g_virstor.c revision 172302
190075Sobrien/*- 290075Sobrien * Copyright (c) 2006-2007 Ivan Voras <ivoras@freebsd.org> 3169689Skan * All rights reserved. 490075Sobrien * 590075Sobrien * Redistribution and use in source and binary forms, with or without 690075Sobrien * modification, are permitted provided that the following conditions 790075Sobrien * are met: 890075Sobrien * 1. Redistributions of source code must retain the above copyright 990075Sobrien * notice, this list of conditions and the following disclaimer. 1090075Sobrien * 2. Redistributions in binary form must reproduce the above copyright 1190075Sobrien * notice, this list of conditions and the following disclaimer in the 1290075Sobrien * documentation and/or other materials provided with the distribution. 1390075Sobrien * 1490075Sobrien * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 1590075Sobrien * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 1690075Sobrien * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 1790075Sobrien * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 1890075Sobrien * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19169689Skan * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20169689Skan * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2190075Sobrien * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2290075Sobrien * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2390075Sobrien * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2490075Sobrien * SUCH DAMAGE. 25132718Skan */ 26132718Skan 2790075Sobrien/* Implementation notes: 28169689Skan * - "Components" are wrappers around providers that make up the 2990075Sobrien * virtual storage (i.e. a virstor has "physical" components) 3090075Sobrien */ 3196263Sobrien 3296263Sobrien#include <sys/cdefs.h> 3390075Sobrien__FBSDID("$FreeBSD: head/sys/geom/virstor/g_virstor.c 172302 2007-09-23 07:34:23Z pjd $"); 3496263Sobrien 35169689Skan#include <sys/param.h> 3690075Sobrien#include <sys/systm.h> 3790075Sobrien#include <sys/kernel.h> 3890075Sobrien#include <sys/module.h> 3990075Sobrien#include <sys/lock.h> 4090075Sobrien#include <sys/mutex.h> 4190075Sobrien#include <sys/sx.h> 4290075Sobrien#include <sys/bio.h> 4390075Sobrien#include <sys/sysctl.h> 4490075Sobrien#include <sys/malloc.h> 4590075Sobrien#include <sys/time.h> 4690075Sobrien#include <sys/proc.h> 4790075Sobrien#include <sys/kthread.h> 4890075Sobrien#include <sys/mutex.h> 4990075Sobrien#include <vm/uma.h> 5090075Sobrien#include <geom/geom.h> 51132718Skan 52132718Skan#include <geom/virstor/g_virstor.h> 53132718Skan#include <geom/virstor/g_virstor_md.h> 54169689Skan 55169689Skan/* Declare malloc(9) label */ 56132718Skanstatic MALLOC_DEFINE(M_GVIRSTOR, "gvirstor", "GEOM_VIRSTOR Data"); 57132718Skan 5890075Sobrien/* GEOM class methods */ 59107590Sobrienstatic g_init_t g_virstor_init; 60107590Sobrienstatic g_fini_t g_virstor_fini; 61107590Sobrienstatic g_taste_t g_virstor_taste; 62107590Sobrienstatic g_ctl_req_t g_virstor_config; 63132718Skanstatic g_ctl_destroy_geom_t g_virstor_destroy_geom; 64107590Sobrien 65107590Sobrien/* Declare & initialize class structure ("geom class") */ 66107590Sobrienstruct g_class g_virstor_class = { 67107590Sobrien .name = G_VIRSTOR_CLASS_NAME, 68132718Skan .version = G_VERSION, 69117395Skan .init = g_virstor_init, 70107590Sobrien .fini = g_virstor_fini, 71107590Sobrien .taste = g_virstor_taste, 72169689Skan .ctlreq = g_virstor_config, 73107590Sobrien .destroy_geom = g_virstor_destroy_geom 74169689Skan /* The .dumpconf and the rest are only usable for a geom instance, so 75107590Sobrien * they will be set when such instance is created. */ 76107590Sobrien}; 77107590Sobrien 78107590Sobrien/* Declare sysctl's and loader tunables */ 79107590SobrienSYSCTL_DECL(_kern_geom); 80107590SobrienSYSCTL_NODE(_kern_geom, OID_AUTO, virstor, CTLFLAG_RW, 0, "GEOM_GVIRSTOR information"); 8190075Sobrien 8290075Sobrienstatic u_int g_virstor_debug = 2; /* XXX: lower to 2 when released to public */ 8390075SobrienTUNABLE_INT("kern.geom.virstor.debug", &g_virstor_debug); 8490075SobrienSYSCTL_UINT(_kern_geom_virstor, OID_AUTO, debug, CTLFLAG_RW, &g_virstor_debug, 85132718Skan 0, "Debug level (2=production, 5=normal, 15=excessive)"); 8690075Sobrien 8790075Sobrienstatic u_int g_virstor_chunk_watermark = 100; 8890075SobrienTUNABLE_INT("kern.geom.virstor.chunk_watermark", &g_virstor_chunk_watermark); 8990075SobrienSYSCTL_UINT(_kern_geom_virstor, OID_AUTO, chunk_watermark, CTLFLAG_RW, 90169689Skan &g_virstor_chunk_watermark, 0, 9190075Sobrien "Minimum number of free chunks before issuing administrative warning"); 9290075Sobrien 93132718Skanstatic u_int g_virstor_component_watermark = 1; 94107590SobrienTUNABLE_INT("kern.geom.virstor.component_watermark", 9590075Sobrien &g_virstor_component_watermark); 9690075SobrienSYSCTL_UINT(_kern_geom_virstor, OID_AUTO, component_watermark, CTLFLAG_RW, 9790075Sobrien &g_virstor_component_watermark, 0, 98169689Skan "Minimum number of free components before issuing administrative warning"); 99107590Sobrien 10090075Sobrienstatic int read_metadata(struct g_consumer *, struct g_virstor_metadata *); 10190075Sobrienstatic int write_metadata(struct g_consumer *, struct g_virstor_metadata *); 10290075Sobrienstatic int clear_metadata(struct g_virstor_component *); 10390075Sobrienstatic int add_provider_to_geom(struct g_virstor_softc *, struct g_provider *, 10490075Sobrien struct g_virstor_metadata *); 105132718Skanstatic struct g_geom *create_virstor_geom(struct g_class *, 10690075Sobrien struct g_virstor_metadata *); 107132718Skanstatic void virstor_check_and_run(struct g_virstor_softc *); 108169689Skanstatic u_int virstor_valid_components(struct g_virstor_softc *); 109169689Skanstatic int virstor_geom_destroy(struct g_virstor_softc *, boolean_t, 110169689Skan boolean_t); 11190075Sobrienstatic void remove_component(struct g_virstor_softc *, 112169689Skan struct g_virstor_component *, boolean_t); 113169689Skanstatic void bioq_dismantle(struct bio_queue_head *); 114117395Skanstatic int allocate_chunk(struct g_virstor_softc *, 115117395Skan struct g_virstor_component **, u_int *, u_int *); 116169689Skanstatic void delay_destroy_consumer(void *, int); 117169689Skanstatic void dump_component(struct g_virstor_component *comp); 118169689Skan#if 0 119169689Skanstatic void dump_me(struct virstor_map_entry *me, unsigned int nr); 120117395Skan#endif 121169689Skan 122132718Skanstatic void virstor_ctl_stop(struct gctl_req *, struct g_class *); 12390075Sobrienstatic void virstor_ctl_add(struct gctl_req *, struct g_class *); 12490075Sobrienstatic void virstor_ctl_remove(struct gctl_req *, struct g_class *); 12590075Sobrienstatic struct g_virstor_softc * virstor_find_geom(const struct g_class *, 12690075Sobrien const char *); 12790075Sobrienstatic void update_metadata(struct g_virstor_softc *); 128169689Skanstatic void fill_metadata(struct g_virstor_softc *, struct g_virstor_metadata *, 129169689Skan u_int, u_int); 130169689Skan 131169689Skanstatic void g_virstor_orphan(struct g_consumer *); 132169689Skanstatic int g_virstor_access(struct g_provider *, int, int, int); 133169689Skanstatic void g_virstor_start(struct bio *); 134169689Skanstatic void g_virstor_dumpconf(struct sbuf *, const char *, struct g_geom *, 135169689Skan struct g_consumer *, struct g_provider *); 136169689Skanstatic void g_virstor_done(struct bio *); 137169689Skan 138169689Skanstatic void invalid_call(void); 139169689Skan/* 140169689Skan * Initialise GEOM class (per-class callback) 141169689Skan */ 142169689Skanstatic void 143169689Skang_virstor_init(struct g_class *mp __unused) 144169689Skan{ 145169689Skan 14690075Sobrien /* Catch map struct size mismatch at compile time; Map entries must 14790075Sobrien * fit into MAXPHYS exactly, with no wasted space. */ 148117395Skan CTASSERT(VIRSTOR_MAP_BLOCK_ENTRIES*VIRSTOR_MAP_ENTRY_SIZE == MAXPHYS); 14990075Sobrien 15090075Sobrien /* Init UMA zones, TAILQ's, other global vars */ 15190075Sobrien} 15290075Sobrien 15390075Sobrien/* 15490075Sobrien * Finalise GEOM class (per-class callback) 155169689Skan */ 15690075Sobrienstatic void 15790075Sobrieng_virstor_fini(struct g_class *mp __unused) 158132718Skan{ 15990075Sobrien 160169689Skan /* Deinit UMA zones & global vars */ 16190075Sobrien} 16290075Sobrien 16390075Sobrien/* 16490075Sobrien * Config (per-class callback) 16590075Sobrien */ 16690075Sobrienstatic void 16790075Sobrieng_virstor_config(struct gctl_req *req, struct g_class *cp, char const *verb) 16890075Sobrien{ 16990075Sobrien uint32_t *version; 170169689Skan 171169689Skan g_topology_assert(); 17290075Sobrien 17390075Sobrien version = gctl_get_paraml(req, "version", sizeof(*version)); 174169689Skan if (version == NULL) { 17590075Sobrien gctl_error(req, "Failed to get 'version' argument"); 17690075Sobrien return; 17790075Sobrien } 178117395Skan if (*version != G_VIRSTOR_VERSION) { 17990075Sobrien gctl_error(req, "Userland and kernel versions out of sync"); 18090075Sobrien return; 18190075Sobrien } 18290075Sobrien 18390075Sobrien g_topology_unlock(); 184169689Skan if (strcmp(verb, "add") == 0) 18590075Sobrien virstor_ctl_add(req, cp); 18690075Sobrien else if (strcmp(verb, "stop") == 0 || strcmp(verb, "destroy") == 0) 18790075Sobrien virstor_ctl_stop(req, cp); 188169689Skan else if (strcmp(verb, "remove") == 0) 18990075Sobrien virstor_ctl_remove(req, cp); 19090075Sobrien else 19190075Sobrien gctl_error(req, "unknown verb: '%s'", verb); 19290075Sobrien g_topology_lock(); 193169689Skan} 194169689Skan 195169689Skan/* 196169689Skan * "stop" verb from userland 19790075Sobrien */ 19890075Sobrienstatic void 19990075Sobrienvirstor_ctl_stop(struct gctl_req *req, struct g_class *cp) 20090075Sobrien{ 20190075Sobrien int *force, *nargs; 20290075Sobrien int i; 20390075Sobrien 20490075Sobrien nargs = gctl_get_paraml(req, "nargs", sizeof *nargs); 205169689Skan if (nargs == NULL) { 20690075Sobrien gctl_error(req, "Error fetching argument '%s'", "nargs"); 20790075Sobrien return; 20890075Sobrien } 209169689Skan if (*nargs < 1) { 21090075Sobrien gctl_error(req, "Invalid number of arguments"); 21190075Sobrien return; 21290075Sobrien } 21390075Sobrien force = gctl_get_paraml(req, "force", sizeof *force); 21490075Sobrien if (force == NULL) { 21590075Sobrien gctl_error(req, "Error fetching argument '%s'", "force"); 21690075Sobrien return; 21790075Sobrien } 21890075Sobrien 219169689Skan g_topology_lock(); 22090075Sobrien for (i = 0; i < *nargs; i++) { 221169689Skan char param[8]; 22290075Sobrien const char *name; 22390075Sobrien struct g_virstor_softc *sc; 224169689Skan int error; 225169689Skan 22690075Sobrien sprintf(param, "arg%d", i); 22790075Sobrien name = gctl_get_asciiparam(req, param); 22890075Sobrien sc = virstor_find_geom(cp, name); 22990075Sobrien LOG_MSG(LVL_INFO, "Stopping %s by the userland command", 23090075Sobrien sc->geom->name); 23190075Sobrien update_metadata(sc); 23290075Sobrien if ((error = virstor_geom_destroy(sc, TRUE, TRUE)) != 0) { 23390075Sobrien LOG_MSG(LVL_ERROR, "Cannot destroy %s: %d", 23490075Sobrien sc->geom->name, error); 23590075Sobrien } 23690075Sobrien } 23790075Sobrien g_topology_unlock(); 23890075Sobrien} 239117395Skan 240117395Skan/* 241117395Skan * "add" verb from userland - add new component(s) to the structure. 242132718Skan * This will be done all at once in here, without going through the 243117395Skan * .taste function for new components. 244117395Skan */ 245117395Skanstatic void 246117395Skanvirstor_ctl_add(struct gctl_req *req, struct g_class *cp) 247117395Skan{ 248117395Skan /* Note: while this is going on, I/O is being done on 249169689Skan * the g_up and g_down threads. The idea is to make changes 250117395Skan * to softc members in a way that can atomically activate 251169689Skan * them all at once. */ 252117395Skan struct g_virstor_softc *sc; 253117395Skan int *hardcode, *nargs; 254117395Skan const char *geom_name; /* geom to add a component to */ 255117395Skan struct g_consumer *fcp; 256117395Skan struct g_virstor_bio_q *bq; 257117395Skan u_int added; 258117395Skan int error; 259117395Skan int i; 260132718Skan 261117395Skan nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs)); 262169689Skan if (nargs == NULL) { 263117395Skan gctl_error(req, "Error fetching argument '%s'", "nargs"); 264132718Skan return; 265117395Skan } 266132718Skan if (*nargs < 2) { 267117395Skan gctl_error(req, "Invalid number of arguments"); 268132718Skan return; 269169689Skan } 270169689Skan hardcode = gctl_get_paraml(req, "hardcode", sizeof(*hardcode)); 271117395Skan if (hardcode == NULL) { 272117395Skan gctl_error(req, "Error fetching argument '%s'", "hardcode"); 273117395Skan return; 27490075Sobrien } 275117395Skan 27690075Sobrien /* Find "our" geom */ 27790075Sobrien geom_name = gctl_get_asciiparam(req, "arg0"); 27890075Sobrien if (geom_name == NULL) { 279132718Skan gctl_error(req, "Error fetching argument '%s'", "geom_name (arg0)"); 28090075Sobrien return; 28190075Sobrien } 282169689Skan sc = virstor_find_geom(cp, geom_name); 283117395Skan if (sc == NULL) { 28490075Sobrien gctl_error(req, "Don't know anything about '%s'", geom_name); 285169689Skan return; 28690075Sobrien } 28790075Sobrien 28890075Sobrien if (virstor_valid_components(sc) != sc->n_components) { 289117395Skan LOG_MSG(LVL_ERROR, "Cannot add components to incomplete " 290117395Skan "virstor %s", sc->geom->name); 29190075Sobrien gctl_error(req, "Virstor %s is incomplete", sc->geom->name); 29290075Sobrien return; 29390075Sobrien } 29490075Sobrien 29590075Sobrien fcp = sc->components[0].gcons; 296169689Skan added = 0; 29790075Sobrien g_topology_lock(); 29890075Sobrien for (i = 1; i < *nargs; i++) { 29990075Sobrien struct g_virstor_metadata md; 30090075Sobrien char aname[8]; 30190075Sobrien const char *prov_name; 30290075Sobrien struct g_provider *pp; 30390075Sobrien struct g_consumer *cp; 30490075Sobrien u_int nc; 30590075Sobrien u_int j; 30690075Sobrien 30790075Sobrien snprintf(aname, sizeof aname, "arg%d", i); 30890075Sobrien prov_name = gctl_get_asciiparam(req, aname); 30990075Sobrien if (strncmp(prov_name, _PATH_DEV, strlen(_PATH_DEV)) == 0) 310169689Skan prov_name += strlen(_PATH_DEV); 311169689Skan 312169689Skan pp = g_provider_by_name(prov_name); 313169689Skan if (pp == NULL) { 314169689Skan /* This is the most common error so be verbose about it */ 315169689Skan if (added != 0) { 316169689Skan gctl_error(req, "Invalid provider: '%s' (added" 317169689Skan " %u components)", prov_name, added); 318169689Skan update_metadata(sc); 319169689Skan } else { 32090075Sobrien gctl_error(req, "Invalid provider: '%s'", 32190075Sobrien prov_name); 32290075Sobrien } 32390075Sobrien g_topology_unlock(); 32490075Sobrien return; 32590075Sobrien } 32690075Sobrien cp = g_new_consumer(sc->geom); 32790075Sobrien if (cp == NULL) { 32890075Sobrien gctl_error(req, "Cannot create consumer"); 32990075Sobrien g_topology_unlock(); 33090075Sobrien return; 33190075Sobrien } 33290075Sobrien error = g_attach(cp, pp); 33390075Sobrien if (error != 0) { 33490075Sobrien gctl_error(req, "Cannot attach a consumer to %s", 33590075Sobrien pp->name); 33690075Sobrien g_destroy_consumer(cp); 33790075Sobrien g_topology_unlock(); 33890075Sobrien return; 339132718Skan } 34090075Sobrien if (fcp->acr != 0 || fcp->acw != 0 || fcp->ace != 0) { 34190075Sobrien error = g_access(cp, fcp->acr, fcp->acw, fcp->ace); 34290075Sobrien if (error != 0) { 34390075Sobrien gctl_error(req, "Access request failed for %s", 34490075Sobrien pp->name); 345117395Skan g_destroy_consumer(cp); 346169689Skan g_topology_unlock(); 34790075Sobrien return; 348169689Skan } 34990075Sobrien } 35090075Sobrien if (fcp->provider->sectorsize != pp->sectorsize) { 35190075Sobrien gctl_error(req, "Sector size doesn't fit for %s", 35290075Sobrien pp->name); 35390075Sobrien g_destroy_consumer(cp); 354117395Skan g_topology_unlock(); 35590075Sobrien return; 356169689Skan } 35790075Sobrien for (j = 0; j < sc->n_components; j++) { 35890075Sobrien if (strcmp(sc->components[j].gcons->provider->name, 359169689Skan pp->name) == 0) { 36090075Sobrien gctl_error(req, "Component %s already in %s", 36190075Sobrien pp->name, sc->geom->name); 362169689Skan g_destroy_consumer(cp); 36390075Sobrien g_topology_unlock(); 36490075Sobrien return; 36590075Sobrien } 366117395Skan } 367117395Skan sc->components = realloc(sc->components, 368169689Skan sizeof(*sc->components) * (sc->n_components + 1), 369117395Skan M_GVIRSTOR, M_WAITOK); 37090075Sobrien 37190075Sobrien nc = sc->n_components; 37290075Sobrien sc->components[nc].gcons = cp; 37390075Sobrien sc->components[nc].sc = sc; 37490075Sobrien sc->components[nc].index = nc; 37590075Sobrien sc->components[nc].chunk_count = cp->provider->mediasize / 37690075Sobrien sc->chunk_size; 377132718Skan sc->components[nc].chunk_next = 0; 37890075Sobrien sc->components[nc].chunk_reserved = 0; 37990075Sobrien 38090075Sobrien if (sc->components[nc].chunk_count < 4) { 38190075Sobrien gctl_error(req, "Provider too small: %s", 38290075Sobrien cp->provider->name); 38390075Sobrien g_destroy_consumer(cp); 38490075Sobrien g_topology_unlock(); 38590075Sobrien return; 38690075Sobrien } 38790075Sobrien fill_metadata(sc, &md, nc, *hardcode); 38890075Sobrien write_metadata(cp, &md); 389132718Skan /* The new component becomes visible when n_components is 39090075Sobrien * incremented */ 39190075Sobrien sc->n_components++; 39290075Sobrien added++; 39390075Sobrien 394169689Skan } 39590075Sobrien /* This call to update_metadata() is critical. In case there's a 39690075Sobrien * power failure in the middle of it and some components are updated 39790075Sobrien * while others are not, there will be trouble on next .taste() iff 39890075Sobrien * a non-updated component is detected first */ 39990075Sobrien update_metadata(sc); 40090075Sobrien g_topology_unlock(); 40190075Sobrien LOG_MSG(LVL_INFO, "Added %d component(s) to %s", added, 40290075Sobrien sc->geom->name); 40390075Sobrien /* Fire off BIOs previously queued because there wasn't any 40490075Sobrien * physical space left. If the BIOs still can't be satisfied 40590075Sobrien * they will again be added to the end of the queue (during 40690075Sobrien * which the mutex will be recursed) */ 40790075Sobrien bq = malloc(sizeof(*bq), M_GVIRSTOR, M_WAITOK); 40890075Sobrien bq->bio = NULL; 40990075Sobrien mtx_lock(&sc->delayed_bio_q_mtx); 41090075Sobrien /* First, insert a sentinel to the queue end, so we don't 41190075Sobrien * end up in an infinite loop if there's still no free 41290075Sobrien * space available. */ 41390075Sobrien STAILQ_INSERT_TAIL(&sc->delayed_bio_q, bq, linkage); 41490075Sobrien while (!STAILQ_EMPTY(&sc->delayed_bio_q)) { 41590075Sobrien bq = STAILQ_FIRST(&sc->delayed_bio_q); 416132718Skan if (bq->bio != NULL) { 41790075Sobrien g_virstor_start(bq->bio); 418117395Skan STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage); 41990075Sobrien free(bq, M_GVIRSTOR); 420117395Skan } else { 421169689Skan STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage); 42290075Sobrien free(bq, M_GVIRSTOR); 423117395Skan break; 42490075Sobrien } 425169689Skan } 42690075Sobrien mtx_unlock(&sc->delayed_bio_q_mtx); 42790075Sobrien 42890075Sobrien} 42990075Sobrien 43090075Sobrien/* 43190075Sobrien * Find a geom handled by the class 43290075Sobrien */ 43390075Sobrienstatic struct g_virstor_softc * 43490075Sobrienvirstor_find_geom(const struct g_class *cp, const char *name) 43590075Sobrien{ 43690075Sobrien struct g_geom *gp; 43790075Sobrien 43890075Sobrien LIST_FOREACH(gp, &cp->geom, geom) { 43990075Sobrien if (strcmp(name, gp->name) == 0) 44090075Sobrien return (gp->softc); 44190075Sobrien } 44290075Sobrien return (NULL); 44390075Sobrien} 44490075Sobrien 445117395Skan/* 44690075Sobrien * Update metadata on all components to reflect the current state 44790075Sobrien * of these fields: 448117395Skan * - chunk_next 449117395Skan * - flags 45090075Sobrien * - md_count 45190075Sobrien * Expects things to be set up so write_metadata() can work, i.e. 45290075Sobrien * the topology lock must be held. 453169689Skan */ 45490075Sobrienstatic void 45590075Sobrienupdate_metadata(struct g_virstor_softc *sc) 45690075Sobrien{ 45790075Sobrien struct g_virstor_metadata md; 45890075Sobrien int n; 45990075Sobrien 460169689Skan if (virstor_valid_components(sc) != sc->n_components) 46190075Sobrien return; /* Incomplete device */ 46290075Sobrien LOG_MSG(LVL_DEBUG, "Updating metadata on components for %s", 46390075Sobrien sc->geom->name); 46490075Sobrien /* Update metadata on components */ 46590075Sobrien g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, 46690075Sobrien sc->geom->class->name, sc->geom->name); 467117395Skan g_topology_assert(); 46890075Sobrien for (n = 0; n < sc->n_components; n++) { 46990075Sobrien read_metadata(sc->components[n].gcons, &md); 470117395Skan md.chunk_next = sc->components[n].chunk_next; 471117395Skan md.flags = sc->components[n].flags; 47290075Sobrien md.md_count = sc->n_components; 47390075Sobrien write_metadata(sc->components[n].gcons, &md); 474117395Skan } 47590075Sobrien} 47690075Sobrien 47790075Sobrien/* 478169689Skan * Fills metadata (struct md) from information stored in softc and the nc'th 479169689Skan * component of virstor 480169689Skan */ 481169689Skanstatic void 482169689Skanfill_metadata(struct g_virstor_softc *sc, struct g_virstor_metadata *md, 483169689Skan u_int nc, u_int hardcode) 484169689Skan{ 485169689Skan struct g_virstor_component *c; 486169689Skan 487169689Skan bzero(md, sizeof *md); 488169689Skan c = &sc->components[nc]; 489169689Skan 490169689Skan strncpy(md->md_magic, G_VIRSTOR_MAGIC, sizeof md->md_magic); 491169689Skan md->md_version = G_VIRSTOR_VERSION; 492169689Skan strncpy(md->md_name, sc->geom->name, sizeof md->md_name); 493169689Skan md->md_id = sc->id; 494169689Skan md->md_virsize = sc->virsize; 495169689Skan md->md_chunk_size = sc->chunk_size; 496169689Skan md->md_count = sc->n_components; 497169689Skan 498169689Skan if (hardcode) { 499169689Skan strncpy(md->provider, c->gcons->provider->name, 500169689Skan sizeof md->provider); 501169689Skan } 502169689Skan md->no = nc; 50390075Sobrien md->provsize = c->gcons->provider->mediasize; 50490075Sobrien md->chunk_count = c->chunk_count; 50590075Sobrien md->chunk_next = c->chunk_next; 50690075Sobrien md->chunk_reserved = c->chunk_reserved; 507132718Skan md->flags = c->flags; 50890075Sobrien} 50990075Sobrien 51090075Sobrien/* 51190075Sobrien * Remove a component from virstor device. 51290075Sobrien * Can only be done if the component is unallocated. 51390075Sobrien */ 51490075Sobrienstatic void 51590075Sobrienvirstor_ctl_remove(struct gctl_req *req, struct g_class *cp) 51690075Sobrien{ 51790075Sobrien /* As this is executed in parallel to I/O, operations on virstor 51890075Sobrien * structures must be as atomic as possible. */ 51990075Sobrien struct g_virstor_softc *sc; 52090075Sobrien int *nargs; 52190075Sobrien const char *geom_name; 522132718Skan u_int removed; 52390075Sobrien int i; 524169689Skan 525169689Skan nargs = gctl_get_paraml(req, "nargs", sizeof(*nargs)); 52690075Sobrien if (nargs == NULL) { 52790075Sobrien gctl_error(req, "Error fetching argument '%s'", "nargs"); 52890075Sobrien return; 52990075Sobrien } 53090075Sobrien if (*nargs < 2) { 531169689Skan gctl_error(req, "Invalid number of arguments"); 532169689Skan return; 53390075Sobrien } 53490075Sobrien /* Find "our" geom */ 53590075Sobrien geom_name = gctl_get_asciiparam(req, "arg0"); 53690075Sobrien if (geom_name == NULL) { 53790075Sobrien gctl_error(req, "Error fetching argument '%s'", 53890075Sobrien "geom_name (arg0)"); 539132718Skan return; 54090075Sobrien } 54190075Sobrien sc = virstor_find_geom(cp, geom_name); 54290075Sobrien if (sc == NULL) { 54390075Sobrien gctl_error(req, "Don't know anything about '%s'", geom_name); 54490075Sobrien return; 54590075Sobrien } 54690075Sobrien 54790075Sobrien if (virstor_valid_components(sc) != sc->n_components) { 54890075Sobrien LOG_MSG(LVL_ERROR, "Cannot remove components from incomplete " 54990075Sobrien "virstor %s", sc->geom->name); 55090075Sobrien gctl_error(req, "Virstor %s is incomplete", sc->geom->name); 55190075Sobrien return; 55290075Sobrien } 55390075Sobrien 55490075Sobrien removed = 0; 555169689Skan for (i = 1; i < *nargs; i++) { 556169689Skan char param[8]; 55790075Sobrien const char *prov_name; 55890075Sobrien int j, found; 55990075Sobrien struct g_virstor_component *newcomp, *compbak; 560169689Skan 56190075Sobrien sprintf(param, "arg%d", i); 56290075Sobrien prov_name = gctl_get_asciiparam(req, param); 563169689Skan if (strncmp(prov_name, _PATH_DEV, strlen(_PATH_DEV)) == 0) 56490075Sobrien prov_name += strlen(_PATH_DEV); 565169689Skan 56690075Sobrien found = -1; 567169689Skan for (j = 0; j < sc->n_components; j++) { 568169689Skan if (strcmp(sc->components[j].gcons->provider->name, 569169689Skan prov_name) == 0) { 570169689Skan found = j; 57190075Sobrien break; 57290075Sobrien } 57390075Sobrien } 57490075Sobrien if (found == -1) { 57590075Sobrien LOG_MSG(LVL_ERROR, "No %s component in %s", 57690075Sobrien prov_name, sc->geom->name); 57790075Sobrien continue; 57890075Sobrien } 579132718Skan 58090075Sobrien compbak = sc->components; 581117395Skan newcomp = malloc(sc->n_components * sizeof(*sc->components), 58290075Sobrien M_GVIRSTOR, M_WAITOK | M_ZERO); 583169689Skan bcopy(sc->components, newcomp, found * sizeof(*sc->components)); 584169689Skan bcopy(&sc->components[found + 1], newcomp + found, 58590075Sobrien found * sizeof(*sc->components)); 58690075Sobrien if ((sc->components[j].flags & VIRSTOR_PROVIDER_ALLOCATED) != 0) { 587169689Skan LOG_MSG(LVL_ERROR, "Allocated provider %s cannot be " 588169689Skan "removed from %s", 589169689Skan prov_name, sc->geom->name); 590169689Skan free(newcomp, M_GVIRSTOR); 591169689Skan /* We'll consider this non-fatal error */ 592169689Skan continue; 593169689Skan } 594169689Skan /* Renumerate unallocated components */ 595169689Skan for (j = 0; j < sc->n_components-1; j++) { 59690075Sobrien if ((sc->components[j].flags & 59790075Sobrien VIRSTOR_PROVIDER_ALLOCATED) == 0) { 59890075Sobrien sc->components[j].index = j; 59990075Sobrien } 60090075Sobrien } 601132718Skan /* This is the critical section. If a component allocation 60290075Sobrien * event happens while both variables are not yet set, 603117395Skan * there will be trouble. Something will panic on encountering 60490075Sobrien * NULL sc->components[x].gcomp member. 605117395Skan * Luckily, component allocation happens very rarely and 606169689Skan * removing components is an abnormal action in any case. */ 607117395Skan sc->components = newcomp; 60890075Sobrien sc->n_components--; 60990075Sobrien /* End critical section */ 61090075Sobrien 61190075Sobrien g_topology_lock(); 61290075Sobrien if (clear_metadata(&compbak[found]) != 0) { 61390075Sobrien LOG_MSG(LVL_WARNING, "Trouble ahead: cannot clear " 61490075Sobrien "metadata on %s", prov_name); 61590075Sobrien } 61690075Sobrien g_detach(compbak[found].gcons); 61790075Sobrien g_destroy_consumer(compbak[found].gcons); 61890075Sobrien g_topology_unlock(); 61990075Sobrien 62090075Sobrien free(compbak, M_GVIRSTOR); 62190075Sobrien 622132718Skan removed++; 62390075Sobrien } 624169689Skan 62590075Sobrien /* This call to update_metadata() is critical. In case there's a 62690075Sobrien * power failure in the middle of it and some components are updated 62790075Sobrien * while others are not, there will be trouble on next .taste() iff 62890075Sobrien * a non-updated component is detected first */ 62990075Sobrien g_topology_lock(); 63090075Sobrien update_metadata(sc); 63190075Sobrien g_topology_unlock(); 63290075Sobrien LOG_MSG(LVL_INFO, "Removed %d component(s) from %s", removed, 63390075Sobrien sc->geom->name); 63490075Sobrien} 635169689Skan 636169689Skan/* 63790075Sobrien * Clear metadata sector on component 63890075Sobrien */ 63990075Sobrienstatic int 64090075Sobrienclear_metadata(struct g_virstor_component *comp) 64190075Sobrien{ 64290075Sobrien char *buf; 64390075Sobrien int error; 64490075Sobrien 64590075Sobrien LOG_MSG(LVL_INFO, "Clearing metadata on %s", 64690075Sobrien comp->gcons->provider->name); 64790075Sobrien g_topology_assert(); 648169689Skan error = g_access(comp->gcons, 0, 1, 0); 649169689Skan if (error != 0) 65090075Sobrien return (error); 651169689Skan buf = malloc(comp->gcons->provider->sectorsize, M_GVIRSTOR, 652169689Skan M_WAITOK | M_ZERO); 65390075Sobrien error = g_write_data(comp->gcons, 654169689Skan comp->gcons->provider->mediasize - 65590075Sobrien comp->gcons->provider->sectorsize, 656169689Skan buf, 65790075Sobrien comp->gcons->provider->sectorsize); 65890075Sobrien free(buf, M_GVIRSTOR); 659169689Skan g_access(comp->gcons, 0, -1, 0); 660169689Skan return (error); 661169689Skan} 66290075Sobrien 663169689Skan/* 66490075Sobrien * Destroy geom forcibly. 66590075Sobrien */ 66690075Sobrienstatic int 667117395Skang_virstor_destroy_geom(struct gctl_req *req __unused, struct g_class *mp, 66890075Sobrien struct g_geom *gp) 66990075Sobrien{ 67090075Sobrien struct g_virstor_softc *sc; 67190075Sobrien int exitval; 67290075Sobrien 673169689Skan sc = gp->softc; 67490075Sobrien KASSERT(sc != NULL, ("%s: NULL sc", __func__)); 67590075Sobrien 67690075Sobrien exitval = 0; 677169689Skan LOG_MSG(LVL_DEBUG, "%s called for %s, sc=%p", __func__, gp->name, 67890075Sobrien gp->softc); 67990075Sobrien 68090075Sobrien if (sc != NULL) { 68190075Sobrien#ifdef INVARIANTS 682169689Skan char *buf; 683169689Skan int error; 684169689Skan off_t off; 68590075Sobrien int isclean, count; 68690075Sobrien int n; 68790075Sobrien 68890075Sobrien LOG_MSG(LVL_INFO, "INVARIANTS detected"); 68990075Sobrien LOG_MSG(LVL_INFO, "Verifying allocation " 69090075Sobrien "table for %s", sc->geom->name); 69190075Sobrien count = 0; 692169689Skan for (n = 0; n < sc->chunk_count; n++) { 69390075Sobrien if (sc->map[n].flags || VIRSTOR_MAP_ALLOCATED != 0) 69490075Sobrien count++; 695169689Skan } 69690075Sobrien LOG_MSG(LVL_INFO, "Device %s has %d allocated chunks", 697169689Skan sc->geom->name, count); 69890075Sobrien n = off = count = 0; 69990075Sobrien isclean = 1; 70090075Sobrien if (virstor_valid_components(sc) != sc->n_components) { 701169689Skan /* This is a incomplete virstor device (not all 702169689Skan * components have been found) */ 70390075Sobrien LOG_MSG(LVL_ERROR, "Device %s is incomplete", 704169689Skan sc->geom->name); 705169689Skan goto bailout; 70690075Sobrien } 70790075Sobrien error = g_access(sc->components[0].gcons, 1, 0, 0); 70890075Sobrien KASSERT(error == 0, ("%s: g_access failed (%d)", __func__, 70990075Sobrien error)); 71090075Sobrien /* Compare the whole on-disk allocation table with what's 711169689Skan * currently in memory */ 712169689Skan while (n < sc->chunk_count) { 713169689Skan buf = g_read_data(sc->components[0].gcons, off, 71490075Sobrien sc->sectorsize, &error); 71590075Sobrien KASSERT(buf != NULL, ("g_read_data returned NULL (%d) " 716169689Skan "for read at %jd", error, off)); 71790075Sobrien if (bcmp(buf, &sc->map[n], sc->sectorsize) != 0) { 71890075Sobrien LOG_MSG(LVL_ERROR, "ERROR in allocation table, " 71990075Sobrien "entry %d, offset %jd", n, off); 720169689Skan isclean = 0; 721169689Skan count++; 72290075Sobrien } 72390075Sobrien n += sc->me_per_sector; 724169689Skan off += sc->sectorsize; 72590075Sobrien g_free(buf); 726169689Skan } 727169689Skan error = g_access(sc->components[0].gcons, -1, 0, 0); 728169689Skan KASSERT(error == 0, ("%s: g_access failed (%d) on exit", 729169689Skan __func__, error)); 73090075Sobrien if (isclean != 1) { 731169689Skan LOG_MSG(LVL_ERROR, "ALLOCATION TABLE CORRUPTED FOR %s " 732169689Skan "(%d sectors don't match, max %d allocations)", 73390075Sobrien sc->geom->name, count, 734169689Skan count * sc->me_per_sector); 73590075Sobrien } else { 736169689Skan LOG_MSG(LVL_INFO, "Allocation table ok for %s", 737169689Skan sc->geom->name); 73890075Sobrien } 73990075Sobrienbailout: 74090075Sobrien#endif 741169689Skan update_metadata(sc); 74290075Sobrien virstor_geom_destroy(sc, FALSE, FALSE); 74390075Sobrien exitval = EAGAIN; 744169689Skan } else 745169689Skan exitval = 0; 746169689Skan return (exitval); 747169689Skan} 748169689Skan 749169689Skan/* 750169689Skan * Taste event (per-class callback) 751169689Skan * Examines a provider and creates geom instances if needed 752169689Skan */ 753169689Skanstatic struct g_geom * 754169689Skang_virstor_taste(struct g_class *mp, struct g_provider *pp, int flags) 75590075Sobrien{ 756117395Skan struct g_virstor_metadata md; 75790075Sobrien struct g_geom *gp; 75890075Sobrien struct g_consumer *cp; 75990075Sobrien struct g_virstor_softc *sc; 76090075Sobrien int error; 76190075Sobrien 762169689Skan g_trace(G_T_TOPOLOGY, "%s(%s, %s)", __func__, mp->name, pp->name); 76390075Sobrien g_topology_assert(); 76490075Sobrien LOG_MSG(LVL_DEBUG, "Tasting %s", pp->name); 76590075Sobrien 766169689Skan /* We need a dummy geom to attach a consumer to the given provider */ 76790075Sobrien gp = g_new_geomf(mp, "virstor:taste.helper"); 76890075Sobrien gp->start = (void *)invalid_call; /* XXX: hacked up so the */ 76990075Sobrien gp->access = (void *)invalid_call; /* compiler doesn't complain. */ 77090075Sobrien gp->orphan = (void *)invalid_call; /* I really want these to fail. */ 771169689Skan 772169689Skan cp = g_new_consumer(gp); 773169689Skan g_attach(cp, pp); 77490075Sobrien error = read_metadata(cp, &md); 77590075Sobrien g_detach(cp); 77690075Sobrien g_destroy_consumer(cp); 77790075Sobrien g_destroy_geom(gp); 77890075Sobrien 77990075Sobrien if (error != 0) 78090075Sobrien return (NULL); 781169689Skan 782169689Skan if (strcmp(md.md_magic, G_VIRSTOR_MAGIC) != 0) 78390075Sobrien return (NULL); 784169689Skan if (md.md_version != G_VIRSTOR_VERSION) { 78590075Sobrien LOG_MSG(LVL_ERROR, "Kernel module version invalid " 786169689Skan "to handle %s (%s) : %d should be %d", 78790075Sobrien md.md_name, pp->name, md.md_version, G_VIRSTOR_VERSION); 78890075Sobrien return (NULL); 789169689Skan } 790169689Skan if (md.provsize != pp->mediasize) 79190075Sobrien return (NULL); 79290075Sobrien 793169689Skan /* If the provider name is hardcoded, use the offered provider only 79490075Sobrien * if it's been offered with its proper name (the one used in 79590075Sobrien * the label command). */ 79690075Sobrien if (md.provider[0] != '\0') { 797169689Skan if (strcmp(md.provider, pp->name) != 0) 798169689Skan return (NULL); 79990075Sobrien } 80090075Sobrien 801169689Skan /* Iterate all geoms this class already knows about to see if a new 80290075Sobrien * geom instance of this class needs to be created (in case the provider 803169689Skan * is first from a (possibly) multi-consumer geom) or it just needs 804169689Skan * to be added to an existing instance. */ 80590075Sobrien sc = NULL; 80690075Sobrien gp = NULL; 80790075Sobrien LIST_FOREACH(gp, &mp->geom, geom) { 80890075Sobrien sc = gp->softc; 80990075Sobrien if (sc == NULL) 81090075Sobrien continue; 81190075Sobrien if (strcmp(md.md_name, sc->geom->name) != 0) 81290075Sobrien continue; 813169689Skan if (md.md_id != sc->id) 81490075Sobrien continue; 815169689Skan break; 816169689Skan } 817169689Skan if (gp != NULL) { /* We found an existing geom instance; add to it */ 818169689Skan LOG_MSG(LVL_INFO, "Adding %s to %s", pp->name, md.md_name); 819169689Skan error = add_provider_to_geom(sc, pp, &md); 820169689Skan if (error != 0) { 821169689Skan LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)", 82290075Sobrien pp->name, md.md_name, error); 823169689Skan return (NULL); 824169689Skan } 825169689Skan } else { /* New geom instance needs to be created */ 826169689Skan gp = create_virstor_geom(mp, &md); 82790075Sobrien if (gp == NULL) { 828169689Skan LOG_MSG(LVL_ERROR, "Error creating new instance of " 82990075Sobrien "class %s: %s", mp->name, md.md_name); 83090075Sobrien LOG_MSG(LVL_DEBUG, "Error creating %s at %s", 83190075Sobrien md.md_name, pp->name); 83290075Sobrien return (NULL); 83390075Sobrien } 83490075Sobrien sc = gp->softc; 83590075Sobrien LOG_MSG(LVL_INFO, "Adding %s to %s (first found)", pp->name, 83690075Sobrien md.md_name); 83790075Sobrien error = add_provider_to_geom(sc, pp, &md); 83890075Sobrien if (error != 0) { 83990075Sobrien LOG_MSG(LVL_ERROR, "Error adding %s to %s (error %d)", 84090075Sobrien pp->name, md.md_name, error); 84190075Sobrien virstor_geom_destroy(sc, TRUE, FALSE); 84290075Sobrien return (NULL); 84390075Sobrien } 84490075Sobrien } 84590075Sobrien 84690075Sobrien return (gp); 84790075Sobrien} 84890075Sobrien 84990075Sobrien/* 85090075Sobrien * Destroyes consumer passed to it in arguments. Used as a callback 85190075Sobrien * on g_event queue. 85290075Sobrien */ 85390075Sobrienstatic void 85490075Sobriendelay_destroy_consumer(void *arg, int flags __unused) 85590075Sobrien{ 85690075Sobrien struct g_consumer *c = arg; 85790075Sobrien KASSERT(c != NULL, ("%s: invalid consumer", __func__)); 858117395Skan LOG_MSG(LVL_DEBUG, "Consumer %s destroyed with delay", 85990075Sobrien c->provider->name); 86090075Sobrien g_detach(c); 86190075Sobrien g_destroy_consumer(c); 862132718Skan} 86390075Sobrien 86490075Sobrien/* 865169689Skan * Remove a component (consumer) from geom instance; If it's the first 86690075Sobrien * component being removed, orphan the provider to announce geom's being 86790075Sobrien * dismantled 86890075Sobrien */ 869169689Skanstatic void 87090075Sobrienremove_component(struct g_virstor_softc *sc, struct g_virstor_component *comp, 87190075Sobrien boolean_t delay) 87290075Sobrien{ 87390075Sobrien struct g_consumer *c; 87490075Sobrien 87590075Sobrien KASSERT(comp->gcons != NULL, ("Component with no consumer in %s", 87690075Sobrien sc->geom->name)); 87790075Sobrien c = comp->gcons; 87890075Sobrien 87990075Sobrien comp->gcons = NULL; 88090075Sobrien KASSERT(c->provider != NULL, ("%s: no provider", __func__)); 88190075Sobrien LOG_MSG(LVL_DEBUG, "Component %s removed from %s", c->provider->name, 882132718Skan sc->geom->name); 88390075Sobrien if (sc->provider != NULL) { 88490075Sobrien /* Whither, GEOM? */ 885169689Skan sc->provider->flags |= G_PF_WITHER; 88690075Sobrien g_orphan_provider(sc->provider, ENXIO); 88790075Sobrien sc->provider = NULL; 88890075Sobrien LOG_MSG(LVL_INFO, "Removing provider %s", sc->geom->name); 88990075Sobrien } 89090075Sobrien 89190075Sobrien if (c->acr > 0 || c->acw > 0 || c->ace > 0) 89290075Sobrien g_access(c, -c->acr, -c->acw, -c->ace); 89390075Sobrien if (delay) { 894169689Skan /* Destroy consumer after it's tasted */ 895169689Skan g_post_event(delay_destroy_consumer, c, M_WAITOK, NULL); 89690075Sobrien } else { 89790075Sobrien g_detach(c); 89890075Sobrien g_destroy_consumer(c); 899169689Skan } 90090075Sobrien} 90190075Sobrien 90290075Sobrien/* 90390075Sobrien * Destroy geom - called internally 90490075Sobrien * See g_virstor_destroy_geom for the other one 90590075Sobrien */ 906169689Skanstatic int 907169689Skanvirstor_geom_destroy(struct g_virstor_softc *sc, boolean_t force, 90890075Sobrien boolean_t delay) 90990075Sobrien{ 91090075Sobrien struct g_provider *pp; 91190075Sobrien struct g_geom *gp; 912169689Skan int n; 913169689Skan 914117395Skan g_topology_assert(); 91590075Sobrien 91690075Sobrien if (sc == NULL) 91790075Sobrien return (ENXIO); 91890075Sobrien 91990075Sobrien pp = sc->provider; 92090075Sobrien if (pp != NULL && (pp->acr != 0 || pp->acw != 0 || pp->ace != 0)) { 92190075Sobrien LOG_MSG(force ? LVL_WARNING : LVL_ERROR, 92290075Sobrien "Device %s is still open.", pp->name); 923132718Skan if (!force) 92490075Sobrien return (EBUSY); 92590075Sobrien } 92690075Sobrien 92790075Sobrien for (n = 0; n < sc->n_components; n++) { 928117395Skan if (sc->components[n].gcons != NULL) 929117395Skan remove_component(sc, &sc->components[n], delay); 930117395Skan } 931117395Skan 932117395Skan gp = sc->geom; 933132718Skan gp->softc = NULL; 934132718Skan 935132718Skan KASSERT(sc->provider == NULL, ("Provider still exists for %s", 936117395Skan gp->name)); 937117395Skan 938117395Skan /* XXX: This might or might not work, since we're called with 939169689Skan * the topology lock held. Also, it might panic the kernel if 940117395Skan * the error'd BIO is in softupdates code. */ 941169689Skan mtx_lock(&sc->delayed_bio_q_mtx); 942169689Skan while (!STAILQ_EMPTY(&sc->delayed_bio_q)) { 943169689Skan struct g_virstor_bio_q *bq; 944169689Skan bq = STAILQ_FIRST(&sc->delayed_bio_q); 945169689Skan bq->bio->bio_error = ENOSPC; 946169689Skan g_io_deliver(bq->bio, EIO); 947169689Skan STAILQ_REMOVE_HEAD(&sc->delayed_bio_q, linkage); 948169689Skan free(bq, M_GVIRSTOR); 949169689Skan } 950169689Skan mtx_unlock(&sc->delayed_bio_q_mtx); 951169689Skan mtx_destroy(&sc->delayed_bio_q_mtx); 952169689Skan 953169689Skan free(sc->map, M_GVIRSTOR); 954169689Skan free(sc->components, M_GVIRSTOR); 955169689Skan bzero(sc, sizeof *sc); 956169689Skan free(sc, M_GVIRSTOR); 957169689Skan 958169689Skan pp = LIST_FIRST(&gp->provider); /* We only offer one provider */ 959169689Skan if (pp == NULL || (pp->acr == 0 && pp->acw == 0 && pp->ace == 0)) 960169689Skan LOG_MSG(LVL_DEBUG, "Device %s destroyed", gp->name); 961169689Skan 962169689Skan g_wither_geom(gp, ENXIO); 963169689Skan 964169689Skan return (0); 965169689Skan} 966169689Skan 967169689Skan/* 968169689Skan * Utility function: read metadata & decode. Wants topology lock to be 969169689Skan * held. 970169689Skan */ 971169689Skanstatic int 972169689Skanread_metadata(struct g_consumer *cp, struct g_virstor_metadata *md) 973169689Skan{ 974169689Skan struct g_provider *pp; 975169689Skan char *buf; 976117395Skan int error; 977169689Skan 978117395Skan g_topology_assert(); 979117395Skan error = g_access(cp, 1, 0, 0); 980117395Skan if (error != 0) 981169689Skan return (error); 982117395Skan pp = cp->provider; 983117395Skan g_topology_unlock(); 984169689Skan buf = g_read_data(cp, pp->mediasize - pp->sectorsize, pp->sectorsize, 985169689Skan &error); 986169689Skan g_topology_lock(); 987117395Skan g_access(cp, -1, 0, 0); 988169689Skan if (buf == NULL) 989169689Skan return (error); 990169689Skan 991117395Skan virstor_metadata_decode(buf, md); 992169689Skan g_free(buf); 993117395Skan 994169689Skan return (0); 995169689Skan} 996169689Skan 997117395Skan/** 998169689Skan * Utility function: encode & write metadata. Assumes topology lock is 999169689Skan * held. 1000169689Skan */ 1001117395Skanstatic int 1002117395Skanwrite_metadata(struct g_consumer *cp, struct g_virstor_metadata *md) 1003117395Skan{ 1004117395Skan struct g_provider *pp; 1005117395Skan char *buf; 1006169689Skan int error; 1007117395Skan 1008169689Skan KASSERT(cp != NULL && md != NULL && cp->provider != NULL, 1009169689Skan ("Something's fishy in %s", __func__)); 1010169689Skan LOG_MSG(LVL_DEBUG, "Writing metadata on %s", cp->provider->name); 1011117395Skan g_topology_assert(); 1012169689Skan error = g_access(cp, 0, 1, 0); 1013169689Skan if (error != 0) 1014169689Skan return (error); 1015169689Skan pp = cp->provider; 1016169689Skan 1017169689Skan buf = malloc(pp->sectorsize, M_GVIRSTOR, M_WAITOK); 1018169689Skan virstor_metadata_encode(md, buf); 1019169689Skan g_topology_unlock(); 1020169689Skan error = g_write_data(cp, pp->mediasize - pp->sectorsize, buf, 1021169689Skan pp->sectorsize); 1022169689Skan g_topology_lock(); 1023169689Skan g_access(cp, 0, -1, 0); 1024169689Skan 1025169689Skan free(buf, M_GVIRSTOR); 1026169689Skan return (0); 1027169689Skan} 1028169689Skan 1029169689Skan/* 1030169689Skan * Creates a new instance of this GEOM class, initialise softc 1031169689Skan */ 1032169689Skanstatic struct g_geom * 1033169689Skancreate_virstor_geom(struct g_class *mp, struct g_virstor_metadata *md) 1034169689Skan{ 1035169689Skan struct g_geom *gp; 1036169689Skan struct g_virstor_softc *sc; 1037169689Skan 1038169689Skan LOG_MSG(LVL_DEBUG, "Creating geom instance for %s (id=%u)", 1039169689Skan md->md_name, md->md_id); 1040169689Skan 1041169689Skan if (md->md_count < 1 || md->md_chunk_size < 1 || 1042169689Skan md->md_virsize < md->md_chunk_size) { 1043169689Skan /* This is bogus configuration, and probably means data is 1044169689Skan * somehow corrupted. Panic, maybe? */ 1045169689Skan LOG_MSG(LVL_ERROR, "Nonsensical metadata information for %s", 1046169689Skan md->md_name); 1047169689Skan return (NULL); 1048169689Skan } 1049169689Skan 1050169689Skan /* Check if it's already created */ 1051169689Skan LIST_FOREACH(gp, &mp->geom, geom) { 1052169689Skan sc = gp->softc; 1053169689Skan if (sc != NULL && strcmp(sc->geom->name, md->md_name) == 0) { 1054169689Skan LOG_MSG(LVL_WARNING, "Geom %s already exists", 1055169689Skan md->md_name); 1056169689Skan if (sc->id != md->md_id) { 1057169689Skan LOG_MSG(LVL_ERROR, 1058169689Skan "Some stale or invalid components " 1059169689Skan "exist for virstor device named %s. " 1060169689Skan "You will need to <CLEAR> all stale " 1061169689Skan "components and maybe reconfigure " 1062169689Skan "the virstor device. Tune " 1063169689Skan "kern.geom.virstor.debug sysctl up " 1064169689Skan "for more information.", 1065169689Skan sc->geom->name); 1066169689Skan } 1067169689Skan return (NULL); 1068169689Skan } 1069169689Skan } 1070169689Skan gp = g_new_geomf(mp, "%s", md->md_name); 1071169689Skan gp->softc = NULL; /* to circumevent races that test softc */ 1072169689Skan 1073169689Skan gp->start = g_virstor_start; 1074169689Skan gp->spoiled = g_virstor_orphan; 1075169689Skan gp->orphan = g_virstor_orphan; 1076169689Skan gp->access = g_virstor_access; 1077169689Skan gp->dumpconf = g_virstor_dumpconf; 1078169689Skan 1079169689Skan sc = malloc(sizeof(*sc), M_GVIRSTOR, M_WAITOK | M_ZERO); 1080 sc->id = md->md_id; 1081 sc->n_components = md->md_count; 1082 sc->components = malloc(sizeof(struct g_virstor_component) * md->md_count, 1083 M_GVIRSTOR, M_WAITOK | M_ZERO); 1084 sc->chunk_size = md->md_chunk_size; 1085 sc->virsize = md->md_virsize; 1086 STAILQ_INIT(&sc->delayed_bio_q); 1087 mtx_init(&sc->delayed_bio_q_mtx, "gvirstor_delayed_bio_q_mtx", 1088 "gvirstor", MTX_DEF | MTX_RECURSE); 1089 1090 sc->geom = gp; 1091 sc->provider = NULL; /* virstor_check_and_run will create it */ 1092 gp->softc = sc; 1093 1094 LOG_MSG(LVL_ANNOUNCE, "Device %s created", sc->geom->name); 1095 1096 return (gp); 1097} 1098 1099/* 1100 * Add provider to a GEOM class instance 1101 */ 1102static int 1103add_provider_to_geom(struct g_virstor_softc *sc, struct g_provider *pp, 1104 struct g_virstor_metadata *md) 1105{ 1106 struct g_virstor_component *component; 1107 struct g_consumer *cp, *fcp; 1108 struct g_geom *gp; 1109 int error; 1110 1111 if (md->no >= sc->n_components) 1112 return (EINVAL); 1113 1114 /* "Current" compontent */ 1115 component = &(sc->components[md->no]); 1116 if (component->gcons != NULL) 1117 return (EEXIST); 1118 1119 gp = sc->geom; 1120 fcp = LIST_FIRST(&gp->consumer); 1121 1122 cp = g_new_consumer(gp); 1123 error = g_attach(cp, pp); 1124 1125 if (error != 0) { 1126 g_destroy_consumer(cp); 1127 return (error); 1128 } 1129 1130 if (fcp != NULL) { 1131 if (fcp->provider->sectorsize != pp->sectorsize) { 1132 /* TODO: this can be made to work */ 1133 LOG_MSG(LVL_ERROR, "Provider %s of %s has invalid " 1134 "sector size (%d)", pp->name, sc->geom->name, 1135 pp->sectorsize); 1136 return (EINVAL); 1137 } 1138 if (fcp->acr > 0 || fcp->acw || fcp->ace > 0) { 1139 /* Replicate access permissions from first "live" consumer 1140 * to the new one */ 1141 error = g_access(cp, fcp->acr, fcp->acw, fcp->ace); 1142 if (error != 0) { 1143 g_detach(cp); 1144 g_destroy_consumer(cp); 1145 return (error); 1146 } 1147 } 1148 } 1149 1150 /* Bring up a new component */ 1151 cp->private = component; 1152 component->gcons = cp; 1153 component->sc = sc; 1154 component->index = md->no; 1155 component->chunk_count = md->chunk_count; 1156 component->chunk_next = md->chunk_next; 1157 component->chunk_reserved = md->chunk_reserved; 1158 component->flags = md->flags; 1159 1160 LOG_MSG(LVL_DEBUG, "%s attached to %s", pp->name, sc->geom->name); 1161 1162 virstor_check_and_run(sc); 1163 return (0); 1164} 1165 1166/* 1167 * Check if everything's ready to create the geom provider & device entry, 1168 * create and start provider. 1169 * Called ultimately by .taste, from g_event thread 1170 */ 1171static void 1172virstor_check_and_run(struct g_virstor_softc *sc) 1173{ 1174 off_t off; 1175 size_t n, count; 1176 int index; 1177 int error; 1178 1179 if (virstor_valid_components(sc) != sc->n_components) 1180 return; 1181 1182 if (virstor_valid_components(sc) == 0) { 1183 /* This is actually a candidate for panic() */ 1184 LOG_MSG(LVL_ERROR, "No valid components for %s?", 1185 sc->provider->name); 1186 return; 1187 } 1188 1189 sc->sectorsize = sc->components[0].gcons->provider->sectorsize; 1190 1191 /* Initialise allocation map from the first consumer */ 1192 sc->chunk_count = sc->virsize / sc->chunk_size; 1193 if (sc->chunk_count * (off_t)sc->chunk_size != sc->virsize) { 1194 LOG_MSG(LVL_WARNING, "Device %s truncated to %ju bytes", 1195 sc->provider->name, 1196 sc->chunk_count * (off_t)sc->chunk_size); 1197 } 1198 sc->map_size = sc->chunk_count * sizeof *(sc->map); 1199 /* The following allocation is in order of 4MB - 8MB */ 1200 sc->map = malloc(sc->map_size, M_GVIRSTOR, M_WAITOK); 1201 KASSERT(sc->map != NULL, ("%s: Memory allocation error (%zu bytes) for %s", 1202 __func__, sc->map_size, sc->provider->name)); 1203 sc->map_sectors = sc->map_size / sc->sectorsize; 1204 1205 count = 0; 1206 for (n = 0; n < sc->n_components; n++) 1207 count += sc->components[n].chunk_count; 1208 LOG_MSG(LVL_INFO, "Device %s has %zu physical chunks and %zu virtual " 1209 "(%zu KB chunks)", 1210 sc->geom->name, count, sc->chunk_count, sc->chunk_size / 1024); 1211 1212 error = g_access(sc->components[0].gcons, 1, 0, 0); 1213 if (error != 0) { 1214 LOG_MSG(LVL_ERROR, "Cannot acquire read access for %s to " 1215 "read allocation map for %s", 1216 sc->components[0].gcons->provider->name, 1217 sc->geom->name); 1218 return; 1219 } 1220 /* Read in the allocation map */ 1221 LOG_MSG(LVL_DEBUG, "Reading map for %s from %s", sc->geom->name, 1222 sc->components[0].gcons->provider->name); 1223 off = count = n = 0; 1224 while (count < sc->map_size) { 1225 struct g_virstor_map_entry *mapbuf; 1226 size_t bs; 1227 1228 bs = MIN(MAXPHYS, sc->map_size - count); 1229 if (bs % sc->sectorsize != 0) { 1230 /* Check for alignment errors */ 1231 bs = (bs / sc->sectorsize) * sc->sectorsize; 1232 if (bs == 0) 1233 break; 1234 LOG_MSG(LVL_ERROR, "Trouble: map is not sector-aligned " 1235 "for %s on %s", sc->geom->name, 1236 sc->components[0].gcons->provider->name); 1237 } 1238 mapbuf = g_read_data(sc->components[0].gcons, off, bs, &error); 1239 if (mapbuf == NULL) { 1240 free(sc->map, M_GVIRSTOR); 1241 LOG_MSG(LVL_ERROR, "Error reading allocation map " 1242 "for %s from %s (offset %ju) (error %d)", 1243 sc->geom->name, 1244 sc->components[0].gcons->provider->name, 1245 off, error); 1246 return; 1247 } 1248 1249 bcopy(mapbuf, &sc->map[n], bs); 1250 off += bs; 1251 count += bs; 1252 n += bs / sizeof *(sc->map); 1253 g_free(mapbuf); 1254 } 1255 g_access(sc->components[0].gcons, -1, 0, 0); 1256 LOG_MSG(LVL_DEBUG, "Read map for %s", sc->geom->name); 1257 1258 /* find first component with allocatable chunks */ 1259 index = -1; 1260 for (n = 0; n < sc->n_components; n++) { 1261 if (sc->components[n].chunk_next < 1262 sc->components[n].chunk_count) { 1263 index = n; 1264 break; 1265 } 1266 } 1267 if (index == -1) 1268 /* not found? set it to the last component and handle it 1269 * later */ 1270 index = sc->n_components - 1; 1271 1272 if (index >= sc->n_components - g_virstor_component_watermark - 1) { 1273 LOG_MSG(LVL_WARNING, "Device %s running out of components " 1274 "(%d/%u: %s)", sc->geom->name, 1275 index+1, 1276 sc->n_components, 1277 sc->components[index].gcons->provider->name); 1278 } 1279 sc->curr_component = index; 1280 1281 if (sc->components[index].chunk_next >= 1282 sc->components[index].chunk_count - g_virstor_chunk_watermark) { 1283 LOG_MSG(LVL_WARNING, 1284 "Component %s of %s is running out of free space " 1285 "(%u chunks left)", 1286 sc->components[index].gcons->provider->name, 1287 sc->geom->name, sc->components[index].chunk_count - 1288 sc->components[index].chunk_next); 1289 } 1290 1291 sc->me_per_sector = sc->sectorsize / sizeof *(sc->map); 1292 if (sc->sectorsize % sizeof *(sc->map) != 0) { 1293 LOG_MSG(LVL_ERROR, 1294 "%s: Map entries don't fit exactly in a sector (%s)", 1295 __func__, sc->geom->name); 1296 return; 1297 } 1298 1299 /* Recalculate allocated chunks in components & at the same time 1300 * verify map data is sane. We could trust metadata on this, but 1301 * we want to make sure. */ 1302 for (n = 0; n < sc->n_components; n++) 1303 sc->components[n].chunk_next = sc->components[n].chunk_reserved; 1304 1305 for (n = 0; n < sc->chunk_count; n++) { 1306 if (sc->map[n].provider_no >= sc->n_components || 1307 sc->map[n].provider_chunk >= 1308 sc->components[sc->map[n].provider_no].chunk_count) { 1309 LOG_MSG(LVL_ERROR, "%s: Invalid entry %u in map for %s", 1310 __func__, (u_int)n, sc->geom->name); 1311 LOG_MSG(LVL_ERROR, "%s: provider_no: %u, n_components: %u" 1312 " provider_chunk: %u, chunk_count: %u", __func__, 1313 sc->map[n].provider_no, sc->n_components, 1314 sc->map[n].provider_chunk, 1315 sc->components[sc->map[n].provider_no].chunk_count); 1316 return; 1317 } 1318 if (sc->map[n].flags & VIRSTOR_MAP_ALLOCATED) 1319 sc->components[sc->map[n].provider_no].chunk_next++; 1320 } 1321 1322 sc->provider = g_new_providerf(sc->geom, "virstor/%s", 1323 sc->geom->name); 1324 1325 sc->provider->sectorsize = sc->sectorsize; 1326 sc->provider->mediasize = sc->virsize; 1327 g_error_provider(sc->provider, 0); 1328 1329 LOG_MSG(LVL_INFO, "%s activated", sc->provider->name); 1330 LOG_MSG(LVL_DEBUG, "%s starting with current component %u, starting " 1331 "chunk %u", sc->provider->name, sc->curr_component, 1332 sc->components[sc->curr_component].chunk_next); 1333} 1334 1335/* 1336 * Returns count of active providers in this geom instance 1337 */ 1338static u_int 1339virstor_valid_components(struct g_virstor_softc *sc) 1340{ 1341 unsigned int nc, i; 1342 1343 nc = 0; 1344 KASSERT(sc != NULL, ("%s: softc is NULL", __func__)); 1345 KASSERT(sc->components != NULL, ("%s: sc->components is NULL", __func__)); 1346 for (i = 0; i < sc->n_components; i++) 1347 if (sc->components[i].gcons != NULL) 1348 nc++; 1349 return (nc); 1350} 1351 1352/* 1353 * Called when the consumer gets orphaned (?) 1354 */ 1355static void 1356g_virstor_orphan(struct g_consumer *cp) 1357{ 1358 struct g_virstor_softc *sc; 1359 struct g_virstor_component *comp; 1360 struct g_geom *gp; 1361 1362 g_topology_assert(); 1363 gp = cp->geom; 1364 sc = gp->softc; 1365 if (sc == NULL) 1366 return; 1367 1368 comp = cp->private; 1369 KASSERT(comp != NULL, ("%s: No component in private part of consumer", 1370 __func__)); 1371 remove_component(sc, comp, FALSE); 1372 if (virstor_valid_components(sc) == 0) 1373 virstor_geom_destroy(sc, TRUE, FALSE); 1374} 1375 1376/* 1377 * Called to notify geom when it's been opened, and for what intent 1378 */ 1379static int 1380g_virstor_access(struct g_provider *pp, int dr, int dw, int de) 1381{ 1382 struct g_consumer *c; 1383 struct g_virstor_softc *sc; 1384 struct g_geom *gp; 1385 int error; 1386 1387 KASSERT(pp != NULL, ("%s: NULL provider", __func__)); 1388 gp = pp->geom; 1389 KASSERT(gp != NULL, ("%s: NULL geom", __func__)); 1390 sc = gp->softc; 1391 1392 if (sc == NULL) { 1393 /* It seems that .access can be called with negative dr,dw,dx 1394 * in this case but I want to check for myself */ 1395 LOG_MSG(LVL_WARNING, "access(%d, %d, %d) for %s", 1396 dr, dw, de, pp->name); 1397 /* This should only happen when geom is withered so 1398 * allow only negative requests */ 1399 KASSERT(dr <= 0 && dw <= 0 && de <= 0, 1400 ("%s: Positive access for %s", __func__, pp->name)); 1401 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) 1402 LOG_MSG(LVL_DEBUG, "Device %s definitely destroyed", 1403 pp->name); 1404 return (0); 1405 } 1406 1407 /* Grab an exclusive bit to propagate on our consumers on first open */ 1408 if (pp->acr == 0 && pp->acw == 0 && pp->ace == 0) 1409 de++; 1410 /* ... drop it on close */ 1411 if (pp->acr + dr == 0 && pp->acw + dw == 0 && pp->ace + de == 0) { 1412 de--; 1413 update_metadata(sc); /* Writes statistical information */ 1414 } 1415 1416 error = ENXIO; 1417 LIST_FOREACH(c, &gp->consumer, consumer) { 1418 KASSERT(c != NULL, ("%s: consumer is NULL", __func__)); 1419 error = g_access(c, dr, dw, de); 1420 if (error != 0) { 1421 struct g_consumer *c2; 1422 1423 /* Backout earlier changes */ 1424 LIST_FOREACH(c2, &gp->consumer, consumer) { 1425 if (c2 == c) /* all eariler components fixed */ 1426 return (error); 1427 g_access(c2, -dr, -dw, -de); 1428 } 1429 } 1430 } 1431 1432 return (error); 1433} 1434 1435/* 1436 * Generate XML dump of current state 1437 */ 1438static void 1439g_virstor_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp, 1440 struct g_consumer *cp, struct g_provider *pp) 1441{ 1442 struct g_virstor_softc *sc; 1443 1444 g_topology_assert(); 1445 sc = gp->softc; 1446 1447 if (sc == NULL || pp != NULL) 1448 return; 1449 1450 if (cp != NULL) { 1451 /* For each component */ 1452 struct g_virstor_component *comp; 1453 1454 comp = cp->private; 1455 if (comp == NULL) 1456 return; 1457 sbuf_printf(sb, "%s<ComponentIndex>%u</ComponentIndex>\n", 1458 indent, comp->index); 1459 sbuf_printf(sb, "%s<ChunkCount>%u</ChunkCount>\n", 1460 indent, comp->chunk_count); 1461 sbuf_printf(sb, "%s<ChunksUsed>%u</ChunksUsed>\n", 1462 indent, comp->chunk_next); 1463 sbuf_printf(sb, "%s<ChunksReserved>%u</ChunksReserved>\n", 1464 indent, comp->chunk_reserved); 1465 sbuf_printf(sb, "%s<StorageFree>%u%%</StorageFree>\n", 1466 indent, 1467 comp->chunk_next > 0 ? 100 - 1468 ((comp->chunk_next + comp->chunk_reserved) * 100) / 1469 comp->chunk_count : 100); 1470 } else { 1471 /* For the whole thing */ 1472 u_int count, used, i; 1473 off_t size; 1474 1475 count = used = size = 0; 1476 for (i = 0; i < sc->n_components; i++) { 1477 if (sc->components[i].gcons != NULL) { 1478 count += sc->components[i].chunk_count; 1479 used += sc->components[i].chunk_next + 1480 sc->components[i].chunk_reserved; 1481 size += sc->components[i].gcons-> 1482 provider->mediasize; 1483 } 1484 } 1485 1486 sbuf_printf(sb, "%s<Status>" 1487 "Components=%u, Online=%u</Status>\n", indent, 1488 sc->n_components, virstor_valid_components(sc)); 1489 sbuf_printf(sb, "%s<State>%u%% physical free</State>\n", 1490 indent, 100-(used * 100) / count); 1491 sbuf_printf(sb, "%s<ChunkSize>%zu</ChunkSize>\n", indent, 1492 sc->chunk_size); 1493 sbuf_printf(sb, "%s<PhysicalFree>%u%%</PhysicalFree>\n", 1494 indent, used > 0 ? 100 - (used * 100) / count : 100); 1495 sbuf_printf(sb, "%s<ChunkPhysicalCount>%u</ChunkPhysicalCount>\n", 1496 indent, count); 1497 sbuf_printf(sb, "%s<ChunkVirtualCount>%zu</ChunkVirtualCount>\n", 1498 indent, sc->chunk_count); 1499 sbuf_printf(sb, "%s<PhysicalBacking>%zu%%</PhysicalBacking>\n", 1500 indent, 1501 (count * 100) / sc->chunk_count); 1502 sbuf_printf(sb, "%s<PhysicalBackingSize>%jd</PhysicalBackingSize>\n", 1503 indent, size); 1504 sbuf_printf(sb, "%s<VirtualSize>%jd</VirtualSize>\n", indent, 1505 sc->virsize); 1506 } 1507} 1508 1509/* 1510 * GEOM .done handler 1511 * Can't use standard handler because one requested IO may 1512 * fork into additional data IOs 1513 */ 1514static void 1515g_virstor_done(struct bio *b) 1516{ 1517 struct g_virstor_softc *sc; 1518 struct bio *parent_b; 1519 1520 parent_b = b->bio_parent; 1521 sc = parent_b->bio_to->geom->softc; 1522 1523 if (b->bio_error != 0) { 1524 LOG_MSG(LVL_ERROR, "Error %d for offset=%ju, length=%ju, %s", 1525 b->bio_error, b->bio_offset, b->bio_length, 1526 b->bio_to->name); 1527 if (parent_b->bio_error == 0) 1528 parent_b->bio_error = b->bio_error; 1529 } 1530 1531 parent_b->bio_inbed++; 1532 parent_b->bio_completed += b->bio_completed; 1533 1534 if (parent_b->bio_children == parent_b->bio_inbed) { 1535 parent_b->bio_completed = parent_b->bio_length; 1536 g_io_deliver(parent_b, parent_b->bio_error); 1537 } 1538 g_destroy_bio(b); 1539} 1540 1541/* 1542 * I/O starts here 1543 * Called in g_down thread 1544 */ 1545static void 1546g_virstor_start(struct bio *b) 1547{ 1548 struct g_virstor_softc *sc; 1549 struct g_virstor_component *comp; 1550 struct bio *cb; 1551 struct g_provider *pp; 1552 char *addr; 1553 off_t offset, length; 1554 struct bio_queue_head bq; 1555 size_t chunk_size; /* cached for convenience */ 1556 u_int count; 1557 1558 pp = b->bio_to; 1559 sc = pp->geom->softc; 1560 KASSERT(sc != NULL, ("%s: no softc (error=%d, device=%s)", __func__, 1561 b->bio_to->error, b->bio_to->name)); 1562 1563 LOG_REQ(LVL_MOREDEBUG, b, "%s", __func__); 1564 1565 switch (b->bio_cmd) { 1566 case BIO_READ: 1567 case BIO_WRITE: 1568 case BIO_DELETE: 1569 break; 1570 default: 1571 g_io_deliver(b, EOPNOTSUPP); 1572 return; 1573 } 1574 1575 LOG_MSG(LVL_DEBUG2, "BIO arrived, size=%ju", b->bio_length); 1576 bioq_init(&bq); 1577 1578 chunk_size = sc->chunk_size; 1579 addr = b->bio_data; 1580 offset = b->bio_offset; /* virtual offset and length */ 1581 length = b->bio_length; 1582 1583 while (length > 0) { 1584 size_t chunk_index, in_chunk_offset, in_chunk_length; 1585 struct virstor_map_entry *me; 1586 1587 chunk_index = offset / chunk_size; /* round downwards */ 1588 in_chunk_offset = offset % chunk_size; 1589 in_chunk_length = min(length, chunk_size - in_chunk_offset); 1590 LOG_MSG(LVL_DEBUG, "Mapped %s(%ju, %ju) to (%zu,%zu,%zu)", 1591 b->bio_cmd == BIO_READ ? "R" : "W", 1592 offset, length, 1593 chunk_index, in_chunk_offset, in_chunk_length); 1594 me = &sc->map[chunk_index]; 1595 1596 if (b->bio_cmd == BIO_READ || b->bio_cmd == BIO_DELETE) { 1597 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) { 1598 /* Reads from unallocated chunks return zeroed 1599 * buffers */ 1600 if (b->bio_cmd == BIO_READ) 1601 bzero(addr, in_chunk_length); 1602 } else { 1603 comp = &sc->components[me->provider_no]; 1604 1605 cb = g_clone_bio(b); 1606 if (cb == NULL) { 1607 bioq_dismantle(&bq); 1608 if (b->bio_error == 0) 1609 b->bio_error = ENOMEM; 1610 g_io_deliver(b, b->bio_error); 1611 return; 1612 } 1613 cb->bio_to = comp->gcons->provider; 1614 cb->bio_done = g_virstor_done; 1615 cb->bio_offset = 1616 (off_t)me->provider_chunk * (off_t)chunk_size 1617 + in_chunk_offset; 1618 cb->bio_length = in_chunk_length; 1619 cb->bio_data = addr; 1620 cb->bio_caller1 = comp; 1621 bioq_disksort(&bq, cb); 1622 } 1623 } else { /* handle BIO_WRITE */ 1624 KASSERT(b->bio_cmd == BIO_WRITE, 1625 ("%s: Unknown command %d", __func__, 1626 b->bio_cmd)); 1627 1628 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) { 1629 /* We have a virtual chunk, represented by 1630 * the "me" entry, but it's not yet allocated 1631 * (tied to) a physical chunk. So do it now. */ 1632 struct virstor_map_entry *data_me; 1633 u_int phys_chunk, comp_no; 1634 off_t s_offset; 1635 int error; 1636 1637 error = allocate_chunk(sc, &comp, &comp_no, 1638 &phys_chunk); 1639 if (error != 0) { 1640 /* We cannot allocate a physical chunk 1641 * to satisfy this request, so we'll 1642 * delay it to when we can... 1643 * XXX: this will prevent the fs from 1644 * being umounted! */ 1645 struct g_virstor_bio_q *biq; 1646 biq = malloc(sizeof *biq, M_GVIRSTOR, 1647 M_NOWAIT); 1648 if (biq == NULL) { 1649 bioq_dismantle(&bq); 1650 if (b->bio_error == 0) 1651 b->bio_error = ENOMEM; 1652 g_io_deliver(b, b->bio_error); 1653 return; 1654 } 1655 biq->bio = b; 1656 mtx_lock(&sc->delayed_bio_q_mtx); 1657 STAILQ_INSERT_TAIL(&sc->delayed_bio_q, 1658 biq, linkage); 1659 mtx_unlock(&sc->delayed_bio_q_mtx); 1660 LOG_MSG(LVL_WARNING, "Delaying BIO " 1661 "(size=%ju) until free physical " 1662 "space can be found on %s", 1663 b->bio_length, 1664 sc->provider->name); 1665 return; 1666 } 1667 LOG_MSG(LVL_DEBUG, "Allocated chunk %u on %s " 1668 "for %s", 1669 phys_chunk, 1670 comp->gcons->provider->name, 1671 sc->provider->name); 1672 1673 me->provider_no = comp_no; 1674 me->provider_chunk = phys_chunk; 1675 me->flags |= VIRSTOR_MAP_ALLOCATED; 1676 1677 cb = g_clone_bio(b); 1678 if (cb == NULL) { 1679 me->flags &= ~VIRSTOR_MAP_ALLOCATED; 1680 me->provider_no = 0; 1681 me->provider_chunk = 0; 1682 bioq_dismantle(&bq); 1683 if (b->bio_error == 0) 1684 b->bio_error = ENOMEM; 1685 g_io_deliver(b, b->bio_error); 1686 return; 1687 } 1688 1689 /* The allocation table is stored continuously 1690 * at the start of the drive. We need to 1691 * calculate the offset of the sector that holds 1692 * this map entry both on the drive and in the 1693 * map array. 1694 * sc_offset will end up pointing to the drive 1695 * sector. */ 1696 s_offset = chunk_index * sizeof *me; 1697 s_offset = (s_offset / sc->sectorsize) * 1698 sc->sectorsize; 1699 1700 /* data_me points to map entry sector 1701 * in memory (analoguos to offset) */ 1702 data_me = &sc->map[(chunk_index / 1703 sc->me_per_sector) * sc->me_per_sector]; 1704 1705 /* Commit sector with map entry to storage */ 1706 cb->bio_to = sc->components[0].gcons->provider; 1707 cb->bio_done = g_virstor_done; 1708 cb->bio_offset = s_offset; 1709 cb->bio_data = (char *)data_me; 1710 cb->bio_length = sc->sectorsize; 1711 cb->bio_caller1 = &sc->components[0]; 1712 bioq_disksort(&bq, cb); 1713 } 1714 1715 comp = &sc->components[me->provider_no]; 1716 cb = g_clone_bio(b); 1717 if (cb == NULL) { 1718 bioq_dismantle(&bq); 1719 if (b->bio_error == 0) 1720 b->bio_error = ENOMEM; 1721 g_io_deliver(b, b->bio_error); 1722 return; 1723 } 1724 /* Finally, handle the data */ 1725 cb->bio_to = comp->gcons->provider; 1726 cb->bio_done = g_virstor_done; 1727 cb->bio_offset = (off_t)me->provider_chunk*(off_t)chunk_size + 1728 in_chunk_offset; 1729 cb->bio_length = in_chunk_length; 1730 cb->bio_data = addr; 1731 cb->bio_caller1 = comp; 1732 bioq_disksort(&bq, cb); 1733 } 1734 addr += in_chunk_length; 1735 length -= in_chunk_length; 1736 offset += in_chunk_length; 1737 } 1738 1739 /* Fire off bio's here */ 1740 count = 0; 1741 for (cb = bioq_first(&bq); cb != NULL; cb = bioq_first(&bq)) { 1742 bioq_remove(&bq, cb); 1743 LOG_REQ(LVL_MOREDEBUG, cb, "Firing request"); 1744 comp = cb->bio_caller1; 1745 cb->bio_caller1 = NULL; 1746 LOG_MSG(LVL_DEBUG, " firing bio, offset=%ju, length=%ju", 1747 cb->bio_offset, cb->bio_length); 1748 g_io_request(cb, comp->gcons); 1749 count++; 1750 } 1751 if (count == 0) { /* We handled everything locally */ 1752 b->bio_completed = b->bio_length; 1753 g_io_deliver(b, 0); 1754 } 1755 1756} 1757 1758/* 1759 * Allocate a chunk from a physical provider. Returns physical component, 1760 * chunk index relative to the component and the component's index. 1761 */ 1762static int 1763allocate_chunk(struct g_virstor_softc *sc, struct g_virstor_component **comp, 1764 u_int *comp_no_p, u_int *chunk) 1765{ 1766 u_int comp_no; 1767 1768 KASSERT(sc->curr_component < sc->n_components, 1769 ("%s: Invalid curr_component: %u", __func__, sc->curr_component)); 1770 1771 comp_no = sc->curr_component; 1772 *comp = &sc->components[comp_no]; 1773 dump_component(*comp); 1774 if ((*comp)->chunk_next >= (*comp)->chunk_count) { 1775 /* This component is full. Allocate next component */ 1776 if (comp_no >= sc->n_components-1) { 1777 LOG_MSG(LVL_ERROR, "All physical space allocated for %s", 1778 sc->geom->name); 1779 return (-1); 1780 } 1781 (*comp)->flags &= ~VIRSTOR_PROVIDER_CURRENT; 1782 sc->curr_component = ++comp_no; 1783 1784 *comp = &sc->components[comp_no]; 1785 if (comp_no >= sc->n_components - g_virstor_component_watermark-1) 1786 LOG_MSG(LVL_WARNING, "Device %s running out of components " 1787 "(switching to %u/%u: %s)", sc->geom->name, 1788 comp_no+1, sc->n_components, 1789 (*comp)->gcons->provider->name); 1790 /* Take care not to overwrite reserved chunks */ 1791 if ( (*comp)->chunk_reserved > 0 && 1792 (*comp)->chunk_next < (*comp)->chunk_reserved) 1793 (*comp)->chunk_next = (*comp)->chunk_reserved; 1794 1795 (*comp)->flags |= 1796 VIRSTOR_PROVIDER_ALLOCATED | VIRSTOR_PROVIDER_CURRENT; 1797 dump_component(*comp); 1798 *comp_no_p = comp_no; 1799 *chunk = (*comp)->chunk_next++; 1800 } else { 1801 *comp_no_p = comp_no; 1802 *chunk = (*comp)->chunk_next++; 1803 } 1804 return (0); 1805} 1806 1807/* Dump a component */ 1808static void 1809dump_component(struct g_virstor_component *comp) 1810{ 1811 1812 if (g_virstor_debug < LVL_DEBUG2) 1813 return; 1814 printf("Component %d: %s\n", comp->index, comp->gcons->provider->name); 1815 printf(" chunk_count: %u\n", comp->chunk_count); 1816 printf(" chunk_next: %u\n", comp->chunk_next); 1817 printf(" flags: %u\n", comp->flags); 1818} 1819 1820#if 0 1821/* Dump a map entry */ 1822static void 1823dump_me(struct virstor_map_entry *me, unsigned int nr) 1824{ 1825 if (g_virstor_debug < LVL_DEBUG) 1826 return; 1827 printf("VIRT. CHUNK #%d: ", nr); 1828 if ((me->flags & VIRSTOR_MAP_ALLOCATED) == 0) 1829 printf("(unallocated)\n"); 1830 else 1831 printf("allocated at provider %u, provider_chunk %u\n", 1832 me->provider_no, me->provider_chunk); 1833} 1834#endif 1835 1836/* 1837 * Dismantle bio_queue and destroy its components 1838 */ 1839static void 1840bioq_dismantle(struct bio_queue_head *bq) 1841{ 1842 struct bio *b; 1843 1844 for (b = bioq_first(bq); b != NULL; b = bioq_first(bq)) { 1845 bioq_remove(bq, b); 1846 g_destroy_bio(b); 1847 } 1848} 1849 1850/* 1851 * The function that shouldn't be called. 1852 * When this is called, the stack is already garbled because of 1853 * argument mismatch. There's nothing to do now but panic, which is 1854 * accidentally the whole purpose of this function. 1855 * Motivation: to guard from accidentally calling geom methods when 1856 * they shouldn't be called. (see g_..._taste) 1857 */ 1858static void 1859invalid_call(void) 1860{ 1861 panic("invalid_call() has just been called. Something's fishy here."); 1862} 1863 1864DECLARE_GEOM_CLASS(g_virstor_class, g_virstor); /* Let there be light */ 1865