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/8.4/sys/geom/virstor/g_virstor.c 239442 2012-08-20 16:13:16Z jimharris $");
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/lock.h>
40 #include <sys/mutex.h>
41 #include <sys/sx.h>
42 #include <sys/bio.h>
43 #include <sys/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 if (sc == NULL) {
235 gctl_error(req, "Don't know anything about '%s'", name);
236 g_topology_unlock();
237 return;
238 }
239
240 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
241 sc->geom->name);
242 update_metadata(sc);
243 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
244 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
245 sc->geom->name, error);
246 }
247 }
248 g_topology_unlock();
249 }
250
251 /*
252 * "add" verb from userland - add new component(s) to the structure.
253 * This will be done all at once in here, without going through the
254 * .taste function for new components.
255 */
256 static void
257 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
258 {
259 /* Note: while this is going on, I/O is being done on
260 * the g_up and g_down threads. The idea is to make changes
261 * to softc members in a way that can atomically activate
262 * them all at once. */
263 struct g_virstor_softc *sc;
264 int *hardcode, *nargs;
265 const char *geom_name; /* geom to add a component to */
266 struct g_consumer *fcp;
267 struct g_virstor_bio_q *bq;
268 u_int added;
269 int error;
270 int i;
271
272 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
273 if (nargs == NULL) {
274 gctl_error(req, "Error fetching argument '%s'", "nargs");
275 return;
276 }
277 if (*nargs < 2) {
278 gctl_error(req, "Invalid number of arguments");
279 return;
280 }
281 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
282 if (hardcode == NULL) {
283 gctl_error(req, "Error fetching argument '%s'", "hardcode");
284 return;
285 }
286
287 /* Find "our" geom */
288 geom_name = gctl_get_asciiparam(req, "arg0");
289 if (geom_name == NULL) {
290 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
291 return;
292 }
293 sc = virstor_find_geom(cp, geom_name);
294 if (sc == NULL) {
295 gctl_error(req, "Don't know anything about '%s'", geom_name);
296 return;
297 }
298
299 if (virstor_valid_components(sc) != sc->n_components) {
300 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
301 "virstor %s", sc->geom->name);
302 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
303 return;
304 }
305
306 fcp = sc->components[0].gcons;
307 added = 0;
308 g_topology_lock();
309 for (i = 1; i < *nargs; i++) {
310 struct g_virstor_metadata md;
311 char aname[8];
312 const char *prov_name;
313 struct g_provider *pp;
314 struct g_consumer *cp;
315 u_int nc;
316 u_int j;
317
318 snprintf(aname, sizeof aname, "arg%d", i);
319 prov_name = gctl_get_asciiparam(req, aname);
320 if (prov_name == NULL) {
321 gctl_error(req, "Error fetching argument '%s'", aname);
322 g_topology_unlock();
323 return;
324 }
325 if (strncmp(prov_name, _PATH_DEV, strlen(_PATH_DEV)) == 0)
326 prov_name += strlen(_PATH_DEV);
327
328 pp = g_provider_by_name(prov_name);
329 if (pp == NULL) {
330 /* This is the most common error so be verbose about it */
331 if (added != 0) {
332 gctl_error(req, "Invalid provider: '%s' (added"
333 " %u components)", prov_name, added);
334 update_metadata(sc);
335 } else {
336 gctl_error(req, "Invalid provider: '%s'",
337 prov_name);
338 }
339 g_topology_unlock();
340 return;
341 }
342 cp = g_new_consumer(sc->geom);
343 if (cp == NULL) {
344 gctl_error(req, "Cannot create consumer");
345 g_topology_unlock();
346 return;
347 }
348 error = g_attach(cp, pp);
349 if (error != 0) {
350 gctl_error(req, "Cannot attach a consumer to %s",
351 pp->name);
352 g_destroy_consumer(cp);
353 g_topology_unlock();
354 return;
355 }
356 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
357 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
358 if (error != 0) {
359 gctl_error(req, "Access request failed for %s",
360 pp->name);
361 g_destroy_consumer(cp);
362 g_topology_unlock();
363 return;
364 }
365 }
366 if (fcp->provider->sectorsize != pp->sectorsize) {
367 gctl_error(req, "Sector size doesn't fit for %s",
368 pp->name);
369 g_destroy_consumer(cp);
370 g_topology_unlock();
371 return;
372 }
373 for (j = 0; j < sc->n_components; j++) {
374 if (strcmp(sc->components[j].gcons->provider->name,
375 pp->name) == 0) {
376 gctl_error(req, "Component %s already in %s",
377 pp->name, sc->geom->name);
378 g_destroy_consumer(cp);
379 g_topology_unlock();
380 return;
381 }
382 }
383 sc->components = realloc(sc->components,
384 sizeof(*sc->components) * (sc->n_components + 1),
385 M_GVIRSTOR, M_WAITOK);
386
387 nc = sc->n_components;
388 sc->components[nc].gcons = cp;
389 sc->components[nc].sc = sc;
390 sc->components[nc].index = nc;
391 sc->components[nc].chunk_count = cp->provider->mediasize /
392 sc->chunk_size;
393 sc->components[nc].chunk_next = 0;
394 sc->components[nc].chunk_reserved = 0;
395
396 if (sc->components[nc].chunk_count < 4) {
397 gctl_error(req, "Provider too small: %s",
398 cp->provider->name);
399 g_destroy_consumer(cp);
400 g_topology_unlock();
401 return;
402 }
403 fill_metadata(sc, &md, nc, *hardcode);
404 write_metadata(cp, &md);
405 /* The new component becomes visible when n_components is
406 * incremented */
407 sc->n_components++;
408 added++;
409
410 }
411 /* This call to update_metadata() is critical. In case there's a
412 * power failure in the middle of it and some components are updated
413 * while others are not, there will be trouble on next .taste() iff
414 * a non-updated component is detected first */
415 update_metadata(sc);
416 g_topology_unlock();
417 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
418 sc->geom->name);
419 /* Fire off BIOs previously queued because there wasn't any
420 * physical space left. If the BIOs still can't be satisfied
421 * they will again be added to the end of the queue (during
422 * which the mutex will be recursed) */
423 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
424 bq->bio = NULL;
425 mtx_lock(&sc->delayed_bio_q_mtx);
426 /* First, insert a sentinel to the queue end, so we don't
427 * end up in an infinite loop if there's still no free
428 * space available. */
429 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
430 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
431 bq = STAILQ_FIRST(&sc->delayed_bio_q);
432 if (bq->bio != NULL) {
433 g_virstor_start(bq->bio);
434 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
435 free(bq, M_GVIRSTOR);
436 } else {
437 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
438 free(bq, M_GVIRSTOR);
439 break;
440 }
441 }
442 mtx_unlock(&sc->delayed_bio_q_mtx);
443
444 }
445
446 /*
447 * Find a geom handled by the class
448 */
449 static struct g_virstor_softc *
450 virstor_find_geom(const struct g_class *cp, const char *name)
451 {
452 struct g_geom *gp;
453
454 LIST_FOREACH(gp, &cp->geom, geom) {
455 if (strcmp(name, gp->name) == 0)
456 return (gp->softc);
457 }
458 return (NULL);
459 }
460
461 /*
462 * Update metadata on all components to reflect the current state
463 * of these fields:
464 * - chunk_next
465 * - flags
466 * - md_count
467 * Expects things to be set up so write_metadata() can work, i.e.
468 * the topology lock must be held.
469 */
470 static void
471 update_metadata(struct g_virstor_softc *sc)
472 {
473 struct g_virstor_metadata md;
474 int n;
475
476 if (virstor_valid_components(sc) != sc->n_components)
477 return; /* Incomplete device */
478 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
479 sc->geom->name);
480 /* Update metadata on components */
481 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
482 sc->geom->class->name, sc->geom->name);
483 g_topology_assert();
484 for (n = 0; n < sc->n_components; n++) {
485 read_metadata(sc->components[n].gcons, &md);
486 md.chunk_next = sc->components[n].chunk_next;
487 md.flags = sc->components[n].flags;
488 md.md_count = sc->n_components;
489 write_metadata(sc->components[n].gcons, &md);
490 }
491 }
492
493 /*
494 * Fills metadata (struct md) from information stored in softc and the nc'th
495 * component of virstor
496 */
497 static void
498 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
499 u_int nc, u_int hardcode)
500 {
501 struct g_virstor_component *c;
502
503 bzero(md, sizeof *md);
504 c = &sc->components[nc];
505
506 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
507 md->md_version = G_VIRSTOR_VERSION;
508 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
509 md->md_id = sc->id;
510 md->md_virsize = sc->virsize;
511 md->md_chunk_size = sc->chunk_size;
512 md->md_count = sc->n_components;
513
514 if (hardcode) {
515 strncpy(md->provider, c->gcons->provider->name,
516 sizeof md->provider);
517 }
518 md->no = nc;
519 md->provsize = c->gcons->provider->mediasize;
520 md->chunk_count = c->chunk_count;
521 md->chunk_next = c->chunk_next;
522 md->chunk_reserved = c->chunk_reserved;
523 md->flags = c->flags;
524 }
525
526 /*
527 * Remove a component from virstor device.
528 * Can only be done if the component is unallocated.
529 */
530 static void
531 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
532 {
533 /* As this is executed in parallel to I/O, operations on virstor
534 * structures must be as atomic as possible. */
535 struct g_virstor_softc *sc;
536 int *nargs;
537 const char *geom_name;
538 u_int removed;
539 int i;
540
541 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
542 if (nargs == NULL) {
543 gctl_error(req, "Error fetching argument '%s'", "nargs");
544 return;
545 }
546 if (*nargs < 2) {
547 gctl_error(req, "Invalid number of arguments");
548 return;
549 }
550 /* Find "our" geom */
551 geom_name = gctl_get_asciiparam(req, "arg0");
552 if (geom_name == NULL) {
553 gctl_error(req, "Error fetching argument '%s'",
554 "geom_name (arg0)");
555 return;
556 }
557 sc = virstor_find_geom(cp, geom_name);
558 if (sc == NULL) {
559 gctl_error(req, "Don't know anything about '%s'", geom_name);
560 return;
561 }
562
563 if (virstor_valid_components(sc) != sc->n_components) {
564 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
565 "virstor %s", sc->geom->name);
566 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
567 return;
568 }
569
570 removed = 0;
571 for (i = 1; i < *nargs; i++) {
572 char param[8];
573 const char *prov_name;
574 int j, found;
575 struct g_virstor_component *newcomp, *compbak;
576
577 sprintf(param, "arg%d", i);
578 prov_name = gctl_get_asciiparam(req, param);
579 if (prov_name == NULL) {
580 gctl_error(req, "Error fetching argument '%s'", param);
581 return;
582 }
583 if (strncmp(prov_name, _PATH_DEV, strlen(_PATH_DEV)) == 0)
584 prov_name += strlen(_PATH_DEV);
585
586 found = -1;
587 for (j = 0; j < sc->n_components; j++) {
588 if (strcmp(sc->components[j].gcons->provider->name,
589 prov_name) == 0) {
590 found = j;
591 break;
592 }
593 }
594 if (found == -1) {
595 LOG_MSG(LVL_ERROR, "No %s component in %s",
596 prov_name, sc->geom->name);
597 continue;
598 }
599
600 compbak = sc->components;
601 newcomp = malloc(sc->n_components * sizeof(*sc->components),
602 M_GVIRSTOR, M_WAITOK | M_ZERO);
603 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
604 bcopy(&sc->components[found + 1], newcomp + found,
605 found * sizeof(*sc->components));
606 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
607 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
608 "removed from %s",
609 prov_name, sc->geom->name);
610 free(newcomp, M_GVIRSTOR);
611 /* We'll consider this non-fatal error */
612 continue;
613 }
614 /* Renumerate unallocated components */
615 for (j = 0; j < sc->n_components-1; j++) {
616 if ((sc->components[j].flags &
617 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
618 sc->components[j].index = j;
619 }
620 }
621 /* This is the critical section. If a component allocation
622 * event happens while both variables are not yet set,
623 * there will be trouble. Something will panic on encountering
624 * NULL sc->components[x].gcomp member.
625 * Luckily, component allocation happens very rarely and
626 * removing components is an abnormal action in any case. */
627 sc->components = newcomp;
628 sc->n_components--;
629 /* End critical section */
630
631 g_topology_lock();
632 if (clear_metadata(&compbak[found]) != 0) {
633 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
634 "metadata on %s", prov_name);
635 }
636 g_detach(compbak[found].gcons);
637 g_destroy_consumer(compbak[found].gcons);
638 g_topology_unlock();
639
640 free(compbak, M_GVIRSTOR);
641
642 removed++;
643 }
644
645 /* This call to update_metadata() is critical. In case there's a
646 * power failure in the middle of it and some components are updated
647 * while others are not, there will be trouble on next .taste() iff
648 * a non-updated component is detected first */
649 g_topology_lock();
650 update_metadata(sc);
651 g_topology_unlock();
652 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
653 sc->geom->name);
654 }
655
656 /*
657 * Clear metadata sector on component
658 */
659 static int
660 clear_metadata(struct g_virstor_component *comp)
661 {
662 char *buf;
663 int error;
664
665 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
666 comp->gcons->provider->name);
667 g_topology_assert();
668 error = g_access(comp->gcons, 0, 1, 0);
669 if (error != 0)
670 return (error);
671 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
672 M_WAITOK | M_ZERO);
673 error = g_write_data(comp->gcons,
674 comp->gcons->provider->mediasize -
675 comp->gcons->provider->sectorsize,
676 buf,
677 comp->gcons->provider->sectorsize);
678 free(buf, M_GVIRSTOR);
679 g_access(comp->gcons, 0, -1, 0);
680 return (error);
681 }
682
683 /*
684 * Destroy geom forcibly.
685 */
686 static int
687 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
688 struct g_geom *gp)
689 {
690 struct g_virstor_softc *sc;
691 int exitval;
692
693 sc = gp->softc;
694 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
695
696 exitval = 0;
697 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
698 gp->softc);
699
700 if (sc != NULL) {
701 #ifdef INVARIANTS
702 char *buf;
703 int error;
704 off_t off;
705 int isclean, count;
706 int n;
707
708 LOG_MSG(LVL_INFO, "INVARIANTS detected");
709 LOG_MSG(LVL_INFO, "Verifying allocation "
710 "table for %s", sc->geom->name);
711 count = 0;
712 for (n = 0; n < sc->chunk_count; n++) {
713 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
714 count++;
715 }
716 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
717 sc->geom->name, count);
718 n = off = count = 0;
719 isclean = 1;
720 if (virstor_valid_components(sc) != sc->n_components) {
721 /* This is a incomplete virstor device (not all
722 * components have been found) */
723 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
724 sc->geom->name);
725 goto bailout;
726 }
727 error = g_access(sc->components[0].gcons, 1, 0, 0);
728 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
729 error));
730 /* Compare the whole on-disk allocation table with what's
731 * currently in memory */
732 while (n < sc->chunk_count) {
733 buf = g_read_data(sc->components[0].gcons, off,
734 sc->sectorsize, &error);
735 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
736 "for read at %jd", error, off));
737 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
738 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
739 "entry %d, offset %jd", n, off);
740 isclean = 0;
741 count++;
742 }
743 n += sc->me_per_sector;
744 off += sc->sectorsize;
745 g_free(buf);
746 }
747 error = g_access(sc->components[0].gcons, -1, 0, 0);
748 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
749 __func__, error));
750 if (isclean != 1) {
751 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
752 "(%d sectors don't match, max %zu allocations)",
753 sc->geom->name, count,
754 count * sc->me_per_sector);
755 } else {
756 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
757 sc->geom->name);
758 }
759 bailout:
760 #endif
761 update_metadata(sc);
762 virstor_geom_destroy(sc, FALSE, FALSE);
763 exitval = EAGAIN;
764 } else
765 exitval = 0;
766 return (exitval);
767 }
768
769 /*
770 * Taste event (per-class callback)
771 * Examines a provider and creates geom instances if needed
772 */
773 static struct g_geom *
774 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
775 {
776 struct g_virstor_metadata md;
777 struct g_geom *gp;
778 struct g_consumer *cp;
779 struct g_virstor_softc *sc;
780 int error;
781
782 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
783 g_topology_assert();
784 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
785
786 /* We need a dummy geom to attach a consumer to the given provider */
787 gp = g_new_geomf(mp, "virstor:taste.helper");
788 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
789 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
790 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
791
792 cp = g_new_consumer(gp);
793 g_attach(cp, pp);
794 error = read_metadata(cp, &md);
795 g_detach(cp);
796 g_destroy_consumer(cp);
797 g_destroy_geom(gp);
798
799 if (error != 0)
800 return (NULL);
801
802 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
803 return (NULL);
804 if (md.md_version != G_VIRSTOR_VERSION) {
805 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
806 "to handle %s (%s) : %d should be %d",
807 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
808 return (NULL);
809 }
810 if (md.provsize != pp->mediasize)
811 return (NULL);
812
813 /* If the provider name is hardcoded, use the offered provider only
814 * if it's been offered with its proper name (the one used in
815 * the label command). */
816 if (md.provider[0] != '\0' &&
817 !g_compare_names(md.provider, pp->name))
818 return (NULL);
819
820 /* Iterate all geoms this class already knows about to see if a new
821 * geom instance of this class needs to be created (in case the provider
822 * is first from a (possibly) multi-consumer geom) or it just needs
823 * to be added to an existing instance. */
824 sc = NULL;
825 gp = NULL;
826 LIST_FOREACH(gp, &mp->geom, geom) {
827 sc = gp->softc;
828 if (sc == NULL)
829 continue;
830 if (strcmp(md.md_name, sc->geom->name) != 0)
831 continue;
832 if (md.md_id != sc->id)
833 continue;
834 break;
835 }
836 if (gp != NULL) { /* We found an existing geom instance; add to it */
837 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
838 error = add_provider_to_geom(sc, pp, &md);
839 if (error != 0) {
840 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
841 pp->name, md.md_name, error);
842 return (NULL);
843 }
844 } else { /* New geom instance needs to be created */
845 gp = create_virstor_geom(mp, &md);
846 if (gp == NULL) {
847 LOG_MSG(LVL_ERROR, "Error creating new instance of "
848 "class %s: %s", mp->name, md.md_name);
849 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
850 md.md_name, pp->name);
851 return (NULL);
852 }
853 sc = gp->softc;
854 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
855 md.md_name);
856 error = add_provider_to_geom(sc, pp, &md);
857 if (error != 0) {
858 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
859 pp->name, md.md_name, error);
860 virstor_geom_destroy(sc, TRUE, FALSE);
861 return (NULL);
862 }
863 }
864
865 return (gp);
866 }
867
868 /*
869 * Destroyes consumer passed to it in arguments. Used as a callback
870 * on g_event queue.
871 */
872 static void
873 delay_destroy_consumer(void *arg, int flags __unused)
874 {
875 struct g_consumer *c = arg;
876 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
877 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
878 c->provider->name);
879 g_detach(c);
880 g_destroy_consumer(c);
881 }
882
883 /*
884 * Remove a component (consumer) from geom instance; If it's the first
885 * component being removed, orphan the provider to announce geom's being
886 * dismantled
887 */
888 static void
889 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
890 boolean_t delay)
891 {
892 struct g_consumer *c;
893
894 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
895 sc->geom->name));
896 c = comp->gcons;
897
898 comp->gcons = NULL;
899 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
900 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
901 sc->geom->name);
902 if (sc->provider != NULL) {
903 /* Whither, GEOM? */
904 sc->provider->flags |= G_PF_WITHER;
905 g_orphan_provider(sc->provider, ENXIO);
906 sc->provider = NULL;
907 LOG_MSG(LVL_INFO, "Removing provider %s", sc->geom->name);
908 }
909
910 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
911 g_access(c, -c->acr, -c->acw, -c->ace);
912 if (delay) {
913 /* Destroy consumer after it's tasted */
914 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
915 } else {
916 g_detach(c);
917 g_destroy_consumer(c);
918 }
919 }
920
921 /*
922 * Destroy geom - called internally
923 * See g_virstor_destroy_geom for the other one
924 */
925 static int
926 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
927 boolean_t delay)
928 {
929 struct g_provider *pp;
930 struct g_geom *gp;
931 int n;
932
933 g_topology_assert();
934
935 if (sc == NULL)
936 return (ENXIO);
937
938 pp = sc->provider;
939 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
940 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
941 "Device %s is still open.", pp->name);
942 if (!force)
943 return (EBUSY);
944 }
945
946 for (n = 0; n < sc->n_components; n++) {
947 if (sc->components[n].gcons != NULL)
948 remove_component(sc, &sc->components[n], delay);
949 }
950
951 gp = sc->geom;
952 gp->softc = NULL;
953
954 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
955 gp->name));
956
957 /* XXX: This might or might not work, since we're called with
958 * the topology lock held. Also, it might panic the kernel if
959 * the error'd BIO is in softupdates code. */
960 mtx_lock(&sc->delayed_bio_q_mtx);
961 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
962 struct g_virstor_bio_q *bq;
963 bq = STAILQ_FIRST(&sc->delayed_bio_q);
964 bq->bio->bio_error = ENOSPC;
965 g_io_deliver(bq->bio, EIO);
966 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
967 free(bq, M_GVIRSTOR);
968 }
969 mtx_unlock(&sc->delayed_bio_q_mtx);
970 mtx_destroy(&sc->delayed_bio_q_mtx);
971
972 free(sc->map, M_GVIRSTOR);
973 free(sc->components, M_GVIRSTOR);
974 bzero(sc, sizeof *sc);
975 free(sc, M_GVIRSTOR);
976
977 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
978 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
979 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
980
981 g_wither_geom(gp, ENXIO);
982
983 return (0);
984 }
985
986 /*
987 * Utility function: read metadata & decode. Wants topology lock to be
988 * held.
989 */
990 static int
991 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
992 {
993 struct g_provider *pp;
994 char *buf;
995 int error;
996
997 g_topology_assert();
998 error = g_access(cp, 1, 0, 0);
999 if (error != 0)
1000 return (error);
1001 pp = cp->provider;
1002 g_topology_unlock();
1003 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
1004 &error);
1005 g_topology_lock();
1006 g_access(cp, -1, 0, 0);
1007 if (buf == NULL)
1008 return (error);
1009
1010 virstor_metadata_decode(buf, md);
1011 g_free(buf);
1012
1013 return (0);
1014 }
1015
1016 /**
1017 * Utility function: encode & write metadata. Assumes topology lock is
1018 * held.
1019 */
1020 static int
1021 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1022 {
1023 struct g_provider *pp;
1024 char *buf;
1025 int error;
1026
1027 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1028 ("Something's fishy in %s", __func__));
1029 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1030 g_topology_assert();
1031 error = g_access(cp, 0, 1, 0);
1032 if (error != 0)
1033 return (error);
1034 pp = cp->provider;
1035
1036 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1037 virstor_metadata_encode(md, buf);
1038 g_topology_unlock();
1039 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1040 pp->sectorsize);
1041 g_topology_lock();
1042 g_access(cp, 0, -1, 0);
1043
1044 free(buf, M_GVIRSTOR);
1045 return (0);
1046 }
1047
1048 /*
1049 * Creates a new instance of this GEOM class, initialise softc
1050 */
1051 static struct g_geom *
1052 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1053 {
1054 struct g_geom *gp;
1055 struct g_virstor_softc *sc;
1056
1057 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1058 md->md_name, md->md_id);
1059
1060 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1061 md->md_virsize < md->md_chunk_size) {
1062 /* This is bogus configuration, and probably means data is
1063 * somehow corrupted. Panic, maybe? */
1064 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1065 md->md_name);
1066 return (NULL);
1067 }
1068
1069 /* Check if it's already created */
1070 LIST_FOREACH(gp, &mp->geom, geom) {
1071 sc = gp->softc;
1072 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1073 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1074 md->md_name);
1075 if (sc->id != md->md_id) {
1076 LOG_MSG(LVL_ERROR,
1077 "Some stale or invalid components "
1078 "exist for virstor device named %s. "
1079 "You will need to <CLEAR> all stale "
1080 "components and maybe reconfigure "
1081 "the virstor device. Tune "
1082 "kern.geom.virstor.debug sysctl up "
1083 "for more information.",
1084 sc->geom->name);
1085 }
1086 return (NULL);
1087 }
1088 }
1089 gp = g_new_geomf(mp, "%s", md->md_name);
1090 gp->softc = NULL; /* to circumevent races that test softc */
1091
1092 gp->start = g_virstor_start;
1093 gp->spoiled = g_virstor_orphan;
1094 gp->orphan = g_virstor_orphan;
1095 gp->access = g_virstor_access;
1096 gp->dumpconf = g_virstor_dumpconf;
1097
1098 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1099 sc->id = md->md_id;
1100 sc->n_components = md->md_count;
1101 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1102 M_GVIRSTOR, M_WAITOK | M_ZERO);
1103 sc->chunk_size = md->md_chunk_size;
1104 sc->virsize = md->md_virsize;
1105 STAILQ_INIT(&sc->delayed_bio_q);
1106 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1107 "gvirstor", MTX_DEF | MTX_RECURSE);
1108
1109 sc->geom = gp;
1110 sc->provider = NULL; /* virstor_check_and_run will create it */
1111 gp->softc = sc;
1112
1113 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1114
1115 return (gp);
1116 }
1117
1118 /*
1119 * Add provider to a GEOM class instance
1120 */
1121 static int
1122 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1123 struct g_virstor_metadata *md)
1124 {
1125 struct g_virstor_component *component;
1126 struct g_consumer *cp, *fcp;
1127 struct g_geom *gp;
1128 int error;
1129
1130 if (md->no >= sc->n_components)
1131 return (EINVAL);
1132
1133 /* "Current" compontent */
1134 component = &(sc->components[md->no]);
1135 if (component->gcons != NULL)
1136 return (EEXIST);
1137
1138 gp = sc->geom;
1139 fcp = LIST_FIRST(&gp->consumer);
1140
1141 cp = g_new_consumer(gp);
1142 error = g_attach(cp, pp);
1143
1144 if (error != 0) {
1145 g_destroy_consumer(cp);
1146 return (error);
1147 }
1148
1149 if (fcp != NULL) {
1150 if (fcp->provider->sectorsize != pp->sectorsize) {
1151 /* TODO: this can be made to work */
1152 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1153 "sector size (%d)", pp->name, sc->geom->name,
1154 pp->sectorsize);
1155 return (EINVAL);
1156 }
1157 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1158 /* Replicate access permissions from first "live" consumer
1159 * to the new one */
1160 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1161 if (error != 0) {
1162 g_detach(cp);
1163 g_destroy_consumer(cp);
1164 return (error);
1165 }
1166 }
1167 }
1168
1169 /* Bring up a new component */
1170 cp->private = component;
1171 component->gcons = cp;
1172 component->sc = sc;
1173 component->index = md->no;
1174 component->chunk_count = md->chunk_count;
1175 component->chunk_next = md->chunk_next;
1176 component->chunk_reserved = md->chunk_reserved;
1177 component->flags = md->flags;
1178
1179 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1180
1181 virstor_check_and_run(sc);
1182 return (0);
1183 }
1184
1185 /*
1186 * Check if everything's ready to create the geom provider & device entry,
1187 * create and start provider.
1188 * Called ultimately by .taste, from g_event thread
1189 */
1190 static void
1191 virstor_check_and_run(struct g_virstor_softc *sc)
1192 {
1193 off_t off;
1194 size_t n, count;
1195 int index;
1196 int error;
1197
1198 if (virstor_valid_components(sc) != sc->n_components)
1199 return;
1200
1201 if (virstor_valid_components(sc) == 0) {
1202 /* This is actually a candidate for panic() */
1203 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1204 sc->provider->name);
1205 return;
1206 }
1207
1208 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1209
1210 /* Initialise allocation map from the first consumer */
1211 sc->chunk_count = sc->virsize / sc->chunk_size;
1212 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1213 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1214 sc->provider->name,
1215 sc->chunk_count * (off_t)sc->chunk_size);
1216 }
1217 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1218 /* The following allocation is in order of 4MB - 8MB */
1219 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1220 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1221 __func__, sc->map_size, sc->provider->name));
1222 sc->map_sectors = sc->map_size / sc->sectorsize;
1223
1224 count = 0;
1225 for (n = 0; n < sc->n_components; n++)
1226 count += sc->components[n].chunk_count;
1227 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1228 "(%zu KB chunks)",
1229 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1230
1231 error = g_access(sc->components[0].gcons, 1, 0, 0);
1232 if (error != 0) {
1233 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1234 "read allocation map for %s",
1235 sc->components[0].gcons->provider->name,
1236 sc->geom->name);
1237 return;
1238 }
1239 /* Read in the allocation map */
1240 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1241 sc->components[0].gcons->provider->name);
1242 off = count = n = 0;
1243 while (count < sc->map_size) {
1244 struct g_virstor_map_entry *mapbuf;
1245 size_t bs;
1246
1247 bs = MIN(MAXPHYS, sc->map_size - count);
1248 if (bs % sc->sectorsize != 0) {
1249 /* Check for alignment errors */
1250 bs = (bs / sc->sectorsize) * sc->sectorsize;
1251 if (bs == 0)
1252 break;
1253 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1254 "for %s on %s", sc->geom->name,
1255 sc->components[0].gcons->provider->name);
1256 }
1257 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1258 if (mapbuf == NULL) {
1259 free(sc->map, M_GVIRSTOR);
1260 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1261 "for %s from %s (offset %ju) (error %d)",
1262 sc->geom->name,
1263 sc->components[0].gcons->provider->name,
1264 off, error);
1265 return;
1266 }
1267
1268 bcopy(mapbuf, &sc->map[n], bs);
1269 off += bs;
1270 count += bs;
1271 n += bs / sizeof *(sc->map);
1272 g_free(mapbuf);
1273 }
1274 g_access(sc->components[0].gcons, -1, 0, 0);
1275 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1276
1277 /* find first component with allocatable chunks */
1278 index = -1;
1279 for (n = 0; n < sc->n_components; n++) {
1280 if (sc->components[n].chunk_next <
1281 sc->components[n].chunk_count) {
1282 index = n;
1283 break;
1284 }
1285 }
1286 if (index == -1)
1287 /* not found? set it to the last component and handle it
1288 * later */
1289 index = sc->n_components - 1;
1290
1291 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1292 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1293 "(%d/%u: %s)", sc->geom->name,
1294 index+1,
1295 sc->n_components,
1296 sc->components[index].gcons->provider->name);
1297 }
1298 sc->curr_component = index;
1299
1300 if (sc->components[index].chunk_next >=
1301 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1302 LOG_MSG(LVL_WARNING,
1303 "Component %s of %s is running out of free space "
1304 "(%u chunks left)",
1305 sc->components[index].gcons->provider->name,
1306 sc->geom->name, sc->components[index].chunk_count -
1307 sc->components[index].chunk_next);
1308 }
1309
1310 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1311 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1312 LOG_MSG(LVL_ERROR,
1313 "%s: Map entries don't fit exactly in a sector (%s)",
1314 __func__, sc->geom->name);
1315 return;
1316 }
1317
1318 /* Recalculate allocated chunks in components & at the same time
1319 * verify map data is sane. We could trust metadata on this, but
1320 * we want to make sure. */
1321 for (n = 0; n < sc->n_components; n++)
1322 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1323
1324 for (n = 0; n < sc->chunk_count; n++) {
1325 if (sc->map[n].provider_no >= sc->n_components ||
1326 sc->map[n].provider_chunk >=
1327 sc->components[sc->map[n].provider_no].chunk_count) {
1328 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1329 __func__, (u_int)n, sc->geom->name);
1330 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1331 " provider_chunk: %u, chunk_count: %u", __func__,
1332 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 return;
1336 }
1337 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1338 sc->components[sc->map[n].provider_no].chunk_next++;
1339 }
1340
1341 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1342 sc->geom->name);
1343
1344 sc->provider->sectorsize = sc->sectorsize;
1345 sc->provider->mediasize = sc->virsize;
1346 g_error_provider(sc->provider, 0);
1347
1348 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1349 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1350 "chunk %u", sc->provider->name, sc->curr_component,
1351 sc->components[sc->curr_component].chunk_next);
1352 }
1353
1354 /*
1355 * Returns count of active providers in this geom instance
1356 */
1357 static u_int
1358 virstor_valid_components(struct g_virstor_softc *sc)
1359 {
1360 unsigned int nc, i;
1361
1362 nc = 0;
1363 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1364 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1365 for (i = 0; i < sc->n_components; i++)
1366 if (sc->components[i].gcons != NULL)
1367 nc++;
1368 return (nc);
1369 }
1370
1371 /*
1372 * Called when the consumer gets orphaned (?)
1373 */
1374 static void
1375 g_virstor_orphan(struct g_consumer *cp)
1376 {
1377 struct g_virstor_softc *sc;
1378 struct g_virstor_component *comp;
1379 struct g_geom *gp;
1380
1381 g_topology_assert();
1382 gp = cp->geom;
1383 sc = gp->softc;
1384 if (sc == NULL)
1385 return;
1386
1387 comp = cp->private;
1388 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1389 __func__));
1390 remove_component(sc, comp, FALSE);
1391 if (virstor_valid_components(sc) == 0)
1392 virstor_geom_destroy(sc, TRUE, FALSE);
1393 }
1394
1395 /*
1396 * Called to notify geom when it's been opened, and for what intent
1397 */
1398 static int
1399 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1400 {
1401 struct g_consumer *c;
1402 struct g_virstor_softc *sc;
1403 struct g_geom *gp;
1404 int error;
1405
1406 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1407 gp = pp->geom;
1408 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1409 sc = gp->softc;
1410
1411 if (sc == NULL) {
1412 /* It seems that .access can be called with negative dr,dw,dx
1413 * in this case but I want to check for myself */
1414 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s",
1415 dr, dw, de, pp->name);
1416 /* This should only happen when geom is withered so
1417 * allow only negative requests */
1418 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
1419 ("%s: Positive access for %s", __func__, pp->name));
1420 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
1421 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed",
1422 pp->name);
1423 return (0);
1424 }
1425
1426 /* Grab an exclusive bit to propagate on our consumers on first open */
1427 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1428 de++;
1429 /* ... drop it on close */
1430 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1431 de--;
1432 update_metadata(sc); /* Writes statistical information */
1433 }
1434
1435 error = ENXIO;
1436 LIST_FOREACH(c, &gp->consumer, consumer) {
1437 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
1438 error = g_access(c, dr, dw, de);
1439 if (error != 0) {
1440 struct g_consumer *c2;
1441
1442 /* Backout earlier changes */
1443 LIST_FOREACH(c2, &gp->consumer, consumer) {
1444 if (c2 == c) /* all eariler components fixed */
1445 return (error);
1446 g_access(c2, -dr, -dw, -de);
1447 }
1448 }
1449 }
1450
1451 return (error);
1452 }
1453
1454 /*
1455 * Generate XML dump of current state
1456 */
1457 static void
1458 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1459 struct g_consumer *cp, struct g_provider *pp)
1460 {
1461 struct g_virstor_softc *sc;
1462
1463 g_topology_assert();
1464 sc = gp->softc;
1465
1466 if (sc == NULL || pp != NULL)
1467 return;
1468
1469 if (cp != NULL) {
1470 /* For each component */
1471 struct g_virstor_component *comp;
1472
1473 comp = cp->private;
1474 if (comp == NULL)
1475 return;
1476 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1477 indent, comp->index);
1478 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1479 indent, comp->chunk_count);
1480 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1481 indent, comp->chunk_next);
1482 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1483 indent, comp->chunk_reserved);
1484 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1485 indent,
1486 comp->chunk_next > 0 ? 100 -
1487 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1488 comp->chunk_count : 100);
1489 } else {
1490 /* For the whole thing */
1491 u_int count, used, i;
1492 off_t size;
1493
1494 count = used = size = 0;
1495 for (i = 0; i < sc->n_components; i++) {
1496 if (sc->components[i].gcons != NULL) {
1497 count += sc->components[i].chunk_count;
1498 used += sc->components[i].chunk_next +
1499 sc->components[i].chunk_reserved;
1500 size += sc->components[i].gcons->
1501 provider->mediasize;
1502 }
1503 }
1504
1505 sbuf_printf(sb, "%s<Status>"
1506 "Components=%u, Online=%u</Status>\n", indent,
1507 sc->n_components, virstor_valid_components(sc));
1508 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1509 indent, 100-(used * 100) / count);
1510 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1511 sc->chunk_size);
1512 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1513 indent, used > 0 ? 100 - (used * 100) / count : 100);
1514 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1515 indent, count);
1516 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1517 indent, sc->chunk_count);
1518 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1519 indent,
1520 (count * 100) / sc->chunk_count);
1521 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1522 indent, size);
1523 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1524 sc->virsize);
1525 }
1526 }
1527
1528 /*
1529 * GEOM .done handler
1530 * Can't use standard handler because one requested IO may
1531 * fork into additional data IOs
1532 */
1533 static void
1534 g_virstor_done(struct bio *b)
1535 {
1536 struct g_virstor_softc *sc;
1537 struct bio *parent_b;
1538
1539 parent_b = b->bio_parent;
1540 sc = parent_b->bio_to->geom->softc;
1541
1542 if (b->bio_error != 0) {
1543 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1544 b->bio_error, b->bio_offset, b->bio_length,
1545 b->bio_to->name);
1546 if (parent_b->bio_error == 0)
1547 parent_b->bio_error = b->bio_error;
1548 }
1549
1550 parent_b->bio_inbed++;
1551 parent_b->bio_completed += b->bio_completed;
1552
1553 if (parent_b->bio_children == parent_b->bio_inbed) {
1554 parent_b->bio_completed = parent_b->bio_length;
1555 g_io_deliver(parent_b, parent_b->bio_error);
1556 }
1557 g_destroy_bio(b);
1558 }
1559
1560 /*
1561 * I/O starts here
1562 * Called in g_down thread
1563 */
1564 static void
1565 g_virstor_start(struct bio *b)
1566 {
1567 struct g_virstor_softc *sc;
1568 struct g_virstor_component *comp;
1569 struct bio *cb;
1570 struct g_provider *pp;
1571 char *addr;
1572 off_t offset, length;
1573 struct bio_queue_head bq;
1574 size_t chunk_size; /* cached for convenience */
1575 u_int count;
1576
1577 pp = b->bio_to;
1578 sc = pp->geom->softc;
1579 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1580 b->bio_to->error, b->bio_to->name));
1581
1582 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1583
1584 switch (b->bio_cmd) {
1585 case BIO_READ:
1586 case BIO_WRITE:
1587 case BIO_DELETE:
1588 break;
1589 default:
1590 g_io_deliver(b, EOPNOTSUPP);
1591 return;
1592 }
1593
1594 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1595 bioq_init(&bq);
1596
1597 chunk_size = sc->chunk_size;
1598 addr = b->bio_data;
1599 offset = b->bio_offset; /* virtual offset and length */
1600 length = b->bio_length;
1601
1602 while (length > 0) {
1603 size_t chunk_index, in_chunk_offset, in_chunk_length;
1604 struct virstor_map_entry *me;
1605
1606 chunk_index = offset / chunk_size; /* round downwards */
1607 in_chunk_offset = offset % chunk_size;
1608 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1609 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1610 b->bio_cmd == BIO_READ ? "R" : "W",
1611 offset, length,
1612 chunk_index, in_chunk_offset, in_chunk_length);
1613 me = &sc->map[chunk_index];
1614
1615 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1616 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1617 /* Reads from unallocated chunks return zeroed
1618 * buffers */
1619 if (b->bio_cmd == BIO_READ)
1620 bzero(addr, in_chunk_length);
1621 } else {
1622 comp = &sc->components[me->provider_no];
1623
1624 cb = g_clone_bio(b);
1625 if (cb == NULL) {
1626 bioq_dismantle(&bq);
1627 if (b->bio_error == 0)
1628 b->bio_error = ENOMEM;
1629 g_io_deliver(b, b->bio_error);
1630 return;
1631 }
1632 cb->bio_to = comp->gcons->provider;
1633 cb->bio_done = g_virstor_done;
1634 cb->bio_offset =
1635 (off_t)me->provider_chunk * (off_t)chunk_size
1636 + in_chunk_offset;
1637 cb->bio_length = in_chunk_length;
1638 cb->bio_data = addr;
1639 cb->bio_caller1 = comp;
1640 bioq_disksort(&bq, cb);
1641 }
1642 } else { /* handle BIO_WRITE */
1643 KASSERT(b->bio_cmd == BIO_WRITE,
1644 ("%s: Unknown command %d", __func__,
1645 b->bio_cmd));
1646
1647 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1648 /* We have a virtual chunk, represented by
1649 * the "me" entry, but it's not yet allocated
1650 * (tied to) a physical chunk. So do it now. */
1651 struct virstor_map_entry *data_me;
1652 u_int phys_chunk, comp_no;
1653 off_t s_offset;
1654 int error;
1655
1656 error = allocate_chunk(sc, &comp, &comp_no,
1657 &phys_chunk);
1658 if (error != 0) {
1659 /* We cannot allocate a physical chunk
1660 * to satisfy this request, so we'll
1661 * delay it to when we can...
1662 * XXX: this will prevent the fs from
1663 * being umounted! */
1664 struct g_virstor_bio_q *biq;
1665 biq = malloc(sizeof *biq, M_GVIRSTOR,
1666 M_NOWAIT);
1667 if (biq == NULL) {
1668 bioq_dismantle(&bq);
1669 if (b->bio_error == 0)
1670 b->bio_error = ENOMEM;
1671 g_io_deliver(b, b->bio_error);
1672 return;
1673 }
1674 biq->bio = b;
1675 mtx_lock(&sc->delayed_bio_q_mtx);
1676 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1677 biq, linkage);
1678 mtx_unlock(&sc->delayed_bio_q_mtx);
1679 LOG_MSG(LVL_WARNING, "Delaying BIO "
1680 "(size=%ju) until free physical "
1681 "space can be found on %s",
1682 b->bio_length,
1683 sc->provider->name);
1684 return;
1685 }
1686 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1687 "for %s",
1688 phys_chunk,
1689 comp->gcons->provider->name,
1690 sc->provider->name);
1691
1692 me->provider_no = comp_no;
1693 me->provider_chunk = phys_chunk;
1694 me->flags |= VIRSTOR_MAP_ALLOCATED;
1695
1696 cb = g_clone_bio(b);
1697 if (cb == NULL) {
1698 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1699 me->provider_no = 0;
1700 me->provider_chunk = 0;
1701 bioq_dismantle(&bq);
1702 if (b->bio_error == 0)
1703 b->bio_error = ENOMEM;
1704 g_io_deliver(b, b->bio_error);
1705 return;
1706 }
1707
1708 /* The allocation table is stored continuously
1709 * at the start of the drive. We need to
1710 * calculate the offset of the sector that holds
1711 * this map entry both on the drive and in the
1712 * map array.
1713 * sc_offset will end up pointing to the drive
1714 * sector. */
1715 s_offset = chunk_index * sizeof *me;
1716 s_offset = (s_offset / sc->sectorsize) *
1717 sc->sectorsize;
1718
1719 /* data_me points to map entry sector
1720 * in memory (analoguos to offset) */
1721 data_me = &sc->map[(chunk_index /
1722 sc->me_per_sector) * sc->me_per_sector];
1723
1724 /* Commit sector with map entry to storage */
1725 cb->bio_to = sc->components[0].gcons->provider;
1726 cb->bio_done = g_virstor_done;
1727 cb->bio_offset = s_offset;
1728 cb->bio_data = (char *)data_me;
1729 cb->bio_length = sc->sectorsize;
1730 cb->bio_caller1 = &sc->components[0];
1731 bioq_disksort(&bq, cb);
1732 }
1733
1734 comp = &sc->components[me->provider_no];
1735 cb = g_clone_bio(b);
1736 if (cb == NULL) {
1737 bioq_dismantle(&bq);
1738 if (b->bio_error == 0)
1739 b->bio_error = ENOMEM;
1740 g_io_deliver(b, b->bio_error);
1741 return;
1742 }
1743 /* Finally, handle the data */
1744 cb->bio_to = comp->gcons->provider;
1745 cb->bio_done = g_virstor_done;
1746 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1747 in_chunk_offset;
1748 cb->bio_length = in_chunk_length;
1749 cb->bio_data = addr;
1750 cb->bio_caller1 = comp;
1751 bioq_disksort(&bq, cb);
1752 }
1753 addr += in_chunk_length;
1754 length -= in_chunk_length;
1755 offset += in_chunk_length;
1756 }
1757
1758 /* Fire off bio's here */
1759 count = 0;
1760 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1761 bioq_remove(&bq, cb);
1762 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1763 comp = cb->bio_caller1;
1764 cb->bio_caller1 = NULL;
1765 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1766 cb->bio_offset, cb->bio_length);
1767 g_io_request(cb, comp->gcons);
1768 count++;
1769 }
1770 if (count == 0) { /* We handled everything locally */
1771 b->bio_completed = b->bio_length;
1772 g_io_deliver(b, 0);
1773 }
1774
1775 }
1776
1777 /*
1778 * Allocate a chunk from a physical provider. Returns physical component,
1779 * chunk index relative to the component and the component's index.
1780 */
1781 static int
1782 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1783 u_int *comp_no_p, u_int *chunk)
1784 {
1785 u_int comp_no;
1786
1787 KASSERT(sc->curr_component < sc->n_components,
1788 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1789
1790 comp_no = sc->curr_component;
1791 *comp = &sc->components[comp_no];
1792 dump_component(*comp);
1793 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1794 /* This component is full. Allocate next component */
1795 if (comp_no >= sc->n_components-1) {
1796 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1797 sc->geom->name);
1798 return (-1);
1799 }
1800 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1801 sc->curr_component = ++comp_no;
1802
1803 *comp = &sc->components[comp_no];
1804 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1805 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1806 "(switching to %u/%u: %s)", sc->geom->name,
1807 comp_no+1, sc->n_components,
1808 (*comp)->gcons->provider->name);
1809 /* Take care not to overwrite reserved chunks */
1810 if ( (*comp)->chunk_reserved > 0 &&
1811 (*comp)->chunk_next < (*comp)->chunk_reserved)
1812 (*comp)->chunk_next = (*comp)->chunk_reserved;
1813
1814 (*comp)->flags |=
1815 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1816 dump_component(*comp);
1817 *comp_no_p = comp_no;
1818 *chunk = (*comp)->chunk_next++;
1819 } else {
1820 *comp_no_p = comp_no;
1821 *chunk = (*comp)->chunk_next++;
1822 }
1823 return (0);
1824 }
1825
1826 /* Dump a component */
1827 static void
1828 dump_component(struct g_virstor_component *comp)
1829 {
1830
1831 if (g_virstor_debug < LVL_DEBUG2)
1832 return;
1833 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1834 printf(" chunk_count: %u\n", comp->chunk_count);
1835 printf(" chunk_next: %u\n", comp->chunk_next);
1836 printf(" flags: %u\n", comp->flags);
1837 }
1838
1839 #if 0
1840 /* Dump a map entry */
1841 static void
1842 dump_me(struct virstor_map_entry *me, unsigned int nr)
1843 {
1844 if (g_virstor_debug < LVL_DEBUG)
1845 return;
1846 printf("VIRT. CHUNK #%d: ", nr);
1847 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1848 printf("(unallocated)\n");
1849 else
1850 printf("allocated at provider %u, provider_chunk %u\n",
1851 me->provider_no, me->provider_chunk);
1852 }
1853 #endif
1854
1855 /*
1856 * Dismantle bio_queue and destroy its components
1857 */
1858 static void
1859 bioq_dismantle(struct bio_queue_head *bq)
1860 {
1861 struct bio *b;
1862
1863 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1864 bioq_remove(bq, b);
1865 g_destroy_bio(b);
1866 }
1867 }
1868
1869 /*
1870 * The function that shouldn't be called.
1871 * When this is called, the stack is already garbled because of
1872 * argument mismatch. There's nothing to do now but panic, which is
1873 * accidentally the whole purpose of this function.
1874 * Motivation: to guard from accidentally calling geom methods when
1875 * they shouldn't be called. (see g_..._taste)
1876 */
1877 static void
1878 invalid_call(void)
1879 {
1880 panic("invalid_call() has just been called. Something's fishy here.");
1881 }
1882
1883 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */
Cache object: 50cff7def098927c2a7ad65b9cd5d525
|