2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2006-2007 Ivan Voras <ivoras@freebsd.org>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 /* Implementation notes:
30 * - "Components" are wrappers around providers that make up the
31 * virtual storage (i.e. a virstor has "physical" components)
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/module.h>
42 #include <sys/mutex.h>
46 #include <sys/sysctl.h>
47 #include <sys/malloc.h>
50 #include <sys/kthread.h>
51 #include <sys/mutex.h>
53 #include <geom/geom.h>
54 #include <geom/geom_dbg.h>
56 #include <geom/virstor/g_virstor.h>
57 #include <geom/virstor/g_virstor_md.h>
59 FEATURE(g_virstor, "GEOM virtual storage support");
61 /* Declare malloc(9) label */
62 static MALLOC_DEFINE(M_GVIRSTOR, "gvirstor", "GEOM_VIRSTOR Data");
64 /* GEOM class methods */
65 static g_init_t g_virstor_init;
66 static g_fini_t g_virstor_fini;
67 static g_taste_t g_virstor_taste;
68 static g_ctl_req_t g_virstor_config;
69 static g_ctl_destroy_geom_t g_virstor_destroy_geom;
71 /* Declare & initialize class structure ("geom class") */
72 struct g_class g_virstor_class = {
73 .name = G_VIRSTOR_CLASS_NAME,
75 .init = g_virstor_init,
76 .fini = g_virstor_fini,
77 .taste = g_virstor_taste,
78 .ctlreq = g_virstor_config,
79 .destroy_geom = g_virstor_destroy_geom
80 /* The .dumpconf and the rest are only usable for a geom instance, so
81 * they will be set when such instance is created. */
84 /* Declare sysctl's and loader tunables */
85 SYSCTL_DECL(_kern_geom);
86 static SYSCTL_NODE(_kern_geom, OID_AUTO, virstor,
87 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
88 "GEOM_GVIRSTOR information");
90 static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */
91 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RWTUN, &g_virstor_debug,
92 0, "Debug level (2=production, 5=normal, 15=excessive)");
94 static u_int g_virstor_chunk_watermark = 100;
95 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RWTUN,
96 &g_virstor_chunk_watermark, 0,
97 "Minimum number of free chunks before issuing administrative warning");
99 static u_int g_virstor_component_watermark = 1;
100 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RWTUN,
101 &g_virstor_component_watermark, 0,
102 "Minimum number of free components before issuing administrative warning");
104 static int read_metadata(struct g_consumer *, struct g_virstor_metadata *);
105 static void write_metadata(struct g_consumer *, struct g_virstor_metadata *);
106 static int clear_metadata(struct g_virstor_component *);
107 static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *,
108 struct g_virstor_metadata *);
109 static struct g_geom *create_virstor_geom(struct g_class *,
110 struct g_virstor_metadata *);
111 static void virstor_check_and_run(struct g_virstor_softc *);
112 static u_int virstor_valid_components(struct g_virstor_softc *);
113 static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t,
115 static void remove_component(struct g_virstor_softc *,
116 struct g_virstor_component *, boolean_t);
117 static void bioq_dismantle(struct bio_queue_head *);
118 static int allocate_chunk(struct g_virstor_softc *,
119 struct g_virstor_component **, u_int *, u_int *);
120 static void delay_destroy_consumer(void *, int);
121 static void dump_component(struct g_virstor_component *comp);
123 static void dump_me(struct virstor_map_entry *me, unsigned int nr);
126 static void virstor_ctl_stop(struct gctl_req *, struct g_class *);
127 static void virstor_ctl_add(struct gctl_req *, struct g_class *);
128 static void virstor_ctl_remove(struct gctl_req *, struct g_class *);
129 static struct g_virstor_softc * virstor_find_geom(const struct g_class *,
131 static void update_metadata(struct g_virstor_softc *);
132 static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *,
135 static void g_virstor_orphan(struct g_consumer *);
136 static int g_virstor_access(struct g_provider *, int, int, int);
137 static void g_virstor_start(struct bio *);
138 static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *,
139 struct g_consumer *, struct g_provider *);
140 static void g_virstor_done(struct bio *);
142 static void invalid_call(void);
144 * Initialise GEOM class (per-class callback)
147 g_virstor_init(struct g_class *mp __unused)
150 /* Catch map struct size mismatch at compile time; Map entries must
151 * fit into maxphys exactly, with no wasted space. */
152 MPASS(VIRSTOR_MAP_BLOCK_ENTRIES * VIRSTOR_MAP_ENTRY_SIZE == maxphys);
154 /* Init UMA zones, TAILQ's, other global vars */
158 * Finalise GEOM class (per-class callback)
161 g_virstor_fini(struct g_class *mp __unused)
164 /* Deinit UMA zones & global vars */
168 * Config (per-class callback)
171 g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb)
177 version = gctl_get_paraml(req, "version", sizeof(*version));
178 if (version == NULL) {
179 gctl_error(req, "Failed to get 'version' argument");
182 if (*version != G_VIRSTOR_VERSION) {
183 gctl_error(req, "Userland and kernel versions out of sync");
188 if (strcmp(verb, "add") == 0)
189 virstor_ctl_add(req, cp);
190 else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0)
191 virstor_ctl_stop(req, cp);
192 else if (strcmp(verb, "remove") == 0)
193 virstor_ctl_remove(req, cp);
195 gctl_error(req, "unknown verb: '%s'", verb);
200 * "stop" verb from userland
203 virstor_ctl_stop(struct gctl_req *req, struct g_class *cp)
208 nargs = gctl_get_paraml(req, "nargs", sizeof *nargs);
210 gctl_error(req, "Error fetching argument '%s'", "nargs");
214 gctl_error(req, "Invalid number of arguments");
217 force = gctl_get_paraml(req, "force", sizeof *force);
219 gctl_error(req, "Error fetching argument '%s'", "force");
224 for (i = 0; i < *nargs; i++) {
227 struct g_virstor_softc *sc;
230 snprintf(param, sizeof(param), "arg%d", i);
231 name = gctl_get_asciiparam(req, param);
233 gctl_error(req, "No 'arg%d' argument", i);
237 sc = virstor_find_geom(cp, name);
239 gctl_error(req, "Don't know anything about '%s'", name);
244 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
247 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
248 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
249 sc->geom->name, error);
256 * "add" verb from userland - add new component(s) to the structure.
257 * This will be done all at once in here, without going through the
258 * .taste function for new components.
261 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
263 /* Note: while this is going on, I/O is being done on
264 * the g_up and g_down threads. The idea is to make changes
265 * to softc members in a way that can atomically activate
266 * them all at once. */
267 struct g_virstor_softc *sc;
268 int *hardcode, *nargs;
269 const char *geom_name; /* geom to add a component to */
270 struct g_consumer *fcp;
271 struct g_virstor_bio_q *bq;
276 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
278 gctl_error(req, "Error fetching argument '%s'", "nargs");
282 gctl_error(req, "Invalid number of arguments");
285 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
286 if (hardcode == NULL) {
287 gctl_error(req, "Error fetching argument '%s'", "hardcode");
291 /* Find "our" geom */
292 geom_name = gctl_get_asciiparam(req, "arg0");
293 if (geom_name == NULL) {
294 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
297 sc = virstor_find_geom(cp, geom_name);
299 gctl_error(req, "Don't know anything about '%s'", geom_name);
303 if (virstor_valid_components(sc) != sc->n_components) {
304 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
305 "virstor %s", sc->geom->name);
306 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
310 fcp = sc->components[0].gcons;
313 for (i = 1; i < *nargs; i++) {
314 struct g_virstor_metadata md;
316 struct g_provider *pp;
317 struct g_consumer *cp;
321 snprintf(aname, sizeof aname, "arg%d", i);
322 pp = gctl_get_provider(req, aname);
324 /* This is the most common error so be verbose about it */
326 gctl_error(req, "Invalid provider. (added"
327 " %u components)", added);
333 cp = g_new_consumer(sc->geom);
335 gctl_error(req, "Cannot create consumer");
339 error = g_attach(cp, pp);
341 gctl_error(req, "Cannot attach a consumer to %s",
343 g_destroy_consumer(cp);
347 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
348 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
350 gctl_error(req, "Access request failed for %s",
352 g_destroy_consumer(cp);
357 if (fcp->provider->sectorsize != pp->sectorsize) {
358 gctl_error(req, "Sector size doesn't fit for %s",
360 g_destroy_consumer(cp);
364 for (j = 0; j < sc->n_components; j++) {
365 if (strcmp(sc->components[j].gcons->provider->name,
367 gctl_error(req, "Component %s already in %s",
368 pp->name, sc->geom->name);
369 g_destroy_consumer(cp);
374 sc->components = realloc(sc->components,
375 sizeof(*sc->components) * (sc->n_components + 1),
376 M_GVIRSTOR, M_WAITOK);
378 nc = sc->n_components;
379 sc->components[nc].gcons = cp;
380 sc->components[nc].sc = sc;
381 sc->components[nc].index = nc;
382 sc->components[nc].chunk_count = cp->provider->mediasize /
384 sc->components[nc].chunk_next = 0;
385 sc->components[nc].chunk_reserved = 0;
387 if (sc->components[nc].chunk_count < 4) {
388 gctl_error(req, "Provider too small: %s",
390 g_destroy_consumer(cp);
394 fill_metadata(sc, &md, nc, *hardcode);
395 write_metadata(cp, &md);
396 /* The new component becomes visible when n_components is
401 /* This call to update_metadata() is critical. In case there's a
402 * power failure in the middle of it and some components are updated
403 * while others are not, there will be trouble on next .taste() iff
404 * a non-updated component is detected first */
407 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
409 /* Fire off BIOs previously queued because there wasn't any
410 * physical space left. If the BIOs still can't be satisfied
411 * they will again be added to the end of the queue (during
412 * which the mutex will be recursed) */
413 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
415 mtx_lock(&sc->delayed_bio_q_mtx);
416 /* First, insert a sentinel to the queue end, so we don't
417 * end up in an infinite loop if there's still no free
418 * space available. */
419 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
420 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
421 bq = STAILQ_FIRST(&sc->delayed_bio_q);
422 if (bq->bio != NULL) {
423 g_virstor_start(bq->bio);
424 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
425 free(bq, M_GVIRSTOR);
427 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
428 free(bq, M_GVIRSTOR);
432 mtx_unlock(&sc->delayed_bio_q_mtx);
437 * Find a geom handled by the class
439 static struct g_virstor_softc *
440 virstor_find_geom(const struct g_class *cp, const char *name)
444 LIST_FOREACH(gp, &cp->geom, geom) {
445 if (strcmp(name, gp->name) == 0)
452 * Update metadata on all components to reflect the current state
457 * Expects things to be set up so write_metadata() can work, i.e.
458 * the topology lock must be held.
461 update_metadata(struct g_virstor_softc *sc)
463 struct g_virstor_metadata md;
466 if (virstor_valid_components(sc) != sc->n_components)
467 return; /* Incomplete device */
468 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
470 /* Update metadata on components */
471 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
472 sc->geom->class->name, sc->geom->name);
474 for (n = 0; n < sc->n_components; n++) {
475 read_metadata(sc->components[n].gcons, &md);
476 md.chunk_next = sc->components[n].chunk_next;
477 md.flags = sc->components[n].flags;
478 md.md_count = sc->n_components;
479 write_metadata(sc->components[n].gcons, &md);
484 * Fills metadata (struct md) from information stored in softc and the nc'th
485 * component of virstor
488 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
489 u_int nc, u_int hardcode)
491 struct g_virstor_component *c;
493 bzero(md, sizeof *md);
494 c = &sc->components[nc];
496 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
497 md->md_version = G_VIRSTOR_VERSION;
498 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
500 md->md_virsize = sc->virsize;
501 md->md_chunk_size = sc->chunk_size;
502 md->md_count = sc->n_components;
505 strncpy(md->provider, c->gcons->provider->name,
506 sizeof md->provider);
509 md->provsize = c->gcons->provider->mediasize;
510 md->chunk_count = c->chunk_count;
511 md->chunk_next = c->chunk_next;
512 md->chunk_reserved = c->chunk_reserved;
513 md->flags = c->flags;
517 * Remove a component from virstor device.
518 * Can only be done if the component is unallocated.
521 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
523 /* As this is executed in parallel to I/O, operations on virstor
524 * structures must be as atomic as possible. */
525 struct g_virstor_softc *sc;
527 const char *geom_name;
531 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
533 gctl_error(req, "Error fetching argument '%s'", "nargs");
537 gctl_error(req, "Invalid number of arguments");
540 /* Find "our" geom */
541 geom_name = gctl_get_asciiparam(req, "arg0");
542 if (geom_name == NULL) {
543 gctl_error(req, "Error fetching argument '%s'",
547 sc = virstor_find_geom(cp, geom_name);
549 gctl_error(req, "Don't know anything about '%s'", geom_name);
553 if (virstor_valid_components(sc) != sc->n_components) {
554 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
555 "virstor %s", sc->geom->name);
556 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
561 for (i = 1; i < *nargs; i++) {
563 const char *prov_name;
565 struct g_virstor_component *newcomp, *compbak;
567 snprintf(param, sizeof(param), "arg%d", i);
568 prov_name = gctl_get_asciiparam(req, param);
569 if (prov_name == NULL) {
570 gctl_error(req, "Error fetching argument '%s'", param);
573 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
574 prov_name += sizeof(_PATH_DEV) - 1;
577 for (j = 0; j < sc->n_components; j++) {
578 if (strcmp(sc->components[j].gcons->provider->name,
585 LOG_MSG(LVL_ERROR, "No %s component in %s",
586 prov_name, sc->geom->name);
590 compbak = sc->components;
591 newcomp = malloc(sc->n_components * sizeof(*sc->components),
592 M_GVIRSTOR, M_WAITOK | M_ZERO);
593 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
594 bcopy(&sc->components[found + 1], newcomp + found,
595 found * sizeof(*sc->components));
596 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
597 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
599 prov_name, sc->geom->name);
600 free(newcomp, M_GVIRSTOR);
601 /* We'll consider this non-fatal error */
604 /* Renumerate unallocated components */
605 for (j = 0; j < sc->n_components-1; j++) {
606 if ((sc->components[j].flags &
607 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
608 sc->components[j].index = j;
611 /* This is the critical section. If a component allocation
612 * event happens while both variables are not yet set,
613 * there will be trouble. Something will panic on encountering
614 * NULL sc->components[x].gcomp member.
615 * Luckily, component allocation happens very rarely and
616 * removing components is an abnormal action in any case. */
617 sc->components = newcomp;
619 /* End critical section */
622 if (clear_metadata(&compbak[found]) != 0) {
623 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
624 "metadata on %s", prov_name);
626 g_detach(compbak[found].gcons);
627 g_destroy_consumer(compbak[found].gcons);
630 free(compbak, M_GVIRSTOR);
635 /* This call to update_metadata() is critical. In case there's a
636 * power failure in the middle of it and some components are updated
637 * while others are not, there will be trouble on next .taste() iff
638 * a non-updated component is detected first */
642 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
647 * Clear metadata sector on component
650 clear_metadata(struct g_virstor_component *comp)
655 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
656 comp->gcons->provider->name);
658 error = g_access(comp->gcons, 0, 1, 0);
661 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
663 error = g_write_data(comp->gcons,
664 comp->gcons->provider->mediasize -
665 comp->gcons->provider->sectorsize,
667 comp->gcons->provider->sectorsize);
668 free(buf, M_GVIRSTOR);
669 g_access(comp->gcons, 0, -1, 0);
674 * Destroy geom forcibly.
677 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
680 struct g_virstor_softc *sc;
684 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
687 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
698 LOG_MSG(LVL_INFO, "INVARIANTS detected");
699 LOG_MSG(LVL_INFO, "Verifying allocation "
700 "table for %s", sc->geom->name);
702 for (n = 0; n < sc->chunk_count; n++) {
703 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
706 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
707 sc->geom->name, count);
710 if (virstor_valid_components(sc) != sc->n_components) {
711 /* This is a incomplete virstor device (not all
712 * components have been found) */
713 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
717 error = g_access(sc->components[0].gcons, 1, 0, 0);
718 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
720 /* Compare the whole on-disk allocation table with what's
721 * currently in memory */
722 while (n < sc->chunk_count) {
723 buf = g_read_data(sc->components[0].gcons, off,
724 sc->sectorsize, &error);
725 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
726 "for read at %jd", error, off));
727 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
728 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
729 "entry %d, offset %jd", n, off);
733 n += sc->me_per_sector;
734 off += sc->sectorsize;
737 error = g_access(sc->components[0].gcons, -1, 0, 0);
738 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
741 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
742 "(%d sectors don't match, max %zu allocations)",
743 sc->geom->name, count,
744 count * sc->me_per_sector);
746 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
752 virstor_geom_destroy(sc, FALSE, FALSE);
760 * Taste event (per-class callback)
761 * Examines a provider and creates geom instances if needed
763 static struct g_geom *
764 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
766 struct g_virstor_metadata md;
768 struct g_consumer *cp;
769 struct g_virstor_softc *sc;
772 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
774 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
776 /* We need a dummy geom to attach a consumer to the given provider */
777 gp = g_new_geomf(mp, "virstor:taste.helper");
778 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
779 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
780 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
782 cp = g_new_consumer(gp);
783 error = g_attach(cp, pp);
785 error = read_metadata(cp, &md);
788 g_destroy_consumer(cp);
794 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
796 if (md.md_version != G_VIRSTOR_VERSION) {
797 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
798 "to handle %s (%s) : %d should be %d",
799 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
802 if (md.provsize != pp->mediasize)
805 /* If the provider name is hardcoded, use the offered provider only
806 * if it's been offered with its proper name (the one used in
807 * the label command). */
808 if (md.provider[0] != '\0' &&
809 !g_compare_names(md.provider, pp->name))
812 /* Iterate all geoms this class already knows about to see if a new
813 * geom instance of this class needs to be created (in case the provider
814 * is first from a (possibly) multi-consumer geom) or it just needs
815 * to be added to an existing instance. */
818 LIST_FOREACH(gp, &mp->geom, geom) {
822 if (strcmp(md.md_name, sc->geom->name) != 0)
824 if (md.md_id != sc->id)
828 if (gp != NULL) { /* We found an existing geom instance; add to it */
829 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
830 error = add_provider_to_geom(sc, pp, &md);
832 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
833 pp->name, md.md_name, error);
836 } else { /* New geom instance needs to be created */
837 gp = create_virstor_geom(mp, &md);
839 LOG_MSG(LVL_ERROR, "Error creating new instance of "
840 "class %s: %s", mp->name, md.md_name);
841 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
842 md.md_name, pp->name);
846 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
848 error = add_provider_to_geom(sc, pp, &md);
850 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
851 pp->name, md.md_name, error);
852 virstor_geom_destroy(sc, TRUE, FALSE);
861 * Destroyes consumer passed to it in arguments. Used as a callback
865 delay_destroy_consumer(void *arg, int flags __unused)
867 struct g_consumer *c = arg;
868 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
869 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
872 g_destroy_consumer(c);
876 * Remove a component (consumer) from geom instance; If it's the first
877 * component being removed, orphan the provider to announce geom's being
881 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
884 struct g_consumer *c;
886 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
891 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
892 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
894 if (sc->provider != NULL) {
895 LOG_MSG(LVL_INFO, "Removing provider %s", sc->provider->name);
896 g_wither_provider(sc->provider, ENXIO);
900 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
903 /* Destroy consumer after it's tasted */
904 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
907 g_destroy_consumer(c);
912 * Destroy geom - called internally
913 * See g_virstor_destroy_geom for the other one
916 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
919 struct g_provider *pp;
929 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
930 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
931 "Device %s is still open.", pp->name);
936 for (n = 0; n < sc->n_components; n++) {
937 if (sc->components[n].gcons != NULL)
938 remove_component(sc, &sc->components[n], delay);
944 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
947 /* XXX: This might or might not work, since we're called with
948 * the topology lock held. Also, it might panic the kernel if
949 * the error'd BIO is in softupdates code. */
950 mtx_lock(&sc->delayed_bio_q_mtx);
951 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
952 struct g_virstor_bio_q *bq;
953 bq = STAILQ_FIRST(&sc->delayed_bio_q);
954 bq->bio->bio_error = ENOSPC;
955 g_io_deliver(bq->bio, EIO);
956 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
957 free(bq, M_GVIRSTOR);
959 mtx_unlock(&sc->delayed_bio_q_mtx);
960 mtx_destroy(&sc->delayed_bio_q_mtx);
962 free(sc->map, M_GVIRSTOR);
963 free(sc->components, M_GVIRSTOR);
964 bzero(sc, sizeof *sc);
965 free(sc, M_GVIRSTOR);
967 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
968 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
969 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
971 g_wither_geom(gp, ENXIO);
977 * Utility function: read metadata & decode. Wants topology lock to be
981 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
983 struct g_provider *pp;
988 error = g_access(cp, 1, 0, 0);
993 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
996 g_access(cp, -1, 0, 0);
1000 virstor_metadata_decode(buf, md);
1007 * Utility function: encode & write metadata. Assumes topology lock is
1010 * There is no useful way of recovering from errors in this function,
1011 * not involving panicking the kernel. If the metadata cannot be written
1012 * the most we can do is notify the operator and hope he spots it and
1013 * replaces the broken drive.
1016 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1018 struct g_provider *pp;
1022 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1023 ("Something's fishy in %s", __func__));
1024 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1025 g_topology_assert();
1026 error = g_access(cp, 0, 1, 0);
1028 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d",
1029 cp->provider->name, error);
1034 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1035 bzero(buf, pp->sectorsize);
1036 virstor_metadata_encode(md, buf);
1037 g_topology_unlock();
1038 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1041 g_access(cp, 0, -1, 0);
1042 free(buf, M_GVIRSTOR);
1045 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s",
1046 error, cp->provider->name);
1050 * Creates a new instance of this GEOM class, initialise softc
1052 static struct g_geom *
1053 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1056 struct g_virstor_softc *sc;
1058 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1059 md->md_name, md->md_id);
1061 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1062 md->md_virsize < md->md_chunk_size) {
1063 /* This is bogus configuration, and probably means data is
1064 * somehow corrupted. Panic, maybe? */
1065 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1070 /* Check if it's already created */
1071 LIST_FOREACH(gp, &mp->geom, geom) {
1073 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1074 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1076 if (sc->id != md->md_id) {
1078 "Some stale or invalid components "
1079 "exist for virstor device named %s. "
1080 "You will need to <CLEAR> all stale "
1081 "components and maybe reconfigure "
1082 "the virstor device. Tune "
1083 "kern.geom.virstor.debug sysctl up "
1084 "for more information.",
1090 gp = g_new_geomf(mp, "%s", md->md_name);
1091 gp->softc = NULL; /* to circumevent races that test softc */
1093 gp->start = g_virstor_start;
1094 gp->spoiled = g_virstor_orphan;
1095 gp->orphan = g_virstor_orphan;
1096 gp->access = g_virstor_access;
1097 gp->dumpconf = g_virstor_dumpconf;
1099 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1101 sc->n_components = md->md_count;
1102 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1103 M_GVIRSTOR, M_WAITOK | M_ZERO);
1104 sc->chunk_size = md->md_chunk_size;
1105 sc->virsize = md->md_virsize;
1106 STAILQ_INIT(&sc->delayed_bio_q);
1107 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1108 "gvirstor", MTX_DEF | MTX_RECURSE);
1111 sc->provider = NULL; /* virstor_check_and_run will create it */
1114 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1120 * Add provider to a GEOM class instance
1123 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1124 struct g_virstor_metadata *md)
1126 struct g_virstor_component *component;
1127 struct g_consumer *cp, *fcp;
1131 if (md->no >= sc->n_components)
1134 /* "Current" compontent */
1135 component = &(sc->components[md->no]);
1136 if (component->gcons != NULL)
1140 fcp = LIST_FIRST(&gp->consumer);
1142 cp = g_new_consumer(gp);
1143 error = g_attach(cp, pp);
1146 g_destroy_consumer(cp);
1151 if (fcp->provider->sectorsize != pp->sectorsize) {
1152 /* TODO: this can be made to work */
1153 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1154 "sector size (%d)", pp->name, sc->geom->name,
1158 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1159 /* Replicate access permissions from first "live" consumer
1161 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1164 g_destroy_consumer(cp);
1170 /* Bring up a new component */
1171 cp->private = component;
1172 component->gcons = cp;
1174 component->index = md->no;
1175 component->chunk_count = md->chunk_count;
1176 component->chunk_next = md->chunk_next;
1177 component->chunk_reserved = md->chunk_reserved;
1178 component->flags = md->flags;
1180 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1182 virstor_check_and_run(sc);
1187 * Check if everything's ready to create the geom provider & device entry,
1188 * create and start provider.
1189 * Called ultimately by .taste, from g_event thread
1192 virstor_check_and_run(struct g_virstor_softc *sc)
1199 if (virstor_valid_components(sc) != sc->n_components)
1202 if (virstor_valid_components(sc) == 0) {
1203 /* This is actually a candidate for panic() */
1204 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1205 sc->provider->name);
1209 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1211 /* Initialise allocation map from the first consumer */
1212 sc->chunk_count = sc->virsize / sc->chunk_size;
1213 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1214 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1216 sc->chunk_count * (off_t)sc->chunk_size);
1218 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1219 /* The following allocation is in order of 4MB - 8MB */
1220 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1221 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1222 __func__, sc->map_size, sc->provider->name));
1223 sc->map_sectors = sc->map_size / sc->sectorsize;
1226 for (n = 0; n < sc->n_components; n++)
1227 count += sc->components[n].chunk_count;
1228 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1230 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1232 error = g_access(sc->components[0].gcons, 1, 0, 0);
1234 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1235 "read allocation map for %s",
1236 sc->components[0].gcons->provider->name,
1240 /* Read in the allocation map */
1241 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1242 sc->components[0].gcons->provider->name);
1243 off = count = n = 0;
1244 while (count < sc->map_size) {
1245 struct g_virstor_map_entry *mapbuf;
1248 bs = MIN(maxphys, sc->map_size - count);
1249 if (bs % sc->sectorsize != 0) {
1250 /* Check for alignment errors */
1251 bs = rounddown(bs, sc->sectorsize);
1254 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1255 "for %s on %s", sc->geom->name,
1256 sc->components[0].gcons->provider->name);
1258 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1259 if (mapbuf == NULL) {
1260 free(sc->map, M_GVIRSTOR);
1261 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1262 "for %s from %s (offset %ju) (error %d)",
1264 sc->components[0].gcons->provider->name,
1269 bcopy(mapbuf, &sc->map[n], bs);
1272 n += bs / sizeof *(sc->map);
1275 g_access(sc->components[0].gcons, -1, 0, 0);
1276 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1278 /* find first component with allocatable chunks */
1280 for (n = 0; n < sc->n_components; n++) {
1281 if (sc->components[n].chunk_next <
1282 sc->components[n].chunk_count) {
1288 /* not found? set it to the last component and handle it
1290 index = sc->n_components - 1;
1292 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1293 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1294 "(%d/%u: %s)", sc->geom->name,
1297 sc->components[index].gcons->provider->name);
1299 sc->curr_component = index;
1301 if (sc->components[index].chunk_next >=
1302 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1303 LOG_MSG(LVL_WARNING,
1304 "Component %s of %s is running out of free space "
1306 sc->components[index].gcons->provider->name,
1307 sc->geom->name, sc->components[index].chunk_count -
1308 sc->components[index].chunk_next);
1311 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1312 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1314 "%s: Map entries don't fit exactly in a sector (%s)",
1315 __func__, sc->geom->name);
1319 /* Recalculate allocated chunks in components & at the same time
1320 * verify map data is sane. We could trust metadata on this, but
1321 * we want to make sure. */
1322 for (n = 0; n < sc->n_components; n++)
1323 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1325 for (n = 0; n < sc->chunk_count; n++) {
1326 if (sc->map[n].provider_no >= sc->n_components ||
1327 sc->map[n].provider_chunk >=
1328 sc->components[sc->map[n].provider_no].chunk_count) {
1329 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1330 __func__, (u_int)n, sc->geom->name);
1331 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1332 " provider_chunk: %u, chunk_count: %u", __func__,
1333 sc->map[n].provider_no, sc->n_components,
1334 sc->map[n].provider_chunk,
1335 sc->components[sc->map[n].provider_no].chunk_count);
1338 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1339 sc->components[sc->map[n].provider_no].chunk_next++;
1342 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1345 sc->provider->sectorsize = sc->sectorsize;
1346 sc->provider->mediasize = sc->virsize;
1347 g_error_provider(sc->provider, 0);
1349 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1350 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1351 "chunk %u", sc->provider->name, sc->curr_component,
1352 sc->components[sc->curr_component].chunk_next);
1356 * Returns count of active providers in this geom instance
1359 virstor_valid_components(struct g_virstor_softc *sc)
1364 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1365 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1366 for (i = 0; i < sc->n_components; i++)
1367 if (sc->components[i].gcons != NULL)
1373 * Called when the consumer gets orphaned (?)
1376 g_virstor_orphan(struct g_consumer *cp)
1378 struct g_virstor_softc *sc;
1379 struct g_virstor_component *comp;
1382 g_topology_assert();
1389 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1391 remove_component(sc, comp, FALSE);
1392 if (LIST_EMPTY(&gp->consumer))
1393 virstor_geom_destroy(sc, TRUE, FALSE);
1397 * Called to notify geom when it's been opened, and for what intent
1400 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1402 struct g_consumer *c, *c2, *tmp;
1403 struct g_virstor_softc *sc;
1407 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1409 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1412 /* Grab an exclusive bit to propagate on our consumers on first open */
1413 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1415 /* ... drop it on close */
1416 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1419 update_metadata(sc);
1423 LIST_FOREACH_SAFE(c, &gp->consumer, consumer, tmp) {
1424 error = g_access(c, dr, dw, de);
1427 if (c->acr == 0 && c->acw == 0 && c->ace == 0 &&
1428 c->flags & G_CF_ORPHAN) {
1430 g_destroy_consumer(c);
1434 if (sc != NULL && LIST_EMPTY(&gp->consumer))
1435 virstor_geom_destroy(sc, TRUE, FALSE);
1440 /* Backout earlier changes */
1441 LIST_FOREACH(c2, &gp->consumer, consumer) {
1444 g_access(c2, -dr, -dw, -de);
1450 * Generate XML dump of current state
1453 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1454 struct g_consumer *cp, struct g_provider *pp)
1456 struct g_virstor_softc *sc;
1458 g_topology_assert();
1461 if (sc == NULL || pp != NULL)
1465 /* For each component */
1466 struct g_virstor_component *comp;
1471 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1472 indent, comp->index);
1473 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1474 indent, comp->chunk_count);
1475 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1476 indent, comp->chunk_next);
1477 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1478 indent, comp->chunk_reserved);
1479 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1481 comp->chunk_next > 0 ? 100 -
1482 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1483 comp->chunk_count : 100);
1485 /* For the whole thing */
1486 u_int count, used, i;
1489 count = used = size = 0;
1490 for (i = 0; i < sc->n_components; i++) {
1491 if (sc->components[i].gcons != NULL) {
1492 count += sc->components[i].chunk_count;
1493 used += sc->components[i].chunk_next +
1494 sc->components[i].chunk_reserved;
1495 size += sc->components[i].gcons->
1496 provider->mediasize;
1500 sbuf_printf(sb, "%s<Status>"
1501 "Components=%u, Online=%u</Status>\n", indent,
1502 sc->n_components, virstor_valid_components(sc));
1503 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1504 indent, 100-(used * 100) / count);
1505 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1507 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1508 indent, used > 0 ? 100 - (used * 100) / count : 100);
1509 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1511 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1512 indent, sc->chunk_count);
1513 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1515 (count * 100) / sc->chunk_count);
1516 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1518 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1524 * GEOM .done handler
1525 * Can't use standard handler because one requested IO may
1526 * fork into additional data IOs
1529 g_virstor_done(struct bio *b)
1531 struct g_virstor_softc *sc;
1532 struct bio *parent_b;
1534 parent_b = b->bio_parent;
1535 sc = parent_b->bio_to->geom->softc;
1537 if (b->bio_error != 0) {
1538 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1539 b->bio_error, b->bio_offset, b->bio_length,
1541 if (parent_b->bio_error == 0)
1542 parent_b->bio_error = b->bio_error;
1545 parent_b->bio_inbed++;
1546 parent_b->bio_completed += b->bio_completed;
1548 if (parent_b->bio_children == parent_b->bio_inbed) {
1549 parent_b->bio_completed = parent_b->bio_length;
1550 g_io_deliver(parent_b, parent_b->bio_error);
1557 * Called in g_down thread
1560 g_virstor_start(struct bio *b)
1562 struct g_virstor_softc *sc;
1563 struct g_virstor_component *comp;
1565 struct g_provider *pp;
1567 off_t offset, length;
1568 struct bio_queue_head bq;
1569 size_t chunk_size; /* cached for convenience */
1573 sc = pp->geom->softc;
1574 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1575 b->bio_to->error, b->bio_to->name));
1577 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1579 switch (b->bio_cmd) {
1585 g_io_deliver(b, EOPNOTSUPP);
1589 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1592 chunk_size = sc->chunk_size;
1594 offset = b->bio_offset; /* virtual offset and length */
1595 length = b->bio_length;
1597 while (length > 0) {
1598 size_t chunk_index, in_chunk_offset, in_chunk_length;
1599 struct virstor_map_entry *me;
1601 chunk_index = offset / chunk_size; /* round downwards */
1602 in_chunk_offset = offset % chunk_size;
1603 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1604 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1605 b->bio_cmd == BIO_READ ? "R" : "W",
1607 chunk_index, in_chunk_offset, in_chunk_length);
1608 me = &sc->map[chunk_index];
1610 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1611 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1612 /* Reads from unallocated chunks return zeroed
1614 if (b->bio_cmd == BIO_READ)
1615 bzero(addr, in_chunk_length);
1617 comp = &sc->components[me->provider_no];
1619 cb = g_clone_bio(b);
1621 bioq_dismantle(&bq);
1622 if (b->bio_error == 0)
1623 b->bio_error = ENOMEM;
1624 g_io_deliver(b, b->bio_error);
1627 cb->bio_to = comp->gcons->provider;
1628 cb->bio_done = g_virstor_done;
1630 (off_t)me->provider_chunk * (off_t)chunk_size
1632 cb->bio_length = in_chunk_length;
1633 cb->bio_data = addr;
1634 cb->bio_caller1 = comp;
1635 bioq_disksort(&bq, cb);
1637 } else { /* handle BIO_WRITE */
1638 KASSERT(b->bio_cmd == BIO_WRITE,
1639 ("%s: Unknown command %d", __func__,
1642 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1643 /* We have a virtual chunk, represented by
1644 * the "me" entry, but it's not yet allocated
1645 * (tied to) a physical chunk. So do it now. */
1646 struct virstor_map_entry *data_me;
1647 u_int phys_chunk, comp_no;
1651 error = allocate_chunk(sc, &comp, &comp_no,
1654 /* We cannot allocate a physical chunk
1655 * to satisfy this request, so we'll
1656 * delay it to when we can...
1657 * XXX: this will prevent the fs from
1658 * being umounted! */
1659 struct g_virstor_bio_q *biq;
1660 biq = malloc(sizeof *biq, M_GVIRSTOR,
1663 bioq_dismantle(&bq);
1664 if (b->bio_error == 0)
1665 b->bio_error = ENOMEM;
1666 g_io_deliver(b, b->bio_error);
1670 mtx_lock(&sc->delayed_bio_q_mtx);
1671 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1673 mtx_unlock(&sc->delayed_bio_q_mtx);
1674 LOG_MSG(LVL_WARNING, "Delaying BIO "
1675 "(size=%ju) until free physical "
1676 "space can be found on %s",
1678 sc->provider->name);
1681 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1684 comp->gcons->provider->name,
1685 sc->provider->name);
1687 me->provider_no = comp_no;
1688 me->provider_chunk = phys_chunk;
1689 me->flags |= VIRSTOR_MAP_ALLOCATED;
1691 cb = g_clone_bio(b);
1693 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1694 me->provider_no = 0;
1695 me->provider_chunk = 0;
1696 bioq_dismantle(&bq);
1697 if (b->bio_error == 0)
1698 b->bio_error = ENOMEM;
1699 g_io_deliver(b, b->bio_error);
1703 /* The allocation table is stored continuously
1704 * at the start of the drive. We need to
1705 * calculate the offset of the sector that holds
1706 * this map entry both on the drive and in the
1708 * sc_offset will end up pointing to the drive
1710 s_offset = chunk_index * sizeof *me;
1711 s_offset = rounddown(s_offset, sc->sectorsize);
1713 /* data_me points to map entry sector
1714 * in memory (analogous to offset) */
1715 data_me = &sc->map[rounddown(chunk_index,
1716 sc->me_per_sector)];
1718 /* Commit sector with map entry to storage */
1719 cb->bio_to = sc->components[0].gcons->provider;
1720 cb->bio_done = g_virstor_done;
1721 cb->bio_offset = s_offset;
1722 cb->bio_data = (char *)data_me;
1723 cb->bio_length = sc->sectorsize;
1724 cb->bio_caller1 = &sc->components[0];
1725 bioq_disksort(&bq, cb);
1728 comp = &sc->components[me->provider_no];
1729 cb = g_clone_bio(b);
1731 bioq_dismantle(&bq);
1732 if (b->bio_error == 0)
1733 b->bio_error = ENOMEM;
1734 g_io_deliver(b, b->bio_error);
1737 /* Finally, handle the data */
1738 cb->bio_to = comp->gcons->provider;
1739 cb->bio_done = g_virstor_done;
1740 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1742 cb->bio_length = in_chunk_length;
1743 cb->bio_data = addr;
1744 cb->bio_caller1 = comp;
1745 bioq_disksort(&bq, cb);
1747 addr += in_chunk_length;
1748 length -= in_chunk_length;
1749 offset += in_chunk_length;
1752 /* Fire off bio's here */
1754 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1755 bioq_remove(&bq, cb);
1756 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1757 comp = cb->bio_caller1;
1758 cb->bio_caller1 = NULL;
1759 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1760 cb->bio_offset, cb->bio_length);
1761 g_io_request(cb, comp->gcons);
1764 if (count == 0) { /* We handled everything locally */
1765 b->bio_completed = b->bio_length;
1772 * Allocate a chunk from a physical provider. Returns physical component,
1773 * chunk index relative to the component and the component's index.
1776 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1777 u_int *comp_no_p, u_int *chunk)
1781 KASSERT(sc->curr_component < sc->n_components,
1782 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1784 comp_no = sc->curr_component;
1785 *comp = &sc->components[comp_no];
1786 dump_component(*comp);
1787 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1788 /* This component is full. Allocate next component */
1789 if (comp_no >= sc->n_components-1) {
1790 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1794 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1795 sc->curr_component = ++comp_no;
1797 *comp = &sc->components[comp_no];
1798 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1799 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1800 "(switching to %u/%u: %s)", sc->geom->name,
1801 comp_no+1, sc->n_components,
1802 (*comp)->gcons->provider->name);
1803 /* Take care not to overwrite reserved chunks */
1804 if ( (*comp)->chunk_reserved > 0 &&
1805 (*comp)->chunk_next < (*comp)->chunk_reserved)
1806 (*comp)->chunk_next = (*comp)->chunk_reserved;
1809 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1810 dump_component(*comp);
1811 *comp_no_p = comp_no;
1812 *chunk = (*comp)->chunk_next++;
1814 *comp_no_p = comp_no;
1815 *chunk = (*comp)->chunk_next++;
1820 /* Dump a component */
1822 dump_component(struct g_virstor_component *comp)
1825 if (g_virstor_debug < LVL_DEBUG2)
1827 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1828 printf(" chunk_count: %u\n", comp->chunk_count);
1829 printf(" chunk_next: %u\n", comp->chunk_next);
1830 printf(" flags: %u\n", comp->flags);
1834 /* Dump a map entry */
1836 dump_me(struct virstor_map_entry *me, unsigned int nr)
1838 if (g_virstor_debug < LVL_DEBUG)
1840 printf("VIRT. CHUNK #%d: ", nr);
1841 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1842 printf("(unallocated)\n");
1844 printf("allocated at provider %u, provider_chunk %u\n",
1845 me->provider_no, me->provider_chunk);
1850 * Dismantle bio_queue and destroy its components
1853 bioq_dismantle(struct bio_queue_head *bq)
1857 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1864 * The function that shouldn't be called.
1865 * When this is called, the stack is already garbled because of
1866 * argument mismatch. There's nothing to do now but panic, which is
1867 * accidentally the whole purpose of this function.
1868 * Motivation: to guard from accidentally calling geom methods when
1869 * they shouldn't be called. (see g_..._taste)
1874 panic("invalid_call() has just been called. Something's fishy here.");
1877 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */
1878 MODULE_VERSION(geom_virstor, 0);