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/sysctl.h>
44 #include <sys/malloc.h>
45 #include <sys/time.h>
46 #include <sys/proc.h>
47 #include <sys/kthread.h>
48 #include <sys/mutex.h>
49 #include <vm/uma.h>
50 #include <geom/geom.h>
51
52 #include <geom/virstor/g_virstor.h>
53 #include <geom/virstor/g_virstor_md.h>
54
55 /* Declare malloc(9) label */
56 static MALLOC_DEFINE(M_GVIRSTOR, "gvirstor", "GEOM_VIRSTOR Data");
57
58 /* GEOM class methods */
59 static g_init_t g_virstor_init;
60 static g_fini_t g_virstor_fini;
61 static g_taste_t g_virstor_taste;
62 static g_ctl_req_t g_virstor_config;
63 static g_ctl_destroy_geom_t g_virstor_destroy_geom;
64
65 /* Declare & initialize class structure ("geom class") */
66 struct g_class g_virstor_class = {
67 .name = G_VIRSTOR_CLASS_NAME,
68 .version = G_VERSION,
69 .init = g_virstor_init,
70 .fini = g_virstor_fini,
71 .taste = g_virstor_taste,
72 .ctlreq = g_virstor_config,
73 .destroy_geom = g_virstor_destroy_geom
74 /* The .dumpconf and the rest are only usable for a geom instance, so
75 * they will be set when such instance is created. */
76 };
77
78 /* Declare sysctl's and loader tunables */
79 SYSCTL_DECL(_kern_geom);
80 SYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0, "GEOM_GVIRSTOR information");
81
82 static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */
83 TUNABLE_INT("kern.geom.virstor.debug", &g_virstor_debug);
84 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RW, &g_virstor_debug,
85 0, "Debug level (2=production, 5=normal, 15=excessive)");
86
87 static u_int g_virstor_chunk_watermark = 100;
88 TUNABLE_INT("kern.geom.virstor.chunk_watermark", &g_virstor_chunk_watermark);
89 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RW,
90 &g_virstor_chunk_watermark, 0,
91 "Minimum number of free chunks before issuing administrative warning");
92
93 static u_int g_virstor_component_watermark = 1;
94 TUNABLE_INT("kern.geom.virstor.component_watermark",
95 &g_virstor_component_watermark);
96 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RW,
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 int 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 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
235 sc->geom->name);
236 update_metadata(sc);
237 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
238 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
239 sc->geom->name, error);
240 }
241 }
242 g_topology_unlock();
243 }
244
245 /*
246 * "add" verb from userland - add new component(s) to the structure.
247 * This will be done all at once in here, without going through the
248 * .taste function for new components.
249 */
250 static void
251 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
252 {
253 /* Note: while this is going on, I/O is being done on
254 * the g_up and g_down threads. The idea is to make changes
255 * to softc members in a way that can atomically activate
256 * them all at once. */
257 struct g_virstor_softc *sc;
258 int *hardcode, *nargs;
259 const char *geom_name; /* geom to add a component to */
260 struct g_consumer *fcp;
261 struct g_virstor_bio_q *bq;
262 u_int added;
263 int error;
264 int i;
265
266 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
267 if (nargs == NULL) {
268 gctl_error(req, "Error fetching argument '%s'", "nargs");
269 return;
270 }
271 if (*nargs < 2) {
272 gctl_error(req, "Invalid number of arguments");
273 return;
274 }
275 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
276 if (hardcode == NULL) {
277 gctl_error(req, "Error fetching argument '%s'", "hardcode");
278 return;
279 }
280
281 /* Find "our" geom */
282 geom_name = gctl_get_asciiparam(req, "arg0");
283 if (geom_name == NULL) {
284 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
285 return;
286 }
287 sc = virstor_find_geom(cp, geom_name);
288 if (sc == NULL) {
289 gctl_error(req, "Don't know anything about '%s'", geom_name);
290 return;
291 }
292
293 if (virstor_valid_components(sc) != sc->n_components) {
294 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
295 "virstor %s", sc->geom->name);
296 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
297 return;
298 }
299
300 fcp = sc->components[0].gcons;
301 added = 0;
302 g_topology_lock();
303 for (i = 1; i < *nargs; i++) {
304 struct g_virstor_metadata md;
305 char aname[8];
306 const char *prov_name;
307 struct g_provider *pp;
308 struct g_consumer *cp;
309 u_int nc;
310 u_int j;
311
312 snprintf(aname, sizeof aname, "arg%d", i);
313 prov_name = gctl_get_asciiparam(req, aname);
314 if (strncmp(prov_name, _PATH_DEV, strlen(_PATH_DEV)) == 0)
315 prov_name += strlen(_PATH_DEV);
316
317 pp = g_provider_by_name(prov_name);
318 if (pp == NULL) {
319 /* This is the most common error so be verbose about it */
320 if (added != 0) {
321 gctl_error(req, "Invalid provider: '%s' (added"
322 " %u components)", prov_name, added);
323 update_metadata(sc);
324 } else {
325 gctl_error(req, "Invalid provider: '%s'",
326 prov_name);
327 }
328 g_topology_unlock();
329 return;
330 }
331 cp = g_new_consumer(sc->geom);
332 if (cp == NULL) {
333 gctl_error(req, "Cannot create consumer");
334 g_topology_unlock();
335 return;
336 }
337 error = g_attach(cp, pp);
338 if (error != 0) {
339 gctl_error(req, "Cannot attach a consumer to %s",
340 pp->name);
341 g_destroy_consumer(cp);
342 g_topology_unlock();
343 return;
344 }
345 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
346 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
347 if (error != 0) {
348 gctl_error(req, "Access request failed for %s",
349 pp->name);
350 g_destroy_consumer(cp);
351 g_topology_unlock();
352 return;
353 }
354 }
355 if (fcp->provider->sectorsize != pp->sectorsize) {
356 gctl_error(req, "Sector size doesn't fit for %s",
357 pp->name);
358 g_destroy_consumer(cp);
359 g_topology_unlock();
360 return;
361 }
362 for (j = 0; j < sc->n_components; j++) {
363 if (strcmp(sc->components[j].gcons->provider->name,
364 pp->name) == 0) {
365 gctl_error(req, "Component %s already in %s",
366 pp->name, sc->geom->name);
367 g_destroy_consumer(cp);
368 g_topology_unlock();
369 return;
370 }
371 }
372 sc->components = realloc(sc->components,
373 sizeof(*sc->components) * (sc->n_components + 1),
374 M_GVIRSTOR, M_WAITOK);
375
376 nc = sc->n_components;
377 sc->components[nc].gcons = cp;
378 sc->components[nc].sc = sc;
379 sc->components[nc].index = nc;
380 sc->components[nc].chunk_count = cp->provider->mediasize /
381 sc->chunk_size;
382 sc->components[nc].chunk_next = 0;
383 sc->components[nc].chunk_reserved = 0;
384
385 if (sc->components[nc].chunk_count < 4) {
386 gctl_error(req, "Provider too small: %s",
387 cp->provider->name);
388 g_destroy_consumer(cp);
389 g_topology_unlock();
390 return;
391 }
392 fill_metadata(sc, &md, nc, *hardcode);
393 write_metadata(cp, &md);
394 /* The new component becomes visible when n_components is
395 * incremented */
396 sc->n_components++;
397 added++;
398
399 }
400 /* This call to update_metadata() is critical. In case there's a
401 * power failure in the middle of it and some components are updated
402 * while others are not, there will be trouble on next .taste() iff
403 * a non-updated component is detected first */
404 update_metadata(sc);
405 g_topology_unlock();
406 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
407 sc->geom->name);
408 /* Fire off BIOs previously queued because there wasn't any
409 * physical space left. If the BIOs still can't be satisfied
410 * they will again be added to the end of the queue (during
411 * which the mutex will be recursed) */
412 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
413 bq->bio = NULL;
414 mtx_lock(&sc->delayed_bio_q_mtx);
415 /* First, insert a sentinel to the queue end, so we don't
416 * end up in an infinite loop if there's still no free
417 * space available. */
418 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
419 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
420 bq = STAILQ_FIRST(&sc->delayed_bio_q);
421 if (bq->bio != NULL) {
422 g_virstor_start(bq->bio);
423 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
424 free(bq, M_GVIRSTOR);
425 } else {
426 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
427 free(bq, M_GVIRSTOR);
428 break;
429 }
430 }
431 mtx_unlock(&sc->delayed_bio_q_mtx);
432
433 }
434
435 /*
436 * Find a geom handled by the class
437 */
438 static struct g_virstor_softc *
439 virstor_find_geom(const struct g_class *cp, const char *name)
440 {
441 struct g_geom *gp;
442
443 LIST_FOREACH(gp, &cp->geom, geom) {
444 if (strcmp(name, gp->name) == 0)
445 return (gp->softc);
446 }
447 return (NULL);
448 }
449
450 /*
451 * Update metadata on all components to reflect the current state
452 * of these fields:
453 * - chunk_next
454 * - flags
455 * - md_count
456 * Expects things to be set up so write_metadata() can work, i.e.
457 * the topology lock must be held.
458 */
459 static void
460 update_metadata(struct g_virstor_softc *sc)
461 {
462 struct g_virstor_metadata md;
463 int n;
464
465 if (virstor_valid_components(sc) != sc->n_components)
466 return; /* Incomplete device */
467 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
468 sc->geom->name);
469 /* Update metadata on components */
470 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
471 sc->geom->class->name, sc->geom->name);
472 g_topology_assert();
473 for (n = 0; n < sc->n_components; n++) {
474 read_metadata(sc->components[n].gcons, &md);
475 md.chunk_next = sc->components[n].chunk_next;
476 md.flags = sc->components[n].flags;
477 md.md_count = sc->n_components;
478 write_metadata(sc->components[n].gcons, &md);
479 }
480 }
481
482 /*
483 * Fills metadata (struct md) from information stored in softc and the nc'th
484 * component of virstor
485 */
486 static void
487 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
488 u_int nc, u_int hardcode)
489 {
490 struct g_virstor_component *c;
491
492 bzero(md, sizeof *md);
493 c = &sc->components[nc];
494
495 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
496 md->md_version = G_VIRSTOR_VERSION;
497 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
498 md->md_id = sc->id;
499 md->md_virsize = sc->virsize;
500 md->md_chunk_size = sc->chunk_size;
501 md->md_count = sc->n_components;
502
503 if (hardcode) {
504 strncpy(md->provider, c->gcons->provider->name,
505 sizeof md->provider);
506 }
507 md->no = nc;
508 md->provsize = c->gcons->provider->mediasize;
509 md->chunk_count = c->chunk_count;
510 md->chunk_next = c->chunk_next;
511 md->chunk_reserved = c->chunk_reserved;
512 md->flags = c->flags;
513 }
514
515 /*
516 * Remove a component from virstor device.
517 * Can only be done if the component is unallocated.
518 */
519 static void
520 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
521 {
522 /* As this is executed in parallel to I/O, operations on virstor
523 * structures must be as atomic as possible. */
524 struct g_virstor_softc *sc;
525 int *nargs;
526 const char *geom_name;
527 u_int removed;
528 int i;
529
530 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
531 if (nargs == NULL) {
532 gctl_error(req, "Error fetching argument '%s'", "nargs");
533 return;
534 }
535 if (*nargs < 2) {
536 gctl_error(req, "Invalid number of arguments");
537 return;
538 }
539 /* Find "our" geom */
540 geom_name = gctl_get_asciiparam(req, "arg0");
541 if (geom_name == NULL) {
542 gctl_error(req, "Error fetching argument '%s'",
543 "geom_name (arg0)");
544 return;
545 }
546 sc = virstor_find_geom(cp, geom_name);
547 if (sc == NULL) {
548 gctl_error(req, "Don't know anything about '%s'", geom_name);
549 return;
550 }
551
552 if (virstor_valid_components(sc) != sc->n_components) {
553 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
554 "virstor %s", sc->geom->name);
555 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
556 return;
557 }
558
559 removed = 0;
560 for (i = 1; i < *nargs; i++) {
561 char param[8];
562 const char *prov_name;
563 int j, found;
564 struct g_virstor_component *newcomp, *compbak;
565
566 sprintf(param, "arg%d", i);
567 prov_name = gctl_get_asciiparam(req, param);
568 if (strncmp(prov_name, _PATH_DEV, strlen(_PATH_DEV)) == 0)
569 prov_name += strlen(_PATH_DEV);
570
571 found = -1;
572 for (j = 0; j < sc->n_components; j++) {
573 if (strcmp(sc->components[j].gcons->provider->name,
574 prov_name) == 0) {
575 found = j;
576 break;
577 }
578 }
579 if (found == -1) {
580 LOG_MSG(LVL_ERROR, "No %s component in %s",
581 prov_name, sc->geom->name);
582 continue;
583 }
584
585 compbak = sc->components;
586 newcomp = malloc(sc->n_components * sizeof(*sc->components),
587 M_GVIRSTOR, M_WAITOK | M_ZERO);
588 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
589 bcopy(&sc->components[found + 1], newcomp + found,
590 found * sizeof(*sc->components));
591 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
592 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
593 "removed from %s",
594 prov_name, sc->geom->name);
595 free(newcomp, M_GVIRSTOR);
596 /* We'll consider this non-fatal error */
597 continue;
598 }
599 /* Renumerate unallocated components */
600 for (j = 0; j < sc->n_components-1; j++) {
601 if ((sc->components[j].flags &
602 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
603 sc->components[j].index = j;
604 }
605 }
606 /* This is the critical section. If a component allocation
607 * event happens while both variables are not yet set,
608 * there will be trouble. Something will panic on encountering
609 * NULL sc->components[x].gcomp member.
610 * Luckily, component allocation happens very rarely and
611 * removing components is an abnormal action in any case. */
612 sc->components = newcomp;
613 sc->n_components--;
614 /* End critical section */
615
616 g_topology_lock();
617 if (clear_metadata(&compbak[found]) != 0) {
618 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
619 "metadata on %s", prov_name);
620 }
621 g_detach(compbak[found].gcons);
622 g_destroy_consumer(compbak[found].gcons);
623 g_topology_unlock();
624
625 free(compbak, M_GVIRSTOR);
626
627 removed++;
628 }
629
630 /* This call to update_metadata() is critical. In case there's a
631 * power failure in the middle of it and some components are updated
632 * while others are not, there will be trouble on next .taste() iff
633 * a non-updated component is detected first */
634 g_topology_lock();
635 update_metadata(sc);
636 g_topology_unlock();
637 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
638 sc->geom->name);
639 }
640
641 /*
642 * Clear metadata sector on component
643 */
644 static int
645 clear_metadata(struct g_virstor_component *comp)
646 {
647 char *buf;
648 int error;
649
650 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
651 comp->gcons->provider->name);
652 g_topology_assert();
653 error = g_access(comp->gcons, 0, 1, 0);
654 if (error != 0)
655 return (error);
656 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
657 M_WAITOK | M_ZERO);
658 error = g_write_data(comp->gcons,
659 comp->gcons->provider->mediasize -
660 comp->gcons->provider->sectorsize,
661 buf,
662 comp->gcons->provider->sectorsize);
663 free(buf, M_GVIRSTOR);
664 g_access(comp->gcons, 0, -1, 0);
665 return (error);
666 }
667
668 /*
669 * Destroy geom forcibly.
670 */
671 static int
672 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
673 struct g_geom *gp)
674 {
675 struct g_virstor_softc *sc;
676 int exitval;
677
678 sc = gp->softc;
679 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
680
681 exitval = 0;
682 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
683 gp->softc);
684
685 if (sc != NULL) {
686 #ifdef INVARIANTS
687 char *buf;
688 int error;
689 off_t off;
690 int isclean, count;
691 int n;
692
693 LOG_MSG(LVL_INFO, "INVARIANTS detected");
694 LOG_MSG(LVL_INFO, "Verifying allocation "
695 "table for %s", sc->geom->name);
696 count = 0;
697 for (n = 0; n < sc->chunk_count; n++) {
698 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
699 count++;
700 }
701 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
702 sc->geom->name, count);
703 n = off = count = 0;
704 isclean = 1;
705 if (virstor_valid_components(sc) != sc->n_components) {
706 /* This is a incomplete virstor device (not all
707 * components have been found) */
708 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
709 sc->geom->name);
710 goto bailout;
711 }
712 error = g_access(sc->components[0].gcons, 1, 0, 0);
713 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
714 error));
715 /* Compare the whole on-disk allocation table with what's
716 * currently in memory */
717 while (n < sc->chunk_count) {
718 buf = g_read_data(sc->components[0].gcons, off,
719 sc->sectorsize, &error);
720 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
721 "for read at %jd", error, off));
722 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
723 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
724 "entry %d, offset %jd", n, off);
725 isclean = 0;
726 count++;
727 }
728 n += sc->me_per_sector;
729 off += sc->sectorsize;
730 g_free(buf);
731 }
732 error = g_access(sc->components[0].gcons, -1, 0, 0);
733 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
734 __func__, error));
735 if (isclean != 1) {
736 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
737 "(%d sectors don't match, max %zu allocations)",
738 sc->geom->name, count,
739 count * sc->me_per_sector);
740 } else {
741 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
742 sc->geom->name);
743 }
744 bailout:
745 #endif
746 update_metadata(sc);
747 virstor_geom_destroy(sc, FALSE, FALSE);
748 exitval = EAGAIN;
749 } else
750 exitval = 0;
751 return (exitval);
752 }
753
754 /*
755 * Taste event (per-class callback)
756 * Examines a provider and creates geom instances if needed
757 */
758 static struct g_geom *
759 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
760 {
761 struct g_virstor_metadata md;
762 struct g_geom *gp;
763 struct g_consumer *cp;
764 struct g_virstor_softc *sc;
765 int error;
766
767 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
768 g_topology_assert();
769 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
770
771 /* We need a dummy geom to attach a consumer to the given provider */
772 gp = g_new_geomf(mp, "virstor:taste.helper");
773 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
774 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
775 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
776
777 cp = g_new_consumer(gp);
778 g_attach(cp, pp);
779 error = read_metadata(cp, &md);
780 g_detach(cp);
781 g_destroy_consumer(cp);
782 g_destroy_geom(gp);
783
784 if (error != 0)
785 return (NULL);
786
787 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
788 return (NULL);
789 if (md.md_version != G_VIRSTOR_VERSION) {
790 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
791 "to handle %s (%s) : %d should be %d",
792 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
793 return (NULL);
794 }
795 if (md.provsize != pp->mediasize)
796 return (NULL);
797
798 /* If the provider name is hardcoded, use the offered provider only
799 * if it's been offered with its proper name (the one used in
800 * the label command). */
801 if (md.provider[0] != '\0') {
802 if (strcmp(md.provider, pp->name) != 0)
803 return (NULL);
804 }
805
806 /* Iterate all geoms this class already knows about to see if a new
807 * geom instance of this class needs to be created (in case the provider
808 * is first from a (possibly) multi-consumer geom) or it just needs
809 * to be added to an existing instance. */
810 sc = NULL;
811 gp = NULL;
812 LIST_FOREACH(gp, &mp->geom, geom) {
813 sc = gp->softc;
814 if (sc == NULL)
815 continue;
816 if (strcmp(md.md_name, sc->geom->name) != 0)
817 continue;
818 if (md.md_id != sc->id)
819 continue;
820 break;
821 }
822 if (gp != NULL) { /* We found an existing geom instance; add to it */
823 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
824 error = add_provider_to_geom(sc, pp, &md);
825 if (error != 0) {
826 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
827 pp->name, md.md_name, error);
828 return (NULL);
829 }
830 } else { /* New geom instance needs to be created */
831 gp = create_virstor_geom(mp, &md);
832 if (gp == NULL) {
833 LOG_MSG(LVL_ERROR, "Error creating new instance of "
834 "class %s: %s", mp->name, md.md_name);
835 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
836 md.md_name, pp->name);
837 return (NULL);
838 }
839 sc = gp->softc;
840 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
841 md.md_name);
842 error = add_provider_to_geom(sc, pp, &md);
843 if (error != 0) {
844 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
845 pp->name, md.md_name, error);
846 virstor_geom_destroy(sc, TRUE, FALSE);
847 return (NULL);
848 }
849 }
850
851 return (gp);
852 }
853
854 /*
855 * Destroyes consumer passed to it in arguments. Used as a callback
856 * on g_event queue.
857 */
858 static void
859 delay_destroy_consumer(void *arg, int flags __unused)
860 {
861 struct g_consumer *c = arg;
862 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
863 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
864 c->provider->name);
865 g_detach(c);
866 g_destroy_consumer(c);
867 }
868
869 /*
870 * Remove a component (consumer) from geom instance; If it's the first
871 * component being removed, orphan the provider to announce geom's being
872 * dismantled
873 */
874 static void
875 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
876 boolean_t delay)
877 {
878 struct g_consumer *c;
879
880 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
881 sc->geom->name));
882 c = comp->gcons;
883
884 comp->gcons = NULL;
885 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
886 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
887 sc->geom->name);
888 if (sc->provider != NULL) {
889 /* Whither, GEOM? */
890 sc->provider->flags |= G_PF_WITHER;
891 g_orphan_provider(sc->provider, ENXIO);
892 sc->provider = NULL;
893 LOG_MSG(LVL_INFO, "Removing provider %s", sc->geom->name);
894 }
895
896 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
897 g_access(c, -c->acr, -c->acw, -c->ace);
898 if (delay) {
899 /* Destroy consumer after it's tasted */
900 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
901 } else {
902 g_detach(c);
903 g_destroy_consumer(c);
904 }
905 }
906
907 /*
908 * Destroy geom - called internally
909 * See g_virstor_destroy_geom for the other one
910 */
911 static int
912 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
913 boolean_t delay)
914 {
915 struct g_provider *pp;
916 struct g_geom *gp;
917 int n;
918
919 g_topology_assert();
920
921 if (sc == NULL)
922 return (ENXIO);
923
924 pp = sc->provider;
925 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
926 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
927 "Device %s is still open.", pp->name);
928 if (!force)
929 return (EBUSY);
930 }
931
932 for (n = 0; n < sc->n_components; n++) {
933 if (sc->components[n].gcons != NULL)
934 remove_component(sc, &sc->components[n], delay);
935 }
936
937 gp = sc->geom;
938 gp->softc = NULL;
939
940 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
941 gp->name));
942
943 /* XXX: This might or might not work, since we're called with
944 * the topology lock held. Also, it might panic the kernel if
945 * the error'd BIO is in softupdates code. */
946 mtx_lock(&sc->delayed_bio_q_mtx);
947 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
948 struct g_virstor_bio_q *bq;
949 bq = STAILQ_FIRST(&sc->delayed_bio_q);
950 bq->bio->bio_error = ENOSPC;
951 g_io_deliver(bq->bio, EIO);
952 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
953 free(bq, M_GVIRSTOR);
954 }
955 mtx_unlock(&sc->delayed_bio_q_mtx);
956 mtx_destroy(&sc->delayed_bio_q_mtx);
957
958 free(sc->map, M_GVIRSTOR);
959 free(sc->components, M_GVIRSTOR);
960 bzero(sc, sizeof *sc);
961 free(sc, M_GVIRSTOR);
962
963 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
964 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
965 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
966
967 g_wither_geom(gp, ENXIO);
968
969 return (0);
970 }
971
972 /*
973 * Utility function: read metadata & decode. Wants topology lock to be
974 * held.
975 */
976 static int
977 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
978 {
979 struct g_provider *pp;
980 char *buf;
981 int error;
982
983 g_topology_assert();
984 error = g_access(cp, 1, 0, 0);
985 if (error != 0)
986 return (error);
987 pp = cp->provider;
988 g_topology_unlock();
989 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
990 &error);
991 g_topology_lock();
992 g_access(cp, -1, 0, 0);
993 if (buf == NULL)
994 return (error);
995
996 virstor_metadata_decode(buf, md);
997 g_free(buf);
998
999 return (0);
1000 }
1001
1002 /**
1003 * Utility function: encode & write metadata. Assumes topology lock is
1004 * held.
1005 */
1006 static int
1007 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1008 {
1009 struct g_provider *pp;
1010 char *buf;
1011 int error;
1012
1013 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1014 ("Something's fishy in %s", __func__));
1015 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1016 g_topology_assert();
1017 error = g_access(cp, 0, 1, 0);
1018 if (error != 0)
1019 return (error);
1020 pp = cp->provider;
1021
1022 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1023 virstor_metadata_encode(md, buf);
1024 g_topology_unlock();
1025 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1026 pp->sectorsize);
1027 g_topology_lock();
1028 g_access(cp, 0, -1, 0);
1029
1030 free(buf, M_GVIRSTOR);
1031 return (0);
1032 }
1033
1034 /*
1035 * Creates a new instance of this GEOM class, initialise softc
1036 */
1037 static struct g_geom *
1038 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1039 {
1040 struct g_geom *gp;
1041 struct g_virstor_softc *sc;
1042
1043 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1044 md->md_name, md->md_id);
1045
1046 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1047 md->md_virsize < md->md_chunk_size) {
1048 /* This is bogus configuration, and probably means data is
1049 * somehow corrupted. Panic, maybe? */
1050 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1051 md->md_name);
1052 return (NULL);
1053 }
1054
1055 /* Check if it's already created */
1056 LIST_FOREACH(gp, &mp->geom, geom) {
1057 sc = gp->softc;
1058 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1059 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1060 md->md_name);
1061 if (sc->id != md->md_id) {
1062 LOG_MSG(LVL_ERROR,
1063 "Some stale or invalid components "
1064 "exist for virstor device named %s. "
1065 "You will need to <CLEAR> all stale "
1066 "components and maybe reconfigure "
1067 "the virstor device. Tune "
1068 "kern.geom.virstor.debug sysctl up "
1069 "for more information.",
1070 sc->geom->name);
1071 }
1072 return (NULL);
1073 }
1074 }
1075 gp = g_new_geomf(mp, "%s", md->md_name);
1076 gp->softc = NULL; /* to circumevent races that test softc */
1077
1078 gp->start = g_virstor_start;
1079 gp->spoiled = g_virstor_orphan;
1080 gp->orphan = g_virstor_orphan;
1081 gp->access = g_virstor_access;
1082 gp->dumpconf = g_virstor_dumpconf;
1083
1084 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1085 sc->id = md->md_id;
1086 sc->n_components = md->md_count;
1087 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1088 M_GVIRSTOR, M_WAITOK | M_ZERO);
1089 sc->chunk_size = md->md_chunk_size;
1090 sc->virsize = md->md_virsize;
1091 STAILQ_INIT(&sc->delayed_bio_q);
1092 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1093 "gvirstor", MTX_DEF | MTX_RECURSE);
1094
1095 sc->geom = gp;
1096 sc->provider = NULL; /* virstor_check_and_run will create it */
1097 gp->softc = sc;
1098
1099 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1100
1101 return (gp);
1102 }
1103
1104 /*
1105 * Add provider to a GEOM class instance
1106 */
1107 static int
1108 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1109 struct g_virstor_metadata *md)
1110 {
1111 struct g_virstor_component *component;
1112 struct g_consumer *cp, *fcp;
1113 struct g_geom *gp;
1114 int error;
1115
1116 if (md->no >= sc->n_components)
1117 return (EINVAL);
1118
1119 /* "Current" compontent */
1120 component = &(sc->components[md->no]);
1121 if (component->gcons != NULL)
1122 return (EEXIST);
1123
1124 gp = sc->geom;
1125 fcp = LIST_FIRST(&gp->consumer);
1126
1127 cp = g_new_consumer(gp);
1128 error = g_attach(cp, pp);
1129
1130 if (error != 0) {
1131 g_destroy_consumer(cp);
1132 return (error);
1133 }
1134
1135 if (fcp != NULL) {
1136 if (fcp->provider->sectorsize != pp->sectorsize) {
1137 /* TODO: this can be made to work */
1138 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1139 "sector size (%d)", pp->name, sc->geom->name,
1140 pp->sectorsize);
1141 return (EINVAL);
1142 }
1143 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1144 /* Replicate access permissions from first "live" consumer
1145 * to the new one */
1146 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1147 if (error != 0) {
1148 g_detach(cp);
1149 g_destroy_consumer(cp);
1150 return (error);
1151 }
1152 }
1153 }
1154
1155 /* Bring up a new component */
1156 cp->private = component;
1157 component->gcons = cp;
1158 component->sc = sc;
1159 component->index = md->no;
1160 component->chunk_count = md->chunk_count;
1161 component->chunk_next = md->chunk_next;
1162 component->chunk_reserved = md->chunk_reserved;
1163 component->flags = md->flags;
1164
1165 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1166
1167 virstor_check_and_run(sc);
1168 return (0);
1169 }
1170
1171 /*
1172 * Check if everything's ready to create the geom provider & device entry,
1173 * create and start provider.
1174 * Called ultimately by .taste, from g_event thread
1175 */
1176 static void
1177 virstor_check_and_run(struct g_virstor_softc *sc)
1178 {
1179 off_t off;
1180 size_t n, count;
1181 int index;
1182 int error;
1183
1184 if (virstor_valid_components(sc) != sc->n_components)
1185 return;
1186
1187 if (virstor_valid_components(sc) == 0) {
1188 /* This is actually a candidate for panic() */
1189 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1190 sc->provider->name);
1191 return;
1192 }
1193
1194 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1195
1196 /* Initialise allocation map from the first consumer */
1197 sc->chunk_count = sc->virsize / sc->chunk_size;
1198 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1199 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1200 sc->provider->name,
1201 sc->chunk_count * (off_t)sc->chunk_size);
1202 }
1203 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1204 /* The following allocation is in order of 4MB - 8MB */
1205 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1206 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1207 __func__, sc->map_size, sc->provider->name));
1208 sc->map_sectors = sc->map_size / sc->sectorsize;
1209
1210 count = 0;
1211 for (n = 0; n < sc->n_components; n++)
1212 count += sc->components[n].chunk_count;
1213 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1214 "(%zu KB chunks)",
1215 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1216
1217 error = g_access(sc->components[0].gcons, 1, 0, 0);
1218 if (error != 0) {
1219 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1220 "read allocation map for %s",
1221 sc->components[0].gcons->provider->name,
1222 sc->geom->name);
1223 return;
1224 }
1225 /* Read in the allocation map */
1226 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1227 sc->components[0].gcons->provider->name);
1228 off = count = n = 0;
1229 while (count < sc->map_size) {
1230 struct g_virstor_map_entry *mapbuf;
1231 size_t bs;
1232
1233 bs = MIN(MAXPHYS, sc->map_size - count);
1234 if (bs % sc->sectorsize != 0) {
1235 /* Check for alignment errors */
1236 bs = (bs / sc->sectorsize) * sc->sectorsize;
1237 if (bs == 0)
1238 break;
1239 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1240 "for %s on %s", sc->geom->name,
1241 sc->components[0].gcons->provider->name);
1242 }
1243 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1244 if (mapbuf == NULL) {
1245 free(sc->map, M_GVIRSTOR);
1246 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1247 "for %s from %s (offset %ju) (error %d)",
1248 sc->geom->name,
1249 sc->components[0].gcons->provider->name,
1250 off, error);
1251 return;
1252 }
1253
1254 bcopy(mapbuf, &sc->map[n], bs);
1255 off += bs;
1256 count += bs;
1257 n += bs / sizeof *(sc->map);
1258 g_free(mapbuf);
1259 }
1260 g_access(sc->components[0].gcons, -1, 0, 0);
1261 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1262
1263 /* find first component with allocatable chunks */
1264 index = -1;
1265 for (n = 0; n < sc->n_components; n++) {
1266 if (sc->components[n].chunk_next <
1267 sc->components[n].chunk_count) {
1268 index = n;
1269 break;
1270 }
1271 }
1272 if (index == -1)
1273 /* not found? set it to the last component and handle it
1274 * later */
1275 index = sc->n_components - 1;
1276
1277 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1278 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1279 "(%d/%u: %s)", sc->geom->name,
1280 index+1,
1281 sc->n_components,
1282 sc->components[index].gcons->provider->name);
1283 }
1284 sc->curr_component = index;
1285
1286 if (sc->components[index].chunk_next >=
1287 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1288 LOG_MSG(LVL_WARNING,
1289 "Component %s of %s is running out of free space "
1290 "(%u chunks left)",
1291 sc->components[index].gcons->provider->name,
1292 sc->geom->name, sc->components[index].chunk_count -
1293 sc->components[index].chunk_next);
1294 }
1295
1296 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1297 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1298 LOG_MSG(LVL_ERROR,
1299 "%s: Map entries don't fit exactly in a sector (%s)",
1300 __func__, sc->geom->name);
1301 return;
1302 }
1303
1304 /* Recalculate allocated chunks in components & at the same time
1305 * verify map data is sane. We could trust metadata on this, but
1306 * we want to make sure. */
1307 for (n = 0; n < sc->n_components; n++)
1308 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1309
1310 for (n = 0; n < sc->chunk_count; n++) {
1311 if (sc->map[n].provider_no >= sc->n_components ||
1312 sc->map[n].provider_chunk >=
1313 sc->components[sc->map[n].provider_no].chunk_count) {
1314 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1315 __func__, (u_int)n, sc->geom->name);
1316 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1317 " provider_chunk: %u, chunk_count: %u", __func__,
1318 sc->map[n].provider_no, sc->n_components,
1319 sc->map[n].provider_chunk,
1320 sc->components[sc->map[n].provider_no].chunk_count);
1321 return;
1322 }
1323 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1324 sc->components[sc->map[n].provider_no].chunk_next++;
1325 }
1326
1327 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1328 sc->geom->name);
1329
1330 sc->provider->sectorsize = sc->sectorsize;
1331 sc->provider->mediasize = sc->virsize;
1332 g_error_provider(sc->provider, 0);
1333
1334 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1335 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1336 "chunk %u", sc->provider->name, sc->curr_component,
1337 sc->components[sc->curr_component].chunk_next);
1338 }
1339
1340 /*
1341 * Returns count of active providers in this geom instance
1342 */
1343 static u_int
1344 virstor_valid_components(struct g_virstor_softc *sc)
1345 {
1346 unsigned int nc, i;
1347
1348 nc = 0;
1349 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1350 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1351 for (i = 0; i < sc->n_components; i++)
1352 if (sc->components[i].gcons != NULL)
1353 nc++;
1354 return (nc);
1355 }
1356
1357 /*
1358 * Called when the consumer gets orphaned (?)
1359 */
1360 static void
1361 g_virstor_orphan(struct g_consumer *cp)
1362 {
1363 struct g_virstor_softc *sc;
1364 struct g_virstor_component *comp;
1365 struct g_geom *gp;
1366
1367 g_topology_assert();
1368 gp = cp->geom;
1369 sc = gp->softc;
1370 if (sc == NULL)
1371 return;
1372
1373 comp = cp->private;
1374 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1375 __func__));
1376 remove_component(sc, comp, FALSE);
1377 if (virstor_valid_components(sc) == 0)
1378 virstor_geom_destroy(sc, TRUE, FALSE);
1379 }
1380
1381 /*
1382 * Called to notify geom when it's been opened, and for what intent
1383 */
1384 static int
1385 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1386 {
1387 struct g_consumer *c;
1388 struct g_virstor_softc *sc;
1389 struct g_geom *gp;
1390 int error;
1391
1392 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1393 gp = pp->geom;
1394 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1395 sc = gp->softc;
1396
1397 if (sc == NULL) {
1398 /* It seems that .access can be called with negative dr,dw,dx
1399 * in this case but I want to check for myself */
1400 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s",
1401 dr, dw, de, pp->name);
1402 /* This should only happen when geom is withered so
1403 * allow only negative requests */
1404 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
1405 ("%s: Positive access for %s", __func__, pp->name));
1406 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
1407 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed",
1408 pp->name);
1409 return (0);
1410 }
1411
1412 /* Grab an exclusive bit to propagate on our consumers on first open */
1413 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1414 de++;
1415 /* ... drop it on close */
1416 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1417 de--;
1418 update_metadata(sc); /* Writes statistical information */
1419 }
1420
1421 error = ENXIO;
1422 LIST_FOREACH(c, &gp->consumer, consumer) {
1423 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
1424 error = g_access(c, dr, dw, de);
1425 if (error != 0) {
1426 struct g_consumer *c2;
1427
1428 /* Backout earlier changes */
1429 LIST_FOREACH(c2, &gp->consumer, consumer) {
1430 if (c2 == c) /* all eariler components fixed */
1431 return (error);
1432 g_access(c2, -dr, -dw, -de);
1433 }
1434 }
1435 }
1436
1437 return (error);
1438 }
1439
1440 /*
1441 * Generate XML dump of current state
1442 */
1443 static void
1444 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1445 struct g_consumer *cp, struct g_provider *pp)
1446 {
1447 struct g_virstor_softc *sc;
1448
1449 g_topology_assert();
1450 sc = gp->softc;
1451
1452 if (sc == NULL || pp != NULL)
1453 return;
1454
1455 if (cp != NULL) {
1456 /* For each component */
1457 struct g_virstor_component *comp;
1458
1459 comp = cp->private;
1460 if (comp == NULL)
1461 return;
1462 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1463 indent, comp->index);
1464 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1465 indent, comp->chunk_count);
1466 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1467 indent, comp->chunk_next);
1468 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1469 indent, comp->chunk_reserved);
1470 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1471 indent,
1472 comp->chunk_next > 0 ? 100 -
1473 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1474 comp->chunk_count : 100);
1475 } else {
1476 /* For the whole thing */
1477 u_int count, used, i;
1478 off_t size;
1479
1480 count = used = size = 0;
1481 for (i = 0; i < sc->n_components; i++) {
1482 if (sc->components[i].gcons != NULL) {
1483 count += sc->components[i].chunk_count;
1484 used += sc->components[i].chunk_next +
1485 sc->components[i].chunk_reserved;
1486 size += sc->components[i].gcons->
1487 provider->mediasize;
1488 }
1489 }
1490
1491 sbuf_printf(sb, "%s<Status>"
1492 "Components=%u, Online=%u</Status>\n", indent,
1493 sc->n_components, virstor_valid_components(sc));
1494 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1495 indent, 100-(used * 100) / count);
1496 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1497 sc->chunk_size);
1498 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1499 indent, used > 0 ? 100 - (used * 100) / count : 100);
1500 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1501 indent, count);
1502 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1503 indent, sc->chunk_count);
1504 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1505 indent,
1506 (count * 100) / sc->chunk_count);
1507 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1508 indent, size);
1509 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1510 sc->virsize);
1511 }
1512 }
1513
1514 /*
1515 * GEOM .done handler
1516 * Can't use standard handler because one requested IO may
1517 * fork into additional data IOs
1518 */
1519 static void
1520 g_virstor_done(struct bio *b)
1521 {
1522 struct g_virstor_softc *sc;
1523 struct bio *parent_b;
1524
1525 parent_b = b->bio_parent;
1526 sc = parent_b->bio_to->geom->softc;
1527
1528 if (b->bio_error != 0) {
1529 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1530 b->bio_error, b->bio_offset, b->bio_length,
1531 b->bio_to->name);
1532 if (parent_b->bio_error == 0)
1533 parent_b->bio_error = b->bio_error;
1534 }
1535
1536 parent_b->bio_inbed++;
1537 parent_b->bio_completed += b->bio_completed;
1538
1539 if (parent_b->bio_children == parent_b->bio_inbed) {
1540 parent_b->bio_completed = parent_b->bio_length;
1541 g_io_deliver(parent_b, parent_b->bio_error);
1542 }
1543 g_destroy_bio(b);
1544 }
1545
1546 /*
1547 * I/O starts here
1548 * Called in g_down thread
1549 */
1550 static void
1551 g_virstor_start(struct bio *b)
1552 {
1553 struct g_virstor_softc *sc;
1554 struct g_virstor_component *comp;
1555 struct bio *cb;
1556 struct g_provider *pp;
1557 char *addr;
1558 off_t offset, length;
1559 struct bio_queue_head bq;
1560 size_t chunk_size; /* cached for convenience */
1561 u_int count;
1562
1563 pp = b->bio_to;
1564 sc = pp->geom->softc;
1565 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1566 b->bio_to->error, b->bio_to->name));
1567
1568 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1569
1570 switch (b->bio_cmd) {
1571 case BIO_READ:
1572 case BIO_WRITE:
1573 case BIO_DELETE:
1574 break;
1575 default:
1576 g_io_deliver(b, EOPNOTSUPP);
1577 return;
1578 }
1579
1580 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1581 bioq_init(&bq);
1582
1583 chunk_size = sc->chunk_size;
1584 addr = b->bio_data;
1585 offset = b->bio_offset; /* virtual offset and length */
1586 length = b->bio_length;
1587
1588 while (length > 0) {
1589 size_t chunk_index, in_chunk_offset, in_chunk_length;
1590 struct virstor_map_entry *me;
1591
1592 chunk_index = offset / chunk_size; /* round downwards */
1593 in_chunk_offset = offset % chunk_size;
1594 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1595 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1596 b->bio_cmd == BIO_READ ? "R" : "W",
1597 offset, length,
1598 chunk_index, in_chunk_offset, in_chunk_length);
1599 me = &sc->map[chunk_index];
1600
1601 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1602 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1603 /* Reads from unallocated chunks return zeroed
1604 * buffers */
1605 if (b->bio_cmd == BIO_READ)
1606 bzero(addr, in_chunk_length);
1607 } else {
1608 comp = &sc->components[me->provider_no];
1609
1610 cb = g_clone_bio(b);
1611 if (cb == NULL) {
1612 bioq_dismantle(&bq);
1613 if (b->bio_error == 0)
1614 b->bio_error = ENOMEM;
1615 g_io_deliver(b, b->bio_error);
1616 return;
1617 }
1618 cb->bio_to = comp->gcons->provider;
1619 cb->bio_done = g_virstor_done;
1620 cb->bio_offset =
1621 (off_t)me->provider_chunk * (off_t)chunk_size
1622 + in_chunk_offset;
1623 cb->bio_length = in_chunk_length;
1624 cb->bio_data = addr;
1625 cb->bio_caller1 = comp;
1626 bioq_disksort(&bq, cb);
1627 }
1628 } else { /* handle BIO_WRITE */
1629 KASSERT(b->bio_cmd == BIO_WRITE,
1630 ("%s: Unknown command %d", __func__,
1631 b->bio_cmd));
1632
1633 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1634 /* We have a virtual chunk, represented by
1635 * the "me" entry, but it's not yet allocated
1636 * (tied to) a physical chunk. So do it now. */
1637 struct virstor_map_entry *data_me;
1638 u_int phys_chunk, comp_no;
1639 off_t s_offset;
1640 int error;
1641
1642 error = allocate_chunk(sc, &comp, &comp_no,
1643 &phys_chunk);
1644 if (error != 0) {
1645 /* We cannot allocate a physical chunk
1646 * to satisfy this request, so we'll
1647 * delay it to when we can...
1648 * XXX: this will prevent the fs from
1649 * being umounted! */
1650 struct g_virstor_bio_q *biq;
1651 biq = malloc(sizeof *biq, M_GVIRSTOR,
1652 M_NOWAIT);
1653 if (biq == NULL) {
1654 bioq_dismantle(&bq);
1655 if (b->bio_error == 0)
1656 b->bio_error = ENOMEM;
1657 g_io_deliver(b, b->bio_error);
1658 return;
1659 }
1660 biq->bio = b;
1661 mtx_lock(&sc->delayed_bio_q_mtx);
1662 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1663 biq, linkage);
1664 mtx_unlock(&sc->delayed_bio_q_mtx);
1665 LOG_MSG(LVL_WARNING, "Delaying BIO "
1666 "(size=%ju) until free physical "
1667 "space can be found on %s",
1668 b->bio_length,
1669 sc->provider->name);
1670 return;
1671 }
1672 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1673 "for %s",
1674 phys_chunk,
1675 comp->gcons->provider->name,
1676 sc->provider->name);
1677
1678 me->provider_no = comp_no;
1679 me->provider_chunk = phys_chunk;
1680 me->flags |= VIRSTOR_MAP_ALLOCATED;
1681
1682 cb = g_clone_bio(b);
1683 if (cb == NULL) {
1684 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1685 me->provider_no = 0;
1686 me->provider_chunk = 0;
1687 bioq_dismantle(&bq);
1688 if (b->bio_error == 0)
1689 b->bio_error = ENOMEM;
1690 g_io_deliver(b, b->bio_error);
1691 return;
1692 }
1693
1694 /* The allocation table is stored continuously
1695 * at the start of the drive. We need to
1696 * calculate the offset of the sector that holds
1697 * this map entry both on the drive and in the
1698 * map array.
1699 * sc_offset will end up pointing to the drive
1700 * sector. */
1701 s_offset = chunk_index * sizeof *me;
1702 s_offset = (s_offset / sc->sectorsize) *
1703 sc->sectorsize;
1704
1705 /* data_me points to map entry sector
1706 * in memory (analoguos to offset) */
1707 data_me = &sc->map[(chunk_index /
1708 sc->me_per_sector) * sc->me_per_sector];
1709
1710 /* Commit sector with map entry to storage */
1711 cb->bio_to = sc->components[0].gcons->provider;
1712 cb->bio_done = g_virstor_done;
1713 cb->bio_offset = s_offset;
1714 cb->bio_data = (char *)data_me;
1715 cb->bio_length = sc->sectorsize;
1716 cb->bio_caller1 = &sc->components[0];
1717 bioq_disksort(&bq, cb);
1718 }
1719
1720 comp = &sc->components[me->provider_no];
1721 cb = g_clone_bio(b);
1722 if (cb == NULL) {
1723 bioq_dismantle(&bq);
1724 if (b->bio_error == 0)
1725 b->bio_error = ENOMEM;
1726 g_io_deliver(b, b->bio_error);
1727 return;
1728 }
1729 /* Finally, handle the data */
1730 cb->bio_to = comp->gcons->provider;
1731 cb->bio_done = g_virstor_done;
1732 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1733 in_chunk_offset;
1734 cb->bio_length = in_chunk_length;
1735 cb->bio_data = addr;
1736 cb->bio_caller1 = comp;
1737 bioq_disksort(&bq, cb);
1738 }
1739 addr += in_chunk_length;
1740 length -= in_chunk_length;
1741 offset += in_chunk_length;
1742 }
1743
1744 /* Fire off bio's here */
1745 count = 0;
1746 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1747 bioq_remove(&bq, cb);
1748 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1749 comp = cb->bio_caller1;
1750 cb->bio_caller1 = NULL;
1751 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1752 cb->bio_offset, cb->bio_length);
1753 g_io_request(cb, comp->gcons);
1754 count++;
1755 }
1756 if (count == 0) { /* We handled everything locally */
1757 b->bio_completed = b->bio_length;
1758 g_io_deliver(b, 0);
1759 }
1760
1761 }
1762
1763 /*
1764 * Allocate a chunk from a physical provider. Returns physical component,
1765 * chunk index relative to the component and the component's index.
1766 */
1767 static int
1768 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1769 u_int *comp_no_p, u_int *chunk)
1770 {
1771 u_int comp_no;
1772
1773 KASSERT(sc->curr_component < sc->n_components,
1774 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1775
1776 comp_no = sc->curr_component;
1777 *comp = &sc->components[comp_no];
1778 dump_component(*comp);
1779 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1780 /* This component is full. Allocate next component */
1781 if (comp_no >= sc->n_components-1) {
1782 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1783 sc->geom->name);
1784 return (-1);
1785 }
1786 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1787 sc->curr_component = ++comp_no;
1788
1789 *comp = &sc->components[comp_no];
1790 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1791 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1792 "(switching to %u/%u: %s)", sc->geom->name,
1793 comp_no+1, sc->n_components,
1794 (*comp)->gcons->provider->name);
1795 /* Take care not to overwrite reserved chunks */
1796 if ( (*comp)->chunk_reserved > 0 &&
1797 (*comp)->chunk_next < (*comp)->chunk_reserved)
1798 (*comp)->chunk_next = (*comp)->chunk_reserved;
1799
1800 (*comp)->flags |=
1801 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1802 dump_component(*comp);
1803 *comp_no_p = comp_no;
1804 *chunk = (*comp)->chunk_next++;
1805 } else {
1806 *comp_no_p = comp_no;
1807 *chunk = (*comp)->chunk_next++;
1808 }
1809 return (0);
1810 }
1811
1812 /* Dump a component */
1813 static void
1814 dump_component(struct g_virstor_component *comp)
1815 {
1816
1817 if (g_virstor_debug < LVL_DEBUG2)
1818 return;
1819 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1820 printf(" chunk_count: %u\n", comp->chunk_count);
1821 printf(" chunk_next: %u\n", comp->chunk_next);
1822 printf(" flags: %u\n", comp->flags);
1823 }
1824
1825 #if 0
1826 /* Dump a map entry */
1827 static void
1828 dump_me(struct virstor_map_entry *me, unsigned int nr)
1829 {
1830 if (g_virstor_debug < LVL_DEBUG)
1831 return;
1832 printf("VIRT. CHUNK #%d: ", nr);
1833 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1834 printf("(unallocated)\n");
1835 else
1836 printf("allocated at provider %u, provider_chunk %u\n",
1837 me->provider_no, me->provider_chunk);
1838 }
1839 #endif
1840
1841 /*
1842 * Dismantle bio_queue and destroy its components
1843 */
1844 static void
1845 bioq_dismantle(struct bio_queue_head *bq)
1846 {
1847 struct bio *b;
1848
1849 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1850 bioq_remove(bq, b);
1851 g_destroy_bio(b);
1852 }
1853 }
1854
1855 /*
1856 * The function that shouldn't be called.
1857 * When this is called, the stack is already garbled because of
1858 * argument mismatch. There's nothing to do now but panic, which is
1859 * accidentally the whole purpose of this function.
1860 * Motivation: to guard from accidentally calling geom methods when
1861 * they shouldn't be called. (see g_..._taste)
1862 */
1863 static void
1864 invalid_call(void)
1865 {
1866 panic("invalid_call() has just been called. Something's fishy here.");
1867 }
1868
1869 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */
Cache object: 4a2952203dba51a7845fdab5fb9064ff
|