primary.c revision 218049
1/*- 2 * Copyright (c) 2009 The FreeBSD Foundation 3 * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org> 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: head/sbin/hastd/primary.c 218049 2011-01-28 22:35:46Z pjd $"); 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 <assert.h> 44#include <err.h> 45#include <errno.h> 46#include <fcntl.h> 47#include <libgeom.h> 48#include <pthread.h> 49#include <signal.h> 50#include <stdint.h> 51#include <stdio.h> 52#include <string.h> 53#include <sysexits.h> 54#include <unistd.h> 55 56#include <activemap.h> 57#include <nv.h> 58#include <rangelock.h> 59 60#include "control.h" 61#include "event.h" 62#include "hast.h" 63#include "hast_proto.h" 64#include "hastd.h" 65#include "hooks.h" 66#include "metadata.h" 67#include "proto.h" 68#include "pjdlog.h" 69#include "subr.h" 70#include "synch.h" 71 72/* The is only one remote component for now. */ 73#define ISREMOTE(no) ((no) == 1) 74 75struct hio { 76 /* 77 * Number of components we are still waiting for. 78 * When this field goes to 0, we can send the request back to the 79 * kernel. Each component has to decrease this counter by one 80 * even on failure. 81 */ 82 unsigned int hio_countdown; 83 /* 84 * Each component has a place to store its own error. 85 * Once the request is handled by all components we can decide if the 86 * request overall is successful or not. 87 */ 88 int *hio_errors; 89 /* 90 * Structure used to comunicate with GEOM Gate class. 91 */ 92 struct g_gate_ctl_io hio_ggio; 93 TAILQ_ENTRY(hio) *hio_next; 94}; 95#define hio_free_next hio_next[0] 96#define hio_done_next hio_next[0] 97 98/* 99 * Free list holds unused structures. When free list is empty, we have to wait 100 * until some in-progress requests are freed. 101 */ 102static TAILQ_HEAD(, hio) hio_free_list; 103static pthread_mutex_t hio_free_list_lock; 104static pthread_cond_t hio_free_list_cond; 105/* 106 * There is one send list for every component. One requests is placed on all 107 * send lists - each component gets the same request, but each component is 108 * responsible for managing his own send list. 109 */ 110static TAILQ_HEAD(, hio) *hio_send_list; 111static pthread_mutex_t *hio_send_list_lock; 112static pthread_cond_t *hio_send_list_cond; 113/* 114 * There is one recv list for every component, although local components don't 115 * use recv lists as local requests are done synchronously. 116 */ 117static TAILQ_HEAD(, hio) *hio_recv_list; 118static pthread_mutex_t *hio_recv_list_lock; 119static pthread_cond_t *hio_recv_list_cond; 120/* 121 * Request is placed on done list by the slowest component (the one that 122 * decreased hio_countdown from 1 to 0). 123 */ 124static TAILQ_HEAD(, hio) hio_done_list; 125static pthread_mutex_t hio_done_list_lock; 126static pthread_cond_t hio_done_list_cond; 127/* 128 * Structure below are for interaction with sync thread. 129 */ 130static bool sync_inprogress; 131static pthread_mutex_t sync_lock; 132static pthread_cond_t sync_cond; 133/* 134 * The lock below allows to synchornize access to remote connections. 135 */ 136static pthread_rwlock_t *hio_remote_lock; 137 138/* 139 * Lock to synchronize metadata updates. Also synchronize access to 140 * hr_primary_localcnt and hr_primary_remotecnt fields. 141 */ 142static pthread_mutex_t metadata_lock; 143 144/* 145 * Maximum number of outstanding I/O requests. 146 */ 147#define HAST_HIO_MAX 256 148/* 149 * Number of components. At this point there are only two components: local 150 * and remote, but in the future it might be possible to use multiple local 151 * and remote components. 152 */ 153#define HAST_NCOMPONENTS 2 154/* 155 * Number of seconds to sleep between reconnect retries or keepalive packets. 156 */ 157#define RETRY_SLEEP 10 158 159#define ISCONNECTED(res, no) \ 160 ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL) 161 162#define QUEUE_INSERT1(hio, name, ncomp) do { \ 163 bool _wakeup; \ 164 \ 165 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 166 _wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]); \ 167 TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio), \ 168 hio_next[(ncomp)]); \ 169 mtx_unlock(&hio_##name##_list_lock[ncomp]); \ 170 if (_wakeup) \ 171 cv_signal(&hio_##name##_list_cond[(ncomp)]); \ 172} while (0) 173#define QUEUE_INSERT2(hio, name) do { \ 174 bool _wakeup; \ 175 \ 176 mtx_lock(&hio_##name##_list_lock); \ 177 _wakeup = TAILQ_EMPTY(&hio_##name##_list); \ 178 TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\ 179 mtx_unlock(&hio_##name##_list_lock); \ 180 if (_wakeup) \ 181 cv_signal(&hio_##name##_list_cond); \ 182} while (0) 183#define QUEUE_TAKE1(hio, name, ncomp, timeout) do { \ 184 bool _last; \ 185 \ 186 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 187 _last = false; \ 188 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \ 189 cv_timedwait(&hio_##name##_list_cond[(ncomp)], \ 190 &hio_##name##_list_lock[(ncomp)], (timeout)); \ 191 if ((timeout) != 0) \ 192 _last = true; \ 193 } \ 194 if (hio != NULL) { \ 195 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \ 196 hio_next[(ncomp)]); \ 197 } \ 198 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 199} while (0) 200#define QUEUE_TAKE2(hio, name) do { \ 201 mtx_lock(&hio_##name##_list_lock); \ 202 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 203 cv_wait(&hio_##name##_list_cond, \ 204 &hio_##name##_list_lock); \ 205 } \ 206 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \ 207 mtx_unlock(&hio_##name##_list_lock); \ 208} while (0) 209 210#define SYNCREQ(hio) do { \ 211 (hio)->hio_ggio.gctl_unit = -1; \ 212 (hio)->hio_ggio.gctl_seq = 1; \ 213} while (0) 214#define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1) 215#define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0) 216#define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2) 217 218static struct hast_resource *gres; 219 220static pthread_mutex_t range_lock; 221static struct rangelocks *range_regular; 222static bool range_regular_wait; 223static pthread_cond_t range_regular_cond; 224static struct rangelocks *range_sync; 225static bool range_sync_wait; 226static pthread_cond_t range_sync_cond; 227 228static void *ggate_recv_thread(void *arg); 229static void *local_send_thread(void *arg); 230static void *remote_send_thread(void *arg); 231static void *remote_recv_thread(void *arg); 232static void *ggate_send_thread(void *arg); 233static void *sync_thread(void *arg); 234static void *guard_thread(void *arg); 235 236static void 237cleanup(struct hast_resource *res) 238{ 239 int rerrno; 240 241 /* Remember errno. */ 242 rerrno = errno; 243 244 /* Destroy ggate provider if we created one. */ 245 if (res->hr_ggateunit >= 0) { 246 struct g_gate_ctl_destroy ggiod; 247 248 bzero(&ggiod, sizeof(ggiod)); 249 ggiod.gctl_version = G_GATE_VERSION; 250 ggiod.gctl_unit = res->hr_ggateunit; 251 ggiod.gctl_force = 1; 252 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) { 253 pjdlog_errno(LOG_WARNING, 254 "Unable to destroy hast/%s device", 255 res->hr_provname); 256 } 257 res->hr_ggateunit = -1; 258 } 259 260 /* Restore errno. */ 261 errno = rerrno; 262} 263 264static __dead2 void 265primary_exit(int exitcode, const char *fmt, ...) 266{ 267 va_list ap; 268 269 assert(exitcode != EX_OK); 270 va_start(ap, fmt); 271 pjdlogv_errno(LOG_ERR, fmt, ap); 272 va_end(ap); 273 cleanup(gres); 274 exit(exitcode); 275} 276 277static __dead2 void 278primary_exitx(int exitcode, const char *fmt, ...) 279{ 280 va_list ap; 281 282 va_start(ap, fmt); 283 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 284 va_end(ap); 285 cleanup(gres); 286 exit(exitcode); 287} 288 289static int 290hast_activemap_flush(struct hast_resource *res) 291{ 292 const unsigned char *buf; 293 size_t size; 294 295 buf = activemap_bitmap(res->hr_amp, &size); 296 assert(buf != NULL); 297 assert((size % res->hr_local_sectorsize) == 0); 298 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 299 (ssize_t)size) { 300 KEEP_ERRNO(pjdlog_errno(LOG_ERR, 301 "Unable to flush activemap to disk")); 302 return (-1); 303 } 304 return (0); 305} 306 307static bool 308real_remote(const struct hast_resource *res) 309{ 310 311 return (strcmp(res->hr_remoteaddr, "none") != 0); 312} 313 314static void 315init_environment(struct hast_resource *res __unused) 316{ 317 struct hio *hio; 318 unsigned int ii, ncomps; 319 320 /* 321 * In the future it might be per-resource value. 322 */ 323 ncomps = HAST_NCOMPONENTS; 324 325 /* 326 * Allocate memory needed by lists. 327 */ 328 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 329 if (hio_send_list == NULL) { 330 primary_exitx(EX_TEMPFAIL, 331 "Unable to allocate %zu bytes of memory for send lists.", 332 sizeof(hio_send_list[0]) * ncomps); 333 } 334 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 335 if (hio_send_list_lock == NULL) { 336 primary_exitx(EX_TEMPFAIL, 337 "Unable to allocate %zu bytes of memory for send list locks.", 338 sizeof(hio_send_list_lock[0]) * ncomps); 339 } 340 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 341 if (hio_send_list_cond == NULL) { 342 primary_exitx(EX_TEMPFAIL, 343 "Unable to allocate %zu bytes of memory for send list condition variables.", 344 sizeof(hio_send_list_cond[0]) * ncomps); 345 } 346 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 347 if (hio_recv_list == NULL) { 348 primary_exitx(EX_TEMPFAIL, 349 "Unable to allocate %zu bytes of memory for recv lists.", 350 sizeof(hio_recv_list[0]) * ncomps); 351 } 352 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 353 if (hio_recv_list_lock == NULL) { 354 primary_exitx(EX_TEMPFAIL, 355 "Unable to allocate %zu bytes of memory for recv list locks.", 356 sizeof(hio_recv_list_lock[0]) * ncomps); 357 } 358 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 359 if (hio_recv_list_cond == NULL) { 360 primary_exitx(EX_TEMPFAIL, 361 "Unable to allocate %zu bytes of memory for recv list condition variables.", 362 sizeof(hio_recv_list_cond[0]) * ncomps); 363 } 364 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 365 if (hio_remote_lock == NULL) { 366 primary_exitx(EX_TEMPFAIL, 367 "Unable to allocate %zu bytes of memory for remote connections locks.", 368 sizeof(hio_remote_lock[0]) * ncomps); 369 } 370 371 /* 372 * Initialize lists, their locks and theirs condition variables. 373 */ 374 TAILQ_INIT(&hio_free_list); 375 mtx_init(&hio_free_list_lock); 376 cv_init(&hio_free_list_cond); 377 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 378 TAILQ_INIT(&hio_send_list[ii]); 379 mtx_init(&hio_send_list_lock[ii]); 380 cv_init(&hio_send_list_cond[ii]); 381 TAILQ_INIT(&hio_recv_list[ii]); 382 mtx_init(&hio_recv_list_lock[ii]); 383 cv_init(&hio_recv_list_cond[ii]); 384 rw_init(&hio_remote_lock[ii]); 385 } 386 TAILQ_INIT(&hio_done_list); 387 mtx_init(&hio_done_list_lock); 388 cv_init(&hio_done_list_cond); 389 mtx_init(&metadata_lock); 390 391 /* 392 * Allocate requests pool and initialize requests. 393 */ 394 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 395 hio = malloc(sizeof(*hio)); 396 if (hio == NULL) { 397 primary_exitx(EX_TEMPFAIL, 398 "Unable to allocate %zu bytes of memory for hio request.", 399 sizeof(*hio)); 400 } 401 hio->hio_countdown = 0; 402 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 403 if (hio->hio_errors == NULL) { 404 primary_exitx(EX_TEMPFAIL, 405 "Unable allocate %zu bytes of memory for hio errors.", 406 sizeof(hio->hio_errors[0]) * ncomps); 407 } 408 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 409 if (hio->hio_next == NULL) { 410 primary_exitx(EX_TEMPFAIL, 411 "Unable allocate %zu bytes of memory for hio_next field.", 412 sizeof(hio->hio_next[0]) * ncomps); 413 } 414 hio->hio_ggio.gctl_version = G_GATE_VERSION; 415 hio->hio_ggio.gctl_data = malloc(MAXPHYS); 416 if (hio->hio_ggio.gctl_data == NULL) { 417 primary_exitx(EX_TEMPFAIL, 418 "Unable to allocate %zu bytes of memory for gctl_data.", 419 MAXPHYS); 420 } 421 hio->hio_ggio.gctl_length = MAXPHYS; 422 hio->hio_ggio.gctl_error = 0; 423 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 424 } 425} 426 427static bool 428init_resuid(struct hast_resource *res) 429{ 430 431 mtx_lock(&metadata_lock); 432 if (res->hr_resuid != 0) { 433 mtx_unlock(&metadata_lock); 434 return (false); 435 } else { 436 /* Initialize unique resource identifier. */ 437 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 438 mtx_unlock(&metadata_lock); 439 if (metadata_write(res) < 0) 440 exit(EX_NOINPUT); 441 return (true); 442 } 443} 444 445static void 446init_local(struct hast_resource *res) 447{ 448 unsigned char *buf; 449 size_t mapsize; 450 451 if (metadata_read(res, true) < 0) 452 exit(EX_NOINPUT); 453 mtx_init(&res->hr_amp_lock); 454 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 455 res->hr_local_sectorsize, res->hr_keepdirty) < 0) { 456 primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 457 } 458 mtx_init(&range_lock); 459 cv_init(&range_regular_cond); 460 if (rangelock_init(&range_regular) < 0) 461 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 462 cv_init(&range_sync_cond); 463 if (rangelock_init(&range_sync) < 0) 464 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 465 mapsize = activemap_ondisk_size(res->hr_amp); 466 buf = calloc(1, mapsize); 467 if (buf == NULL) { 468 primary_exitx(EX_TEMPFAIL, 469 "Unable to allocate buffer for activemap."); 470 } 471 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 472 (ssize_t)mapsize) { 473 primary_exit(EX_NOINPUT, "Unable to read activemap"); 474 } 475 activemap_copyin(res->hr_amp, buf, mapsize); 476 free(buf); 477 if (res->hr_resuid != 0) 478 return; 479 /* 480 * We're using provider for the first time. Initialize local and remote 481 * counters. We don't initialize resuid here, as we want to do it just 482 * in time. The reason for this is that we want to inform secondary 483 * that there were no writes yet, so there is no need to synchronize 484 * anything. 485 */ 486 res->hr_primary_localcnt = 1; 487 res->hr_primary_remotecnt = 0; 488 if (metadata_write(res) < 0) 489 exit(EX_NOINPUT); 490} 491 492static bool 493init_remote(struct hast_resource *res, struct proto_conn **inp, 494 struct proto_conn **outp) 495{ 496 struct proto_conn *in, *out; 497 struct nv *nvout, *nvin; 498 const unsigned char *token; 499 unsigned char *map; 500 const char *errmsg; 501 int32_t extentsize; 502 int64_t datasize; 503 uint32_t mapsize; 504 size_t size; 505 506 assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 507 assert(real_remote(res)); 508 509 in = out = NULL; 510 errmsg = NULL; 511 512 /* Prepare outgoing connection with remote node. */ 513 if (proto_client(res->hr_remoteaddr, &out) < 0) { 514 primary_exit(EX_TEMPFAIL, 515 "Unable to create outgoing connection to %s", 516 res->hr_remoteaddr); 517 } 518 /* Try to connect, but accept failure. */ 519 if (proto_connect(out) < 0) { 520 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 521 res->hr_remoteaddr); 522 goto close; 523 } 524 /* Error in setting timeout is not critical, but why should it fail? */ 525 if (proto_timeout(out, res->hr_timeout) < 0) 526 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 527 /* 528 * First handshake step. 529 * Setup outgoing connection with remote node. 530 */ 531 nvout = nv_alloc(); 532 nv_add_string(nvout, res->hr_name, "resource"); 533 if (nv_error(nvout) != 0) { 534 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 535 "Unable to allocate header for connection with %s", 536 res->hr_remoteaddr); 537 nv_free(nvout); 538 goto close; 539 } 540 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) { 541 pjdlog_errno(LOG_WARNING, 542 "Unable to send handshake header to %s", 543 res->hr_remoteaddr); 544 nv_free(nvout); 545 goto close; 546 } 547 nv_free(nvout); 548 if (hast_proto_recv_hdr(out, &nvin) < 0) { 549 pjdlog_errno(LOG_WARNING, 550 "Unable to receive handshake header from %s", 551 res->hr_remoteaddr); 552 goto close; 553 } 554 errmsg = nv_get_string(nvin, "errmsg"); 555 if (errmsg != NULL) { 556 pjdlog_warning("%s", errmsg); 557 nv_free(nvin); 558 goto close; 559 } 560 token = nv_get_uint8_array(nvin, &size, "token"); 561 if (token == NULL) { 562 pjdlog_warning("Handshake header from %s has no 'token' field.", 563 res->hr_remoteaddr); 564 nv_free(nvin); 565 goto close; 566 } 567 if (size != sizeof(res->hr_token)) { 568 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 569 res->hr_remoteaddr, size, sizeof(res->hr_token)); 570 nv_free(nvin); 571 goto close; 572 } 573 bcopy(token, res->hr_token, sizeof(res->hr_token)); 574 nv_free(nvin); 575 576 /* 577 * Second handshake step. 578 * Setup incoming connection with remote node. 579 */ 580 if (proto_client(res->hr_remoteaddr, &in) < 0) { 581 primary_exit(EX_TEMPFAIL, 582 "Unable to create incoming connection to %s", 583 res->hr_remoteaddr); 584 } 585 /* Try to connect, but accept failure. */ 586 if (proto_connect(in) < 0) { 587 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 588 res->hr_remoteaddr); 589 goto close; 590 } 591 /* Error in setting timeout is not critical, but why should it fail? */ 592 if (proto_timeout(in, res->hr_timeout) < 0) 593 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 594 nvout = nv_alloc(); 595 nv_add_string(nvout, res->hr_name, "resource"); 596 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 597 "token"); 598 if (res->hr_resuid == 0) { 599 /* 600 * The resuid field was not yet initialized. 601 * Because we do synchronization inside init_resuid(), it is 602 * possible that someone already initialized it, the function 603 * will return false then, but if we successfully initialized 604 * it, we will get true. True means that there were no writes 605 * to this resource yet and we want to inform secondary that 606 * synchronization is not needed by sending "virgin" argument. 607 */ 608 if (init_resuid(res)) 609 nv_add_int8(nvout, 1, "virgin"); 610 } 611 nv_add_uint64(nvout, res->hr_resuid, "resuid"); 612 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 613 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 614 if (nv_error(nvout) != 0) { 615 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 616 "Unable to allocate header for connection with %s", 617 res->hr_remoteaddr); 618 nv_free(nvout); 619 goto close; 620 } 621 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) { 622 pjdlog_errno(LOG_WARNING, 623 "Unable to send handshake header to %s", 624 res->hr_remoteaddr); 625 nv_free(nvout); 626 goto close; 627 } 628 nv_free(nvout); 629 if (hast_proto_recv_hdr(out, &nvin) < 0) { 630 pjdlog_errno(LOG_WARNING, 631 "Unable to receive handshake header from %s", 632 res->hr_remoteaddr); 633 goto close; 634 } 635 errmsg = nv_get_string(nvin, "errmsg"); 636 if (errmsg != NULL) { 637 pjdlog_warning("%s", errmsg); 638 nv_free(nvin); 639 goto close; 640 } 641 datasize = nv_get_int64(nvin, "datasize"); 642 if (datasize != res->hr_datasize) { 643 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 644 (intmax_t)res->hr_datasize, (intmax_t)datasize); 645 nv_free(nvin); 646 goto close; 647 } 648 extentsize = nv_get_int32(nvin, "extentsize"); 649 if (extentsize != res->hr_extentsize) { 650 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 651 (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 652 nv_free(nvin); 653 goto close; 654 } 655 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 656 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 657 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 658 map = NULL; 659 mapsize = nv_get_uint32(nvin, "mapsize"); 660 if (mapsize > 0) { 661 map = malloc(mapsize); 662 if (map == NULL) { 663 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 664 (uintmax_t)mapsize); 665 nv_free(nvin); 666 goto close; 667 } 668 /* 669 * Remote node have some dirty extents on its own, lets 670 * download its activemap. 671 */ 672 if (hast_proto_recv_data(res, out, nvin, map, 673 mapsize) < 0) { 674 pjdlog_errno(LOG_ERR, 675 "Unable to receive remote activemap"); 676 nv_free(nvin); 677 free(map); 678 goto close; 679 } 680 /* 681 * Merge local and remote bitmaps. 682 */ 683 activemap_merge(res->hr_amp, map, mapsize); 684 free(map); 685 /* 686 * Now that we merged bitmaps from both nodes, flush it to the 687 * disk before we start to synchronize. 688 */ 689 (void)hast_activemap_flush(res); 690 } 691 nv_free(nvin); 692 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 693 if (inp != NULL && outp != NULL) { 694 *inp = in; 695 *outp = out; 696 } else { 697 res->hr_remotein = in; 698 res->hr_remoteout = out; 699 } 700 event_send(res, EVENT_CONNECT); 701 return (true); 702close: 703 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 704 event_send(res, EVENT_SPLITBRAIN); 705 proto_close(out); 706 if (in != NULL) 707 proto_close(in); 708 return (false); 709} 710 711static void 712sync_start(void) 713{ 714 715 mtx_lock(&sync_lock); 716 sync_inprogress = true; 717 mtx_unlock(&sync_lock); 718 cv_signal(&sync_cond); 719} 720 721static void 722sync_stop(void) 723{ 724 725 mtx_lock(&sync_lock); 726 if (sync_inprogress) 727 sync_inprogress = false; 728 mtx_unlock(&sync_lock); 729} 730 731static void 732init_ggate(struct hast_resource *res) 733{ 734 struct g_gate_ctl_create ggiocreate; 735 struct g_gate_ctl_cancel ggiocancel; 736 737 /* 738 * We communicate with ggate via /dev/ggctl. Open it. 739 */ 740 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 741 if (res->hr_ggatefd < 0) 742 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 743 /* 744 * Create provider before trying to connect, as connection failure 745 * is not critical, but may take some time. 746 */ 747 bzero(&ggiocreate, sizeof(ggiocreate)); 748 ggiocreate.gctl_version = G_GATE_VERSION; 749 ggiocreate.gctl_mediasize = res->hr_datasize; 750 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 751 ggiocreate.gctl_flags = 0; 752 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE; 753 ggiocreate.gctl_timeout = 0; 754 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 755 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 756 res->hr_provname); 757 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 758 pjdlog_info("Device hast/%s created.", res->hr_provname); 759 res->hr_ggateunit = ggiocreate.gctl_unit; 760 return; 761 } 762 if (errno != EEXIST) { 763 primary_exit(EX_OSERR, "Unable to create hast/%s device", 764 res->hr_provname); 765 } 766 pjdlog_debug(1, 767 "Device hast/%s already exists, we will try to take it over.", 768 res->hr_provname); 769 /* 770 * If we received EEXIST, we assume that the process who created the 771 * provider died and didn't clean up. In that case we will start from 772 * where he left of. 773 */ 774 bzero(&ggiocancel, sizeof(ggiocancel)); 775 ggiocancel.gctl_version = G_GATE_VERSION; 776 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 777 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 778 res->hr_provname); 779 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 780 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 781 res->hr_ggateunit = ggiocancel.gctl_unit; 782 return; 783 } 784 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 785 res->hr_provname); 786} 787 788void 789hastd_primary(struct hast_resource *res) 790{ 791 pthread_t td; 792 pid_t pid; 793 int error, mode; 794 795 /* 796 * Create communication channel between parent and child. 797 */ 798 if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 799 /* TODO: There's no need for this to be fatal error. */ 800 KEEP_ERRNO((void)pidfile_remove(pfh)); 801 pjdlog_exit(EX_OSERR, 802 "Unable to create control sockets between parent and child"); 803 } 804 /* 805 * Create communication channel between child and parent. 806 */ 807 if (proto_client("socketpair://", &res->hr_event) < 0) { 808 /* TODO: There's no need for this to be fatal error. */ 809 KEEP_ERRNO((void)pidfile_remove(pfh)); 810 pjdlog_exit(EX_OSERR, 811 "Unable to create event sockets between child and parent"); 812 } 813 814 pid = fork(); 815 if (pid < 0) { 816 /* TODO: There's no need for this to be fatal error. */ 817 KEEP_ERRNO((void)pidfile_remove(pfh)); 818 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 819 } 820 821 if (pid > 0) { 822 /* This is parent. */ 823 /* Declare that we are receiver. */ 824 proto_recv(res->hr_event, NULL, 0); 825 /* Declare that we are sender. */ 826 proto_send(res->hr_ctrl, NULL, 0); 827 res->hr_workerpid = pid; 828 return; 829 } 830 831 gres = res; 832 mode = pjdlog_mode_get(); 833 834 /* Declare that we are sender. */ 835 proto_send(res->hr_event, NULL, 0); 836 /* Declare that we are receiver. */ 837 proto_recv(res->hr_ctrl, NULL, 0); 838 descriptors_cleanup(res); 839 840 descriptors_assert(res, mode); 841 842 pjdlog_init(mode); 843 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 844 setproctitle("%s (primary)", res->hr_name); 845 846 init_local(res); 847 init_ggate(res); 848 init_environment(res); 849 850 if (drop_privs() != 0) { 851 cleanup(res); 852 exit(EX_CONFIG); 853 } 854 855 /* 856 * Create the guard thread first, so we can handle signals from the 857 * very begining. 858 */ 859 error = pthread_create(&td, NULL, guard_thread, res); 860 assert(error == 0); 861 /* 862 * Create the control thread before sending any event to the parent, 863 * as we can deadlock when parent sends control request to worker, 864 * but worker has no control thread started yet, so parent waits. 865 * In the meantime worker sends an event to the parent, but parent 866 * is unable to handle the event, because it waits for control 867 * request response. 868 */ 869 error = pthread_create(&td, NULL, ctrl_thread, res); 870 assert(error == 0); 871 if (real_remote(res) && init_remote(res, NULL, NULL)) 872 sync_start(); 873 error = pthread_create(&td, NULL, ggate_recv_thread, res); 874 assert(error == 0); 875 error = pthread_create(&td, NULL, local_send_thread, res); 876 assert(error == 0); 877 error = pthread_create(&td, NULL, remote_send_thread, res); 878 assert(error == 0); 879 error = pthread_create(&td, NULL, remote_recv_thread, res); 880 assert(error == 0); 881 error = pthread_create(&td, NULL, ggate_send_thread, res); 882 assert(error == 0); 883 (void)sync_thread(res); 884} 885 886static void 887reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 888{ 889 char msg[1024]; 890 va_list ap; 891 int len; 892 893 va_start(ap, fmt); 894 len = vsnprintf(msg, sizeof(msg), fmt, ap); 895 va_end(ap); 896 if ((size_t)len < sizeof(msg)) { 897 switch (ggio->gctl_cmd) { 898 case BIO_READ: 899 (void)snprintf(msg + len, sizeof(msg) - len, 900 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 901 (uintmax_t)ggio->gctl_length); 902 break; 903 case BIO_DELETE: 904 (void)snprintf(msg + len, sizeof(msg) - len, 905 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 906 (uintmax_t)ggio->gctl_length); 907 break; 908 case BIO_FLUSH: 909 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 910 break; 911 case BIO_WRITE: 912 (void)snprintf(msg + len, sizeof(msg) - len, 913 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 914 (uintmax_t)ggio->gctl_length); 915 break; 916 default: 917 (void)snprintf(msg + len, sizeof(msg) - len, 918 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 919 break; 920 } 921 } 922 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 923} 924 925static void 926remote_close(struct hast_resource *res, int ncomp) 927{ 928 929 rw_wlock(&hio_remote_lock[ncomp]); 930 /* 931 * A race is possible between dropping rlock and acquiring wlock - 932 * another thread can close connection in-between. 933 */ 934 if (!ISCONNECTED(res, ncomp)) { 935 assert(res->hr_remotein == NULL); 936 assert(res->hr_remoteout == NULL); 937 rw_unlock(&hio_remote_lock[ncomp]); 938 return; 939 } 940 941 assert(res->hr_remotein != NULL); 942 assert(res->hr_remoteout != NULL); 943 944 pjdlog_debug(2, "Closing incoming connection to %s.", 945 res->hr_remoteaddr); 946 proto_close(res->hr_remotein); 947 res->hr_remotein = NULL; 948 pjdlog_debug(2, "Closing outgoing connection to %s.", 949 res->hr_remoteaddr); 950 proto_close(res->hr_remoteout); 951 res->hr_remoteout = NULL; 952 953 rw_unlock(&hio_remote_lock[ncomp]); 954 955 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 956 957 /* 958 * Stop synchronization if in-progress. 959 */ 960 sync_stop(); 961 962 event_send(res, EVENT_DISCONNECT); 963} 964 965/* 966 * Thread receives ggate I/O requests from the kernel and passes them to 967 * appropriate threads: 968 * WRITE - always goes to both local_send and remote_send threads 969 * READ (when the block is up-to-date on local component) - 970 * only local_send thread 971 * READ (when the block isn't up-to-date on local component) - 972 * only remote_send thread 973 * DELETE - always goes to both local_send and remote_send threads 974 * FLUSH - always goes to both local_send and remote_send threads 975 */ 976static void * 977ggate_recv_thread(void *arg) 978{ 979 struct hast_resource *res = arg; 980 struct g_gate_ctl_io *ggio; 981 struct hio *hio; 982 unsigned int ii, ncomp, ncomps; 983 int error; 984 985 ncomps = HAST_NCOMPONENTS; 986 987 for (;;) { 988 pjdlog_debug(2, "ggate_recv: Taking free request."); 989 QUEUE_TAKE2(hio, free); 990 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 991 ggio = &hio->hio_ggio; 992 ggio->gctl_unit = res->hr_ggateunit; 993 ggio->gctl_length = MAXPHYS; 994 ggio->gctl_error = 0; 995 pjdlog_debug(2, 996 "ggate_recv: (%p) Waiting for request from the kernel.", 997 hio); 998 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 999 if (sigexit_received) 1000 pthread_exit(NULL); 1001 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 1002 } 1003 error = ggio->gctl_error; 1004 switch (error) { 1005 case 0: 1006 break; 1007 case ECANCELED: 1008 /* Exit gracefully. */ 1009 if (!sigexit_received) { 1010 pjdlog_debug(2, 1011 "ggate_recv: (%p) Received cancel from the kernel.", 1012 hio); 1013 pjdlog_info("Received cancel from the kernel, exiting."); 1014 } 1015 pthread_exit(NULL); 1016 case ENOMEM: 1017 /* 1018 * Buffer too small? Impossible, we allocate MAXPHYS 1019 * bytes - request can't be bigger than that. 1020 */ 1021 /* FALLTHROUGH */ 1022 case ENXIO: 1023 default: 1024 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 1025 strerror(error)); 1026 } 1027 for (ii = 0; ii < ncomps; ii++) 1028 hio->hio_errors[ii] = EINVAL; 1029 reqlog(LOG_DEBUG, 2, ggio, 1030 "ggate_recv: (%p) Request received from the kernel: ", 1031 hio); 1032 /* 1033 * Inform all components about new write request. 1034 * For read request prefer local component unless the given 1035 * range is out-of-date, then use remote component. 1036 */ 1037 switch (ggio->gctl_cmd) { 1038 case BIO_READ: 1039 pjdlog_debug(2, 1040 "ggate_recv: (%p) Moving request to the send queue.", 1041 hio); 1042 refcount_init(&hio->hio_countdown, 1); 1043 mtx_lock(&metadata_lock); 1044 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 1045 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1046 /* 1047 * This range is up-to-date on local component, 1048 * so handle request locally. 1049 */ 1050 /* Local component is 0 for now. */ 1051 ncomp = 0; 1052 } else /* if (res->hr_syncsrc == 1053 HAST_SYNCSRC_SECONDARY) */ { 1054 assert(res->hr_syncsrc == 1055 HAST_SYNCSRC_SECONDARY); 1056 /* 1057 * This range is out-of-date on local component, 1058 * so send request to the remote node. 1059 */ 1060 /* Remote component is 1 for now. */ 1061 ncomp = 1; 1062 } 1063 mtx_unlock(&metadata_lock); 1064 QUEUE_INSERT1(hio, send, ncomp); 1065 break; 1066 case BIO_WRITE: 1067 if (res->hr_resuid == 0) { 1068 /* This is first write, initialize resuid. */ 1069 (void)init_resuid(res); 1070 } 1071 for (;;) { 1072 mtx_lock(&range_lock); 1073 if (rangelock_islocked(range_sync, 1074 ggio->gctl_offset, ggio->gctl_length)) { 1075 pjdlog_debug(2, 1076 "regular: Range offset=%jd length=%zu locked.", 1077 (intmax_t)ggio->gctl_offset, 1078 (size_t)ggio->gctl_length); 1079 range_regular_wait = true; 1080 cv_wait(&range_regular_cond, &range_lock); 1081 range_regular_wait = false; 1082 mtx_unlock(&range_lock); 1083 continue; 1084 } 1085 if (rangelock_add(range_regular, 1086 ggio->gctl_offset, ggio->gctl_length) < 0) { 1087 mtx_unlock(&range_lock); 1088 pjdlog_debug(2, 1089 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1090 (intmax_t)ggio->gctl_offset, 1091 (size_t)ggio->gctl_length); 1092 sleep(1); 1093 continue; 1094 } 1095 mtx_unlock(&range_lock); 1096 break; 1097 } 1098 mtx_lock(&res->hr_amp_lock); 1099 if (activemap_write_start(res->hr_amp, 1100 ggio->gctl_offset, ggio->gctl_length)) { 1101 (void)hast_activemap_flush(res); 1102 } 1103 mtx_unlock(&res->hr_amp_lock); 1104 /* FALLTHROUGH */ 1105 case BIO_DELETE: 1106 case BIO_FLUSH: 1107 pjdlog_debug(2, 1108 "ggate_recv: (%p) Moving request to the send queues.", 1109 hio); 1110 refcount_init(&hio->hio_countdown, ncomps); 1111 for (ii = 0; ii < ncomps; ii++) 1112 QUEUE_INSERT1(hio, send, ii); 1113 break; 1114 } 1115 } 1116 /* NOTREACHED */ 1117 return (NULL); 1118} 1119 1120/* 1121 * Thread reads from or writes to local component. 1122 * If local read fails, it redirects it to remote_send thread. 1123 */ 1124static void * 1125local_send_thread(void *arg) 1126{ 1127 struct hast_resource *res = arg; 1128 struct g_gate_ctl_io *ggio; 1129 struct hio *hio; 1130 unsigned int ncomp, rncomp; 1131 ssize_t ret; 1132 1133 /* Local component is 0 for now. */ 1134 ncomp = 0; 1135 /* Remote component is 1 for now. */ 1136 rncomp = 1; 1137 1138 for (;;) { 1139 pjdlog_debug(2, "local_send: Taking request."); 1140 QUEUE_TAKE1(hio, send, ncomp, 0); 1141 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1142 ggio = &hio->hio_ggio; 1143 switch (ggio->gctl_cmd) { 1144 case BIO_READ: 1145 ret = pread(res->hr_localfd, ggio->gctl_data, 1146 ggio->gctl_length, 1147 ggio->gctl_offset + res->hr_localoff); 1148 if (ret == ggio->gctl_length) 1149 hio->hio_errors[ncomp] = 0; 1150 else { 1151 /* 1152 * If READ failed, try to read from remote node. 1153 */ 1154 if (ret < 0) { 1155 reqlog(LOG_WARNING, 0, ggio, 1156 "Local request failed (%s), trying remote node. ", 1157 strerror(errno)); 1158 } else if (ret != ggio->gctl_length) { 1159 reqlog(LOG_WARNING, 0, ggio, 1160 "Local request failed (%zd != %jd), trying remote node. ", 1161 ret, (intmax_t)ggio->gctl_length); 1162 } 1163 QUEUE_INSERT1(hio, send, rncomp); 1164 continue; 1165 } 1166 break; 1167 case BIO_WRITE: 1168 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1169 ggio->gctl_length, 1170 ggio->gctl_offset + res->hr_localoff); 1171 if (ret < 0) { 1172 hio->hio_errors[ncomp] = errno; 1173 reqlog(LOG_WARNING, 0, ggio, 1174 "Local request failed (%s): ", 1175 strerror(errno)); 1176 } else if (ret != ggio->gctl_length) { 1177 hio->hio_errors[ncomp] = EIO; 1178 reqlog(LOG_WARNING, 0, ggio, 1179 "Local request failed (%zd != %jd): ", 1180 ret, (intmax_t)ggio->gctl_length); 1181 } else { 1182 hio->hio_errors[ncomp] = 0; 1183 } 1184 break; 1185 case BIO_DELETE: 1186 ret = g_delete(res->hr_localfd, 1187 ggio->gctl_offset + res->hr_localoff, 1188 ggio->gctl_length); 1189 if (ret < 0) { 1190 hio->hio_errors[ncomp] = errno; 1191 reqlog(LOG_WARNING, 0, ggio, 1192 "Local request failed (%s): ", 1193 strerror(errno)); 1194 } else { 1195 hio->hio_errors[ncomp] = 0; 1196 } 1197 break; 1198 case BIO_FLUSH: 1199 ret = g_flush(res->hr_localfd); 1200 if (ret < 0) { 1201 hio->hio_errors[ncomp] = errno; 1202 reqlog(LOG_WARNING, 0, ggio, 1203 "Local request failed (%s): ", 1204 strerror(errno)); 1205 } else { 1206 hio->hio_errors[ncomp] = 0; 1207 } 1208 break; 1209 } 1210 if (refcount_release(&hio->hio_countdown)) { 1211 if (ISSYNCREQ(hio)) { 1212 mtx_lock(&sync_lock); 1213 SYNCREQDONE(hio); 1214 mtx_unlock(&sync_lock); 1215 cv_signal(&sync_cond); 1216 } else { 1217 pjdlog_debug(2, 1218 "local_send: (%p) Moving request to the done queue.", 1219 hio); 1220 QUEUE_INSERT2(hio, done); 1221 } 1222 } 1223 } 1224 /* NOTREACHED */ 1225 return (NULL); 1226} 1227 1228static void 1229keepalive_send(struct hast_resource *res, unsigned int ncomp) 1230{ 1231 struct nv *nv; 1232 1233 if (!ISCONNECTED(res, ncomp)) 1234 return; 1235 1236 assert(res->hr_remotein != NULL); 1237 assert(res->hr_remoteout != NULL); 1238 1239 nv = nv_alloc(); 1240 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1241 if (nv_error(nv) != 0) { 1242 nv_free(nv); 1243 pjdlog_debug(1, 1244 "keepalive_send: Unable to prepare header to send."); 1245 return; 1246 } 1247 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1248 pjdlog_common(LOG_DEBUG, 1, errno, 1249 "keepalive_send: Unable to send request"); 1250 nv_free(nv); 1251 rw_unlock(&hio_remote_lock[ncomp]); 1252 remote_close(res, ncomp); 1253 rw_rlock(&hio_remote_lock[ncomp]); 1254 return; 1255 } 1256 nv_free(nv); 1257 pjdlog_debug(2, "keepalive_send: Request sent."); 1258} 1259 1260/* 1261 * Thread sends request to secondary node. 1262 */ 1263static void * 1264remote_send_thread(void *arg) 1265{ 1266 struct hast_resource *res = arg; 1267 struct g_gate_ctl_io *ggio; 1268 time_t lastcheck, now; 1269 struct hio *hio; 1270 struct nv *nv; 1271 unsigned int ncomp; 1272 bool wakeup; 1273 uint64_t offset, length; 1274 uint8_t cmd; 1275 void *data; 1276 1277 /* Remote component is 1 for now. */ 1278 ncomp = 1; 1279 lastcheck = time(NULL); 1280 1281 for (;;) { 1282 pjdlog_debug(2, "remote_send: Taking request."); 1283 QUEUE_TAKE1(hio, send, ncomp, RETRY_SLEEP); 1284 if (hio == NULL) { 1285 now = time(NULL); 1286 if (lastcheck + RETRY_SLEEP <= now) { 1287 keepalive_send(res, ncomp); 1288 lastcheck = now; 1289 } 1290 continue; 1291 } 1292 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1293 ggio = &hio->hio_ggio; 1294 switch (ggio->gctl_cmd) { 1295 case BIO_READ: 1296 cmd = HIO_READ; 1297 data = NULL; 1298 offset = ggio->gctl_offset; 1299 length = ggio->gctl_length; 1300 break; 1301 case BIO_WRITE: 1302 cmd = HIO_WRITE; 1303 data = ggio->gctl_data; 1304 offset = ggio->gctl_offset; 1305 length = ggio->gctl_length; 1306 break; 1307 case BIO_DELETE: 1308 cmd = HIO_DELETE; 1309 data = NULL; 1310 offset = ggio->gctl_offset; 1311 length = ggio->gctl_length; 1312 break; 1313 case BIO_FLUSH: 1314 cmd = HIO_FLUSH; 1315 data = NULL; 1316 offset = 0; 1317 length = 0; 1318 break; 1319 default: 1320 assert(!"invalid condition"); 1321 abort(); 1322 } 1323 nv = nv_alloc(); 1324 nv_add_uint8(nv, cmd, "cmd"); 1325 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1326 nv_add_uint64(nv, offset, "offset"); 1327 nv_add_uint64(nv, length, "length"); 1328 if (nv_error(nv) != 0) { 1329 hio->hio_errors[ncomp] = nv_error(nv); 1330 pjdlog_debug(2, 1331 "remote_send: (%p) Unable to prepare header to send.", 1332 hio); 1333 reqlog(LOG_ERR, 0, ggio, 1334 "Unable to prepare header to send (%s): ", 1335 strerror(nv_error(nv))); 1336 /* Move failed request immediately to the done queue. */ 1337 goto done_queue; 1338 } 1339 pjdlog_debug(2, 1340 "remote_send: (%p) Moving request to the recv queue.", 1341 hio); 1342 /* 1343 * Protect connection from disappearing. 1344 */ 1345 rw_rlock(&hio_remote_lock[ncomp]); 1346 if (!ISCONNECTED(res, ncomp)) { 1347 rw_unlock(&hio_remote_lock[ncomp]); 1348 hio->hio_errors[ncomp] = ENOTCONN; 1349 goto done_queue; 1350 } 1351 /* 1352 * Move the request to recv queue before sending it, because 1353 * in different order we can get reply before we move request 1354 * to recv queue. 1355 */ 1356 mtx_lock(&hio_recv_list_lock[ncomp]); 1357 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1358 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1359 mtx_unlock(&hio_recv_list_lock[ncomp]); 1360 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1361 data != NULL ? length : 0) < 0) { 1362 hio->hio_errors[ncomp] = errno; 1363 rw_unlock(&hio_remote_lock[ncomp]); 1364 pjdlog_debug(2, 1365 "remote_send: (%p) Unable to send request.", hio); 1366 reqlog(LOG_ERR, 0, ggio, 1367 "Unable to send request (%s): ", 1368 strerror(hio->hio_errors[ncomp])); 1369 remote_close(res, ncomp); 1370 /* 1371 * Take request back from the receive queue and move 1372 * it immediately to the done queue. 1373 */ 1374 mtx_lock(&hio_recv_list_lock[ncomp]); 1375 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1376 mtx_unlock(&hio_recv_list_lock[ncomp]); 1377 goto done_queue; 1378 } 1379 rw_unlock(&hio_remote_lock[ncomp]); 1380 nv_free(nv); 1381 if (wakeup) 1382 cv_signal(&hio_recv_list_cond[ncomp]); 1383 continue; 1384done_queue: 1385 nv_free(nv); 1386 if (ISSYNCREQ(hio)) { 1387 if (!refcount_release(&hio->hio_countdown)) 1388 continue; 1389 mtx_lock(&sync_lock); 1390 SYNCREQDONE(hio); 1391 mtx_unlock(&sync_lock); 1392 cv_signal(&sync_cond); 1393 continue; 1394 } 1395 if (ggio->gctl_cmd == BIO_WRITE) { 1396 mtx_lock(&res->hr_amp_lock); 1397 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1398 ggio->gctl_length)) { 1399 (void)hast_activemap_flush(res); 1400 } 1401 mtx_unlock(&res->hr_amp_lock); 1402 } 1403 if (!refcount_release(&hio->hio_countdown)) 1404 continue; 1405 pjdlog_debug(2, 1406 "remote_send: (%p) Moving request to the done queue.", 1407 hio); 1408 QUEUE_INSERT2(hio, done); 1409 } 1410 /* NOTREACHED */ 1411 return (NULL); 1412} 1413 1414/* 1415 * Thread receives answer from secondary node and passes it to ggate_send 1416 * thread. 1417 */ 1418static void * 1419remote_recv_thread(void *arg) 1420{ 1421 struct hast_resource *res = arg; 1422 struct g_gate_ctl_io *ggio; 1423 struct hio *hio; 1424 struct nv *nv; 1425 unsigned int ncomp; 1426 uint64_t seq; 1427 int error; 1428 1429 /* Remote component is 1 for now. */ 1430 ncomp = 1; 1431 1432 for (;;) { 1433 /* Wait until there is anything to receive. */ 1434 mtx_lock(&hio_recv_list_lock[ncomp]); 1435 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1436 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1437 cv_wait(&hio_recv_list_cond[ncomp], 1438 &hio_recv_list_lock[ncomp]); 1439 } 1440 mtx_unlock(&hio_recv_list_lock[ncomp]); 1441 rw_rlock(&hio_remote_lock[ncomp]); 1442 if (!ISCONNECTED(res, ncomp)) { 1443 rw_unlock(&hio_remote_lock[ncomp]); 1444 /* 1445 * Connection is dead, so move all pending requests to 1446 * the done queue (one-by-one). 1447 */ 1448 mtx_lock(&hio_recv_list_lock[ncomp]); 1449 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1450 assert(hio != NULL); 1451 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1452 hio_next[ncomp]); 1453 mtx_unlock(&hio_recv_list_lock[ncomp]); 1454 goto done_queue; 1455 } 1456 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1457 pjdlog_errno(LOG_ERR, 1458 "Unable to receive reply header"); 1459 rw_unlock(&hio_remote_lock[ncomp]); 1460 remote_close(res, ncomp); 1461 continue; 1462 } 1463 rw_unlock(&hio_remote_lock[ncomp]); 1464 seq = nv_get_uint64(nv, "seq"); 1465 if (seq == 0) { 1466 pjdlog_error("Header contains no 'seq' field."); 1467 nv_free(nv); 1468 continue; 1469 } 1470 mtx_lock(&hio_recv_list_lock[ncomp]); 1471 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1472 if (hio->hio_ggio.gctl_seq == seq) { 1473 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1474 hio_next[ncomp]); 1475 break; 1476 } 1477 } 1478 mtx_unlock(&hio_recv_list_lock[ncomp]); 1479 if (hio == NULL) { 1480 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1481 (uintmax_t)seq); 1482 nv_free(nv); 1483 continue; 1484 } 1485 error = nv_get_int16(nv, "error"); 1486 if (error != 0) { 1487 /* Request failed on remote side. */ 1488 hio->hio_errors[ncomp] = error; 1489 reqlog(LOG_WARNING, 0, &hio->hio_ggio, 1490 "Remote request failed (%s): ", strerror(error)); 1491 nv_free(nv); 1492 goto done_queue; 1493 } 1494 ggio = &hio->hio_ggio; 1495 switch (ggio->gctl_cmd) { 1496 case BIO_READ: 1497 rw_rlock(&hio_remote_lock[ncomp]); 1498 if (!ISCONNECTED(res, ncomp)) { 1499 rw_unlock(&hio_remote_lock[ncomp]); 1500 nv_free(nv); 1501 goto done_queue; 1502 } 1503 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1504 ggio->gctl_data, ggio->gctl_length) < 0) { 1505 hio->hio_errors[ncomp] = errno; 1506 pjdlog_errno(LOG_ERR, 1507 "Unable to receive reply data"); 1508 rw_unlock(&hio_remote_lock[ncomp]); 1509 nv_free(nv); 1510 remote_close(res, ncomp); 1511 goto done_queue; 1512 } 1513 rw_unlock(&hio_remote_lock[ncomp]); 1514 break; 1515 case BIO_WRITE: 1516 case BIO_DELETE: 1517 case BIO_FLUSH: 1518 break; 1519 default: 1520 assert(!"invalid condition"); 1521 abort(); 1522 } 1523 hio->hio_errors[ncomp] = 0; 1524 nv_free(nv); 1525done_queue: 1526 if (refcount_release(&hio->hio_countdown)) { 1527 if (ISSYNCREQ(hio)) { 1528 mtx_lock(&sync_lock); 1529 SYNCREQDONE(hio); 1530 mtx_unlock(&sync_lock); 1531 cv_signal(&sync_cond); 1532 } else { 1533 pjdlog_debug(2, 1534 "remote_recv: (%p) Moving request to the done queue.", 1535 hio); 1536 QUEUE_INSERT2(hio, done); 1537 } 1538 } 1539 } 1540 /* NOTREACHED */ 1541 return (NULL); 1542} 1543 1544/* 1545 * Thread sends answer to the kernel. 1546 */ 1547static void * 1548ggate_send_thread(void *arg) 1549{ 1550 struct hast_resource *res = arg; 1551 struct g_gate_ctl_io *ggio; 1552 struct hio *hio; 1553 unsigned int ii, ncomp, ncomps; 1554 1555 ncomps = HAST_NCOMPONENTS; 1556 1557 for (;;) { 1558 pjdlog_debug(2, "ggate_send: Taking request."); 1559 QUEUE_TAKE2(hio, done); 1560 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1561 ggio = &hio->hio_ggio; 1562 for (ii = 0; ii < ncomps; ii++) { 1563 if (hio->hio_errors[ii] == 0) { 1564 /* 1565 * One successful request is enough to declare 1566 * success. 1567 */ 1568 ggio->gctl_error = 0; 1569 break; 1570 } 1571 } 1572 if (ii == ncomps) { 1573 /* 1574 * None of the requests were successful. 1575 * Use first error. 1576 */ 1577 ggio->gctl_error = hio->hio_errors[0]; 1578 } 1579 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1580 mtx_lock(&res->hr_amp_lock); 1581 activemap_write_complete(res->hr_amp, 1582 ggio->gctl_offset, ggio->gctl_length); 1583 mtx_unlock(&res->hr_amp_lock); 1584 } 1585 if (ggio->gctl_cmd == BIO_WRITE) { 1586 /* 1587 * Unlock range we locked. 1588 */ 1589 mtx_lock(&range_lock); 1590 rangelock_del(range_regular, ggio->gctl_offset, 1591 ggio->gctl_length); 1592 if (range_sync_wait) 1593 cv_signal(&range_sync_cond); 1594 mtx_unlock(&range_lock); 1595 /* 1596 * Bump local count if this is first write after 1597 * connection failure with remote node. 1598 */ 1599 ncomp = 1; 1600 rw_rlock(&hio_remote_lock[ncomp]); 1601 if (!ISCONNECTED(res, ncomp)) { 1602 mtx_lock(&metadata_lock); 1603 if (res->hr_primary_localcnt == 1604 res->hr_secondary_remotecnt) { 1605 res->hr_primary_localcnt++; 1606 pjdlog_debug(1, 1607 "Increasing localcnt to %ju.", 1608 (uintmax_t)res->hr_primary_localcnt); 1609 (void)metadata_write(res); 1610 } 1611 mtx_unlock(&metadata_lock); 1612 } 1613 rw_unlock(&hio_remote_lock[ncomp]); 1614 } 1615 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1616 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1617 pjdlog_debug(2, 1618 "ggate_send: (%p) Moving request to the free queue.", hio); 1619 QUEUE_INSERT2(hio, free); 1620 } 1621 /* NOTREACHED */ 1622 return (NULL); 1623} 1624 1625/* 1626 * Thread synchronize local and remote components. 1627 */ 1628static void * 1629sync_thread(void *arg __unused) 1630{ 1631 struct hast_resource *res = arg; 1632 struct hio *hio; 1633 struct g_gate_ctl_io *ggio; 1634 unsigned int ii, ncomp, ncomps; 1635 off_t offset, length, synced; 1636 bool dorewind; 1637 int syncext; 1638 1639 ncomps = HAST_NCOMPONENTS; 1640 dorewind = true; 1641 synced = 0; 1642 offset = -1; 1643 1644 for (;;) { 1645 mtx_lock(&sync_lock); 1646 if (offset >= 0 && !sync_inprogress) { 1647 pjdlog_info("Synchronization interrupted. " 1648 "%jd bytes synchronized so far.", 1649 (intmax_t)synced); 1650 event_send(res, EVENT_SYNCINTR); 1651 } 1652 while (!sync_inprogress) { 1653 dorewind = true; 1654 synced = 0; 1655 cv_wait(&sync_cond, &sync_lock); 1656 } 1657 mtx_unlock(&sync_lock); 1658 /* 1659 * Obtain offset at which we should synchronize. 1660 * Rewind synchronization if needed. 1661 */ 1662 mtx_lock(&res->hr_amp_lock); 1663 if (dorewind) 1664 activemap_sync_rewind(res->hr_amp); 1665 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1666 if (syncext != -1) { 1667 /* 1668 * We synchronized entire syncext extent, we can mark 1669 * it as clean now. 1670 */ 1671 if (activemap_extent_complete(res->hr_amp, syncext)) 1672 (void)hast_activemap_flush(res); 1673 } 1674 mtx_unlock(&res->hr_amp_lock); 1675 if (dorewind) { 1676 dorewind = false; 1677 if (offset < 0) 1678 pjdlog_info("Nodes are in sync."); 1679 else { 1680 pjdlog_info("Synchronization started. %ju bytes to go.", 1681 (uintmax_t)(res->hr_extentsize * 1682 activemap_ndirty(res->hr_amp))); 1683 event_send(res, EVENT_SYNCSTART); 1684 } 1685 } 1686 if (offset < 0) { 1687 sync_stop(); 1688 pjdlog_debug(1, "Nothing to synchronize."); 1689 /* 1690 * Synchronization complete, make both localcnt and 1691 * remotecnt equal. 1692 */ 1693 ncomp = 1; 1694 rw_rlock(&hio_remote_lock[ncomp]); 1695 if (ISCONNECTED(res, ncomp)) { 1696 if (synced > 0) { 1697 pjdlog_info("Synchronization complete. " 1698 "%jd bytes synchronized.", 1699 (intmax_t)synced); 1700 event_send(res, EVENT_SYNCDONE); 1701 } 1702 mtx_lock(&metadata_lock); 1703 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1704 res->hr_primary_localcnt = 1705 res->hr_secondary_localcnt; 1706 res->hr_primary_remotecnt = 1707 res->hr_secondary_remotecnt; 1708 pjdlog_debug(1, 1709 "Setting localcnt to %ju and remotecnt to %ju.", 1710 (uintmax_t)res->hr_primary_localcnt, 1711 (uintmax_t)res->hr_secondary_localcnt); 1712 (void)metadata_write(res); 1713 mtx_unlock(&metadata_lock); 1714 } 1715 rw_unlock(&hio_remote_lock[ncomp]); 1716 continue; 1717 } 1718 pjdlog_debug(2, "sync: Taking free request."); 1719 QUEUE_TAKE2(hio, free); 1720 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1721 /* 1722 * Lock the range we are going to synchronize. We don't want 1723 * race where someone writes between our read and write. 1724 */ 1725 for (;;) { 1726 mtx_lock(&range_lock); 1727 if (rangelock_islocked(range_regular, offset, length)) { 1728 pjdlog_debug(2, 1729 "sync: Range offset=%jd length=%jd locked.", 1730 (intmax_t)offset, (intmax_t)length); 1731 range_sync_wait = true; 1732 cv_wait(&range_sync_cond, &range_lock); 1733 range_sync_wait = false; 1734 mtx_unlock(&range_lock); 1735 continue; 1736 } 1737 if (rangelock_add(range_sync, offset, length) < 0) { 1738 mtx_unlock(&range_lock); 1739 pjdlog_debug(2, 1740 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1741 (intmax_t)offset, (intmax_t)length); 1742 sleep(1); 1743 continue; 1744 } 1745 mtx_unlock(&range_lock); 1746 break; 1747 } 1748 /* 1749 * First read the data from synchronization source. 1750 */ 1751 SYNCREQ(hio); 1752 ggio = &hio->hio_ggio; 1753 ggio->gctl_cmd = BIO_READ; 1754 ggio->gctl_offset = offset; 1755 ggio->gctl_length = length; 1756 ggio->gctl_error = 0; 1757 for (ii = 0; ii < ncomps; ii++) 1758 hio->hio_errors[ii] = EINVAL; 1759 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1760 hio); 1761 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1762 hio); 1763 mtx_lock(&metadata_lock); 1764 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1765 /* 1766 * This range is up-to-date on local component, 1767 * so handle request locally. 1768 */ 1769 /* Local component is 0 for now. */ 1770 ncomp = 0; 1771 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1772 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1773 /* 1774 * This range is out-of-date on local component, 1775 * so send request to the remote node. 1776 */ 1777 /* Remote component is 1 for now. */ 1778 ncomp = 1; 1779 } 1780 mtx_unlock(&metadata_lock); 1781 refcount_init(&hio->hio_countdown, 1); 1782 QUEUE_INSERT1(hio, send, ncomp); 1783 1784 /* 1785 * Let's wait for READ to finish. 1786 */ 1787 mtx_lock(&sync_lock); 1788 while (!ISSYNCREQDONE(hio)) 1789 cv_wait(&sync_cond, &sync_lock); 1790 mtx_unlock(&sync_lock); 1791 1792 if (hio->hio_errors[ncomp] != 0) { 1793 pjdlog_error("Unable to read synchronization data: %s.", 1794 strerror(hio->hio_errors[ncomp])); 1795 goto free_queue; 1796 } 1797 1798 /* 1799 * We read the data from synchronization source, now write it 1800 * to synchronization target. 1801 */ 1802 SYNCREQ(hio); 1803 ggio->gctl_cmd = BIO_WRITE; 1804 for (ii = 0; ii < ncomps; ii++) 1805 hio->hio_errors[ii] = EINVAL; 1806 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1807 hio); 1808 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1809 hio); 1810 mtx_lock(&metadata_lock); 1811 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1812 /* 1813 * This range is up-to-date on local component, 1814 * so we update remote component. 1815 */ 1816 /* Remote component is 1 for now. */ 1817 ncomp = 1; 1818 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1819 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1820 /* 1821 * This range is out-of-date on local component, 1822 * so we update it. 1823 */ 1824 /* Local component is 0 for now. */ 1825 ncomp = 0; 1826 } 1827 mtx_unlock(&metadata_lock); 1828 1829 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1830 hio); 1831 refcount_init(&hio->hio_countdown, 1); 1832 QUEUE_INSERT1(hio, send, ncomp); 1833 1834 /* 1835 * Let's wait for WRITE to finish. 1836 */ 1837 mtx_lock(&sync_lock); 1838 while (!ISSYNCREQDONE(hio)) 1839 cv_wait(&sync_cond, &sync_lock); 1840 mtx_unlock(&sync_lock); 1841 1842 if (hio->hio_errors[ncomp] != 0) { 1843 pjdlog_error("Unable to write synchronization data: %s.", 1844 strerror(hio->hio_errors[ncomp])); 1845 goto free_queue; 1846 } 1847 1848 synced += length; 1849free_queue: 1850 mtx_lock(&range_lock); 1851 rangelock_del(range_sync, offset, length); 1852 if (range_regular_wait) 1853 cv_signal(&range_regular_cond); 1854 mtx_unlock(&range_lock); 1855 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1856 hio); 1857 QUEUE_INSERT2(hio, free); 1858 } 1859 /* NOTREACHED */ 1860 return (NULL); 1861} 1862 1863void 1864primary_config_reload(struct hast_resource *res, struct nv *nv) 1865{ 1866 unsigned int ii, ncomps; 1867 int modified, vint; 1868 const char *vstr; 1869 1870 pjdlog_info("Reloading configuration..."); 1871 1872 assert(res->hr_role == HAST_ROLE_PRIMARY); 1873 assert(gres == res); 1874 nv_assert(nv, "remoteaddr"); 1875 nv_assert(nv, "replication"); 1876 nv_assert(nv, "timeout"); 1877 nv_assert(nv, "exec"); 1878 1879 ncomps = HAST_NCOMPONENTS; 1880 1881#define MODIFIED_REMOTEADDR 0x1 1882#define MODIFIED_REPLICATION 0x2 1883#define MODIFIED_TIMEOUT 0x4 1884#define MODIFIED_EXEC 0x8 1885 modified = 0; 1886 1887 vstr = nv_get_string(nv, "remoteaddr"); 1888 if (strcmp(gres->hr_remoteaddr, vstr) != 0) { 1889 /* 1890 * Don't copy res->hr_remoteaddr to gres just yet. 1891 * We want remote_close() to log disconnect from the old 1892 * addresses, not from the new ones. 1893 */ 1894 modified |= MODIFIED_REMOTEADDR; 1895 } 1896 vint = nv_get_int32(nv, "replication"); 1897 if (gres->hr_replication != vint) { 1898 gres->hr_replication = vint; 1899 modified |= MODIFIED_REPLICATION; 1900 } 1901 vint = nv_get_int32(nv, "timeout"); 1902 if (gres->hr_timeout != vint) { 1903 gres->hr_timeout = vint; 1904 modified |= MODIFIED_TIMEOUT; 1905 } 1906 vstr = nv_get_string(nv, "exec"); 1907 if (strcmp(gres->hr_exec, vstr) != 0) { 1908 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec)); 1909 modified |= MODIFIED_EXEC; 1910 } 1911 1912 /* 1913 * If only timeout was modified we only need to change it without 1914 * reconnecting. 1915 */ 1916 if (modified == MODIFIED_TIMEOUT) { 1917 for (ii = 0; ii < ncomps; ii++) { 1918 if (!ISREMOTE(ii)) 1919 continue; 1920 rw_rlock(&hio_remote_lock[ii]); 1921 if (!ISCONNECTED(gres, ii)) { 1922 rw_unlock(&hio_remote_lock[ii]); 1923 continue; 1924 } 1925 rw_unlock(&hio_remote_lock[ii]); 1926 if (proto_timeout(gres->hr_remotein, 1927 gres->hr_timeout) < 0) { 1928 pjdlog_errno(LOG_WARNING, 1929 "Unable to set connection timeout"); 1930 } 1931 if (proto_timeout(gres->hr_remoteout, 1932 gres->hr_timeout) < 0) { 1933 pjdlog_errno(LOG_WARNING, 1934 "Unable to set connection timeout"); 1935 } 1936 } 1937 } else if ((modified & 1938 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) { 1939 for (ii = 0; ii < ncomps; ii++) { 1940 if (!ISREMOTE(ii)) 1941 continue; 1942 remote_close(gres, ii); 1943 } 1944 if (modified & MODIFIED_REMOTEADDR) { 1945 vstr = nv_get_string(nv, "remoteaddr"); 1946 strlcpy(gres->hr_remoteaddr, vstr, 1947 sizeof(gres->hr_remoteaddr)); 1948 } 1949 } 1950#undef MODIFIED_REMOTEADDR 1951#undef MODIFIED_REPLICATION 1952#undef MODIFIED_TIMEOUT 1953#undef MODIFIED_EXEC 1954 1955 pjdlog_info("Configuration reloaded successfully."); 1956} 1957 1958static void 1959guard_one(struct hast_resource *res, unsigned int ncomp) 1960{ 1961 struct proto_conn *in, *out; 1962 1963 if (!ISREMOTE(ncomp)) 1964 return; 1965 1966 rw_rlock(&hio_remote_lock[ncomp]); 1967 1968 if (!real_remote(res)) { 1969 rw_unlock(&hio_remote_lock[ncomp]); 1970 return; 1971 } 1972 1973 if (ISCONNECTED(res, ncomp)) { 1974 assert(res->hr_remotein != NULL); 1975 assert(res->hr_remoteout != NULL); 1976 rw_unlock(&hio_remote_lock[ncomp]); 1977 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 1978 res->hr_remoteaddr); 1979 return; 1980 } 1981 1982 assert(res->hr_remotein == NULL); 1983 assert(res->hr_remoteout == NULL); 1984 /* 1985 * Upgrade the lock. It doesn't have to be atomic as no other thread 1986 * can change connection status from disconnected to connected. 1987 */ 1988 rw_unlock(&hio_remote_lock[ncomp]); 1989 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 1990 res->hr_remoteaddr); 1991 in = out = NULL; 1992 if (init_remote(res, &in, &out)) { 1993 rw_wlock(&hio_remote_lock[ncomp]); 1994 assert(res->hr_remotein == NULL); 1995 assert(res->hr_remoteout == NULL); 1996 assert(in != NULL && out != NULL); 1997 res->hr_remotein = in; 1998 res->hr_remoteout = out; 1999 rw_unlock(&hio_remote_lock[ncomp]); 2000 pjdlog_info("Successfully reconnected to %s.", 2001 res->hr_remoteaddr); 2002 sync_start(); 2003 } else { 2004 /* Both connections should be NULL. */ 2005 assert(res->hr_remotein == NULL); 2006 assert(res->hr_remoteout == NULL); 2007 assert(in == NULL && out == NULL); 2008 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 2009 res->hr_remoteaddr); 2010 } 2011} 2012 2013/* 2014 * Thread guards remote connections and reconnects when needed, handles 2015 * signals, etc. 2016 */ 2017static void * 2018guard_thread(void *arg) 2019{ 2020 struct hast_resource *res = arg; 2021 unsigned int ii, ncomps; 2022 struct timespec timeout; 2023 time_t lastcheck, now; 2024 sigset_t mask; 2025 int signo; 2026 2027 ncomps = HAST_NCOMPONENTS; 2028 lastcheck = time(NULL); 2029 2030 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 2031 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 2032 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 2033 2034 timeout.tv_sec = RETRY_SLEEP; 2035 timeout.tv_nsec = 0; 2036 signo = -1; 2037 2038 for (;;) { 2039 switch (signo) { 2040 case SIGINT: 2041 case SIGTERM: 2042 sigexit_received = true; 2043 primary_exitx(EX_OK, 2044 "Termination signal received, exiting."); 2045 break; 2046 default: 2047 break; 2048 } 2049 2050 pjdlog_debug(2, "remote_guard: Checking connections."); 2051 now = time(NULL); 2052 if (lastcheck + RETRY_SLEEP <= now) { 2053 for (ii = 0; ii < ncomps; ii++) 2054 guard_one(res, ii); 2055 lastcheck = now; 2056 } 2057 signo = sigtimedwait(&mask, NULL, &timeout); 2058 } 2059 /* NOTREACHED */ 2060 return (NULL); 2061} 2062