primary.c revision 204076
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 204076 2010-02-18 23:16:19Z 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 463204076Spjdstatic void 464204076Spjdinit_remote(struct hast_resource *res) 465204076Spjd{ 466204076Spjd struct nv *nvout, *nvin; 467204076Spjd const unsigned char *token; 468204076Spjd unsigned char *map; 469204076Spjd const char *errmsg; 470204076Spjd int32_t extentsize; 471204076Spjd int64_t datasize; 472204076Spjd uint32_t mapsize; 473204076Spjd size_t size; 474204076Spjd 475204076Spjd /* Prepare outgoing connection with remote node. */ 476204076Spjd if (proto_client(res->hr_remoteaddr, &res->hr_remoteout) < 0) { 477204076Spjd primary_exit(EX_OSERR, "Unable to create connection to %s", 478204076Spjd res->hr_remoteaddr); 479204076Spjd } 480204076Spjd /* Try to connect, but accept failure. */ 481204076Spjd if (proto_connect(res->hr_remoteout) < 0) { 482204076Spjd pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 483204076Spjd res->hr_remoteaddr); 484204076Spjd goto close; 485204076Spjd } 486204076Spjd /* 487204076Spjd * First handshake step. 488204076Spjd * Setup outgoing connection with remote node. 489204076Spjd */ 490204076Spjd nvout = nv_alloc(); 491204076Spjd nv_add_string(nvout, res->hr_name, "resource"); 492204076Spjd if (nv_error(nvout) != 0) { 493204076Spjd pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 494204076Spjd "Unable to allocate header for connection with %s", 495204076Spjd res->hr_remoteaddr); 496204076Spjd nv_free(nvout); 497204076Spjd goto close; 498204076Spjd } 499204076Spjd if (hast_proto_send(res, res->hr_remoteout, nvout, NULL, 0) < 0) { 500204076Spjd pjdlog_errno(LOG_WARNING, 501204076Spjd "Unable to send handshake header to %s", 502204076Spjd res->hr_remoteaddr); 503204076Spjd nv_free(nvout); 504204076Spjd goto close; 505204076Spjd } 506204076Spjd nv_free(nvout); 507204076Spjd if (hast_proto_recv_hdr(res->hr_remoteout, &nvin) < 0) { 508204076Spjd pjdlog_errno(LOG_WARNING, 509204076Spjd "Unable to receive handshake header from %s", 510204076Spjd res->hr_remoteaddr); 511204076Spjd goto close; 512204076Spjd } 513204076Spjd errmsg = nv_get_string(nvin, "errmsg"); 514204076Spjd if (errmsg != NULL) { 515204076Spjd pjdlog_warning("%s", errmsg); 516204076Spjd nv_free(nvin); 517204076Spjd goto close; 518204076Spjd } 519204076Spjd token = nv_get_uint8_array(nvin, &size, "token"); 520204076Spjd if (token == NULL) { 521204076Spjd pjdlog_warning("Handshake header from %s has no 'token' field.", 522204076Spjd res->hr_remoteaddr); 523204076Spjd nv_free(nvin); 524204076Spjd goto close; 525204076Spjd } 526204076Spjd if (size != sizeof(res->hr_token)) { 527204076Spjd pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 528204076Spjd res->hr_remoteaddr, size, sizeof(res->hr_token)); 529204076Spjd nv_free(nvin); 530204076Spjd goto close; 531204076Spjd } 532204076Spjd bcopy(token, res->hr_token, sizeof(res->hr_token)); 533204076Spjd nv_free(nvin); 534204076Spjd 535204076Spjd /* 536204076Spjd * Second handshake step. 537204076Spjd * Setup incoming connection with remote node. 538204076Spjd */ 539204076Spjd if (proto_client(res->hr_remoteaddr, &res->hr_remotein) < 0) { 540204076Spjd pjdlog_errno(LOG_WARNING, "Unable to create connection to %s", 541204076Spjd res->hr_remoteaddr); 542204076Spjd } 543204076Spjd /* Try to connect, but accept failure. */ 544204076Spjd if (proto_connect(res->hr_remotein) < 0) { 545204076Spjd pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 546204076Spjd res->hr_remoteaddr); 547204076Spjd goto close; 548204076Spjd } 549204076Spjd nvout = nv_alloc(); 550204076Spjd nv_add_string(nvout, res->hr_name, "resource"); 551204076Spjd nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 552204076Spjd "token"); 553204076Spjd nv_add_uint64(nvout, res->hr_resuid, "resuid"); 554204076Spjd nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 555204076Spjd nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 556204076Spjd if (nv_error(nvout) != 0) { 557204076Spjd pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 558204076Spjd "Unable to allocate header for connection with %s", 559204076Spjd res->hr_remoteaddr); 560204076Spjd nv_free(nvout); 561204076Spjd goto close; 562204076Spjd } 563204076Spjd if (hast_proto_send(res, res->hr_remotein, nvout, NULL, 0) < 0) { 564204076Spjd pjdlog_errno(LOG_WARNING, 565204076Spjd "Unable to send handshake header to %s", 566204076Spjd res->hr_remoteaddr); 567204076Spjd nv_free(nvout); 568204076Spjd goto close; 569204076Spjd } 570204076Spjd nv_free(nvout); 571204076Spjd if (hast_proto_recv_hdr(res->hr_remoteout, &nvin) < 0) { 572204076Spjd pjdlog_errno(LOG_WARNING, 573204076Spjd "Unable to receive handshake header from %s", 574204076Spjd res->hr_remoteaddr); 575204076Spjd goto close; 576204076Spjd } 577204076Spjd errmsg = nv_get_string(nvin, "errmsg"); 578204076Spjd if (errmsg != NULL) { 579204076Spjd pjdlog_warning("%s", errmsg); 580204076Spjd nv_free(nvin); 581204076Spjd goto close; 582204076Spjd } 583204076Spjd datasize = nv_get_int64(nvin, "datasize"); 584204076Spjd if (datasize != res->hr_datasize) { 585204076Spjd pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 586204076Spjd (intmax_t)res->hr_datasize, (intmax_t)datasize); 587204076Spjd nv_free(nvin); 588204076Spjd goto close; 589204076Spjd } 590204076Spjd extentsize = nv_get_int32(nvin, "extentsize"); 591204076Spjd if (extentsize != res->hr_extentsize) { 592204076Spjd pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 593204076Spjd (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 594204076Spjd nv_free(nvin); 595204076Spjd goto close; 596204076Spjd } 597204076Spjd res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 598204076Spjd res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 599204076Spjd res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 600204076Spjd map = NULL; 601204076Spjd mapsize = nv_get_uint32(nvin, "mapsize"); 602204076Spjd if (mapsize > 0) { 603204076Spjd map = malloc(mapsize); 604204076Spjd if (map == NULL) { 605204076Spjd pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 606204076Spjd (uintmax_t)mapsize); 607204076Spjd nv_free(nvin); 608204076Spjd goto close; 609204076Spjd } 610204076Spjd /* 611204076Spjd * Remote node have some dirty extents on its own, lets 612204076Spjd * download its activemap. 613204076Spjd */ 614204076Spjd if (hast_proto_recv_data(res, res->hr_remoteout, nvin, map, 615204076Spjd mapsize) < 0) { 616204076Spjd pjdlog_errno(LOG_ERR, 617204076Spjd "Unable to receive remote activemap"); 618204076Spjd nv_free(nvin); 619204076Spjd free(map); 620204076Spjd goto close; 621204076Spjd } 622204076Spjd /* 623204076Spjd * Merge local and remote bitmaps. 624204076Spjd */ 625204076Spjd activemap_merge(res->hr_amp, map, mapsize); 626204076Spjd free(map); 627204076Spjd /* 628204076Spjd * Now that we merged bitmaps from both nodes, flush it to the 629204076Spjd * disk before we start to synchronize. 630204076Spjd */ 631204076Spjd (void)hast_activemap_flush(res); 632204076Spjd } 633204076Spjd pjdlog_info("Connected to %s.", res->hr_remoteaddr); 634204076Spjd mtx_lock(&sync_lock); 635204076Spjd sync_inprogress = true; 636204076Spjd mtx_unlock(&sync_lock); 637204076Spjd cv_signal(&sync_cond); 638204076Spjd return; 639204076Spjdclose: 640204076Spjd proto_close(res->hr_remoteout); 641204076Spjd res->hr_remoteout = NULL; 642204076Spjd if (res->hr_remotein != NULL) { 643204076Spjd proto_close(res->hr_remotein); 644204076Spjd res->hr_remotein = NULL; 645204076Spjd } 646204076Spjd} 647204076Spjd 648204076Spjdstatic void 649204076Spjdinit_ggate(struct hast_resource *res) 650204076Spjd{ 651204076Spjd struct g_gate_ctl_create ggiocreate; 652204076Spjd struct g_gate_ctl_cancel ggiocancel; 653204076Spjd 654204076Spjd /* 655204076Spjd * We communicate with ggate via /dev/ggctl. Open it. 656204076Spjd */ 657204076Spjd res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 658204076Spjd if (res->hr_ggatefd < 0) 659204076Spjd primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 660204076Spjd /* 661204076Spjd * Create provider before trying to connect, as connection failure 662204076Spjd * is not critical, but may take some time. 663204076Spjd */ 664204076Spjd ggiocreate.gctl_version = G_GATE_VERSION; 665204076Spjd ggiocreate.gctl_mediasize = res->hr_datasize; 666204076Spjd ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 667204076Spjd ggiocreate.gctl_flags = 0; 668204076Spjd ggiocreate.gctl_maxcount = 128; 669204076Spjd ggiocreate.gctl_timeout = 0; 670204076Spjd ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 671204076Spjd snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 672204076Spjd res->hr_provname); 673204076Spjd bzero(ggiocreate.gctl_info, sizeof(ggiocreate.gctl_info)); 674204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 675204076Spjd pjdlog_info("Device hast/%s created.", res->hr_provname); 676204076Spjd res->hr_ggateunit = ggiocreate.gctl_unit; 677204076Spjd return; 678204076Spjd } 679204076Spjd if (errno != EEXIST) { 680204076Spjd primary_exit(EX_OSERR, "Unable to create hast/%s device", 681204076Spjd res->hr_provname); 682204076Spjd } 683204076Spjd pjdlog_debug(1, 684204076Spjd "Device hast/%s already exists, we will try to take it over.", 685204076Spjd res->hr_provname); 686204076Spjd /* 687204076Spjd * If we received EEXIST, we assume that the process who created the 688204076Spjd * provider died and didn't clean up. In that case we will start from 689204076Spjd * where he left of. 690204076Spjd */ 691204076Spjd ggiocancel.gctl_version = G_GATE_VERSION; 692204076Spjd ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 693204076Spjd snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 694204076Spjd res->hr_provname); 695204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 696204076Spjd pjdlog_info("Device hast/%s recovered.", res->hr_provname); 697204076Spjd res->hr_ggateunit = ggiocancel.gctl_unit; 698204076Spjd return; 699204076Spjd } 700204076Spjd primary_exit(EX_OSERR, "Unable to take over hast/%s device", 701204076Spjd res->hr_provname); 702204076Spjd} 703204076Spjd 704204076Spjdvoid 705204076Spjdhastd_primary(struct hast_resource *res) 706204076Spjd{ 707204076Spjd pthread_t td; 708204076Spjd pid_t pid; 709204076Spjd int error; 710204076Spjd 711204076Spjd gres = res; 712204076Spjd 713204076Spjd /* 714204076Spjd * Create communication channel between parent and child. 715204076Spjd */ 716204076Spjd if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 717204076Spjd KEEP_ERRNO((void)pidfile_remove(pfh)); 718204076Spjd primary_exit(EX_OSERR, 719204076Spjd "Unable to create control sockets between parent and child"); 720204076Spjd } 721204076Spjd 722204076Spjd pid = fork(); 723204076Spjd if (pid < 0) { 724204076Spjd KEEP_ERRNO((void)pidfile_remove(pfh)); 725204076Spjd primary_exit(EX_OSERR, "Unable to fork"); 726204076Spjd } 727204076Spjd 728204076Spjd if (pid > 0) { 729204076Spjd /* This is parent. */ 730204076Spjd res->hr_workerpid = pid; 731204076Spjd return; 732204076Spjd } 733204076Spjd (void)pidfile_close(pfh); 734204076Spjd 735204076Spjd setproctitle("%s (primary)", res->hr_name); 736204076Spjd 737204076Spjd init_local(res); 738204076Spjd init_remote(res); 739204076Spjd init_ggate(res); 740204076Spjd init_environment(res); 741204076Spjd error = pthread_create(&td, NULL, ggate_recv_thread, res); 742204076Spjd assert(error == 0); 743204076Spjd error = pthread_create(&td, NULL, local_send_thread, res); 744204076Spjd assert(error == 0); 745204076Spjd error = pthread_create(&td, NULL, remote_send_thread, res); 746204076Spjd assert(error == 0); 747204076Spjd error = pthread_create(&td, NULL, remote_recv_thread, res); 748204076Spjd assert(error == 0); 749204076Spjd error = pthread_create(&td, NULL, ggate_send_thread, res); 750204076Spjd assert(error == 0); 751204076Spjd error = pthread_create(&td, NULL, sync_thread, res); 752204076Spjd assert(error == 0); 753204076Spjd error = pthread_create(&td, NULL, ctrl_thread, res); 754204076Spjd assert(error == 0); 755204076Spjd (void)guard_thread(res); 756204076Spjd} 757204076Spjd 758204076Spjdstatic void 759204076Spjdreqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 760204076Spjd{ 761204076Spjd char msg[1024]; 762204076Spjd va_list ap; 763204076Spjd int len; 764204076Spjd 765204076Spjd va_start(ap, fmt); 766204076Spjd len = vsnprintf(msg, sizeof(msg), fmt, ap); 767204076Spjd va_end(ap); 768204076Spjd if ((size_t)len < sizeof(msg)) { 769204076Spjd switch (ggio->gctl_cmd) { 770204076Spjd case BIO_READ: 771204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 772204076Spjd "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 773204076Spjd (uintmax_t)ggio->gctl_length); 774204076Spjd break; 775204076Spjd case BIO_DELETE: 776204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 777204076Spjd "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 778204076Spjd (uintmax_t)ggio->gctl_length); 779204076Spjd break; 780204076Spjd case BIO_FLUSH: 781204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 782204076Spjd break; 783204076Spjd case BIO_WRITE: 784204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 785204076Spjd "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 786204076Spjd (uintmax_t)ggio->gctl_length); 787204076Spjd break; 788204076Spjd default: 789204076Spjd (void)snprintf(msg + len, sizeof(msg) - len, 790204076Spjd "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 791204076Spjd break; 792204076Spjd } 793204076Spjd } 794204076Spjd pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 795204076Spjd} 796204076Spjd 797204076Spjdstatic void 798204076Spjdremote_close(struct hast_resource *res, int ncomp) 799204076Spjd{ 800204076Spjd 801204076Spjd rw_wlock(&hio_remote_lock[ncomp]); 802204076Spjd /* 803204076Spjd * A race is possible between dropping rlock and acquiring wlock - 804204076Spjd * another thread can close connection in-between. 805204076Spjd */ 806204076Spjd if (!ISCONNECTED(res, ncomp)) { 807204076Spjd assert(res->hr_remotein == NULL); 808204076Spjd assert(res->hr_remoteout == NULL); 809204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 810204076Spjd return; 811204076Spjd } 812204076Spjd 813204076Spjd assert(res->hr_remotein != NULL); 814204076Spjd assert(res->hr_remoteout != NULL); 815204076Spjd 816204076Spjd pjdlog_debug(2, "Closing old incoming connection to %s.", 817204076Spjd res->hr_remoteaddr); 818204076Spjd proto_close(res->hr_remotein); 819204076Spjd res->hr_remotein = NULL; 820204076Spjd pjdlog_debug(2, "Closing old outgoing connection to %s.", 821204076Spjd res->hr_remoteaddr); 822204076Spjd proto_close(res->hr_remoteout); 823204076Spjd res->hr_remoteout = NULL; 824204076Spjd 825204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 826204076Spjd 827204076Spjd /* 828204076Spjd * Stop synchronization if in-progress. 829204076Spjd */ 830204076Spjd mtx_lock(&sync_lock); 831204076Spjd if (sync_inprogress) 832204076Spjd sync_inprogress = false; 833204076Spjd mtx_unlock(&sync_lock); 834204076Spjd 835204076Spjd /* 836204076Spjd * Wake up guard thread, so it can immediately start reconnect. 837204076Spjd */ 838204076Spjd mtx_lock(&hio_guard_lock); 839204076Spjd cv_signal(&hio_guard_cond); 840204076Spjd mtx_unlock(&hio_guard_lock); 841204076Spjd} 842204076Spjd 843204076Spjd/* 844204076Spjd * Thread receives ggate I/O requests from the kernel and passes them to 845204076Spjd * appropriate threads: 846204076Spjd * WRITE - always goes to both local_send and remote_send threads 847204076Spjd * READ (when the block is up-to-date on local component) - 848204076Spjd * only local_send thread 849204076Spjd * READ (when the block isn't up-to-date on local component) - 850204076Spjd * only remote_send thread 851204076Spjd * DELETE - always goes to both local_send and remote_send threads 852204076Spjd * FLUSH - always goes to both local_send and remote_send threads 853204076Spjd */ 854204076Spjdstatic void * 855204076Spjdggate_recv_thread(void *arg) 856204076Spjd{ 857204076Spjd struct hast_resource *res = arg; 858204076Spjd struct g_gate_ctl_io *ggio; 859204076Spjd struct hio *hio; 860204076Spjd unsigned int ii, ncomp, ncomps; 861204076Spjd int error; 862204076Spjd 863204076Spjd ncomps = HAST_NCOMPONENTS; 864204076Spjd 865204076Spjd for (;;) { 866204076Spjd pjdlog_debug(2, "ggate_recv: Taking free request."); 867204076Spjd QUEUE_TAKE2(hio, free); 868204076Spjd pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 869204076Spjd ggio = &hio->hio_ggio; 870204076Spjd ggio->gctl_unit = res->hr_ggateunit; 871204076Spjd ggio->gctl_length = MAXPHYS; 872204076Spjd ggio->gctl_error = 0; 873204076Spjd pjdlog_debug(2, 874204076Spjd "ggate_recv: (%p) Waiting for request from the kernel.", 875204076Spjd hio); 876204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 877204076Spjd if (sigexit_received) 878204076Spjd pthread_exit(NULL); 879204076Spjd primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 880204076Spjd } 881204076Spjd error = ggio->gctl_error; 882204076Spjd switch (error) { 883204076Spjd case 0: 884204076Spjd break; 885204076Spjd case ECANCELED: 886204076Spjd /* Exit gracefully. */ 887204076Spjd if (!sigexit_received) { 888204076Spjd pjdlog_debug(2, 889204076Spjd "ggate_recv: (%p) Received cancel from the kernel.", 890204076Spjd hio); 891204076Spjd pjdlog_info("Received cancel from the kernel, exiting."); 892204076Spjd } 893204076Spjd pthread_exit(NULL); 894204076Spjd case ENOMEM: 895204076Spjd /* 896204076Spjd * Buffer too small? Impossible, we allocate MAXPHYS 897204076Spjd * bytes - request can't be bigger than that. 898204076Spjd */ 899204076Spjd /* FALLTHROUGH */ 900204076Spjd case ENXIO: 901204076Spjd default: 902204076Spjd primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 903204076Spjd strerror(error)); 904204076Spjd } 905204076Spjd for (ii = 0; ii < ncomps; ii++) 906204076Spjd hio->hio_errors[ii] = EINVAL; 907204076Spjd reqlog(LOG_DEBUG, 2, ggio, 908204076Spjd "ggate_recv: (%p) Request received from the kernel: ", 909204076Spjd hio); 910204076Spjd /* 911204076Spjd * Inform all components about new write request. 912204076Spjd * For read request prefer local component unless the given 913204076Spjd * range is out-of-date, then use remote component. 914204076Spjd */ 915204076Spjd switch (ggio->gctl_cmd) { 916204076Spjd case BIO_READ: 917204076Spjd pjdlog_debug(2, 918204076Spjd "ggate_recv: (%p) Moving request to the send queue.", 919204076Spjd hio); 920204076Spjd refcount_init(&hio->hio_countdown, 1); 921204076Spjd mtx_lock(&metadata_lock); 922204076Spjd if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 923204076Spjd res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 924204076Spjd /* 925204076Spjd * This range is up-to-date on local component, 926204076Spjd * so handle request locally. 927204076Spjd */ 928204076Spjd /* Local component is 0 for now. */ 929204076Spjd ncomp = 0; 930204076Spjd } else /* if (res->hr_syncsrc == 931204076Spjd HAST_SYNCSRC_SECONDARY) */ { 932204076Spjd assert(res->hr_syncsrc == 933204076Spjd HAST_SYNCSRC_SECONDARY); 934204076Spjd /* 935204076Spjd * This range is out-of-date on local component, 936204076Spjd * so send request to the remote node. 937204076Spjd */ 938204076Spjd /* Remote component is 1 for now. */ 939204076Spjd ncomp = 1; 940204076Spjd } 941204076Spjd mtx_unlock(&metadata_lock); 942204076Spjd QUEUE_INSERT1(hio, send, ncomp); 943204076Spjd break; 944204076Spjd case BIO_WRITE: 945204076Spjd for (;;) { 946204076Spjd mtx_lock(&range_lock); 947204076Spjd if (rangelock_islocked(range_sync, 948204076Spjd ggio->gctl_offset, ggio->gctl_length)) { 949204076Spjd pjdlog_debug(2, 950204076Spjd "regular: Range offset=%jd length=%zu locked.", 951204076Spjd (intmax_t)ggio->gctl_offset, 952204076Spjd (size_t)ggio->gctl_length); 953204076Spjd range_regular_wait = true; 954204076Spjd cv_wait(&range_regular_cond, &range_lock); 955204076Spjd range_regular_wait = false; 956204076Spjd mtx_unlock(&range_lock); 957204076Spjd continue; 958204076Spjd } 959204076Spjd if (rangelock_add(range_regular, 960204076Spjd ggio->gctl_offset, ggio->gctl_length) < 0) { 961204076Spjd mtx_unlock(&range_lock); 962204076Spjd pjdlog_debug(2, 963204076Spjd "regular: Range offset=%jd length=%zu is already locked, waiting.", 964204076Spjd (intmax_t)ggio->gctl_offset, 965204076Spjd (size_t)ggio->gctl_length); 966204076Spjd sleep(1); 967204076Spjd continue; 968204076Spjd } 969204076Spjd mtx_unlock(&range_lock); 970204076Spjd break; 971204076Spjd } 972204076Spjd mtx_lock(&res->hr_amp_lock); 973204076Spjd if (activemap_write_start(res->hr_amp, 974204076Spjd ggio->gctl_offset, ggio->gctl_length)) { 975204076Spjd (void)hast_activemap_flush(res); 976204076Spjd } 977204076Spjd mtx_unlock(&res->hr_amp_lock); 978204076Spjd /* FALLTHROUGH */ 979204076Spjd case BIO_DELETE: 980204076Spjd case BIO_FLUSH: 981204076Spjd pjdlog_debug(2, 982204076Spjd "ggate_recv: (%p) Moving request to the send queues.", 983204076Spjd hio); 984204076Spjd refcount_init(&hio->hio_countdown, ncomps); 985204076Spjd for (ii = 0; ii < ncomps; ii++) 986204076Spjd QUEUE_INSERT1(hio, send, ii); 987204076Spjd break; 988204076Spjd } 989204076Spjd } 990204076Spjd /* NOTREACHED */ 991204076Spjd return (NULL); 992204076Spjd} 993204076Spjd 994204076Spjd/* 995204076Spjd * Thread reads from or writes to local component. 996204076Spjd * If local read fails, it redirects it to remote_send thread. 997204076Spjd */ 998204076Spjdstatic void * 999204076Spjdlocal_send_thread(void *arg) 1000204076Spjd{ 1001204076Spjd struct hast_resource *res = arg; 1002204076Spjd struct g_gate_ctl_io *ggio; 1003204076Spjd struct hio *hio; 1004204076Spjd unsigned int ncomp, rncomp; 1005204076Spjd ssize_t ret; 1006204076Spjd 1007204076Spjd /* Local component is 0 for now. */ 1008204076Spjd ncomp = 0; 1009204076Spjd /* Remote component is 1 for now. */ 1010204076Spjd rncomp = 1; 1011204076Spjd 1012204076Spjd for (;;) { 1013204076Spjd pjdlog_debug(2, "local_send: Taking request."); 1014204076Spjd QUEUE_TAKE1(hio, send, ncomp); 1015204076Spjd pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1016204076Spjd ggio = &hio->hio_ggio; 1017204076Spjd switch (ggio->gctl_cmd) { 1018204076Spjd case BIO_READ: 1019204076Spjd ret = pread(res->hr_localfd, ggio->gctl_data, 1020204076Spjd ggio->gctl_length, 1021204076Spjd ggio->gctl_offset + res->hr_localoff); 1022204076Spjd if (ret == ggio->gctl_length) 1023204076Spjd hio->hio_errors[ncomp] = 0; 1024204076Spjd else { 1025204076Spjd /* 1026204076Spjd * If READ failed, try to read from remote node. 1027204076Spjd */ 1028204076Spjd QUEUE_INSERT1(hio, send, rncomp); 1029204076Spjd continue; 1030204076Spjd } 1031204076Spjd break; 1032204076Spjd case BIO_WRITE: 1033204076Spjd ret = pwrite(res->hr_localfd, ggio->gctl_data, 1034204076Spjd ggio->gctl_length, 1035204076Spjd ggio->gctl_offset + res->hr_localoff); 1036204076Spjd if (ret < 0) 1037204076Spjd hio->hio_errors[ncomp] = errno; 1038204076Spjd else if (ret != ggio->gctl_length) 1039204076Spjd hio->hio_errors[ncomp] = EIO; 1040204076Spjd else 1041204076Spjd hio->hio_errors[ncomp] = 0; 1042204076Spjd break; 1043204076Spjd case BIO_DELETE: 1044204076Spjd ret = g_delete(res->hr_localfd, 1045204076Spjd ggio->gctl_offset + res->hr_localoff, 1046204076Spjd ggio->gctl_length); 1047204076Spjd if (ret < 0) 1048204076Spjd hio->hio_errors[ncomp] = errno; 1049204076Spjd else 1050204076Spjd hio->hio_errors[ncomp] = 0; 1051204076Spjd break; 1052204076Spjd case BIO_FLUSH: 1053204076Spjd ret = g_flush(res->hr_localfd); 1054204076Spjd if (ret < 0) 1055204076Spjd hio->hio_errors[ncomp] = errno; 1056204076Spjd else 1057204076Spjd hio->hio_errors[ncomp] = 0; 1058204076Spjd break; 1059204076Spjd } 1060204076Spjd if (refcount_release(&hio->hio_countdown)) { 1061204076Spjd if (ISSYNCREQ(hio)) { 1062204076Spjd mtx_lock(&sync_lock); 1063204076Spjd SYNCREQDONE(hio); 1064204076Spjd mtx_unlock(&sync_lock); 1065204076Spjd cv_signal(&sync_cond); 1066204076Spjd } else { 1067204076Spjd pjdlog_debug(2, 1068204076Spjd "local_send: (%p) Moving request to the done queue.", 1069204076Spjd hio); 1070204076Spjd QUEUE_INSERT2(hio, done); 1071204076Spjd } 1072204076Spjd } 1073204076Spjd } 1074204076Spjd /* NOTREACHED */ 1075204076Spjd return (NULL); 1076204076Spjd} 1077204076Spjd 1078204076Spjd/* 1079204076Spjd * Thread sends request to secondary node. 1080204076Spjd */ 1081204076Spjdstatic void * 1082204076Spjdremote_send_thread(void *arg) 1083204076Spjd{ 1084204076Spjd struct hast_resource *res = arg; 1085204076Spjd struct g_gate_ctl_io *ggio; 1086204076Spjd struct hio *hio; 1087204076Spjd struct nv *nv; 1088204076Spjd unsigned int ncomp; 1089204076Spjd bool wakeup; 1090204076Spjd uint64_t offset, length; 1091204076Spjd uint8_t cmd; 1092204076Spjd void *data; 1093204076Spjd 1094204076Spjd /* Remote component is 1 for now. */ 1095204076Spjd ncomp = 1; 1096204076Spjd 1097204076Spjd for (;;) { 1098204076Spjd pjdlog_debug(2, "remote_send: Taking request."); 1099204076Spjd QUEUE_TAKE1(hio, send, ncomp); 1100204076Spjd pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1101204076Spjd ggio = &hio->hio_ggio; 1102204076Spjd switch (ggio->gctl_cmd) { 1103204076Spjd case BIO_READ: 1104204076Spjd cmd = HIO_READ; 1105204076Spjd data = NULL; 1106204076Spjd offset = ggio->gctl_offset; 1107204076Spjd length = ggio->gctl_length; 1108204076Spjd break; 1109204076Spjd case BIO_WRITE: 1110204076Spjd cmd = HIO_WRITE; 1111204076Spjd data = ggio->gctl_data; 1112204076Spjd offset = ggio->gctl_offset; 1113204076Spjd length = ggio->gctl_length; 1114204076Spjd break; 1115204076Spjd case BIO_DELETE: 1116204076Spjd cmd = HIO_DELETE; 1117204076Spjd data = NULL; 1118204076Spjd offset = ggio->gctl_offset; 1119204076Spjd length = ggio->gctl_length; 1120204076Spjd break; 1121204076Spjd case BIO_FLUSH: 1122204076Spjd cmd = HIO_FLUSH; 1123204076Spjd data = NULL; 1124204076Spjd offset = 0; 1125204076Spjd length = 0; 1126204076Spjd break; 1127204076Spjd default: 1128204076Spjd assert(!"invalid condition"); 1129204076Spjd abort(); 1130204076Spjd } 1131204076Spjd nv = nv_alloc(); 1132204076Spjd nv_add_uint8(nv, cmd, "cmd"); 1133204076Spjd nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1134204076Spjd nv_add_uint64(nv, offset, "offset"); 1135204076Spjd nv_add_uint64(nv, length, "length"); 1136204076Spjd if (nv_error(nv) != 0) { 1137204076Spjd hio->hio_errors[ncomp] = nv_error(nv); 1138204076Spjd pjdlog_debug(2, 1139204076Spjd "remote_send: (%p) Unable to prepare header to send.", 1140204076Spjd hio); 1141204076Spjd reqlog(LOG_ERR, 0, ggio, 1142204076Spjd "Unable to prepare header to send (%s): ", 1143204076Spjd strerror(nv_error(nv))); 1144204076Spjd /* Move failed request immediately to the done queue. */ 1145204076Spjd goto done_queue; 1146204076Spjd } 1147204076Spjd pjdlog_debug(2, 1148204076Spjd "remote_send: (%p) Moving request to the recv queue.", 1149204076Spjd hio); 1150204076Spjd /* 1151204076Spjd * Protect connection from disappearing. 1152204076Spjd */ 1153204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1154204076Spjd if (!ISCONNECTED(res, ncomp)) { 1155204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1156204076Spjd hio->hio_errors[ncomp] = ENOTCONN; 1157204076Spjd goto done_queue; 1158204076Spjd } 1159204076Spjd /* 1160204076Spjd * Move the request to recv queue before sending it, because 1161204076Spjd * in different order we can get reply before we move request 1162204076Spjd * to recv queue. 1163204076Spjd */ 1164204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1165204076Spjd wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1166204076Spjd TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1167204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1168204076Spjd if (hast_proto_send(res, res->hr_remoteout, nv, data, 1169204076Spjd data != NULL ? length : 0) < 0) { 1170204076Spjd hio->hio_errors[ncomp] = errno; 1171204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1172204076Spjd remote_close(res, ncomp); 1173204076Spjd pjdlog_debug(2, 1174204076Spjd "remote_send: (%p) Unable to send request.", hio); 1175204076Spjd reqlog(LOG_ERR, 0, ggio, 1176204076Spjd "Unable to send request (%s): ", 1177204076Spjd strerror(hio->hio_errors[ncomp])); 1178204076Spjd /* 1179204076Spjd * Take request back from the receive queue and move 1180204076Spjd * it immediately to the done queue. 1181204076Spjd */ 1182204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1183204076Spjd TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1184204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1185204076Spjd goto done_queue; 1186204076Spjd } 1187204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1188204076Spjd nv_free(nv); 1189204076Spjd if (wakeup) 1190204076Spjd cv_signal(&hio_recv_list_cond[ncomp]); 1191204076Spjd continue; 1192204076Spjddone_queue: 1193204076Spjd nv_free(nv); 1194204076Spjd if (ISSYNCREQ(hio)) { 1195204076Spjd if (!refcount_release(&hio->hio_countdown)) 1196204076Spjd continue; 1197204076Spjd mtx_lock(&sync_lock); 1198204076Spjd SYNCREQDONE(hio); 1199204076Spjd mtx_unlock(&sync_lock); 1200204076Spjd cv_signal(&sync_cond); 1201204076Spjd continue; 1202204076Spjd } 1203204076Spjd if (ggio->gctl_cmd == BIO_WRITE) { 1204204076Spjd mtx_lock(&res->hr_amp_lock); 1205204076Spjd if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1206204076Spjd ggio->gctl_length)) { 1207204076Spjd (void)hast_activemap_flush(res); 1208204076Spjd } 1209204076Spjd mtx_unlock(&res->hr_amp_lock); 1210204076Spjd } 1211204076Spjd if (!refcount_release(&hio->hio_countdown)) 1212204076Spjd continue; 1213204076Spjd pjdlog_debug(2, 1214204076Spjd "remote_send: (%p) Moving request to the done queue.", 1215204076Spjd hio); 1216204076Spjd QUEUE_INSERT2(hio, done); 1217204076Spjd } 1218204076Spjd /* NOTREACHED */ 1219204076Spjd return (NULL); 1220204076Spjd} 1221204076Spjd 1222204076Spjd/* 1223204076Spjd * Thread receives answer from secondary node and passes it to ggate_send 1224204076Spjd * thread. 1225204076Spjd */ 1226204076Spjdstatic void * 1227204076Spjdremote_recv_thread(void *arg) 1228204076Spjd{ 1229204076Spjd struct hast_resource *res = arg; 1230204076Spjd struct g_gate_ctl_io *ggio; 1231204076Spjd struct hio *hio; 1232204076Spjd struct nv *nv; 1233204076Spjd unsigned int ncomp; 1234204076Spjd uint64_t seq; 1235204076Spjd int error; 1236204076Spjd 1237204076Spjd /* Remote component is 1 for now. */ 1238204076Spjd ncomp = 1; 1239204076Spjd 1240204076Spjd for (;;) { 1241204076Spjd /* Wait until there is anything to receive. */ 1242204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1243204076Spjd while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1244204076Spjd pjdlog_debug(2, "remote_recv: No requests, waiting."); 1245204076Spjd cv_wait(&hio_recv_list_cond[ncomp], 1246204076Spjd &hio_recv_list_lock[ncomp]); 1247204076Spjd } 1248204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1249204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1250204076Spjd if (!ISCONNECTED(res, ncomp)) { 1251204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1252204076Spjd /* 1253204076Spjd * Connection is dead, so move all pending requests to 1254204076Spjd * the done queue (one-by-one). 1255204076Spjd */ 1256204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1257204076Spjd hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1258204076Spjd assert(hio != NULL); 1259204076Spjd TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1260204076Spjd hio_next[ncomp]); 1261204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1262204076Spjd goto done_queue; 1263204076Spjd } 1264204076Spjd if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1265204076Spjd pjdlog_errno(LOG_ERR, 1266204076Spjd "Unable to receive reply header"); 1267204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1268204076Spjd remote_close(res, ncomp); 1269204076Spjd continue; 1270204076Spjd } 1271204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1272204076Spjd seq = nv_get_uint64(nv, "seq"); 1273204076Spjd if (seq == 0) { 1274204076Spjd pjdlog_error("Header contains no 'seq' field."); 1275204076Spjd nv_free(nv); 1276204076Spjd continue; 1277204076Spjd } 1278204076Spjd mtx_lock(&hio_recv_list_lock[ncomp]); 1279204076Spjd TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1280204076Spjd if (hio->hio_ggio.gctl_seq == seq) { 1281204076Spjd TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1282204076Spjd hio_next[ncomp]); 1283204076Spjd break; 1284204076Spjd } 1285204076Spjd } 1286204076Spjd mtx_unlock(&hio_recv_list_lock[ncomp]); 1287204076Spjd if (hio == NULL) { 1288204076Spjd pjdlog_error("Found no request matching received 'seq' field (%ju).", 1289204076Spjd (uintmax_t)seq); 1290204076Spjd nv_free(nv); 1291204076Spjd continue; 1292204076Spjd } 1293204076Spjd error = nv_get_int16(nv, "error"); 1294204076Spjd if (error != 0) { 1295204076Spjd /* Request failed on remote side. */ 1296204076Spjd hio->hio_errors[ncomp] = 0; 1297204076Spjd nv_free(nv); 1298204076Spjd goto done_queue; 1299204076Spjd } 1300204076Spjd ggio = &hio->hio_ggio; 1301204076Spjd switch (ggio->gctl_cmd) { 1302204076Spjd case BIO_READ: 1303204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1304204076Spjd if (!ISCONNECTED(res, ncomp)) { 1305204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1306204076Spjd nv_free(nv); 1307204076Spjd goto done_queue; 1308204076Spjd } 1309204076Spjd if (hast_proto_recv_data(res, res->hr_remotein, nv, 1310204076Spjd ggio->gctl_data, ggio->gctl_length) < 0) { 1311204076Spjd hio->hio_errors[ncomp] = errno; 1312204076Spjd pjdlog_errno(LOG_ERR, 1313204076Spjd "Unable to receive reply data"); 1314204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1315204076Spjd nv_free(nv); 1316204076Spjd remote_close(res, ncomp); 1317204076Spjd goto done_queue; 1318204076Spjd } 1319204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1320204076Spjd break; 1321204076Spjd case BIO_WRITE: 1322204076Spjd case BIO_DELETE: 1323204076Spjd case BIO_FLUSH: 1324204076Spjd break; 1325204076Spjd default: 1326204076Spjd assert(!"invalid condition"); 1327204076Spjd abort(); 1328204076Spjd } 1329204076Spjd hio->hio_errors[ncomp] = 0; 1330204076Spjd nv_free(nv); 1331204076Spjddone_queue: 1332204076Spjd if (refcount_release(&hio->hio_countdown)) { 1333204076Spjd if (ISSYNCREQ(hio)) { 1334204076Spjd mtx_lock(&sync_lock); 1335204076Spjd SYNCREQDONE(hio); 1336204076Spjd mtx_unlock(&sync_lock); 1337204076Spjd cv_signal(&sync_cond); 1338204076Spjd } else { 1339204076Spjd pjdlog_debug(2, 1340204076Spjd "remote_recv: (%p) Moving request to the done queue.", 1341204076Spjd hio); 1342204076Spjd QUEUE_INSERT2(hio, done); 1343204076Spjd } 1344204076Spjd } 1345204076Spjd } 1346204076Spjd /* NOTREACHED */ 1347204076Spjd return (NULL); 1348204076Spjd} 1349204076Spjd 1350204076Spjd/* 1351204076Spjd * Thread sends answer to the kernel. 1352204076Spjd */ 1353204076Spjdstatic void * 1354204076Spjdggate_send_thread(void *arg) 1355204076Spjd{ 1356204076Spjd struct hast_resource *res = arg; 1357204076Spjd struct g_gate_ctl_io *ggio; 1358204076Spjd struct hio *hio; 1359204076Spjd unsigned int ii, ncomp, ncomps; 1360204076Spjd 1361204076Spjd ncomps = HAST_NCOMPONENTS; 1362204076Spjd 1363204076Spjd for (;;) { 1364204076Spjd pjdlog_debug(2, "ggate_send: Taking request."); 1365204076Spjd QUEUE_TAKE2(hio, done); 1366204076Spjd pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1367204076Spjd ggio = &hio->hio_ggio; 1368204076Spjd for (ii = 0; ii < ncomps; ii++) { 1369204076Spjd if (hio->hio_errors[ii] == 0) { 1370204076Spjd /* 1371204076Spjd * One successful request is enough to declare 1372204076Spjd * success. 1373204076Spjd */ 1374204076Spjd ggio->gctl_error = 0; 1375204076Spjd break; 1376204076Spjd } 1377204076Spjd } 1378204076Spjd if (ii == ncomps) { 1379204076Spjd /* 1380204076Spjd * None of the requests were successful. 1381204076Spjd * Use first error. 1382204076Spjd */ 1383204076Spjd ggio->gctl_error = hio->hio_errors[0]; 1384204076Spjd } 1385204076Spjd if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1386204076Spjd mtx_lock(&res->hr_amp_lock); 1387204076Spjd activemap_write_complete(res->hr_amp, 1388204076Spjd ggio->gctl_offset, ggio->gctl_length); 1389204076Spjd mtx_unlock(&res->hr_amp_lock); 1390204076Spjd } 1391204076Spjd if (ggio->gctl_cmd == BIO_WRITE) { 1392204076Spjd /* 1393204076Spjd * Unlock range we locked. 1394204076Spjd */ 1395204076Spjd mtx_lock(&range_lock); 1396204076Spjd rangelock_del(range_regular, ggio->gctl_offset, 1397204076Spjd ggio->gctl_length); 1398204076Spjd if (range_sync_wait) 1399204076Spjd cv_signal(&range_sync_cond); 1400204076Spjd mtx_unlock(&range_lock); 1401204076Spjd /* 1402204076Spjd * Bump local count if this is first write after 1403204076Spjd * connection failure with remote node. 1404204076Spjd */ 1405204076Spjd ncomp = 1; 1406204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1407204076Spjd if (!ISCONNECTED(res, ncomp)) { 1408204076Spjd mtx_lock(&metadata_lock); 1409204076Spjd if (res->hr_primary_localcnt == 1410204076Spjd res->hr_secondary_remotecnt) { 1411204076Spjd res->hr_primary_localcnt++; 1412204076Spjd pjdlog_debug(1, 1413204076Spjd "Increasing localcnt to %ju.", 1414204076Spjd (uintmax_t)res->hr_primary_localcnt); 1415204076Spjd (void)metadata_write(res); 1416204076Spjd } 1417204076Spjd mtx_unlock(&metadata_lock); 1418204076Spjd } 1419204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1420204076Spjd } 1421204076Spjd if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1422204076Spjd primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1423204076Spjd pjdlog_debug(2, 1424204076Spjd "ggate_send: (%p) Moving request to the free queue.", hio); 1425204076Spjd QUEUE_INSERT2(hio, free); 1426204076Spjd } 1427204076Spjd /* NOTREACHED */ 1428204076Spjd return (NULL); 1429204076Spjd} 1430204076Spjd 1431204076Spjd/* 1432204076Spjd * Thread synchronize local and remote components. 1433204076Spjd */ 1434204076Spjdstatic void * 1435204076Spjdsync_thread(void *arg __unused) 1436204076Spjd{ 1437204076Spjd struct hast_resource *res = arg; 1438204076Spjd struct hio *hio; 1439204076Spjd struct g_gate_ctl_io *ggio; 1440204076Spjd unsigned int ii, ncomp, ncomps; 1441204076Spjd off_t offset, length, synced; 1442204076Spjd bool dorewind; 1443204076Spjd int syncext; 1444204076Spjd 1445204076Spjd ncomps = HAST_NCOMPONENTS; 1446204076Spjd dorewind = true; 1447204076Spjd synced = 0; 1448204076Spjd 1449204076Spjd for (;;) { 1450204076Spjd mtx_lock(&sync_lock); 1451204076Spjd while (!sync_inprogress) { 1452204076Spjd dorewind = true; 1453204076Spjd synced = 0; 1454204076Spjd cv_wait(&sync_cond, &sync_lock); 1455204076Spjd } 1456204076Spjd mtx_unlock(&sync_lock); 1457204076Spjd /* 1458204076Spjd * Obtain offset at which we should synchronize. 1459204076Spjd * Rewind synchronization if needed. 1460204076Spjd */ 1461204076Spjd mtx_lock(&res->hr_amp_lock); 1462204076Spjd if (dorewind) 1463204076Spjd activemap_sync_rewind(res->hr_amp); 1464204076Spjd offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1465204076Spjd if (syncext != -1) { 1466204076Spjd /* 1467204076Spjd * We synchronized entire syncext extent, we can mark 1468204076Spjd * it as clean now. 1469204076Spjd */ 1470204076Spjd if (activemap_extent_complete(res->hr_amp, syncext)) 1471204076Spjd (void)hast_activemap_flush(res); 1472204076Spjd } 1473204076Spjd mtx_unlock(&res->hr_amp_lock); 1474204076Spjd if (dorewind) { 1475204076Spjd dorewind = false; 1476204076Spjd if (offset < 0) 1477204076Spjd pjdlog_info("Nodes are in sync."); 1478204076Spjd else { 1479204076Spjd pjdlog_info("Synchronization started. %ju bytes to go.", 1480204076Spjd (uintmax_t)(res->hr_extentsize * 1481204076Spjd activemap_ndirty(res->hr_amp))); 1482204076Spjd } 1483204076Spjd } 1484204076Spjd if (offset < 0) { 1485204076Spjd mtx_lock(&sync_lock); 1486204076Spjd sync_inprogress = false; 1487204076Spjd mtx_unlock(&sync_lock); 1488204076Spjd pjdlog_debug(1, "Nothing to synchronize."); 1489204076Spjd /* 1490204076Spjd * Synchronization complete, make both localcnt and 1491204076Spjd * remotecnt equal. 1492204076Spjd */ 1493204076Spjd ncomp = 1; 1494204076Spjd rw_rlock(&hio_remote_lock[ncomp]); 1495204076Spjd if (ISCONNECTED(res, ncomp)) { 1496204076Spjd if (synced > 0) { 1497204076Spjd pjdlog_info("Synchronization complete. " 1498204076Spjd "%jd bytes synchronized.", 1499204076Spjd (intmax_t)synced); 1500204076Spjd } 1501204076Spjd mtx_lock(&metadata_lock); 1502204076Spjd res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1503204076Spjd res->hr_primary_localcnt = 1504204076Spjd res->hr_secondary_localcnt; 1505204076Spjd res->hr_primary_remotecnt = 1506204076Spjd res->hr_secondary_remotecnt; 1507204076Spjd pjdlog_debug(1, 1508204076Spjd "Setting localcnt to %ju and remotecnt to %ju.", 1509204076Spjd (uintmax_t)res->hr_primary_localcnt, 1510204076Spjd (uintmax_t)res->hr_secondary_localcnt); 1511204076Spjd (void)metadata_write(res); 1512204076Spjd mtx_unlock(&metadata_lock); 1513204076Spjd } else if (synced > 0) { 1514204076Spjd pjdlog_info("Synchronization interrupted. " 1515204076Spjd "%jd bytes synchronized so far.", 1516204076Spjd (intmax_t)synced); 1517204076Spjd } 1518204076Spjd rw_unlock(&hio_remote_lock[ncomp]); 1519204076Spjd continue; 1520204076Spjd } 1521204076Spjd pjdlog_debug(2, "sync: Taking free request."); 1522204076Spjd QUEUE_TAKE2(hio, free); 1523204076Spjd pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1524204076Spjd /* 1525204076Spjd * Lock the range we are going to synchronize. We don't want 1526204076Spjd * race where someone writes between our read and write. 1527204076Spjd */ 1528204076Spjd for (;;) { 1529204076Spjd mtx_lock(&range_lock); 1530204076Spjd if (rangelock_islocked(range_regular, offset, length)) { 1531204076Spjd pjdlog_debug(2, 1532204076Spjd "sync: Range offset=%jd length=%jd locked.", 1533204076Spjd (intmax_t)offset, (intmax_t)length); 1534204076Spjd range_sync_wait = true; 1535204076Spjd cv_wait(&range_sync_cond, &range_lock); 1536204076Spjd range_sync_wait = false; 1537204076Spjd mtx_unlock(&range_lock); 1538204076Spjd continue; 1539204076Spjd } 1540204076Spjd if (rangelock_add(range_sync, offset, length) < 0) { 1541204076Spjd mtx_unlock(&range_lock); 1542204076Spjd pjdlog_debug(2, 1543204076Spjd "sync: Range offset=%jd length=%jd is already locked, waiting.", 1544204076Spjd (intmax_t)offset, (intmax_t)length); 1545204076Spjd sleep(1); 1546204076Spjd continue; 1547204076Spjd } 1548204076Spjd mtx_unlock(&range_lock); 1549204076Spjd break; 1550204076Spjd } 1551204076Spjd /* 1552204076Spjd * First read the data from synchronization source. 1553204076Spjd */ 1554204076Spjd SYNCREQ(hio); 1555204076Spjd ggio = &hio->hio_ggio; 1556204076Spjd ggio->gctl_cmd = BIO_READ; 1557204076Spjd ggio->gctl_offset = offset; 1558204076Spjd ggio->gctl_length = length; 1559204076Spjd ggio->gctl_error = 0; 1560204076Spjd for (ii = 0; ii < ncomps; ii++) 1561204076Spjd hio->hio_errors[ii] = EINVAL; 1562204076Spjd reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1563204076Spjd hio); 1564204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1565204076Spjd hio); 1566204076Spjd mtx_lock(&metadata_lock); 1567204076Spjd if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1568204076Spjd /* 1569204076Spjd * This range is up-to-date on local component, 1570204076Spjd * so handle request locally. 1571204076Spjd */ 1572204076Spjd /* Local component is 0 for now. */ 1573204076Spjd ncomp = 0; 1574204076Spjd } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1575204076Spjd assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1576204076Spjd /* 1577204076Spjd * This range is out-of-date on local component, 1578204076Spjd * so send request to the remote node. 1579204076Spjd */ 1580204076Spjd /* Remote component is 1 for now. */ 1581204076Spjd ncomp = 1; 1582204076Spjd } 1583204076Spjd mtx_unlock(&metadata_lock); 1584204076Spjd refcount_init(&hio->hio_countdown, 1); 1585204076Spjd QUEUE_INSERT1(hio, send, ncomp); 1586204076Spjd 1587204076Spjd /* 1588204076Spjd * Let's wait for READ to finish. 1589204076Spjd */ 1590204076Spjd mtx_lock(&sync_lock); 1591204076Spjd while (!ISSYNCREQDONE(hio)) 1592204076Spjd cv_wait(&sync_cond, &sync_lock); 1593204076Spjd mtx_unlock(&sync_lock); 1594204076Spjd 1595204076Spjd if (hio->hio_errors[ncomp] != 0) { 1596204076Spjd pjdlog_error("Unable to read synchronization data: %s.", 1597204076Spjd strerror(hio->hio_errors[ncomp])); 1598204076Spjd goto free_queue; 1599204076Spjd } 1600204076Spjd 1601204076Spjd /* 1602204076Spjd * We read the data from synchronization source, now write it 1603204076Spjd * to synchronization target. 1604204076Spjd */ 1605204076Spjd SYNCREQ(hio); 1606204076Spjd ggio->gctl_cmd = BIO_WRITE; 1607204076Spjd for (ii = 0; ii < ncomps; ii++) 1608204076Spjd hio->hio_errors[ii] = EINVAL; 1609204076Spjd reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1610204076Spjd hio); 1611204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1612204076Spjd hio); 1613204076Spjd mtx_lock(&metadata_lock); 1614204076Spjd if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1615204076Spjd /* 1616204076Spjd * This range is up-to-date on local component, 1617204076Spjd * so we update remote component. 1618204076Spjd */ 1619204076Spjd /* Remote component is 1 for now. */ 1620204076Spjd ncomp = 1; 1621204076Spjd } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1622204076Spjd assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1623204076Spjd /* 1624204076Spjd * This range is out-of-date on local component, 1625204076Spjd * so we update it. 1626204076Spjd */ 1627204076Spjd /* Local component is 0 for now. */ 1628204076Spjd ncomp = 0; 1629204076Spjd } 1630204076Spjd mtx_unlock(&metadata_lock); 1631204076Spjd 1632204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1633204076Spjd hio); 1634204076Spjd refcount_init(&hio->hio_countdown, 1); 1635204076Spjd QUEUE_INSERT1(hio, send, ncomp); 1636204076Spjd 1637204076Spjd /* 1638204076Spjd * Let's wait for WRITE to finish. 1639204076Spjd */ 1640204076Spjd mtx_lock(&sync_lock); 1641204076Spjd while (!ISSYNCREQDONE(hio)) 1642204076Spjd cv_wait(&sync_cond, &sync_lock); 1643204076Spjd mtx_unlock(&sync_lock); 1644204076Spjd 1645204076Spjd if (hio->hio_errors[ncomp] != 0) { 1646204076Spjd pjdlog_error("Unable to write synchronization data: %s.", 1647204076Spjd strerror(hio->hio_errors[ncomp])); 1648204076Spjd goto free_queue; 1649204076Spjd } 1650204076Spjdfree_queue: 1651204076Spjd mtx_lock(&range_lock); 1652204076Spjd rangelock_del(range_sync, offset, length); 1653204076Spjd if (range_regular_wait) 1654204076Spjd cv_signal(&range_regular_cond); 1655204076Spjd mtx_unlock(&range_lock); 1656204076Spjd 1657204076Spjd synced += length; 1658204076Spjd 1659204076Spjd pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1660204076Spjd hio); 1661204076Spjd QUEUE_INSERT2(hio, free); 1662204076Spjd } 1663204076Spjd /* NOTREACHED */ 1664204076Spjd return (NULL); 1665204076Spjd} 1666204076Spjd 1667204076Spjdstatic void 1668204076Spjdsighandler(int sig) 1669204076Spjd{ 1670204076Spjd bool unlock; 1671204076Spjd 1672204076Spjd switch (sig) { 1673204076Spjd case SIGINT: 1674204076Spjd case SIGTERM: 1675204076Spjd sigexit_received = true; 1676204076Spjd break; 1677204076Spjd default: 1678204076Spjd assert(!"invalid condition"); 1679204076Spjd } 1680204076Spjd /* 1681204076Spjd * XXX: Racy, but if we cannot obtain hio_guard_lock here, we don't 1682204076Spjd * want to risk deadlock. 1683204076Spjd */ 1684204076Spjd unlock = mtx_trylock(&hio_guard_lock); 1685204076Spjd cv_signal(&hio_guard_cond); 1686204076Spjd if (unlock) 1687204076Spjd mtx_unlock(&hio_guard_lock); 1688204076Spjd} 1689204076Spjd 1690204076Spjd/* 1691204076Spjd * Thread guards remote connections and reconnects when needed, handles 1692204076Spjd * signals, etc. 1693204076Spjd */ 1694204076Spjdstatic void * 1695204076Spjdguard_thread(void *arg) 1696204076Spjd{ 1697204076Spjd struct hast_resource *res = arg; 1698204076Spjd unsigned int ii, ncomps; 1699204076Spjd int timeout; 1700204076Spjd 1701204076Spjd ncomps = HAST_NCOMPONENTS; 1702204076Spjd /* The is only one remote component for now. */ 1703204076Spjd#define ISREMOTE(no) ((no) == 1) 1704204076Spjd 1705204076Spjd for (;;) { 1706204076Spjd if (sigexit_received) { 1707204076Spjd primary_exitx(EX_OK, 1708204076Spjd "Termination signal received, exiting."); 1709204076Spjd } 1710204076Spjd /* 1711204076Spjd * If all the connection will be fine, we will sleep until 1712204076Spjd * someone wakes us up. 1713204076Spjd * If any of the connections will be broken and we won't be 1714204076Spjd * able to connect, we will sleep only for RECONNECT_SLEEP 1715204076Spjd * seconds so we can retry soon. 1716204076Spjd */ 1717204076Spjd timeout = 0; 1718204076Spjd pjdlog_debug(2, "remote_guard: Checking connections."); 1719204076Spjd mtx_lock(&hio_guard_lock); 1720204076Spjd for (ii = 0; ii < ncomps; ii++) { 1721204076Spjd if (!ISREMOTE(ii)) 1722204076Spjd continue; 1723204076Spjd rw_rlock(&hio_remote_lock[ii]); 1724204076Spjd if (ISCONNECTED(res, ii)) { 1725204076Spjd assert(res->hr_remotein != NULL); 1726204076Spjd assert(res->hr_remoteout != NULL); 1727204076Spjd rw_unlock(&hio_remote_lock[ii]); 1728204076Spjd pjdlog_debug(2, 1729204076Spjd "remote_guard: Connection to %s is ok.", 1730204076Spjd res->hr_remoteaddr); 1731204076Spjd } else { 1732204076Spjd assert(res->hr_remotein == NULL); 1733204076Spjd assert(res->hr_remoteout == NULL); 1734204076Spjd /* 1735204076Spjd * Upgrade the lock. It doesn't have to be 1736204076Spjd * atomic as no other thread can change 1737204076Spjd * connection status from disconnected to 1738204076Spjd * connected. 1739204076Spjd */ 1740204076Spjd rw_unlock(&hio_remote_lock[ii]); 1741204076Spjd rw_wlock(&hio_remote_lock[ii]); 1742204076Spjd assert(res->hr_remotein == NULL); 1743204076Spjd assert(res->hr_remoteout == NULL); 1744204076Spjd pjdlog_debug(2, 1745204076Spjd "remote_guard: Reconnecting to %s.", 1746204076Spjd res->hr_remoteaddr); 1747204076Spjd init_remote(res); 1748204076Spjd if (ISCONNECTED(res, ii)) { 1749204076Spjd pjdlog_info("Successfully reconnected to %s.", 1750204076Spjd res->hr_remoteaddr); 1751204076Spjd } else { 1752204076Spjd /* Both connections should be NULL. */ 1753204076Spjd assert(res->hr_remotein == NULL); 1754204076Spjd assert(res->hr_remoteout == NULL); 1755204076Spjd pjdlog_debug(2, 1756204076Spjd "remote_guard: Reconnect to %s failed.", 1757204076Spjd res->hr_remoteaddr); 1758204076Spjd timeout = RECONNECT_SLEEP; 1759204076Spjd } 1760204076Spjd rw_unlock(&hio_remote_lock[ii]); 1761204076Spjd } 1762204076Spjd } 1763204076Spjd (void)cv_timedwait(&hio_guard_cond, &hio_guard_lock, timeout); 1764204076Spjd mtx_unlock(&hio_guard_lock); 1765204076Spjd } 1766204076Spjd#undef ISREMOTE 1767204076Spjd /* NOTREACHED */ 1768204076Spjd return (NULL); 1769204076Spjd} 1770