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 static SYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0,
84 "GEOM_GVIRSTOR information");
86 static u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */
87 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RWTUN, &g_virstor_debug,
88 0, "Debug level (2=production, 5=normal, 15=excessive)");
90 static u_int g_virstor_chunk_watermark = 100;
91 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RWTUN,
92 &g_virstor_chunk_watermark, 0,
93 "Minimum number of free chunks before issuing administrative warning");
95 static u_int g_virstor_component_watermark = 1;
96 SYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RWTUN,
97 &g_virstor_component_watermark, 0,
98 "Minimum number of free components before issuing administrative warning");
100 static int read_metadata(struct g_consumer *, struct g_virstor_metadata *);
101 static void write_metadata(struct g_consumer *, struct g_virstor_metadata *);
102 static int clear_metadata(struct g_virstor_component *);
103 static int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *,
104 struct g_virstor_metadata *);
105 static struct g_geom *create_virstor_geom(struct g_class *,
106 struct g_virstor_metadata *);
107 static void virstor_check_and_run(struct g_virstor_softc *);
108 static u_int virstor_valid_components(struct g_virstor_softc *);
109 static int virstor_geom_destroy(struct g_virstor_softc *, boolean_t,
111 static void remove_component(struct g_virstor_softc *,
112 struct g_virstor_component *, boolean_t);
113 static void bioq_dismantle(struct bio_queue_head *);
114 static int allocate_chunk(struct g_virstor_softc *,
115 struct g_virstor_component **, u_int *, u_int *);
116 static void delay_destroy_consumer(void *, int);
117 static void dump_component(struct g_virstor_component *comp);
119 static void dump_me(struct virstor_map_entry *me, unsigned int nr);
122 static void virstor_ctl_stop(struct gctl_req *, struct g_class *);
123 static void virstor_ctl_add(struct gctl_req *, struct g_class *);
124 static void virstor_ctl_remove(struct gctl_req *, struct g_class *);
125 static struct g_virstor_softc * virstor_find_geom(const struct g_class *,
127 static void update_metadata(struct g_virstor_softc *);
128 static void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *,
131 static void g_virstor_orphan(struct g_consumer *);
132 static int g_virstor_access(struct g_provider *, int, int, int);
133 static void g_virstor_start(struct bio *);
134 static void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *,
135 struct g_consumer *, struct g_provider *);
136 static void g_virstor_done(struct bio *);
138 static void invalid_call(void);
140 * Initialise GEOM class (per-class callback)
143 g_virstor_init(struct g_class *mp __unused)
146 /* Catch map struct size mismatch at compile time; Map entries must
147 * fit into MAXPHYS exactly, with no wasted space. */
148 CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS);
150 /* Init UMA zones, TAILQ's, other global vars */
154 * Finalise GEOM class (per-class callback)
157 g_virstor_fini(struct g_class *mp __unused)
160 /* Deinit UMA zones & global vars */
164 * Config (per-class callback)
167 g_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb)
173 version = gctl_get_paraml(req, "version", sizeof(*version));
174 if (version == NULL) {
175 gctl_error(req, "Failed to get 'version' argument");
178 if (*version != G_VIRSTOR_VERSION) {
179 gctl_error(req, "Userland and kernel versions out of sync");
184 if (strcmp(verb, "add") == 0)
185 virstor_ctl_add(req, cp);
186 else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0)
187 virstor_ctl_stop(req, cp);
188 else if (strcmp(verb, "remove") == 0)
189 virstor_ctl_remove(req, cp);
191 gctl_error(req, "unknown verb: '%s'", verb);
196 * "stop" verb from userland
199 virstor_ctl_stop(struct gctl_req *req, struct g_class *cp)
204 nargs = gctl_get_paraml(req, "nargs", sizeof *nargs);
206 gctl_error(req, "Error fetching argument '%s'", "nargs");
210 gctl_error(req, "Invalid number of arguments");
213 force = gctl_get_paraml(req, "force", sizeof *force);
215 gctl_error(req, "Error fetching argument '%s'", "force");
220 for (i = 0; i < *nargs; i++) {
223 struct g_virstor_softc *sc;
226 sprintf(param, "arg%d", i);
227 name = gctl_get_asciiparam(req, param);
229 gctl_error(req, "No 'arg%d' argument", i);
233 sc = virstor_find_geom(cp, name);
235 gctl_error(req, "Don't know anything about '%s'", name);
240 LOG_MSG(LVL_INFO, "Stopping %s by the userland command",
243 if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) {
244 LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d",
245 sc->geom->name, error);
252 * "add" verb from userland - add new component(s) to the structure.
253 * This will be done all at once in here, without going through the
254 * .taste function for new components.
257 virstor_ctl_add(struct gctl_req *req, struct g_class *cp)
259 /* Note: while this is going on, I/O is being done on
260 * the g_up and g_down threads. The idea is to make changes
261 * to softc members in a way that can atomically activate
262 * them all at once. */
263 struct g_virstor_softc *sc;
264 int *hardcode, *nargs;
265 const char *geom_name; /* geom to add a component to */
266 struct g_consumer *fcp;
267 struct g_virstor_bio_q *bq;
272 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
274 gctl_error(req, "Error fetching argument '%s'", "nargs");
278 gctl_error(req, "Invalid number of arguments");
281 hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode));
282 if (hardcode == NULL) {
283 gctl_error(req, "Error fetching argument '%s'", "hardcode");
287 /* Find "our" geom */
288 geom_name = gctl_get_asciiparam(req, "arg0");
289 if (geom_name == NULL) {
290 gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)");
293 sc = virstor_find_geom(cp, geom_name);
295 gctl_error(req, "Don't know anything about '%s'", geom_name);
299 if (virstor_valid_components(sc) != sc->n_components) {
300 LOG_MSG(LVL_ERROR, "Cannot add components to incomplete "
301 "virstor %s", sc->geom->name);
302 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
306 fcp = sc->components[0].gcons;
309 for (i = 1; i < *nargs; i++) {
310 struct g_virstor_metadata md;
312 const char *prov_name;
313 struct g_provider *pp;
314 struct g_consumer *cp;
318 snprintf(aname, sizeof aname, "arg%d", i);
319 prov_name = gctl_get_asciiparam(req, aname);
320 if (prov_name == NULL) {
321 gctl_error(req, "Error fetching argument '%s'", aname);
325 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
326 prov_name += sizeof(_PATH_DEV) - 1;
328 pp = g_provider_by_name(prov_name);
330 /* This is the most common error so be verbose about it */
332 gctl_error(req, "Invalid provider: '%s' (added"
333 " %u components)", prov_name, added);
336 gctl_error(req, "Invalid provider: '%s'",
342 cp = g_new_consumer(sc->geom);
344 gctl_error(req, "Cannot create consumer");
348 error = g_attach(cp, pp);
350 gctl_error(req, "Cannot attach a consumer to %s",
352 g_destroy_consumer(cp);
356 if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) {
357 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
359 gctl_error(req, "Access request failed for %s",
361 g_destroy_consumer(cp);
366 if (fcp->provider->sectorsize != pp->sectorsize) {
367 gctl_error(req, "Sector size doesn't fit for %s",
369 g_destroy_consumer(cp);
373 for (j = 0; j < sc->n_components; j++) {
374 if (strcmp(sc->components[j].gcons->provider->name,
376 gctl_error(req, "Component %s already in %s",
377 pp->name, sc->geom->name);
378 g_destroy_consumer(cp);
383 sc->components = realloc(sc->components,
384 sizeof(*sc->components) * (sc->n_components + 1),
385 M_GVIRSTOR, M_WAITOK);
387 nc = sc->n_components;
388 sc->components[nc].gcons = cp;
389 sc->components[nc].sc = sc;
390 sc->components[nc].index = nc;
391 sc->components[nc].chunk_count = cp->provider->mediasize /
393 sc->components[nc].chunk_next = 0;
394 sc->components[nc].chunk_reserved = 0;
396 if (sc->components[nc].chunk_count < 4) {
397 gctl_error(req, "Provider too small: %s",
399 g_destroy_consumer(cp);
403 fill_metadata(sc, &md, nc, *hardcode);
404 write_metadata(cp, &md);
405 /* The new component becomes visible when n_components is
411 /* This call to update_metadata() is critical. In case there's a
412 * power failure in the middle of it and some components are updated
413 * while others are not, there will be trouble on next .taste() iff
414 * a non-updated component is detected first */
417 LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added,
419 /* Fire off BIOs previously queued because there wasn't any
420 * physical space left. If the BIOs still can't be satisfied
421 * they will again be added to the end of the queue (during
422 * which the mutex will be recursed) */
423 bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK);
425 mtx_lock(&sc->delayed_bio_q_mtx);
426 /* First, insert a sentinel to the queue end, so we don't
427 * end up in an infinite loop if there's still no free
428 * space available. */
429 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage);
430 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
431 bq = STAILQ_FIRST(&sc->delayed_bio_q);
432 if (bq->bio != NULL) {
433 g_virstor_start(bq->bio);
434 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
435 free(bq, M_GVIRSTOR);
437 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
438 free(bq, M_GVIRSTOR);
442 mtx_unlock(&sc->delayed_bio_q_mtx);
447 * Find a geom handled by the class
449 static struct g_virstor_softc *
450 virstor_find_geom(const struct g_class *cp, const char *name)
454 LIST_FOREACH(gp, &cp->geom, geom) {
455 if (strcmp(name, gp->name) == 0)
462 * Update metadata on all components to reflect the current state
467 * Expects things to be set up so write_metadata() can work, i.e.
468 * the topology lock must be held.
471 update_metadata(struct g_virstor_softc *sc)
473 struct g_virstor_metadata md;
476 if (virstor_valid_components(sc) != sc->n_components)
477 return; /* Incomplete device */
478 LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s",
480 /* Update metadata on components */
481 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__,
482 sc->geom->class->name, sc->geom->name);
484 for (n = 0; n < sc->n_components; n++) {
485 read_metadata(sc->components[n].gcons, &md);
486 md.chunk_next = sc->components[n].chunk_next;
487 md.flags = sc->components[n].flags;
488 md.md_count = sc->n_components;
489 write_metadata(sc->components[n].gcons, &md);
494 * Fills metadata (struct md) from information stored in softc and the nc'th
495 * component of virstor
498 fill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md,
499 u_int nc, u_int hardcode)
501 struct g_virstor_component *c;
503 bzero(md, sizeof *md);
504 c = &sc->components[nc];
506 strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic);
507 md->md_version = G_VIRSTOR_VERSION;
508 strncpy(md->md_name, sc->geom->name, sizeof md->md_name);
510 md->md_virsize = sc->virsize;
511 md->md_chunk_size = sc->chunk_size;
512 md->md_count = sc->n_components;
515 strncpy(md->provider, c->gcons->provider->name,
516 sizeof md->provider);
519 md->provsize = c->gcons->provider->mediasize;
520 md->chunk_count = c->chunk_count;
521 md->chunk_next = c->chunk_next;
522 md->chunk_reserved = c->chunk_reserved;
523 md->flags = c->flags;
527 * Remove a component from virstor device.
528 * Can only be done if the component is unallocated.
531 virstor_ctl_remove(struct gctl_req *req, struct g_class *cp)
533 /* As this is executed in parallel to I/O, operations on virstor
534 * structures must be as atomic as possible. */
535 struct g_virstor_softc *sc;
537 const char *geom_name;
541 nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs));
543 gctl_error(req, "Error fetching argument '%s'", "nargs");
547 gctl_error(req, "Invalid number of arguments");
550 /* Find "our" geom */
551 geom_name = gctl_get_asciiparam(req, "arg0");
552 if (geom_name == NULL) {
553 gctl_error(req, "Error fetching argument '%s'",
557 sc = virstor_find_geom(cp, geom_name);
559 gctl_error(req, "Don't know anything about '%s'", geom_name);
563 if (virstor_valid_components(sc) != sc->n_components) {
564 LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete "
565 "virstor %s", sc->geom->name);
566 gctl_error(req, "Virstor %s is incomplete", sc->geom->name);
571 for (i = 1; i < *nargs; i++) {
573 const char *prov_name;
575 struct g_virstor_component *newcomp, *compbak;
577 sprintf(param, "arg%d", i);
578 prov_name = gctl_get_asciiparam(req, param);
579 if (prov_name == NULL) {
580 gctl_error(req, "Error fetching argument '%s'", param);
583 if (strncmp(prov_name, _PATH_DEV, sizeof(_PATH_DEV) - 1) == 0)
584 prov_name += sizeof(_PATH_DEV) - 1;
587 for (j = 0; j < sc->n_components; j++) {
588 if (strcmp(sc->components[j].gcons->provider->name,
595 LOG_MSG(LVL_ERROR, "No %s component in %s",
596 prov_name, sc->geom->name);
600 compbak = sc->components;
601 newcomp = malloc(sc->n_components * sizeof(*sc->components),
602 M_GVIRSTOR, M_WAITOK | M_ZERO);
603 bcopy(sc->components, newcomp, found * sizeof(*sc->components));
604 bcopy(&sc->components[found + 1], newcomp + found,
605 found * sizeof(*sc->components));
606 if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) {
607 LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be "
609 prov_name, sc->geom->name);
610 free(newcomp, M_GVIRSTOR);
611 /* We'll consider this non-fatal error */
614 /* Renumerate unallocated components */
615 for (j = 0; j < sc->n_components-1; j++) {
616 if ((sc->components[j].flags &
617 VIRSTOR_PROVIDER_ALLOCATED) == 0) {
618 sc->components[j].index = j;
621 /* This is the critical section. If a component allocation
622 * event happens while both variables are not yet set,
623 * there will be trouble. Something will panic on encountering
624 * NULL sc->components[x].gcomp member.
625 * Luckily, component allocation happens very rarely and
626 * removing components is an abnormal action in any case. */
627 sc->components = newcomp;
629 /* End critical section */
632 if (clear_metadata(&compbak[found]) != 0) {
633 LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear "
634 "metadata on %s", prov_name);
636 g_detach(compbak[found].gcons);
637 g_destroy_consumer(compbak[found].gcons);
640 free(compbak, M_GVIRSTOR);
645 /* This call to update_metadata() is critical. In case there's a
646 * power failure in the middle of it and some components are updated
647 * while others are not, there will be trouble on next .taste() iff
648 * a non-updated component is detected first */
652 LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed,
657 * Clear metadata sector on component
660 clear_metadata(struct g_virstor_component *comp)
665 LOG_MSG(LVL_INFO, "Clearing metadata on %s",
666 comp->gcons->provider->name);
668 error = g_access(comp->gcons, 0, 1, 0);
671 buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR,
673 error = g_write_data(comp->gcons,
674 comp->gcons->provider->mediasize -
675 comp->gcons->provider->sectorsize,
677 comp->gcons->provider->sectorsize);
678 free(buf, M_GVIRSTOR);
679 g_access(comp->gcons, 0, -1, 0);
684 * Destroy geom forcibly.
687 g_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp,
690 struct g_virstor_softc *sc;
694 KASSERT(sc != NULL, ("%s: NULL sc", __func__));
697 LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name,
708 LOG_MSG(LVL_INFO, "INVARIANTS detected");
709 LOG_MSG(LVL_INFO, "Verifying allocation "
710 "table for %s", sc->geom->name);
712 for (n = 0; n < sc->chunk_count; n++) {
713 if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0)
716 LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks",
717 sc->geom->name, count);
720 if (virstor_valid_components(sc) != sc->n_components) {
721 /* This is a incomplete virstor device (not all
722 * components have been found) */
723 LOG_MSG(LVL_ERROR, "Device %s is incomplete",
727 error = g_access(sc->components[0].gcons, 1, 0, 0);
728 KASSERT(error == 0, ("%s: g_access failed (%d)", __func__,
730 /* Compare the whole on-disk allocation table with what's
731 * currently in memory */
732 while (n < sc->chunk_count) {
733 buf = g_read_data(sc->components[0].gcons, off,
734 sc->sectorsize, &error);
735 KASSERT(buf != NULL, ("g_read_data returned NULL (%d) "
736 "for read at %jd", error, off));
737 if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) {
738 LOG_MSG(LVL_ERROR, "ERROR in allocation table, "
739 "entry %d, offset %jd", n, off);
743 n += sc->me_per_sector;
744 off += sc->sectorsize;
747 error = g_access(sc->components[0].gcons, -1, 0, 0);
748 KASSERT(error == 0, ("%s: g_access failed (%d) on exit",
751 LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s "
752 "(%d sectors don't match, max %zu allocations)",
753 sc->geom->name, count,
754 count * sc->me_per_sector);
756 LOG_MSG(LVL_INFO, "Allocation table ok for %s",
762 virstor_geom_destroy(sc, FALSE, FALSE);
770 * Taste event (per-class callback)
771 * Examines a provider and creates geom instances if needed
773 static struct g_geom *
774 g_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags)
776 struct g_virstor_metadata md;
778 struct g_consumer *cp;
779 struct g_virstor_softc *sc;
782 g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name);
784 LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name);
786 /* We need a dummy geom to attach a consumer to the given provider */
787 gp = g_new_geomf(mp, "virstor:taste.helper");
788 gp->start = (void *)invalid_call; /* XXX: hacked up so the */
789 gp->access = (void *)invalid_call; /* compiler doesn't complain. */
790 gp->orphan = (void *)invalid_call; /* I really want these to fail. */
792 cp = g_new_consumer(gp);
794 error = read_metadata(cp, &md);
796 g_destroy_consumer(cp);
802 if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0)
804 if (md.md_version != G_VIRSTOR_VERSION) {
805 LOG_MSG(LVL_ERROR, "Kernel module version invalid "
806 "to handle %s (%s) : %d should be %d",
807 md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION);
810 if (md.provsize != pp->mediasize)
813 /* If the provider name is hardcoded, use the offered provider only
814 * if it's been offered with its proper name (the one used in
815 * the label command). */
816 if (md.provider[0] != '\0' &&
817 !g_compare_names(md.provider, pp->name))
820 /* Iterate all geoms this class already knows about to see if a new
821 * geom instance of this class needs to be created (in case the provider
822 * is first from a (possibly) multi-consumer geom) or it just needs
823 * to be added to an existing instance. */
826 LIST_FOREACH(gp, &mp->geom, geom) {
830 if (strcmp(md.md_name, sc->geom->name) != 0)
832 if (md.md_id != sc->id)
836 if (gp != NULL) { /* We found an existing geom instance; add to it */
837 LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name);
838 error = add_provider_to_geom(sc, pp, &md);
840 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
841 pp->name, md.md_name, error);
844 } else { /* New geom instance needs to be created */
845 gp = create_virstor_geom(mp, &md);
847 LOG_MSG(LVL_ERROR, "Error creating new instance of "
848 "class %s: %s", mp->name, md.md_name);
849 LOG_MSG(LVL_DEBUG, "Error creating %s at %s",
850 md.md_name, pp->name);
854 LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name,
856 error = add_provider_to_geom(sc, pp, &md);
858 LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)",
859 pp->name, md.md_name, error);
860 virstor_geom_destroy(sc, TRUE, FALSE);
869 * Destroyes consumer passed to it in arguments. Used as a callback
873 delay_destroy_consumer(void *arg, int flags __unused)
875 struct g_consumer *c = arg;
876 KASSERT(c != NULL, ("%s: invalid consumer", __func__));
877 LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay",
880 g_destroy_consumer(c);
884 * Remove a component (consumer) from geom instance; If it's the first
885 * component being removed, orphan the provider to announce geom's being
889 remove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp,
892 struct g_consumer *c;
894 KASSERT(comp->gcons != NULL, ("Component with no consumer in %s",
899 KASSERT(c->provider != NULL, ("%s: no provider", __func__));
900 LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name,
902 if (sc->provider != NULL) {
904 sc->provider->flags |= G_PF_WITHER;
905 g_orphan_provider(sc->provider, ENXIO);
907 LOG_MSG(LVL_INFO, "Removing provider %s", sc->geom->name);
910 if (c->acr > 0 || c->acw > 0 || c->ace > 0)
911 g_access(c, -c->acr, -c->acw, -c->ace);
913 /* Destroy consumer after it's tasted */
914 g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL);
917 g_destroy_consumer(c);
922 * Destroy geom - called internally
923 * See g_virstor_destroy_geom for the other one
926 virstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force,
929 struct g_provider *pp;
939 if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) {
940 LOG_MSG(force ? LVL_WARNING : LVL_ERROR,
941 "Device %s is still open.", pp->name);
946 for (n = 0; n < sc->n_components; n++) {
947 if (sc->components[n].gcons != NULL)
948 remove_component(sc, &sc->components[n], delay);
954 KASSERT(sc->provider == NULL, ("Provider still exists for %s",
957 /* XXX: This might or might not work, since we're called with
958 * the topology lock held. Also, it might panic the kernel if
959 * the error'd BIO is in softupdates code. */
960 mtx_lock(&sc->delayed_bio_q_mtx);
961 while (!STAILQ_EMPTY(&sc->delayed_bio_q)) {
962 struct g_virstor_bio_q *bq;
963 bq = STAILQ_FIRST(&sc->delayed_bio_q);
964 bq->bio->bio_error = ENOSPC;
965 g_io_deliver(bq->bio, EIO);
966 STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage);
967 free(bq, M_GVIRSTOR);
969 mtx_unlock(&sc->delayed_bio_q_mtx);
970 mtx_destroy(&sc->delayed_bio_q_mtx);
972 free(sc->map, M_GVIRSTOR);
973 free(sc->components, M_GVIRSTOR);
974 bzero(sc, sizeof *sc);
975 free(sc, M_GVIRSTOR);
977 pp = LIST_FIRST(&gp->provider); /* We only offer one provider */
978 if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0))
979 LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name);
981 g_wither_geom(gp, ENXIO);
987 * Utility function: read metadata & decode. Wants topology lock to be
991 read_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
993 struct g_provider *pp;
998 error = g_access(cp, 1, 0, 0);
1002 g_topology_unlock();
1003 buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize,
1006 g_access(cp, -1, 0, 0);
1010 virstor_metadata_decode(buf, md);
1017 * Utility function: encode & write metadata. Assumes topology lock is
1020 * There is no useful way of recovering from errors in this function,
1021 * not involving panicking the kernel. If the metadata cannot be written
1022 * the most we can do is notify the operator and hope he spots it and
1023 * replaces the broken drive.
1026 write_metadata(struct g_consumer *cp, struct g_virstor_metadata *md)
1028 struct g_provider *pp;
1032 KASSERT(cp != NULL && md != NULL && cp->provider != NULL,
1033 ("Something's fishy in %s", __func__));
1034 LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name);
1035 g_topology_assert();
1036 error = g_access(cp, 0, 1, 0);
1038 LOG_MSG(LVL_ERROR, "g_access(0,1,0) failed for %s: %d",
1039 cp->provider->name, error);
1044 buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK);
1045 virstor_metadata_encode(md, buf);
1046 g_topology_unlock();
1047 error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf,
1050 g_access(cp, 0, -1, 0);
1051 free(buf, M_GVIRSTOR);
1054 LOG_MSG(LVL_ERROR, "Error %d writing metadata to %s",
1055 error, cp->provider->name);
1059 * Creates a new instance of this GEOM class, initialise softc
1061 static struct g_geom *
1062 create_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md)
1065 struct g_virstor_softc *sc;
1067 LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)",
1068 md->md_name, md->md_id);
1070 if (md->md_count < 1 || md->md_chunk_size < 1 ||
1071 md->md_virsize < md->md_chunk_size) {
1072 /* This is bogus configuration, and probably means data is
1073 * somehow corrupted. Panic, maybe? */
1074 LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s",
1079 /* Check if it's already created */
1080 LIST_FOREACH(gp, &mp->geom, geom) {
1082 if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) {
1083 LOG_MSG(LVL_WARNING, "Geom %s already exists",
1085 if (sc->id != md->md_id) {
1087 "Some stale or invalid components "
1088 "exist for virstor device named %s. "
1089 "You will need to <CLEAR> all stale "
1090 "components and maybe reconfigure "
1091 "the virstor device. Tune "
1092 "kern.geom.virstor.debug sysctl up "
1093 "for more information.",
1099 gp = g_new_geomf(mp, "%s", md->md_name);
1100 gp->softc = NULL; /* to circumevent races that test softc */
1102 gp->start = g_virstor_start;
1103 gp->spoiled = g_virstor_orphan;
1104 gp->orphan = g_virstor_orphan;
1105 gp->access = g_virstor_access;
1106 gp->dumpconf = g_virstor_dumpconf;
1108 sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO);
1110 sc->n_components = md->md_count;
1111 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count,
1112 M_GVIRSTOR, M_WAITOK | M_ZERO);
1113 sc->chunk_size = md->md_chunk_size;
1114 sc->virsize = md->md_virsize;
1115 STAILQ_INIT(&sc->delayed_bio_q);
1116 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx",
1117 "gvirstor", MTX_DEF | MTX_RECURSE);
1120 sc->provider = NULL; /* virstor_check_and_run will create it */
1123 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name);
1129 * Add provider to a GEOM class instance
1132 add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp,
1133 struct g_virstor_metadata *md)
1135 struct g_virstor_component *component;
1136 struct g_consumer *cp, *fcp;
1140 if (md->no >= sc->n_components)
1143 /* "Current" compontent */
1144 component = &(sc->components[md->no]);
1145 if (component->gcons != NULL)
1149 fcp = LIST_FIRST(&gp->consumer);
1151 cp = g_new_consumer(gp);
1152 error = g_attach(cp, pp);
1155 g_destroy_consumer(cp);
1160 if (fcp->provider->sectorsize != pp->sectorsize) {
1161 /* TODO: this can be made to work */
1162 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid "
1163 "sector size (%d)", pp->name, sc->geom->name,
1167 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) {
1168 /* Replicate access permissions from first "live" consumer
1170 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace);
1173 g_destroy_consumer(cp);
1179 /* Bring up a new component */
1180 cp->private = component;
1181 component->gcons = cp;
1183 component->index = md->no;
1184 component->chunk_count = md->chunk_count;
1185 component->chunk_next = md->chunk_next;
1186 component->chunk_reserved = md->chunk_reserved;
1187 component->flags = md->flags;
1189 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name);
1191 virstor_check_and_run(sc);
1196 * Check if everything's ready to create the geom provider & device entry,
1197 * create and start provider.
1198 * Called ultimately by .taste, from g_event thread
1201 virstor_check_and_run(struct g_virstor_softc *sc)
1208 if (virstor_valid_components(sc) != sc->n_components)
1211 if (virstor_valid_components(sc) == 0) {
1212 /* This is actually a candidate for panic() */
1213 LOG_MSG(LVL_ERROR, "No valid components for %s?",
1214 sc->provider->name);
1218 sc->sectorsize = sc->components[0].gcons->provider->sectorsize;
1220 /* Initialise allocation map from the first consumer */
1221 sc->chunk_count = sc->virsize / sc->chunk_size;
1222 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) {
1223 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes",
1225 sc->chunk_count * (off_t)sc->chunk_size);
1227 sc->map_size = sc->chunk_count * sizeof *(sc->map);
1228 /* The following allocation is in order of 4MB - 8MB */
1229 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK);
1230 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s",
1231 __func__, sc->map_size, sc->provider->name));
1232 sc->map_sectors = sc->map_size / sc->sectorsize;
1235 for (n = 0; n < sc->n_components; n++)
1236 count += sc->components[n].chunk_count;
1237 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual "
1239 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024);
1241 error = g_access(sc->components[0].gcons, 1, 0, 0);
1243 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to "
1244 "read allocation map for %s",
1245 sc->components[0].gcons->provider->name,
1249 /* Read in the allocation map */
1250 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name,
1251 sc->components[0].gcons->provider->name);
1252 off = count = n = 0;
1253 while (count < sc->map_size) {
1254 struct g_virstor_map_entry *mapbuf;
1257 bs = MIN(MAXPHYS, sc->map_size - count);
1258 if (bs % sc->sectorsize != 0) {
1259 /* Check for alignment errors */
1260 bs = (bs / sc->sectorsize) * sc->sectorsize;
1263 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned "
1264 "for %s on %s", sc->geom->name,
1265 sc->components[0].gcons->provider->name);
1267 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error);
1268 if (mapbuf == NULL) {
1269 free(sc->map, M_GVIRSTOR);
1270 LOG_MSG(LVL_ERROR, "Error reading allocation map "
1271 "for %s from %s (offset %ju) (error %d)",
1273 sc->components[0].gcons->provider->name,
1278 bcopy(mapbuf, &sc->map[n], bs);
1281 n += bs / sizeof *(sc->map);
1284 g_access(sc->components[0].gcons, -1, 0, 0);
1285 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name);
1287 /* find first component with allocatable chunks */
1289 for (n = 0; n < sc->n_components; n++) {
1290 if (sc->components[n].chunk_next <
1291 sc->components[n].chunk_count) {
1297 /* not found? set it to the last component and handle it
1299 index = sc->n_components - 1;
1301 if (index >= sc->n_components - g_virstor_component_watermark - 1) {
1302 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1303 "(%d/%u: %s)", sc->geom->name,
1306 sc->components[index].gcons->provider->name);
1308 sc->curr_component = index;
1310 if (sc->components[index].chunk_next >=
1311 sc->components[index].chunk_count - g_virstor_chunk_watermark) {
1312 LOG_MSG(LVL_WARNING,
1313 "Component %s of %s is running out of free space "
1315 sc->components[index].gcons->provider->name,
1316 sc->geom->name, sc->components[index].chunk_count -
1317 sc->components[index].chunk_next);
1320 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map);
1321 if (sc->sectorsize % sizeof *(sc->map) != 0) {
1323 "%s: Map entries don't fit exactly in a sector (%s)",
1324 __func__, sc->geom->name);
1328 /* Recalculate allocated chunks in components & at the same time
1329 * verify map data is sane. We could trust metadata on this, but
1330 * we want to make sure. */
1331 for (n = 0; n < sc->n_components; n++)
1332 sc->components[n].chunk_next = sc->components[n].chunk_reserved;
1334 for (n = 0; n < sc->chunk_count; n++) {
1335 if (sc->map[n].provider_no >= sc->n_components ||
1336 sc->map[n].provider_chunk >=
1337 sc->components[sc->map[n].provider_no].chunk_count) {
1338 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s",
1339 __func__, (u_int)n, sc->geom->name);
1340 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u"
1341 " provider_chunk: %u, chunk_count: %u", __func__,
1342 sc->map[n].provider_no, sc->n_components,
1343 sc->map[n].provider_chunk,
1344 sc->components[sc->map[n].provider_no].chunk_count);
1347 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED)
1348 sc->components[sc->map[n].provider_no].chunk_next++;
1351 sc->provider = g_new_providerf(sc->geom, "virstor/%s",
1354 sc->provider->sectorsize = sc->sectorsize;
1355 sc->provider->mediasize = sc->virsize;
1356 g_error_provider(sc->provider, 0);
1358 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name);
1359 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting "
1360 "chunk %u", sc->provider->name, sc->curr_component,
1361 sc->components[sc->curr_component].chunk_next);
1365 * Returns count of active providers in this geom instance
1368 virstor_valid_components(struct g_virstor_softc *sc)
1373 KASSERT(sc != NULL, ("%s: softc is NULL", __func__));
1374 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__));
1375 for (i = 0; i < sc->n_components; i++)
1376 if (sc->components[i].gcons != NULL)
1382 * Called when the consumer gets orphaned (?)
1385 g_virstor_orphan(struct g_consumer *cp)
1387 struct g_virstor_softc *sc;
1388 struct g_virstor_component *comp;
1391 g_topology_assert();
1398 KASSERT(comp != NULL, ("%s: No component in private part of consumer",
1400 remove_component(sc, comp, FALSE);
1401 if (virstor_valid_components(sc) == 0)
1402 virstor_geom_destroy(sc, TRUE, FALSE);
1406 * Called to notify geom when it's been opened, and for what intent
1409 g_virstor_access(struct g_provider *pp, int dr, int dw, int de)
1411 struct g_consumer *c;
1412 struct g_virstor_softc *sc;
1416 KASSERT(pp != NULL, ("%s: NULL provider", __func__));
1418 KASSERT(gp != NULL, ("%s: NULL geom", __func__));
1422 /* It seems that .access can be called with negative dr,dw,dx
1423 * in this case but I want to check for myself */
1424 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s",
1425 dr, dw, de, pp->name);
1426 /* This should only happen when geom is withered so
1427 * allow only negative requests */
1428 KASSERT(dr <= 0 && dw <= 0 && de <= 0,
1429 ("%s: Positive access for %s", __func__, pp->name));
1430 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0)
1431 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed",
1436 /* Grab an exclusive bit to propagate on our consumers on first open */
1437 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)
1439 /* ... drop it on close */
1440 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) {
1442 update_metadata(sc); /* Writes statistical information */
1446 LIST_FOREACH(c, &gp->consumer, consumer) {
1447 KASSERT(c != NULL, ("%s: consumer is NULL", __func__));
1448 error = g_access(c, dr, dw, de);
1450 struct g_consumer *c2;
1452 /* Backout earlier changes */
1453 LIST_FOREACH(c2, &gp->consumer, consumer) {
1454 if (c2 == c) /* all eariler components fixed */
1456 g_access(c2, -dr, -dw, -de);
1465 * Generate XML dump of current state
1468 g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
1469 struct g_consumer *cp, struct g_provider *pp)
1471 struct g_virstor_softc *sc;
1473 g_topology_assert();
1476 if (sc == NULL || pp != NULL)
1480 /* For each component */
1481 struct g_virstor_component *comp;
1486 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n",
1487 indent, comp->index);
1488 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n",
1489 indent, comp->chunk_count);
1490 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n",
1491 indent, comp->chunk_next);
1492 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n",
1493 indent, comp->chunk_reserved);
1494 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n",
1496 comp->chunk_next > 0 ? 100 -
1497 ((comp->chunk_next + comp->chunk_reserved) * 100) /
1498 comp->chunk_count : 100);
1500 /* For the whole thing */
1501 u_int count, used, i;
1504 count = used = size = 0;
1505 for (i = 0; i < sc->n_components; i++) {
1506 if (sc->components[i].gcons != NULL) {
1507 count += sc->components[i].chunk_count;
1508 used += sc->components[i].chunk_next +
1509 sc->components[i].chunk_reserved;
1510 size += sc->components[i].gcons->
1511 provider->mediasize;
1515 sbuf_printf(sb, "%s<Status>"
1516 "Components=%u, Online=%u</Status>\n", indent,
1517 sc->n_components, virstor_valid_components(sc));
1518 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n",
1519 indent, 100-(used * 100) / count);
1520 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent,
1522 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n",
1523 indent, used > 0 ? 100 - (used * 100) / count : 100);
1524 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n",
1526 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n",
1527 indent, sc->chunk_count);
1528 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n",
1530 (count * 100) / sc->chunk_count);
1531 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n",
1533 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent,
1539 * GEOM .done handler
1540 * Can't use standard handler because one requested IO may
1541 * fork into additional data IOs
1544 g_virstor_done(struct bio *b)
1546 struct g_virstor_softc *sc;
1547 struct bio *parent_b;
1549 parent_b = b->bio_parent;
1550 sc = parent_b->bio_to->geom->softc;
1552 if (b->bio_error != 0) {
1553 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s",
1554 b->bio_error, b->bio_offset, b->bio_length,
1556 if (parent_b->bio_error == 0)
1557 parent_b->bio_error = b->bio_error;
1560 parent_b->bio_inbed++;
1561 parent_b->bio_completed += b->bio_completed;
1563 if (parent_b->bio_children == parent_b->bio_inbed) {
1564 parent_b->bio_completed = parent_b->bio_length;
1565 g_io_deliver(parent_b, parent_b->bio_error);
1572 * Called in g_down thread
1575 g_virstor_start(struct bio *b)
1577 struct g_virstor_softc *sc;
1578 struct g_virstor_component *comp;
1580 struct g_provider *pp;
1582 off_t offset, length;
1583 struct bio_queue_head bq;
1584 size_t chunk_size; /* cached for convenience */
1588 sc = pp->geom->softc;
1589 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__,
1590 b->bio_to->error, b->bio_to->name));
1592 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__);
1594 switch (b->bio_cmd) {
1600 g_io_deliver(b, EOPNOTSUPP);
1604 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length);
1607 chunk_size = sc->chunk_size;
1609 offset = b->bio_offset; /* virtual offset and length */
1610 length = b->bio_length;
1612 while (length > 0) {
1613 size_t chunk_index, in_chunk_offset, in_chunk_length;
1614 struct virstor_map_entry *me;
1616 chunk_index = offset / chunk_size; /* round downwards */
1617 in_chunk_offset = offset % chunk_size;
1618 in_chunk_length = min(length, chunk_size - in_chunk_offset);
1619 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)",
1620 b->bio_cmd == BIO_READ ? "R" : "W",
1622 chunk_index, in_chunk_offset, in_chunk_length);
1623 me = &sc->map[chunk_index];
1625 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) {
1626 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1627 /* Reads from unallocated chunks return zeroed
1629 if (b->bio_cmd == BIO_READ)
1630 bzero(addr, in_chunk_length);
1632 comp = &sc->components[me->provider_no];
1634 cb = g_clone_bio(b);
1636 bioq_dismantle(&bq);
1637 if (b->bio_error == 0)
1638 b->bio_error = ENOMEM;
1639 g_io_deliver(b, b->bio_error);
1642 cb->bio_to = comp->gcons->provider;
1643 cb->bio_done = g_virstor_done;
1645 (off_t)me->provider_chunk * (off_t)chunk_size
1647 cb->bio_length = in_chunk_length;
1648 cb->bio_data = addr;
1649 cb->bio_caller1 = comp;
1650 bioq_disksort(&bq, cb);
1652 } else { /* handle BIO_WRITE */
1653 KASSERT(b->bio_cmd == BIO_WRITE,
1654 ("%s: Unknown command %d", __func__,
1657 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) {
1658 /* We have a virtual chunk, represented by
1659 * the "me" entry, but it's not yet allocated
1660 * (tied to) a physical chunk. So do it now. */
1661 struct virstor_map_entry *data_me;
1662 u_int phys_chunk, comp_no;
1666 error = allocate_chunk(sc, &comp, &comp_no,
1669 /* We cannot allocate a physical chunk
1670 * to satisfy this request, so we'll
1671 * delay it to when we can...
1672 * XXX: this will prevent the fs from
1673 * being umounted! */
1674 struct g_virstor_bio_q *biq;
1675 biq = malloc(sizeof *biq, M_GVIRSTOR,
1678 bioq_dismantle(&bq);
1679 if (b->bio_error == 0)
1680 b->bio_error = ENOMEM;
1681 g_io_deliver(b, b->bio_error);
1685 mtx_lock(&sc->delayed_bio_q_mtx);
1686 STAILQ_INSERT_TAIL(&sc->delayed_bio_q,
1688 mtx_unlock(&sc->delayed_bio_q_mtx);
1689 LOG_MSG(LVL_WARNING, "Delaying BIO "
1690 "(size=%ju) until free physical "
1691 "space can be found on %s",
1693 sc->provider->name);
1696 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s "
1699 comp->gcons->provider->name,
1700 sc->provider->name);
1702 me->provider_no = comp_no;
1703 me->provider_chunk = phys_chunk;
1704 me->flags |= VIRSTOR_MAP_ALLOCATED;
1706 cb = g_clone_bio(b);
1708 me->flags &= ~VIRSTOR_MAP_ALLOCATED;
1709 me->provider_no = 0;
1710 me->provider_chunk = 0;
1711 bioq_dismantle(&bq);
1712 if (b->bio_error == 0)
1713 b->bio_error = ENOMEM;
1714 g_io_deliver(b, b->bio_error);
1718 /* The allocation table is stored continuously
1719 * at the start of the drive. We need to
1720 * calculate the offset of the sector that holds
1721 * this map entry both on the drive and in the
1723 * sc_offset will end up pointing to the drive
1725 s_offset = chunk_index * sizeof *me;
1726 s_offset = (s_offset / sc->sectorsize) *
1729 /* data_me points to map entry sector
1730 * in memory (analoguos to offset) */
1731 data_me = &sc->map[(chunk_index /
1732 sc->me_per_sector) * sc->me_per_sector];
1734 /* Commit sector with map entry to storage */
1735 cb->bio_to = sc->components[0].gcons->provider;
1736 cb->bio_done = g_virstor_done;
1737 cb->bio_offset = s_offset;
1738 cb->bio_data = (char *)data_me;
1739 cb->bio_length = sc->sectorsize;
1740 cb->bio_caller1 = &sc->components[0];
1741 bioq_disksort(&bq, cb);
1744 comp = &sc->components[me->provider_no];
1745 cb = g_clone_bio(b);
1747 bioq_dismantle(&bq);
1748 if (b->bio_error == 0)
1749 b->bio_error = ENOMEM;
1750 g_io_deliver(b, b->bio_error);
1753 /* Finally, handle the data */
1754 cb->bio_to = comp->gcons->provider;
1755 cb->bio_done = g_virstor_done;
1756 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size +
1758 cb->bio_length = in_chunk_length;
1759 cb->bio_data = addr;
1760 cb->bio_caller1 = comp;
1761 bioq_disksort(&bq, cb);
1763 addr += in_chunk_length;
1764 length -= in_chunk_length;
1765 offset += in_chunk_length;
1768 /* Fire off bio's here */
1770 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) {
1771 bioq_remove(&bq, cb);
1772 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request");
1773 comp = cb->bio_caller1;
1774 cb->bio_caller1 = NULL;
1775 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju",
1776 cb->bio_offset, cb->bio_length);
1777 g_io_request(cb, comp->gcons);
1780 if (count == 0) { /* We handled everything locally */
1781 b->bio_completed = b->bio_length;
1788 * Allocate a chunk from a physical provider. Returns physical component,
1789 * chunk index relative to the component and the component's index.
1792 allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp,
1793 u_int *comp_no_p, u_int *chunk)
1797 KASSERT(sc->curr_component < sc->n_components,
1798 ("%s: Invalid curr_component: %u", __func__, sc->curr_component));
1800 comp_no = sc->curr_component;
1801 *comp = &sc->components[comp_no];
1802 dump_component(*comp);
1803 if ((*comp)->chunk_next >= (*comp)->chunk_count) {
1804 /* This component is full. Allocate next component */
1805 if (comp_no >= sc->n_components-1) {
1806 LOG_MSG(LVL_ERROR, "All physical space allocated for %s",
1810 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT;
1811 sc->curr_component = ++comp_no;
1813 *comp = &sc->components[comp_no];
1814 if (comp_no >= sc->n_components - g_virstor_component_watermark-1)
1815 LOG_MSG(LVL_WARNING, "Device %s running out of components "
1816 "(switching to %u/%u: %s)", sc->geom->name,
1817 comp_no+1, sc->n_components,
1818 (*comp)->gcons->provider->name);
1819 /* Take care not to overwrite reserved chunks */
1820 if ( (*comp)->chunk_reserved > 0 &&
1821 (*comp)->chunk_next < (*comp)->chunk_reserved)
1822 (*comp)->chunk_next = (*comp)->chunk_reserved;
1825 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT;
1826 dump_component(*comp);
1827 *comp_no_p = comp_no;
1828 *chunk = (*comp)->chunk_next++;
1830 *comp_no_p = comp_no;
1831 *chunk = (*comp)->chunk_next++;
1836 /* Dump a component */
1838 dump_component(struct g_virstor_component *comp)
1841 if (g_virstor_debug < LVL_DEBUG2)
1843 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name);
1844 printf(" chunk_count: %u\n", comp->chunk_count);
1845 printf(" chunk_next: %u\n", comp->chunk_next);
1846 printf(" flags: %u\n", comp->flags);
1850 /* Dump a map entry */
1852 dump_me(struct virstor_map_entry *me, unsigned int nr)
1854 if (g_virstor_debug < LVL_DEBUG)
1856 printf("VIRT. CHUNK #%d: ", nr);
1857 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0)
1858 printf("(unallocated)\n");
1860 printf("allocated at provider %u, provider_chunk %u\n",
1861 me->provider_no, me->provider_chunk);
1866 * Dismantle bio_queue and destroy its components
1869 bioq_dismantle(struct bio_queue_head *bq)
1873 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) {
1880 * The function that shouldn't be called.
1881 * When this is called, the stack is already garbled because of
1882 * argument mismatch. There's nothing to do now but panic, which is
1883 * accidentally the whole purpose of this function.
1884 * Motivation: to guard from accidentally calling geom methods when
1885 * they shouldn't be called. (see g_..._taste)
1890 panic("invalid_call() has just been called. Something's fishy here.");
1893 DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */