1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2009 The FreeBSD Foundation 5 * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net> 6 * All rights reserved. 7 * 8 * This software was developed by Pawel Jakub Dawidek under sponsorship from 9 * the FreeBSD Foundation. 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in the 18 * documentation and/or other materials provided with the distribution. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 */ 32 33 #include <sys/cdefs.h> 34 __FBSDID("$FreeBSD$"); 35 36 #include <sys/types.h> 37 #include <sys/time.h> 38 #include <sys/bio.h> 39 #include <sys/disk.h> 40 #include <sys/stat.h> 41 42 #include <geom/gate/g_gate.h> 43 44 #include <err.h> 45 #include <errno.h> 46 #include <fcntl.h> 47 #include <libgeom.h> 48 #include <pthread.h> 49 #include <signal.h> 50 #include <stdint.h> 51 #include <stdio.h> 52 #include <string.h> 53 #include <sysexits.h> 54 #include <unistd.h> 55 56 #include <activemap.h> 57 #include <nv.h> 58 #include <rangelock.h> 59 60 #include "control.h" 61 #include "event.h" 62 #include "hast.h" 63 #include "hast_proto.h" 64 #include "hastd.h" 65 #include "hooks.h" 66 #include "metadata.h" 67 #include "proto.h" 68 #include "pjdlog.h" 69 #include "refcnt.h" 70 #include "subr.h" 71 #include "synch.h" 72 73 /* The is only one remote component for now. */ 74 #define ISREMOTE(no) ((no) == 1) 75 76 struct hio { 77 /* 78 * Number of components we are still waiting for. 79 * When this field goes to 0, we can send the request back to the 80 * kernel. Each component has to decrease this counter by one 81 * even on failure. 82 */ 83 refcnt_t hio_countdown; 84 /* 85 * Each component has a place to store its own error. 86 * Once the request is handled by all components we can decide if the 87 * request overall is successful or not. 88 */ 89 int *hio_errors; 90 /* 91 * Structure used to communicate with GEOM Gate class. 92 */ 93 struct g_gate_ctl_io hio_ggio; 94 /* 95 * Request was already confirmed to GEOM Gate. 96 */ 97 bool hio_done; 98 /* 99 * Number of components we are still waiting before sending write 100 * completion ack to GEOM Gate. Used for memsync. 101 */ 102 refcnt_t hio_writecount; 103 /* 104 * Memsync request was acknowleged by remote. 105 */ 106 bool hio_memsyncacked; 107 /* 108 * Remember replication from the time the request was initiated, 109 * so we won't get confused when replication changes on reload. 110 */ 111 int hio_replication; 112 TAILQ_ENTRY(hio) *hio_next; 113 }; 114 #define hio_free_next hio_next[0] 115 #define hio_done_next hio_next[0] 116 117 /* 118 * Free list holds unused structures. When free list is empty, we have to wait 119 * until some in-progress requests are freed. 120 */ 121 static TAILQ_HEAD(, hio) hio_free_list; 122 static size_t hio_free_list_size; 123 static pthread_mutex_t hio_free_list_lock; 124 static pthread_cond_t hio_free_list_cond; 125 /* 126 * There is one send list for every component. One requests is placed on all 127 * send lists - each component gets the same request, but each component is 128 * responsible for managing his own send list. 129 */ 130 static TAILQ_HEAD(, hio) *hio_send_list; 131 static size_t *hio_send_list_size; 132 static pthread_mutex_t *hio_send_list_lock; 133 static pthread_cond_t *hio_send_list_cond; 134 #define hio_send_local_list_size hio_send_list_size[0] 135 #define hio_send_remote_list_size hio_send_list_size[1] 136 /* 137 * There is one recv list for every component, although local components don't 138 * use recv lists as local requests are done synchronously. 139 */ 140 static TAILQ_HEAD(, hio) *hio_recv_list; 141 static size_t *hio_recv_list_size; 142 static pthread_mutex_t *hio_recv_list_lock; 143 static pthread_cond_t *hio_recv_list_cond; 144 #define hio_recv_remote_list_size hio_recv_list_size[1] 145 /* 146 * Request is placed on done list by the slowest component (the one that 147 * decreased hio_countdown from 1 to 0). 148 */ 149 static TAILQ_HEAD(, hio) hio_done_list; 150 static size_t hio_done_list_size; 151 static pthread_mutex_t hio_done_list_lock; 152 static pthread_cond_t hio_done_list_cond; 153 /* 154 * Structure below are for interaction with sync thread. 155 */ 156 static bool sync_inprogress; 157 static pthread_mutex_t sync_lock; 158 static pthread_cond_t sync_cond; 159 /* 160 * The lock below allows to synchornize access to remote connections. 161 */ 162 static pthread_rwlock_t *hio_remote_lock; 163 164 /* 165 * Lock to synchronize metadata updates. Also synchronize access to 166 * hr_primary_localcnt and hr_primary_remotecnt fields. 167 */ 168 static pthread_mutex_t metadata_lock; 169 170 /* 171 * Maximum number of outstanding I/O requests. 172 */ 173 #define HAST_HIO_MAX 256 174 /* 175 * Number of components. At this point there are only two components: local 176 * and remote, but in the future it might be possible to use multiple local 177 * and remote components. 178 */ 179 #define HAST_NCOMPONENTS 2 180 181 #define ISCONNECTED(res, no) \ 182 ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL) 183 184 #define QUEUE_INSERT1(hio, name, ncomp) do { \ 185 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 186 if (TAILQ_EMPTY(&hio_##name##_list[(ncomp)])) \ 187 cv_broadcast(&hio_##name##_list_cond[(ncomp)]); \ 188 TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio), \ 189 hio_next[(ncomp)]); \ 190 hio_##name##_list_size[(ncomp)]++; \ 191 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 192 } while (0) 193 #define QUEUE_INSERT2(hio, name) do { \ 194 mtx_lock(&hio_##name##_list_lock); \ 195 if (TAILQ_EMPTY(&hio_##name##_list)) \ 196 cv_broadcast(&hio_##name##_list_cond); \ 197 TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\ 198 hio_##name##_list_size++; \ 199 mtx_unlock(&hio_##name##_list_lock); \ 200 } while (0) 201 #define QUEUE_TAKE1(hio, name, ncomp, timeout) do { \ 202 bool _last; \ 203 \ 204 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 205 _last = false; \ 206 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \ 207 cv_timedwait(&hio_##name##_list_cond[(ncomp)], \ 208 &hio_##name##_list_lock[(ncomp)], (timeout)); \ 209 if ((timeout) != 0) \ 210 _last = true; \ 211 } \ 212 if (hio != NULL) { \ 213 PJDLOG_ASSERT(hio_##name##_list_size[(ncomp)] != 0); \ 214 hio_##name##_list_size[(ncomp)]--; \ 215 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \ 216 hio_next[(ncomp)]); \ 217 } \ 218 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 219 } while (0) 220 #define QUEUE_TAKE2(hio, name) do { \ 221 mtx_lock(&hio_##name##_list_lock); \ 222 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 223 cv_wait(&hio_##name##_list_cond, \ 224 &hio_##name##_list_lock); \ 225 } \ 226 PJDLOG_ASSERT(hio_##name##_list_size != 0); \ 227 hio_##name##_list_size--; \ 228 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \ 229 mtx_unlock(&hio_##name##_list_lock); \ 230 } while (0) 231 232 #define ISFULLSYNC(hio) ((hio)->hio_replication == HAST_REPLICATION_FULLSYNC) 233 #define ISMEMSYNC(hio) ((hio)->hio_replication == HAST_REPLICATION_MEMSYNC) 234 #define ISASYNC(hio) ((hio)->hio_replication == HAST_REPLICATION_ASYNC) 235 236 #define SYNCREQ(hio) do { \ 237 (hio)->hio_ggio.gctl_unit = -1; \ 238 (hio)->hio_ggio.gctl_seq = 1; \ 239 } while (0) 240 #define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1) 241 #define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0) 242 #define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2) 243 244 #define ISMEMSYNCWRITE(hio) (ISMEMSYNC(hio) && \ 245 (hio)->hio_ggio.gctl_cmd == BIO_WRITE && !ISSYNCREQ(hio)) 246 247 static struct hast_resource *gres; 248 249 static pthread_mutex_t range_lock; 250 static struct rangelocks *range_regular; 251 static bool range_regular_wait; 252 static pthread_cond_t range_regular_cond; 253 static struct rangelocks *range_sync; 254 static bool range_sync_wait; 255 static pthread_cond_t range_sync_cond; 256 static bool fullystarted; 257 258 static void *ggate_recv_thread(void *arg); 259 static void *local_send_thread(void *arg); 260 static void *remote_send_thread(void *arg); 261 static void *remote_recv_thread(void *arg); 262 static void *ggate_send_thread(void *arg); 263 static void *sync_thread(void *arg); 264 static void *guard_thread(void *arg); 265 266 static void 267 output_status_aux(struct nv *nvout) 268 { 269 270 nv_add_uint64(nvout, (uint64_t)hio_free_list_size, 271 "idle_queue_size"); 272 nv_add_uint64(nvout, (uint64_t)hio_send_local_list_size, 273 "local_queue_size"); 274 nv_add_uint64(nvout, (uint64_t)hio_send_remote_list_size, 275 "send_queue_size"); 276 nv_add_uint64(nvout, (uint64_t)hio_recv_remote_list_size, 277 "recv_queue_size"); 278 nv_add_uint64(nvout, (uint64_t)hio_done_list_size, 279 "done_queue_size"); 280 } 281 282 static void 283 cleanup(struct hast_resource *res) 284 { 285 int rerrno; 286 287 /* Remember errno. */ 288 rerrno = errno; 289 290 /* Destroy ggate provider if we created one. */ 291 if (res->hr_ggateunit >= 0) { 292 struct g_gate_ctl_destroy ggiod; 293 294 bzero(&ggiod, sizeof(ggiod)); 295 ggiod.gctl_version = G_GATE_VERSION; 296 ggiod.gctl_unit = res->hr_ggateunit; 297 ggiod.gctl_force = 1; 298 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) == -1) { 299 pjdlog_errno(LOG_WARNING, 300 "Unable to destroy hast/%s device", 301 res->hr_provname); 302 } 303 res->hr_ggateunit = -1; 304 } 305 306 /* Restore errno. */ 307 errno = rerrno; 308 } 309 310 static __dead2 void 311 primary_exit(int exitcode, const char *fmt, ...) 312 { 313 va_list ap; 314 315 PJDLOG_ASSERT(exitcode != EX_OK); 316 va_start(ap, fmt); 317 pjdlogv_errno(LOG_ERR, fmt, ap); 318 va_end(ap); 319 cleanup(gres); 320 exit(exitcode); 321 } 322 323 static __dead2 void 324 primary_exitx(int exitcode, const char *fmt, ...) 325 { 326 va_list ap; 327 328 va_start(ap, fmt); 329 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 330 va_end(ap); 331 cleanup(gres); 332 exit(exitcode); 333 } 334 335 static int 336 hast_activemap_flush(struct hast_resource *res) __unlocks(res->hr_amp_lock) 337 { 338 const unsigned char *buf; 339 size_t size; 340 int ret; 341 342 mtx_lock(&res->hr_amp_diskmap_lock); 343 buf = activemap_bitmap(res->hr_amp, &size); 344 mtx_unlock(&res->hr_amp_lock); 345 PJDLOG_ASSERT(buf != NULL); 346 PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0); 347 ret = 0; 348 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 349 (ssize_t)size) { 350 pjdlog_errno(LOG_ERR, "Unable to flush activemap to disk"); 351 res->hr_stat_activemap_write_error++; 352 ret = -1; 353 } 354 if (ret == 0 && res->hr_metaflush == 1 && 355 g_flush(res->hr_localfd) == -1) { 356 if (errno == EOPNOTSUPP) { 357 pjdlog_warning("The %s provider doesn't support flushing write cache. Disabling it.", 358 res->hr_localpath); 359 res->hr_metaflush = 0; 360 } else { 361 pjdlog_errno(LOG_ERR, 362 "Unable to flush disk cache on activemap update"); 363 res->hr_stat_activemap_flush_error++; 364 ret = -1; 365 } 366 } 367 mtx_unlock(&res->hr_amp_diskmap_lock); 368 return (ret); 369 } 370 371 static bool 372 real_remote(const struct hast_resource *res) 373 { 374 375 return (strcmp(res->hr_remoteaddr, "none") != 0); 376 } 377 378 static void 379 init_environment(struct hast_resource *res __unused) 380 { 381 struct hio *hio; 382 unsigned int ii, ncomps; 383 384 /* 385 * In the future it might be per-resource value. 386 */ 387 ncomps = HAST_NCOMPONENTS; 388 389 /* 390 * Allocate memory needed by lists. 391 */ 392 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 393 if (hio_send_list == NULL) { 394 primary_exitx(EX_TEMPFAIL, 395 "Unable to allocate %zu bytes of memory for send lists.", 396 sizeof(hio_send_list[0]) * ncomps); 397 } 398 hio_send_list_size = malloc(sizeof(hio_send_list_size[0]) * ncomps); 399 if (hio_send_list_size == NULL) { 400 primary_exitx(EX_TEMPFAIL, 401 "Unable to allocate %zu bytes of memory for send list counters.", 402 sizeof(hio_send_list_size[0]) * ncomps); 403 } 404 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 405 if (hio_send_list_lock == NULL) { 406 primary_exitx(EX_TEMPFAIL, 407 "Unable to allocate %zu bytes of memory for send list locks.", 408 sizeof(hio_send_list_lock[0]) * ncomps); 409 } 410 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 411 if (hio_send_list_cond == NULL) { 412 primary_exitx(EX_TEMPFAIL, 413 "Unable to allocate %zu bytes of memory for send list condition variables.", 414 sizeof(hio_send_list_cond[0]) * ncomps); 415 } 416 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 417 if (hio_recv_list == NULL) { 418 primary_exitx(EX_TEMPFAIL, 419 "Unable to allocate %zu bytes of memory for recv lists.", 420 sizeof(hio_recv_list[0]) * ncomps); 421 } 422 hio_recv_list_size = malloc(sizeof(hio_recv_list_size[0]) * ncomps); 423 if (hio_recv_list_size == NULL) { 424 primary_exitx(EX_TEMPFAIL, 425 "Unable to allocate %zu bytes of memory for recv list counters.", 426 sizeof(hio_recv_list_size[0]) * ncomps); 427 } 428 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 429 if (hio_recv_list_lock == NULL) { 430 primary_exitx(EX_TEMPFAIL, 431 "Unable to allocate %zu bytes of memory for recv list locks.", 432 sizeof(hio_recv_list_lock[0]) * ncomps); 433 } 434 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 435 if (hio_recv_list_cond == NULL) { 436 primary_exitx(EX_TEMPFAIL, 437 "Unable to allocate %zu bytes of memory for recv list condition variables.", 438 sizeof(hio_recv_list_cond[0]) * ncomps); 439 } 440 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 441 if (hio_remote_lock == NULL) { 442 primary_exitx(EX_TEMPFAIL, 443 "Unable to allocate %zu bytes of memory for remote connections locks.", 444 sizeof(hio_remote_lock[0]) * ncomps); 445 } 446 447 /* 448 * Initialize lists, their counters, locks and condition variables. 449 */ 450 TAILQ_INIT(&hio_free_list); 451 mtx_init(&hio_free_list_lock); 452 cv_init(&hio_free_list_cond); 453 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 454 TAILQ_INIT(&hio_send_list[ii]); 455 hio_send_list_size[ii] = 0; 456 mtx_init(&hio_send_list_lock[ii]); 457 cv_init(&hio_send_list_cond[ii]); 458 TAILQ_INIT(&hio_recv_list[ii]); 459 hio_recv_list_size[ii] = 0; 460 mtx_init(&hio_recv_list_lock[ii]); 461 cv_init(&hio_recv_list_cond[ii]); 462 rw_init(&hio_remote_lock[ii]); 463 } 464 TAILQ_INIT(&hio_done_list); 465 mtx_init(&hio_done_list_lock); 466 cv_init(&hio_done_list_cond); 467 mtx_init(&metadata_lock); 468 469 /* 470 * Allocate requests pool and initialize requests. 471 */ 472 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 473 hio = malloc(sizeof(*hio)); 474 if (hio == NULL) { 475 primary_exitx(EX_TEMPFAIL, 476 "Unable to allocate %zu bytes of memory for hio request.", 477 sizeof(*hio)); 478 } 479 refcnt_init(&hio->hio_countdown, 0); 480 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 481 if (hio->hio_errors == NULL) { 482 primary_exitx(EX_TEMPFAIL, 483 "Unable allocate %zu bytes of memory for hio errors.", 484 sizeof(hio->hio_errors[0]) * ncomps); 485 } 486 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 487 if (hio->hio_next == NULL) { 488 primary_exitx(EX_TEMPFAIL, 489 "Unable allocate %zu bytes of memory for hio_next field.", 490 sizeof(hio->hio_next[0]) * ncomps); 491 } 492 hio->hio_ggio.gctl_version = G_GATE_VERSION; 493 hio->hio_ggio.gctl_data = malloc(MAXPHYS); 494 if (hio->hio_ggio.gctl_data == NULL) { 495 primary_exitx(EX_TEMPFAIL, 496 "Unable to allocate %zu bytes of memory for gctl_data.", 497 MAXPHYS); 498 } 499 hio->hio_ggio.gctl_length = MAXPHYS; 500 hio->hio_ggio.gctl_error = 0; 501 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 502 hio_free_list_size++; 503 } 504 } 505 506 static bool 507 init_resuid(struct hast_resource *res) 508 { 509 510 mtx_lock(&metadata_lock); 511 if (res->hr_resuid != 0) { 512 mtx_unlock(&metadata_lock); 513 return (false); 514 } else { 515 /* Initialize unique resource identifier. */ 516 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 517 mtx_unlock(&metadata_lock); 518 if (metadata_write(res) == -1) 519 exit(EX_NOINPUT); 520 return (true); 521 } 522 } 523 524 static void 525 init_local(struct hast_resource *res) 526 { 527 unsigned char *buf; 528 size_t mapsize; 529 530 if (metadata_read(res, true) == -1) 531 exit(EX_NOINPUT); 532 mtx_init(&res->hr_amp_lock); 533 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 534 res->hr_local_sectorsize, res->hr_keepdirty) == -1) { 535 primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 536 } 537 mtx_init(&range_lock); 538 cv_init(&range_regular_cond); 539 if (rangelock_init(&range_regular) == -1) 540 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 541 cv_init(&range_sync_cond); 542 if (rangelock_init(&range_sync) == -1) 543 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 544 mapsize = activemap_ondisk_size(res->hr_amp); 545 buf = calloc(1, mapsize); 546 if (buf == NULL) { 547 primary_exitx(EX_TEMPFAIL, 548 "Unable to allocate buffer for activemap."); 549 } 550 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 551 (ssize_t)mapsize) { 552 primary_exit(EX_NOINPUT, "Unable to read activemap"); 553 } 554 activemap_copyin(res->hr_amp, buf, mapsize); 555 free(buf); 556 if (res->hr_resuid != 0) 557 return; 558 /* 559 * We're using provider for the first time. Initialize local and remote 560 * counters. We don't initialize resuid here, as we want to do it just 561 * in time. The reason for this is that we want to inform secondary 562 * that there were no writes yet, so there is no need to synchronize 563 * anything. 564 */ 565 res->hr_primary_localcnt = 0; 566 res->hr_primary_remotecnt = 0; 567 if (metadata_write(res) == -1) 568 exit(EX_NOINPUT); 569 } 570 571 static int 572 primary_connect(struct hast_resource *res, struct proto_conn **connp) 573 { 574 struct proto_conn *conn; 575 int16_t val; 576 577 val = 1; 578 if (proto_send(res->hr_conn, &val, sizeof(val)) == -1) { 579 primary_exit(EX_TEMPFAIL, 580 "Unable to send connection request to parent"); 581 } 582 if (proto_recv(res->hr_conn, &val, sizeof(val)) == -1) { 583 primary_exit(EX_TEMPFAIL, 584 "Unable to receive reply to connection request from parent"); 585 } 586 if (val != 0) { 587 errno = val; 588 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 589 res->hr_remoteaddr); 590 return (-1); 591 } 592 if (proto_connection_recv(res->hr_conn, true, &conn) == -1) { 593 primary_exit(EX_TEMPFAIL, 594 "Unable to receive connection from parent"); 595 } 596 if (proto_connect_wait(conn, res->hr_timeout) == -1) { 597 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 598 res->hr_remoteaddr); 599 proto_close(conn); 600 return (-1); 601 } 602 /* Error in setting timeout is not critical, but why should it fail? */ 603 if (proto_timeout(conn, res->hr_timeout) == -1) 604 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 605 606 *connp = conn; 607 608 return (0); 609 } 610 611 /* 612 * Function instructs GEOM_GATE to handle reads directly from within the kernel. 613 */ 614 static void 615 enable_direct_reads(struct hast_resource *res) 616 { 617 struct g_gate_ctl_modify ggiomodify; 618 619 bzero(&ggiomodify, sizeof(ggiomodify)); 620 ggiomodify.gctl_version = G_GATE_VERSION; 621 ggiomodify.gctl_unit = res->hr_ggateunit; 622 ggiomodify.gctl_modify = GG_MODIFY_READPROV | GG_MODIFY_READOFFSET; 623 strlcpy(ggiomodify.gctl_readprov, res->hr_localpath, 624 sizeof(ggiomodify.gctl_readprov)); 625 ggiomodify.gctl_readoffset = res->hr_localoff; 626 if (ioctl(res->hr_ggatefd, G_GATE_CMD_MODIFY, &ggiomodify) == 0) 627 pjdlog_debug(1, "Direct reads enabled."); 628 else 629 pjdlog_errno(LOG_WARNING, "Failed to enable direct reads"); 630 } 631 632 static int 633 init_remote(struct hast_resource *res, struct proto_conn **inp, 634 struct proto_conn **outp) 635 { 636 struct proto_conn *in, *out; 637 struct nv *nvout, *nvin; 638 const unsigned char *token; 639 unsigned char *map; 640 const char *errmsg; 641 int32_t extentsize; 642 int64_t datasize; 643 uint32_t mapsize; 644 uint8_t version; 645 size_t size; 646 int error; 647 648 PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 649 PJDLOG_ASSERT(real_remote(res)); 650 651 in = out = NULL; 652 errmsg = NULL; 653 654 if (primary_connect(res, &out) == -1) 655 return (ECONNREFUSED); 656 657 error = ECONNABORTED; 658 659 /* 660 * First handshake step. 661 * Setup outgoing connection with remote node. 662 */ 663 nvout = nv_alloc(); 664 nv_add_string(nvout, res->hr_name, "resource"); 665 nv_add_uint8(nvout, HAST_PROTO_VERSION, "version"); 666 if (nv_error(nvout) != 0) { 667 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 668 "Unable to allocate header for connection with %s", 669 res->hr_remoteaddr); 670 nv_free(nvout); 671 goto close; 672 } 673 if (hast_proto_send(res, out, nvout, NULL, 0) == -1) { 674 pjdlog_errno(LOG_WARNING, 675 "Unable to send handshake header to %s", 676 res->hr_remoteaddr); 677 nv_free(nvout); 678 goto close; 679 } 680 nv_free(nvout); 681 if (hast_proto_recv_hdr(out, &nvin) == -1) { 682 pjdlog_errno(LOG_WARNING, 683 "Unable to receive handshake header from %s", 684 res->hr_remoteaddr); 685 goto close; 686 } 687 errmsg = nv_get_string(nvin, "errmsg"); 688 if (errmsg != NULL) { 689 pjdlog_warning("%s", errmsg); 690 if (nv_exists(nvin, "wait")) 691 error = EBUSY; 692 nv_free(nvin); 693 goto close; 694 } 695 version = nv_get_uint8(nvin, "version"); 696 if (version == 0) { 697 /* 698 * If no version is sent, it means this is protocol version 1. 699 */ 700 version = 1; 701 } 702 if (version > HAST_PROTO_VERSION) { 703 pjdlog_warning("Invalid version received (%hhu).", version); 704 nv_free(nvin); 705 goto close; 706 } 707 res->hr_version = version; 708 pjdlog_debug(1, "Negotiated protocol version %d.", res->hr_version); 709 token = nv_get_uint8_array(nvin, &size, "token"); 710 if (token == NULL) { 711 pjdlog_warning("Handshake header from %s has no 'token' field.", 712 res->hr_remoteaddr); 713 nv_free(nvin); 714 goto close; 715 } 716 if (size != sizeof(res->hr_token)) { 717 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 718 res->hr_remoteaddr, size, sizeof(res->hr_token)); 719 nv_free(nvin); 720 goto close; 721 } 722 bcopy(token, res->hr_token, sizeof(res->hr_token)); 723 nv_free(nvin); 724 725 /* 726 * Second handshake step. 727 * Setup incoming connection with remote node. 728 */ 729 if (primary_connect(res, &in) == -1) 730 goto close; 731 732 nvout = nv_alloc(); 733 nv_add_string(nvout, res->hr_name, "resource"); 734 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 735 "token"); 736 if (res->hr_resuid == 0) { 737 /* 738 * The resuid field was not yet initialized. 739 * Because we do synchronization inside init_resuid(), it is 740 * possible that someone already initialized it, the function 741 * will return false then, but if we successfully initialized 742 * it, we will get true. True means that there were no writes 743 * to this resource yet and we want to inform secondary that 744 * synchronization is not needed by sending "virgin" argument. 745 */ 746 if (init_resuid(res)) 747 nv_add_int8(nvout, 1, "virgin"); 748 } 749 nv_add_uint64(nvout, res->hr_resuid, "resuid"); 750 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 751 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 752 if (nv_error(nvout) != 0) { 753 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 754 "Unable to allocate header for connection with %s", 755 res->hr_remoteaddr); 756 nv_free(nvout); 757 goto close; 758 } 759 if (hast_proto_send(res, in, nvout, NULL, 0) == -1) { 760 pjdlog_errno(LOG_WARNING, 761 "Unable to send handshake header to %s", 762 res->hr_remoteaddr); 763 nv_free(nvout); 764 goto close; 765 } 766 nv_free(nvout); 767 if (hast_proto_recv_hdr(out, &nvin) == -1) { 768 pjdlog_errno(LOG_WARNING, 769 "Unable to receive handshake header from %s", 770 res->hr_remoteaddr); 771 goto close; 772 } 773 errmsg = nv_get_string(nvin, "errmsg"); 774 if (errmsg != NULL) { 775 pjdlog_warning("%s", errmsg); 776 nv_free(nvin); 777 goto close; 778 } 779 datasize = nv_get_int64(nvin, "datasize"); 780 if (datasize != res->hr_datasize) { 781 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 782 (intmax_t)res->hr_datasize, (intmax_t)datasize); 783 nv_free(nvin); 784 goto close; 785 } 786 extentsize = nv_get_int32(nvin, "extentsize"); 787 if (extentsize != res->hr_extentsize) { 788 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 789 (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 790 nv_free(nvin); 791 goto close; 792 } 793 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 794 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 795 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 796 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) 797 enable_direct_reads(res); 798 if (nv_exists(nvin, "virgin")) { 799 /* 800 * Secondary was reinitialized, bump localcnt if it is 0 as 801 * only we have the data. 802 */ 803 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY); 804 PJDLOG_ASSERT(res->hr_secondary_localcnt == 0); 805 806 if (res->hr_primary_localcnt == 0) { 807 PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0); 808 809 mtx_lock(&metadata_lock); 810 res->hr_primary_localcnt++; 811 pjdlog_debug(1, "Increasing localcnt to %ju.", 812 (uintmax_t)res->hr_primary_localcnt); 813 (void)metadata_write(res); 814 mtx_unlock(&metadata_lock); 815 } 816 } 817 map = NULL; 818 mapsize = nv_get_uint32(nvin, "mapsize"); 819 if (mapsize > 0) { 820 map = malloc(mapsize); 821 if (map == NULL) { 822 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 823 (uintmax_t)mapsize); 824 nv_free(nvin); 825 goto close; 826 } 827 /* 828 * Remote node have some dirty extents on its own, lets 829 * download its activemap. 830 */ 831 if (hast_proto_recv_data(res, out, nvin, map, 832 mapsize) == -1) { 833 pjdlog_errno(LOG_ERR, 834 "Unable to receive remote activemap"); 835 nv_free(nvin); 836 free(map); 837 goto close; 838 } 839 mtx_lock(&res->hr_amp_lock); 840 /* 841 * Merge local and remote bitmaps. 842 */ 843 activemap_merge(res->hr_amp, map, mapsize); 844 free(map); 845 /* 846 * Now that we merged bitmaps from both nodes, flush it to the 847 * disk before we start to synchronize. 848 */ 849 (void)hast_activemap_flush(res); 850 } 851 nv_free(nvin); 852 #ifdef notyet 853 /* Setup directions. */ 854 if (proto_send(out, NULL, 0) == -1) 855 pjdlog_errno(LOG_WARNING, "Unable to set connection direction"); 856 if (proto_recv(in, NULL, 0) == -1) 857 pjdlog_errno(LOG_WARNING, "Unable to set connection direction"); 858 #endif 859 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 860 if (res->hr_original_replication == HAST_REPLICATION_MEMSYNC && 861 res->hr_version < 2) { 862 pjdlog_warning("The 'memsync' replication mode is not supported by the remote node, falling back to 'fullsync' mode."); 863 res->hr_replication = HAST_REPLICATION_FULLSYNC; 864 } else if (res->hr_replication != res->hr_original_replication) { 865 /* 866 * This is in case hastd disconnected and was upgraded. 867 */ 868 res->hr_replication = res->hr_original_replication; 869 } 870 if (inp != NULL && outp != NULL) { 871 *inp = in; 872 *outp = out; 873 } else { 874 res->hr_remotein = in; 875 res->hr_remoteout = out; 876 } 877 event_send(res, EVENT_CONNECT); 878 return (0); 879 close: 880 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 881 event_send(res, EVENT_SPLITBRAIN); 882 proto_close(out); 883 if (in != NULL) 884 proto_close(in); 885 return (error); 886 } 887 888 static void 889 sync_start(void) 890 { 891 892 mtx_lock(&sync_lock); 893 sync_inprogress = true; 894 mtx_unlock(&sync_lock); 895 cv_signal(&sync_cond); 896 } 897 898 static void 899 sync_stop(void) 900 { 901 902 mtx_lock(&sync_lock); 903 if (sync_inprogress) 904 sync_inprogress = false; 905 mtx_unlock(&sync_lock); 906 } 907 908 static void 909 init_ggate(struct hast_resource *res) 910 { 911 struct g_gate_ctl_create ggiocreate; 912 struct g_gate_ctl_cancel ggiocancel; 913 914 /* 915 * We communicate with ggate via /dev/ggctl. Open it. 916 */ 917 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 918 if (res->hr_ggatefd == -1) 919 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 920 /* 921 * Create provider before trying to connect, as connection failure 922 * is not critical, but may take some time. 923 */ 924 bzero(&ggiocreate, sizeof(ggiocreate)); 925 ggiocreate.gctl_version = G_GATE_VERSION; 926 ggiocreate.gctl_mediasize = res->hr_datasize; 927 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 928 ggiocreate.gctl_flags = 0; 929 ggiocreate.gctl_maxcount = 0; 930 ggiocreate.gctl_timeout = 0; 931 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 932 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 933 res->hr_provname); 934 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 935 pjdlog_info("Device hast/%s created.", res->hr_provname); 936 res->hr_ggateunit = ggiocreate.gctl_unit; 937 return; 938 } 939 if (errno != EEXIST) { 940 primary_exit(EX_OSERR, "Unable to create hast/%s device", 941 res->hr_provname); 942 } 943 pjdlog_debug(1, 944 "Device hast/%s already exists, we will try to take it over.", 945 res->hr_provname); 946 /* 947 * If we received EEXIST, we assume that the process who created the 948 * provider died and didn't clean up. In that case we will start from 949 * where he left of. 950 */ 951 bzero(&ggiocancel, sizeof(ggiocancel)); 952 ggiocancel.gctl_version = G_GATE_VERSION; 953 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 954 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 955 res->hr_provname); 956 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 957 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 958 res->hr_ggateunit = ggiocancel.gctl_unit; 959 return; 960 } 961 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 962 res->hr_provname); 963 } 964 965 void 966 hastd_primary(struct hast_resource *res) 967 { 968 pthread_t td; 969 pid_t pid; 970 int error, mode, debuglevel; 971 972 /* 973 * Create communication channel for sending control commands from 974 * parent to child. 975 */ 976 if (proto_client(NULL, "socketpair://", &res->hr_ctrl) == -1) { 977 /* TODO: There's no need for this to be fatal error. */ 978 KEEP_ERRNO((void)pidfile_remove(pfh)); 979 pjdlog_exit(EX_OSERR, 980 "Unable to create control sockets between parent and child"); 981 } 982 /* 983 * Create communication channel for sending events from child to parent. 984 */ 985 if (proto_client(NULL, "socketpair://", &res->hr_event) == -1) { 986 /* TODO: There's no need for this to be fatal error. */ 987 KEEP_ERRNO((void)pidfile_remove(pfh)); 988 pjdlog_exit(EX_OSERR, 989 "Unable to create event sockets between child and parent"); 990 } 991 /* 992 * Create communication channel for sending connection requests from 993 * child to parent. 994 */ 995 if (proto_client(NULL, "socketpair://", &res->hr_conn) == -1) { 996 /* TODO: There's no need for this to be fatal error. */ 997 KEEP_ERRNO((void)pidfile_remove(pfh)); 998 pjdlog_exit(EX_OSERR, 999 "Unable to create connection sockets between child and parent"); 1000 } 1001 1002 pid = fork(); 1003 if (pid == -1) { 1004 /* TODO: There's no need for this to be fatal error. */ 1005 KEEP_ERRNO((void)pidfile_remove(pfh)); 1006 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 1007 } 1008 1009 if (pid > 0) { 1010 /* This is parent. */ 1011 /* Declare that we are receiver. */ 1012 proto_recv(res->hr_event, NULL, 0); 1013 proto_recv(res->hr_conn, NULL, 0); 1014 /* Declare that we are sender. */ 1015 proto_send(res->hr_ctrl, NULL, 0); 1016 res->hr_workerpid = pid; 1017 return; 1018 } 1019 1020 gres = res; 1021 res->output_status_aux = output_status_aux; 1022 mode = pjdlog_mode_get(); 1023 debuglevel = pjdlog_debug_get(); 1024 1025 /* Declare that we are sender. */ 1026 proto_send(res->hr_event, NULL, 0); 1027 proto_send(res->hr_conn, NULL, 0); 1028 /* Declare that we are receiver. */ 1029 proto_recv(res->hr_ctrl, NULL, 0); 1030 descriptors_cleanup(res); 1031 1032 descriptors_assert(res, mode); 1033 1034 pjdlog_init(mode); 1035 pjdlog_debug_set(debuglevel); 1036 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 1037 setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role)); 1038 1039 init_local(res); 1040 init_ggate(res); 1041 init_environment(res); 1042 1043 if (drop_privs(res) != 0) { 1044 cleanup(res); 1045 exit(EX_CONFIG); 1046 } 1047 pjdlog_info("Privileges successfully dropped."); 1048 1049 /* 1050 * Create the guard thread first, so we can handle signals from the 1051 * very beginning. 1052 */ 1053 error = pthread_create(&td, NULL, guard_thread, res); 1054 PJDLOG_ASSERT(error == 0); 1055 /* 1056 * Create the control thread before sending any event to the parent, 1057 * as we can deadlock when parent sends control request to worker, 1058 * but worker has no control thread started yet, so parent waits. 1059 * In the meantime worker sends an event to the parent, but parent 1060 * is unable to handle the event, because it waits for control 1061 * request response. 1062 */ 1063 error = pthread_create(&td, NULL, ctrl_thread, res); 1064 PJDLOG_ASSERT(error == 0); 1065 if (real_remote(res)) { 1066 error = init_remote(res, NULL, NULL); 1067 if (error == 0) { 1068 sync_start(); 1069 } else if (error == EBUSY) { 1070 time_t start = time(NULL); 1071 1072 pjdlog_warning("Waiting for remote node to become %s for %ds.", 1073 role2str(HAST_ROLE_SECONDARY), 1074 res->hr_timeout); 1075 for (;;) { 1076 sleep(1); 1077 error = init_remote(res, NULL, NULL); 1078 if (error != EBUSY) 1079 break; 1080 if (time(NULL) > start + res->hr_timeout) 1081 break; 1082 } 1083 if (error == EBUSY) { 1084 pjdlog_warning("Remote node is still %s, starting anyway.", 1085 role2str(HAST_ROLE_PRIMARY)); 1086 } 1087 } 1088 } 1089 error = pthread_create(&td, NULL, ggate_recv_thread, res); 1090 PJDLOG_ASSERT(error == 0); 1091 error = pthread_create(&td, NULL, local_send_thread, res); 1092 PJDLOG_ASSERT(error == 0); 1093 error = pthread_create(&td, NULL, remote_send_thread, res); 1094 PJDLOG_ASSERT(error == 0); 1095 error = pthread_create(&td, NULL, remote_recv_thread, res); 1096 PJDLOG_ASSERT(error == 0); 1097 error = pthread_create(&td, NULL, ggate_send_thread, res); 1098 PJDLOG_ASSERT(error == 0); 1099 fullystarted = true; 1100 (void)sync_thread(res); 1101 } 1102 1103 static void 1104 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, 1105 const char *fmt, ...) 1106 { 1107 char msg[1024]; 1108 va_list ap; 1109 1110 va_start(ap, fmt); 1111 (void)vsnprintf(msg, sizeof(msg), fmt, ap); 1112 va_end(ap); 1113 switch (ggio->gctl_cmd) { 1114 case BIO_READ: 1115 (void)snprlcat(msg, sizeof(msg), "READ(%ju, %ju).", 1116 (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length); 1117 break; 1118 case BIO_DELETE: 1119 (void)snprlcat(msg, sizeof(msg), "DELETE(%ju, %ju).", 1120 (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length); 1121 break; 1122 case BIO_FLUSH: 1123 (void)snprlcat(msg, sizeof(msg), "FLUSH."); 1124 break; 1125 case BIO_WRITE: 1126 (void)snprlcat(msg, sizeof(msg), "WRITE(%ju, %ju).", 1127 (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length); 1128 break; 1129 default: 1130 (void)snprlcat(msg, sizeof(msg), "UNKNOWN(%u).", 1131 (unsigned int)ggio->gctl_cmd); 1132 break; 1133 } 1134 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 1135 } 1136 1137 static void 1138 remote_close(struct hast_resource *res, int ncomp) 1139 { 1140 1141 rw_wlock(&hio_remote_lock[ncomp]); 1142 /* 1143 * Check for a race between dropping rlock and acquiring wlock - 1144 * another thread can close connection in-between. 1145 */ 1146 if (!ISCONNECTED(res, ncomp)) { 1147 PJDLOG_ASSERT(res->hr_remotein == NULL); 1148 PJDLOG_ASSERT(res->hr_remoteout == NULL); 1149 rw_unlock(&hio_remote_lock[ncomp]); 1150 return; 1151 } 1152 1153 PJDLOG_ASSERT(res->hr_remotein != NULL); 1154 PJDLOG_ASSERT(res->hr_remoteout != NULL); 1155 1156 pjdlog_debug(2, "Closing incoming connection to %s.", 1157 res->hr_remoteaddr); 1158 proto_close(res->hr_remotein); 1159 res->hr_remotein = NULL; 1160 pjdlog_debug(2, "Closing outgoing connection to %s.", 1161 res->hr_remoteaddr); 1162 proto_close(res->hr_remoteout); 1163 res->hr_remoteout = NULL; 1164 1165 rw_unlock(&hio_remote_lock[ncomp]); 1166 1167 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 1168 1169 /* 1170 * Stop synchronization if in-progress. 1171 */ 1172 sync_stop(); 1173 1174 event_send(res, EVENT_DISCONNECT); 1175 } 1176 1177 /* 1178 * Acknowledge write completion to the kernel, but don't update activemap yet. 1179 */ 1180 static void 1181 write_complete(struct hast_resource *res, struct hio *hio) 1182 { 1183 struct g_gate_ctl_io *ggio; 1184 unsigned int ncomp; 1185 1186 PJDLOG_ASSERT(!hio->hio_done); 1187 1188 ggio = &hio->hio_ggio; 1189 PJDLOG_ASSERT(ggio->gctl_cmd == BIO_WRITE); 1190 1191 /* 1192 * Bump local count if this is first write after 1193 * connection failure with remote node. 1194 */ 1195 ncomp = 1; 1196 rw_rlock(&hio_remote_lock[ncomp]); 1197 if (!ISCONNECTED(res, ncomp)) { 1198 mtx_lock(&metadata_lock); 1199 if (res->hr_primary_localcnt == res->hr_secondary_remotecnt) { 1200 res->hr_primary_localcnt++; 1201 pjdlog_debug(1, "Increasing localcnt to %ju.", 1202 (uintmax_t)res->hr_primary_localcnt); 1203 (void)metadata_write(res); 1204 } 1205 mtx_unlock(&metadata_lock); 1206 } 1207 rw_unlock(&hio_remote_lock[ncomp]); 1208 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1) 1209 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1210 hio->hio_done = true; 1211 } 1212 1213 /* 1214 * Thread receives ggate I/O requests from the kernel and passes them to 1215 * appropriate threads: 1216 * WRITE - always goes to both local_send and remote_send threads 1217 * READ (when the block is up-to-date on local component) - 1218 * only local_send thread 1219 * READ (when the block isn't up-to-date on local component) - 1220 * only remote_send thread 1221 * DELETE - always goes to both local_send and remote_send threads 1222 * FLUSH - always goes to both local_send and remote_send threads 1223 */ 1224 static void * 1225 ggate_recv_thread(void *arg) 1226 { 1227 struct hast_resource *res = arg; 1228 struct g_gate_ctl_io *ggio; 1229 struct hio *hio; 1230 unsigned int ii, ncomp, ncomps; 1231 int error; 1232 1233 for (;;) { 1234 pjdlog_debug(2, "ggate_recv: Taking free request."); 1235 QUEUE_TAKE2(hio, free); 1236 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 1237 ggio = &hio->hio_ggio; 1238 ggio->gctl_unit = res->hr_ggateunit; 1239 ggio->gctl_length = MAXPHYS; 1240 ggio->gctl_error = 0; 1241 hio->hio_done = false; 1242 hio->hio_replication = res->hr_replication; 1243 pjdlog_debug(2, 1244 "ggate_recv: (%p) Waiting for request from the kernel.", 1245 hio); 1246 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) == -1) { 1247 if (sigexit_received) 1248 pthread_exit(NULL); 1249 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 1250 } 1251 error = ggio->gctl_error; 1252 switch (error) { 1253 case 0: 1254 break; 1255 case ECANCELED: 1256 /* Exit gracefully. */ 1257 if (!sigexit_received) { 1258 pjdlog_debug(2, 1259 "ggate_recv: (%p) Received cancel from the kernel.", 1260 hio); 1261 pjdlog_info("Received cancel from the kernel, exiting."); 1262 } 1263 pthread_exit(NULL); 1264 case ENOMEM: 1265 /* 1266 * Buffer too small? Impossible, we allocate MAXPHYS 1267 * bytes - request can't be bigger than that. 1268 */ 1269 /* FALLTHROUGH */ 1270 case ENXIO: 1271 default: 1272 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 1273 strerror(error)); 1274 } 1275 1276 ncomp = 0; 1277 ncomps = HAST_NCOMPONENTS; 1278 1279 for (ii = 0; ii < ncomps; ii++) 1280 hio->hio_errors[ii] = EINVAL; 1281 reqlog(LOG_DEBUG, 2, ggio, 1282 "ggate_recv: (%p) Request received from the kernel: ", 1283 hio); 1284 1285 /* 1286 * Inform all components about new write request. 1287 * For read request prefer local component unless the given 1288 * range is out-of-date, then use remote component. 1289 */ 1290 switch (ggio->gctl_cmd) { 1291 case BIO_READ: 1292 res->hr_stat_read++; 1293 ncomps = 1; 1294 mtx_lock(&metadata_lock); 1295 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 1296 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1297 /* 1298 * This range is up-to-date on local component, 1299 * so handle request locally. 1300 */ 1301 /* Local component is 0 for now. */ 1302 ncomp = 0; 1303 } else /* if (res->hr_syncsrc == 1304 HAST_SYNCSRC_SECONDARY) */ { 1305 PJDLOG_ASSERT(res->hr_syncsrc == 1306 HAST_SYNCSRC_SECONDARY); 1307 /* 1308 * This range is out-of-date on local component, 1309 * so send request to the remote node. 1310 */ 1311 /* Remote component is 1 for now. */ 1312 ncomp = 1; 1313 } 1314 mtx_unlock(&metadata_lock); 1315 break; 1316 case BIO_WRITE: 1317 res->hr_stat_write++; 1318 if (res->hr_resuid == 0 && 1319 res->hr_primary_localcnt == 0) { 1320 /* This is first write. */ 1321 res->hr_primary_localcnt = 1; 1322 } 1323 for (;;) { 1324 mtx_lock(&range_lock); 1325 if (rangelock_islocked(range_sync, 1326 ggio->gctl_offset, ggio->gctl_length)) { 1327 pjdlog_debug(2, 1328 "regular: Range offset=%jd length=%zu locked.", 1329 (intmax_t)ggio->gctl_offset, 1330 (size_t)ggio->gctl_length); 1331 range_regular_wait = true; 1332 cv_wait(&range_regular_cond, &range_lock); 1333 range_regular_wait = false; 1334 mtx_unlock(&range_lock); 1335 continue; 1336 } 1337 if (rangelock_add(range_regular, 1338 ggio->gctl_offset, ggio->gctl_length) == -1) { 1339 mtx_unlock(&range_lock); 1340 pjdlog_debug(2, 1341 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1342 (intmax_t)ggio->gctl_offset, 1343 (size_t)ggio->gctl_length); 1344 sleep(1); 1345 continue; 1346 } 1347 mtx_unlock(&range_lock); 1348 break; 1349 } 1350 mtx_lock(&res->hr_amp_lock); 1351 if (activemap_write_start(res->hr_amp, 1352 ggio->gctl_offset, ggio->gctl_length)) { 1353 res->hr_stat_activemap_update++; 1354 (void)hast_activemap_flush(res); 1355 } else { 1356 mtx_unlock(&res->hr_amp_lock); 1357 } 1358 if (ISMEMSYNC(hio)) { 1359 hio->hio_memsyncacked = false; 1360 refcnt_init(&hio->hio_writecount, ncomps); 1361 } 1362 break; 1363 case BIO_DELETE: 1364 res->hr_stat_delete++; 1365 break; 1366 case BIO_FLUSH: 1367 res->hr_stat_flush++; 1368 break; 1369 } 1370 pjdlog_debug(2, 1371 "ggate_recv: (%p) Moving request to the send queues.", hio); 1372 refcnt_init(&hio->hio_countdown, ncomps); 1373 for (ii = ncomp; ii < ncomps; ii++) 1374 QUEUE_INSERT1(hio, send, ii); 1375 } 1376 /* NOTREACHED */ 1377 return (NULL); 1378 } 1379 1380 /* 1381 * Thread reads from or writes to local component. 1382 * If local read fails, it redirects it to remote_send thread. 1383 */ 1384 static void * 1385 local_send_thread(void *arg) 1386 { 1387 struct hast_resource *res = arg; 1388 struct g_gate_ctl_io *ggio; 1389 struct hio *hio; 1390 unsigned int ncomp, rncomp; 1391 ssize_t ret; 1392 1393 /* Local component is 0 for now. */ 1394 ncomp = 0; 1395 /* Remote component is 1 for now. */ 1396 rncomp = 1; 1397 1398 for (;;) { 1399 pjdlog_debug(2, "local_send: Taking request."); 1400 QUEUE_TAKE1(hio, send, ncomp, 0); 1401 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1402 ggio = &hio->hio_ggio; 1403 switch (ggio->gctl_cmd) { 1404 case BIO_READ: 1405 ret = pread(res->hr_localfd, ggio->gctl_data, 1406 ggio->gctl_length, 1407 ggio->gctl_offset + res->hr_localoff); 1408 if (ret == ggio->gctl_length) 1409 hio->hio_errors[ncomp] = 0; 1410 else if (!ISSYNCREQ(hio)) { 1411 /* 1412 * If READ failed, try to read from remote node. 1413 */ 1414 if (ret == -1) { 1415 reqlog(LOG_WARNING, 0, ggio, 1416 "Local request failed (%s), trying remote node. ", 1417 strerror(errno)); 1418 } else if (ret != ggio->gctl_length) { 1419 reqlog(LOG_WARNING, 0, ggio, 1420 "Local request failed (%zd != %jd), trying remote node. ", 1421 ret, (intmax_t)ggio->gctl_length); 1422 } 1423 QUEUE_INSERT1(hio, send, rncomp); 1424 continue; 1425 } 1426 break; 1427 case BIO_WRITE: 1428 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1429 ggio->gctl_length, 1430 ggio->gctl_offset + res->hr_localoff); 1431 if (ret == -1) { 1432 hio->hio_errors[ncomp] = errno; 1433 reqlog(LOG_WARNING, 0, ggio, 1434 "Local request failed (%s): ", 1435 strerror(errno)); 1436 } else if (ret != ggio->gctl_length) { 1437 hio->hio_errors[ncomp] = EIO; 1438 reqlog(LOG_WARNING, 0, ggio, 1439 "Local request failed (%zd != %jd): ", 1440 ret, (intmax_t)ggio->gctl_length); 1441 } else { 1442 hio->hio_errors[ncomp] = 0; 1443 if (ISASYNC(hio)) { 1444 ggio->gctl_error = 0; 1445 write_complete(res, hio); 1446 } 1447 } 1448 break; 1449 case BIO_DELETE: 1450 ret = g_delete(res->hr_localfd, 1451 ggio->gctl_offset + res->hr_localoff, 1452 ggio->gctl_length); 1453 if (ret == -1) { 1454 hio->hio_errors[ncomp] = errno; 1455 reqlog(LOG_WARNING, 0, ggio, 1456 "Local request failed (%s): ", 1457 strerror(errno)); 1458 } else { 1459 hio->hio_errors[ncomp] = 0; 1460 } 1461 break; 1462 case BIO_FLUSH: 1463 if (!res->hr_localflush) { 1464 ret = -1; 1465 errno = EOPNOTSUPP; 1466 break; 1467 } 1468 ret = g_flush(res->hr_localfd); 1469 if (ret == -1) { 1470 if (errno == EOPNOTSUPP) 1471 res->hr_localflush = false; 1472 hio->hio_errors[ncomp] = errno; 1473 reqlog(LOG_WARNING, 0, ggio, 1474 "Local request failed (%s): ", 1475 strerror(errno)); 1476 } else { 1477 hio->hio_errors[ncomp] = 0; 1478 } 1479 break; 1480 } 1481 if (ISMEMSYNCWRITE(hio)) { 1482 if (refcnt_release(&hio->hio_writecount) == 0) { 1483 write_complete(res, hio); 1484 } 1485 } 1486 if (refcnt_release(&hio->hio_countdown) > 0) 1487 continue; 1488 if (ISSYNCREQ(hio)) { 1489 mtx_lock(&sync_lock); 1490 SYNCREQDONE(hio); 1491 mtx_unlock(&sync_lock); 1492 cv_signal(&sync_cond); 1493 } else { 1494 pjdlog_debug(2, 1495 "local_send: (%p) Moving request to the done queue.", 1496 hio); 1497 QUEUE_INSERT2(hio, done); 1498 } 1499 } 1500 /* NOTREACHED */ 1501 return (NULL); 1502 } 1503 1504 static void 1505 keepalive_send(struct hast_resource *res, unsigned int ncomp) 1506 { 1507 struct nv *nv; 1508 1509 rw_rlock(&hio_remote_lock[ncomp]); 1510 1511 if (!ISCONNECTED(res, ncomp)) { 1512 rw_unlock(&hio_remote_lock[ncomp]); 1513 return; 1514 } 1515 1516 PJDLOG_ASSERT(res->hr_remotein != NULL); 1517 PJDLOG_ASSERT(res->hr_remoteout != NULL); 1518 1519 nv = nv_alloc(); 1520 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1521 if (nv_error(nv) != 0) { 1522 rw_unlock(&hio_remote_lock[ncomp]); 1523 nv_free(nv); 1524 pjdlog_debug(1, 1525 "keepalive_send: Unable to prepare header to send."); 1526 return; 1527 } 1528 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) == -1) { 1529 rw_unlock(&hio_remote_lock[ncomp]); 1530 pjdlog_common(LOG_DEBUG, 1, errno, 1531 "keepalive_send: Unable to send request"); 1532 nv_free(nv); 1533 remote_close(res, ncomp); 1534 return; 1535 } 1536 1537 rw_unlock(&hio_remote_lock[ncomp]); 1538 nv_free(nv); 1539 pjdlog_debug(2, "keepalive_send: Request sent."); 1540 } 1541 1542 /* 1543 * Thread sends request to secondary node. 1544 */ 1545 static void * 1546 remote_send_thread(void *arg) 1547 { 1548 struct hast_resource *res = arg; 1549 struct g_gate_ctl_io *ggio; 1550 time_t lastcheck, now; 1551 struct hio *hio; 1552 struct nv *nv; 1553 unsigned int ncomp; 1554 bool wakeup; 1555 uint64_t offset, length; 1556 uint8_t cmd; 1557 void *data; 1558 1559 /* Remote component is 1 for now. */ 1560 ncomp = 1; 1561 lastcheck = time(NULL); 1562 1563 for (;;) { 1564 pjdlog_debug(2, "remote_send: Taking request."); 1565 QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE); 1566 if (hio == NULL) { 1567 now = time(NULL); 1568 if (lastcheck + HAST_KEEPALIVE <= now) { 1569 keepalive_send(res, ncomp); 1570 lastcheck = now; 1571 } 1572 continue; 1573 } 1574 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1575 ggio = &hio->hio_ggio; 1576 switch (ggio->gctl_cmd) { 1577 case BIO_READ: 1578 cmd = HIO_READ; 1579 data = NULL; 1580 offset = ggio->gctl_offset; 1581 length = ggio->gctl_length; 1582 break; 1583 case BIO_WRITE: 1584 cmd = HIO_WRITE; 1585 data = ggio->gctl_data; 1586 offset = ggio->gctl_offset; 1587 length = ggio->gctl_length; 1588 break; 1589 case BIO_DELETE: 1590 cmd = HIO_DELETE; 1591 data = NULL; 1592 offset = ggio->gctl_offset; 1593 length = ggio->gctl_length; 1594 break; 1595 case BIO_FLUSH: 1596 cmd = HIO_FLUSH; 1597 data = NULL; 1598 offset = 0; 1599 length = 0; 1600 break; 1601 default: 1602 PJDLOG_ABORT("invalid condition"); 1603 } 1604 nv = nv_alloc(); 1605 nv_add_uint8(nv, cmd, "cmd"); 1606 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1607 nv_add_uint64(nv, offset, "offset"); 1608 nv_add_uint64(nv, length, "length"); 1609 if (ISMEMSYNCWRITE(hio)) 1610 nv_add_uint8(nv, 1, "memsync"); 1611 if (nv_error(nv) != 0) { 1612 hio->hio_errors[ncomp] = nv_error(nv); 1613 pjdlog_debug(2, 1614 "remote_send: (%p) Unable to prepare header to send.", 1615 hio); 1616 reqlog(LOG_ERR, 0, ggio, 1617 "Unable to prepare header to send (%s): ", 1618 strerror(nv_error(nv))); 1619 /* Move failed request immediately to the done queue. */ 1620 goto done_queue; 1621 } 1622 /* 1623 * Protect connection from disappearing. 1624 */ 1625 rw_rlock(&hio_remote_lock[ncomp]); 1626 if (!ISCONNECTED(res, ncomp)) { 1627 rw_unlock(&hio_remote_lock[ncomp]); 1628 hio->hio_errors[ncomp] = ENOTCONN; 1629 goto done_queue; 1630 } 1631 /* 1632 * Move the request to recv queue before sending it, because 1633 * in different order we can get reply before we move request 1634 * to recv queue. 1635 */ 1636 pjdlog_debug(2, 1637 "remote_send: (%p) Moving request to the recv queue.", 1638 hio); 1639 mtx_lock(&hio_recv_list_lock[ncomp]); 1640 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1641 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1642 hio_recv_list_size[ncomp]++; 1643 mtx_unlock(&hio_recv_list_lock[ncomp]); 1644 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1645 data != NULL ? length : 0) == -1) { 1646 hio->hio_errors[ncomp] = errno; 1647 rw_unlock(&hio_remote_lock[ncomp]); 1648 pjdlog_debug(2, 1649 "remote_send: (%p) Unable to send request.", hio); 1650 reqlog(LOG_ERR, 0, ggio, 1651 "Unable to send request (%s): ", 1652 strerror(hio->hio_errors[ncomp])); 1653 remote_close(res, ncomp); 1654 } else { 1655 rw_unlock(&hio_remote_lock[ncomp]); 1656 } 1657 nv_free(nv); 1658 if (wakeup) 1659 cv_signal(&hio_recv_list_cond[ncomp]); 1660 continue; 1661 done_queue: 1662 nv_free(nv); 1663 if (ISSYNCREQ(hio)) { 1664 if (refcnt_release(&hio->hio_countdown) > 0) 1665 continue; 1666 mtx_lock(&sync_lock); 1667 SYNCREQDONE(hio); 1668 mtx_unlock(&sync_lock); 1669 cv_signal(&sync_cond); 1670 continue; 1671 } 1672 if (ggio->gctl_cmd == BIO_WRITE) { 1673 mtx_lock(&res->hr_amp_lock); 1674 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1675 ggio->gctl_length)) { 1676 (void)hast_activemap_flush(res); 1677 } else { 1678 mtx_unlock(&res->hr_amp_lock); 1679 } 1680 if (ISMEMSYNCWRITE(hio)) { 1681 if (refcnt_release(&hio->hio_writecount) == 0) { 1682 if (hio->hio_errors[0] == 0) 1683 write_complete(res, hio); 1684 } 1685 } 1686 } 1687 if (refcnt_release(&hio->hio_countdown) > 0) 1688 continue; 1689 pjdlog_debug(2, 1690 "remote_send: (%p) Moving request to the done queue.", 1691 hio); 1692 QUEUE_INSERT2(hio, done); 1693 } 1694 /* NOTREACHED */ 1695 return (NULL); 1696 } 1697 1698 /* 1699 * Thread receives answer from secondary node and passes it to ggate_send 1700 * thread. 1701 */ 1702 static void * 1703 remote_recv_thread(void *arg) 1704 { 1705 struct hast_resource *res = arg; 1706 struct g_gate_ctl_io *ggio; 1707 struct hio *hio; 1708 struct nv *nv; 1709 unsigned int ncomp; 1710 uint64_t seq; 1711 bool memsyncack; 1712 int error; 1713 1714 /* Remote component is 1 for now. */ 1715 ncomp = 1; 1716 1717 for (;;) { 1718 /* Wait until there is anything to receive. */ 1719 mtx_lock(&hio_recv_list_lock[ncomp]); 1720 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1721 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1722 cv_wait(&hio_recv_list_cond[ncomp], 1723 &hio_recv_list_lock[ncomp]); 1724 } 1725 mtx_unlock(&hio_recv_list_lock[ncomp]); 1726 1727 memsyncack = false; 1728 1729 rw_rlock(&hio_remote_lock[ncomp]); 1730 if (!ISCONNECTED(res, ncomp)) { 1731 rw_unlock(&hio_remote_lock[ncomp]); 1732 /* 1733 * Connection is dead, so move all pending requests to 1734 * the done queue (one-by-one). 1735 */ 1736 mtx_lock(&hio_recv_list_lock[ncomp]); 1737 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1738 PJDLOG_ASSERT(hio != NULL); 1739 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1740 hio_next[ncomp]); 1741 hio_recv_list_size[ncomp]--; 1742 mtx_unlock(&hio_recv_list_lock[ncomp]); 1743 hio->hio_errors[ncomp] = ENOTCONN; 1744 goto done_queue; 1745 } 1746 if (hast_proto_recv_hdr(res->hr_remotein, &nv) == -1) { 1747 pjdlog_errno(LOG_ERR, 1748 "Unable to receive reply header"); 1749 rw_unlock(&hio_remote_lock[ncomp]); 1750 remote_close(res, ncomp); 1751 continue; 1752 } 1753 rw_unlock(&hio_remote_lock[ncomp]); 1754 seq = nv_get_uint64(nv, "seq"); 1755 if (seq == 0) { 1756 pjdlog_error("Header contains no 'seq' field."); 1757 nv_free(nv); 1758 continue; 1759 } 1760 memsyncack = nv_exists(nv, "received"); 1761 mtx_lock(&hio_recv_list_lock[ncomp]); 1762 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1763 if (hio->hio_ggio.gctl_seq == seq) { 1764 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1765 hio_next[ncomp]); 1766 hio_recv_list_size[ncomp]--; 1767 break; 1768 } 1769 } 1770 mtx_unlock(&hio_recv_list_lock[ncomp]); 1771 if (hio == NULL) { 1772 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1773 (uintmax_t)seq); 1774 nv_free(nv); 1775 continue; 1776 } 1777 ggio = &hio->hio_ggio; 1778 error = nv_get_int16(nv, "error"); 1779 if (error != 0) { 1780 /* Request failed on remote side. */ 1781 hio->hio_errors[ncomp] = error; 1782 reqlog(LOG_WARNING, 0, ggio, 1783 "Remote request failed (%s): ", strerror(error)); 1784 nv_free(nv); 1785 goto done_queue; 1786 } 1787 switch (ggio->gctl_cmd) { 1788 case BIO_READ: 1789 rw_rlock(&hio_remote_lock[ncomp]); 1790 if (!ISCONNECTED(res, ncomp)) { 1791 rw_unlock(&hio_remote_lock[ncomp]); 1792 nv_free(nv); 1793 goto done_queue; 1794 } 1795 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1796 ggio->gctl_data, ggio->gctl_length) == -1) { 1797 hio->hio_errors[ncomp] = errno; 1798 pjdlog_errno(LOG_ERR, 1799 "Unable to receive reply data"); 1800 rw_unlock(&hio_remote_lock[ncomp]); 1801 nv_free(nv); 1802 remote_close(res, ncomp); 1803 goto done_queue; 1804 } 1805 rw_unlock(&hio_remote_lock[ncomp]); 1806 break; 1807 case BIO_WRITE: 1808 case BIO_DELETE: 1809 case BIO_FLUSH: 1810 break; 1811 default: 1812 PJDLOG_ABORT("invalid condition"); 1813 } 1814 hio->hio_errors[ncomp] = 0; 1815 nv_free(nv); 1816 done_queue: 1817 if (ISMEMSYNCWRITE(hio)) { 1818 if (!hio->hio_memsyncacked) { 1819 PJDLOG_ASSERT(memsyncack || 1820 hio->hio_errors[ncomp] != 0); 1821 /* Remote ack arrived. */ 1822 if (refcnt_release(&hio->hio_writecount) == 0) { 1823 if (hio->hio_errors[0] == 0) 1824 write_complete(res, hio); 1825 } 1826 hio->hio_memsyncacked = true; 1827 if (hio->hio_errors[ncomp] == 0) { 1828 pjdlog_debug(2, 1829 "remote_recv: (%p) Moving request " 1830 "back to the recv queue.", hio); 1831 mtx_lock(&hio_recv_list_lock[ncomp]); 1832 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], 1833 hio, hio_next[ncomp]); 1834 hio_recv_list_size[ncomp]++; 1835 mtx_unlock(&hio_recv_list_lock[ncomp]); 1836 continue; 1837 } 1838 } else { 1839 PJDLOG_ASSERT(!memsyncack); 1840 /* Remote final reply arrived. */ 1841 } 1842 } 1843 if (refcnt_release(&hio->hio_countdown) > 0) 1844 continue; 1845 if (ISSYNCREQ(hio)) { 1846 mtx_lock(&sync_lock); 1847 SYNCREQDONE(hio); 1848 mtx_unlock(&sync_lock); 1849 cv_signal(&sync_cond); 1850 } else { 1851 pjdlog_debug(2, 1852 "remote_recv: (%p) Moving request to the done queue.", 1853 hio); 1854 QUEUE_INSERT2(hio, done); 1855 } 1856 } 1857 /* NOTREACHED */ 1858 return (NULL); 1859 } 1860 1861 /* 1862 * Thread sends answer to the kernel. 1863 */ 1864 static void * 1865 ggate_send_thread(void *arg) 1866 { 1867 struct hast_resource *res = arg; 1868 struct g_gate_ctl_io *ggio; 1869 struct hio *hio; 1870 unsigned int ii, ncomps; 1871 1872 ncomps = HAST_NCOMPONENTS; 1873 1874 for (;;) { 1875 pjdlog_debug(2, "ggate_send: Taking request."); 1876 QUEUE_TAKE2(hio, done); 1877 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1878 ggio = &hio->hio_ggio; 1879 for (ii = 0; ii < ncomps; ii++) { 1880 if (hio->hio_errors[ii] == 0) { 1881 /* 1882 * One successful request is enough to declare 1883 * success. 1884 */ 1885 ggio->gctl_error = 0; 1886 break; 1887 } 1888 } 1889 if (ii == ncomps) { 1890 /* 1891 * None of the requests were successful. 1892 * Use the error from local component except the 1893 * case when we did only remote request. 1894 */ 1895 if (ggio->gctl_cmd == BIO_READ && 1896 res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) 1897 ggio->gctl_error = hio->hio_errors[1]; 1898 else 1899 ggio->gctl_error = hio->hio_errors[0]; 1900 } 1901 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1902 mtx_lock(&res->hr_amp_lock); 1903 if (activemap_write_complete(res->hr_amp, 1904 ggio->gctl_offset, ggio->gctl_length)) { 1905 res->hr_stat_activemap_update++; 1906 (void)hast_activemap_flush(res); 1907 } else { 1908 mtx_unlock(&res->hr_amp_lock); 1909 } 1910 } 1911 if (ggio->gctl_cmd == BIO_WRITE) { 1912 /* 1913 * Unlock range we locked. 1914 */ 1915 mtx_lock(&range_lock); 1916 rangelock_del(range_regular, ggio->gctl_offset, 1917 ggio->gctl_length); 1918 if (range_sync_wait) 1919 cv_signal(&range_sync_cond); 1920 mtx_unlock(&range_lock); 1921 if (!hio->hio_done) 1922 write_complete(res, hio); 1923 } else { 1924 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1) { 1925 primary_exit(EX_OSERR, 1926 "G_GATE_CMD_DONE failed"); 1927 } 1928 } 1929 if (hio->hio_errors[0]) { 1930 switch (ggio->gctl_cmd) { 1931 case BIO_READ: 1932 res->hr_stat_read_error++; 1933 break; 1934 case BIO_WRITE: 1935 res->hr_stat_write_error++; 1936 break; 1937 case BIO_DELETE: 1938 res->hr_stat_delete_error++; 1939 break; 1940 case BIO_FLUSH: 1941 res->hr_stat_flush_error++; 1942 break; 1943 } 1944 } 1945 pjdlog_debug(2, 1946 "ggate_send: (%p) Moving request to the free queue.", hio); 1947 QUEUE_INSERT2(hio, free); 1948 } 1949 /* NOTREACHED */ 1950 return (NULL); 1951 } 1952 1953 /* 1954 * Thread synchronize local and remote components. 1955 */ 1956 static void * 1957 sync_thread(void *arg __unused) 1958 { 1959 struct hast_resource *res = arg; 1960 struct hio *hio; 1961 struct g_gate_ctl_io *ggio; 1962 struct timeval tstart, tend, tdiff; 1963 unsigned int ii, ncomp, ncomps; 1964 off_t offset, length, synced; 1965 bool dorewind, directreads; 1966 int syncext; 1967 1968 ncomps = HAST_NCOMPONENTS; 1969 dorewind = true; 1970 synced = 0; 1971 offset = -1; 1972 directreads = false; 1973 1974 for (;;) { 1975 mtx_lock(&sync_lock); 1976 if (offset >= 0 && !sync_inprogress) { 1977 gettimeofday(&tend, NULL); 1978 timersub(&tend, &tstart, &tdiff); 1979 pjdlog_info("Synchronization interrupted after %#.0T. " 1980 "%NB synchronized so far.", &tdiff, 1981 (intmax_t)synced); 1982 event_send(res, EVENT_SYNCINTR); 1983 } 1984 while (!sync_inprogress) { 1985 dorewind = true; 1986 synced = 0; 1987 cv_wait(&sync_cond, &sync_lock); 1988 } 1989 mtx_unlock(&sync_lock); 1990 /* 1991 * Obtain offset at which we should synchronize. 1992 * Rewind synchronization if needed. 1993 */ 1994 mtx_lock(&res->hr_amp_lock); 1995 if (dorewind) 1996 activemap_sync_rewind(res->hr_amp); 1997 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1998 if (syncext != -1) { 1999 /* 2000 * We synchronized entire syncext extent, we can mark 2001 * it as clean now. 2002 */ 2003 if (activemap_extent_complete(res->hr_amp, syncext)) 2004 (void)hast_activemap_flush(res); 2005 else 2006 mtx_unlock(&res->hr_amp_lock); 2007 } else { 2008 mtx_unlock(&res->hr_amp_lock); 2009 } 2010 if (dorewind) { 2011 dorewind = false; 2012 if (offset == -1) 2013 pjdlog_info("Nodes are in sync."); 2014 else { 2015 pjdlog_info("Synchronization started. %NB to go.", 2016 (intmax_t)(res->hr_extentsize * 2017 activemap_ndirty(res->hr_amp))); 2018 event_send(res, EVENT_SYNCSTART); 2019 gettimeofday(&tstart, NULL); 2020 } 2021 } 2022 if (offset == -1) { 2023 sync_stop(); 2024 pjdlog_debug(1, "Nothing to synchronize."); 2025 /* 2026 * Synchronization complete, make both localcnt and 2027 * remotecnt equal. 2028 */ 2029 ncomp = 1; 2030 rw_rlock(&hio_remote_lock[ncomp]); 2031 if (ISCONNECTED(res, ncomp)) { 2032 if (synced > 0) { 2033 int64_t bps; 2034 2035 gettimeofday(&tend, NULL); 2036 timersub(&tend, &tstart, &tdiff); 2037 bps = (int64_t)((double)synced / 2038 ((double)tdiff.tv_sec + 2039 (double)tdiff.tv_usec / 1000000)); 2040 pjdlog_info("Synchronization complete. " 2041 "%NB synchronized in %#.0lT (%NB/sec).", 2042 (intmax_t)synced, &tdiff, 2043 (intmax_t)bps); 2044 event_send(res, EVENT_SYNCDONE); 2045 } 2046 mtx_lock(&metadata_lock); 2047 if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) 2048 directreads = true; 2049 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 2050 res->hr_primary_localcnt = 2051 res->hr_secondary_remotecnt; 2052 res->hr_primary_remotecnt = 2053 res->hr_secondary_localcnt; 2054 pjdlog_debug(1, 2055 "Setting localcnt to %ju and remotecnt to %ju.", 2056 (uintmax_t)res->hr_primary_localcnt, 2057 (uintmax_t)res->hr_primary_remotecnt); 2058 (void)metadata_write(res); 2059 mtx_unlock(&metadata_lock); 2060 } 2061 rw_unlock(&hio_remote_lock[ncomp]); 2062 if (directreads) { 2063 directreads = false; 2064 enable_direct_reads(res); 2065 } 2066 continue; 2067 } 2068 pjdlog_debug(2, "sync: Taking free request."); 2069 QUEUE_TAKE2(hio, free); 2070 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 2071 /* 2072 * Lock the range we are going to synchronize. We don't want 2073 * race where someone writes between our read and write. 2074 */ 2075 for (;;) { 2076 mtx_lock(&range_lock); 2077 if (rangelock_islocked(range_regular, offset, length)) { 2078 pjdlog_debug(2, 2079 "sync: Range offset=%jd length=%jd locked.", 2080 (intmax_t)offset, (intmax_t)length); 2081 range_sync_wait = true; 2082 cv_wait(&range_sync_cond, &range_lock); 2083 range_sync_wait = false; 2084 mtx_unlock(&range_lock); 2085 continue; 2086 } 2087 if (rangelock_add(range_sync, offset, length) == -1) { 2088 mtx_unlock(&range_lock); 2089 pjdlog_debug(2, 2090 "sync: Range offset=%jd length=%jd is already locked, waiting.", 2091 (intmax_t)offset, (intmax_t)length); 2092 sleep(1); 2093 continue; 2094 } 2095 mtx_unlock(&range_lock); 2096 break; 2097 } 2098 /* 2099 * First read the data from synchronization source. 2100 */ 2101 SYNCREQ(hio); 2102 ggio = &hio->hio_ggio; 2103 ggio->gctl_cmd = BIO_READ; 2104 ggio->gctl_offset = offset; 2105 ggio->gctl_length = length; 2106 ggio->gctl_error = 0; 2107 hio->hio_done = false; 2108 hio->hio_replication = res->hr_replication; 2109 for (ii = 0; ii < ncomps; ii++) 2110 hio->hio_errors[ii] = EINVAL; 2111 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 2112 hio); 2113 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 2114 hio); 2115 mtx_lock(&metadata_lock); 2116 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 2117 /* 2118 * This range is up-to-date on local component, 2119 * so handle request locally. 2120 */ 2121 /* Local component is 0 for now. */ 2122 ncomp = 0; 2123 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 2124 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 2125 /* 2126 * This range is out-of-date on local component, 2127 * so send request to the remote node. 2128 */ 2129 /* Remote component is 1 for now. */ 2130 ncomp = 1; 2131 } 2132 mtx_unlock(&metadata_lock); 2133 refcnt_init(&hio->hio_countdown, 1); 2134 QUEUE_INSERT1(hio, send, ncomp); 2135 2136 /* 2137 * Let's wait for READ to finish. 2138 */ 2139 mtx_lock(&sync_lock); 2140 while (!ISSYNCREQDONE(hio)) 2141 cv_wait(&sync_cond, &sync_lock); 2142 mtx_unlock(&sync_lock); 2143 2144 if (hio->hio_errors[ncomp] != 0) { 2145 pjdlog_error("Unable to read synchronization data: %s.", 2146 strerror(hio->hio_errors[ncomp])); 2147 goto free_queue; 2148 } 2149 2150 /* 2151 * We read the data from synchronization source, now write it 2152 * to synchronization target. 2153 */ 2154 SYNCREQ(hio); 2155 ggio->gctl_cmd = BIO_WRITE; 2156 for (ii = 0; ii < ncomps; ii++) 2157 hio->hio_errors[ii] = EINVAL; 2158 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 2159 hio); 2160 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 2161 hio); 2162 mtx_lock(&metadata_lock); 2163 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 2164 /* 2165 * This range is up-to-date on local component, 2166 * so we update remote component. 2167 */ 2168 /* Remote component is 1 for now. */ 2169 ncomp = 1; 2170 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 2171 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 2172 /* 2173 * This range is out-of-date on local component, 2174 * so we update it. 2175 */ 2176 /* Local component is 0 for now. */ 2177 ncomp = 0; 2178 } 2179 mtx_unlock(&metadata_lock); 2180 2181 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 2182 hio); 2183 refcnt_init(&hio->hio_countdown, 1); 2184 QUEUE_INSERT1(hio, send, ncomp); 2185 2186 /* 2187 * Let's wait for WRITE to finish. 2188 */ 2189 mtx_lock(&sync_lock); 2190 while (!ISSYNCREQDONE(hio)) 2191 cv_wait(&sync_cond, &sync_lock); 2192 mtx_unlock(&sync_lock); 2193 2194 if (hio->hio_errors[ncomp] != 0) { 2195 pjdlog_error("Unable to write synchronization data: %s.", 2196 strerror(hio->hio_errors[ncomp])); 2197 goto free_queue; 2198 } 2199 2200 synced += length; 2201 free_queue: 2202 mtx_lock(&range_lock); 2203 rangelock_del(range_sync, offset, length); 2204 if (range_regular_wait) 2205 cv_signal(&range_regular_cond); 2206 mtx_unlock(&range_lock); 2207 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 2208 hio); 2209 QUEUE_INSERT2(hio, free); 2210 } 2211 /* NOTREACHED */ 2212 return (NULL); 2213 } 2214 2215 void 2216 primary_config_reload(struct hast_resource *res, struct nv *nv) 2217 { 2218 unsigned int ii, ncomps; 2219 int modified, vint; 2220 const char *vstr; 2221 2222 pjdlog_info("Reloading configuration..."); 2223 2224 PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY); 2225 PJDLOG_ASSERT(gres == res); 2226 nv_assert(nv, "remoteaddr"); 2227 nv_assert(nv, "sourceaddr"); 2228 nv_assert(nv, "replication"); 2229 nv_assert(nv, "checksum"); 2230 nv_assert(nv, "compression"); 2231 nv_assert(nv, "timeout"); 2232 nv_assert(nv, "exec"); 2233 nv_assert(nv, "metaflush"); 2234 2235 ncomps = HAST_NCOMPONENTS; 2236 2237 #define MODIFIED_REMOTEADDR 0x01 2238 #define MODIFIED_SOURCEADDR 0x02 2239 #define MODIFIED_REPLICATION 0x04 2240 #define MODIFIED_CHECKSUM 0x08 2241 #define MODIFIED_COMPRESSION 0x10 2242 #define MODIFIED_TIMEOUT 0x20 2243 #define MODIFIED_EXEC 0x40 2244 #define MODIFIED_METAFLUSH 0x80 2245 modified = 0; 2246 2247 vstr = nv_get_string(nv, "remoteaddr"); 2248 if (strcmp(gres->hr_remoteaddr, vstr) != 0) { 2249 /* 2250 * Don't copy res->hr_remoteaddr to gres just yet. 2251 * We want remote_close() to log disconnect from the old 2252 * addresses, not from the new ones. 2253 */ 2254 modified |= MODIFIED_REMOTEADDR; 2255 } 2256 vstr = nv_get_string(nv, "sourceaddr"); 2257 if (strcmp(gres->hr_sourceaddr, vstr) != 0) { 2258 strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr)); 2259 modified |= MODIFIED_SOURCEADDR; 2260 } 2261 vint = nv_get_int32(nv, "replication"); 2262 if (gres->hr_replication != vint) { 2263 gres->hr_replication = vint; 2264 modified |= MODIFIED_REPLICATION; 2265 } 2266 vint = nv_get_int32(nv, "checksum"); 2267 if (gres->hr_checksum != vint) { 2268 gres->hr_checksum = vint; 2269 modified |= MODIFIED_CHECKSUM; 2270 } 2271 vint = nv_get_int32(nv, "compression"); 2272 if (gres->hr_compression != vint) { 2273 gres->hr_compression = vint; 2274 modified |= MODIFIED_COMPRESSION; 2275 } 2276 vint = nv_get_int32(nv, "timeout"); 2277 if (gres->hr_timeout != vint) { 2278 gres->hr_timeout = vint; 2279 modified |= MODIFIED_TIMEOUT; 2280 } 2281 vstr = nv_get_string(nv, "exec"); 2282 if (strcmp(gres->hr_exec, vstr) != 0) { 2283 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec)); 2284 modified |= MODIFIED_EXEC; 2285 } 2286 vint = nv_get_int32(nv, "metaflush"); 2287 if (gres->hr_metaflush != vint) { 2288 gres->hr_metaflush = vint; 2289 modified |= MODIFIED_METAFLUSH; 2290 } 2291 2292 /* 2293 * Change timeout for connected sockets. 2294 * Don't bother if we need to reconnect. 2295 */ 2296 if ((modified & MODIFIED_TIMEOUT) != 0 && 2297 (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) == 0) { 2298 for (ii = 0; ii < ncomps; ii++) { 2299 if (!ISREMOTE(ii)) 2300 continue; 2301 rw_rlock(&hio_remote_lock[ii]); 2302 if (!ISCONNECTED(gres, ii)) { 2303 rw_unlock(&hio_remote_lock[ii]); 2304 continue; 2305 } 2306 rw_unlock(&hio_remote_lock[ii]); 2307 if (proto_timeout(gres->hr_remotein, 2308 gres->hr_timeout) == -1) { 2309 pjdlog_errno(LOG_WARNING, 2310 "Unable to set connection timeout"); 2311 } 2312 if (proto_timeout(gres->hr_remoteout, 2313 gres->hr_timeout) == -1) { 2314 pjdlog_errno(LOG_WARNING, 2315 "Unable to set connection timeout"); 2316 } 2317 } 2318 } 2319 if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) != 0) { 2320 for (ii = 0; ii < ncomps; ii++) { 2321 if (!ISREMOTE(ii)) 2322 continue; 2323 remote_close(gres, ii); 2324 } 2325 if (modified & MODIFIED_REMOTEADDR) { 2326 vstr = nv_get_string(nv, "remoteaddr"); 2327 strlcpy(gres->hr_remoteaddr, vstr, 2328 sizeof(gres->hr_remoteaddr)); 2329 } 2330 } 2331 #undef MODIFIED_REMOTEADDR 2332 #undef MODIFIED_SOURCEADDR 2333 #undef MODIFIED_REPLICATION 2334 #undef MODIFIED_CHECKSUM 2335 #undef MODIFIED_COMPRESSION 2336 #undef MODIFIED_TIMEOUT 2337 #undef MODIFIED_EXEC 2338 #undef MODIFIED_METAFLUSH 2339 2340 pjdlog_info("Configuration reloaded successfully."); 2341 } 2342 2343 static void 2344 guard_one(struct hast_resource *res, unsigned int ncomp) 2345 { 2346 struct proto_conn *in, *out; 2347 2348 if (!ISREMOTE(ncomp)) 2349 return; 2350 2351 rw_rlock(&hio_remote_lock[ncomp]); 2352 2353 if (!real_remote(res)) { 2354 rw_unlock(&hio_remote_lock[ncomp]); 2355 return; 2356 } 2357 2358 if (ISCONNECTED(res, ncomp)) { 2359 PJDLOG_ASSERT(res->hr_remotein != NULL); 2360 PJDLOG_ASSERT(res->hr_remoteout != NULL); 2361 rw_unlock(&hio_remote_lock[ncomp]); 2362 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 2363 res->hr_remoteaddr); 2364 return; 2365 } 2366 2367 PJDLOG_ASSERT(res->hr_remotein == NULL); 2368 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2369 /* 2370 * Upgrade the lock. It doesn't have to be atomic as no other thread 2371 * can change connection status from disconnected to connected. 2372 */ 2373 rw_unlock(&hio_remote_lock[ncomp]); 2374 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 2375 res->hr_remoteaddr); 2376 in = out = NULL; 2377 if (init_remote(res, &in, &out) == 0) { 2378 rw_wlock(&hio_remote_lock[ncomp]); 2379 PJDLOG_ASSERT(res->hr_remotein == NULL); 2380 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2381 PJDLOG_ASSERT(in != NULL && out != NULL); 2382 res->hr_remotein = in; 2383 res->hr_remoteout = out; 2384 rw_unlock(&hio_remote_lock[ncomp]); 2385 pjdlog_info("Successfully reconnected to %s.", 2386 res->hr_remoteaddr); 2387 sync_start(); 2388 } else { 2389 /* Both connections should be NULL. */ 2390 PJDLOG_ASSERT(res->hr_remotein == NULL); 2391 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2392 PJDLOG_ASSERT(in == NULL && out == NULL); 2393 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 2394 res->hr_remoteaddr); 2395 } 2396 } 2397 2398 /* 2399 * Thread guards remote connections and reconnects when needed, handles 2400 * signals, etc. 2401 */ 2402 static void * 2403 guard_thread(void *arg) 2404 { 2405 struct hast_resource *res = arg; 2406 unsigned int ii, ncomps; 2407 struct timespec timeout; 2408 time_t lastcheck, now; 2409 sigset_t mask; 2410 int signo; 2411 2412 ncomps = HAST_NCOMPONENTS; 2413 lastcheck = time(NULL); 2414 2415 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 2416 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 2417 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 2418 2419 timeout.tv_sec = HAST_KEEPALIVE; 2420 timeout.tv_nsec = 0; 2421 signo = -1; 2422 2423 for (;;) { 2424 switch (signo) { 2425 case SIGINT: 2426 case SIGTERM: 2427 sigexit_received = true; 2428 primary_exitx(EX_OK, 2429 "Termination signal received, exiting."); 2430 break; 2431 default: 2432 break; 2433 } 2434 2435 /* 2436 * Don't check connections until we fully started, 2437 * as we may still be looping, waiting for remote node 2438 * to switch from primary to secondary. 2439 */ 2440 if (fullystarted) { 2441 pjdlog_debug(2, "remote_guard: Checking connections."); 2442 now = time(NULL); 2443 if (lastcheck + HAST_KEEPALIVE <= now) { 2444 for (ii = 0; ii < ncomps; ii++) 2445 guard_one(res, ii); 2446 lastcheck = now; 2447 } 2448 } 2449 signo = sigtimedwait(&mask, NULL, &timeout); 2450 } 2451 /* NOTREACHED */ 2452 return (NULL); 2453 } 2454