);
849 }
850 sc = gp->softc;
851 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
852 md.md_name);
853 error = add_provider_to_geom(sc, pp, &md);
854 if (error != 0) {
855 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
856 pp->name, md.md_name, error);
857 virstor_geom_destroy(sc, TRUE, FALSE);
858 return (NULL);
859 }
860 }
861
862 return (gp);
863 }
864
865 /*
866 * Destroyes consumer passed to it in arguments. Used as a callback
867 * on g_event queue.
868 */
869 static void
870 delay_destroy_consumer(void *arg, int flags __unused)
871 {
872 struct g_consumer *c = arg;
873 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
874 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
875 c->provider->name);
876 g_detach(c);
877 g_destroy_consumer(c);
878 }
879
880 /*
881 * Remove a component (consumer) from geom instance; If it's the first
882 * component being removed, orphan the provider to announce geom's being
883 * dismantled
884 */
885 static void
886 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
887 boolean_t delay)
888 {
889 struct g_consumer *c;
890
891 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
892 sc->geom->name));
893 c = comp->gcons;
894
895 comp->gcons = NULL;
896 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
897 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
898 sc->geom->name);
899 if (sc->provider != NULL) {
900 /* Whither, GEOM? */
901 sc->provider->flags |= G_PF_WITHER;
902 g_orphan_provider(sc->provider, ENXIO);
903 sc->provider = NULL;
904 LOG_MSG(LVL_INFO, "Removing provider %s", sc->geom->name);
905 }
906
907 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
908 g_access(c, -c->acr, -c->acw, -c->ace);
909 if (delay) {
910 /* Destroy consumer after it's tasted */
911 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
912 } else {
913 g_detach(c);
914 g_destroy_consumer(c);
915 }
916 }
917
918 /*
919 * Destroy geom - called internally
920 * See g_virstor_destroy_geom for the other one
921 */
922 static int
923 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
924 boolean_t delay)
925 {
926 struct g_provider *pp;
927 struct g_geom *gp;
928 int n;
929
930 g_topology_assert();
931
932 if (sc == NULL)
933 return (ENXIO);
934
935 pp = sc->provider;
936 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
937 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
938 "Device %s is still open.", pp->name);
939 if (!force)
940 return (EBUSY);
941 }
942
943 for (n = 0; n < sc->n_components; n++) {
944 if (sc->components[n].gcons != NULL)
945 remove_component(sc, &sc->components[n], delay);
946 }
947
948 gp = sc->geom;
949 gp->softc = NULL;
950
951 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
952 gp->name));
953
954 /* XXX: This might or might not work, since we're called with
955 * the topology lock held. Also, it might panic the kernel if
956 * the error'd BIO is in softupdates code. */
957 mtx_lock(&sc->delayed_bio_q_mtx);
958 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
959 struct g_virstor_bio_q *bq;
960 bq = STAILQ_FIRST(&sc->delayed_bio_q);
961 bq->bio->bio_error = ENOSPC;
962 g_io_deliver(bq->bio, EIO);
963 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
964 free(bq, M_GVIRSTOR);
965 }
966 mtx_unlock(&sc->delayed_bio_q_mtx);
967 mtx_destroy(&sc->delayed_bio_q_mtx);
968
969 free(sc->map, M_GVIRSTOR);
970 free(sc->components, M_GVIRSTOR);
971 bzero(sc, sizeof *sc);
972 free(sc, M_GVIRSTOR);
973
974 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
975 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
976 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
977
978 g_wither_geom(gp, ENXIO);
979
980 return (0);
981 }
982
983 /*
984 * Utility function: read metadata & decode. Wants topology lock to be
985 * held.
986 */
987 static int
988 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
989 {
990 struct g_provider *pp;
991 char *buf;
992 int error;
993
994 g_topology_assert();
995 error = g_access(cp, 1, 0, 0);
996 if (error != 0)
997 return (error);
998 pp = cp->provider;
999 g_topology_unlock();
1000 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
1001 &error);
1002 g_topology_lock();
1003 g_access(cp, -1, 0, 0);
1004 if (buf == NULL)
1005 return (error);
1006
1007 virstor_metadata_decode(buf, md);
1008 g_free(buf);
1009
1010 return (0);
1011 }
1012
1013 /**
1014 * Utility function: encode & write metadata. Assumes topology lock is
1015 * held.
1016 *
1017 * There is no useful way of recovering from errors in this function,
1018 * not involving panicking the kernel. If the metadata cannot be written
1019 * the most we can do is notify the operator and hope he spots it and
1020 * replaces the broken drive.
1021 */
1022 static void
1023 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1024 {
1025 struct g_provider *pp;
1026 char *buf;
1027 int error;
1028
1029 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1030 ("Something's fishy in %s", __func__));
1031 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1032 g_topology_assert();
1033 error = g_access(cp, 0, 1, 0);
1034 if (error != 0) {
1035 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d",
1036 cp->provider->name, error);
1037 return;
1038 }
1039 pp = cp->provider;
1040
1041 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1042 virstor_metadata_encode(md, buf);
1043 g_topology_unlock();
1044 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1045 pp->sectorsize);
1046 g_topology_lock();
1047 g_access(cp, 0, -1, 0);
1048 free(buf, M_GVIRSTOR);
1049
1050 if (error != 0)
1051 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s",
1052 error, cp->provider->name);
1053 }
1054
1055 /*
1056 * Creates a new instance of this GEOM class, initialise softc
1057 */
1058 static struct g_geom *
1059 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1060 {
1061 struct g_geom *gp;
1062 struct g_virstor_softc *sc;
1063
1064 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1065 md->md_name, md->md_id);
1066
1067 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1068 md->md_virsize < md->md_chunk_size) {
1069 /* This is bogus configuration, and probably means data is
1070 * somehow corrupted. Panic, maybe? */
1071 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1072 md->md_name);
1073 return (NULL);
1074 }
1075
1076 /* Check if it's already created */
1077 LIST_FOREACH(gp, &mp->geom, geom) {
1078 sc = gp->softc;
1079 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1080 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1081 md->md_name);
1082 if (sc->id != md->md_id) {
1083 LOG_MSG(LVL_ERROR,
1084 "Some stale or invalid components "
1085 "exist for virstor device named %s. "
1086 "You will need to <CLEAR> all stale "
1087 "components and maybe reconfigure "
1088 "the virstor device. Tune "
1089 "kern.geom.virstor.debug sysctl up "
1090 "for more information.",
1091 sc->geom->name);
1092 }
1093 return (NULL);
1094 }
1095 }
1096 gp = g_new_geomf(mp, "%s", md->md_name);
1097 gp->softc = NULL; /* to circumevent races that test softc */
1098
1099 gp->start = g_virstor_start;
1100 gp->spoiled = g_virstor_orphan;
1101 gp->orphan = g_virstor_orphan;
1102 gp->access = g_virstor_access;
1103 gp->dumpconf = g_virstor_dumpconf;
1104
1105 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1106 sc->id = md->md_id;
1107 sc->n_components = md->md_count;
1108 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1109 M_GVIRSTOR, M_WAITOK | M_ZERO);
1110 sc->chunk_size = md->md_chunk_size;
1111 sc->virsize = md->md_virsize;
1112 STAILQ_INIT(&sc->delayed_bio_q);
1113 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1114 "gvirstor", MTX_DEF | MTX_RECURSE);
1115
1116 sc->geom = gp;
1117 sc->provider = NULL; /* virstor_check_and_run will create it */
1118 gp->softc = sc;
1119
1120 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1121
1122 return (gp);
1123 }
1124
1125 /*
1126 * Add provider to a GEOM class instance
1127 */
1128 static int
1129 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1130 struct g_virstor_metadata *md)
1131 {
1132 struct g_virstor_component *component;
1133 struct g_consumer *cp, *fcp;
1134 struct g_geom *gp;
1135 int error;
1136
1137 if (md->no >= sc->n_components)
1138 return (EINVAL);
1139
1140 /* "Current" compontent */
1141 component = &(sc->components[md->no]);
1142 if (component->gcons != NULL)
1143 return (EEXIST);
1144
1145 gp = sc->geom;
1146 fcp = LIST_FIRST(&gp->consumer);
1147
1148 cp = g_new_consumer(gp);
1149 error = g_attach(cp, pp);
1150
1151 if (error != 0) {
1152 g_destroy_consumer(cp);
1153 return (error);
1154 }
1155
1156 if (fcp != NULL) {
1157 if (fcp->provider->sectorsize != pp->sectorsize) {
1158 /* TODO: this can be made to work */
1159 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1160 "sector size (%d)", pp->name, sc->geom->name,
1161 pp->sectorsize);
1162 return (EINVAL);
1163 }
1164 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1165 /* Replicate access permissions from first "live" consumer
1166 * to the new one */
1167 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1168 if (error != 0) {
1169 g_detach(cp);
1170 g_destroy_consumer(cp);
1171 return (error);
1172 }
1173 }
1174 }
1175
1176 /* Bring up a new component */
1177 cp->private = component;
1178 component->gcons = cp;
1179 component->sc = sc;
1180 component->index = md->no;
1181 component->chunk_count = md->chunk_count;
1182 component->chunk_next = md->chunk_next;
1183 component->chunk_reserved = md->chunk_reserved;
1184 component->flags = md->flags;
1185
1186 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1187
1188 virstor_check_and_run(sc);
1189 return (0);
1190 }
1191
1192 /*
1193 * Check if everything's ready to create the geom provider & device entry,
1194 * create and start provider.
1195 * Called ultimately by .taste, from g_event thread
1196 */
1197 static void
1198 virstor_check_and_run(struct g_virstor_softc *sc)
1199 {
1200 off_t off;
1201 size_t n, count;
1202 int index;
1203 int error;
1204
1205 if (virstor_valid_components(sc) != sc->n_components)
1206 return;
1207
1208 if (virstor_valid_components(sc) == 0) {
1209 /* This is actually a candidate for panic() */
1210 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1211 sc->provider->name);
1212 return;
1213 }
1214
1215 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1216
1217 /* Initialise allocation map from the first consumer */
1218 sc->chunk_count = sc->virsize / sc->chunk_size;
1219 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1220 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1221 sc->provider->name,
1222 sc->chunk_count * (off_t)sc->chunk_size);
1223 }
1224 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1225 /* The following allocation is in order of 4MB - 8MB */
1226 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1227 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1228 __func__, sc->map_size, sc->provider->name));
1229 sc->map_sectors = sc->map_size / sc->sectorsize;
1230
1231 count = 0;
1232 for (n = 0; n < sc->n_components; n++)
1233 count += sc->components[n].chunk_count;
1234 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1235 "(%zu KB chunks)",
1236 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1237
1238 error = g_access(sc->components[0].gcons, 1, 0, 0);
1239 if (error != 0) {
1240 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1241 "read allocation map for %s",
1242 sc->components[0].gcons->provider->name,
1243 sc->geom->name);
1244 return;
1245 }
1246 /* Read in the allocation map */
1247 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1248 sc->components[0].gcons->provider->name);
1249 off = count = n = 0;
1250 while (count < sc->map_size) {
1251 struct g_virstor_map_entry *mapbuf;
1252 size_t bs;
1253
1254 bs = MIN(MAXPHYS, sc->map_size - count);
1255 if (bs % sc->sectorsize != 0) {
1256 /* Check for alignment errors */
1257 bs = (bs / sc->sectorsize) * sc->sectorsize;
1258 if (bs == 0)
1259 break;
1260 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1261 "for %s on %s", sc->geom->name,
1262 sc->components[0].gcons->provider->name);
1263 }
1264 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1265 if (mapbuf == NULL) {
1266 free(sc->map, M_GVIRSTOR);
1267 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1268 "for %s from %s (offset %ju) (error %d)",
1269 sc->geom->name,
1270 sc->components[0].gcons->provider->name,
1271 off, error);
1272 return;
1273 }
1274
1275 bcopy(mapbuf, &sc->map[n], bs);
1276 off += bs;
1277 count += bs;
1278 n += bs / sizeof *(sc->map);
1279 g_free(mapbuf);
1280 }
1281 g_access(sc->components[0].gcons, -1, 0, 0);
1282 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1283
1284 /* find first component with allocatable chunks */
1285 index = -1;
1286 for (n = 0; n < sc->n_components; n++) {
1287 if (sc->components[n].chunk_next <
1288 sc->components[n].chunk_count) {
1289 index = n;
1290 break;
1291 }
1292 }
1293 if (index == -1)
1294 /* not found? set it to the last component and handle it
1295 * later */
1296 index = sc->n_components - 1;
1297
1298 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1299 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1300 "(%d/%u: %s)", sc->geom->name,
1301 index+1,
1302 sc->n_components,
1303 sc->components[index].gcons->provider->name);
1304 }
1305 sc->curr_component = index;
1306
1307 if (sc->components[index].chunk_next >=
1308 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1309 LOG_MSG(LVL_WARNING,
1310 "Component %s of %s is running out of free space "
1311 "(%u chunks left)",
1312 sc->components[index].gcons->provider->name,
1313 sc->geom->name, sc->components[index].chunk_count -
1314 sc->components[index].chunk_next);
1315 }
1316
1317 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1318 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1319 LOG_MSG(LVL_ERROR,
1320 "%s: Map entries don't fit exactly in a sector (%s)",
1321 __func__, sc->geom->name);
1322 return;
1323 }
1324
1325 /* Recalculate allocated chunks in components & at the same time
1326 * verify map data is sane. We could trust metadata on this, but
1327 * we want to make sure. */
1328 for (n = 0; n < sc->n_components; n++)
1329 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1330
1331 for (n = 0; n < sc->chunk_count; n++) {
1332 if (sc->map[n].provider_no >= sc->n_components ||
1333 sc->map[n].provider_chunk >=
1334 sc->components[sc->map[n].provider_no].chunk_count) {
1335 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1336 __func__, (u_int)n, sc->geom->name);
1337 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1338 " provider_chunk: %u, chunk_count: %u", __func__,
1339 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 return;
1343 }
1344 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1345 sc->components[sc->map[n].provider_no].chunk_next++;
1346 }
1347
1348 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1349 sc->geom->name);
1350
1351 sc->provider->sectorsize = sc->sectorsize;
1352 sc->provider->mediasize = sc->virsize;
1353 g_error_provider(sc->provider, 0);
1354
1355 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1356 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1357 "chunk %u", sc->provider->name, sc->curr_component,
1358 sc->components[sc->curr_component].chunk_next);
1359 }
1360
1361 /*
1362 * Returns count of active providers in this geom instance
1363 */
1364 static u_int
1365 virstor_valid_components(struct g_virstor_softc *sc)
1366 {
1367 unsigned int nc, i;
1368
1369 nc = 0;
1370 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1371 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1372 for (i = 0; i < sc->n_components; i++)
1373 if (sc->components[i].gcons != NULL)
1374 nc++;
1375 return (nc);
1376 }
1377
1378 /*
1379 * Called when the consumer gets orphaned (?)
1380 */
1381 static void
1382 g_virstor_orphan(struct g_consumer *cp)
1383 {
1384 struct g_virstor_softc *sc;
1385 struct g_virstor_component *comp;
1386 struct g_geom *gp;
1387
1388 g_topology_assert();
1389 gp = cp->geom;
1390 sc = gp->softc;
1391 if (sc == NULL)
1392 return;
1393
1394 comp = cp->private;
1395 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1396 __func__));
1397 remove_component(sc, comp, FALSE);
1398 if (virstor_valid_components(sc) == 0)
1399 virstor_geom_destroy(sc, TRUE, FALSE);
1400 }
1401
1402 /*
1403 * Called to notify geom when it's been opened, and for what intent
1404 */
1405 static int
1406 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1407 {
1408 struct g_consumer *c;
1409 struct g_virstor_softc *sc;
1410 struct g_geom *gp;
1411 int error;
1412
1413 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1414 gp = pp->geom;
1415 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1416 sc = gp->softc;
1417
1418 if (sc == NULL) {
1419 /* It seems that .access can be called with negative dr,dw,dx
1420 * in this case but I want to check for myself */
1421 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s",
1422 dr, dw, de, pp->name);
1423 /* This should only happen when geom is withered so
1424 * allow only negative requests */
1425 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
1426 ("%s: Positive access for %s", __func__, pp->name));
1427 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
1428 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed",
1429 pp->name);
1430 return (0);
1431 }
1432
1433 /* Grab an exclusive bit to propagate on our consumers on first open */
1434 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1435 de++;
1436 /* ... drop it on close */
1437 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1438 de--;
1439 update_metadata(sc); /* Writes statistical information */
1440 }
1441
1442 error = ENXIO;
1443 LIST_FOREACH(c, &gp->consumer, consumer) {
1444 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
1445 error = g_access(c, dr, dw, de);
1446 if (error != 0) {
1447 struct g_consumer *c2;
1448
1449 /* Backout earlier changes */
1450 LIST_FOREACH(c2, &gp->consumer, consumer) {
1451 if (c2 == c) /* all eariler components fixed */
1452 return (error);
1453 g_access(c2, -dr, -dw, -de);
1454 }
1455 }
1456 }
1457
1458 return (error);
1459 }
1460
1461 /*
1462 * Generate XML dump of current state
1463 */
1464 static void
1465 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1466 struct g_consumer *cp, struct g_provider *pp)
1467 {
1468 struct g_virstor_softc *sc;
1469
1470 g_topology_assert();
1471 sc = gp->softc;
1472
1473 if (sc == NULL || pp != NULL)
1474 return;
1475
1476 if (cp != NULL) {
1477 /* For each component */
1478 struct g_virstor_component *comp;
1479
1480 comp = cp->private;
1481 if (comp == NULL)
1482 return;
1483 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1484 indent, comp->index);
1485 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1486 indent, comp->chunk_count);
1487 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1488 indent, comp->chunk_next);
1489 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1490 indent, comp->chunk_reserved);
1491 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1492 indent,
1493 comp->chunk_next > 0 ? 100 -
1494 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1495 comp->chunk_count : 100);
1496 } else {
1497 /* For the whole thing */
1498 u_int count, used, i;
1499 off_t size;
1500
1501 count = used = size = 0;
1502 for (i = 0; i < sc->n_components; i++) {
1503 if (sc->components[i].gcons != NULL) {
1504 count += sc->components[i].chunk_count;
1505 used += sc->components[i].chunk_next +
1506 sc->components[i].chunk_reserved;
1507 size += sc->components[i].gcons->
1508 provider->mediasize;
1509 }
1510 }
1511
1512 sbuf_printf(sb, "%s<Status>"
1513 "Components=%u, Online=%u</Status>\n", indent,
1514 sc->n_components, virstor_valid_components(sc));
1515 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1516 indent, 100-(used * 100) / count);
1517 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1518 sc->chunk_size);
1519 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1520 indent, used > 0 ? 100 - (used * 100) / count : 100);
1521 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1522 indent, count);
1523 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1524 indent, sc->chunk_count);
1525 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1526 indent,
1527 (count * 100) / sc->chunk_count);
1528 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1529 indent, size);
1530 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1531 sc->virsize);
1532 }
1533 }
1534
1535 /*
1536 * GEOM .done handler
1537 * Can't use standard handler because one requested IO may
1538 * fork into additional data IOs
1539 */
1540 static void
1541 g_virstor_done(struct bio *b)
1542 {
1543 struct g_virstor_softc *sc;
1544 struct bio *parent_b;
1545
1546 parent_b = b->bio_parent;
1547 sc = parent_b->bio_to->geom->softc;
1548
1549 if (b->bio_error != 0) {
1550 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1551 b->bio_error, b->bio_offset, b->bio_length,
1552 b->bio_to->name);
1553 if (parent_b->bio_error == 0)
1554 parent_b->bio_error = b->bio_error;
1555 }
1556
1557 parent_b->bio_inbed++;
1558 parent_b->bio_completed += b->bio_completed;
1559
1560 if (parent_b->bio_children == parent_b->bio_inbed) {
1561 parent_b->bio_completed = parent_b->bio_length;
1562 g_io_deliver(parent_b, parent_b->bio_error);
1563 }
1564 g_destroy_bio(b);
1565 }
1566
1567 /*
1568 * I/O starts here
1569 * Called in g_down thread
1570 */
1571 static void
1572 g_virstor_start(struct bio *b)
1573 {
1574 struct g_virstor_softc *sc;
1575 struct g_virstor_component *comp;
1576 struct bio *cb;
1577 struct g_provider *pp;
1578 char *addr;
1579 off_t offset, length;
1580 struct bio_queue_head bq;
1581 size_t chunk_size; /* cached for convenience */
1582 u_int count;
1583
1584 pp = b->bio_to;
1585 sc = pp->geom->softc;
1586 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1587 b->bio_to->error, b->bio_to->name));
1588
1589 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1590
1591 switch (b->bio_cmd) {
1592 case BIO_READ:
1593 case BIO_WRITE:
1594 case BIO_DELETE:
1595 break;
1596 default:
1597 g_io_deliver(b, EOPNOTSUPP);
1598 return;
1599 }
1600
1601 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1602 bioq_init(&bq);
1603
1604 chunk_size = sc->chunk_size;
1605 addr = b->bio_data;
1606 offset = b->bio_offset; /* virtual offset and length */
1607 length = b->bio_length;
1608
1609 while (length > 0) {
1610 size_t chunk_index, in_chunk_offset, in_chunk_length;
1611 struct virstor_map_entry *me;
1612
1613 chunk_index = offset / chunk_size; /* round downwards */
1614 in_chunk_offset = offset % chunk_size;
1615 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1616 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1617 b->bio_cmd == BIO_READ ? "R" : "W",
1618 offset, length,
1619 chunk_index, in_chunk_offset, in_chunk_length);
1620 me = &sc->map[chunk_index];
1621
1622 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1623 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1624 /* Reads from unallocated chunks return zeroed
1625 * buffers */
1626 if (b->bio_cmd == BIO_READ)
1627 bzero(addr, in_chunk_length);
1628 } else {
1629 comp = &sc->components[me->provider_no];
1630
1631 cb = g_clone_bio(b);
1632 if (cb == NULL) {
1633 bioq_dismantle(&bq);
1634 if (b->bio_error == 0)
1635 b->bio_error = ENOMEM;
1636 g_io_deliver(b, b->bio_error);
1637 return;
1638 }
1639 cb->bio_to = comp->gcons->provider;
1640 cb->bio_done = g_virstor_done;
1641 cb->bio_offset =
1642 (off_t)me->provider_chunk * (off_t)chunk_size
1643 + in_chunk_offset;
1644 cb->bio_length = in_chunk_length;
1645 cb->bio_data = addr;
1646 cb->bio_caller1 = comp;
1647 bioq_disksort(&bq, cb);
1648 }
1649 } else { /* handle BIO_WRITE */
1650 KASSERT(b->bio_cmd == BIO_WRITE,
1651 ("%s: Unknown command %d", __func__,
1652 b->bio_cmd));
1653
1654 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1655 /* We have a virtual chunk, represented by
1656 * the "me" entry, but it's not yet allocated
1657 * (tied to) a physical chunk. So do it now. */
1658 struct virstor_map_entry *data_me;
1659 u_int phys_chunk, comp_no;
1660 off_t s_offset;
1661 int error;
1662
1663 error = allocate_chunk(sc, &comp, &comp_no,
1664 &phys_chunk);
1665 if (error != 0) {
1666 /* We cannot allocate a physical chunk
1667 * to satisfy this request, so we'll
1668 * delay it to when we can...
1669 * XXX: this will prevent the fs from
1670 * being umounted! */
1671 struct g_virstor_bio_q *biq;
1672 biq = malloc(sizeof *biq, M_GVIRSTOR,
1673 M_NOWAIT);
1674 if (biq == NULL) {
1675 bioq_dismantle(&bq);
1676 if (b->bio_error == 0)
1677 b->bio_error = ENOMEM;
1678 g_io_deliver(b, b->bio_error);
1679 return;
1680 }
1681 biq->bio = b;
1682 mtx_lock(&sc->delayed_bio_q_mtx);
1683 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1684 biq, linkage);
1685 mtx_unlock(&sc->delayed_bio_q_mtx);
1686 LOG_MSG(LVL_WARNING, "Delaying BIO "
1687 "(size=%ju) until free physical "
1688 "space can be found on %s",
1689 b->bio_length,
1690 sc->provider->name);
1691 return;
1692 }
1693 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1694 "for %s",
1695 phys_chunk,
1696 comp->gcons->provider->name,
1697 sc->provider->name);
1698
1699 me->provider_no = comp_no;
1700 me->provider_chunk = phys_chunk;
1701 me->flags |= VIRSTOR_MAP_ALLOCATED;
1702
1703 cb = g_clone_bio(b);
1704 if (cb == NULL) {
1705 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1706 me->provider_no = 0;
1707 me->provider_chunk = 0;
1708 bioq_dismantle(&bq);
1709 if (b->bio_error == 0)
1710 b->bio_error = ENOMEM;
1711 g_io_deliver(b, b->bio_error);
1712 return;
1713 }
1714
1715 /* The allocation table is stored continuously
1716 * at the start of the drive. We need to
1717 * calculate the offset of the sector that holds
1718 * this map entry both on the drive and in the
1719 * map array.
1720 * sc_offset will end up pointing to the drive
1721 * sector. */
1722 s_offset = chunk_index * sizeof *me;
1723 s_offset = (s_offset / sc->sectorsize) *
1724 sc->sectorsize;
1725
1726 /* data_me points to map entry sector
1727 * in memory (analoguos to offset) */
1728 data_me = &sc->map[(chunk_index /
1729 sc->me_per_sector) * sc->me_per_sector];
1730
1731 /* Commit sector with map entry to storage */
1732 cb->bio_to = sc->components[0].gcons->provider;
1733 cb->bio_done = g_virstor_done;
1734 cb->bio_offset = s_offset;
1735 cb->bio_data = (char *)data_me;
1736 cb->bio_length = sc->sectorsize;
1737 cb->bio_caller1 = &sc->components[0];
1738 bioq_disksort(&bq, cb);
1739 }
1740
1741 comp = &sc->components[me->provider_no];
1742 cb = g_clone_bio(b);
1743 if (cb == NULL) {
1744 bioq_dismantle(&bq);
1745 if (b->bio_error == 0)
1746 b->bio_error = ENOMEM;
1747 g_io_deliver(b, b->bio_error);
1748 return;
1749 }
1750 /* Finally, handle the data */
1751 cb->bio_to = comp->gcons->provider;
1752 cb->bio_done = g_virstor_done;
1753 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1754 in_chunk_offset;
1755 cb->bio_length = in_chunk_length;
1756 cb->bio_data = addr;
1757 cb->bio_caller1 = comp;
1758 bioq_disksort(&bq, cb);
1759 }
1760 addr += in_chunk_length;
1761 length -= in_chunk_length;
1762 offset += in_chunk_length;
1763 }
1764
1765 /* Fire off bio's here */
1766 count = 0;
1767 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1768 bioq_remove(&bq, cb);
1769 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1770 comp = cb->bio_caller1;
1771 cb->bio_caller1 = NULL;
1772 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1773 cb->bio_offset, cb->bio_length);
1774 g_io_request(cb, comp->gcons);
1775 count++;
1776 }
1777 if (count == 0) { /* We handled everything locally */
1778 b->bio_completed = b->bio_length;
1779 g_io_deliver(b, 0);
1780 }
1781
1782 }
1783
1784 /*
1785 * Allocate a chunk from a physical provider. Returns physical component,
1786 * chunk index relative to the component and the component's index.
1787 */
1788 static int
1789 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1790 u_int *comp_no_p, u_int *chunk)
1791 {
1792 u_int comp_no;
1793
1794 KASSERT(sc->curr_component < sc->n_components,
1795 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1796
1797 comp_no = sc->curr_component;
1798 *comp = &sc->components[comp_no];
1799 dump_component(*comp);
1800 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1801 /* This component is full. Allocate next component */
1802 if (comp_no >= sc->n_components-1) {
1803 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1804 sc->geom->name);
1805 return (-1);
1806 }
1807 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1808 sc->curr_component = ++comp_no;
1809
1810 *comp = &sc->components[comp_no];
1811 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1812 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1813 "(switching to %u/%u: %s)", sc->geom->name,
1814 comp_no+1, sc->n_components,
1815 (*comp)->gcons->provider->name);
1816 /* Take care not to overwrite reserved chunks */
1817 if ( (*comp)->chunk_reserved > 0 &&
1818 (*comp)->chunk_next < (*comp)->chunk_reserved)
1819 (*comp)->chunk_next = (*comp)->chunk_reserved;
1820
1821 (*comp)->flags |=
1822 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1823 dump_component(*comp);
1824 *comp_no_p = comp_no;
1825 *chunk = (*comp)->chunk_next++;
1826 } else {
1827 *comp_no_p = comp_no;
1828 *chunk = (*comp)->chunk_next++;
1829 }
1830 return (0);
1831 }
1832
1833 /* Dump a component */
1834 static void
1835 dump_component(struct g_virstor_component *comp)
1836 {
1837
1838 if (g_virstor_debug < LVL_DEBUG2)
1839 return;
1840 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1841 printf(" chunk_count: %u\n", comp->chunk_count);
1842 printf(" chunk_next: %u\n", comp->chunk_next);
1843 printf(" flags: %u\n", comp->flags);
1844 }
1845
1846 #if 0
1847 /* Dump a map entry */
1848 static void
1849 dump_me(struct virstor_map_entry *me, unsigned int nr)
1850 {
1851 if (g_virstor_debug < LVL_DEBUG)
1852 return;
1853 printf("VIRT. CHUNK #%d: ", nr);
1854 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1855 printf("(unallocated)\n");
1856 else
1857 printf("allocated at provider %u, provider_chunk %u\n",
1858 me->provider_no, me->provider_chunk);
1859 }
1860 #endif
1861
1862 /*
1863 * Dismantle bio_queue and destroy its components
1864 */
1865 static void
1866 bioq_dismantle(struct bio_queue_head *bq)
1867 {
1868 struct bio *b;
1869
1870 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1871 bioq_remove(bq, b);
1872 g_destroy_bio(b);
1873 }
1874 }
1875
1876 /*
1877 * The function that shouldn't be called.
1878 * When this is called, the stack is already garbled because of
1879 * argument mismatch. There's nothing to do now but panic, which is
1880 * accidentally the whole purpose of this function.
1881 * Motivation: to guard from accidentally calling geom methods when
1882 * they shouldn't be called. (see g_..._taste)
1883 */
1884 static void
1885 invalid_call(void)
1886 {
1887 panic("invalid_call() has just been called. Something's fishy here.");
1888 }
1889
1890 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */
Cache object: 0e12e365bda2242f2fbdf3420f3b14f5
|