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 = 1; 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, HAST_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 map = NULL; 671 mapsize = nv_get_uint32(nvin, "mapsize"); 672 if (mapsize > 0) { 673 map = malloc(mapsize); 674 if (map == NULL) { 675 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 676 (uintmax_t)mapsize); 677 nv_free(nvin); 678 goto close; 679 } 680 /* 681 * Remote node have some dirty extents on its own, lets 682 * download its activemap. 683 */ 684 if (hast_proto_recv_data(res, out, nvin, map, 685 mapsize) < 0) { 686 pjdlog_errno(LOG_ERR, 687 "Unable to receive remote activemap"); 688 nv_free(nvin); 689 free(map); 690 goto close; 691 } 692 /* 693 * Merge local and remote bitmaps. 694 */ 695 activemap_merge(res->hr_amp, map, mapsize); 696 free(map); 697 /* 698 * Now that we merged bitmaps from both nodes, flush it to the 699 * disk before we start to synchronize. 700 */ 701 (void)hast_activemap_flush(res); 702 } 703 nv_free(nvin); 704 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 705 if (inp != NULL && outp != NULL) { 706 *inp = in; 707 *outp = out; 708 } else { 709 res->hr_remotein = in; 710 res->hr_remoteout = out; 711 } 712 event_send(res, EVENT_CONNECT); 713 return (true); 714 close: 715 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 716 event_send(res, EVENT_SPLITBRAIN); 717 proto_close(out); 718 if (in != NULL) 719 proto_close(in); 720 return (false); 721 } 722 723 static void 724 sync_start(void) 725 { 726 727 mtx_lock(&sync_lock); 728 sync_inprogress = true; 729 mtx_unlock(&sync_lock); 730 cv_signal(&sync_cond); 731 } 732 733 static void 734 sync_stop(void) 735 { 736 737 mtx_lock(&sync_lock); 738 if (sync_inprogress) 739 sync_inprogress = false; 740 mtx_unlock(&sync_lock); 741 } 742 743 static void 744 init_ggate(struct hast_resource *res) 745 { 746 struct g_gate_ctl_create ggiocreate; 747 struct g_gate_ctl_cancel ggiocancel; 748 749 /* 750 * We communicate with ggate via /dev/ggctl. Open it. 751 */ 752 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 753 if (res->hr_ggatefd < 0) 754 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 755 /* 756 * Create provider before trying to connect, as connection failure 757 * is not critical, but may take some time. 758 */ 759 bzero(&ggiocreate, sizeof(ggiocreate)); 760 ggiocreate.gctl_version = G_GATE_VERSION; 761 ggiocreate.gctl_mediasize = res->hr_datasize; 762 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 763 ggiocreate.gctl_flags = 0; 764 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE; 765 ggiocreate.gctl_timeout = 0; 766 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 767 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 768 res->hr_provname); 769 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 770 pjdlog_info("Device hast/%s created.", res->hr_provname); 771 res->hr_ggateunit = ggiocreate.gctl_unit; 772 return; 773 } 774 if (errno != EEXIST) { 775 primary_exit(EX_OSERR, "Unable to create hast/%s device", 776 res->hr_provname); 777 } 778 pjdlog_debug(1, 779 "Device hast/%s already exists, we will try to take it over.", 780 res->hr_provname); 781 /* 782 * If we received EEXIST, we assume that the process who created the 783 * provider died and didn't clean up. In that case we will start from 784 * where he left of. 785 */ 786 bzero(&ggiocancel, sizeof(ggiocancel)); 787 ggiocancel.gctl_version = G_GATE_VERSION; 788 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 789 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 790 res->hr_provname); 791 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 792 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 793 res->hr_ggateunit = ggiocancel.gctl_unit; 794 return; 795 } 796 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 797 res->hr_provname); 798 } 799 800 void 801 hastd_primary(struct hast_resource *res) 802 { 803 pthread_t td; 804 pid_t pid; 805 int error, mode, debuglevel; 806 807 /* 808 * Create communication channel for sending control commands from 809 * parent to child. 810 */ 811 if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) { 812 /* TODO: There's no need for this to be fatal error. */ 813 KEEP_ERRNO((void)pidfile_remove(pfh)); 814 pjdlog_exit(EX_OSERR, 815 "Unable to create control sockets between parent and child"); 816 } 817 /* 818 * Create communication channel for sending events from child to parent. 819 */ 820 if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) { 821 /* TODO: There's no need for this to be fatal error. */ 822 KEEP_ERRNO((void)pidfile_remove(pfh)); 823 pjdlog_exit(EX_OSERR, 824 "Unable to create event sockets between child and parent"); 825 } 826 /* 827 * Create communication channel for sending connection requests from 828 * child to parent. 829 */ 830 if (proto_client(NULL, "socketpair://", &res->hr_conn) < 0) { 831 /* TODO: There's no need for this to be fatal error. */ 832 KEEP_ERRNO((void)pidfile_remove(pfh)); 833 pjdlog_exit(EX_OSERR, 834 "Unable to create connection sockets between child and parent"); 835 } 836 837 pid = fork(); 838 if (pid < 0) { 839 /* TODO: There's no need for this to be fatal error. */ 840 KEEP_ERRNO((void)pidfile_remove(pfh)); 841 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 842 } 843 844 if (pid > 0) { 845 /* This is parent. */ 846 /* Declare that we are receiver. */ 847 proto_recv(res->hr_event, NULL, 0); 848 proto_recv(res->hr_conn, NULL, 0); 849 /* Declare that we are sender. */ 850 proto_send(res->hr_ctrl, NULL, 0); 851 res->hr_workerpid = pid; 852 return; 853 } 854 855 gres = res; 856 mode = pjdlog_mode_get(); 857 debuglevel = pjdlog_debug_get(); 858 859 /* Declare that we are sender. */ 860 proto_send(res->hr_event, NULL, 0); 861 proto_send(res->hr_conn, NULL, 0); 862 /* Declare that we are receiver. */ 863 proto_recv(res->hr_ctrl, NULL, 0); 864 descriptors_cleanup(res); 865 866 descriptors_assert(res, mode); 867 868 pjdlog_init(mode); 869 pjdlog_debug_set(debuglevel); 870 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 871 setproctitle("%s (primary)", res->hr_name); 872 873 init_local(res); 874 init_ggate(res); 875 init_environment(res); 876 877 if (drop_privs() != 0) { 878 cleanup(res); 879 exit(EX_CONFIG); 880 } 881 pjdlog_info("Privileges successfully dropped."); 882 883 /* 884 * Create the guard thread first, so we can handle signals from the 885 * very begining. 886 */ 887 error = pthread_create(&td, NULL, guard_thread, res); 888 PJDLOG_ASSERT(error == 0); 889 /* 890 * Create the control thread before sending any event to the parent, 891 * as we can deadlock when parent sends control request to worker, 892 * but worker has no control thread started yet, so parent waits. 893 * In the meantime worker sends an event to the parent, but parent 894 * is unable to handle the event, because it waits for control 895 * request response. 896 */ 897 error = pthread_create(&td, NULL, ctrl_thread, res); 898 PJDLOG_ASSERT(error == 0); 899 if (real_remote(res) && init_remote(res, NULL, NULL)) 900 sync_start(); 901 error = pthread_create(&td, NULL, ggate_recv_thread, res); 902 PJDLOG_ASSERT(error == 0); 903 error = pthread_create(&td, NULL, local_send_thread, res); 904 PJDLOG_ASSERT(error == 0); 905 error = pthread_create(&td, NULL, remote_send_thread, res); 906 PJDLOG_ASSERT(error == 0); 907 error = pthread_create(&td, NULL, remote_recv_thread, res); 908 PJDLOG_ASSERT(error == 0); 909 error = pthread_create(&td, NULL, ggate_send_thread, res); 910 PJDLOG_ASSERT(error == 0); 911 (void)sync_thread(res); 912 } 913 914 static void 915 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 916 { 917 char msg[1024]; 918 va_list ap; 919 int len; 920 921 va_start(ap, fmt); 922 len = vsnprintf(msg, sizeof(msg), fmt, ap); 923 va_end(ap); 924 if ((size_t)len < sizeof(msg)) { 925 switch (ggio->gctl_cmd) { 926 case BIO_READ: 927 (void)snprintf(msg + len, sizeof(msg) - len, 928 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 929 (uintmax_t)ggio->gctl_length); 930 break; 931 case BIO_DELETE: 932 (void)snprintf(msg + len, sizeof(msg) - len, 933 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 934 (uintmax_t)ggio->gctl_length); 935 break; 936 case BIO_FLUSH: 937 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 938 break; 939 case BIO_WRITE: 940 (void)snprintf(msg + len, sizeof(msg) - len, 941 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 942 (uintmax_t)ggio->gctl_length); 943 break; 944 default: 945 (void)snprintf(msg + len, sizeof(msg) - len, 946 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 947 break; 948 } 949 } 950 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 951 } 952 953 static void 954 remote_close(struct hast_resource *res, int ncomp) 955 { 956 957 rw_wlock(&hio_remote_lock[ncomp]); 958 /* 959 * A race is possible between dropping rlock and acquiring wlock - 960 * another thread can close connection in-between. 961 */ 962 if (!ISCONNECTED(res, ncomp)) { 963 PJDLOG_ASSERT(res->hr_remotein == NULL); 964 PJDLOG_ASSERT(res->hr_remoteout == NULL); 965 rw_unlock(&hio_remote_lock[ncomp]); 966 return; 967 } 968 969 PJDLOG_ASSERT(res->hr_remotein != NULL); 970 PJDLOG_ASSERT(res->hr_remoteout != NULL); 971 972 pjdlog_debug(2, "Closing incoming connection to %s.", 973 res->hr_remoteaddr); 974 proto_close(res->hr_remotein); 975 res->hr_remotein = NULL; 976 pjdlog_debug(2, "Closing outgoing connection to %s.", 977 res->hr_remoteaddr); 978 proto_close(res->hr_remoteout); 979 res->hr_remoteout = NULL; 980 981 rw_unlock(&hio_remote_lock[ncomp]); 982 983 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 984 985 /* 986 * Stop synchronization if in-progress. 987 */ 988 sync_stop(); 989 990 event_send(res, EVENT_DISCONNECT); 991 } 992 993 /* 994 * Thread receives ggate I/O requests from the kernel and passes them to 995 * appropriate threads: 996 * WRITE - always goes to both local_send and remote_send threads 997 * READ (when the block is up-to-date on local component) - 998 * only local_send thread 999 * READ (when the block isn't up-to-date on local component) - 1000 * only remote_send thread 1001 * DELETE - always goes to both local_send and remote_send threads 1002 * FLUSH - always goes to both local_send and remote_send threads 1003 */ 1004 static void * 1005 ggate_recv_thread(void *arg) 1006 { 1007 struct hast_resource *res = arg; 1008 struct g_gate_ctl_io *ggio; 1009 struct hio *hio; 1010 unsigned int ii, ncomp, ncomps; 1011 int error; 1012 1013 ncomps = HAST_NCOMPONENTS; 1014 1015 for (;;) { 1016 pjdlog_debug(2, "ggate_recv: Taking free request."); 1017 QUEUE_TAKE2(hio, free); 1018 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 1019 ggio = &hio->hio_ggio; 1020 ggio->gctl_unit = res->hr_ggateunit; 1021 ggio->gctl_length = MAXPHYS; 1022 ggio->gctl_error = 0; 1023 pjdlog_debug(2, 1024 "ggate_recv: (%p) Waiting for request from the kernel.", 1025 hio); 1026 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 1027 if (sigexit_received) 1028 pthread_exit(NULL); 1029 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 1030 } 1031 error = ggio->gctl_error; 1032 switch (error) { 1033 case 0: 1034 break; 1035 case ECANCELED: 1036 /* Exit gracefully. */ 1037 if (!sigexit_received) { 1038 pjdlog_debug(2, 1039 "ggate_recv: (%p) Received cancel from the kernel.", 1040 hio); 1041 pjdlog_info("Received cancel from the kernel, exiting."); 1042 } 1043 pthread_exit(NULL); 1044 case ENOMEM: 1045 /* 1046 * Buffer too small? Impossible, we allocate MAXPHYS 1047 * bytes - request can't be bigger than that. 1048 */ 1049 /* FALLTHROUGH */ 1050 case ENXIO: 1051 default: 1052 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 1053 strerror(error)); 1054 } 1055 for (ii = 0; ii < ncomps; ii++) 1056 hio->hio_errors[ii] = EINVAL; 1057 reqlog(LOG_DEBUG, 2, ggio, 1058 "ggate_recv: (%p) Request received from the kernel: ", 1059 hio); 1060 /* 1061 * Inform all components about new write request. 1062 * For read request prefer local component unless the given 1063 * range is out-of-date, then use remote component. 1064 */ 1065 switch (ggio->gctl_cmd) { 1066 case BIO_READ: 1067 pjdlog_debug(2, 1068 "ggate_recv: (%p) Moving request to the send queue.", 1069 hio); 1070 refcount_init(&hio->hio_countdown, 1); 1071 mtx_lock(&metadata_lock); 1072 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 1073 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1074 /* 1075 * This range is up-to-date on local component, 1076 * so handle request locally. 1077 */ 1078 /* Local component is 0 for now. */ 1079 ncomp = 0; 1080 } else /* if (res->hr_syncsrc == 1081 HAST_SYNCSRC_SECONDARY) */ { 1082 PJDLOG_ASSERT(res->hr_syncsrc == 1083 HAST_SYNCSRC_SECONDARY); 1084 /* 1085 * This range is out-of-date on local component, 1086 * so send request to the remote node. 1087 */ 1088 /* Remote component is 1 for now. */ 1089 ncomp = 1; 1090 } 1091 mtx_unlock(&metadata_lock); 1092 QUEUE_INSERT1(hio, send, ncomp); 1093 break; 1094 case BIO_WRITE: 1095 if (res->hr_resuid == 0) { 1096 /* This is first write, initialize resuid. */ 1097 (void)init_resuid(res); 1098 } 1099 for (;;) { 1100 mtx_lock(&range_lock); 1101 if (rangelock_islocked(range_sync, 1102 ggio->gctl_offset, ggio->gctl_length)) { 1103 pjdlog_debug(2, 1104 "regular: Range offset=%jd length=%zu locked.", 1105 (intmax_t)ggio->gctl_offset, 1106 (size_t)ggio->gctl_length); 1107 range_regular_wait = true; 1108 cv_wait(&range_regular_cond, &range_lock); 1109 range_regular_wait = false; 1110 mtx_unlock(&range_lock); 1111 continue; 1112 } 1113 if (rangelock_add(range_regular, 1114 ggio->gctl_offset, ggio->gctl_length) < 0) { 1115 mtx_unlock(&range_lock); 1116 pjdlog_debug(2, 1117 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1118 (intmax_t)ggio->gctl_offset, 1119 (size_t)ggio->gctl_length); 1120 sleep(1); 1121 continue; 1122 } 1123 mtx_unlock(&range_lock); 1124 break; 1125 } 1126 mtx_lock(&res->hr_amp_lock); 1127 if (activemap_write_start(res->hr_amp, 1128 ggio->gctl_offset, ggio->gctl_length)) { 1129 (void)hast_activemap_flush(res); 1130 } 1131 mtx_unlock(&res->hr_amp_lock); 1132 /* FALLTHROUGH */ 1133 case BIO_DELETE: 1134 case BIO_FLUSH: 1135 pjdlog_debug(2, 1136 "ggate_recv: (%p) Moving request to the send queues.", 1137 hio); 1138 refcount_init(&hio->hio_countdown, ncomps); 1139 for (ii = 0; ii < ncomps; ii++) 1140 QUEUE_INSERT1(hio, send, ii); 1141 break; 1142 } 1143 } 1144 /* NOTREACHED */ 1145 return (NULL); 1146 } 1147 1148 /* 1149 * Thread reads from or writes to local component. 1150 * If local read fails, it redirects it to remote_send thread. 1151 */ 1152 static void * 1153 local_send_thread(void *arg) 1154 { 1155 struct hast_resource *res = arg; 1156 struct g_gate_ctl_io *ggio; 1157 struct hio *hio; 1158 unsigned int ncomp, rncomp; 1159 ssize_t ret; 1160 1161 /* Local component is 0 for now. */ 1162 ncomp = 0; 1163 /* Remote component is 1 for now. */ 1164 rncomp = 1; 1165 1166 for (;;) { 1167 pjdlog_debug(2, "local_send: Taking request."); 1168 QUEUE_TAKE1(hio, send, ncomp, 0); 1169 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1170 ggio = &hio->hio_ggio; 1171 switch (ggio->gctl_cmd) { 1172 case BIO_READ: 1173 ret = pread(res->hr_localfd, ggio->gctl_data, 1174 ggio->gctl_length, 1175 ggio->gctl_offset + res->hr_localoff); 1176 if (ret == ggio->gctl_length) 1177 hio->hio_errors[ncomp] = 0; 1178 else { 1179 /* 1180 * If READ failed, try to read from remote node. 1181 */ 1182 if (ret < 0) { 1183 reqlog(LOG_WARNING, 0, ggio, 1184 "Local request failed (%s), trying remote node. ", 1185 strerror(errno)); 1186 } else if (ret != ggio->gctl_length) { 1187 reqlog(LOG_WARNING, 0, ggio, 1188 "Local request failed (%zd != %jd), trying remote node. ", 1189 ret, (intmax_t)ggio->gctl_length); 1190 } 1191 QUEUE_INSERT1(hio, send, rncomp); 1192 continue; 1193 } 1194 break; 1195 case BIO_WRITE: 1196 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1197 ggio->gctl_length, 1198 ggio->gctl_offset + res->hr_localoff); 1199 if (ret < 0) { 1200 hio->hio_errors[ncomp] = errno; 1201 reqlog(LOG_WARNING, 0, ggio, 1202 "Local request failed (%s): ", 1203 strerror(errno)); 1204 } else if (ret != ggio->gctl_length) { 1205 hio->hio_errors[ncomp] = EIO; 1206 reqlog(LOG_WARNING, 0, ggio, 1207 "Local request failed (%zd != %jd): ", 1208 ret, (intmax_t)ggio->gctl_length); 1209 } else { 1210 hio->hio_errors[ncomp] = 0; 1211 } 1212 break; 1213 case BIO_DELETE: 1214 ret = g_delete(res->hr_localfd, 1215 ggio->gctl_offset + res->hr_localoff, 1216 ggio->gctl_length); 1217 if (ret < 0) { 1218 hio->hio_errors[ncomp] = errno; 1219 reqlog(LOG_WARNING, 0, ggio, 1220 "Local request failed (%s): ", 1221 strerror(errno)); 1222 } else { 1223 hio->hio_errors[ncomp] = 0; 1224 } 1225 break; 1226 case BIO_FLUSH: 1227 ret = g_flush(res->hr_localfd); 1228 if (ret < 0) { 1229 hio->hio_errors[ncomp] = errno; 1230 reqlog(LOG_WARNING, 0, ggio, 1231 "Local request failed (%s): ", 1232 strerror(errno)); 1233 } else { 1234 hio->hio_errors[ncomp] = 0; 1235 } 1236 break; 1237 } 1238 if (refcount_release(&hio->hio_countdown)) { 1239 if (ISSYNCREQ(hio)) { 1240 mtx_lock(&sync_lock); 1241 SYNCREQDONE(hio); 1242 mtx_unlock(&sync_lock); 1243 cv_signal(&sync_cond); 1244 } else { 1245 pjdlog_debug(2, 1246 "local_send: (%p) Moving request to the done queue.", 1247 hio); 1248 QUEUE_INSERT2(hio, done); 1249 } 1250 } 1251 } 1252 /* NOTREACHED */ 1253 return (NULL); 1254 } 1255 1256 static void 1257 keepalive_send(struct hast_resource *res, unsigned int ncomp) 1258 { 1259 struct nv *nv; 1260 1261 rw_rlock(&hio_remote_lock[ncomp]); 1262 1263 if (!ISCONNECTED(res, ncomp)) { 1264 rw_unlock(&hio_remote_lock[ncomp]); 1265 return; 1266 } 1267 1268 PJDLOG_ASSERT(res->hr_remotein != NULL); 1269 PJDLOG_ASSERT(res->hr_remoteout != NULL); 1270 1271 nv = nv_alloc(); 1272 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1273 if (nv_error(nv) != 0) { 1274 rw_unlock(&hio_remote_lock[ncomp]); 1275 nv_free(nv); 1276 pjdlog_debug(1, 1277 "keepalive_send: Unable to prepare header to send."); 1278 return; 1279 } 1280 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1281 rw_unlock(&hio_remote_lock[ncomp]); 1282 pjdlog_common(LOG_DEBUG, 1, errno, 1283 "keepalive_send: Unable to send request"); 1284 nv_free(nv); 1285 remote_close(res, ncomp); 1286 return; 1287 } 1288 1289 rw_unlock(&hio_remote_lock[ncomp]); 1290 nv_free(nv); 1291 pjdlog_debug(2, "keepalive_send: Request sent."); 1292 } 1293 1294 /* 1295 * Thread sends request to secondary node. 1296 */ 1297 static void * 1298 remote_send_thread(void *arg) 1299 { 1300 struct hast_resource *res = arg; 1301 struct g_gate_ctl_io *ggio; 1302 time_t lastcheck, now; 1303 struct hio *hio; 1304 struct nv *nv; 1305 unsigned int ncomp; 1306 bool wakeup; 1307 uint64_t offset, length; 1308 uint8_t cmd; 1309 void *data; 1310 1311 /* Remote component is 1 for now. */ 1312 ncomp = 1; 1313 lastcheck = time(NULL); 1314 1315 for (;;) { 1316 pjdlog_debug(2, "remote_send: Taking request."); 1317 QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE); 1318 if (hio == NULL) { 1319 now = time(NULL); 1320 if (lastcheck + HAST_KEEPALIVE <= now) { 1321 keepalive_send(res, ncomp); 1322 lastcheck = now; 1323 } 1324 continue; 1325 } 1326 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1327 ggio = &hio->hio_ggio; 1328 switch (ggio->gctl_cmd) { 1329 case BIO_READ: 1330 cmd = HIO_READ; 1331 data = NULL; 1332 offset = ggio->gctl_offset; 1333 length = ggio->gctl_length; 1334 break; 1335 case BIO_WRITE: 1336 cmd = HIO_WRITE; 1337 data = ggio->gctl_data; 1338 offset = ggio->gctl_offset; 1339 length = ggio->gctl_length; 1340 break; 1341 case BIO_DELETE: 1342 cmd = HIO_DELETE; 1343 data = NULL; 1344 offset = ggio->gctl_offset; 1345 length = ggio->gctl_length; 1346 break; 1347 case BIO_FLUSH: 1348 cmd = HIO_FLUSH; 1349 data = NULL; 1350 offset = 0; 1351 length = 0; 1352 break; 1353 default: 1354 PJDLOG_ASSERT(!"invalid condition"); 1355 abort(); 1356 } 1357 nv = nv_alloc(); 1358 nv_add_uint8(nv, cmd, "cmd"); 1359 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1360 nv_add_uint64(nv, offset, "offset"); 1361 nv_add_uint64(nv, length, "length"); 1362 if (nv_error(nv) != 0) { 1363 hio->hio_errors[ncomp] = nv_error(nv); 1364 pjdlog_debug(2, 1365 "remote_send: (%p) Unable to prepare header to send.", 1366 hio); 1367 reqlog(LOG_ERR, 0, ggio, 1368 "Unable to prepare header to send (%s): ", 1369 strerror(nv_error(nv))); 1370 /* Move failed request immediately to the done queue. */ 1371 goto done_queue; 1372 } 1373 pjdlog_debug(2, 1374 "remote_send: (%p) Moving request to the recv queue.", 1375 hio); 1376 /* 1377 * Protect connection from disappearing. 1378 */ 1379 rw_rlock(&hio_remote_lock[ncomp]); 1380 if (!ISCONNECTED(res, ncomp)) { 1381 rw_unlock(&hio_remote_lock[ncomp]); 1382 hio->hio_errors[ncomp] = ENOTCONN; 1383 goto done_queue; 1384 } 1385 /* 1386 * Move the request to recv queue before sending it, because 1387 * in different order we can get reply before we move request 1388 * to recv queue. 1389 */ 1390 mtx_lock(&hio_recv_list_lock[ncomp]); 1391 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1392 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1393 mtx_unlock(&hio_recv_list_lock[ncomp]); 1394 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1395 data != NULL ? length : 0) < 0) { 1396 hio->hio_errors[ncomp] = errno; 1397 rw_unlock(&hio_remote_lock[ncomp]); 1398 pjdlog_debug(2, 1399 "remote_send: (%p) Unable to send request.", hio); 1400 reqlog(LOG_ERR, 0, ggio, 1401 "Unable to send request (%s): ", 1402 strerror(hio->hio_errors[ncomp])); 1403 remote_close(res, ncomp); 1404 /* 1405 * Take request back from the receive queue and move 1406 * it immediately to the done queue. 1407 */ 1408 mtx_lock(&hio_recv_list_lock[ncomp]); 1409 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1410 mtx_unlock(&hio_recv_list_lock[ncomp]); 1411 goto done_queue; 1412 } 1413 rw_unlock(&hio_remote_lock[ncomp]); 1414 nv_free(nv); 1415 if (wakeup) 1416 cv_signal(&hio_recv_list_cond[ncomp]); 1417 continue; 1418 done_queue: 1419 nv_free(nv); 1420 if (ISSYNCREQ(hio)) { 1421 if (!refcount_release(&hio->hio_countdown)) 1422 continue; 1423 mtx_lock(&sync_lock); 1424 SYNCREQDONE(hio); 1425 mtx_unlock(&sync_lock); 1426 cv_signal(&sync_cond); 1427 continue; 1428 } 1429 if (ggio->gctl_cmd == BIO_WRITE) { 1430 mtx_lock(&res->hr_amp_lock); 1431 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1432 ggio->gctl_length)) { 1433 (void)hast_activemap_flush(res); 1434 } 1435 mtx_unlock(&res->hr_amp_lock); 1436 } 1437 if (!refcount_release(&hio->hio_countdown)) 1438 continue; 1439 pjdlog_debug(2, 1440 "remote_send: (%p) Moving request to the done queue.", 1441 hio); 1442 QUEUE_INSERT2(hio, done); 1443 } 1444 /* NOTREACHED */ 1445 return (NULL); 1446 } 1447 1448 /* 1449 * Thread receives answer from secondary node and passes it to ggate_send 1450 * thread. 1451 */ 1452 static void * 1453 remote_recv_thread(void *arg) 1454 { 1455 struct hast_resource *res = arg; 1456 struct g_gate_ctl_io *ggio; 1457 struct hio *hio; 1458 struct nv *nv; 1459 unsigned int ncomp; 1460 uint64_t seq; 1461 int error; 1462 1463 /* Remote component is 1 for now. */ 1464 ncomp = 1; 1465 1466 for (;;) { 1467 /* Wait until there is anything to receive. */ 1468 mtx_lock(&hio_recv_list_lock[ncomp]); 1469 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1470 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1471 cv_wait(&hio_recv_list_cond[ncomp], 1472 &hio_recv_list_lock[ncomp]); 1473 } 1474 mtx_unlock(&hio_recv_list_lock[ncomp]); 1475 rw_rlock(&hio_remote_lock[ncomp]); 1476 if (!ISCONNECTED(res, ncomp)) { 1477 rw_unlock(&hio_remote_lock[ncomp]); 1478 /* 1479 * Connection is dead, so move all pending requests to 1480 * the done queue (one-by-one). 1481 */ 1482 mtx_lock(&hio_recv_list_lock[ncomp]); 1483 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1484 PJDLOG_ASSERT(hio != NULL); 1485 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1486 hio_next[ncomp]); 1487 mtx_unlock(&hio_recv_list_lock[ncomp]); 1488 goto done_queue; 1489 } 1490 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1491 pjdlog_errno(LOG_ERR, 1492 "Unable to receive reply header"); 1493 rw_unlock(&hio_remote_lock[ncomp]); 1494 remote_close(res, ncomp); 1495 continue; 1496 } 1497 rw_unlock(&hio_remote_lock[ncomp]); 1498 seq = nv_get_uint64(nv, "seq"); 1499 if (seq == 0) { 1500 pjdlog_error("Header contains no 'seq' field."); 1501 nv_free(nv); 1502 continue; 1503 } 1504 mtx_lock(&hio_recv_list_lock[ncomp]); 1505 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1506 if (hio->hio_ggio.gctl_seq == seq) { 1507 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1508 hio_next[ncomp]); 1509 break; 1510 } 1511 } 1512 mtx_unlock(&hio_recv_list_lock[ncomp]); 1513 if (hio == NULL) { 1514 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1515 (uintmax_t)seq); 1516 nv_free(nv); 1517 continue; 1518 } 1519 error = nv_get_int16(nv, "error"); 1520 if (error != 0) { 1521 /* Request failed on remote side. */ 1522 hio->hio_errors[ncomp] = error; 1523 reqlog(LOG_WARNING, 0, &hio->hio_ggio, 1524 "Remote request failed (%s): ", strerror(error)); 1525 nv_free(nv); 1526 goto done_queue; 1527 } 1528 ggio = &hio->hio_ggio; 1529 switch (ggio->gctl_cmd) { 1530 case BIO_READ: 1531 rw_rlock(&hio_remote_lock[ncomp]); 1532 if (!ISCONNECTED(res, ncomp)) { 1533 rw_unlock(&hio_remote_lock[ncomp]); 1534 nv_free(nv); 1535 goto done_queue; 1536 } 1537 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1538 ggio->gctl_data, ggio->gctl_length) < 0) { 1539 hio->hio_errors[ncomp] = errno; 1540 pjdlog_errno(LOG_ERR, 1541 "Unable to receive reply data"); 1542 rw_unlock(&hio_remote_lock[ncomp]); 1543 nv_free(nv); 1544 remote_close(res, ncomp); 1545 goto done_queue; 1546 } 1547 rw_unlock(&hio_remote_lock[ncomp]); 1548 break; 1549 case BIO_WRITE: 1550 case BIO_DELETE: 1551 case BIO_FLUSH: 1552 break; 1553 default: 1554 PJDLOG_ASSERT(!"invalid condition"); 1555 abort(); 1556 } 1557 hio->hio_errors[ncomp] = 0; 1558 nv_free(nv); 1559 done_queue: 1560 if (refcount_release(&hio->hio_countdown)) { 1561 if (ISSYNCREQ(hio)) { 1562 mtx_lock(&sync_lock); 1563 SYNCREQDONE(hio); 1564 mtx_unlock(&sync_lock); 1565 cv_signal(&sync_cond); 1566 } else { 1567 pjdlog_debug(2, 1568 "remote_recv: (%p) Moving request to the done queue.", 1569 hio); 1570 QUEUE_INSERT2(hio, done); 1571 } 1572 } 1573 } 1574 /* NOTREACHED */ 1575 return (NULL); 1576 } 1577 1578 /* 1579 * Thread sends answer to the kernel. 1580 */ 1581 static void * 1582 ggate_send_thread(void *arg) 1583 { 1584 struct hast_resource *res = arg; 1585 struct g_gate_ctl_io *ggio; 1586 struct hio *hio; 1587 unsigned int ii, ncomp, ncomps; 1588 1589 ncomps = HAST_NCOMPONENTS; 1590 1591 for (;;) { 1592 pjdlog_debug(2, "ggate_send: Taking request."); 1593 QUEUE_TAKE2(hio, done); 1594 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1595 ggio = &hio->hio_ggio; 1596 for (ii = 0; ii < ncomps; ii++) { 1597 if (hio->hio_errors[ii] == 0) { 1598 /* 1599 * One successful request is enough to declare 1600 * success. 1601 */ 1602 ggio->gctl_error = 0; 1603 break; 1604 } 1605 } 1606 if (ii == ncomps) { 1607 /* 1608 * None of the requests were successful. 1609 * Use first error. 1610 */ 1611 ggio->gctl_error = hio->hio_errors[0]; 1612 } 1613 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1614 mtx_lock(&res->hr_amp_lock); 1615 activemap_write_complete(res->hr_amp, 1616 ggio->gctl_offset, ggio->gctl_length); 1617 mtx_unlock(&res->hr_amp_lock); 1618 } 1619 if (ggio->gctl_cmd == BIO_WRITE) { 1620 /* 1621 * Unlock range we locked. 1622 */ 1623 mtx_lock(&range_lock); 1624 rangelock_del(range_regular, ggio->gctl_offset, 1625 ggio->gctl_length); 1626 if (range_sync_wait) 1627 cv_signal(&range_sync_cond); 1628 mtx_unlock(&range_lock); 1629 /* 1630 * Bump local count if this is first write after 1631 * connection failure with remote node. 1632 */ 1633 ncomp = 1; 1634 rw_rlock(&hio_remote_lock[ncomp]); 1635 if (!ISCONNECTED(res, ncomp)) { 1636 mtx_lock(&metadata_lock); 1637 if (res->hr_primary_localcnt == 1638 res->hr_secondary_remotecnt) { 1639 res->hr_primary_localcnt++; 1640 pjdlog_debug(1, 1641 "Increasing localcnt to %ju.", 1642 (uintmax_t)res->hr_primary_localcnt); 1643 (void)metadata_write(res); 1644 } 1645 mtx_unlock(&metadata_lock); 1646 } 1647 rw_unlock(&hio_remote_lock[ncomp]); 1648 } 1649 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1650 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1651 pjdlog_debug(2, 1652 "ggate_send: (%p) Moving request to the free queue.", hio); 1653 QUEUE_INSERT2(hio, free); 1654 } 1655 /* NOTREACHED */ 1656 return (NULL); 1657 } 1658 1659 /* 1660 * Thread synchronize local and remote components. 1661 */ 1662 static void * 1663 sync_thread(void *arg __unused) 1664 { 1665 struct hast_resource *res = arg; 1666 struct hio *hio; 1667 struct g_gate_ctl_io *ggio; 1668 struct timeval tstart, tend, tdiff; 1669 unsigned int ii, ncomp, ncomps; 1670 off_t offset, length, synced; 1671 bool dorewind; 1672 int syncext; 1673 1674 ncomps = HAST_NCOMPONENTS; 1675 dorewind = true; 1676 synced = 0; 1677 offset = -1; 1678 1679 for (;;) { 1680 mtx_lock(&sync_lock); 1681 if (offset >= 0 && !sync_inprogress) { 1682 gettimeofday(&tend, NULL); 1683 timersub(&tend, &tstart, &tdiff); 1684 pjdlog_info("Synchronization interrupted after %#.0T. " 1685 "%NB synchronized so far.", &tdiff, 1686 (intmax_t)synced); 1687 event_send(res, EVENT_SYNCINTR); 1688 } 1689 while (!sync_inprogress) { 1690 dorewind = true; 1691 synced = 0; 1692 cv_wait(&sync_cond, &sync_lock); 1693 } 1694 mtx_unlock(&sync_lock); 1695 /* 1696 * Obtain offset at which we should synchronize. 1697 * Rewind synchronization if needed. 1698 */ 1699 mtx_lock(&res->hr_amp_lock); 1700 if (dorewind) 1701 activemap_sync_rewind(res->hr_amp); 1702 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1703 if (syncext != -1) { 1704 /* 1705 * We synchronized entire syncext extent, we can mark 1706 * it as clean now. 1707 */ 1708 if (activemap_extent_complete(res->hr_amp, syncext)) 1709 (void)hast_activemap_flush(res); 1710 } 1711 mtx_unlock(&res->hr_amp_lock); 1712 if (dorewind) { 1713 dorewind = false; 1714 if (offset < 0) 1715 pjdlog_info("Nodes are in sync."); 1716 else { 1717 pjdlog_info("Synchronization started. %NB to go.", 1718 (intmax_t)(res->hr_extentsize * 1719 activemap_ndirty(res->hr_amp))); 1720 event_send(res, EVENT_SYNCSTART); 1721 gettimeofday(&tstart, NULL); 1722 } 1723 } 1724 if (offset < 0) { 1725 sync_stop(); 1726 pjdlog_debug(1, "Nothing to synchronize."); 1727 /* 1728 * Synchronization complete, make both localcnt and 1729 * remotecnt equal. 1730 */ 1731 ncomp = 1; 1732 rw_rlock(&hio_remote_lock[ncomp]); 1733 if (ISCONNECTED(res, ncomp)) { 1734 if (synced > 0) { 1735 int64_t bps; 1736 1737 gettimeofday(&tend, NULL); 1738 timersub(&tend, &tstart, &tdiff); 1739 bps = (int64_t)((double)synced / 1740 ((double)tdiff.tv_sec + 1741 (double)tdiff.tv_usec / 1000000)); 1742 pjdlog_info("Synchronization complete. " 1743 "%NB synchronized in %#.0lT (%NB/sec).", 1744 (intmax_t)synced, &tdiff, 1745 (intmax_t)bps); 1746 event_send(res, EVENT_SYNCDONE); 1747 } 1748 mtx_lock(&metadata_lock); 1749 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1750 res->hr_primary_localcnt = 1751 res->hr_secondary_localcnt; 1752 res->hr_primary_remotecnt = 1753 res->hr_secondary_remotecnt; 1754 pjdlog_debug(1, 1755 "Setting localcnt to %ju and remotecnt to %ju.", 1756 (uintmax_t)res->hr_primary_localcnt, 1757 (uintmax_t)res->hr_secondary_localcnt); 1758 (void)metadata_write(res); 1759 mtx_unlock(&metadata_lock); 1760 } 1761 rw_unlock(&hio_remote_lock[ncomp]); 1762 continue; 1763 } 1764 pjdlog_debug(2, "sync: Taking free request."); 1765 QUEUE_TAKE2(hio, free); 1766 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1767 /* 1768 * Lock the range we are going to synchronize. We don't want 1769 * race where someone writes between our read and write. 1770 */ 1771 for (;;) { 1772 mtx_lock(&range_lock); 1773 if (rangelock_islocked(range_regular, offset, length)) { 1774 pjdlog_debug(2, 1775 "sync: Range offset=%jd length=%jd locked.", 1776 (intmax_t)offset, (intmax_t)length); 1777 range_sync_wait = true; 1778 cv_wait(&range_sync_cond, &range_lock); 1779 range_sync_wait = false; 1780 mtx_unlock(&range_lock); 1781 continue; 1782 } 1783 if (rangelock_add(range_sync, offset, length) < 0) { 1784 mtx_unlock(&range_lock); 1785 pjdlog_debug(2, 1786 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1787 (intmax_t)offset, (intmax_t)length); 1788 sleep(1); 1789 continue; 1790 } 1791 mtx_unlock(&range_lock); 1792 break; 1793 } 1794 /* 1795 * First read the data from synchronization source. 1796 */ 1797 SYNCREQ(hio); 1798 ggio = &hio->hio_ggio; 1799 ggio->gctl_cmd = BIO_READ; 1800 ggio->gctl_offset = offset; 1801 ggio->gctl_length = length; 1802 ggio->gctl_error = 0; 1803 for (ii = 0; ii < ncomps; ii++) 1804 hio->hio_errors[ii] = EINVAL; 1805 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1806 hio); 1807 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1808 hio); 1809 mtx_lock(&metadata_lock); 1810 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1811 /* 1812 * This range is up-to-date on local component, 1813 * so handle request locally. 1814 */ 1815 /* Local component is 0 for now. */ 1816 ncomp = 0; 1817 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1818 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1819 /* 1820 * This range is out-of-date on local component, 1821 * so send request to the remote node. 1822 */ 1823 /* Remote component is 1 for now. */ 1824 ncomp = 1; 1825 } 1826 mtx_unlock(&metadata_lock); 1827 refcount_init(&hio->hio_countdown, 1); 1828 QUEUE_INSERT1(hio, send, ncomp); 1829 1830 /* 1831 * Let's wait for READ to finish. 1832 */ 1833 mtx_lock(&sync_lock); 1834 while (!ISSYNCREQDONE(hio)) 1835 cv_wait(&sync_cond, &sync_lock); 1836 mtx_unlock(&sync_lock); 1837 1838 if (hio->hio_errors[ncomp] != 0) { 1839 pjdlog_error("Unable to read synchronization data: %s.", 1840 strerror(hio->hio_errors[ncomp])); 1841 goto free_queue; 1842 } 1843 1844 /* 1845 * We read the data from synchronization source, now write it 1846 * to synchronization target. 1847 */ 1848 SYNCREQ(hio); 1849 ggio->gctl_cmd = BIO_WRITE; 1850 for (ii = 0; ii < ncomps; ii++) 1851 hio->hio_errors[ii] = EINVAL; 1852 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1853 hio); 1854 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1855 hio); 1856 mtx_lock(&metadata_lock); 1857 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1858 /* 1859 * This range is up-to-date on local component, 1860 * so we update remote component. 1861 */ 1862 /* Remote component is 1 for now. */ 1863 ncomp = 1; 1864 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1865 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1866 /* 1867 * This range is out-of-date on local component, 1868 * so we update it. 1869 */ 1870 /* Local component is 0 for now. */ 1871 ncomp = 0; 1872 } 1873 mtx_unlock(&metadata_lock); 1874 1875 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1876 hio); 1877 refcount_init(&hio->hio_countdown, 1); 1878 QUEUE_INSERT1(hio, send, ncomp); 1879 1880 /* 1881 * Let's wait for WRITE to finish. 1882 */ 1883 mtx_lock(&sync_lock); 1884 while (!ISSYNCREQDONE(hio)) 1885 cv_wait(&sync_cond, &sync_lock); 1886 mtx_unlock(&sync_lock); 1887 1888 if (hio->hio_errors[ncomp] != 0) { 1889 pjdlog_error("Unable to write synchronization data: %s.", 1890 strerror(hio->hio_errors[ncomp])); 1891 goto free_queue; 1892 } 1893 1894 synced += length; 1895 free_queue: 1896 mtx_lock(&range_lock); 1897 rangelock_del(range_sync, offset, length); 1898 if (range_regular_wait) 1899 cv_signal(&range_regular_cond); 1900 mtx_unlock(&range_lock); 1901 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1902 hio); 1903 QUEUE_INSERT2(hio, free); 1904 } 1905 /* NOTREACHED */ 1906 return (NULL); 1907 } 1908 1909 void 1910 primary_config_reload(struct hast_resource *res, struct nv *nv) 1911 { 1912 unsigned int ii, ncomps; 1913 int modified, vint; 1914 const char *vstr; 1915 1916 pjdlog_info("Reloading configuration..."); 1917 1918 PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY); 1919 PJDLOG_ASSERT(gres == res); 1920 nv_assert(nv, "remoteaddr"); 1921 nv_assert(nv, "sourceaddr"); 1922 nv_assert(nv, "replication"); 1923 nv_assert(nv, "checksum"); 1924 nv_assert(nv, "compression"); 1925 nv_assert(nv, "timeout"); 1926 nv_assert(nv, "exec"); 1927 1928 ncomps = HAST_NCOMPONENTS; 1929 1930 #define MODIFIED_REMOTEADDR 0x01 1931 #define MODIFIED_SOURCEADDR 0x02 1932 #define MODIFIED_REPLICATION 0x04 1933 #define MODIFIED_CHECKSUM 0x08 1934 #define MODIFIED_COMPRESSION 0x10 1935 #define MODIFIED_TIMEOUT 0x20 1936 #define MODIFIED_EXEC 0x40 1937 modified = 0; 1938 1939 vstr = nv_get_string(nv, "remoteaddr"); 1940 if (strcmp(gres->hr_remoteaddr, vstr) != 0) { 1941 /* 1942 * Don't copy res->hr_remoteaddr to gres just yet. 1943 * We want remote_close() to log disconnect from the old 1944 * addresses, not from the new ones. 1945 */ 1946 modified |= MODIFIED_REMOTEADDR; 1947 } 1948 vstr = nv_get_string(nv, "sourceaddr"); 1949 if (strcmp(gres->hr_sourceaddr, vstr) != 0) { 1950 strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr)); 1951 modified |= MODIFIED_SOURCEADDR; 1952 } 1953 vint = nv_get_int32(nv, "replication"); 1954 if (gres->hr_replication != vint) { 1955 gres->hr_replication = vint; 1956 modified |= MODIFIED_REPLICATION; 1957 } 1958 vint = nv_get_int32(nv, "checksum"); 1959 if (gres->hr_checksum != vint) { 1960 gres->hr_checksum = vint; 1961 modified |= MODIFIED_CHECKSUM; 1962 } 1963 vint = nv_get_int32(nv, "compression"); 1964 if (gres->hr_compression != vint) { 1965 gres->hr_compression = vint; 1966 modified |= MODIFIED_COMPRESSION; 1967 } 1968 vint = nv_get_int32(nv, "timeout"); 1969 if (gres->hr_timeout != vint) { 1970 gres->hr_timeout = vint; 1971 modified |= MODIFIED_TIMEOUT; 1972 } 1973 vstr = nv_get_string(nv, "exec"); 1974 if (strcmp(gres->hr_exec, vstr) != 0) { 1975 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec)); 1976 modified |= MODIFIED_EXEC; 1977 } 1978 1979 /* 1980 * Change timeout for connected sockets. 1981 * Don't bother if we need to reconnect. 1982 */ 1983 if ((modified & MODIFIED_TIMEOUT) != 0 && 1984 (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR | 1985 MODIFIED_REPLICATION)) == 0) { 1986 for (ii = 0; ii < ncomps; ii++) { 1987 if (!ISREMOTE(ii)) 1988 continue; 1989 rw_rlock(&hio_remote_lock[ii]); 1990 if (!ISCONNECTED(gres, ii)) { 1991 rw_unlock(&hio_remote_lock[ii]); 1992 continue; 1993 } 1994 rw_unlock(&hio_remote_lock[ii]); 1995 if (proto_timeout(gres->hr_remotein, 1996 gres->hr_timeout) < 0) { 1997 pjdlog_errno(LOG_WARNING, 1998 "Unable to set connection timeout"); 1999 } 2000 if (proto_timeout(gres->hr_remoteout, 2001 gres->hr_timeout) < 0) { 2002 pjdlog_errno(LOG_WARNING, 2003 "Unable to set connection timeout"); 2004 } 2005 } 2006 } 2007 if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR | 2008 MODIFIED_REPLICATION)) != 0) { 2009 for (ii = 0; ii < ncomps; ii++) { 2010 if (!ISREMOTE(ii)) 2011 continue; 2012 remote_close(gres, ii); 2013 } 2014 if (modified & MODIFIED_REMOTEADDR) { 2015 vstr = nv_get_string(nv, "remoteaddr"); 2016 strlcpy(gres->hr_remoteaddr, vstr, 2017 sizeof(gres->hr_remoteaddr)); 2018 } 2019 } 2020 #undef MODIFIED_REMOTEADDR 2021 #undef MODIFIED_SOURCEADDR 2022 #undef MODIFIED_REPLICATION 2023 #undef MODIFIED_CHECKSUM 2024 #undef MODIFIED_COMPRESSION 2025 #undef MODIFIED_TIMEOUT 2026 #undef MODIFIED_EXEC 2027 2028 pjdlog_info("Configuration reloaded successfully."); 2029 } 2030 2031 static void 2032 guard_one(struct hast_resource *res, unsigned int ncomp) 2033 { 2034 struct proto_conn *in, *out; 2035 2036 if (!ISREMOTE(ncomp)) 2037 return; 2038 2039 rw_rlock(&hio_remote_lock[ncomp]); 2040 2041 if (!real_remote(res)) { 2042 rw_unlock(&hio_remote_lock[ncomp]); 2043 return; 2044 } 2045 2046 if (ISCONNECTED(res, ncomp)) { 2047 PJDLOG_ASSERT(res->hr_remotein != NULL); 2048 PJDLOG_ASSERT(res->hr_remoteout != NULL); 2049 rw_unlock(&hio_remote_lock[ncomp]); 2050 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 2051 res->hr_remoteaddr); 2052 return; 2053 } 2054 2055 PJDLOG_ASSERT(res->hr_remotein == NULL); 2056 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2057 /* 2058 * Upgrade the lock. It doesn't have to be atomic as no other thread 2059 * can change connection status from disconnected to connected. 2060 */ 2061 rw_unlock(&hio_remote_lock[ncomp]); 2062 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 2063 res->hr_remoteaddr); 2064 in = out = NULL; 2065 if (init_remote(res, &in, &out)) { 2066 rw_wlock(&hio_remote_lock[ncomp]); 2067 PJDLOG_ASSERT(res->hr_remotein == NULL); 2068 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2069 PJDLOG_ASSERT(in != NULL && out != NULL); 2070 res->hr_remotein = in; 2071 res->hr_remoteout = out; 2072 rw_unlock(&hio_remote_lock[ncomp]); 2073 pjdlog_info("Successfully reconnected to %s.", 2074 res->hr_remoteaddr); 2075 sync_start(); 2076 } else { 2077 /* Both connections should be NULL. */ 2078 PJDLOG_ASSERT(res->hr_remotein == NULL); 2079 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2080 PJDLOG_ASSERT(in == NULL && out == NULL); 2081 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 2082 res->hr_remoteaddr); 2083 } 2084 } 2085 2086 /* 2087 * Thread guards remote connections and reconnects when needed, handles 2088 * signals, etc. 2089 */ 2090 static void * 2091 guard_thread(void *arg) 2092 { 2093 struct hast_resource *res = arg; 2094 unsigned int ii, ncomps; 2095 struct timespec timeout; 2096 time_t lastcheck, now; 2097 sigset_t mask; 2098 int signo; 2099 2100 ncomps = HAST_NCOMPONENTS; 2101 lastcheck = time(NULL); 2102 2103 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 2104 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 2105 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 2106 2107 timeout.tv_sec = HAST_KEEPALIVE; 2108 timeout.tv_nsec = 0; 2109 signo = -1; 2110 2111 for (;;) { 2112 switch (signo) { 2113 case SIGINT: 2114 case SIGTERM: 2115 sigexit_received = true; 2116 primary_exitx(EX_OK, 2117 "Termination signal received, exiting."); 2118 break; 2119 default: 2120 break; 2121 } 2122 2123 pjdlog_debug(2, "remote_guard: Checking connections."); 2124 now = time(NULL); 2125 if (lastcheck + HAST_KEEPALIVE <= now) { 2126 for (ii = 0; ii < ncomps; ii++) 2127 guard_one(res, ii); 2128 lastcheck = now; 2129 } 2130 signo = sigtimedwait(&mask, NULL, &timeout); 2131 } 2132 /* NOTREACHED */ 2133 return (NULL); 2134 } 2135