primary.c revision 218045
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 218045 2011-01-28 21:57:42Z 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 /* 851 * Create the guard thread first, so we can handle signals from the 852 * very begining. 853 */ 854 error = pthread_create(&td, NULL, guard_thread, res); 855 assert(error == 0); 856 /* 857 * Create the control thread before sending any event to the parent, 858 * as we can deadlock when parent sends control request to worker, 859 * but worker has no control thread started yet, so parent waits. 860 * In the meantime worker sends an event to the parent, but parent 861 * is unable to handle the event, because it waits for control 862 * request response. 863 */ 864 error = pthread_create(&td, NULL, ctrl_thread, res); 865 assert(error == 0); 866 if (real_remote(res) && init_remote(res, NULL, NULL)) 867 sync_start(); 868 error = pthread_create(&td, NULL, ggate_recv_thread, res); 869 assert(error == 0); 870 error = pthread_create(&td, NULL, local_send_thread, res); 871 assert(error == 0); 872 error = pthread_create(&td, NULL, remote_send_thread, res); 873 assert(error == 0); 874 error = pthread_create(&td, NULL, remote_recv_thread, res); 875 assert(error == 0); 876 error = pthread_create(&td, NULL, ggate_send_thread, res); 877 assert(error == 0); 878 (void)sync_thread(res); 879} 880 881static void 882reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 883{ 884 char msg[1024]; 885 va_list ap; 886 int len; 887 888 va_start(ap, fmt); 889 len = vsnprintf(msg, sizeof(msg), fmt, ap); 890 va_end(ap); 891 if ((size_t)len < sizeof(msg)) { 892 switch (ggio->gctl_cmd) { 893 case BIO_READ: 894 (void)snprintf(msg + len, sizeof(msg) - len, 895 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 896 (uintmax_t)ggio->gctl_length); 897 break; 898 case BIO_DELETE: 899 (void)snprintf(msg + len, sizeof(msg) - len, 900 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 901 (uintmax_t)ggio->gctl_length); 902 break; 903 case BIO_FLUSH: 904 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 905 break; 906 case BIO_WRITE: 907 (void)snprintf(msg + len, sizeof(msg) - len, 908 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 909 (uintmax_t)ggio->gctl_length); 910 break; 911 default: 912 (void)snprintf(msg + len, sizeof(msg) - len, 913 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 914 break; 915 } 916 } 917 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 918} 919 920static void 921remote_close(struct hast_resource *res, int ncomp) 922{ 923 924 rw_wlock(&hio_remote_lock[ncomp]); 925 /* 926 * A race is possible between dropping rlock and acquiring wlock - 927 * another thread can close connection in-between. 928 */ 929 if (!ISCONNECTED(res, ncomp)) { 930 assert(res->hr_remotein == NULL); 931 assert(res->hr_remoteout == NULL); 932 rw_unlock(&hio_remote_lock[ncomp]); 933 return; 934 } 935 936 assert(res->hr_remotein != NULL); 937 assert(res->hr_remoteout != NULL); 938 939 pjdlog_debug(2, "Closing incoming connection to %s.", 940 res->hr_remoteaddr); 941 proto_close(res->hr_remotein); 942 res->hr_remotein = NULL; 943 pjdlog_debug(2, "Closing outgoing connection to %s.", 944 res->hr_remoteaddr); 945 proto_close(res->hr_remoteout); 946 res->hr_remoteout = NULL; 947 948 rw_unlock(&hio_remote_lock[ncomp]); 949 950 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 951 952 /* 953 * Stop synchronization if in-progress. 954 */ 955 sync_stop(); 956 957 event_send(res, EVENT_DISCONNECT); 958} 959 960/* 961 * Thread receives ggate I/O requests from the kernel and passes them to 962 * appropriate threads: 963 * WRITE - always goes to both local_send and remote_send threads 964 * READ (when the block is up-to-date on local component) - 965 * only local_send thread 966 * READ (when the block isn't up-to-date on local component) - 967 * only remote_send thread 968 * DELETE - always goes to both local_send and remote_send threads 969 * FLUSH - always goes to both local_send and remote_send threads 970 */ 971static void * 972ggate_recv_thread(void *arg) 973{ 974 struct hast_resource *res = arg; 975 struct g_gate_ctl_io *ggio; 976 struct hio *hio; 977 unsigned int ii, ncomp, ncomps; 978 int error; 979 980 ncomps = HAST_NCOMPONENTS; 981 982 for (;;) { 983 pjdlog_debug(2, "ggate_recv: Taking free request."); 984 QUEUE_TAKE2(hio, free); 985 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 986 ggio = &hio->hio_ggio; 987 ggio->gctl_unit = res->hr_ggateunit; 988 ggio->gctl_length = MAXPHYS; 989 ggio->gctl_error = 0; 990 pjdlog_debug(2, 991 "ggate_recv: (%p) Waiting for request from the kernel.", 992 hio); 993 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 994 if (sigexit_received) 995 pthread_exit(NULL); 996 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 997 } 998 error = ggio->gctl_error; 999 switch (error) { 1000 case 0: 1001 break; 1002 case ECANCELED: 1003 /* Exit gracefully. */ 1004 if (!sigexit_received) { 1005 pjdlog_debug(2, 1006 "ggate_recv: (%p) Received cancel from the kernel.", 1007 hio); 1008 pjdlog_info("Received cancel from the kernel, exiting."); 1009 } 1010 pthread_exit(NULL); 1011 case ENOMEM: 1012 /* 1013 * Buffer too small? Impossible, we allocate MAXPHYS 1014 * bytes - request can't be bigger than that. 1015 */ 1016 /* FALLTHROUGH */ 1017 case ENXIO: 1018 default: 1019 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 1020 strerror(error)); 1021 } 1022 for (ii = 0; ii < ncomps; ii++) 1023 hio->hio_errors[ii] = EINVAL; 1024 reqlog(LOG_DEBUG, 2, ggio, 1025 "ggate_recv: (%p) Request received from the kernel: ", 1026 hio); 1027 /* 1028 * Inform all components about new write request. 1029 * For read request prefer local component unless the given 1030 * range is out-of-date, then use remote component. 1031 */ 1032 switch (ggio->gctl_cmd) { 1033 case BIO_READ: 1034 pjdlog_debug(2, 1035 "ggate_recv: (%p) Moving request to the send queue.", 1036 hio); 1037 refcount_init(&hio->hio_countdown, 1); 1038 mtx_lock(&metadata_lock); 1039 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 1040 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1041 /* 1042 * This range is up-to-date on local component, 1043 * so handle request locally. 1044 */ 1045 /* Local component is 0 for now. */ 1046 ncomp = 0; 1047 } else /* if (res->hr_syncsrc == 1048 HAST_SYNCSRC_SECONDARY) */ { 1049 assert(res->hr_syncsrc == 1050 HAST_SYNCSRC_SECONDARY); 1051 /* 1052 * This range is out-of-date on local component, 1053 * so send request to the remote node. 1054 */ 1055 /* Remote component is 1 for now. */ 1056 ncomp = 1; 1057 } 1058 mtx_unlock(&metadata_lock); 1059 QUEUE_INSERT1(hio, send, ncomp); 1060 break; 1061 case BIO_WRITE: 1062 if (res->hr_resuid == 0) { 1063 /* This is first write, initialize resuid. */ 1064 (void)init_resuid(res); 1065 } 1066 for (;;) { 1067 mtx_lock(&range_lock); 1068 if (rangelock_islocked(range_sync, 1069 ggio->gctl_offset, ggio->gctl_length)) { 1070 pjdlog_debug(2, 1071 "regular: Range offset=%jd length=%zu locked.", 1072 (intmax_t)ggio->gctl_offset, 1073 (size_t)ggio->gctl_length); 1074 range_regular_wait = true; 1075 cv_wait(&range_regular_cond, &range_lock); 1076 range_regular_wait = false; 1077 mtx_unlock(&range_lock); 1078 continue; 1079 } 1080 if (rangelock_add(range_regular, 1081 ggio->gctl_offset, ggio->gctl_length) < 0) { 1082 mtx_unlock(&range_lock); 1083 pjdlog_debug(2, 1084 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1085 (intmax_t)ggio->gctl_offset, 1086 (size_t)ggio->gctl_length); 1087 sleep(1); 1088 continue; 1089 } 1090 mtx_unlock(&range_lock); 1091 break; 1092 } 1093 mtx_lock(&res->hr_amp_lock); 1094 if (activemap_write_start(res->hr_amp, 1095 ggio->gctl_offset, ggio->gctl_length)) { 1096 (void)hast_activemap_flush(res); 1097 } 1098 mtx_unlock(&res->hr_amp_lock); 1099 /* FALLTHROUGH */ 1100 case BIO_DELETE: 1101 case BIO_FLUSH: 1102 pjdlog_debug(2, 1103 "ggate_recv: (%p) Moving request to the send queues.", 1104 hio); 1105 refcount_init(&hio->hio_countdown, ncomps); 1106 for (ii = 0; ii < ncomps; ii++) 1107 QUEUE_INSERT1(hio, send, ii); 1108 break; 1109 } 1110 } 1111 /* NOTREACHED */ 1112 return (NULL); 1113} 1114 1115/* 1116 * Thread reads from or writes to local component. 1117 * If local read fails, it redirects it to remote_send thread. 1118 */ 1119static void * 1120local_send_thread(void *arg) 1121{ 1122 struct hast_resource *res = arg; 1123 struct g_gate_ctl_io *ggio; 1124 struct hio *hio; 1125 unsigned int ncomp, rncomp; 1126 ssize_t ret; 1127 1128 /* Local component is 0 for now. */ 1129 ncomp = 0; 1130 /* Remote component is 1 for now. */ 1131 rncomp = 1; 1132 1133 for (;;) { 1134 pjdlog_debug(2, "local_send: Taking request."); 1135 QUEUE_TAKE1(hio, send, ncomp, 0); 1136 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1137 ggio = &hio->hio_ggio; 1138 switch (ggio->gctl_cmd) { 1139 case BIO_READ: 1140 ret = pread(res->hr_localfd, ggio->gctl_data, 1141 ggio->gctl_length, 1142 ggio->gctl_offset + res->hr_localoff); 1143 if (ret == ggio->gctl_length) 1144 hio->hio_errors[ncomp] = 0; 1145 else { 1146 /* 1147 * If READ failed, try to read from remote node. 1148 */ 1149 if (ret < 0) { 1150 reqlog(LOG_WARNING, 0, ggio, 1151 "Local request failed (%s), trying remote node. ", 1152 strerror(errno)); 1153 } else if (ret != ggio->gctl_length) { 1154 reqlog(LOG_WARNING, 0, ggio, 1155 "Local request failed (%zd != %jd), trying remote node. ", 1156 ret, (intmax_t)ggio->gctl_length); 1157 } 1158 QUEUE_INSERT1(hio, send, rncomp); 1159 continue; 1160 } 1161 break; 1162 case BIO_WRITE: 1163 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1164 ggio->gctl_length, 1165 ggio->gctl_offset + res->hr_localoff); 1166 if (ret < 0) { 1167 hio->hio_errors[ncomp] = errno; 1168 reqlog(LOG_WARNING, 0, ggio, 1169 "Local request failed (%s): ", 1170 strerror(errno)); 1171 } else if (ret != ggio->gctl_length) { 1172 hio->hio_errors[ncomp] = EIO; 1173 reqlog(LOG_WARNING, 0, ggio, 1174 "Local request failed (%zd != %jd): ", 1175 ret, (intmax_t)ggio->gctl_length); 1176 } else { 1177 hio->hio_errors[ncomp] = 0; 1178 } 1179 break; 1180 case BIO_DELETE: 1181 ret = g_delete(res->hr_localfd, 1182 ggio->gctl_offset + res->hr_localoff, 1183 ggio->gctl_length); 1184 if (ret < 0) { 1185 hio->hio_errors[ncomp] = errno; 1186 reqlog(LOG_WARNING, 0, ggio, 1187 "Local request failed (%s): ", 1188 strerror(errno)); 1189 } else { 1190 hio->hio_errors[ncomp] = 0; 1191 } 1192 break; 1193 case BIO_FLUSH: 1194 ret = g_flush(res->hr_localfd); 1195 if (ret < 0) { 1196 hio->hio_errors[ncomp] = errno; 1197 reqlog(LOG_WARNING, 0, ggio, 1198 "Local request failed (%s): ", 1199 strerror(errno)); 1200 } else { 1201 hio->hio_errors[ncomp] = 0; 1202 } 1203 break; 1204 } 1205 if (refcount_release(&hio->hio_countdown)) { 1206 if (ISSYNCREQ(hio)) { 1207 mtx_lock(&sync_lock); 1208 SYNCREQDONE(hio); 1209 mtx_unlock(&sync_lock); 1210 cv_signal(&sync_cond); 1211 } else { 1212 pjdlog_debug(2, 1213 "local_send: (%p) Moving request to the done queue.", 1214 hio); 1215 QUEUE_INSERT2(hio, done); 1216 } 1217 } 1218 } 1219 /* NOTREACHED */ 1220 return (NULL); 1221} 1222 1223static void 1224keepalive_send(struct hast_resource *res, unsigned int ncomp) 1225{ 1226 struct nv *nv; 1227 1228 if (!ISCONNECTED(res, ncomp)) 1229 return; 1230 1231 assert(res->hr_remotein != NULL); 1232 assert(res->hr_remoteout != NULL); 1233 1234 nv = nv_alloc(); 1235 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1236 if (nv_error(nv) != 0) { 1237 nv_free(nv); 1238 pjdlog_debug(1, 1239 "keepalive_send: Unable to prepare header to send."); 1240 return; 1241 } 1242 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1243 pjdlog_common(LOG_DEBUG, 1, errno, 1244 "keepalive_send: Unable to send request"); 1245 nv_free(nv); 1246 rw_unlock(&hio_remote_lock[ncomp]); 1247 remote_close(res, ncomp); 1248 rw_rlock(&hio_remote_lock[ncomp]); 1249 return; 1250 } 1251 nv_free(nv); 1252 pjdlog_debug(2, "keepalive_send: Request sent."); 1253} 1254 1255/* 1256 * Thread sends request to secondary node. 1257 */ 1258static void * 1259remote_send_thread(void *arg) 1260{ 1261 struct hast_resource *res = arg; 1262 struct g_gate_ctl_io *ggio; 1263 time_t lastcheck, now; 1264 struct hio *hio; 1265 struct nv *nv; 1266 unsigned int ncomp; 1267 bool wakeup; 1268 uint64_t offset, length; 1269 uint8_t cmd; 1270 void *data; 1271 1272 /* Remote component is 1 for now. */ 1273 ncomp = 1; 1274 lastcheck = time(NULL); 1275 1276 for (;;) { 1277 pjdlog_debug(2, "remote_send: Taking request."); 1278 QUEUE_TAKE1(hio, send, ncomp, RETRY_SLEEP); 1279 if (hio == NULL) { 1280 now = time(NULL); 1281 if (lastcheck + RETRY_SLEEP <= now) { 1282 keepalive_send(res, ncomp); 1283 lastcheck = now; 1284 } 1285 continue; 1286 } 1287 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1288 ggio = &hio->hio_ggio; 1289 switch (ggio->gctl_cmd) { 1290 case BIO_READ: 1291 cmd = HIO_READ; 1292 data = NULL; 1293 offset = ggio->gctl_offset; 1294 length = ggio->gctl_length; 1295 break; 1296 case BIO_WRITE: 1297 cmd = HIO_WRITE; 1298 data = ggio->gctl_data; 1299 offset = ggio->gctl_offset; 1300 length = ggio->gctl_length; 1301 break; 1302 case BIO_DELETE: 1303 cmd = HIO_DELETE; 1304 data = NULL; 1305 offset = ggio->gctl_offset; 1306 length = ggio->gctl_length; 1307 break; 1308 case BIO_FLUSH: 1309 cmd = HIO_FLUSH; 1310 data = NULL; 1311 offset = 0; 1312 length = 0; 1313 break; 1314 default: 1315 assert(!"invalid condition"); 1316 abort(); 1317 } 1318 nv = nv_alloc(); 1319 nv_add_uint8(nv, cmd, "cmd"); 1320 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1321 nv_add_uint64(nv, offset, "offset"); 1322 nv_add_uint64(nv, length, "length"); 1323 if (nv_error(nv) != 0) { 1324 hio->hio_errors[ncomp] = nv_error(nv); 1325 pjdlog_debug(2, 1326 "remote_send: (%p) Unable to prepare header to send.", 1327 hio); 1328 reqlog(LOG_ERR, 0, ggio, 1329 "Unable to prepare header to send (%s): ", 1330 strerror(nv_error(nv))); 1331 /* Move failed request immediately to the done queue. */ 1332 goto done_queue; 1333 } 1334 pjdlog_debug(2, 1335 "remote_send: (%p) Moving request to the recv queue.", 1336 hio); 1337 /* 1338 * Protect connection from disappearing. 1339 */ 1340 rw_rlock(&hio_remote_lock[ncomp]); 1341 if (!ISCONNECTED(res, ncomp)) { 1342 rw_unlock(&hio_remote_lock[ncomp]); 1343 hio->hio_errors[ncomp] = ENOTCONN; 1344 goto done_queue; 1345 } 1346 /* 1347 * Move the request to recv queue before sending it, because 1348 * in different order we can get reply before we move request 1349 * to recv queue. 1350 */ 1351 mtx_lock(&hio_recv_list_lock[ncomp]); 1352 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1353 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1354 mtx_unlock(&hio_recv_list_lock[ncomp]); 1355 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1356 data != NULL ? length : 0) < 0) { 1357 hio->hio_errors[ncomp] = errno; 1358 rw_unlock(&hio_remote_lock[ncomp]); 1359 pjdlog_debug(2, 1360 "remote_send: (%p) Unable to send request.", hio); 1361 reqlog(LOG_ERR, 0, ggio, 1362 "Unable to send request (%s): ", 1363 strerror(hio->hio_errors[ncomp])); 1364 remote_close(res, ncomp); 1365 /* 1366 * Take request back from the receive queue and move 1367 * it immediately to the done queue. 1368 */ 1369 mtx_lock(&hio_recv_list_lock[ncomp]); 1370 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1371 mtx_unlock(&hio_recv_list_lock[ncomp]); 1372 goto done_queue; 1373 } 1374 rw_unlock(&hio_remote_lock[ncomp]); 1375 nv_free(nv); 1376 if (wakeup) 1377 cv_signal(&hio_recv_list_cond[ncomp]); 1378 continue; 1379done_queue: 1380 nv_free(nv); 1381 if (ISSYNCREQ(hio)) { 1382 if (!refcount_release(&hio->hio_countdown)) 1383 continue; 1384 mtx_lock(&sync_lock); 1385 SYNCREQDONE(hio); 1386 mtx_unlock(&sync_lock); 1387 cv_signal(&sync_cond); 1388 continue; 1389 } 1390 if (ggio->gctl_cmd == BIO_WRITE) { 1391 mtx_lock(&res->hr_amp_lock); 1392 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1393 ggio->gctl_length)) { 1394 (void)hast_activemap_flush(res); 1395 } 1396 mtx_unlock(&res->hr_amp_lock); 1397 } 1398 if (!refcount_release(&hio->hio_countdown)) 1399 continue; 1400 pjdlog_debug(2, 1401 "remote_send: (%p) Moving request to the done queue.", 1402 hio); 1403 QUEUE_INSERT2(hio, done); 1404 } 1405 /* NOTREACHED */ 1406 return (NULL); 1407} 1408 1409/* 1410 * Thread receives answer from secondary node and passes it to ggate_send 1411 * thread. 1412 */ 1413static void * 1414remote_recv_thread(void *arg) 1415{ 1416 struct hast_resource *res = arg; 1417 struct g_gate_ctl_io *ggio; 1418 struct hio *hio; 1419 struct nv *nv; 1420 unsigned int ncomp; 1421 uint64_t seq; 1422 int error; 1423 1424 /* Remote component is 1 for now. */ 1425 ncomp = 1; 1426 1427 for (;;) { 1428 /* Wait until there is anything to receive. */ 1429 mtx_lock(&hio_recv_list_lock[ncomp]); 1430 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1431 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1432 cv_wait(&hio_recv_list_cond[ncomp], 1433 &hio_recv_list_lock[ncomp]); 1434 } 1435 mtx_unlock(&hio_recv_list_lock[ncomp]); 1436 rw_rlock(&hio_remote_lock[ncomp]); 1437 if (!ISCONNECTED(res, ncomp)) { 1438 rw_unlock(&hio_remote_lock[ncomp]); 1439 /* 1440 * Connection is dead, so move all pending requests to 1441 * the done queue (one-by-one). 1442 */ 1443 mtx_lock(&hio_recv_list_lock[ncomp]); 1444 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1445 assert(hio != NULL); 1446 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1447 hio_next[ncomp]); 1448 mtx_unlock(&hio_recv_list_lock[ncomp]); 1449 goto done_queue; 1450 } 1451 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1452 pjdlog_errno(LOG_ERR, 1453 "Unable to receive reply header"); 1454 rw_unlock(&hio_remote_lock[ncomp]); 1455 remote_close(res, ncomp); 1456 continue; 1457 } 1458 rw_unlock(&hio_remote_lock[ncomp]); 1459 seq = nv_get_uint64(nv, "seq"); 1460 if (seq == 0) { 1461 pjdlog_error("Header contains no 'seq' field."); 1462 nv_free(nv); 1463 continue; 1464 } 1465 mtx_lock(&hio_recv_list_lock[ncomp]); 1466 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1467 if (hio->hio_ggio.gctl_seq == seq) { 1468 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1469 hio_next[ncomp]); 1470 break; 1471 } 1472 } 1473 mtx_unlock(&hio_recv_list_lock[ncomp]); 1474 if (hio == NULL) { 1475 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1476 (uintmax_t)seq); 1477 nv_free(nv); 1478 continue; 1479 } 1480 error = nv_get_int16(nv, "error"); 1481 if (error != 0) { 1482 /* Request failed on remote side. */ 1483 hio->hio_errors[ncomp] = error; 1484 reqlog(LOG_WARNING, 0, &hio->hio_ggio, 1485 "Remote request failed (%s): ", strerror(error)); 1486 nv_free(nv); 1487 goto done_queue; 1488 } 1489 ggio = &hio->hio_ggio; 1490 switch (ggio->gctl_cmd) { 1491 case BIO_READ: 1492 rw_rlock(&hio_remote_lock[ncomp]); 1493 if (!ISCONNECTED(res, ncomp)) { 1494 rw_unlock(&hio_remote_lock[ncomp]); 1495 nv_free(nv); 1496 goto done_queue; 1497 } 1498 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1499 ggio->gctl_data, ggio->gctl_length) < 0) { 1500 hio->hio_errors[ncomp] = errno; 1501 pjdlog_errno(LOG_ERR, 1502 "Unable to receive reply data"); 1503 rw_unlock(&hio_remote_lock[ncomp]); 1504 nv_free(nv); 1505 remote_close(res, ncomp); 1506 goto done_queue; 1507 } 1508 rw_unlock(&hio_remote_lock[ncomp]); 1509 break; 1510 case BIO_WRITE: 1511 case BIO_DELETE: 1512 case BIO_FLUSH: 1513 break; 1514 default: 1515 assert(!"invalid condition"); 1516 abort(); 1517 } 1518 hio->hio_errors[ncomp] = 0; 1519 nv_free(nv); 1520done_queue: 1521 if (refcount_release(&hio->hio_countdown)) { 1522 if (ISSYNCREQ(hio)) { 1523 mtx_lock(&sync_lock); 1524 SYNCREQDONE(hio); 1525 mtx_unlock(&sync_lock); 1526 cv_signal(&sync_cond); 1527 } else { 1528 pjdlog_debug(2, 1529 "remote_recv: (%p) Moving request to the done queue.", 1530 hio); 1531 QUEUE_INSERT2(hio, done); 1532 } 1533 } 1534 } 1535 /* NOTREACHED */ 1536 return (NULL); 1537} 1538 1539/* 1540 * Thread sends answer to the kernel. 1541 */ 1542static void * 1543ggate_send_thread(void *arg) 1544{ 1545 struct hast_resource *res = arg; 1546 struct g_gate_ctl_io *ggio; 1547 struct hio *hio; 1548 unsigned int ii, ncomp, ncomps; 1549 1550 ncomps = HAST_NCOMPONENTS; 1551 1552 for (;;) { 1553 pjdlog_debug(2, "ggate_send: Taking request."); 1554 QUEUE_TAKE2(hio, done); 1555 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1556 ggio = &hio->hio_ggio; 1557 for (ii = 0; ii < ncomps; ii++) { 1558 if (hio->hio_errors[ii] == 0) { 1559 /* 1560 * One successful request is enough to declare 1561 * success. 1562 */ 1563 ggio->gctl_error = 0; 1564 break; 1565 } 1566 } 1567 if (ii == ncomps) { 1568 /* 1569 * None of the requests were successful. 1570 * Use first error. 1571 */ 1572 ggio->gctl_error = hio->hio_errors[0]; 1573 } 1574 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1575 mtx_lock(&res->hr_amp_lock); 1576 activemap_write_complete(res->hr_amp, 1577 ggio->gctl_offset, ggio->gctl_length); 1578 mtx_unlock(&res->hr_amp_lock); 1579 } 1580 if (ggio->gctl_cmd == BIO_WRITE) { 1581 /* 1582 * Unlock range we locked. 1583 */ 1584 mtx_lock(&range_lock); 1585 rangelock_del(range_regular, ggio->gctl_offset, 1586 ggio->gctl_length); 1587 if (range_sync_wait) 1588 cv_signal(&range_sync_cond); 1589 mtx_unlock(&range_lock); 1590 /* 1591 * Bump local count if this is first write after 1592 * connection failure with remote node. 1593 */ 1594 ncomp = 1; 1595 rw_rlock(&hio_remote_lock[ncomp]); 1596 if (!ISCONNECTED(res, ncomp)) { 1597 mtx_lock(&metadata_lock); 1598 if (res->hr_primary_localcnt == 1599 res->hr_secondary_remotecnt) { 1600 res->hr_primary_localcnt++; 1601 pjdlog_debug(1, 1602 "Increasing localcnt to %ju.", 1603 (uintmax_t)res->hr_primary_localcnt); 1604 (void)metadata_write(res); 1605 } 1606 mtx_unlock(&metadata_lock); 1607 } 1608 rw_unlock(&hio_remote_lock[ncomp]); 1609 } 1610 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1611 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1612 pjdlog_debug(2, 1613 "ggate_send: (%p) Moving request to the free queue.", hio); 1614 QUEUE_INSERT2(hio, free); 1615 } 1616 /* NOTREACHED */ 1617 return (NULL); 1618} 1619 1620/* 1621 * Thread synchronize local and remote components. 1622 */ 1623static void * 1624sync_thread(void *arg __unused) 1625{ 1626 struct hast_resource *res = arg; 1627 struct hio *hio; 1628 struct g_gate_ctl_io *ggio; 1629 unsigned int ii, ncomp, ncomps; 1630 off_t offset, length, synced; 1631 bool dorewind; 1632 int syncext; 1633 1634 ncomps = HAST_NCOMPONENTS; 1635 dorewind = true; 1636 synced = 0; 1637 offset = -1; 1638 1639 for (;;) { 1640 mtx_lock(&sync_lock); 1641 if (offset >= 0 && !sync_inprogress) { 1642 pjdlog_info("Synchronization interrupted. " 1643 "%jd bytes synchronized so far.", 1644 (intmax_t)synced); 1645 event_send(res, EVENT_SYNCINTR); 1646 } 1647 while (!sync_inprogress) { 1648 dorewind = true; 1649 synced = 0; 1650 cv_wait(&sync_cond, &sync_lock); 1651 } 1652 mtx_unlock(&sync_lock); 1653 /* 1654 * Obtain offset at which we should synchronize. 1655 * Rewind synchronization if needed. 1656 */ 1657 mtx_lock(&res->hr_amp_lock); 1658 if (dorewind) 1659 activemap_sync_rewind(res->hr_amp); 1660 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1661 if (syncext != -1) { 1662 /* 1663 * We synchronized entire syncext extent, we can mark 1664 * it as clean now. 1665 */ 1666 if (activemap_extent_complete(res->hr_amp, syncext)) 1667 (void)hast_activemap_flush(res); 1668 } 1669 mtx_unlock(&res->hr_amp_lock); 1670 if (dorewind) { 1671 dorewind = false; 1672 if (offset < 0) 1673 pjdlog_info("Nodes are in sync."); 1674 else { 1675 pjdlog_info("Synchronization started. %ju bytes to go.", 1676 (uintmax_t)(res->hr_extentsize * 1677 activemap_ndirty(res->hr_amp))); 1678 event_send(res, EVENT_SYNCSTART); 1679 } 1680 } 1681 if (offset < 0) { 1682 sync_stop(); 1683 pjdlog_debug(1, "Nothing to synchronize."); 1684 /* 1685 * Synchronization complete, make both localcnt and 1686 * remotecnt equal. 1687 */ 1688 ncomp = 1; 1689 rw_rlock(&hio_remote_lock[ncomp]); 1690 if (ISCONNECTED(res, ncomp)) { 1691 if (synced > 0) { 1692 pjdlog_info("Synchronization complete. " 1693 "%jd bytes synchronized.", 1694 (intmax_t)synced); 1695 event_send(res, EVENT_SYNCDONE); 1696 } 1697 mtx_lock(&metadata_lock); 1698 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1699 res->hr_primary_localcnt = 1700 res->hr_secondary_localcnt; 1701 res->hr_primary_remotecnt = 1702 res->hr_secondary_remotecnt; 1703 pjdlog_debug(1, 1704 "Setting localcnt to %ju and remotecnt to %ju.", 1705 (uintmax_t)res->hr_primary_localcnt, 1706 (uintmax_t)res->hr_secondary_localcnt); 1707 (void)metadata_write(res); 1708 mtx_unlock(&metadata_lock); 1709 } 1710 rw_unlock(&hio_remote_lock[ncomp]); 1711 continue; 1712 } 1713 pjdlog_debug(2, "sync: Taking free request."); 1714 QUEUE_TAKE2(hio, free); 1715 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1716 /* 1717 * Lock the range we are going to synchronize. We don't want 1718 * race where someone writes between our read and write. 1719 */ 1720 for (;;) { 1721 mtx_lock(&range_lock); 1722 if (rangelock_islocked(range_regular, offset, length)) { 1723 pjdlog_debug(2, 1724 "sync: Range offset=%jd length=%jd locked.", 1725 (intmax_t)offset, (intmax_t)length); 1726 range_sync_wait = true; 1727 cv_wait(&range_sync_cond, &range_lock); 1728 range_sync_wait = false; 1729 mtx_unlock(&range_lock); 1730 continue; 1731 } 1732 if (rangelock_add(range_sync, offset, length) < 0) { 1733 mtx_unlock(&range_lock); 1734 pjdlog_debug(2, 1735 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1736 (intmax_t)offset, (intmax_t)length); 1737 sleep(1); 1738 continue; 1739 } 1740 mtx_unlock(&range_lock); 1741 break; 1742 } 1743 /* 1744 * First read the data from synchronization source. 1745 */ 1746 SYNCREQ(hio); 1747 ggio = &hio->hio_ggio; 1748 ggio->gctl_cmd = BIO_READ; 1749 ggio->gctl_offset = offset; 1750 ggio->gctl_length = length; 1751 ggio->gctl_error = 0; 1752 for (ii = 0; ii < ncomps; ii++) 1753 hio->hio_errors[ii] = EINVAL; 1754 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1755 hio); 1756 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1757 hio); 1758 mtx_lock(&metadata_lock); 1759 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1760 /* 1761 * This range is up-to-date on local component, 1762 * so handle request locally. 1763 */ 1764 /* Local component is 0 for now. */ 1765 ncomp = 0; 1766 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1767 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1768 /* 1769 * This range is out-of-date on local component, 1770 * so send request to the remote node. 1771 */ 1772 /* Remote component is 1 for now. */ 1773 ncomp = 1; 1774 } 1775 mtx_unlock(&metadata_lock); 1776 refcount_init(&hio->hio_countdown, 1); 1777 QUEUE_INSERT1(hio, send, ncomp); 1778 1779 /* 1780 * Let's wait for READ to finish. 1781 */ 1782 mtx_lock(&sync_lock); 1783 while (!ISSYNCREQDONE(hio)) 1784 cv_wait(&sync_cond, &sync_lock); 1785 mtx_unlock(&sync_lock); 1786 1787 if (hio->hio_errors[ncomp] != 0) { 1788 pjdlog_error("Unable to read synchronization data: %s.", 1789 strerror(hio->hio_errors[ncomp])); 1790 goto free_queue; 1791 } 1792 1793 /* 1794 * We read the data from synchronization source, now write it 1795 * to synchronization target. 1796 */ 1797 SYNCREQ(hio); 1798 ggio->gctl_cmd = BIO_WRITE; 1799 for (ii = 0; ii < ncomps; ii++) 1800 hio->hio_errors[ii] = EINVAL; 1801 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1802 hio); 1803 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1804 hio); 1805 mtx_lock(&metadata_lock); 1806 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1807 /* 1808 * This range is up-to-date on local component, 1809 * so we update remote component. 1810 */ 1811 /* Remote component is 1 for now. */ 1812 ncomp = 1; 1813 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1814 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1815 /* 1816 * This range is out-of-date on local component, 1817 * so we update it. 1818 */ 1819 /* Local component is 0 for now. */ 1820 ncomp = 0; 1821 } 1822 mtx_unlock(&metadata_lock); 1823 1824 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1825 hio); 1826 refcount_init(&hio->hio_countdown, 1); 1827 QUEUE_INSERT1(hio, send, ncomp); 1828 1829 /* 1830 * Let's wait for WRITE to finish. 1831 */ 1832 mtx_lock(&sync_lock); 1833 while (!ISSYNCREQDONE(hio)) 1834 cv_wait(&sync_cond, &sync_lock); 1835 mtx_unlock(&sync_lock); 1836 1837 if (hio->hio_errors[ncomp] != 0) { 1838 pjdlog_error("Unable to write synchronization data: %s.", 1839 strerror(hio->hio_errors[ncomp])); 1840 goto free_queue; 1841 } 1842 1843 synced += length; 1844free_queue: 1845 mtx_lock(&range_lock); 1846 rangelock_del(range_sync, offset, length); 1847 if (range_regular_wait) 1848 cv_signal(&range_regular_cond); 1849 mtx_unlock(&range_lock); 1850 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1851 hio); 1852 QUEUE_INSERT2(hio, free); 1853 } 1854 /* NOTREACHED */ 1855 return (NULL); 1856} 1857 1858void 1859primary_config_reload(struct hast_resource *res, struct nv *nv) 1860{ 1861 unsigned int ii, ncomps; 1862 int modified, vint; 1863 const char *vstr; 1864 1865 pjdlog_info("Reloading configuration..."); 1866 1867 assert(res->hr_role == HAST_ROLE_PRIMARY); 1868 assert(gres == res); 1869 nv_assert(nv, "remoteaddr"); 1870 nv_assert(nv, "replication"); 1871 nv_assert(nv, "timeout"); 1872 nv_assert(nv, "exec"); 1873 1874 ncomps = HAST_NCOMPONENTS; 1875 1876#define MODIFIED_REMOTEADDR 0x1 1877#define MODIFIED_REPLICATION 0x2 1878#define MODIFIED_TIMEOUT 0x4 1879#define MODIFIED_EXEC 0x8 1880 modified = 0; 1881 1882 vstr = nv_get_string(nv, "remoteaddr"); 1883 if (strcmp(gres->hr_remoteaddr, vstr) != 0) { 1884 /* 1885 * Don't copy res->hr_remoteaddr to gres just yet. 1886 * We want remote_close() to log disconnect from the old 1887 * addresses, not from the new ones. 1888 */ 1889 modified |= MODIFIED_REMOTEADDR; 1890 } 1891 vint = nv_get_int32(nv, "replication"); 1892 if (gres->hr_replication != vint) { 1893 gres->hr_replication = vint; 1894 modified |= MODIFIED_REPLICATION; 1895 } 1896 vint = nv_get_int32(nv, "timeout"); 1897 if (gres->hr_timeout != vint) { 1898 gres->hr_timeout = vint; 1899 modified |= MODIFIED_TIMEOUT; 1900 } 1901 vstr = nv_get_string(nv, "exec"); 1902 if (strcmp(gres->hr_exec, vstr) != 0) { 1903 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec)); 1904 modified |= MODIFIED_EXEC; 1905 } 1906 1907 /* 1908 * If only timeout was modified we only need to change it without 1909 * reconnecting. 1910 */ 1911 if (modified == MODIFIED_TIMEOUT) { 1912 for (ii = 0; ii < ncomps; ii++) { 1913 if (!ISREMOTE(ii)) 1914 continue; 1915 rw_rlock(&hio_remote_lock[ii]); 1916 if (!ISCONNECTED(gres, ii)) { 1917 rw_unlock(&hio_remote_lock[ii]); 1918 continue; 1919 } 1920 rw_unlock(&hio_remote_lock[ii]); 1921 if (proto_timeout(gres->hr_remotein, 1922 gres->hr_timeout) < 0) { 1923 pjdlog_errno(LOG_WARNING, 1924 "Unable to set connection timeout"); 1925 } 1926 if (proto_timeout(gres->hr_remoteout, 1927 gres->hr_timeout) < 0) { 1928 pjdlog_errno(LOG_WARNING, 1929 "Unable to set connection timeout"); 1930 } 1931 } 1932 } else if ((modified & 1933 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) { 1934 for (ii = 0; ii < ncomps; ii++) { 1935 if (!ISREMOTE(ii)) 1936 continue; 1937 remote_close(gres, ii); 1938 } 1939 if (modified & MODIFIED_REMOTEADDR) { 1940 vstr = nv_get_string(nv, "remoteaddr"); 1941 strlcpy(gres->hr_remoteaddr, vstr, 1942 sizeof(gres->hr_remoteaddr)); 1943 } 1944 } 1945#undef MODIFIED_REMOTEADDR 1946#undef MODIFIED_REPLICATION 1947#undef MODIFIED_TIMEOUT 1948#undef MODIFIED_EXEC 1949 1950 pjdlog_info("Configuration reloaded successfully."); 1951} 1952 1953static void 1954guard_one(struct hast_resource *res, unsigned int ncomp) 1955{ 1956 struct proto_conn *in, *out; 1957 1958 if (!ISREMOTE(ncomp)) 1959 return; 1960 1961 rw_rlock(&hio_remote_lock[ncomp]); 1962 1963 if (!real_remote(res)) { 1964 rw_unlock(&hio_remote_lock[ncomp]); 1965 return; 1966 } 1967 1968 if (ISCONNECTED(res, ncomp)) { 1969 assert(res->hr_remotein != NULL); 1970 assert(res->hr_remoteout != NULL); 1971 rw_unlock(&hio_remote_lock[ncomp]); 1972 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 1973 res->hr_remoteaddr); 1974 return; 1975 } 1976 1977 assert(res->hr_remotein == NULL); 1978 assert(res->hr_remoteout == NULL); 1979 /* 1980 * Upgrade the lock. It doesn't have to be atomic as no other thread 1981 * can change connection status from disconnected to connected. 1982 */ 1983 rw_unlock(&hio_remote_lock[ncomp]); 1984 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 1985 res->hr_remoteaddr); 1986 in = out = NULL; 1987 if (init_remote(res, &in, &out)) { 1988 rw_wlock(&hio_remote_lock[ncomp]); 1989 assert(res->hr_remotein == NULL); 1990 assert(res->hr_remoteout == NULL); 1991 assert(in != NULL && out != NULL); 1992 res->hr_remotein = in; 1993 res->hr_remoteout = out; 1994 rw_unlock(&hio_remote_lock[ncomp]); 1995 pjdlog_info("Successfully reconnected to %s.", 1996 res->hr_remoteaddr); 1997 sync_start(); 1998 } else { 1999 /* Both connections should be NULL. */ 2000 assert(res->hr_remotein == NULL); 2001 assert(res->hr_remoteout == NULL); 2002 assert(in == NULL && out == NULL); 2003 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 2004 res->hr_remoteaddr); 2005 } 2006} 2007 2008/* 2009 * Thread guards remote connections and reconnects when needed, handles 2010 * signals, etc. 2011 */ 2012static void * 2013guard_thread(void *arg) 2014{ 2015 struct hast_resource *res = arg; 2016 unsigned int ii, ncomps; 2017 struct timespec timeout; 2018 time_t lastcheck, now; 2019 sigset_t mask; 2020 int signo; 2021 2022 ncomps = HAST_NCOMPONENTS; 2023 lastcheck = time(NULL); 2024 2025 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 2026 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 2027 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 2028 2029 timeout.tv_sec = RETRY_SLEEP; 2030 timeout.tv_nsec = 0; 2031 signo = -1; 2032 2033 for (;;) { 2034 switch (signo) { 2035 case SIGINT: 2036 case SIGTERM: 2037 sigexit_received = true; 2038 primary_exitx(EX_OK, 2039 "Termination signal received, exiting."); 2040 break; 2041 default: 2042 break; 2043 } 2044 2045 pjdlog_debug(2, "remote_guard: Checking connections."); 2046 now = time(NULL); 2047 if (lastcheck + RETRY_SLEEP <= now) { 2048 for (ii = 0; ii < ncomps; ii++) 2049 guard_one(res, ii); 2050 lastcheck = now; 2051 } 2052 signo = sigtimedwait(&mask, NULL, &timeout); 2053 } 2054 /* NOTREACHED */ 2055 return (NULL); 2056} 2057