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