2 * Copyright (c) 2006-2007 Ivan Voras <ivoras@freebsd.org>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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.
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
27 /* Implementation notes:
28 * - "Components" are wrappers around providers that make up the
29 * virtual storage (i.e. a virstor has "physical" components)
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/module.h>
40 #include <sys/mutex.h>
44 #include <sys/sysctl.h>
45 #include <sys/malloc.h>
48 #include <sys/kthread.h>
49 #include <sys/mutex.h>
51 #include <geom/geom.h>
53 #include <geom/virstor/g_virstor.h>
54 #include <geom/virstor/g_virstor_md.h>
56 FEATURE(g_virstor, "GEOM virtual storage support");
58 /* Declare malloc(9) label */
59 static MALLOC_DEFINE(M_GVIRSTOR, "gvirstor", "GEOM_VIRSTOR Data");
61 /* GEOM class methods */
62 static g_init_t g_virstor_init;
63 static g_fini_t g_virstor_fini;
64 static g_taste_t g_virstor_taste;
65 static g_ctl_req_t g_virstor_config;
66 static g_ctl_destroy_geom_t g_virstor_destroy_geom;
68 /* Declare & initialize class structure ("geom class") */
69 struct g_class g_virstor_class = {
70 .name = G_VIRSTOR_CLASS_NAME,
72 .init = g_virstor_init,
73 .fini = g_virstor_fini,
74 .taste = g_virstor_taste,
75 .ctlreq = g_virstor_config,
76 .destroy_geom = g_virstor_destroy_geom
77 /* The .dumpconf and the rest are only usable for a geom instance, so
78 * they will be set when such instance is created. */
81 /* Declare sysctl's and loader tunables */
82 SYSCTL_DECL(_kern_geom);
83 SYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0, "GEOM_GVIRSTOR information");
85 static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */
86 TUNABLE_INT("kern.geom.virstor.debug", &g_virstor_debug);
87 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RW, &g_virstor_debug,
88 0, "Debug level (2=production, 5=normal, 15=excessive)");
90 static u_int g_virstor_chunk_watermark = 100;
91 TUNABLE_INT("kern.geom.virstor.chunk_watermark", &g_virstor_chunk_watermark);
92 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RW,
93 &g_virstor_chunk_watermark, 0,
94 "Minimum number of free chunks before issuing administrative warning");
96 static u_int g_virstor_component_watermark = 1;
97 TUNABLE_INT("kern.geom.virstor.component_watermark",
98 &g_virstor_component_watermark);
99 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RW,
100 &g_virstor_component_watermark, 0,
101 "Minimum number of free components before issuing administrative warning");
103 static int read_metadata(struct g_consumer *, struct g_virstor_metadata *);
104 static void write_metadata(struct g_consumer *, struct g_virstor_metadata *);
105 static int clear_metadata(struct g_virstor_component *);
106 static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *,
107 struct g_virstor_metadata *);
108 static struct g_geom *create_virstor_geom(struct g_class *,
109 struct g_virstor_metadata *);
110 static void virstor_check_and_run(struct g_virstor_softc *);
111 static u_int virstor_valid_components(struct g_virstor_softc *);
112 static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t,
114 static void remove_component(struct g_virstor_softc *,
115 struct g_virstor_component *, boolean_t);
116 static void bioq_dismantle(struct bio_queue_head *);
117 static int allocate_chunk(struct g_virstor_softc *,
118 struct g_virstor_component **, u_int *, u_int *);
119 static void delay_destroy_consumer(void *, int);
120 static void dump_component(struct g_virstor_component *comp);
122 static void dump_me(struct virstor_map_entry *me, unsigned int nr);
125 static void virstor_ctl_stop(struct gctl_req *, struct g_class *);
126 static void virstor_ctl_add(struct gctl_req *, struct g_class *);
127 static void virstor_ctl_remove(struct gctl_req *, struct g_class *);
128 static struct g_virstor_softc * virstor_find_geom(const struct g_class *,
130 static void update_metadata(struct g_virstor_softc *);
131 static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *,
134 static void g_virstor_orphan(struct g_consumer *);
135 static int g_virstor_access(struct g_provider *, int, int, int);
136 static void g_virstor_start(struct bio *);
137 static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *,
138 struct g_consumer *, struct g_provider *);
139 static void g_virstor_done(struct bio *);
141 static void invalid_call(void);
143 * Initialise GEOM class (per-class callback)
146 g_virstor_init(struct g_class *mp __unused)
149 /* Catch map struct size mismatch at compile time; Map entries must
150 * fit into MAXPHYS exactly, with no wasted space. */
151 CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS);
153 /* Init UMA zones, TAILQ's, other global vars */
157 * Finalise GEOM class (per-class callback)
160 g_virstor_fini(struct g_class *mp __unused)
163 /* Deinit UMA zones & global vars */
167 * Config (per-class callback)
170 g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb)
176 version = gctl_get_paraml(req, "version", sizeof(*version));
177 if (version == NULL) {
178 gctl_error(req, "Failed to get 'version' argument");
181 if (*version != G_VIRSTOR_VERSION) {
182 gctl_error(req, "Userland and kernel versions out of sync");
187 if (strcmp(verb, "add") == 0)
188 virstor_ctl_add(req, cp);
189 else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0)
190 virstor_ctl_stop(req, cp);
191 else if (strcmp(verb, "remove") == 0)
192 virstor_ctl_remove(req, cp);
194 gctl_error(req, "unknown verb: '%s'", verb);
199 * "stop" verb from userland
202 virstor_ctl_stop(struct gctl_req *req, struct g_class *cp)
207 nargs = gctl_get_paraml(req, "nargs", sizeof *nargs);
209 gctl_error(req, "Error fetching argument '%s'", "nargs");
213 gctl_error(req, "Invalid number of arguments");
216 force = gctl_get_paraml(req, "force", sizeof *force);
218 gctl_error(req, "Error fetching argument '%s'", "force");
223 for (i = 0; i < *nargs; i++) {
226 struct g_virstor_softc *sc;
229 sprintf(param, "arg%d", i);
230 name = gctl_get_asciiparam(req, param);
232 gctl_error(req, "No 'arg%d' argument", i);
236 sc = virstor_find_geom(cp, name);
238 gctl_error(req, "Don't know anything about '%s'", name);
243 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
246 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
247 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
248 sc->geom->name, error);
255 * "add" verb from userland - add new component(s) to the structure.
256 * This will be done all at once in here, without going through the
257 * .taste function for new components.
260 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
262 /* Note: while this is going on, I/O is being done on
263 * the g_up and g_down threads. The idea is to make changes
264 * to softc members in a way that can atomically activate
265 * them all at once. */
266 struct g_virstor_softc *sc;
267 int *hardcode, *nargs;
268 const char *geom_name; /* geom to add a component to */
269 struct g_consumer *fcp;
270 struct g_virstor_bio_q *bq;
275 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
277 gctl_error(req, "Error fetching argument '%s'", "nargs");
281 gctl_error(req, "Invalid number of arguments");
284 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
285 if (hardcode == NULL) {
286 gctl_error(req, "Error fetching argument '%s'", "hardcode");
290 /* Find "our" geom */
291 geom_name = gctl_get_asciiparam(req, "arg0");
292 if (geom_name == NULL) {
293 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
296 sc = virstor_find_geom(cp, geom_name);
298 gctl_error(req, "Don't know anything about '%s'", geom_name);
302 if (virstor_valid_components(sc) != sc->n_components) {
303 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
304 "virstor %s", sc->geom->name);
305 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
309 fcp = sc->components[0].gcons;
312 for (i = 1; i < *nargs; i++) {
313 struct g_virstor_metadata md;
315 const char *prov_name;
316 struct g_provider *pp;
317 struct g_consumer *cp;
321 snprintf(aname, sizeof aname, "arg%d", i);
322 prov_name = gctl_get_asciiparam(req, aname);
323 if (prov_name == NULL) {
324 gctl_error(req, "Error fetching argument '%s'", aname);
328 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
329 prov_name += sizeof(_PATH_DEV) - 1;
331 pp = g_provider_by_name(prov_name);
333 /* This is the most common error so be verbose about it */
335 gctl_error(req, "Invalid provider: '%s' (added"
336 " %u components)", prov_name, added);
339 gctl_error(req, "Invalid provider: '%s'",
345 cp = g_new_consumer(sc->geom);
347 gctl_error(req, "Cannot create consumer");
351 error = g_attach(cp, pp);
353 gctl_error(req, "Cannot attach a consumer to %s",
355 g_destroy_consumer(cp);
359 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
360 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
362 gctl_error(req, "Access request failed for %s",
364 g_destroy_consumer(cp);
369 if (fcp->provider->sectorsize != pp->sectorsize) {
370 gctl_error(req, "Sector size doesn't fit for %s",
372 g_destroy_consumer(cp);
376 for (j = 0; j < sc->n_components; j++) {
377 if (strcmp(sc->components[j].gcons->provider->name,
379 gctl_error(req, "Component %s already in %s",
380 pp->name, sc->geom->name);
381 g_destroy_consumer(cp);
386 sc->components = realloc(sc->components,
387 sizeof(*sc->components) * (sc->n_components + 1),
388 M_GVIRSTOR, M_WAITOK);
390 nc = sc->n_components;
391 sc->components[nc].gcons = cp;
392 sc->components[nc].sc = sc;
393 sc->components[nc].index = nc;
394 sc->components[nc].chunk_count = cp->provider->mediasize /
396 sc->components[nc].chunk_next = 0;
397 sc->components[nc].chunk_reserved = 0;
399 if (sc->components[nc].chunk_count < 4) {
400 gctl_error(req, "Provider too small: %s",
402 g_destroy_consumer(cp);
406 fill_metadata(sc, &md, nc, *hardcode);
407 write_metadata(cp, &md);
408 /* The new component becomes visible when n_components is
414 /* This call to update_metadata() is critical. In case there's a
415 * power failure in the middle of it and some components are updated
416 * while others are not, there will be trouble on next .taste() iff
417 * a non-updated component is detected first */
420 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
422 /* Fire off BIOs previously queued because there wasn't any
423 * physical space left. If the BIOs still can't be satisfied
424 * they will again be added to the end of the queue (during
425 * which the mutex will be recursed) */
426 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
428 mtx_lock(&sc->delayed_bio_q_mtx);
429 /* First, insert a sentinel to the queue end, so we don't
430 * end up in an infinite loop if there's still no free
431 * space available. */
432 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
433 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
434 bq = STAILQ_FIRST(&sc->delayed_bio_q);
435 if (bq->bio != NULL) {
436 g_virstor_start(bq->bio);
437 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
438 free(bq, M_GVIRSTOR);
440 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
441 free(bq, M_GVIRSTOR);
445 mtx_unlock(&sc->delayed_bio_q_mtx);
450 * Find a geom handled by the class
452 static struct g_virstor_softc *
453 virstor_find_geom(const struct g_class *cp, const char *name)
457 LIST_FOREACH(gp, &cp->geom, geom) {
458 if (strcmp(name, gp->name) == 0)
465 * Update metadata on all components to reflect the current state
470 * Expects things to be set up so write_metadata() can work, i.e.
471 * the topology lock must be held.
474 update_metadata(struct g_virstor_softc *sc)
476 struct g_virstor_metadata md;
479 if (virstor_valid_components(sc) != sc->n_components)
480 return; /* Incomplete device */
481 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
483 /* Update metadata on components */
484 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
485 sc->geom->class->name, sc->geom->name);
487 for (n = 0; n < sc->n_components; n++) {
488 read_metadata(sc->components[n].gcons, &md);
489 md.chunk_next = sc->components[n].chunk_next;
490 md.flags = sc->components[n].flags;
491 md.md_count = sc->n_components;
492 write_metadata(sc->components[n].gcons, &md);
497 * Fills metadata (struct md) from information stored in softc and the nc'th
498 * component of virstor
501 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
502 u_int nc, u_int hardcode)
504 struct g_virstor_component *c;
506 bzero(md, sizeof *md);
507 c = &sc->components[nc];
509 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
510 md->md_version = G_VIRSTOR_VERSION;
511 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
513 md->md_virsize = sc->virsize;
514 md->md_chunk_size = sc->chunk_size;
515 md->md_count = sc->n_components;
518 strncpy(md->provider, c->gcons->provider->name,
519 sizeof md->provider);
522 md->provsize = c->gcons->provider->mediasize;
523 md->chunk_count = c->chunk_count;
524 md->chunk_next = c->chunk_next;
525 md->chunk_reserved = c->chunk_reserved;
526 md->flags = c->flags;
530 * Remove a component from virstor device.
531 * Can only be done if the component is unallocated.
534 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
536 /* As this is executed in parallel to I/O, operations on virstor
537 * structures must be as atomic as possible. */
538 struct g_virstor_softc *sc;
540 const char *geom_name;
544 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
546 gctl_error(req, "Error fetching argument '%s'", "nargs");
550 gctl_error(req, "Invalid number of arguments");
553 /* Find "our" geom */
554 geom_name = gctl_get_asciiparam(req, "arg0");
555 if (geom_name == NULL) {
556 gctl_error(req, "Error fetching argument '%s'",
560 sc = virstor_find_geom(cp, geom_name);
562 gctl_error(req, "Don't know anything about '%s'", geom_name);
566 if (virstor_valid_components(sc) != sc->n_components) {
567 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
568 "virstor %s", sc->geom->name);
569 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
574 for (i = 1; i < *nargs; i++) {
576 const char *prov_name;
578 struct g_virstor_component *newcomp, *compbak;
580 sprintf(param, "arg%d", i);
581 prov_name = gctl_get_asciiparam(req, param);
582 if (prov_name == NULL) {
583 gctl_error(req, "Error fetching argument '%s'", param);
586 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
587 prov_name += sizeof(_PATH_DEV) - 1;
590 for (j = 0; j < sc->n_components; j++) {
591 if (strcmp(sc->components[j].gcons->provider->name,
598 LOG_MSG(LVL_ERROR, "No %s component in %s",
599 prov_name, sc->geom->name);
603 compbak = sc->components;
604 newcomp = malloc(sc->n_components * sizeof(*sc->components),
605 M_GVIRSTOR, M_WAITOK | M_ZERO);
606 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
607 bcopy(&sc->components[found + 1], newcomp + found,
608 found * sizeof(*sc->components));
609 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
610 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
612 prov_name, sc->geom->name);
613 free(newcomp, M_GVIRSTOR);
614 /* We'll consider this non-fatal error */
617 /* Renumerate unallocated components */
618 for (j = 0; j < sc->n_components-1; j++) {
619 if ((sc->components[j].flags &
620 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
621 sc->components[j].index = j;
624 /* This is the critical section. If a component allocation
625 * event happens while both variables are not yet set,
626 * there will be trouble. Something will panic on encountering
627 * NULL sc->components[x].gcomp member.
628 * Luckily, component allocation happens very rarely and
629 * removing components is an abnormal action in any case. */
630 sc->components = newcomp;
632 /* End critical section */
635 if (clear_metadata(&compbak[found]) != 0) {
636 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
637 "metadata on %s", prov_name);
639 g_detach(compbak[found].gcons);
640 g_destroy_consumer(compbak[found].gcons);
643 free(compbak, M_GVIRSTOR);
648 /* This call to update_metadata() is critical. In case there's a
649 * power failure in the middle of it and some components are updated
650 * while others are not, there will be trouble on next .taste() iff
651 * a non-updated component is detected first */
655 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
660 * Clear metadata sector on component
663 clear_metadata(struct g_virstor_component *comp)
668 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
669 comp->gcons->provider->name);
671 error = g_access(comp->gcons, 0, 1, 0);
674 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
676 error = g_write_data(comp->gcons,
677 comp->gcons->provider->mediasize -
678 comp->gcons->provider->sectorsize,
680 comp->gcons->provider->sectorsize);
681 free(buf, M_GVIRSTOR);
682 g_access(comp->gcons, 0, -1, 0);
687 * Destroy geom forcibly.
690 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
693 struct g_virstor_softc *sc;
697 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
700 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
711 LOG_MSG(LVL_INFO, "INVARIANTS detected");
712 LOG_MSG(LVL_INFO, "Verifying allocation "
713 "table for %s", sc->geom->name);
715 for (n = 0; n < sc->chunk_count; n++) {
716 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
719 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
720 sc->geom->name, count);
723 if (virstor_valid_components(sc) != sc->n_components) {
724 /* This is a incomplete virstor device (not all
725 * components have been found) */
726 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
730 error = g_access(sc->components[0].gcons, 1, 0, 0);
731 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
733 /* Compare the whole on-disk allocation table with what's
734 * currently in memory */
735 while (n < sc->chunk_count) {
736 buf = g_read_data(sc->components[0].gcons, off,
737 sc->sectorsize, &error);
738 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
739 "for read at %jd", error, off));
740 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
741 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
742 "entry %d, offset %jd", n, off);
746 n += sc->me_per_sector;
747 off += sc->sectorsize;
750 error = g_access(sc->components[0].gcons, -1, 0, 0);
751 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
754 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
755 "(%d sectors don't match, max %zu allocations)",
756 sc->geom->name, count,
757 count * sc->me_per_sector);
759 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
765 virstor_geom_destroy(sc, FALSE, FALSE);
773 * Taste event (per-class callback)
774 * Examines a provider and creates geom instances if needed
776 static struct g_geom *
777 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
779 struct g_virstor_metadata md;
781 struct g_consumer *cp;
782 struct g_virstor_softc *sc;
785 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
787 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
789 /* We need a dummy geom to attach a consumer to the given provider */
790 gp = g_new_geomf(mp, "virstor:taste.helper");
791 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
792 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
793 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
795 cp = g_new_consumer(gp);
797 error = read_metadata(cp, &md);
799 g_destroy_consumer(cp);
805 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
807 if (md.md_version != G_VIRSTOR_VERSION) {
808 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
809 "to handle %s (%s) : %d should be %d",
810 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
813 if (md.provsize != pp->mediasize)
816 /* If the provider name is hardcoded, use the offered provider only
817 * if it's been offered with its proper name (the one used in
818 * the label command). */
819 if (md.provider[0] != '\0' &&
820 !g_compare_names(md.provider, pp->name))
823 /* Iterate all geoms this class already knows about to see if a new
824 * geom instance of this class needs to be created (in case the provider
825 * is first from a (possibly) multi-consumer geom) or it just needs
826 * to be added to an existing instance. */
829 LIST_FOREACH(gp, &mp->geom, geom) {
833 if (strcmp(md.md_name, sc->geom->name) != 0)
835 if (md.md_id != sc->id)
839 if (gp != NULL) { /* We found an existing geom instance; add to it */
840 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
841 error = add_provider_to_geom(sc, pp, &md);
843 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
844 pp->name, md.md_name, error);
847 } else { /* New geom instance needs to be created */
848 gp = create_virstor_geom(mp, &md);
850 LOG_MSG(LVL_ERROR, "Error creating new instance of "
851 "class %s: %s", mp->name, md.md_name);
852 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
853 md.md_name, pp->name);
857 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
859 error = add_provider_to_geom(sc, pp, &md);
861 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
862 pp->name, md.md_name, error);
863 virstor_geom_destroy(sc, TRUE, FALSE);
872 * Destroyes consumer passed to it in arguments. Used as a callback
876 delay_destroy_consumer(void *arg, int flags __unused)
878 struct g_consumer *c = arg;
879 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
880 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
883 g_destroy_consumer(c);
887 * Remove a component (consumer) from geom instance; If it's the first
888 * component being removed, orphan the provider to announce geom's being
892 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
895 struct g_consumer *c;
897 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
902 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
903 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
905 if (sc->provider != NULL) {
907 sc->provider->flags |= G_PF_WITHER;
908 g_orphan_provider(sc->provider, ENXIO);
910 LOG_MSG(LVL_INFO, "Removing provider %s", sc->geom->name);
913 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
914 g_access(c, -c->acr, -c->acw, -c->ace);
916 /* Destroy consumer after it's tasted */
917 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
920 g_destroy_consumer(c);
925 * Destroy geom - called internally
926 * See g_virstor_destroy_geom for the other one
929 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
932 struct g_provider *pp;
942 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
943 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
944 "Device %s is still open.", pp->name);
949 for (n = 0; n < sc->n_components; n++) {
950 if (sc->components[n].gcons != NULL)
951 remove_component(sc, &sc->components[n], delay);
957 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
960 /* XXX: This might or might not work, since we're called with
961 * the topology lock held. Also, it might panic the kernel if
962 * the error'd BIO is in softupdates code. */
963 mtx_lock(&sc->delayed_bio_q_mtx);
964 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
965 struct g_virstor_bio_q *bq;
966 bq = STAILQ_FIRST(&sc->delayed_bio_q);
967 bq->bio->bio_error = ENOSPC;
968 g_io_deliver(bq->bio, EIO);
969 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
970 free(bq, M_GVIRSTOR);
972 mtx_unlock(&sc->delayed_bio_q_mtx);
973 mtx_destroy(&sc->delayed_bio_q_mtx);
975 free(sc->map, M_GVIRSTOR);
976 free(sc->components, M_GVIRSTOR);
977 bzero(sc, sizeof *sc);
978 free(sc, M_GVIRSTOR);
980 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
981 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
982 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
984 g_wither_geom(gp, ENXIO);
990 * Utility function: read metadata & decode. Wants topology lock to be
994 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
996 struct g_provider *pp;
1000 g_topology_assert();
1001 error = g_access(cp, 1, 0, 0);
1005 g_topology_unlock();
1006 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
1009 g_access(cp, -1, 0, 0);
1013 virstor_metadata_decode(buf, md);
1020 * Utility function: encode & write metadata. Assumes topology lock is
1023 * There is no useful way of recovering from errors in this function,
1024 * not involving panicking the kernel. If the metadata cannot be written
1025 * the most we can do is notify the operator and hope he spots it and
1026 * replaces the broken drive.
1029 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1031 struct g_provider *pp;
1035 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1036 ("Something's fishy in %s", __func__));
1037 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1038 g_topology_assert();
1039 error = g_access(cp, 0, 1, 0);
1041 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d",
1042 cp->provider->name, error);
1047 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1048 virstor_metadata_encode(md, buf);
1049 g_topology_unlock();
1050 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1053 g_access(cp, 0, -1, 0);
1054 free(buf, M_GVIRSTOR);
1057 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s",
1058 error, cp->provider->name);
1062 * Creates a new instance of this GEOM class, initialise softc
1064 static struct g_geom *
1065 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1068 struct g_virstor_softc *sc;
1070 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1071 md->md_name, md->md_id);
1073 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1074 md->md_virsize < md->md_chunk_size) {
1075 /* This is bogus configuration, and probably means data is
1076 * somehow corrupted. Panic, maybe? */
1077 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1082 /* Check if it's already created */
1083 LIST_FOREACH(gp, &mp->geom, geom) {
1085 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1086 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1088 if (sc->id != md->md_id) {
1090 "Some stale or invalid components "
1091 "exist for virstor device named %s. "
1092 "You will need to <CLEAR> all stale "
1093 "components and maybe reconfigure "
1094 "the virstor device. Tune "
1095 "kern.geom.virstor.debug sysctl up "
1096 "for more information.",
1102 gp = g_new_geomf(mp, "%s", md->md_name);
1103 gp->softc = NULL; /* to circumevent races that test softc */
1105 gp->start = g_virstor_start;
1106 gp->spoiled = g_virstor_orphan;
1107 gp->orphan = g_virstor_orphan;
1108 gp->access = g_virstor_access;
1109 gp->dumpconf = g_virstor_dumpconf;
1111 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1113 sc->n_components = md->md_count;
1114 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1115 M_GVIRSTOR, M_WAITOK | M_ZERO);
1116 sc->chunk_size = md->md_chunk_size;
1117 sc->virsize = md->md_virsize;
1118 STAILQ_INIT(&sc->delayed_bio_q);
1119 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1120 "gvirstor", MTX_DEF | MTX_RECURSE);
1123 sc->provider = NULL; /* virstor_check_and_run will create it */
1126 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1132 * Add provider to a GEOM class instance
1135 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1136 struct g_virstor_metadata *md)
1138 struct g_virstor_component *component;
1139 struct g_consumer *cp, *fcp;
1143 if (md->no >= sc->n_components)
1146 /* "Current" compontent */
1147 component = &(sc->components[md->no]);
1148 if (component->gcons != NULL)
1152 fcp = LIST_FIRST(&gp->consumer);
1154 cp = g_new_consumer(gp);
1155 error = g_attach(cp, pp);
1158 g_destroy_consumer(cp);
1163 if (fcp->provider->sectorsize != pp->sectorsize) {
1164 /* TODO: this can be made to work */
1165 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1166 "sector size (%d)", pp->name, sc->geom->name,
1170 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1171 /* Replicate access permissions from first "live" consumer
1173 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1176 g_destroy_consumer(cp);
1182 /* Bring up a new component */
1183 cp->private = component;
1184 component->gcons = cp;
1186 component->index = md->no;
1187 component->chunk_count = md->chunk_count;
1188 component->chunk_next = md->chunk_next;
1189 component->chunk_reserved = md->chunk_reserved;
1190 component->flags = md->flags;
1192 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1194 virstor_check_and_run(sc);
1199 * Check if everything's ready to create the geom provider & device entry,
1200 * create and start provider.
1201 * Called ultimately by .taste, from g_event thread
1204 virstor_check_and_run(struct g_virstor_softc *sc)
1211 if (virstor_valid_components(sc) != sc->n_components)
1214 if (virstor_valid_components(sc) == 0) {
1215 /* This is actually a candidate for panic() */
1216 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1217 sc->provider->name);
1221 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1223 /* Initialise allocation map from the first consumer */
1224 sc->chunk_count = sc->virsize / sc->chunk_size;
1225 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1226 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1228 sc->chunk_count * (off_t)sc->chunk_size);
1230 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1231 /* The following allocation is in order of 4MB - 8MB */
1232 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1233 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1234 __func__, sc->map_size, sc->provider->name));
1235 sc->map_sectors = sc->map_size / sc->sectorsize;
1238 for (n = 0; n < sc->n_components; n++)
1239 count += sc->components[n].chunk_count;
1240 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1242 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1244 error = g_access(sc->components[0].gcons, 1, 0, 0);
1246 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1247 "read allocation map for %s",
1248 sc->components[0].gcons->provider->name,
1252 /* Read in the allocation map */
1253 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1254 sc->components[0].gcons->provider->name);
1255 off = count = n = 0;
1256 while (count < sc->map_size) {
1257 struct g_virstor_map_entry *mapbuf;
1260 bs = MIN(MAXPHYS, sc->map_size - count);
1261 if (bs % sc->sectorsize != 0) {
1262 /* Check for alignment errors */
1263 bs = (bs / sc->sectorsize) * sc->sectorsize;
1266 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1267 "for %s on %s", sc->geom->name,
1268 sc->components[0].gcons->provider->name);
1270 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1271 if (mapbuf == NULL) {
1272 free(sc->map, M_GVIRSTOR);
1273 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1274 "for %s from %s (offset %ju) (error %d)",
1276 sc->components[0].gcons->provider->name,
1281 bcopy(mapbuf, &sc->map[n], bs);
1284 n += bs / sizeof *(sc->map);
1287 g_access(sc->components[0].gcons, -1, 0, 0);
1288 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1290 /* find first component with allocatable chunks */
1292 for (n = 0; n < sc->n_components; n++) {
1293 if (sc->components[n].chunk_next <
1294 sc->components[n].chunk_count) {
1300 /* not found? set it to the last component and handle it
1302 index = sc->n_components - 1;
1304 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1305 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1306 "(%d/%u: %s)", sc->geom->name,
1309 sc->components[index].gcons->provider->name);
1311 sc->curr_component = index;
1313 if (sc->components[index].chunk_next >=
1314 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1315 LOG_MSG(LVL_WARNING,
1316 "Component %s of %s is running out of free space "
1318 sc->components[index].gcons->provider->name,
1319 sc->geom->name, sc->components[index].chunk_count -
1320 sc->components[index].chunk_next);
1323 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1324 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1326 "%s: Map entries don't fit exactly in a sector (%s)",
1327 __func__, sc->geom->name);
1331 /* Recalculate allocated chunks in components & at the same time
1332 * verify map data is sane. We could trust metadata on this, but
1333 * we want to make sure. */
1334 for (n = 0; n < sc->n_components; n++)
1335 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1337 for (n = 0; n < sc->chunk_count; n++) {
1338 if (sc->map[n].provider_no >= sc->n_components ||
1339 sc->map[n].provider_chunk >=
1340 sc->components[sc->map[n].provider_no].chunk_count) {
1341 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1342 __func__, (u_int)n, sc->geom->name);
1343 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1344 " provider_chunk: %u, chunk_count: %u", __func__,
1345 sc->map[n].provider_no, sc->n_components,
1346 sc->map[n].provider_chunk,
1347 sc->components[sc->map[n].provider_no].chunk_count);
1350 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1351 sc->components[sc->map[n].provider_no].chunk_next++;
1354 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1357 sc->provider->sectorsize = sc->sectorsize;
1358 sc->provider->mediasize = sc->virsize;
1359 g_error_provider(sc->provider, 0);
1361 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1362 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1363 "chunk %u", sc->provider->name, sc->curr_component,
1364 sc->components[sc->curr_component].chunk_next);
1368 * Returns count of active providers in this geom instance
1371 virstor_valid_components(struct g_virstor_softc *sc)
1376 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1377 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1378 for (i = 0; i < sc->n_components; i++)
1379 if (sc->components[i].gcons != NULL)
1385 * Called when the consumer gets orphaned (?)
1388 g_virstor_orphan(struct g_consumer *cp)
1390 struct g_virstor_softc *sc;
1391 struct g_virstor_component *comp;
1394 g_topology_assert();
1401 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1403 remove_component(sc, comp, FALSE);
1404 if (virstor_valid_components(sc) == 0)
1405 virstor_geom_destroy(sc, TRUE, FALSE);
1409 * Called to notify geom when it's been opened, and for what intent
1412 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1414 struct g_consumer *c;
1415 struct g_virstor_softc *sc;
1419 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1421 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1425 /* It seems that .access can be called with negative dr,dw,dx
1426 * in this case but I want to check for myself */
1427 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s",
1428 dr, dw, de, pp->name);
1429 /* This should only happen when geom is withered so
1430 * allow only negative requests */
1431 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
1432 ("%s: Positive access for %s", __func__, pp->name));
1433 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
1434 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed",
1439 /* Grab an exclusive bit to propagate on our consumers on first open */
1440 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1442 /* ... drop it on close */
1443 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1445 update_metadata(sc); /* Writes statistical information */
1449 LIST_FOREACH(c, &gp->consumer, consumer) {
1450 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
1451 error = g_access(c, dr, dw, de);
1453 struct g_consumer *c2;
1455 /* Backout earlier changes */
1456 LIST_FOREACH(c2, &gp->consumer, consumer) {
1457 if (c2 == c) /* all eariler components fixed */
1459 g_access(c2, -dr, -dw, -de);
1468 * Generate XML dump of current state
1471 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1472 struct g_consumer *cp, struct g_provider *pp)
1474 struct g_virstor_softc *sc;
1476 g_topology_assert();
1479 if (sc == NULL || pp != NULL)
1483 /* For each component */
1484 struct g_virstor_component *comp;
1489 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1490 indent, comp->index);
1491 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1492 indent, comp->chunk_count);
1493 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1494 indent, comp->chunk_next);
1495 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1496 indent, comp->chunk_reserved);
1497 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1499 comp->chunk_next > 0 ? 100 -
1500 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1501 comp->chunk_count : 100);
1503 /* For the whole thing */
1504 u_int count, used, i;
1507 count = used = size = 0;
1508 for (i = 0; i < sc->n_components; i++) {
1509 if (sc->components[i].gcons != NULL) {
1510 count += sc->components[i].chunk_count;
1511 used += sc->components[i].chunk_next +
1512 sc->components[i].chunk_reserved;
1513 size += sc->components[i].gcons->
1514 provider->mediasize;
1518 sbuf_printf(sb, "%s<Status>"
1519 "Components=%u, Online=%u</Status>\n", indent,
1520 sc->n_components, virstor_valid_components(sc));
1521 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1522 indent, 100-(used * 100) / count);
1523 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1525 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1526 indent, used > 0 ? 100 - (used * 100) / count : 100);
1527 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1529 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1530 indent, sc->chunk_count);
1531 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1533 (count * 100) / sc->chunk_count);
1534 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1536 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1542 * GEOM .done handler
1543 * Can't use standard handler because one requested IO may
1544 * fork into additional data IOs
1547 g_virstor_done(struct bio *b)
1549 struct g_virstor_softc *sc;
1550 struct bio *parent_b;
1552 parent_b = b->bio_parent;
1553 sc = parent_b->bio_to->geom->softc;
1555 if (b->bio_error != 0) {
1556 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1557 b->bio_error, b->bio_offset, b->bio_length,
1559 if (parent_b->bio_error == 0)
1560 parent_b->bio_error = b->bio_error;
1563 parent_b->bio_inbed++;
1564 parent_b->bio_completed += b->bio_completed;
1566 if (parent_b->bio_children == parent_b->bio_inbed) {
1567 parent_b->bio_completed = parent_b->bio_length;
1568 g_io_deliver(parent_b, parent_b->bio_error);
1575 * Called in g_down thread
1578 g_virstor_start(struct bio *b)
1580 struct g_virstor_softc *sc;
1581 struct g_virstor_component *comp;
1583 struct g_provider *pp;
1585 off_t offset, length;
1586 struct bio_queue_head bq;
1587 size_t chunk_size; /* cached for convenience */
1591 sc = pp->geom->softc;
1592 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1593 b->bio_to->error, b->bio_to->name));
1595 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1597 switch (b->bio_cmd) {
1603 g_io_deliver(b, EOPNOTSUPP);
1607 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1610 chunk_size = sc->chunk_size;
1612 offset = b->bio_offset; /* virtual offset and length */
1613 length = b->bio_length;
1615 while (length > 0) {
1616 size_t chunk_index, in_chunk_offset, in_chunk_length;
1617 struct virstor_map_entry *me;
1619 chunk_index = offset / chunk_size; /* round downwards */
1620 in_chunk_offset = offset % chunk_size;
1621 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1622 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1623 b->bio_cmd == BIO_READ ? "R" : "W",
1625 chunk_index, in_chunk_offset, in_chunk_length);
1626 me = &sc->map[chunk_index];
1628 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1629 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1630 /* Reads from unallocated chunks return zeroed
1632 if (b->bio_cmd == BIO_READ)
1633 bzero(addr, in_chunk_length);
1635 comp = &sc->components[me->provider_no];
1637 cb = g_clone_bio(b);
1639 bioq_dismantle(&bq);
1640 if (b->bio_error == 0)
1641 b->bio_error = ENOMEM;
1642 g_io_deliver(b, b->bio_error);
1645 cb->bio_to = comp->gcons->provider;
1646 cb->bio_done = g_virstor_done;
1648 (off_t)me->provider_chunk * (off_t)chunk_size
1650 cb->bio_length = in_chunk_length;
1651 cb->bio_data = addr;
1652 cb->bio_caller1 = comp;
1653 bioq_disksort(&bq, cb);
1655 } else { /* handle BIO_WRITE */
1656 KASSERT(b->bio_cmd == BIO_WRITE,
1657 ("%s: Unknown command %d", __func__,
1660 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1661 /* We have a virtual chunk, represented by
1662 * the "me" entry, but it's not yet allocated
1663 * (tied to) a physical chunk. So do it now. */
1664 struct virstor_map_entry *data_me;
1665 u_int phys_chunk, comp_no;
1669 error = allocate_chunk(sc, &comp, &comp_no,
1672 /* We cannot allocate a physical chunk
1673 * to satisfy this request, so we'll
1674 * delay it to when we can...
1675 * XXX: this will prevent the fs from
1676 * being umounted! */
1677 struct g_virstor_bio_q *biq;
1678 biq = malloc(sizeof *biq, M_GVIRSTOR,
1681 bioq_dismantle(&bq);
1682 if (b->bio_error == 0)
1683 b->bio_error = ENOMEM;
1684 g_io_deliver(b, b->bio_error);
1688 mtx_lock(&sc->delayed_bio_q_mtx);
1689 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1691 mtx_unlock(&sc->delayed_bio_q_mtx);
1692 LOG_MSG(LVL_WARNING, "Delaying BIO "
1693 "(size=%ju) until free physical "
1694 "space can be found on %s",
1696 sc->provider->name);
1699 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1702 comp->gcons->provider->name,
1703 sc->provider->name);
1705 me->provider_no = comp_no;
1706 me->provider_chunk = phys_chunk;
1707 me->flags |= VIRSTOR_MAP_ALLOCATED;
1709 cb = g_clone_bio(b);
1711 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1712 me->provider_no = 0;
1713 me->provider_chunk = 0;
1714 bioq_dismantle(&bq);
1715 if (b->bio_error == 0)
1716 b->bio_error = ENOMEM;
1717 g_io_deliver(b, b->bio_error);
1721 /* The allocation table is stored continuously
1722 * at the start of the drive. We need to
1723 * calculate the offset of the sector that holds
1724 * this map entry both on the drive and in the
1726 * sc_offset will end up pointing to the drive
1728 s_offset = chunk_index * sizeof *me;
1729 s_offset = (s_offset / sc->sectorsize) *
1732 /* data_me points to map entry sector
1733 * in memory (analoguos to offset) */
1734 data_me = &sc->map[(chunk_index /
1735 sc->me_per_sector) * sc->me_per_sector];
1737 /* Commit sector with map entry to storage */
1738 cb->bio_to = sc->components[0].gcons->provider;
1739 cb->bio_done = g_virstor_done;
1740 cb->bio_offset = s_offset;
1741 cb->bio_data = (char *)data_me;
1742 cb->bio_length = sc->sectorsize;
1743 cb->bio_caller1 = &sc->components[0];
1744 bioq_disksort(&bq, cb);
1747 comp = &sc->components[me->provider_no];
1748 cb = g_clone_bio(b);
1750 bioq_dismantle(&bq);
1751 if (b->bio_error == 0)
1752 b->bio_error = ENOMEM;
1753 g_io_deliver(b, b->bio_error);
1756 /* Finally, handle the data */
1757 cb->bio_to = comp->gcons->provider;
1758 cb->bio_done = g_virstor_done;
1759 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1761 cb->bio_length = in_chunk_length;
1762 cb->bio_data = addr;
1763 cb->bio_caller1 = comp;
1764 bioq_disksort(&bq, cb);
1766 addr += in_chunk_length;
1767 length -= in_chunk_length;
1768 offset += in_chunk_length;
1771 /* Fire off bio's here */
1773 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1774 bioq_remove(&bq, cb);
1775 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1776 comp = cb->bio_caller1;
1777 cb->bio_caller1 = NULL;
1778 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1779 cb->bio_offset, cb->bio_length);
1780 g_io_request(cb, comp->gcons);
1783 if (count == 0) { /* We handled everything locally */
1784 b->bio_completed = b->bio_length;
1791 * Allocate a chunk from a physical provider. Returns physical component,
1792 * chunk index relative to the component and the component's index.
1795 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1796 u_int *comp_no_p, u_int *chunk)
1800 KASSERT(sc->curr_component < sc->n_components,
1801 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1803 comp_no = sc->curr_component;
1804 *comp = &sc->components[comp_no];
1805 dump_component(*comp);
1806 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1807 /* This component is full. Allocate next component */
1808 if (comp_no >= sc->n_components-1) {
1809 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1813 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1814 sc->curr_component = ++comp_no;
1816 *comp = &sc->components[comp_no];
1817 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1818 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1819 "(switching to %u/%u: %s)", sc->geom->name,
1820 comp_no+1, sc->n_components,
1821 (*comp)->gcons->provider->name);
1822 /* Take care not to overwrite reserved chunks */
1823 if ( (*comp)->chunk_reserved > 0 &&
1824 (*comp)->chunk_next < (*comp)->chunk_reserved)
1825 (*comp)->chunk_next = (*comp)->chunk_reserved;
1828 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1829 dump_component(*comp);
1830 *comp_no_p = comp_no;
1831 *chunk = (*comp)->chunk_next++;
1833 *comp_no_p = comp_no;
1834 *chunk = (*comp)->chunk_next++;
1839 /* Dump a component */
1841 dump_component(struct g_virstor_component *comp)
1844 if (g_virstor_debug < LVL_DEBUG2)
1846 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1847 printf(" chunk_count: %u\n", comp->chunk_count);
1848 printf(" chunk_next: %u\n", comp->chunk_next);
1849 printf(" flags: %u\n", comp->flags);
1853 /* Dump a map entry */
1855 dump_me(struct virstor_map_entry *me, unsigned int nr)
1857 if (g_virstor_debug < LVL_DEBUG)
1859 printf("VIRT. CHUNK #%d: ", nr);
1860 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1861 printf("(unallocated)\n");
1863 printf("allocated at provider %u, provider_chunk %u\n",
1864 me->provider_no, me->provider_chunk);
1869 * Dismantle bio_queue and destroy its components
1872 bioq_dismantle(struct bio_queue_head *bq)
1876 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1883 * The function that shouldn't be called.
1884 * When this is called, the stack is already garbled because of
1885 * argument mismatch. There's nothing to do now but panic, which is
1886 * accidentally the whole purpose of this function.
1887 * Motivation: to guard from accidentally calling geom methods when
1888 * they shouldn't be called. (see g_..._taste)
1893 panic("invalid_call() has just been called. Something's fishy here.");
1896 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */