primary.c revision 212899
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 212899 2010-09-20 13:23:43Z 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) do { \ 184 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 185 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL) { \ 186 cv_wait(&hio_##name##_list_cond[(ncomp)], \ 187 &hio_##name##_list_lock[(ncomp)]); \ 188 } \ 189 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \ 190 hio_next[(ncomp)]); \ 191 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 192} while (0) 193#define QUEUE_TAKE2(hio, name) do { \ 194 mtx_lock(&hio_##name##_list_lock); \ 195 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 196 cv_wait(&hio_##name##_list_cond, \ 197 &hio_##name##_list_lock); \ 198 } \ 199 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \ 200 mtx_unlock(&hio_##name##_list_lock); \ 201} while (0) 202 203#define SYNCREQ(hio) do { \ 204 (hio)->hio_ggio.gctl_unit = -1; \ 205 (hio)->hio_ggio.gctl_seq = 1; \ 206} while (0) 207#define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1) 208#define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0) 209#define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2) 210 211static struct hast_resource *gres; 212 213static pthread_mutex_t range_lock; 214static struct rangelocks *range_regular; 215static bool range_regular_wait; 216static pthread_cond_t range_regular_cond; 217static struct rangelocks *range_sync; 218static bool range_sync_wait; 219static pthread_cond_t range_sync_cond; 220 221static void *ggate_recv_thread(void *arg); 222static void *local_send_thread(void *arg); 223static void *remote_send_thread(void *arg); 224static void *remote_recv_thread(void *arg); 225static void *ggate_send_thread(void *arg); 226static void *sync_thread(void *arg); 227static void *guard_thread(void *arg); 228 229static void 230cleanup(struct hast_resource *res) 231{ 232 int rerrno; 233 234 /* Remember errno. */ 235 rerrno = errno; 236 237 /* 238 * Close descriptor to /dev/hast/<name> 239 * to work-around race in the kernel. 240 */ 241 close(res->hr_localfd); 242 243 /* Destroy ggate provider if we created one. */ 244 if (res->hr_ggateunit >= 0) { 245 struct g_gate_ctl_destroy ggiod; 246 247 ggiod.gctl_version = G_GATE_VERSION; 248 ggiod.gctl_unit = res->hr_ggateunit; 249 ggiod.gctl_force = 1; 250 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) { 251 pjdlog_warning("Unable to destroy hast/%s device", 252 res->hr_provname); 253 } 254 res->hr_ggateunit = -1; 255 } 256 257 /* Restore errno. */ 258 errno = rerrno; 259} 260 261static __dead2 void 262primary_exit(int exitcode, const char *fmt, ...) 263{ 264 va_list ap; 265 266 assert(exitcode != EX_OK); 267 va_start(ap, fmt); 268 pjdlogv_errno(LOG_ERR, fmt, ap); 269 va_end(ap); 270 cleanup(gres); 271 exit(exitcode); 272} 273 274static __dead2 void 275primary_exitx(int exitcode, const char *fmt, ...) 276{ 277 va_list ap; 278 279 va_start(ap, fmt); 280 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 281 va_end(ap); 282 cleanup(gres); 283 exit(exitcode); 284} 285 286static int 287hast_activemap_flush(struct hast_resource *res) 288{ 289 const unsigned char *buf; 290 size_t size; 291 292 buf = activemap_bitmap(res->hr_amp, &size); 293 assert(buf != NULL); 294 assert((size % res->hr_local_sectorsize) == 0); 295 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 296 (ssize_t)size) { 297 KEEP_ERRNO(pjdlog_errno(LOG_ERR, 298 "Unable to flush activemap to disk")); 299 return (-1); 300 } 301 return (0); 302} 303 304static bool 305real_remote(const struct hast_resource *res) 306{ 307 308 return (strcmp(res->hr_remoteaddr, "none") != 0); 309} 310 311static void 312init_environment(struct hast_resource *res __unused) 313{ 314 struct hio *hio; 315 unsigned int ii, ncomps; 316 sigset_t mask; 317 318 /* 319 * In the future it might be per-resource value. 320 */ 321 ncomps = HAST_NCOMPONENTS; 322 323 /* 324 * Allocate memory needed by lists. 325 */ 326 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 327 if (hio_send_list == NULL) { 328 primary_exitx(EX_TEMPFAIL, 329 "Unable to allocate %zu bytes of memory for send lists.", 330 sizeof(hio_send_list[0]) * ncomps); 331 } 332 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 333 if (hio_send_list_lock == NULL) { 334 primary_exitx(EX_TEMPFAIL, 335 "Unable to allocate %zu bytes of memory for send list locks.", 336 sizeof(hio_send_list_lock[0]) * ncomps); 337 } 338 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 339 if (hio_send_list_cond == NULL) { 340 primary_exitx(EX_TEMPFAIL, 341 "Unable to allocate %zu bytes of memory for send list condition variables.", 342 sizeof(hio_send_list_cond[0]) * ncomps); 343 } 344 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 345 if (hio_recv_list == NULL) { 346 primary_exitx(EX_TEMPFAIL, 347 "Unable to allocate %zu bytes of memory for recv lists.", 348 sizeof(hio_recv_list[0]) * ncomps); 349 } 350 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 351 if (hio_recv_list_lock == NULL) { 352 primary_exitx(EX_TEMPFAIL, 353 "Unable to allocate %zu bytes of memory for recv list locks.", 354 sizeof(hio_recv_list_lock[0]) * ncomps); 355 } 356 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 357 if (hio_recv_list_cond == NULL) { 358 primary_exitx(EX_TEMPFAIL, 359 "Unable to allocate %zu bytes of memory for recv list condition variables.", 360 sizeof(hio_recv_list_cond[0]) * ncomps); 361 } 362 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 363 if (hio_remote_lock == NULL) { 364 primary_exitx(EX_TEMPFAIL, 365 "Unable to allocate %zu bytes of memory for remote connections locks.", 366 sizeof(hio_remote_lock[0]) * ncomps); 367 } 368 369 /* 370 * Initialize lists, their locks and theirs condition variables. 371 */ 372 TAILQ_INIT(&hio_free_list); 373 mtx_init(&hio_free_list_lock); 374 cv_init(&hio_free_list_cond); 375 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 376 TAILQ_INIT(&hio_send_list[ii]); 377 mtx_init(&hio_send_list_lock[ii]); 378 cv_init(&hio_send_list_cond[ii]); 379 TAILQ_INIT(&hio_recv_list[ii]); 380 mtx_init(&hio_recv_list_lock[ii]); 381 cv_init(&hio_recv_list_cond[ii]); 382 rw_init(&hio_remote_lock[ii]); 383 } 384 TAILQ_INIT(&hio_done_list); 385 mtx_init(&hio_done_list_lock); 386 cv_init(&hio_done_list_cond); 387 mtx_init(&metadata_lock); 388 389 /* 390 * Allocate requests pool and initialize requests. 391 */ 392 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 393 hio = malloc(sizeof(*hio)); 394 if (hio == NULL) { 395 primary_exitx(EX_TEMPFAIL, 396 "Unable to allocate %zu bytes of memory for hio request.", 397 sizeof(*hio)); 398 } 399 hio->hio_countdown = 0; 400 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 401 if (hio->hio_errors == NULL) { 402 primary_exitx(EX_TEMPFAIL, 403 "Unable allocate %zu bytes of memory for hio errors.", 404 sizeof(hio->hio_errors[0]) * ncomps); 405 } 406 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 407 if (hio->hio_next == NULL) { 408 primary_exitx(EX_TEMPFAIL, 409 "Unable allocate %zu bytes of memory for hio_next field.", 410 sizeof(hio->hio_next[0]) * ncomps); 411 } 412 hio->hio_ggio.gctl_version = G_GATE_VERSION; 413 hio->hio_ggio.gctl_data = malloc(MAXPHYS); 414 if (hio->hio_ggio.gctl_data == NULL) { 415 primary_exitx(EX_TEMPFAIL, 416 "Unable to allocate %zu bytes of memory for gctl_data.", 417 MAXPHYS); 418 } 419 hio->hio_ggio.gctl_length = MAXPHYS; 420 hio->hio_ggio.gctl_error = 0; 421 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 422 } 423 424 /* 425 * Turn on signals handling. 426 */ 427 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 428 PJDLOG_VERIFY(sigaddset(&mask, SIGHUP) == 0); 429 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 430 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 431 PJDLOG_VERIFY(sigprocmask(SIG_SETMASK, &mask, NULL) == 0); 432} 433 434static void 435init_local(struct hast_resource *res) 436{ 437 unsigned char *buf; 438 size_t mapsize; 439 440 if (metadata_read(res, true) < 0) 441 exit(EX_NOINPUT); 442 mtx_init(&res->hr_amp_lock); 443 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 444 res->hr_local_sectorsize, res->hr_keepdirty) < 0) { 445 primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 446 } 447 mtx_init(&range_lock); 448 cv_init(&range_regular_cond); 449 if (rangelock_init(&range_regular) < 0) 450 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 451 cv_init(&range_sync_cond); 452 if (rangelock_init(&range_sync) < 0) 453 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 454 mapsize = activemap_ondisk_size(res->hr_amp); 455 buf = calloc(1, mapsize); 456 if (buf == NULL) { 457 primary_exitx(EX_TEMPFAIL, 458 "Unable to allocate buffer for activemap."); 459 } 460 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 461 (ssize_t)mapsize) { 462 primary_exit(EX_NOINPUT, "Unable to read activemap"); 463 } 464 activemap_copyin(res->hr_amp, buf, mapsize); 465 free(buf); 466 if (res->hr_resuid != 0) 467 return; 468 /* 469 * We're using provider for the first time, so we have to generate 470 * resource unique identifier and initialize local and remote counts. 471 */ 472 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 473 res->hr_primary_localcnt = 1; 474 res->hr_primary_remotecnt = 0; 475 if (metadata_write(res) < 0) 476 exit(EX_NOINPUT); 477} 478 479static bool 480init_remote(struct hast_resource *res, struct proto_conn **inp, 481 struct proto_conn **outp) 482{ 483 struct proto_conn *in, *out; 484 struct nv *nvout, *nvin; 485 const unsigned char *token; 486 unsigned char *map; 487 const char *errmsg; 488 int32_t extentsize; 489 int64_t datasize; 490 uint32_t mapsize; 491 size_t size; 492 493 assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 494 assert(real_remote(res)); 495 496 in = out = NULL; 497 errmsg = NULL; 498 499 /* Prepare outgoing connection with remote node. */ 500 if (proto_client(res->hr_remoteaddr, &out) < 0) { 501 primary_exit(EX_TEMPFAIL, "Unable to create connection to %s", 502 res->hr_remoteaddr); 503 } 504 /* Try to connect, but accept failure. */ 505 if (proto_connect(out) < 0) { 506 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 507 res->hr_remoteaddr); 508 goto close; 509 } 510 /* Error in setting timeout is not critical, but why should it fail? */ 511 if (proto_timeout(out, res->hr_timeout) < 0) 512 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 513 /* 514 * First handshake step. 515 * Setup outgoing connection with remote node. 516 */ 517 nvout = nv_alloc(); 518 nv_add_string(nvout, res->hr_name, "resource"); 519 if (nv_error(nvout) != 0) { 520 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 521 "Unable to allocate header for connection with %s", 522 res->hr_remoteaddr); 523 nv_free(nvout); 524 goto close; 525 } 526 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) { 527 pjdlog_errno(LOG_WARNING, 528 "Unable to send handshake header to %s", 529 res->hr_remoteaddr); 530 nv_free(nvout); 531 goto close; 532 } 533 nv_free(nvout); 534 if (hast_proto_recv_hdr(out, &nvin) < 0) { 535 pjdlog_errno(LOG_WARNING, 536 "Unable to receive handshake header from %s", 537 res->hr_remoteaddr); 538 goto close; 539 } 540 errmsg = nv_get_string(nvin, "errmsg"); 541 if (errmsg != NULL) { 542 pjdlog_warning("%s", errmsg); 543 nv_free(nvin); 544 goto close; 545 } 546 token = nv_get_uint8_array(nvin, &size, "token"); 547 if (token == NULL) { 548 pjdlog_warning("Handshake header from %s has no 'token' field.", 549 res->hr_remoteaddr); 550 nv_free(nvin); 551 goto close; 552 } 553 if (size != sizeof(res->hr_token)) { 554 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 555 res->hr_remoteaddr, size, sizeof(res->hr_token)); 556 nv_free(nvin); 557 goto close; 558 } 559 bcopy(token, res->hr_token, sizeof(res->hr_token)); 560 nv_free(nvin); 561 562 /* 563 * Second handshake step. 564 * Setup incoming connection with remote node. 565 */ 566 if (proto_client(res->hr_remoteaddr, &in) < 0) { 567 pjdlog_errno(LOG_WARNING, "Unable to create connection to %s", 568 res->hr_remoteaddr); 569 } 570 /* Try to connect, but accept failure. */ 571 if (proto_connect(in) < 0) { 572 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 573 res->hr_remoteaddr); 574 goto close; 575 } 576 /* Error in setting timeout is not critical, but why should it fail? */ 577 if (proto_timeout(in, res->hr_timeout) < 0) 578 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 579 nvout = nv_alloc(); 580 nv_add_string(nvout, res->hr_name, "resource"); 581 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 582 "token"); 583 nv_add_uint64(nvout, res->hr_resuid, "resuid"); 584 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 585 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 586 if (nv_error(nvout) != 0) { 587 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 588 "Unable to allocate header for connection with %s", 589 res->hr_remoteaddr); 590 nv_free(nvout); 591 goto close; 592 } 593 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) { 594 pjdlog_errno(LOG_WARNING, 595 "Unable to send handshake header to %s", 596 res->hr_remoteaddr); 597 nv_free(nvout); 598 goto close; 599 } 600 nv_free(nvout); 601 if (hast_proto_recv_hdr(out, &nvin) < 0) { 602 pjdlog_errno(LOG_WARNING, 603 "Unable to receive handshake header from %s", 604 res->hr_remoteaddr); 605 goto close; 606 } 607 errmsg = nv_get_string(nvin, "errmsg"); 608 if (errmsg != NULL) { 609 pjdlog_warning("%s", errmsg); 610 nv_free(nvin); 611 goto close; 612 } 613 datasize = nv_get_int64(nvin, "datasize"); 614 if (datasize != res->hr_datasize) { 615 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 616 (intmax_t)res->hr_datasize, (intmax_t)datasize); 617 nv_free(nvin); 618 goto close; 619 } 620 extentsize = nv_get_int32(nvin, "extentsize"); 621 if (extentsize != res->hr_extentsize) { 622 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 623 (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 624 nv_free(nvin); 625 goto close; 626 } 627 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 628 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 629 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 630 map = NULL; 631 mapsize = nv_get_uint32(nvin, "mapsize"); 632 if (mapsize > 0) { 633 map = malloc(mapsize); 634 if (map == NULL) { 635 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 636 (uintmax_t)mapsize); 637 nv_free(nvin); 638 goto close; 639 } 640 /* 641 * Remote node have some dirty extents on its own, lets 642 * download its activemap. 643 */ 644 if (hast_proto_recv_data(res, out, nvin, map, 645 mapsize) < 0) { 646 pjdlog_errno(LOG_ERR, 647 "Unable to receive remote activemap"); 648 nv_free(nvin); 649 free(map); 650 goto close; 651 } 652 /* 653 * Merge local and remote bitmaps. 654 */ 655 activemap_merge(res->hr_amp, map, mapsize); 656 free(map); 657 /* 658 * Now that we merged bitmaps from both nodes, flush it to the 659 * disk before we start to synchronize. 660 */ 661 (void)hast_activemap_flush(res); 662 } 663 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 664 if (inp != NULL && outp != NULL) { 665 *inp = in; 666 *outp = out; 667 } else { 668 res->hr_remotein = in; 669 res->hr_remoteout = out; 670 } 671 event_send(res, EVENT_CONNECT); 672 return (true); 673close: 674 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 675 event_send(res, EVENT_SPLITBRAIN); 676 proto_close(out); 677 if (in != NULL) 678 proto_close(in); 679 return (false); 680} 681 682static void 683sync_start(void) 684{ 685 686 mtx_lock(&sync_lock); 687 sync_inprogress = true; 688 mtx_unlock(&sync_lock); 689 cv_signal(&sync_cond); 690} 691 692static void 693sync_stop(void) 694{ 695 696 mtx_lock(&sync_lock); 697 if (sync_inprogress) 698 sync_inprogress = false; 699 mtx_unlock(&sync_lock); 700} 701 702static void 703init_ggate(struct hast_resource *res) 704{ 705 struct g_gate_ctl_create ggiocreate; 706 struct g_gate_ctl_cancel ggiocancel; 707 708 /* 709 * We communicate with ggate via /dev/ggctl. Open it. 710 */ 711 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 712 if (res->hr_ggatefd < 0) 713 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 714 /* 715 * Create provider before trying to connect, as connection failure 716 * is not critical, but may take some time. 717 */ 718 ggiocreate.gctl_version = G_GATE_VERSION; 719 ggiocreate.gctl_mediasize = res->hr_datasize; 720 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 721 ggiocreate.gctl_flags = 0; 722 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE; 723 ggiocreate.gctl_timeout = 0; 724 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 725 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 726 res->hr_provname); 727 bzero(ggiocreate.gctl_info, sizeof(ggiocreate.gctl_info)); 728 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 729 pjdlog_info("Device hast/%s created.", res->hr_provname); 730 res->hr_ggateunit = ggiocreate.gctl_unit; 731 return; 732 } 733 if (errno != EEXIST) { 734 primary_exit(EX_OSERR, "Unable to create hast/%s device", 735 res->hr_provname); 736 } 737 pjdlog_debug(1, 738 "Device hast/%s already exists, we will try to take it over.", 739 res->hr_provname); 740 /* 741 * If we received EEXIST, we assume that the process who created the 742 * provider died and didn't clean up. In that case we will start from 743 * where he left of. 744 */ 745 ggiocancel.gctl_version = G_GATE_VERSION; 746 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 747 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 748 res->hr_provname); 749 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 750 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 751 res->hr_ggateunit = ggiocancel.gctl_unit; 752 return; 753 } 754 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 755 res->hr_provname); 756} 757 758void 759hastd_primary(struct hast_resource *res) 760{ 761 pthread_t td; 762 pid_t pid; 763 int error; 764 765 /* 766 * Create communication channel between parent and child. 767 */ 768 if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 769 KEEP_ERRNO((void)pidfile_remove(pfh)); 770 pjdlog_exit(EX_OSERR, 771 "Unable to create control sockets between parent and child"); 772 } 773 /* 774 * Create communication channel between child and parent. 775 */ 776 if (proto_client("socketpair://", &res->hr_event) < 0) { 777 KEEP_ERRNO((void)pidfile_remove(pfh)); 778 pjdlog_exit(EX_OSERR, 779 "Unable to create event sockets between child and parent"); 780 } 781 782 pid = fork(); 783 if (pid < 0) { 784 KEEP_ERRNO((void)pidfile_remove(pfh)); 785 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 786 } 787 788 if (pid > 0) { 789 /* This is parent. */ 790 /* Declare that we are receiver. */ 791 proto_recv(res->hr_event, NULL, 0); 792 res->hr_workerpid = pid; 793 return; 794 } 795 796 gres = res; 797 798 (void)pidfile_close(pfh); 799 hook_fini(); 800 801 setproctitle("%s (primary)", res->hr_name); 802 803 signal(SIGHUP, SIG_DFL); 804 signal(SIGCHLD, SIG_DFL); 805 806 /* Declare that we are sender. */ 807 proto_send(res->hr_event, NULL, 0); 808 809 init_local(res); 810 if (real_remote(res) && init_remote(res, NULL, NULL)) 811 sync_start(); 812 init_ggate(res); 813 init_environment(res); 814 error = pthread_create(&td, NULL, ggate_recv_thread, res); 815 assert(error == 0); 816 error = pthread_create(&td, NULL, local_send_thread, res); 817 assert(error == 0); 818 error = pthread_create(&td, NULL, remote_send_thread, res); 819 assert(error == 0); 820 error = pthread_create(&td, NULL, remote_recv_thread, res); 821 assert(error == 0); 822 error = pthread_create(&td, NULL, ggate_send_thread, res); 823 assert(error == 0); 824 error = pthread_create(&td, NULL, sync_thread, res); 825 assert(error == 0); 826 error = pthread_create(&td, NULL, ctrl_thread, res); 827 assert(error == 0); 828 (void)guard_thread(res); 829} 830 831static void 832reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 833{ 834 char msg[1024]; 835 va_list ap; 836 int len; 837 838 va_start(ap, fmt); 839 len = vsnprintf(msg, sizeof(msg), fmt, ap); 840 va_end(ap); 841 if ((size_t)len < sizeof(msg)) { 842 switch (ggio->gctl_cmd) { 843 case BIO_READ: 844 (void)snprintf(msg + len, sizeof(msg) - len, 845 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 846 (uintmax_t)ggio->gctl_length); 847 break; 848 case BIO_DELETE: 849 (void)snprintf(msg + len, sizeof(msg) - len, 850 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 851 (uintmax_t)ggio->gctl_length); 852 break; 853 case BIO_FLUSH: 854 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 855 break; 856 case BIO_WRITE: 857 (void)snprintf(msg + len, sizeof(msg) - len, 858 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 859 (uintmax_t)ggio->gctl_length); 860 break; 861 default: 862 (void)snprintf(msg + len, sizeof(msg) - len, 863 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 864 break; 865 } 866 } 867 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 868} 869 870static void 871remote_close(struct hast_resource *res, int ncomp) 872{ 873 874 rw_wlock(&hio_remote_lock[ncomp]); 875 /* 876 * A race is possible between dropping rlock and acquiring wlock - 877 * another thread can close connection in-between. 878 */ 879 if (!ISCONNECTED(res, ncomp)) { 880 assert(res->hr_remotein == NULL); 881 assert(res->hr_remoteout == NULL); 882 rw_unlock(&hio_remote_lock[ncomp]); 883 return; 884 } 885 886 assert(res->hr_remotein != NULL); 887 assert(res->hr_remoteout != NULL); 888 889 pjdlog_debug(2, "Closing incoming connection to %s.", 890 res->hr_remoteaddr); 891 proto_close(res->hr_remotein); 892 res->hr_remotein = NULL; 893 pjdlog_debug(2, "Closing outgoing connection to %s.", 894 res->hr_remoteaddr); 895 proto_close(res->hr_remoteout); 896 res->hr_remoteout = NULL; 897 898 rw_unlock(&hio_remote_lock[ncomp]); 899 900 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 901 902 /* 903 * Stop synchronization if in-progress. 904 */ 905 sync_stop(); 906 907 event_send(res, EVENT_DISCONNECT); 908} 909 910/* 911 * Thread receives ggate I/O requests from the kernel and passes them to 912 * appropriate threads: 913 * WRITE - always goes to both local_send and remote_send threads 914 * READ (when the block is up-to-date on local component) - 915 * only local_send thread 916 * READ (when the block isn't up-to-date on local component) - 917 * only remote_send thread 918 * DELETE - always goes to both local_send and remote_send threads 919 * FLUSH - always goes to both local_send and remote_send threads 920 */ 921static void * 922ggate_recv_thread(void *arg) 923{ 924 struct hast_resource *res = arg; 925 struct g_gate_ctl_io *ggio; 926 struct hio *hio; 927 unsigned int ii, ncomp, ncomps; 928 int error; 929 930 ncomps = HAST_NCOMPONENTS; 931 932 for (;;) { 933 pjdlog_debug(2, "ggate_recv: Taking free request."); 934 QUEUE_TAKE2(hio, free); 935 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 936 ggio = &hio->hio_ggio; 937 ggio->gctl_unit = res->hr_ggateunit; 938 ggio->gctl_length = MAXPHYS; 939 ggio->gctl_error = 0; 940 pjdlog_debug(2, 941 "ggate_recv: (%p) Waiting for request from the kernel.", 942 hio); 943 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 944 if (sigexit_received) 945 pthread_exit(NULL); 946 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 947 } 948 error = ggio->gctl_error; 949 switch (error) { 950 case 0: 951 break; 952 case ECANCELED: 953 /* Exit gracefully. */ 954 if (!sigexit_received) { 955 pjdlog_debug(2, 956 "ggate_recv: (%p) Received cancel from the kernel.", 957 hio); 958 pjdlog_info("Received cancel from the kernel, exiting."); 959 } 960 pthread_exit(NULL); 961 case ENOMEM: 962 /* 963 * Buffer too small? Impossible, we allocate MAXPHYS 964 * bytes - request can't be bigger than that. 965 */ 966 /* FALLTHROUGH */ 967 case ENXIO: 968 default: 969 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 970 strerror(error)); 971 } 972 for (ii = 0; ii < ncomps; ii++) 973 hio->hio_errors[ii] = EINVAL; 974 reqlog(LOG_DEBUG, 2, ggio, 975 "ggate_recv: (%p) Request received from the kernel: ", 976 hio); 977 /* 978 * Inform all components about new write request. 979 * For read request prefer local component unless the given 980 * range is out-of-date, then use remote component. 981 */ 982 switch (ggio->gctl_cmd) { 983 case BIO_READ: 984 pjdlog_debug(2, 985 "ggate_recv: (%p) Moving request to the send queue.", 986 hio); 987 refcount_init(&hio->hio_countdown, 1); 988 mtx_lock(&metadata_lock); 989 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 990 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 991 /* 992 * This range is up-to-date on local component, 993 * so handle request locally. 994 */ 995 /* Local component is 0 for now. */ 996 ncomp = 0; 997 } else /* if (res->hr_syncsrc == 998 HAST_SYNCSRC_SECONDARY) */ { 999 assert(res->hr_syncsrc == 1000 HAST_SYNCSRC_SECONDARY); 1001 /* 1002 * This range is out-of-date on local component, 1003 * so send request to the remote node. 1004 */ 1005 /* Remote component is 1 for now. */ 1006 ncomp = 1; 1007 } 1008 mtx_unlock(&metadata_lock); 1009 QUEUE_INSERT1(hio, send, ncomp); 1010 break; 1011 case BIO_WRITE: 1012 for (;;) { 1013 mtx_lock(&range_lock); 1014 if (rangelock_islocked(range_sync, 1015 ggio->gctl_offset, ggio->gctl_length)) { 1016 pjdlog_debug(2, 1017 "regular: Range offset=%jd length=%zu locked.", 1018 (intmax_t)ggio->gctl_offset, 1019 (size_t)ggio->gctl_length); 1020 range_regular_wait = true; 1021 cv_wait(&range_regular_cond, &range_lock); 1022 range_regular_wait = false; 1023 mtx_unlock(&range_lock); 1024 continue; 1025 } 1026 if (rangelock_add(range_regular, 1027 ggio->gctl_offset, ggio->gctl_length) < 0) { 1028 mtx_unlock(&range_lock); 1029 pjdlog_debug(2, 1030 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1031 (intmax_t)ggio->gctl_offset, 1032 (size_t)ggio->gctl_length); 1033 sleep(1); 1034 continue; 1035 } 1036 mtx_unlock(&range_lock); 1037 break; 1038 } 1039 mtx_lock(&res->hr_amp_lock); 1040 if (activemap_write_start(res->hr_amp, 1041 ggio->gctl_offset, ggio->gctl_length)) { 1042 (void)hast_activemap_flush(res); 1043 } 1044 mtx_unlock(&res->hr_amp_lock); 1045 /* FALLTHROUGH */ 1046 case BIO_DELETE: 1047 case BIO_FLUSH: 1048 pjdlog_debug(2, 1049 "ggate_recv: (%p) Moving request to the send queues.", 1050 hio); 1051 refcount_init(&hio->hio_countdown, ncomps); 1052 for (ii = 0; ii < ncomps; ii++) 1053 QUEUE_INSERT1(hio, send, ii); 1054 break; 1055 } 1056 } 1057 /* NOTREACHED */ 1058 return (NULL); 1059} 1060 1061/* 1062 * Thread reads from or writes to local component. 1063 * If local read fails, it redirects it to remote_send thread. 1064 */ 1065static void * 1066local_send_thread(void *arg) 1067{ 1068 struct hast_resource *res = arg; 1069 struct g_gate_ctl_io *ggio; 1070 struct hio *hio; 1071 unsigned int ncomp, rncomp; 1072 ssize_t ret; 1073 1074 /* Local component is 0 for now. */ 1075 ncomp = 0; 1076 /* Remote component is 1 for now. */ 1077 rncomp = 1; 1078 1079 for (;;) { 1080 pjdlog_debug(2, "local_send: Taking request."); 1081 QUEUE_TAKE1(hio, send, ncomp); 1082 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1083 ggio = &hio->hio_ggio; 1084 switch (ggio->gctl_cmd) { 1085 case BIO_READ: 1086 ret = pread(res->hr_localfd, ggio->gctl_data, 1087 ggio->gctl_length, 1088 ggio->gctl_offset + res->hr_localoff); 1089 if (ret == ggio->gctl_length) 1090 hio->hio_errors[ncomp] = 0; 1091 else { 1092 /* 1093 * If READ failed, try to read from remote node. 1094 */ 1095 QUEUE_INSERT1(hio, send, rncomp); 1096 continue; 1097 } 1098 break; 1099 case BIO_WRITE: 1100 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1101 ggio->gctl_length, 1102 ggio->gctl_offset + res->hr_localoff); 1103 if (ret < 0) 1104 hio->hio_errors[ncomp] = errno; 1105 else if (ret != ggio->gctl_length) 1106 hio->hio_errors[ncomp] = EIO; 1107 else 1108 hio->hio_errors[ncomp] = 0; 1109 break; 1110 case BIO_DELETE: 1111 ret = g_delete(res->hr_localfd, 1112 ggio->gctl_offset + res->hr_localoff, 1113 ggio->gctl_length); 1114 if (ret < 0) 1115 hio->hio_errors[ncomp] = errno; 1116 else 1117 hio->hio_errors[ncomp] = 0; 1118 break; 1119 case BIO_FLUSH: 1120 ret = g_flush(res->hr_localfd); 1121 if (ret < 0) 1122 hio->hio_errors[ncomp] = errno; 1123 else 1124 hio->hio_errors[ncomp] = 0; 1125 break; 1126 } 1127 if (refcount_release(&hio->hio_countdown)) { 1128 if (ISSYNCREQ(hio)) { 1129 mtx_lock(&sync_lock); 1130 SYNCREQDONE(hio); 1131 mtx_unlock(&sync_lock); 1132 cv_signal(&sync_cond); 1133 } else { 1134 pjdlog_debug(2, 1135 "local_send: (%p) Moving request to the done queue.", 1136 hio); 1137 QUEUE_INSERT2(hio, done); 1138 } 1139 } 1140 } 1141 /* NOTREACHED */ 1142 return (NULL); 1143} 1144 1145/* 1146 * Thread sends request to secondary node. 1147 */ 1148static void * 1149remote_send_thread(void *arg) 1150{ 1151 struct hast_resource *res = arg; 1152 struct g_gate_ctl_io *ggio; 1153 struct hio *hio; 1154 struct nv *nv; 1155 unsigned int ncomp; 1156 bool wakeup; 1157 uint64_t offset, length; 1158 uint8_t cmd; 1159 void *data; 1160 1161 /* Remote component is 1 for now. */ 1162 ncomp = 1; 1163 1164 for (;;) { 1165 pjdlog_debug(2, "remote_send: Taking request."); 1166 QUEUE_TAKE1(hio, send, ncomp); 1167 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1168 ggio = &hio->hio_ggio; 1169 switch (ggio->gctl_cmd) { 1170 case BIO_READ: 1171 cmd = HIO_READ; 1172 data = NULL; 1173 offset = ggio->gctl_offset; 1174 length = ggio->gctl_length; 1175 break; 1176 case BIO_WRITE: 1177 cmd = HIO_WRITE; 1178 data = ggio->gctl_data; 1179 offset = ggio->gctl_offset; 1180 length = ggio->gctl_length; 1181 break; 1182 case BIO_DELETE: 1183 cmd = HIO_DELETE; 1184 data = NULL; 1185 offset = ggio->gctl_offset; 1186 length = ggio->gctl_length; 1187 break; 1188 case BIO_FLUSH: 1189 cmd = HIO_FLUSH; 1190 data = NULL; 1191 offset = 0; 1192 length = 0; 1193 break; 1194 default: 1195 assert(!"invalid condition"); 1196 abort(); 1197 } 1198 nv = nv_alloc(); 1199 nv_add_uint8(nv, cmd, "cmd"); 1200 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1201 nv_add_uint64(nv, offset, "offset"); 1202 nv_add_uint64(nv, length, "length"); 1203 if (nv_error(nv) != 0) { 1204 hio->hio_errors[ncomp] = nv_error(nv); 1205 pjdlog_debug(2, 1206 "remote_send: (%p) Unable to prepare header to send.", 1207 hio); 1208 reqlog(LOG_ERR, 0, ggio, 1209 "Unable to prepare header to send (%s): ", 1210 strerror(nv_error(nv))); 1211 /* Move failed request immediately to the done queue. */ 1212 goto done_queue; 1213 } 1214 pjdlog_debug(2, 1215 "remote_send: (%p) Moving request to the recv queue.", 1216 hio); 1217 /* 1218 * Protect connection from disappearing. 1219 */ 1220 rw_rlock(&hio_remote_lock[ncomp]); 1221 if (!ISCONNECTED(res, ncomp)) { 1222 rw_unlock(&hio_remote_lock[ncomp]); 1223 hio->hio_errors[ncomp] = ENOTCONN; 1224 goto done_queue; 1225 } 1226 /* 1227 * Move the request to recv queue before sending it, because 1228 * in different order we can get reply before we move request 1229 * to recv queue. 1230 */ 1231 mtx_lock(&hio_recv_list_lock[ncomp]); 1232 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1233 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1234 mtx_unlock(&hio_recv_list_lock[ncomp]); 1235 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1236 data != NULL ? length : 0) < 0) { 1237 hio->hio_errors[ncomp] = errno; 1238 rw_unlock(&hio_remote_lock[ncomp]); 1239 pjdlog_debug(2, 1240 "remote_send: (%p) Unable to send request.", hio); 1241 reqlog(LOG_ERR, 0, ggio, 1242 "Unable to send request (%s): ", 1243 strerror(hio->hio_errors[ncomp])); 1244 remote_close(res, ncomp); 1245 /* 1246 * Take request back from the receive queue and move 1247 * it immediately to the done queue. 1248 */ 1249 mtx_lock(&hio_recv_list_lock[ncomp]); 1250 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1251 mtx_unlock(&hio_recv_list_lock[ncomp]); 1252 goto done_queue; 1253 } 1254 rw_unlock(&hio_remote_lock[ncomp]); 1255 nv_free(nv); 1256 if (wakeup) 1257 cv_signal(&hio_recv_list_cond[ncomp]); 1258 continue; 1259done_queue: 1260 nv_free(nv); 1261 if (ISSYNCREQ(hio)) { 1262 if (!refcount_release(&hio->hio_countdown)) 1263 continue; 1264 mtx_lock(&sync_lock); 1265 SYNCREQDONE(hio); 1266 mtx_unlock(&sync_lock); 1267 cv_signal(&sync_cond); 1268 continue; 1269 } 1270 if (ggio->gctl_cmd == BIO_WRITE) { 1271 mtx_lock(&res->hr_amp_lock); 1272 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1273 ggio->gctl_length)) { 1274 (void)hast_activemap_flush(res); 1275 } 1276 mtx_unlock(&res->hr_amp_lock); 1277 } 1278 if (!refcount_release(&hio->hio_countdown)) 1279 continue; 1280 pjdlog_debug(2, 1281 "remote_send: (%p) Moving request to the done queue.", 1282 hio); 1283 QUEUE_INSERT2(hio, done); 1284 } 1285 /* NOTREACHED */ 1286 return (NULL); 1287} 1288 1289/* 1290 * Thread receives answer from secondary node and passes it to ggate_send 1291 * thread. 1292 */ 1293static void * 1294remote_recv_thread(void *arg) 1295{ 1296 struct hast_resource *res = arg; 1297 struct g_gate_ctl_io *ggio; 1298 struct hio *hio; 1299 struct nv *nv; 1300 unsigned int ncomp; 1301 uint64_t seq; 1302 int error; 1303 1304 /* Remote component is 1 for now. */ 1305 ncomp = 1; 1306 1307 for (;;) { 1308 /* Wait until there is anything to receive. */ 1309 mtx_lock(&hio_recv_list_lock[ncomp]); 1310 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1311 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1312 cv_wait(&hio_recv_list_cond[ncomp], 1313 &hio_recv_list_lock[ncomp]); 1314 } 1315 mtx_unlock(&hio_recv_list_lock[ncomp]); 1316 rw_rlock(&hio_remote_lock[ncomp]); 1317 if (!ISCONNECTED(res, ncomp)) { 1318 rw_unlock(&hio_remote_lock[ncomp]); 1319 /* 1320 * Connection is dead, so move all pending requests to 1321 * the done queue (one-by-one). 1322 */ 1323 mtx_lock(&hio_recv_list_lock[ncomp]); 1324 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1325 assert(hio != NULL); 1326 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1327 hio_next[ncomp]); 1328 mtx_unlock(&hio_recv_list_lock[ncomp]); 1329 goto done_queue; 1330 } 1331 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1332 pjdlog_errno(LOG_ERR, 1333 "Unable to receive reply header"); 1334 rw_unlock(&hio_remote_lock[ncomp]); 1335 remote_close(res, ncomp); 1336 continue; 1337 } 1338 rw_unlock(&hio_remote_lock[ncomp]); 1339 seq = nv_get_uint64(nv, "seq"); 1340 if (seq == 0) { 1341 pjdlog_error("Header contains no 'seq' field."); 1342 nv_free(nv); 1343 continue; 1344 } 1345 mtx_lock(&hio_recv_list_lock[ncomp]); 1346 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1347 if (hio->hio_ggio.gctl_seq == seq) { 1348 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1349 hio_next[ncomp]); 1350 break; 1351 } 1352 } 1353 mtx_unlock(&hio_recv_list_lock[ncomp]); 1354 if (hio == NULL) { 1355 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1356 (uintmax_t)seq); 1357 nv_free(nv); 1358 continue; 1359 } 1360 error = nv_get_int16(nv, "error"); 1361 if (error != 0) { 1362 /* Request failed on remote side. */ 1363 hio->hio_errors[ncomp] = 0; 1364 nv_free(nv); 1365 goto done_queue; 1366 } 1367 ggio = &hio->hio_ggio; 1368 switch (ggio->gctl_cmd) { 1369 case BIO_READ: 1370 rw_rlock(&hio_remote_lock[ncomp]); 1371 if (!ISCONNECTED(res, ncomp)) { 1372 rw_unlock(&hio_remote_lock[ncomp]); 1373 nv_free(nv); 1374 goto done_queue; 1375 } 1376 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1377 ggio->gctl_data, ggio->gctl_length) < 0) { 1378 hio->hio_errors[ncomp] = errno; 1379 pjdlog_errno(LOG_ERR, 1380 "Unable to receive reply data"); 1381 rw_unlock(&hio_remote_lock[ncomp]); 1382 nv_free(nv); 1383 remote_close(res, ncomp); 1384 goto done_queue; 1385 } 1386 rw_unlock(&hio_remote_lock[ncomp]); 1387 break; 1388 case BIO_WRITE: 1389 case BIO_DELETE: 1390 case BIO_FLUSH: 1391 break; 1392 default: 1393 assert(!"invalid condition"); 1394 abort(); 1395 } 1396 hio->hio_errors[ncomp] = 0; 1397 nv_free(nv); 1398done_queue: 1399 if (refcount_release(&hio->hio_countdown)) { 1400 if (ISSYNCREQ(hio)) { 1401 mtx_lock(&sync_lock); 1402 SYNCREQDONE(hio); 1403 mtx_unlock(&sync_lock); 1404 cv_signal(&sync_cond); 1405 } else { 1406 pjdlog_debug(2, 1407 "remote_recv: (%p) Moving request to the done queue.", 1408 hio); 1409 QUEUE_INSERT2(hio, done); 1410 } 1411 } 1412 } 1413 /* NOTREACHED */ 1414 return (NULL); 1415} 1416 1417/* 1418 * Thread sends answer to the kernel. 1419 */ 1420static void * 1421ggate_send_thread(void *arg) 1422{ 1423 struct hast_resource *res = arg; 1424 struct g_gate_ctl_io *ggio; 1425 struct hio *hio; 1426 unsigned int ii, ncomp, ncomps; 1427 1428 ncomps = HAST_NCOMPONENTS; 1429 1430 for (;;) { 1431 pjdlog_debug(2, "ggate_send: Taking request."); 1432 QUEUE_TAKE2(hio, done); 1433 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1434 ggio = &hio->hio_ggio; 1435 for (ii = 0; ii < ncomps; ii++) { 1436 if (hio->hio_errors[ii] == 0) { 1437 /* 1438 * One successful request is enough to declare 1439 * success. 1440 */ 1441 ggio->gctl_error = 0; 1442 break; 1443 } 1444 } 1445 if (ii == ncomps) { 1446 /* 1447 * None of the requests were successful. 1448 * Use first error. 1449 */ 1450 ggio->gctl_error = hio->hio_errors[0]; 1451 } 1452 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1453 mtx_lock(&res->hr_amp_lock); 1454 activemap_write_complete(res->hr_amp, 1455 ggio->gctl_offset, ggio->gctl_length); 1456 mtx_unlock(&res->hr_amp_lock); 1457 } 1458 if (ggio->gctl_cmd == BIO_WRITE) { 1459 /* 1460 * Unlock range we locked. 1461 */ 1462 mtx_lock(&range_lock); 1463 rangelock_del(range_regular, ggio->gctl_offset, 1464 ggio->gctl_length); 1465 if (range_sync_wait) 1466 cv_signal(&range_sync_cond); 1467 mtx_unlock(&range_lock); 1468 /* 1469 * Bump local count if this is first write after 1470 * connection failure with remote node. 1471 */ 1472 ncomp = 1; 1473 rw_rlock(&hio_remote_lock[ncomp]); 1474 if (!ISCONNECTED(res, ncomp)) { 1475 mtx_lock(&metadata_lock); 1476 if (res->hr_primary_localcnt == 1477 res->hr_secondary_remotecnt) { 1478 res->hr_primary_localcnt++; 1479 pjdlog_debug(1, 1480 "Increasing localcnt to %ju.", 1481 (uintmax_t)res->hr_primary_localcnt); 1482 (void)metadata_write(res); 1483 } 1484 mtx_unlock(&metadata_lock); 1485 } 1486 rw_unlock(&hio_remote_lock[ncomp]); 1487 } 1488 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1489 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1490 pjdlog_debug(2, 1491 "ggate_send: (%p) Moving request to the free queue.", hio); 1492 QUEUE_INSERT2(hio, free); 1493 } 1494 /* NOTREACHED */ 1495 return (NULL); 1496} 1497 1498/* 1499 * Thread synchronize local and remote components. 1500 */ 1501static void * 1502sync_thread(void *arg __unused) 1503{ 1504 struct hast_resource *res = arg; 1505 struct hio *hio; 1506 struct g_gate_ctl_io *ggio; 1507 unsigned int ii, ncomp, ncomps; 1508 off_t offset, length, synced; 1509 bool dorewind; 1510 int syncext; 1511 1512 ncomps = HAST_NCOMPONENTS; 1513 dorewind = true; 1514 synced = 0; 1515 offset = -1; 1516 1517 for (;;) { 1518 mtx_lock(&sync_lock); 1519 if (offset >= 0 && !sync_inprogress) { 1520 pjdlog_info("Synchronization interrupted. " 1521 "%jd bytes synchronized so far.", 1522 (intmax_t)synced); 1523 event_send(res, EVENT_SYNCINTR); 1524 } 1525 while (!sync_inprogress) { 1526 dorewind = true; 1527 synced = 0; 1528 cv_wait(&sync_cond, &sync_lock); 1529 } 1530 mtx_unlock(&sync_lock); 1531 /* 1532 * Obtain offset at which we should synchronize. 1533 * Rewind synchronization if needed. 1534 */ 1535 mtx_lock(&res->hr_amp_lock); 1536 if (dorewind) 1537 activemap_sync_rewind(res->hr_amp); 1538 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1539 if (syncext != -1) { 1540 /* 1541 * We synchronized entire syncext extent, we can mark 1542 * it as clean now. 1543 */ 1544 if (activemap_extent_complete(res->hr_amp, syncext)) 1545 (void)hast_activemap_flush(res); 1546 } 1547 mtx_unlock(&res->hr_amp_lock); 1548 if (dorewind) { 1549 dorewind = false; 1550 if (offset < 0) 1551 pjdlog_info("Nodes are in sync."); 1552 else { 1553 pjdlog_info("Synchronization started. %ju bytes to go.", 1554 (uintmax_t)(res->hr_extentsize * 1555 activemap_ndirty(res->hr_amp))); 1556 event_send(res, EVENT_SYNCSTART); 1557 } 1558 } 1559 if (offset < 0) { 1560 sync_stop(); 1561 pjdlog_debug(1, "Nothing to synchronize."); 1562 /* 1563 * Synchronization complete, make both localcnt and 1564 * remotecnt equal. 1565 */ 1566 ncomp = 1; 1567 rw_rlock(&hio_remote_lock[ncomp]); 1568 if (ISCONNECTED(res, ncomp)) { 1569 if (synced > 0) { 1570 pjdlog_info("Synchronization complete. " 1571 "%jd bytes synchronized.", 1572 (intmax_t)synced); 1573 event_send(res, EVENT_SYNCDONE); 1574 } 1575 mtx_lock(&metadata_lock); 1576 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1577 res->hr_primary_localcnt = 1578 res->hr_secondary_localcnt; 1579 res->hr_primary_remotecnt = 1580 res->hr_secondary_remotecnt; 1581 pjdlog_debug(1, 1582 "Setting localcnt to %ju and remotecnt to %ju.", 1583 (uintmax_t)res->hr_primary_localcnt, 1584 (uintmax_t)res->hr_secondary_localcnt); 1585 (void)metadata_write(res); 1586 mtx_unlock(&metadata_lock); 1587 } 1588 rw_unlock(&hio_remote_lock[ncomp]); 1589 continue; 1590 } 1591 pjdlog_debug(2, "sync: Taking free request."); 1592 QUEUE_TAKE2(hio, free); 1593 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1594 /* 1595 * Lock the range we are going to synchronize. We don't want 1596 * race where someone writes between our read and write. 1597 */ 1598 for (;;) { 1599 mtx_lock(&range_lock); 1600 if (rangelock_islocked(range_regular, offset, length)) { 1601 pjdlog_debug(2, 1602 "sync: Range offset=%jd length=%jd locked.", 1603 (intmax_t)offset, (intmax_t)length); 1604 range_sync_wait = true; 1605 cv_wait(&range_sync_cond, &range_lock); 1606 range_sync_wait = false; 1607 mtx_unlock(&range_lock); 1608 continue; 1609 } 1610 if (rangelock_add(range_sync, offset, length) < 0) { 1611 mtx_unlock(&range_lock); 1612 pjdlog_debug(2, 1613 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1614 (intmax_t)offset, (intmax_t)length); 1615 sleep(1); 1616 continue; 1617 } 1618 mtx_unlock(&range_lock); 1619 break; 1620 } 1621 /* 1622 * First read the data from synchronization source. 1623 */ 1624 SYNCREQ(hio); 1625 ggio = &hio->hio_ggio; 1626 ggio->gctl_cmd = BIO_READ; 1627 ggio->gctl_offset = offset; 1628 ggio->gctl_length = length; 1629 ggio->gctl_error = 0; 1630 for (ii = 0; ii < ncomps; ii++) 1631 hio->hio_errors[ii] = EINVAL; 1632 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1633 hio); 1634 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1635 hio); 1636 mtx_lock(&metadata_lock); 1637 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1638 /* 1639 * This range is up-to-date on local component, 1640 * so handle request locally. 1641 */ 1642 /* Local component is 0 for now. */ 1643 ncomp = 0; 1644 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1645 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1646 /* 1647 * This range is out-of-date on local component, 1648 * so send request to the remote node. 1649 */ 1650 /* Remote component is 1 for now. */ 1651 ncomp = 1; 1652 } 1653 mtx_unlock(&metadata_lock); 1654 refcount_init(&hio->hio_countdown, 1); 1655 QUEUE_INSERT1(hio, send, ncomp); 1656 1657 /* 1658 * Let's wait for READ to finish. 1659 */ 1660 mtx_lock(&sync_lock); 1661 while (!ISSYNCREQDONE(hio)) 1662 cv_wait(&sync_cond, &sync_lock); 1663 mtx_unlock(&sync_lock); 1664 1665 if (hio->hio_errors[ncomp] != 0) { 1666 pjdlog_error("Unable to read synchronization data: %s.", 1667 strerror(hio->hio_errors[ncomp])); 1668 goto free_queue; 1669 } 1670 1671 /* 1672 * We read the data from synchronization source, now write it 1673 * to synchronization target. 1674 */ 1675 SYNCREQ(hio); 1676 ggio->gctl_cmd = BIO_WRITE; 1677 for (ii = 0; ii < ncomps; ii++) 1678 hio->hio_errors[ii] = EINVAL; 1679 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1680 hio); 1681 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1682 hio); 1683 mtx_lock(&metadata_lock); 1684 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1685 /* 1686 * This range is up-to-date on local component, 1687 * so we update remote component. 1688 */ 1689 /* Remote component is 1 for now. */ 1690 ncomp = 1; 1691 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1692 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1693 /* 1694 * This range is out-of-date on local component, 1695 * so we update it. 1696 */ 1697 /* Local component is 0 for now. */ 1698 ncomp = 0; 1699 } 1700 mtx_unlock(&metadata_lock); 1701 1702 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1703 hio); 1704 refcount_init(&hio->hio_countdown, 1); 1705 QUEUE_INSERT1(hio, send, ncomp); 1706 1707 /* 1708 * Let's wait for WRITE to finish. 1709 */ 1710 mtx_lock(&sync_lock); 1711 while (!ISSYNCREQDONE(hio)) 1712 cv_wait(&sync_cond, &sync_lock); 1713 mtx_unlock(&sync_lock); 1714 1715 if (hio->hio_errors[ncomp] != 0) { 1716 pjdlog_error("Unable to write synchronization data: %s.", 1717 strerror(hio->hio_errors[ncomp])); 1718 goto free_queue; 1719 } 1720 1721 synced += length; 1722free_queue: 1723 mtx_lock(&range_lock); 1724 rangelock_del(range_sync, offset, length); 1725 if (range_regular_wait) 1726 cv_signal(&range_regular_cond); 1727 mtx_unlock(&range_lock); 1728 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1729 hio); 1730 QUEUE_INSERT2(hio, free); 1731 } 1732 /* NOTREACHED */ 1733 return (NULL); 1734} 1735 1736static void 1737config_reload(void) 1738{ 1739 struct hastd_config *newcfg; 1740 struct hast_resource *res; 1741 unsigned int ii, ncomps; 1742 int modified; 1743 1744 pjdlog_info("Reloading configuration..."); 1745 1746 ncomps = HAST_NCOMPONENTS; 1747 1748 newcfg = yy_config_parse(cfgpath, false); 1749 if (newcfg == NULL) 1750 goto failed; 1751 1752 TAILQ_FOREACH(res, &newcfg->hc_resources, hr_next) { 1753 if (strcmp(res->hr_name, gres->hr_name) == 0) 1754 break; 1755 } 1756 /* 1757 * If resource was removed from the configuration file, resource 1758 * name, provider name or path to local component was modified we 1759 * shouldn't be here. This means that someone modified configuration 1760 * file and send SIGHUP to us instead of main hastd process. 1761 * Log advice and ignore the signal. 1762 */ 1763 if (res == NULL || strcmp(gres->hr_name, res->hr_name) != 0 || 1764 strcmp(gres->hr_provname, res->hr_provname) != 0 || 1765 strcmp(gres->hr_localpath, res->hr_localpath) != 0) { 1766 pjdlog_warning("To reload configuration send SIGHUP to the main hastd process (pid %u).", 1767 (unsigned int)getppid()); 1768 goto failed; 1769 } 1770 1771#define MODIFIED_REMOTEADDR 0x1 1772#define MODIFIED_REPLICATION 0x2 1773#define MODIFIED_TIMEOUT 0x4 1774#define MODIFIED_EXEC 0x8 1775 modified = 0; 1776 if (strcmp(gres->hr_remoteaddr, res->hr_remoteaddr) != 0) { 1777 /* 1778 * Don't copy res->hr_remoteaddr to gres just yet. 1779 * We want remote_close() to log disconnect from the old 1780 * addresses, not from the new ones. 1781 */ 1782 modified |= MODIFIED_REMOTEADDR; 1783 } 1784 if (gres->hr_replication != res->hr_replication) { 1785 gres->hr_replication = res->hr_replication; 1786 modified |= MODIFIED_REPLICATION; 1787 } 1788 if (gres->hr_timeout != res->hr_timeout) { 1789 gres->hr_timeout = res->hr_timeout; 1790 modified |= MODIFIED_TIMEOUT; 1791 } 1792 if (strcmp(gres->hr_exec, res->hr_exec) != 0) { 1793 strlcpy(gres->hr_exec, res->hr_exec, sizeof(gres->hr_exec)); 1794 modified |= MODIFIED_EXEC; 1795 } 1796 /* 1797 * If only timeout was modified we only need to change it without 1798 * reconnecting. 1799 */ 1800 if (modified == MODIFIED_TIMEOUT) { 1801 for (ii = 0; ii < ncomps; ii++) { 1802 if (!ISREMOTE(ii)) 1803 continue; 1804 rw_rlock(&hio_remote_lock[ii]); 1805 if (!ISCONNECTED(gres, ii)) { 1806 rw_unlock(&hio_remote_lock[ii]); 1807 continue; 1808 } 1809 rw_unlock(&hio_remote_lock[ii]); 1810 if (proto_timeout(gres->hr_remotein, 1811 gres->hr_timeout) < 0) { 1812 pjdlog_errno(LOG_WARNING, 1813 "Unable to set connection timeout"); 1814 } 1815 if (proto_timeout(gres->hr_remoteout, 1816 gres->hr_timeout) < 0) { 1817 pjdlog_errno(LOG_WARNING, 1818 "Unable to set connection timeout"); 1819 } 1820 } 1821 } else if ((modified & 1822 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) { 1823 for (ii = 0; ii < ncomps; ii++) { 1824 if (!ISREMOTE(ii)) 1825 continue; 1826 remote_close(gres, ii); 1827 } 1828 if (modified & MODIFIED_REMOTEADDR) { 1829 strlcpy(gres->hr_remoteaddr, res->hr_remoteaddr, 1830 sizeof(gres->hr_remoteaddr)); 1831 } 1832 } 1833#undef MODIFIED_REMOTEADDR 1834#undef MODIFIED_REPLICATION 1835#undef MODIFIED_TIMEOUT 1836#undef MODIFIED_EXEC 1837 1838 pjdlog_info("Configuration reloaded successfully."); 1839 return; 1840failed: 1841 if (newcfg != NULL) { 1842 if (newcfg->hc_controlconn != NULL) 1843 proto_close(newcfg->hc_controlconn); 1844 if (newcfg->hc_listenconn != NULL) 1845 proto_close(newcfg->hc_listenconn); 1846 yy_config_free(newcfg); 1847 } 1848 pjdlog_warning("Configuration not reloaded."); 1849} 1850 1851static void 1852keepalive_send(struct hast_resource *res, unsigned int ncomp) 1853{ 1854 struct nv *nv; 1855 1856 nv = nv_alloc(); 1857 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1858 if (nv_error(nv) != 0) { 1859 nv_free(nv); 1860 pjdlog_debug(1, 1861 "keepalive_send: Unable to prepare header to send."); 1862 return; 1863 } 1864 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1865 pjdlog_common(LOG_DEBUG, 1, errno, 1866 "keepalive_send: Unable to send request"); 1867 nv_free(nv); 1868 rw_unlock(&hio_remote_lock[ncomp]); 1869 remote_close(res, ncomp); 1870 rw_rlock(&hio_remote_lock[ncomp]); 1871 return; 1872 } 1873 nv_free(nv); 1874 pjdlog_debug(2, "keepalive_send: Request sent."); 1875} 1876 1877static void 1878guard_one(struct hast_resource *res, unsigned int ncomp) 1879{ 1880 struct proto_conn *in, *out; 1881 1882 if (!ISREMOTE(ncomp)) 1883 return; 1884 1885 rw_rlock(&hio_remote_lock[ncomp]); 1886 1887 if (!real_remote(res)) { 1888 rw_unlock(&hio_remote_lock[ncomp]); 1889 return; 1890 } 1891 1892 if (ISCONNECTED(res, ncomp)) { 1893 assert(res->hr_remotein != NULL); 1894 assert(res->hr_remoteout != NULL); 1895 keepalive_send(res, ncomp); 1896 } 1897 1898 if (ISCONNECTED(res, ncomp)) { 1899 assert(res->hr_remotein != NULL); 1900 assert(res->hr_remoteout != NULL); 1901 rw_unlock(&hio_remote_lock[ncomp]); 1902 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 1903 res->hr_remoteaddr); 1904 return; 1905 } 1906 1907 assert(res->hr_remotein == NULL); 1908 assert(res->hr_remoteout == NULL); 1909 /* 1910 * Upgrade the lock. It doesn't have to be atomic as no other thread 1911 * can change connection status from disconnected to connected. 1912 */ 1913 rw_unlock(&hio_remote_lock[ncomp]); 1914 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 1915 res->hr_remoteaddr); 1916 in = out = NULL; 1917 if (init_remote(res, &in, &out)) { 1918 rw_wlock(&hio_remote_lock[ncomp]); 1919 assert(res->hr_remotein == NULL); 1920 assert(res->hr_remoteout == NULL); 1921 assert(in != NULL && out != NULL); 1922 res->hr_remotein = in; 1923 res->hr_remoteout = out; 1924 rw_unlock(&hio_remote_lock[ncomp]); 1925 pjdlog_info("Successfully reconnected to %s.", 1926 res->hr_remoteaddr); 1927 sync_start(); 1928 } else { 1929 /* Both connections should be NULL. */ 1930 assert(res->hr_remotein == NULL); 1931 assert(res->hr_remoteout == NULL); 1932 assert(in == NULL && out == NULL); 1933 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 1934 res->hr_remoteaddr); 1935 } 1936} 1937 1938/* 1939 * Thread guards remote connections and reconnects when needed, handles 1940 * signals, etc. 1941 */ 1942static void * 1943guard_thread(void *arg) 1944{ 1945 struct hast_resource *res = arg; 1946 unsigned int ii, ncomps; 1947 struct timespec timeout; 1948 time_t lastcheck, now; 1949 sigset_t mask; 1950 int signo; 1951 1952 ncomps = HAST_NCOMPONENTS; 1953 lastcheck = time(NULL); 1954 1955 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 1956 PJDLOG_VERIFY(sigaddset(&mask, SIGHUP) == 0); 1957 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 1958 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 1959 1960 timeout.tv_nsec = 0; 1961 signo = -1; 1962 1963 for (;;) { 1964 switch (signo) { 1965 case SIGHUP: 1966 config_reload(); 1967 break; 1968 case SIGINT: 1969 case SIGTERM: 1970 sigexit_received = true; 1971 primary_exitx(EX_OK, 1972 "Termination signal received, exiting."); 1973 break; 1974 default: 1975 break; 1976 } 1977 1978 pjdlog_debug(2, "remote_guard: Checking connections."); 1979 now = time(NULL); 1980 if (lastcheck + RETRY_SLEEP <= now) { 1981 for (ii = 0; ii < ncomps; ii++) 1982 guard_one(res, ii); 1983 lastcheck = now; 1984 } 1985 timeout.tv_sec = RETRY_SLEEP; 1986 signo = sigtimedwait(&mask, NULL, &timeout); 1987 } 1988 /* NOTREACHED */ 1989 return (NULL); 1990} 1991