primary.c revision 205738
1204076Spjd/*- 2204076Spjd * Copyright (c) 2009 The FreeBSD Foundation 3204076Spjd * All rights reserved. 4204076Spjd * 5204076Spjd * This software was developed by Pawel Jakub Dawidek under sponsorship from 6204076Spjd * the FreeBSD Foundation. 7204076Spjd * 8204076Spjd * Redistribution and use in source and binary forms, with or without 9204076Spjd * modification, are permitted provided that the following conditions 10204076Spjd * are met: 11204076Spjd * 1. Redistributions of source code must retain the above copyright 12204076Spjd * notice, this list of conditions and the following disclaimer. 13204076Spjd * 2. Redistributions in binary form must reproduce the above copyright 14204076Spjd * notice, this list of conditions and the following disclaimer in the 15204076Spjd * documentation and/or other materials provided with the distribution. 16204076Spjd * 17204076Spjd * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 18204076Spjd * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19204076Spjd * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20204076Spjd * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 21204076Spjd * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22204076Spjd * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23204076Spjd * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24204076Spjd * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25204076Spjd * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26204076Spjd * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27204076Spjd * SUCH DAMAGE. 28204076Spjd */ 29204076Spjd 30204076Spjd#include <sys/cdefs.h> 31204076Spjd__FBSDID("$FreeBSD: head/sbin/hastd/primary.c 205738 2010-03-27 16:35:07Z pjd $"); 32204076Spjd 33204076Spjd#include <sys/types.h> 34204076Spjd#include <sys/time.h> 35204076Spjd#include <sys/bio.h> 36204076Spjd#include <sys/disk.h> 37204076Spjd#include <sys/refcount.h> 38204076Spjd#include <sys/stat.h> 39204076Spjd 40204076Spjd#include <geom/gate/g_gate.h> 41204076Spjd 42204076Spjd#include <assert.h> 43204076Spjd#include <err.h> 44204076Spjd#include <errno.h> 45204076Spjd#include <fcntl.h> 46204076Spjd#include <libgeom.h> 47204076Spjd#include <pthread.h> 48204076Spjd#include <stdint.h> 49204076Spjd#include <stdio.h> 50204076Spjd#include <string.h> 51204076Spjd#include <sysexits.h> 52204076Spjd#include <unistd.h> 53204076Spjd 54204076Spjd#include <activemap.h> 55204076Spjd#include <nv.h> 56204076Spjd#include <rangelock.h> 57204076Spjd 58204076Spjd#include "control.h" 59204076Spjd#include "hast.h" 60204076Spjd#include "hast_proto.h" 61204076Spjd#include "hastd.h" 62204076Spjd#include "metadata.h" 63204076Spjd#include "proto.h" 64204076Spjd#include "pjdlog.h" 65204076Spjd#include "subr.h" 66204076Spjd#include "synch.h" 67204076Spjd 68204076Spjdstruct hio { 69204076Spjd /* 70204076Spjd * Number of components we are still waiting for. 71204076Spjd * When this field goes to 0, we can send the request back to the 72204076Spjd * kernel. Each component has to decrease this counter by one 73204076Spjd * even on failure. 74204076Spjd */ 75204076Spjd unsigned int hio_countdown; 76204076Spjd /* 77204076Spjd * Each component has a place to store its own error. 78204076Spjd * Once the request is handled by all components we can decide if the 79204076Spjd * request overall is successful or not. 80204076Spjd */ 81204076Spjd int *hio_errors; 82204076Spjd /* 83204076Spjd * Structure used to comunicate with GEOM Gate class. 84204076Spjd */ 85204076Spjd struct g_gate_ctl_io hio_ggio; 86204076Spjd TAILQ_ENTRY(hio) *hio_next; 87204076Spjd}; 88204076Spjd#define hio_free_next hio_next[0] 89204076Spjd#define hio_done_next hio_next[0] 90204076Spjd 91204076Spjd/* 92204076Spjd * Free list holds unused structures. When free list is empty, we have to wait 93204076Spjd * until some in-progress requests are freed. 94204076Spjd */ 95204076Spjdstatic TAILQ_HEAD(, hio) hio_free_list; 96204076Spjdstatic pthread_mutex_t hio_free_list_lock; 97204076Spjdstatic pthread_cond_t hio_free_list_cond; 98204076Spjd/* 99204076Spjd * There is one send list for every component. One requests is placed on all 100204076Spjd * send lists - each component gets the same request, but each component is 101204076Spjd * responsible for managing his own send list. 102204076Spjd */ 103204076Spjdstatic TAILQ_HEAD(, hio) *hio_send_list; 104204076Spjdstatic pthread_mutex_t *hio_send_list_lock; 105204076Spjdstatic pthread_cond_t *hio_send_list_cond; 106204076Spjd/* 107204076Spjd * There is one recv list for every component, although local components don't 108204076Spjd * use recv lists as local requests are done synchronously. 109204076Spjd */ 110204076Spjdstatic TAILQ_HEAD(, hio) *hio_recv_list; 111204076Spjdstatic pthread_mutex_t *hio_recv_list_lock; 112204076Spjdstatic pthread_cond_t *hio_recv_list_cond; 113204076Spjd/* 114204076Spjd * Request is placed on done list by the slowest component (the one that 115204076Spjd * decreased hio_countdown from 1 to 0). 116204076Spjd */ 117204076Spjdstatic TAILQ_HEAD(, hio) hio_done_list; 118204076Spjdstatic pthread_mutex_t hio_done_list_lock; 119204076Spjdstatic pthread_cond_t hio_done_list_cond; 120204076Spjd/* 121204076Spjd * Structure below are for interaction with sync thread. 122204076Spjd */ 123204076Spjdstatic bool sync_inprogress; 124204076Spjdstatic pthread_mutex_t sync_lock; 125204076Spjdstatic pthread_cond_t sync_cond; 126204076Spjd/* 127204076Spjd * The lock below allows to synchornize access to remote connections. 128204076Spjd */ 129204076Spjdstatic pthread_rwlock_t *hio_remote_lock; 130204076Spjdstatic pthread_mutex_t hio_guard_lock; 131204076Spjdstatic pthread_cond_t hio_guard_cond; 132204076Spjd 133204076Spjd/* 134204076Spjd * Lock to synchronize metadata updates. Also synchronize access to 135204076Spjd * hr_primary_localcnt and hr_primary_remotecnt fields. 136204076Spjd */ 137204076Spjdstatic pthread_mutex_t metadata_lock; 138204076Spjd 139204076Spjd/* 140204076Spjd * Maximum number of outstanding I/O requests. 141204076Spjd */ 142204076Spjd#define HAST_HIO_MAX 256 143204076Spjd/* 144204076Spjd * Number of components. At this point there are only two components: local 145204076Spjd * and remote, but in the future it might be possible to use multiple local 146204076Spjd * and remote components. 147204076Spjd */ 148204076Spjd#define HAST_NCOMPONENTS 2 149204076Spjd/* 150204076Spjd * Number of seconds to sleep before next reconnect try. 151204076Spjd */ 152204076Spjd#define RECONNECT_SLEEP 5 153204076Spjd 154204076Spjd#define ISCONNECTED(res, no) \ 155204076Spjd ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL) 156204076Spjd 157204076Spjd#define QUEUE_INSERT1(hio, name, ncomp) do { \ 158204076Spjd bool _wakeup; \ 159204076Spjd \ 160204076Spjd mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 161204076Spjd _wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]); \ 162204076Spjd TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio), \ 163204076Spjd hio_next[(ncomp)]); \ 164204076Spjd mtx_unlock(&hio_##name##_list_lock[ncomp]); \ 165204076Spjd if (_wakeup) \ 166204076Spjd cv_signal(&hio_##name##_list_cond[(ncomp)]); \ 167204076Spjd} while (0) 168204076Spjd#define QUEUE_INSERT2(hio, name) do { \ 169204076Spjd bool _wakeup; \ 170204076Spjd \ 171204076Spjd mtx_lock(&hio_##name##_list_lock); \ 172204076Spjd _wakeup = TAILQ_EMPTY(&hio_##name##_list); \ 173204076Spjd TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\ 174204076Spjd mtx_unlock(&hio_##name##_list_lock); \ 175204076Spjd if (_wakeup) \ 176204076Spjd cv_signal(&hio_##name##_list_cond); \ 177204076Spjd} while (0) 178204076Spjd#define QUEUE_TAKE1(hio, name, ncomp) do { \ 179204076Spjd mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 180204076Spjd while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL) { \ 181204076Spjd cv_wait(&hio_##name##_list_cond[(ncomp)], \ 182204076Spjd &hio_##name##_list_lock[(ncomp)]); \ 183204076Spjd } \ 184204076Spjd TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \ 185204076Spjd hio_next[(ncomp)]); \ 186204076Spjd mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 187204076Spjd} while (0) 188204076Spjd#define QUEUE_TAKE2(hio, name) do { \ 189204076Spjd mtx_lock(&hio_##name##_list_lock); \ 190204076Spjd while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 191204076Spjd cv_wait(&hio_##name##_list_cond, \ 192204076Spjd &hio_##name##_list_lock); \ 193204076Spjd } \ 194204076Spjd TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \ 195204076Spjd mtx_unlock(&hio_##name##_list_lock); \ 196204076Spjd} while (0) 197204076Spjd 198204076Spjd#define SYNCREQ(hio) do { (hio)->hio_ggio.gctl_unit = -1; } while (0) 199204076Spjd#define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1) 200204076Spjd#define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0) 201204076Spjd#define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2) 202204076Spjd 203204076Spjdstatic struct hast_resource *gres; 204204076Spjd 205204076Spjdstatic pthread_mutex_t range_lock; 206204076Spjdstatic struct rangelocks *range_regular; 207204076Spjdstatic bool range_regular_wait; 208204076Spjdstatic pthread_cond_t range_regular_cond; 209204076Spjdstatic struct rangelocks *range_sync; 210204076Spjdstatic bool range_sync_wait; 211204076Spjdstatic pthread_cond_t range_sync_cond; 212204076Spjd 213204076Spjdstatic void *ggate_recv_thread(void *arg); 214204076Spjdstatic void *local_send_thread(void *arg); 215204076Spjdstatic void *remote_send_thread(void *arg); 216204076Spjdstatic void *remote_recv_thread(void *arg); 217204076Spjdstatic void *ggate_send_thread(void *arg); 218204076Spjdstatic void *sync_thread(void *arg); 219204076Spjdstatic void *guard_thread(void *arg); 220204076Spjd 221204076Spjdstatic void sighandler(int sig); 222204076Spjd 223204076Spjdstatic void 224204076Spjdcleanup(struct hast_resource *res) 225204076Spjd{ 226204076Spjd int rerrno; 227204076Spjd 228204076Spjd /* Remember errno. */ 229204076Spjd rerrno = errno; 230204076Spjd 231204076Spjd /* 232204076Spjd * Close descriptor to /dev/hast/<name> 233204076Spjd * to work-around race in the kernel. 234204076Spjd */ 235204076Spjd close(res->hr_localfd); 236204076Spjd 237204076Spjd /* Destroy ggate provider if we created one. */ 238204076Spjd if (res->hr_ggateunit >= 0) { 239204076Spjd struct g_gate_ctl_destroy ggiod; 240204076Spjd 241204076Spjd ggiod.gctl_version = G_GATE_VERSION; 242204076Spjd ggiod.gctl_unit = res->hr_ggateunit; 243204076Spjd ggiod.gctl_force = 1; 244204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) { 245204076Spjd pjdlog_warning("Unable to destroy hast/%s device", 246204076Spjd res->hr_provname); 247204076Spjd } 248204076Spjd res->hr_ggateunit = -1; 249204076Spjd } 250204076Spjd 251204076Spjd /* Restore errno. */ 252204076Spjd errno = rerrno; 253204076Spjd} 254204076Spjd 255204076Spjdstatic void 256204076Spjdprimary_exit(int exitcode, const char *fmt, ...) 257204076Spjd{ 258204076Spjd va_list ap; 259204076Spjd 260204076Spjd assert(exitcode != EX_OK); 261204076Spjd va_start(ap, fmt); 262204076Spjd pjdlogv_errno(LOG_ERR, fmt, ap); 263204076Spjd va_end(ap); 264204076Spjd cleanup(gres); 265204076Spjd exit(exitcode); 266204076Spjd} 267204076Spjd 268204076Spjdstatic void 269204076Spjdprimary_exitx(int exitcode, const char *fmt, ...) 270204076Spjd{ 271204076Spjd va_list ap; 272204076Spjd 273204076Spjd va_start(ap, fmt); 274204076Spjd pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 275204076Spjd va_end(ap); 276204076Spjd cleanup(gres); 277204076Spjd exit(exitcode); 278204076Spjd} 279204076Spjd 280204076Spjdstatic int 281204076Spjdhast_activemap_flush(struct hast_resource *res) 282204076Spjd{ 283204076Spjd const unsigned char *buf; 284204076Spjd size_t size; 285204076Spjd 286204076Spjd buf = activemap_bitmap(res->hr_amp, &size); 287204076Spjd assert(buf != NULL); 288204076Spjd assert((size % res->hr_local_sectorsize) == 0); 289204076Spjd if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 290204076Spjd (ssize_t)size) { 291204076Spjd KEEP_ERRNO(pjdlog_errno(LOG_ERR, 292204076Spjd "Unable to flush activemap to disk")); 293204076Spjd return (-1); 294204076Spjd } 295204076Spjd return (0); 296204076Spjd} 297204076Spjd 298204076Spjdstatic void 299204076Spjdinit_environment(struct hast_resource *res __unused) 300204076Spjd{ 301204076Spjd struct hio *hio; 302204076Spjd unsigned int ii, ncomps; 303204076Spjd 304204076Spjd /* 305204076Spjd * In the future it might be per-resource value. 306204076Spjd */ 307204076Spjd ncomps = HAST_NCOMPONENTS; 308204076Spjd 309204076Spjd /* 310204076Spjd * Allocate memory needed by lists. 311204076Spjd */ 312204076Spjd hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 313204076Spjd if (hio_send_list == NULL) { 314204076Spjd primary_exitx(EX_TEMPFAIL, 315204076Spjd "Unable to allocate %zu bytes of memory for send lists.", 316204076Spjd sizeof(hio_send_list[0]) * ncomps); 317204076Spjd } 318204076Spjd hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 319204076Spjd if (hio_send_list_lock == NULL) { 320204076Spjd primary_exitx(EX_TEMPFAIL, 321204076Spjd "Unable to allocate %zu bytes of memory for send list locks.", 322204076Spjd sizeof(hio_send_list_lock[0]) * ncomps); 323204076Spjd } 324204076Spjd hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 325204076Spjd if (hio_send_list_cond == NULL) { 326204076Spjd primary_exitx(EX_TEMPFAIL, 327204076Spjd "Unable to allocate %zu bytes of memory for send list condition variables.", 328204076Spjd sizeof(hio_send_list_cond[0]) * ncomps); 329204076Spjd } 330204076Spjd hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 331204076Spjd if (hio_recv_list == NULL) { 332204076Spjd primary_exitx(EX_TEMPFAIL, 333204076Spjd "Unable to allocate %zu bytes of memory for recv lists.", 334204076Spjd sizeof(hio_recv_list[0]) * ncomps); 335204076Spjd } 336204076Spjd hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 337204076Spjd if (hio_recv_list_lock == NULL) { 338204076Spjd primary_exitx(EX_TEMPFAIL, 339204076Spjd "Unable to allocate %zu bytes of memory for recv list locks.", 340204076Spjd sizeof(hio_recv_list_lock[0]) * ncomps); 341204076Spjd } 342204076Spjd hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 343204076Spjd if (hio_recv_list_cond == NULL) { 344204076Spjd primary_exitx(EX_TEMPFAIL, 345204076Spjd "Unable to allocate %zu bytes of memory for recv list condition variables.", 346204076Spjd sizeof(hio_recv_list_cond[0]) * ncomps); 347204076Spjd } 348204076Spjd hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 349204076Spjd if (hio_remote_lock == NULL) { 350204076Spjd primary_exitx(EX_TEMPFAIL, 351204076Spjd "Unable to allocate %zu bytes of memory for remote connections locks.", 352204076Spjd sizeof(hio_remote_lock[0]) * ncomps); 353204076Spjd } 354204076Spjd 355204076Spjd /* 356204076Spjd * Initialize lists, their locks and theirs condition variables. 357204076Spjd */ 358204076Spjd TAILQ_INIT(&hio_free_list); 359204076Spjd mtx_init(&hio_free_list_lock); 360204076Spjd cv_init(&hio_free_list_cond); 361204076Spjd for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 362204076Spjd TAILQ_INIT(&hio_send_list[ii]); 363204076Spjd mtx_init(&hio_send_list_lock[ii]); 364204076Spjd cv_init(&hio_send_list_cond[ii]); 365204076Spjd TAILQ_INIT(&hio_recv_list[ii]); 366204076Spjd mtx_init(&hio_recv_list_lock[ii]); 367204076Spjd cv_init(&hio_recv_list_cond[ii]); 368204076Spjd rw_init(&hio_remote_lock[ii]); 369204076Spjd } 370204076Spjd TAILQ_INIT(&hio_done_list); 371204076Spjd mtx_init(&hio_done_list_lock); 372204076Spjd cv_init(&hio_done_list_cond); 373204076Spjd mtx_init(&hio_guard_lock); 374204076Spjd cv_init(&hio_guard_cond); 375204076Spjd mtx_init(&metadata_lock); 376204076Spjd 377204076Spjd /* 378204076Spjd * Allocate requests pool and initialize requests. 379204076Spjd */ 380204076Spjd for (ii = 0; ii < HAST_HIO_MAX; ii++) { 381204076Spjd hio = malloc(sizeof(*hio)); 382204076Spjd if (hio == NULL) { 383204076Spjd primary_exitx(EX_TEMPFAIL, 384204076Spjd "Unable to allocate %zu bytes of memory for hio request.", 385204076Spjd sizeof(*hio)); 386204076Spjd } 387204076Spjd hio->hio_countdown = 0; 388204076Spjd hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 389204076Spjd if (hio->hio_errors == NULL) { 390204076Spjd primary_exitx(EX_TEMPFAIL, 391204076Spjd "Unable allocate %zu bytes of memory for hio errors.", 392204076Spjd sizeof(hio->hio_errors[0]) * ncomps); 393204076Spjd } 394204076Spjd hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 395204076Spjd if (hio->hio_next == NULL) { 396204076Spjd primary_exitx(EX_TEMPFAIL, 397204076Spjd "Unable allocate %zu bytes of memory for hio_next field.", 398204076Spjd sizeof(hio->hio_next[0]) * ncomps); 399204076Spjd } 400204076Spjd hio->hio_ggio.gctl_version = G_GATE_VERSION; 401204076Spjd hio->hio_ggio.gctl_data = malloc(MAXPHYS); 402204076Spjd if (hio->hio_ggio.gctl_data == NULL) { 403204076Spjd primary_exitx(EX_TEMPFAIL, 404204076Spjd "Unable to allocate %zu bytes of memory for gctl_data.", 405204076Spjd MAXPHYS); 406204076Spjd } 407204076Spjd hio->hio_ggio.gctl_length = MAXPHYS; 408204076Spjd hio->hio_ggio.gctl_error = 0; 409204076Spjd TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 410204076Spjd } 411204076Spjd 412204076Spjd /* 413204076Spjd * Turn on signals handling. 414204076Spjd */ 415204076Spjd signal(SIGINT, sighandler); 416204076Spjd signal(SIGTERM, sighandler); 417204076Spjd} 418204076Spjd 419204076Spjdstatic void 420204076Spjdinit_local(struct hast_resource *res) 421204076Spjd{ 422204076Spjd unsigned char *buf; 423204076Spjd size_t mapsize; 424204076Spjd 425204076Spjd if (metadata_read(res, true) < 0) 426204076Spjd exit(EX_NOINPUT); 427204076Spjd mtx_init(&res->hr_amp_lock); 428204076Spjd if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 429204076Spjd res->hr_local_sectorsize, res->hr_keepdirty) < 0) { 430204076Spjd primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 431204076Spjd } 432204076Spjd mtx_init(&range_lock); 433204076Spjd cv_init(&range_regular_cond); 434204076Spjd if (rangelock_init(&range_regular) < 0) 435204076Spjd primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 436204076Spjd cv_init(&range_sync_cond); 437204076Spjd if (rangelock_init(&range_sync) < 0) 438204076Spjd primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 439204076Spjd mapsize = activemap_ondisk_size(res->hr_amp); 440204076Spjd buf = calloc(1, mapsize); 441204076Spjd if (buf == NULL) { 442204076Spjd primary_exitx(EX_TEMPFAIL, 443204076Spjd "Unable to allocate buffer for activemap."); 444204076Spjd } 445204076Spjd if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 446204076Spjd (ssize_t)mapsize) { 447204076Spjd primary_exit(EX_NOINPUT, "Unable to read activemap"); 448204076Spjd } 449204076Spjd activemap_copyin(res->hr_amp, buf, mapsize); 450204076Spjd if (res->hr_resuid != 0) 451204076Spjd return; 452204076Spjd /* 453204076Spjd * We're using provider for the first time, so we have to generate 454204076Spjd * resource unique identifier and initialize local and remote counts. 455204076Spjd */ 456204076Spjd arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 457204076Spjd res->hr_primary_localcnt = 1; 458204076Spjd res->hr_primary_remotecnt = 0; 459204076Spjd if (metadata_write(res) < 0) 460204076Spjd exit(EX_NOINPUT); 461204076Spjd} 462204076Spjd 463205738Spjdstatic bool 464205738Spjdinit_remote(struct hast_resource *res, struct proto_conn **inp, 465205738Spjd struct proto_conn **outp) 466204076Spjd{ 467205738Spjd struct proto_conn *in, *out; 468204076Spjd struct nv *nvout, *nvin; 469204076Spjd const unsigned char *token; 470204076Spjd unsigned char *map; 471204076Spjd const char *errmsg; 472204076Spjd int32_t extentsize; 473204076Spjd int64_t datasize; 474204076Spjd uint32_t mapsize; 475204076Spjd size_t size; 476204076Spjd 477205738Spjd assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 478205738Spjd 479205738Spjd in = out = NULL; 480205738Spjd 481204076Spjd /* Prepare outgoing connection with remote node. */ 482205738Spjd if (proto_client(res->hr_remoteaddr, &out) < 0) { 483204076Spjd primary_exit(EX_OSERR, "Unable to create connection to %s", 484204076Spjd res->hr_remoteaddr); 485204076Spjd } 486204076Spjd /* Try to connect, but accept failure. */ 487205738Spjd if (proto_connect(out) < 0) { 488204076Spjd pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 489204076Spjd res->hr_remoteaddr); 490204076Spjd goto close; 491204076Spjd } 492204076Spjd /* 493204076Spjd * First handshake step. 494204076Spjd * Setup outgoing connection with remote node. 495204076Spjd */ 496204076Spjd nvout = nv_alloc(); 497204076Spjd nv_add_string(nvout, res->hr_name, "resource"); 498204076Spjd if (nv_error(nvout) != 0) { 499204076Spjd pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 500204076Spjd "Unable to allocate header for connection with %s", 501204076Spjd res->hr_remoteaddr); 502204076Spjd nv_free(nvout); 503204076Spjd goto close; 504204076Spjd } 505205738Spjd if (hast_proto_send(res, out, nvout, NULL, 0) < 0) { 506204076Spjd pjdlog_errno(LOG_WARNING, 507204076Spjd "Unable to send handshake header to %s", 508204076Spjd res->hr_remoteaddr); 509204076Spjd nv_free(nvout); 510204076Spjd goto close; 511204076Spjd } 512204076Spjd nv_free(nvout); 513205738Spjd if (hast_proto_recv_hdr(out, &nvin) < 0) { 514204076Spjd pjdlog_errno(LOG_WARNING, 515204076Spjd "Unable to receive handshake header from %s", 516204076Spjd res->hr_remoteaddr); 517204076Spjd goto close; 518204076Spjd } 519204076Spjd errmsg = nv_get_string(nvin, "errmsg"); 520204076Spjd if (errmsg != NULL) { 521204076Spjd pjdlog_warning("%s", errmsg); 522204076Spjd nv_free(nvin); 523204076Spjd goto close; 524204076Spjd } 525204076Spjd token = nv_get_uint8_array(nvin, &size, "token"); 526204076Spjd if (token == NULL) { 527204076Spjd pjdlog_warning("Handshake header from %s has no 'token' field.", 528204076Spjd res->hr_remoteaddr); 529204076Spjd nv_free(nvin); 530204076Spjd goto close; 531204076Spjd } 532204076Spjd if (size != sizeof(res->hr_token)) { 533204076Spjd pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 534204076Spjd res->hr_remoteaddr, size, sizeof(res->hr_token)); 535204076Spjd nv_free(nvin); 536204076Spjd goto close; 537204076Spjd } 538204076Spjd bcopy(token, res->hr_token, sizeof(res->hr_token)); 539204076Spjd nv_free(nvin); 540204076Spjd 541204076Spjd /* 542204076Spjd * Second handshake step. 543204076Spjd * Setup incoming connection with remote node. 544204076Spjd */ 545205738Spjd if (proto_client(res->hr_remoteaddr, &in) < 0) { 546204076Spjd pjdlog_errno(LOG_WARNING, "Unable to create connection to %s", 547204076Spjd res->hr_remoteaddr); 548204076Spjd } 549204076Spjd /* Try to connect, but accept failure. */ 550205738Spjd if (proto_connect(in) < 0) { 551204076Spjd pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 552204076Spjd res->hr_remoteaddr); 553204076Spjd goto close; 554204076Spjd } 555204076Spjd nvout = nv_alloc(); 556204076Spjd nv_add_string(nvout, res->hr_name, "resource"); 557204076Spjd nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 558204076Spjd "token"); 559204076Spjd nv_add_uint64(nvout, res->hr_resuid, "resuid"); 560204076Spjd nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 561204076Spjd nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 562204076Spjd if (nv_error(nvout) != 0) { 563204076Spjd pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 564204076Spjd "Unable to allocate header for connection with %s", 565204076Spjd res->hr_remoteaddr); 566204076Spjd nv_free(nvout); 567204076Spjd goto close; 568204076Spjd } 569205738Spjd if (hast_proto_send(res, in, nvout, NULL, 0) < 0) { 570204076Spjd pjdlog_errno(LOG_WARNING, 571204076Spjd "Unable to send handshake header to %s", 572204076Spjd res->hr_remoteaddr); 573204076Spjd nv_free(nvout); 574204076Spjd goto close; 575204076Spjd } 576204076Spjd nv_free(nvout); 577205738Spjd if (hast_proto_recv_hdr(out, &nvin) < 0) { 578204076Spjd pjdlog_errno(LOG_WARNING, 579204076Spjd "Unable to receive handshake header from %s", 580204076Spjd res->hr_remoteaddr); 581204076Spjd goto close; 582204076Spjd } 583204076Spjd errmsg = nv_get_string(nvin, "errmsg"); 584204076Spjd if (errmsg != NULL) { 585204076Spjd pjdlog_warning("%s", errmsg); 586204076Spjd nv_free(nvin); 587204076Spjd goto close; 588204076Spjd } 589204076Spjd datasize = nv_get_int64(nvin, "datasize"); 590204076Spjd if (datasize != res->hr_datasize) { 591204076Spjd pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 592204076Spjd (intmax_t)res->hr_datasize, (intmax_t)datasize); 593204076Spjd nv_free(nvin); 594204076Spjd goto close; 595204076Spjd } 596204076Spjd extentsize = nv_get_int32(nvin, "extentsize"); 597204076Spjd if (extentsize != res->hr_extentsize) { 598204076Spjd pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 599204076Spjd (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 600204076Spjd nv_free(nvin); 601204076Spjd goto close; 602204076Spjd } 603204076Spjd res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 604204076Spjd res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 605204076Spjd res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 606204076Spjd map = NULL; 607204076Spjd mapsize = nv_get_uint32(nvin, "mapsize"); 608204076Spjd if (mapsize > 0) { 609204076Spjd map = malloc(mapsize); 610204076Spjd if (map == NULL) { 611204076Spjd pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 612204076Spjd (uintmax_t)mapsize); 613204076Spjd nv_free(nvin); 614204076Spjd goto close; 615204076Spjd } 616204076Spjd /* 617204076Spjd * Remote node have some dirty extents on its own, lets 618204076Spjd * download its activemap. 619204076Spjd */ 620205738Spjd if (hast_proto_recv_data(res, out, nvin, map, 621204076Spjd mapsize) < 0) { 622204076Spjd pjdlog_errno(LOG_ERR, 623204076Spjd "Unable to receive remote activemap"); 624204076Spjd nv_free(nvin); 625204076Spjd free(map); 626204076Spjd goto close; 627204076Spjd } 628204076Spjd /* 629204076Spjd * Merge local and remote bitmaps. 630204076Spjd */ 631204076Spjd activemap_merge(res->hr_amp, map, mapsize); 632204076Spjd free(map); 633204076Spjd /* 634204076Spjd * Now that we merged bitmaps from both nodes, flush it to the 635204076Spjd * disk before we start to synchronize. 636204076Spjd */ 637204076Spjd (void)hast_activemap_flush(res); 638204076Spjd } 639204076Spjd pjdlog_info("Connected to %s.", res->hr_remoteaddr); 640205738Spjd if (inp != NULL && outp != NULL) { 641205738Spjd *inp = in; 642205738Spjd *outp = out; 643205738Spjd } else { 644205738Spjd res->hr_remotein = in; 645205738Spjd res->hr_remoteout = out; 646205738Spjd } 647205738Spjd return (true); 648205738Spjdclose: 649205738Spjd proto_close(out); 650205738Spjd if (in != NULL) 651205738Spjd proto_close(in); 652205738Spjd return (false); 653205738Spjd} 654205738Spjd 655205738Spjdstatic void 656205738Spjdsync_start(void) 657205738Spjd{ 658205738Spjd 659204076Spjd mtx_lock(&sync_lock); 660204076Spjd sync_inprogress = true; 661204076Spjd mtx_unlock(&sync_lock); 662204076Spjd cv_signal(&sync_cond); 663204076Spjd} 664204076Spjd 665204076Spjdstatic void 666204076Spjdinit_ggate(struct hast_resource *res) 667204076Spjd{ 668204076Spjd struct g_gate_ctl_create ggiocreate; 669204076Spjd struct g_gate_ctl_cancel ggiocancel; 670204076Spjd 671204076Spjd /* 672204076Spjd * We communicate with ggate via /dev/ggctl. Open it. 673204076Spjd */ 674204076Spjd res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 675204076Spjd if (res->hr_ggatefd < 0) 676204076Spjd primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 677204076Spjd /* 678204076Spjd * Create provider before trying to connect, as connection failure 679204076Spjd * is not critical, but may take some time. 680204076Spjd */ 681204076Spjd ggiocreate.gctl_version = G_GATE_VERSION; 682204076Spjd ggiocreate.gctl_mediasize = res->hr_datasize; 683204076Spjd ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 684204076Spjd ggiocreate.gctl_flags = 0; 685204076Spjd ggiocreate.gctl_maxcount = 128; 686204076Spjd ggiocreate.gctl_timeout = 0; 687204076Spjd ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 688204076Spjd snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 689204076Spjd res->hr_provname); 690204076Spjd bzero(ggiocreate.gctl_info, sizeof(ggiocreate.gctl_info)); 691204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 692204076Spjd pjdlog_info("Device hast/%s created.", res->hr_provname); 693204076Spjd res->hr_ggateunit = ggiocreate.gctl_unit; 694204076Spjd return; 695204076Spjd } 696204076Spjd if (errno != EEXIST) { 697204076Spjd primary_exit(EX_OSERR, "Unable to create hast/%s device", 698204076Spjd res->hr_provname); 699204076Spjd } 700204076Spjd pjdlog_debug(1, 701204076Spjd "Device hast/%s already exists, we will try to take it over.", 702204076Spjd res->hr_provname); 703204076Spjd /* 704204076Spjd * If we received EEXIST, we assume that the process who created the 705204076Spjd * provider died and didn't clean up. In that case we will start from 706204076Spjd * where he left of. 707204076Spjd */ 708204076Spjd ggiocancel.gctl_version = G_GATE_VERSION; 709204076Spjd ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 710204076Spjd snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 711204076Spjd res->hr_provname); 712204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 713204076Spjd pjdlog_info("Device hast/%s recovered.", res->hr_provname); 714204076Spjd res->hr_ggateunit = ggiocancel.gctl_unit; 715204076Spjd return; 716204076Spjd } 717204076Spjd primary_exit(EX_OSERR, "Unable to take over hast/%s device", 718204076Spjd res->hr_provname); 719204076Spjd} 720204076Spjd 721204076Spjdvoid 722204076Spjdhastd_primary(struct hast_resource *res) 723204076Spjd{ 724204076Spjd pthread_t td; 725204076Spjd pid_t pid; 726204076Spjd int error; 727204076Spjd 728204076Spjd gres = res; 729204076Spjd 730204076Spjd /* 731204076Spjd * Create communication channel between parent and child. 732204076Spjd */ 733204076Spjd if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 734204076Spjd KEEP_ERRNO((void)pidfile_remove(pfh)); 735204076Spjd primary_exit(EX_OSERR, 736204076Spjd "Unable to create control sockets between parent and child"); 737204076Spjd } 738204076Spjd 739204076Spjd pid = fork(); 740204076Spjd if (pid < 0) { 741204076Spjd KEEP_ERRNO((void)pidfile_remove(pfh)); 742204076Spjd primary_exit(EX_OSERR, "Unable to fork"); 743204076Spjd } 744204076Spjd 745204076Spjd if (pid > 0) { 746204076Spjd /* This is parent. */ 747204076Spjd res->hr_workerpid = pid; 748204076Spjd return; 749204076Spjd } 750204076Spjd (void)pidfile_close(pfh); 751204076Spjd 752204076Spjd setproctitle("%s (primary)", res->hr_name); 753204076Spjd 754204076Spjd init_local(res); 755205738Spjd if (init_remote(res, NULL, NULL)) 756205738Spjd sync_start(); 757204076Spjd init_ggate(res); 758204076Spjd init_environment(res); 759204076Spjd error = pthread_create(&td, NULL, ggate_recv_thread, res); 760204076Spjd assert(error == 0); 761204076Spjd error = pthread_create(&td, NULL, local_send_thread, res); 762204076Spjd assert(error == 0); 763204076Spjd error = pthread_create(&td, NULL, remote_send_thread, res); 764204076Spjd assert(error == 0); 765204076Spjd error = pthread_create(&td, NULL, remote_recv_thread, res); 766204076Spjd assert(error == 0); 767204076Spjd error = pthread_create(&td, NULL, ggate_send_thread, res); 768204076Spjd assert(error == 0); 769204076Spjd error = pthread_create(&td, NULL, sync_thread, res); 770204076Spjd assert(error == 0); 771204076Spjd error = pthread_create(&td, NULL, ctrl_thread, res); 772204076Spjd assert(error == 0); 773204076Spjd (void)guard_thread(res); 774204076Spjd} 775204076Spjd 776204076Spjdstatic void 777204076Spjdreqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 778204076Spjd{ 779204076Spjd char msg[1024]; 780204076Spjd va_list ap; 781204076Spjd int len; 782204076Spjd 783204076Spjd va_start(ap, fmt); 784204076Spjd len = vsnprintf(msg, sizeof(msg), fmt, ap); 785204076Spjd va_end(ap); 786204076Spjd if ((size_t)len < sizeof(msg)) { 787204076Spjd switch (ggio->gctl_cmd) { 788204076Spjd case BIO_READ: 789204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 790204076Spjd "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 791204076Spjd (uintmax_t)ggio->gctl_length); 792204076Spjd break; 793204076Spjd case BIO_DELETE: 794204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 795204076Spjd "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 796204076Spjd (uintmax_t)ggio->gctl_length); 797204076Spjd break; 798204076Spjd case BIO_FLUSH: 799204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 800204076Spjd break; 801204076Spjd case BIO_WRITE: 802204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 803204076Spjd "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 804204076Spjd (uintmax_t)ggio->gctl_length); 805204076Spjd break; 806204076Spjd default: 807204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 808204076Spjd "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 809204076Spjd break; 810204076Spjd } 811204076Spjd } 812204076Spjd pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 813204076Spjd} 814204076Spjd 815204076Spjdstatic void 816204076Spjdremote_close(struct hast_resource *res, int ncomp) 817204076Spjd{ 818204076Spjd 819204076Spjd rw_wlock(&hio_remote_lock[ncomp]); 820204076Spjd /* 821204076Spjd * A race is possible between dropping rlock and acquiring wlock - 822204076Spjd * another thread can close connection in-between. 823204076Spjd */ 824204076Spjd if (!ISCONNECTED(res, ncomp)) { 825204076Spjd assert(res->hr_remotein == NULL); 826204076Spjd assert(res->hr_remoteout == NULL); 827204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 828204076Spjd return; 829204076Spjd } 830204076Spjd 831204076Spjd assert(res->hr_remotein != NULL); 832204076Spjd assert(res->hr_remoteout != NULL); 833204076Spjd 834204076Spjd pjdlog_debug(2, "Closing old incoming connection to %s.", 835204076Spjd res->hr_remoteaddr); 836204076Spjd proto_close(res->hr_remotein); 837204076Spjd res->hr_remotein = NULL; 838204076Spjd pjdlog_debug(2, "Closing old outgoing connection to %s.", 839204076Spjd res->hr_remoteaddr); 840204076Spjd proto_close(res->hr_remoteout); 841204076Spjd res->hr_remoteout = NULL; 842204076Spjd 843204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 844204076Spjd 845204076Spjd /* 846204076Spjd * Stop synchronization if in-progress. 847204076Spjd */ 848204076Spjd mtx_lock(&sync_lock); 849204076Spjd if (sync_inprogress) 850204076Spjd sync_inprogress = false; 851204076Spjd mtx_unlock(&sync_lock); 852204076Spjd 853204076Spjd /* 854204076Spjd * Wake up guard thread, so it can immediately start reconnect. 855204076Spjd */ 856204076Spjd mtx_lock(&hio_guard_lock); 857204076Spjd cv_signal(&hio_guard_cond); 858204076Spjd mtx_unlock(&hio_guard_lock); 859204076Spjd} 860204076Spjd 861204076Spjd/* 862204076Spjd * Thread receives ggate I/O requests from the kernel and passes them to 863204076Spjd * appropriate threads: 864204076Spjd * WRITE - always goes to both local_send and remote_send threads 865204076Spjd * READ (when the block is up-to-date on local component) - 866204076Spjd * only local_send thread 867204076Spjd * READ (when the block isn't up-to-date on local component) - 868204076Spjd * only remote_send thread 869204076Spjd * DELETE - always goes to both local_send and remote_send threads 870204076Spjd * FLUSH - always goes to both local_send and remote_send threads 871204076Spjd */ 872204076Spjdstatic void * 873204076Spjdggate_recv_thread(void *arg) 874204076Spjd{ 875204076Spjd struct hast_resource *res = arg; 876204076Spjd struct g_gate_ctl_io *ggio; 877204076Spjd struct hio *hio; 878204076Spjd unsigned int ii, ncomp, ncomps; 879204076Spjd int error; 880204076Spjd 881204076Spjd ncomps = HAST_NCOMPONENTS; 882204076Spjd 883204076Spjd for (;;) { 884204076Spjd pjdlog_debug(2, "ggate_recv: Taking free request."); 885204076Spjd QUEUE_TAKE2(hio, free); 886204076Spjd pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 887204076Spjd ggio = &hio->hio_ggio; 888204076Spjd ggio->gctl_unit = res->hr_ggateunit; 889204076Spjd ggio->gctl_length = MAXPHYS; 890204076Spjd ggio->gctl_error = 0; 891204076Spjd pjdlog_debug(2, 892204076Spjd "ggate_recv: (%p) Waiting for request from the kernel.", 893204076Spjd hio); 894204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 895204076Spjd if (sigexit_received) 896204076Spjd pthread_exit(NULL); 897204076Spjd primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 898204076Spjd } 899204076Spjd error = ggio->gctl_error; 900204076Spjd switch (error) { 901204076Spjd case 0: 902204076Spjd break; 903204076Spjd case ECANCELED: 904204076Spjd /* Exit gracefully. */ 905204076Spjd if (!sigexit_received) { 906204076Spjd pjdlog_debug(2, 907204076Spjd "ggate_recv: (%p) Received cancel from the kernel.", 908204076Spjd hio); 909204076Spjd pjdlog_info("Received cancel from the kernel, exiting."); 910204076Spjd } 911204076Spjd pthread_exit(NULL); 912204076Spjd case ENOMEM: 913204076Spjd /* 914204076Spjd * Buffer too small? Impossible, we allocate MAXPHYS 915204076Spjd * bytes - request can't be bigger than that. 916204076Spjd */ 917204076Spjd /* FALLTHROUGH */ 918204076Spjd case ENXIO: 919204076Spjd default: 920204076Spjd primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 921204076Spjd strerror(error)); 922204076Spjd } 923204076Spjd for (ii = 0; ii < ncomps; ii++) 924204076Spjd hio->hio_errors[ii] = EINVAL; 925204076Spjd reqlog(LOG_DEBUG, 2, ggio, 926204076Spjd "ggate_recv: (%p) Request received from the kernel: ", 927204076Spjd hio); 928204076Spjd /* 929204076Spjd * Inform all components about new write request. 930204076Spjd * For read request prefer local component unless the given 931204076Spjd * range is out-of-date, then use remote component. 932204076Spjd */ 933204076Spjd switch (ggio->gctl_cmd) { 934204076Spjd case BIO_READ: 935204076Spjd pjdlog_debug(2, 936204076Spjd "ggate_recv: (%p) Moving request to the send queue.", 937204076Spjd hio); 938204076Spjd refcount_init(&hio->hio_countdown, 1); 939204076Spjd mtx_lock(&metadata_lock); 940204076Spjd if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 941204076Spjd res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 942204076Spjd /* 943204076Spjd * This range is up-to-date on local component, 944204076Spjd * so handle request locally. 945204076Spjd */ 946204076Spjd /* Local component is 0 for now. */ 947204076Spjd ncomp = 0; 948204076Spjd } else /* if (res->hr_syncsrc == 949204076Spjd HAST_SYNCSRC_SECONDARY) */ { 950204076Spjd assert(res->hr_syncsrc == 951204076Spjd HAST_SYNCSRC_SECONDARY); 952204076Spjd /* 953204076Spjd * This range is out-of-date on local component, 954204076Spjd * so send request to the remote node. 955204076Spjd */ 956204076Spjd /* Remote component is 1 for now. */ 957204076Spjd ncomp = 1; 958204076Spjd } 959204076Spjd mtx_unlock(&metadata_lock); 960204076Spjd QUEUE_INSERT1(hio, send, ncomp); 961204076Spjd break; 962204076Spjd case BIO_WRITE: 963204076Spjd for (;;) { 964204076Spjd mtx_lock(&range_lock); 965204076Spjd if (rangelock_islocked(range_sync, 966204076Spjd ggio->gctl_offset, ggio->gctl_length)) { 967204076Spjd pjdlog_debug(2, 968204076Spjd "regular: Range offset=%jd length=%zu locked.", 969204076Spjd (intmax_t)ggio->gctl_offset, 970204076Spjd (size_t)ggio->gctl_length); 971204076Spjd range_regular_wait = true; 972204076Spjd cv_wait(&range_regular_cond, &range_lock); 973204076Spjd range_regular_wait = false; 974204076Spjd mtx_unlock(&range_lock); 975204076Spjd continue; 976204076Spjd } 977204076Spjd if (rangelock_add(range_regular, 978204076Spjd ggio->gctl_offset, ggio->gctl_length) < 0) { 979204076Spjd mtx_unlock(&range_lock); 980204076Spjd pjdlog_debug(2, 981204076Spjd "regular: Range offset=%jd length=%zu is already locked, waiting.", 982204076Spjd (intmax_t)ggio->gctl_offset, 983204076Spjd (size_t)ggio->gctl_length); 984204076Spjd sleep(1); 985204076Spjd continue; 986204076Spjd } 987204076Spjd mtx_unlock(&range_lock); 988204076Spjd break; 989204076Spjd } 990204076Spjd mtx_lock(&res->hr_amp_lock); 991204076Spjd if (activemap_write_start(res->hr_amp, 992204076Spjd ggio->gctl_offset, ggio->gctl_length)) { 993204076Spjd (void)hast_activemap_flush(res); 994204076Spjd } 995204076Spjd mtx_unlock(&res->hr_amp_lock); 996204076Spjd /* FALLTHROUGH */ 997204076Spjd case BIO_DELETE: 998204076Spjd case BIO_FLUSH: 999204076Spjd pjdlog_debug(2, 1000204076Spjd "ggate_recv: (%p) Moving request to the send queues.", 1001204076Spjd hio); 1002204076Spjd refcount_init(&hio->hio_countdown, ncomps); 1003204076Spjd for (ii = 0; ii < ncomps; ii++) 1004204076Spjd QUEUE_INSERT1(hio, send, ii); 1005204076Spjd break; 1006204076Spjd } 1007204076Spjd } 1008204076Spjd /* NOTREACHED */ 1009204076Spjd return (NULL); 1010204076Spjd} 1011204076Spjd 1012204076Spjd/* 1013204076Spjd * Thread reads from or writes to local component. 1014204076Spjd * If local read fails, it redirects it to remote_send thread. 1015204076Spjd */ 1016204076Spjdstatic void * 1017204076Spjdlocal_send_thread(void *arg) 1018204076Spjd{ 1019204076Spjd struct hast_resource *res = arg; 1020204076Spjd struct g_gate_ctl_io *ggio; 1021204076Spjd struct hio *hio; 1022204076Spjd unsigned int ncomp, rncomp; 1023204076Spjd ssize_t ret; 1024204076Spjd 1025204076Spjd /* Local component is 0 for now. */ 1026204076Spjd ncomp = 0; 1027204076Spjd /* Remote component is 1 for now. */ 1028204076Spjd rncomp = 1; 1029204076Spjd 1030204076Spjd for (;;) { 1031204076Spjd pjdlog_debug(2, "local_send: Taking request."); 1032204076Spjd QUEUE_TAKE1(hio, send, ncomp); 1033204076Spjd pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1034204076Spjd ggio = &hio->hio_ggio; 1035204076Spjd switch (ggio->gctl_cmd) { 1036204076Spjd case BIO_READ: 1037204076Spjd ret = pread(res->hr_localfd, ggio->gctl_data, 1038204076Spjd ggio->gctl_length, 1039204076Spjd ggio->gctl_offset + res->hr_localoff); 1040204076Spjd if (ret == ggio->gctl_length) 1041204076Spjd hio->hio_errors[ncomp] = 0; 1042204076Spjd else { 1043204076Spjd /* 1044204076Spjd * If READ failed, try to read from remote node. 1045204076Spjd */ 1046204076Spjd QUEUE_INSERT1(hio, send, rncomp); 1047204076Spjd continue; 1048204076Spjd } 1049204076Spjd break; 1050204076Spjd case BIO_WRITE: 1051204076Spjd ret = pwrite(res->hr_localfd, ggio->gctl_data, 1052204076Spjd ggio->gctl_length, 1053204076Spjd ggio->gctl_offset + res->hr_localoff); 1054204076Spjd if (ret < 0) 1055204076Spjd hio->hio_errors[ncomp] = errno; 1056204076Spjd else if (ret != ggio->gctl_length) 1057204076Spjd hio->hio_errors[ncomp] = EIO; 1058204076Spjd else 1059204076Spjd hio->hio_errors[ncomp] = 0; 1060204076Spjd break; 1061204076Spjd case BIO_DELETE: 1062204076Spjd ret = g_delete(res->hr_localfd, 1063204076Spjd ggio->gctl_offset + res->hr_localoff, 1064204076Spjd ggio->gctl_length); 1065204076Spjd if (ret < 0) 1066204076Spjd hio->hio_errors[ncomp] = errno; 1067204076Spjd else 1068204076Spjd hio->hio_errors[ncomp] = 0; 1069204076Spjd break; 1070204076Spjd case BIO_FLUSH: 1071204076Spjd ret = g_flush(res->hr_localfd); 1072204076Spjd if (ret < 0) 1073204076Spjd hio->hio_errors[ncomp] = errno; 1074204076Spjd else 1075204076Spjd hio->hio_errors[ncomp] = 0; 1076204076Spjd break; 1077204076Spjd } 1078204076Spjd if (refcount_release(&hio->hio_countdown)) { 1079204076Spjd if (ISSYNCREQ(hio)) { 1080204076Spjd mtx_lock(&sync_lock); 1081204076Spjd SYNCREQDONE(hio); 1082204076Spjd mtx_unlock(&sync_lock); 1083204076Spjd cv_signal(&sync_cond); 1084204076Spjd } else { 1085204076Spjd pjdlog_debug(2, 1086204076Spjd "local_send: (%p) Moving request to the done queue.", 1087204076Spjd hio); 1088204076Spjd QUEUE_INSERT2(hio, done); 1089204076Spjd } 1090204076Spjd } 1091204076Spjd } 1092204076Spjd /* NOTREACHED */ 1093204076Spjd return (NULL); 1094204076Spjd} 1095204076Spjd 1096204076Spjd/* 1097204076Spjd * Thread sends request to secondary node. 1098204076Spjd */ 1099204076Spjdstatic void * 1100204076Spjdremote_send_thread(void *arg) 1101204076Spjd{ 1102204076Spjd struct hast_resource *res = arg; 1103204076Spjd struct g_gate_ctl_io *ggio; 1104204076Spjd struct hio *hio; 1105204076Spjd struct nv *nv; 1106204076Spjd unsigned int ncomp; 1107204076Spjd bool wakeup; 1108204076Spjd uint64_t offset, length; 1109204076Spjd uint8_t cmd; 1110204076Spjd void *data; 1111204076Spjd 1112204076Spjd /* Remote component is 1 for now. */ 1113204076Spjd ncomp = 1; 1114204076Spjd 1115204076Spjd for (;;) { 1116204076Spjd pjdlog_debug(2, "remote_send: Taking request."); 1117204076Spjd QUEUE_TAKE1(hio, send, ncomp); 1118204076Spjd pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1119204076Spjd ggio = &hio->hio_ggio; 1120204076Spjd switch (ggio->gctl_cmd) { 1121204076Spjd case BIO_READ: 1122204076Spjd cmd = HIO_READ; 1123204076Spjd data = NULL; 1124204076Spjd offset = ggio->gctl_offset; 1125204076Spjd length = ggio->gctl_length; 1126204076Spjd break; 1127204076Spjd case BIO_WRITE: 1128204076Spjd cmd = HIO_WRITE; 1129204076Spjd data = ggio->gctl_data; 1130204076Spjd offset = ggio->gctl_offset; 1131204076Spjd length = ggio->gctl_length; 1132204076Spjd break; 1133204076Spjd case BIO_DELETE: 1134204076Spjd cmd = HIO_DELETE; 1135204076Spjd data = NULL; 1136204076Spjd offset = ggio->gctl_offset; 1137204076Spjd length = ggio->gctl_length; 1138204076Spjd break; 1139204076Spjd case BIO_FLUSH: 1140204076Spjd cmd = HIO_FLUSH; 1141204076Spjd data = NULL; 1142204076Spjd offset = 0; 1143204076Spjd length = 0; 1144204076Spjd break; 1145204076Spjd default: 1146204076Spjd assert(!"invalid condition"); 1147204076Spjd abort(); 1148204076Spjd } 1149204076Spjd nv = nv_alloc(); 1150204076Spjd nv_add_uint8(nv, cmd, "cmd"); 1151204076Spjd nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1152204076Spjd nv_add_uint64(nv, offset, "offset"); 1153204076Spjd nv_add_uint64(nv, length, "length"); 1154204076Spjd if (nv_error(nv) != 0) { 1155204076Spjd hio->hio_errors[ncomp] = nv_error(nv); 1156204076Spjd pjdlog_debug(2, 1157204076Spjd "remote_send: (%p) Unable to prepare header to send.", 1158204076Spjd hio); 1159204076Spjd reqlog(LOG_ERR, 0, ggio, 1160204076Spjd "Unable to prepare header to send (%s): ", 1161204076Spjd strerror(nv_error(nv))); 1162204076Spjd /* Move failed request immediately to the done queue. */ 1163204076Spjd goto done_queue; 1164204076Spjd } 1165204076Spjd pjdlog_debug(2, 1166204076Spjd "remote_send: (%p) Moving request to the recv queue.", 1167204076Spjd hio); 1168204076Spjd /* 1169204076Spjd * Protect connection from disappearing. 1170204076Spjd */ 1171204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1172204076Spjd if (!ISCONNECTED(res, ncomp)) { 1173204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1174204076Spjd hio->hio_errors[ncomp] = ENOTCONN; 1175204076Spjd goto done_queue; 1176204076Spjd } 1177204076Spjd /* 1178204076Spjd * Move the request to recv queue before sending it, because 1179204076Spjd * in different order we can get reply before we move request 1180204076Spjd * to recv queue. 1181204076Spjd */ 1182204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1183204076Spjd wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1184204076Spjd TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1185204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1186204076Spjd if (hast_proto_send(res, res->hr_remoteout, nv, data, 1187204076Spjd data != NULL ? length : 0) < 0) { 1188204076Spjd hio->hio_errors[ncomp] = errno; 1189204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1190204076Spjd remote_close(res, ncomp); 1191204076Spjd pjdlog_debug(2, 1192204076Spjd "remote_send: (%p) Unable to send request.", hio); 1193204076Spjd reqlog(LOG_ERR, 0, ggio, 1194204076Spjd "Unable to send request (%s): ", 1195204076Spjd strerror(hio->hio_errors[ncomp])); 1196204076Spjd /* 1197204076Spjd * Take request back from the receive queue and move 1198204076Spjd * it immediately to the done queue. 1199204076Spjd */ 1200204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1201204076Spjd TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1202204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1203204076Spjd goto done_queue; 1204204076Spjd } 1205204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1206204076Spjd nv_free(nv); 1207204076Spjd if (wakeup) 1208204076Spjd cv_signal(&hio_recv_list_cond[ncomp]); 1209204076Spjd continue; 1210204076Spjddone_queue: 1211204076Spjd nv_free(nv); 1212204076Spjd if (ISSYNCREQ(hio)) { 1213204076Spjd if (!refcount_release(&hio->hio_countdown)) 1214204076Spjd continue; 1215204076Spjd mtx_lock(&sync_lock); 1216204076Spjd SYNCREQDONE(hio); 1217204076Spjd mtx_unlock(&sync_lock); 1218204076Spjd cv_signal(&sync_cond); 1219204076Spjd continue; 1220204076Spjd } 1221204076Spjd if (ggio->gctl_cmd == BIO_WRITE) { 1222204076Spjd mtx_lock(&res->hr_amp_lock); 1223204076Spjd if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1224204076Spjd ggio->gctl_length)) { 1225204076Spjd (void)hast_activemap_flush(res); 1226204076Spjd } 1227204076Spjd mtx_unlock(&res->hr_amp_lock); 1228204076Spjd } 1229204076Spjd if (!refcount_release(&hio->hio_countdown)) 1230204076Spjd continue; 1231204076Spjd pjdlog_debug(2, 1232204076Spjd "remote_send: (%p) Moving request to the done queue.", 1233204076Spjd hio); 1234204076Spjd QUEUE_INSERT2(hio, done); 1235204076Spjd } 1236204076Spjd /* NOTREACHED */ 1237204076Spjd return (NULL); 1238204076Spjd} 1239204076Spjd 1240204076Spjd/* 1241204076Spjd * Thread receives answer from secondary node and passes it to ggate_send 1242204076Spjd * thread. 1243204076Spjd */ 1244204076Spjdstatic void * 1245204076Spjdremote_recv_thread(void *arg) 1246204076Spjd{ 1247204076Spjd struct hast_resource *res = arg; 1248204076Spjd struct g_gate_ctl_io *ggio; 1249204076Spjd struct hio *hio; 1250204076Spjd struct nv *nv; 1251204076Spjd unsigned int ncomp; 1252204076Spjd uint64_t seq; 1253204076Spjd int error; 1254204076Spjd 1255204076Spjd /* Remote component is 1 for now. */ 1256204076Spjd ncomp = 1; 1257204076Spjd 1258204076Spjd for (;;) { 1259204076Spjd /* Wait until there is anything to receive. */ 1260204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1261204076Spjd while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1262204076Spjd pjdlog_debug(2, "remote_recv: No requests, waiting."); 1263204076Spjd cv_wait(&hio_recv_list_cond[ncomp], 1264204076Spjd &hio_recv_list_lock[ncomp]); 1265204076Spjd } 1266204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1267204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1268204076Spjd if (!ISCONNECTED(res, ncomp)) { 1269204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1270204076Spjd /* 1271204076Spjd * Connection is dead, so move all pending requests to 1272204076Spjd * the done queue (one-by-one). 1273204076Spjd */ 1274204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1275204076Spjd hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1276204076Spjd assert(hio != NULL); 1277204076Spjd TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1278204076Spjd hio_next[ncomp]); 1279204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1280204076Spjd goto done_queue; 1281204076Spjd } 1282204076Spjd if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1283204076Spjd pjdlog_errno(LOG_ERR, 1284204076Spjd "Unable to receive reply header"); 1285204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1286204076Spjd remote_close(res, ncomp); 1287204076Spjd continue; 1288204076Spjd } 1289204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1290204076Spjd seq = nv_get_uint64(nv, "seq"); 1291204076Spjd if (seq == 0) { 1292204076Spjd pjdlog_error("Header contains no 'seq' field."); 1293204076Spjd nv_free(nv); 1294204076Spjd continue; 1295204076Spjd } 1296204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1297204076Spjd TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1298204076Spjd if (hio->hio_ggio.gctl_seq == seq) { 1299204076Spjd TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1300204076Spjd hio_next[ncomp]); 1301204076Spjd break; 1302204076Spjd } 1303204076Spjd } 1304204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1305204076Spjd if (hio == NULL) { 1306204076Spjd pjdlog_error("Found no request matching received 'seq' field (%ju).", 1307204076Spjd (uintmax_t)seq); 1308204076Spjd nv_free(nv); 1309204076Spjd continue; 1310204076Spjd } 1311204076Spjd error = nv_get_int16(nv, "error"); 1312204076Spjd if (error != 0) { 1313204076Spjd /* Request failed on remote side. */ 1314204076Spjd hio->hio_errors[ncomp] = 0; 1315204076Spjd nv_free(nv); 1316204076Spjd goto done_queue; 1317204076Spjd } 1318204076Spjd ggio = &hio->hio_ggio; 1319204076Spjd switch (ggio->gctl_cmd) { 1320204076Spjd case BIO_READ: 1321204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1322204076Spjd if (!ISCONNECTED(res, ncomp)) { 1323204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1324204076Spjd nv_free(nv); 1325204076Spjd goto done_queue; 1326204076Spjd } 1327204076Spjd if (hast_proto_recv_data(res, res->hr_remotein, nv, 1328204076Spjd ggio->gctl_data, ggio->gctl_length) < 0) { 1329204076Spjd hio->hio_errors[ncomp] = errno; 1330204076Spjd pjdlog_errno(LOG_ERR, 1331204076Spjd "Unable to receive reply data"); 1332204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1333204076Spjd nv_free(nv); 1334204076Spjd remote_close(res, ncomp); 1335204076Spjd goto done_queue; 1336204076Spjd } 1337204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1338204076Spjd break; 1339204076Spjd case BIO_WRITE: 1340204076Spjd case BIO_DELETE: 1341204076Spjd case BIO_FLUSH: 1342204076Spjd break; 1343204076Spjd default: 1344204076Spjd assert(!"invalid condition"); 1345204076Spjd abort(); 1346204076Spjd } 1347204076Spjd hio->hio_errors[ncomp] = 0; 1348204076Spjd nv_free(nv); 1349204076Spjddone_queue: 1350204076Spjd if (refcount_release(&hio->hio_countdown)) { 1351204076Spjd if (ISSYNCREQ(hio)) { 1352204076Spjd mtx_lock(&sync_lock); 1353204076Spjd SYNCREQDONE(hio); 1354204076Spjd mtx_unlock(&sync_lock); 1355204076Spjd cv_signal(&sync_cond); 1356204076Spjd } else { 1357204076Spjd pjdlog_debug(2, 1358204076Spjd "remote_recv: (%p) Moving request to the done queue.", 1359204076Spjd hio); 1360204076Spjd QUEUE_INSERT2(hio, done); 1361204076Spjd } 1362204076Spjd } 1363204076Spjd } 1364204076Spjd /* NOTREACHED */ 1365204076Spjd return (NULL); 1366204076Spjd} 1367204076Spjd 1368204076Spjd/* 1369204076Spjd * Thread sends answer to the kernel. 1370204076Spjd */ 1371204076Spjdstatic void * 1372204076Spjdggate_send_thread(void *arg) 1373204076Spjd{ 1374204076Spjd struct hast_resource *res = arg; 1375204076Spjd struct g_gate_ctl_io *ggio; 1376204076Spjd struct hio *hio; 1377204076Spjd unsigned int ii, ncomp, ncomps; 1378204076Spjd 1379204076Spjd ncomps = HAST_NCOMPONENTS; 1380204076Spjd 1381204076Spjd for (;;) { 1382204076Spjd pjdlog_debug(2, "ggate_send: Taking request."); 1383204076Spjd QUEUE_TAKE2(hio, done); 1384204076Spjd pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1385204076Spjd ggio = &hio->hio_ggio; 1386204076Spjd for (ii = 0; ii < ncomps; ii++) { 1387204076Spjd if (hio->hio_errors[ii] == 0) { 1388204076Spjd /* 1389204076Spjd * One successful request is enough to declare 1390204076Spjd * success. 1391204076Spjd */ 1392204076Spjd ggio->gctl_error = 0; 1393204076Spjd break; 1394204076Spjd } 1395204076Spjd } 1396204076Spjd if (ii == ncomps) { 1397204076Spjd /* 1398204076Spjd * None of the requests were successful. 1399204076Spjd * Use first error. 1400204076Spjd */ 1401204076Spjd ggio->gctl_error = hio->hio_errors[0]; 1402204076Spjd } 1403204076Spjd if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1404204076Spjd mtx_lock(&res->hr_amp_lock); 1405204076Spjd activemap_write_complete(res->hr_amp, 1406204076Spjd ggio->gctl_offset, ggio->gctl_length); 1407204076Spjd mtx_unlock(&res->hr_amp_lock); 1408204076Spjd } 1409204076Spjd if (ggio->gctl_cmd == BIO_WRITE) { 1410204076Spjd /* 1411204076Spjd * Unlock range we locked. 1412204076Spjd */ 1413204076Spjd mtx_lock(&range_lock); 1414204076Spjd rangelock_del(range_regular, ggio->gctl_offset, 1415204076Spjd ggio->gctl_length); 1416204076Spjd if (range_sync_wait) 1417204076Spjd cv_signal(&range_sync_cond); 1418204076Spjd mtx_unlock(&range_lock); 1419204076Spjd /* 1420204076Spjd * Bump local count if this is first write after 1421204076Spjd * connection failure with remote node. 1422204076Spjd */ 1423204076Spjd ncomp = 1; 1424204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1425204076Spjd if (!ISCONNECTED(res, ncomp)) { 1426204076Spjd mtx_lock(&metadata_lock); 1427204076Spjd if (res->hr_primary_localcnt == 1428204076Spjd res->hr_secondary_remotecnt) { 1429204076Spjd res->hr_primary_localcnt++; 1430204076Spjd pjdlog_debug(1, 1431204076Spjd "Increasing localcnt to %ju.", 1432204076Spjd (uintmax_t)res->hr_primary_localcnt); 1433204076Spjd (void)metadata_write(res); 1434204076Spjd } 1435204076Spjd mtx_unlock(&metadata_lock); 1436204076Spjd } 1437204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1438204076Spjd } 1439204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1440204076Spjd primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1441204076Spjd pjdlog_debug(2, 1442204076Spjd "ggate_send: (%p) Moving request to the free queue.", hio); 1443204076Spjd QUEUE_INSERT2(hio, free); 1444204076Spjd } 1445204076Spjd /* NOTREACHED */ 1446204076Spjd return (NULL); 1447204076Spjd} 1448204076Spjd 1449204076Spjd/* 1450204076Spjd * Thread synchronize local and remote components. 1451204076Spjd */ 1452204076Spjdstatic void * 1453204076Spjdsync_thread(void *arg __unused) 1454204076Spjd{ 1455204076Spjd struct hast_resource *res = arg; 1456204076Spjd struct hio *hio; 1457204076Spjd struct g_gate_ctl_io *ggio; 1458204076Spjd unsigned int ii, ncomp, ncomps; 1459204076Spjd off_t offset, length, synced; 1460204076Spjd bool dorewind; 1461204076Spjd int syncext; 1462204076Spjd 1463204076Spjd ncomps = HAST_NCOMPONENTS; 1464204076Spjd dorewind = true; 1465204076Spjd synced = 0; 1466204076Spjd 1467204076Spjd for (;;) { 1468204076Spjd mtx_lock(&sync_lock); 1469204076Spjd while (!sync_inprogress) { 1470204076Spjd dorewind = true; 1471204076Spjd synced = 0; 1472204076Spjd cv_wait(&sync_cond, &sync_lock); 1473204076Spjd } 1474204076Spjd mtx_unlock(&sync_lock); 1475204076Spjd /* 1476204076Spjd * Obtain offset at which we should synchronize. 1477204076Spjd * Rewind synchronization if needed. 1478204076Spjd */ 1479204076Spjd mtx_lock(&res->hr_amp_lock); 1480204076Spjd if (dorewind) 1481204076Spjd activemap_sync_rewind(res->hr_amp); 1482204076Spjd offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1483204076Spjd if (syncext != -1) { 1484204076Spjd /* 1485204076Spjd * We synchronized entire syncext extent, we can mark 1486204076Spjd * it as clean now. 1487204076Spjd */ 1488204076Spjd if (activemap_extent_complete(res->hr_amp, syncext)) 1489204076Spjd (void)hast_activemap_flush(res); 1490204076Spjd } 1491204076Spjd mtx_unlock(&res->hr_amp_lock); 1492204076Spjd if (dorewind) { 1493204076Spjd dorewind = false; 1494204076Spjd if (offset < 0) 1495204076Spjd pjdlog_info("Nodes are in sync."); 1496204076Spjd else { 1497204076Spjd pjdlog_info("Synchronization started. %ju bytes to go.", 1498204076Spjd (uintmax_t)(res->hr_extentsize * 1499204076Spjd activemap_ndirty(res->hr_amp))); 1500204076Spjd } 1501204076Spjd } 1502204076Spjd if (offset < 0) { 1503204076Spjd mtx_lock(&sync_lock); 1504204076Spjd sync_inprogress = false; 1505204076Spjd mtx_unlock(&sync_lock); 1506204076Spjd pjdlog_debug(1, "Nothing to synchronize."); 1507204076Spjd /* 1508204076Spjd * Synchronization complete, make both localcnt and 1509204076Spjd * remotecnt equal. 1510204076Spjd */ 1511204076Spjd ncomp = 1; 1512204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1513204076Spjd if (ISCONNECTED(res, ncomp)) { 1514204076Spjd if (synced > 0) { 1515204076Spjd pjdlog_info("Synchronization complete. " 1516204076Spjd "%jd bytes synchronized.", 1517204076Spjd (intmax_t)synced); 1518204076Spjd } 1519204076Spjd mtx_lock(&metadata_lock); 1520204076Spjd res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1521204076Spjd res->hr_primary_localcnt = 1522204076Spjd res->hr_secondary_localcnt; 1523204076Spjd res->hr_primary_remotecnt = 1524204076Spjd res->hr_secondary_remotecnt; 1525204076Spjd pjdlog_debug(1, 1526204076Spjd "Setting localcnt to %ju and remotecnt to %ju.", 1527204076Spjd (uintmax_t)res->hr_primary_localcnt, 1528204076Spjd (uintmax_t)res->hr_secondary_localcnt); 1529204076Spjd (void)metadata_write(res); 1530204076Spjd mtx_unlock(&metadata_lock); 1531204076Spjd } else if (synced > 0) { 1532204076Spjd pjdlog_info("Synchronization interrupted. " 1533204076Spjd "%jd bytes synchronized so far.", 1534204076Spjd (intmax_t)synced); 1535204076Spjd } 1536204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1537204076Spjd continue; 1538204076Spjd } 1539204076Spjd pjdlog_debug(2, "sync: Taking free request."); 1540204076Spjd QUEUE_TAKE2(hio, free); 1541204076Spjd pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1542204076Spjd /* 1543204076Spjd * Lock the range we are going to synchronize. We don't want 1544204076Spjd * race where someone writes between our read and write. 1545204076Spjd */ 1546204076Spjd for (;;) { 1547204076Spjd mtx_lock(&range_lock); 1548204076Spjd if (rangelock_islocked(range_regular, offset, length)) { 1549204076Spjd pjdlog_debug(2, 1550204076Spjd "sync: Range offset=%jd length=%jd locked.", 1551204076Spjd (intmax_t)offset, (intmax_t)length); 1552204076Spjd range_sync_wait = true; 1553204076Spjd cv_wait(&range_sync_cond, &range_lock); 1554204076Spjd range_sync_wait = false; 1555204076Spjd mtx_unlock(&range_lock); 1556204076Spjd continue; 1557204076Spjd } 1558204076Spjd if (rangelock_add(range_sync, offset, length) < 0) { 1559204076Spjd mtx_unlock(&range_lock); 1560204076Spjd pjdlog_debug(2, 1561204076Spjd "sync: Range offset=%jd length=%jd is already locked, waiting.", 1562204076Spjd (intmax_t)offset, (intmax_t)length); 1563204076Spjd sleep(1); 1564204076Spjd continue; 1565204076Spjd } 1566204076Spjd mtx_unlock(&range_lock); 1567204076Spjd break; 1568204076Spjd } 1569204076Spjd /* 1570204076Spjd * First read the data from synchronization source. 1571204076Spjd */ 1572204076Spjd SYNCREQ(hio); 1573204076Spjd ggio = &hio->hio_ggio; 1574204076Spjd ggio->gctl_cmd = BIO_READ; 1575204076Spjd ggio->gctl_offset = offset; 1576204076Spjd ggio->gctl_length = length; 1577204076Spjd ggio->gctl_error = 0; 1578204076Spjd for (ii = 0; ii < ncomps; ii++) 1579204076Spjd hio->hio_errors[ii] = EINVAL; 1580204076Spjd reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1581204076Spjd hio); 1582204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1583204076Spjd hio); 1584204076Spjd mtx_lock(&metadata_lock); 1585204076Spjd if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1586204076Spjd /* 1587204076Spjd * This range is up-to-date on local component, 1588204076Spjd * so handle request locally. 1589204076Spjd */ 1590204076Spjd /* Local component is 0 for now. */ 1591204076Spjd ncomp = 0; 1592204076Spjd } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1593204076Spjd assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1594204076Spjd /* 1595204076Spjd * This range is out-of-date on local component, 1596204076Spjd * so send request to the remote node. 1597204076Spjd */ 1598204076Spjd /* Remote component is 1 for now. */ 1599204076Spjd ncomp = 1; 1600204076Spjd } 1601204076Spjd mtx_unlock(&metadata_lock); 1602204076Spjd refcount_init(&hio->hio_countdown, 1); 1603204076Spjd QUEUE_INSERT1(hio, send, ncomp); 1604204076Spjd 1605204076Spjd /* 1606204076Spjd * Let's wait for READ to finish. 1607204076Spjd */ 1608204076Spjd mtx_lock(&sync_lock); 1609204076Spjd while (!ISSYNCREQDONE(hio)) 1610204076Spjd cv_wait(&sync_cond, &sync_lock); 1611204076Spjd mtx_unlock(&sync_lock); 1612204076Spjd 1613204076Spjd if (hio->hio_errors[ncomp] != 0) { 1614204076Spjd pjdlog_error("Unable to read synchronization data: %s.", 1615204076Spjd strerror(hio->hio_errors[ncomp])); 1616204076Spjd goto free_queue; 1617204076Spjd } 1618204076Spjd 1619204076Spjd /* 1620204076Spjd * We read the data from synchronization source, now write it 1621204076Spjd * to synchronization target. 1622204076Spjd */ 1623204076Spjd SYNCREQ(hio); 1624204076Spjd ggio->gctl_cmd = BIO_WRITE; 1625204076Spjd for (ii = 0; ii < ncomps; ii++) 1626204076Spjd hio->hio_errors[ii] = EINVAL; 1627204076Spjd reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1628204076Spjd hio); 1629204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1630204076Spjd hio); 1631204076Spjd mtx_lock(&metadata_lock); 1632204076Spjd if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1633204076Spjd /* 1634204076Spjd * This range is up-to-date on local component, 1635204076Spjd * so we update remote component. 1636204076Spjd */ 1637204076Spjd /* Remote component is 1 for now. */ 1638204076Spjd ncomp = 1; 1639204076Spjd } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1640204076Spjd assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1641204076Spjd /* 1642204076Spjd * This range is out-of-date on local component, 1643204076Spjd * so we update it. 1644204076Spjd */ 1645204076Spjd /* Local component is 0 for now. */ 1646204076Spjd ncomp = 0; 1647204076Spjd } 1648204076Spjd mtx_unlock(&metadata_lock); 1649204076Spjd 1650204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1651204076Spjd hio); 1652204076Spjd refcount_init(&hio->hio_countdown, 1); 1653204076Spjd QUEUE_INSERT1(hio, send, ncomp); 1654204076Spjd 1655204076Spjd /* 1656204076Spjd * Let's wait for WRITE to finish. 1657204076Spjd */ 1658204076Spjd mtx_lock(&sync_lock); 1659204076Spjd while (!ISSYNCREQDONE(hio)) 1660204076Spjd cv_wait(&sync_cond, &sync_lock); 1661204076Spjd mtx_unlock(&sync_lock); 1662204076Spjd 1663204076Spjd if (hio->hio_errors[ncomp] != 0) { 1664204076Spjd pjdlog_error("Unable to write synchronization data: %s.", 1665204076Spjd strerror(hio->hio_errors[ncomp])); 1666204076Spjd goto free_queue; 1667204076Spjd } 1668204076Spjdfree_queue: 1669204076Spjd mtx_lock(&range_lock); 1670204076Spjd rangelock_del(range_sync, offset, length); 1671204076Spjd if (range_regular_wait) 1672204076Spjd cv_signal(&range_regular_cond); 1673204076Spjd mtx_unlock(&range_lock); 1674204076Spjd 1675204076Spjd synced += length; 1676204076Spjd 1677204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1678204076Spjd hio); 1679204076Spjd QUEUE_INSERT2(hio, free); 1680204076Spjd } 1681204076Spjd /* NOTREACHED */ 1682204076Spjd return (NULL); 1683204076Spjd} 1684204076Spjd 1685204076Spjdstatic void 1686204076Spjdsighandler(int sig) 1687204076Spjd{ 1688204076Spjd bool unlock; 1689204076Spjd 1690204076Spjd switch (sig) { 1691204076Spjd case SIGINT: 1692204076Spjd case SIGTERM: 1693204076Spjd sigexit_received = true; 1694204076Spjd break; 1695204076Spjd default: 1696204076Spjd assert(!"invalid condition"); 1697204076Spjd } 1698204076Spjd /* 1699204076Spjd * XXX: Racy, but if we cannot obtain hio_guard_lock here, we don't 1700204076Spjd * want to risk deadlock. 1701204076Spjd */ 1702204076Spjd unlock = mtx_trylock(&hio_guard_lock); 1703204076Spjd cv_signal(&hio_guard_cond); 1704204076Spjd if (unlock) 1705204076Spjd mtx_unlock(&hio_guard_lock); 1706204076Spjd} 1707204076Spjd 1708204076Spjd/* 1709204076Spjd * Thread guards remote connections and reconnects when needed, handles 1710204076Spjd * signals, etc. 1711204076Spjd */ 1712204076Spjdstatic void * 1713204076Spjdguard_thread(void *arg) 1714204076Spjd{ 1715204076Spjd struct hast_resource *res = arg; 1716205738Spjd struct proto_conn *in, *out; 1717204076Spjd unsigned int ii, ncomps; 1718204076Spjd int timeout; 1719204076Spjd 1720204076Spjd ncomps = HAST_NCOMPONENTS; 1721204076Spjd /* The is only one remote component for now. */ 1722204076Spjd#define ISREMOTE(no) ((no) == 1) 1723204076Spjd 1724204076Spjd for (;;) { 1725204076Spjd if (sigexit_received) { 1726204076Spjd primary_exitx(EX_OK, 1727204076Spjd "Termination signal received, exiting."); 1728204076Spjd } 1729204076Spjd /* 1730204076Spjd * If all the connection will be fine, we will sleep until 1731204076Spjd * someone wakes us up. 1732204076Spjd * If any of the connections will be broken and we won't be 1733204076Spjd * able to connect, we will sleep only for RECONNECT_SLEEP 1734204076Spjd * seconds so we can retry soon. 1735204076Spjd */ 1736204076Spjd timeout = 0; 1737204076Spjd pjdlog_debug(2, "remote_guard: Checking connections."); 1738204076Spjd mtx_lock(&hio_guard_lock); 1739204076Spjd for (ii = 0; ii < ncomps; ii++) { 1740204076Spjd if (!ISREMOTE(ii)) 1741204076Spjd continue; 1742204076Spjd rw_rlock(&hio_remote_lock[ii]); 1743204076Spjd if (ISCONNECTED(res, ii)) { 1744204076Spjd assert(res->hr_remotein != NULL); 1745204076Spjd assert(res->hr_remoteout != NULL); 1746204076Spjd rw_unlock(&hio_remote_lock[ii]); 1747204076Spjd pjdlog_debug(2, 1748204076Spjd "remote_guard: Connection to %s is ok.", 1749204076Spjd res->hr_remoteaddr); 1750204076Spjd } else { 1751204076Spjd assert(res->hr_remotein == NULL); 1752204076Spjd assert(res->hr_remoteout == NULL); 1753204076Spjd /* 1754204076Spjd * Upgrade the lock. It doesn't have to be 1755204076Spjd * atomic as no other thread can change 1756204076Spjd * connection status from disconnected to 1757204076Spjd * connected. 1758204076Spjd */ 1759204076Spjd rw_unlock(&hio_remote_lock[ii]); 1760204076Spjd pjdlog_debug(2, 1761204076Spjd "remote_guard: Reconnecting to %s.", 1762204076Spjd res->hr_remoteaddr); 1763205738Spjd in = out = NULL; 1764205738Spjd if (init_remote(res, &in, &out)) { 1765205738Spjd rw_wlock(&hio_remote_lock[ii]); 1766205738Spjd assert(res->hr_remotein == NULL); 1767205738Spjd assert(res->hr_remoteout == NULL); 1768205738Spjd assert(in != NULL && out != NULL); 1769205738Spjd res->hr_remotein = in; 1770205738Spjd res->hr_remoteout = out; 1771205738Spjd rw_unlock(&hio_remote_lock[ii]); 1772204076Spjd pjdlog_info("Successfully reconnected to %s.", 1773204076Spjd res->hr_remoteaddr); 1774205738Spjd sync_start(); 1775204076Spjd } else { 1776204076Spjd /* Both connections should be NULL. */ 1777204076Spjd assert(res->hr_remotein == NULL); 1778204076Spjd assert(res->hr_remoteout == NULL); 1779205738Spjd assert(in == NULL && out == NULL); 1780204076Spjd pjdlog_debug(2, 1781204076Spjd "remote_guard: Reconnect to %s failed.", 1782204076Spjd res->hr_remoteaddr); 1783204076Spjd timeout = RECONNECT_SLEEP; 1784204076Spjd } 1785204076Spjd } 1786204076Spjd } 1787204076Spjd (void)cv_timedwait(&hio_guard_cond, &hio_guard_lock, timeout); 1788204076Spjd mtx_unlock(&hio_guard_lock); 1789204076Spjd } 1790204076Spjd#undef ISREMOTE 1791204076Spjd /* NOTREACHED */ 1792204076Spjd return (NULL); 1793204076Spjd} 1794