primary.c revision 213531
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 213531 2010-10-07 18:20:16Z 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 /* Destroy ggate provider if we created one. */ 238 if (res->hr_ggateunit >= 0) { 239 struct g_gate_ctl_destroy ggiod; 240 241 ggiod.gctl_version = G_GATE_VERSION; 242 ggiod.gctl_unit = res->hr_ggateunit; 243 ggiod.gctl_force = 1; 244 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) { 245 pjdlog_errno(LOG_WARNING, 246 "Unable to destroy hast/%s device", 247 res->hr_provname); 248 } 249 res->hr_ggateunit = -1; 250 } 251 252 /* Restore errno. */ 253 errno = rerrno; 254} 255 256static __dead2 void 257primary_exit(int exitcode, const char *fmt, ...) 258{ 259 va_list ap; 260 261 assert(exitcode != EX_OK); 262 va_start(ap, fmt); 263 pjdlogv_errno(LOG_ERR, fmt, ap); 264 va_end(ap); 265 cleanup(gres); 266 exit(exitcode); 267} 268 269static __dead2 void 270primary_exitx(int exitcode, const char *fmt, ...) 271{ 272 va_list ap; 273 274 va_start(ap, fmt); 275 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 276 va_end(ap); 277 cleanup(gres); 278 exit(exitcode); 279} 280 281static int 282hast_activemap_flush(struct hast_resource *res) 283{ 284 const unsigned char *buf; 285 size_t size; 286 287 buf = activemap_bitmap(res->hr_amp, &size); 288 assert(buf != NULL); 289 assert((size % res->hr_local_sectorsize) == 0); 290 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 291 (ssize_t)size) { 292 KEEP_ERRNO(pjdlog_errno(LOG_ERR, 293 "Unable to flush activemap to disk")); 294 return (-1); 295 } 296 return (0); 297} 298 299static bool 300real_remote(const struct hast_resource *res) 301{ 302 303 return (strcmp(res->hr_remoteaddr, "none") != 0); 304} 305 306static void 307init_environment(struct hast_resource *res __unused) 308{ 309 struct hio *hio; 310 unsigned int ii, ncomps; 311 312 /* 313 * In the future it might be per-resource value. 314 */ 315 ncomps = HAST_NCOMPONENTS; 316 317 /* 318 * Allocate memory needed by lists. 319 */ 320 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 321 if (hio_send_list == NULL) { 322 primary_exitx(EX_TEMPFAIL, 323 "Unable to allocate %zu bytes of memory for send lists.", 324 sizeof(hio_send_list[0]) * ncomps); 325 } 326 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 327 if (hio_send_list_lock == NULL) { 328 primary_exitx(EX_TEMPFAIL, 329 "Unable to allocate %zu bytes of memory for send list locks.", 330 sizeof(hio_send_list_lock[0]) * ncomps); 331 } 332 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 333 if (hio_send_list_cond == NULL) { 334 primary_exitx(EX_TEMPFAIL, 335 "Unable to allocate %zu bytes of memory for send list condition variables.", 336 sizeof(hio_send_list_cond[0]) * ncomps); 337 } 338 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 339 if (hio_recv_list == NULL) { 340 primary_exitx(EX_TEMPFAIL, 341 "Unable to allocate %zu bytes of memory for recv lists.", 342 sizeof(hio_recv_list[0]) * ncomps); 343 } 344 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 345 if (hio_recv_list_lock == NULL) { 346 primary_exitx(EX_TEMPFAIL, 347 "Unable to allocate %zu bytes of memory for recv list locks.", 348 sizeof(hio_recv_list_lock[0]) * ncomps); 349 } 350 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 351 if (hio_recv_list_cond == NULL) { 352 primary_exitx(EX_TEMPFAIL, 353 "Unable to allocate %zu bytes of memory for recv list condition variables.", 354 sizeof(hio_recv_list_cond[0]) * ncomps); 355 } 356 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 357 if (hio_remote_lock == NULL) { 358 primary_exitx(EX_TEMPFAIL, 359 "Unable to allocate %zu bytes of memory for remote connections locks.", 360 sizeof(hio_remote_lock[0]) * ncomps); 361 } 362 363 /* 364 * Initialize lists, their locks and theirs condition variables. 365 */ 366 TAILQ_INIT(&hio_free_list); 367 mtx_init(&hio_free_list_lock); 368 cv_init(&hio_free_list_cond); 369 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 370 TAILQ_INIT(&hio_send_list[ii]); 371 mtx_init(&hio_send_list_lock[ii]); 372 cv_init(&hio_send_list_cond[ii]); 373 TAILQ_INIT(&hio_recv_list[ii]); 374 mtx_init(&hio_recv_list_lock[ii]); 375 cv_init(&hio_recv_list_cond[ii]); 376 rw_init(&hio_remote_lock[ii]); 377 } 378 TAILQ_INIT(&hio_done_list); 379 mtx_init(&hio_done_list_lock); 380 cv_init(&hio_done_list_cond); 381 mtx_init(&metadata_lock); 382 383 /* 384 * Allocate requests pool and initialize requests. 385 */ 386 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 387 hio = malloc(sizeof(*hio)); 388 if (hio == NULL) { 389 primary_exitx(EX_TEMPFAIL, 390 "Unable to allocate %zu bytes of memory for hio request.", 391 sizeof(*hio)); 392 } 393 hio->hio_countdown = 0; 394 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 395 if (hio->hio_errors == NULL) { 396 primary_exitx(EX_TEMPFAIL, 397 "Unable allocate %zu bytes of memory for hio errors.", 398 sizeof(hio->hio_errors[0]) * ncomps); 399 } 400 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 401 if (hio->hio_next == NULL) { 402 primary_exitx(EX_TEMPFAIL, 403 "Unable allocate %zu bytes of memory for hio_next field.", 404 sizeof(hio->hio_next[0]) * ncomps); 405 } 406 hio->hio_ggio.gctl_version = G_GATE_VERSION; 407 hio->hio_ggio.gctl_data = malloc(MAXPHYS); 408 if (hio->hio_ggio.gctl_data == NULL) { 409 primary_exitx(EX_TEMPFAIL, 410 "Unable to allocate %zu bytes of memory for gctl_data.", 411 MAXPHYS); 412 } 413 hio->hio_ggio.gctl_length = MAXPHYS; 414 hio->hio_ggio.gctl_error = 0; 415 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 416 } 417} 418 419static void 420init_local(struct hast_resource *res) 421{ 422 unsigned char *buf; 423 size_t mapsize; 424 425 if (metadata_read(res, true) < 0) 426 exit(EX_NOINPUT); 427 mtx_init(&res->hr_amp_lock); 428 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 429 res->hr_local_sectorsize, res->hr_keepdirty) < 0) { 430 primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 431 } 432 mtx_init(&range_lock); 433 cv_init(&range_regular_cond); 434 if (rangelock_init(&range_regular) < 0) 435 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 436 cv_init(&range_sync_cond); 437 if (rangelock_init(&range_sync) < 0) 438 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 439 mapsize = activemap_ondisk_size(res->hr_amp); 440 buf = calloc(1, mapsize); 441 if (buf == NULL) { 442 primary_exitx(EX_TEMPFAIL, 443 "Unable to allocate buffer for activemap."); 444 } 445 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 446 (ssize_t)mapsize) { 447 primary_exit(EX_NOINPUT, "Unable to read activemap"); 448 } 449 activemap_copyin(res->hr_amp, buf, mapsize); 450 free(buf); 451 if (res->hr_resuid != 0) 452 return; 453 /* 454 * We're using provider for the first time, so we have to generate 455 * resource unique identifier and initialize local and remote counts. 456 */ 457 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 458 res->hr_primary_localcnt = 1; 459 res->hr_primary_remotecnt = 0; 460 if (metadata_write(res) < 0) 461 exit(EX_NOINPUT); 462} 463 464static bool 465init_remote(struct hast_resource *res, struct proto_conn **inp, 466 struct proto_conn **outp) 467{ 468 struct proto_conn *in, *out; 469 struct nv *nvout, *nvin; 470 const unsigned char *token; 471 unsigned char *map; 472 const char *errmsg; 473 int32_t extentsize; 474 int64_t datasize; 475 uint32_t mapsize; 476 size_t size; 477 478 assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 479 assert(real_remote(res)); 480 481 in = out = NULL; 482 errmsg = NULL; 483 484 /* Prepare outgoing connection with remote node. */ 485 if (proto_client(res->hr_remoteaddr, &out) < 0) { 486 primary_exit(EX_TEMPFAIL, "Unable to create connection to %s", 487 res->hr_remoteaddr); 488 } 489 /* Try to connect, but accept failure. */ 490 if (proto_connect(out) < 0) { 491 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 492 res->hr_remoteaddr); 493 goto close; 494 } 495 /* Error in setting timeout is not critical, but why should it fail? */ 496 if (proto_timeout(out, res->hr_timeout) < 0) 497 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 498 /* 499 * First handshake step. 500 * Setup outgoing connection with remote node. 501 */ 502 nvout = nv_alloc(); 503 nv_add_string(nvout, res->hr_name, "resource"); 504 if (nv_error(nvout) != 0) { 505 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 506 "Unable to allocate header for connection with %s", 507 res->hr_remoteaddr); 508 nv_free(nvout); 509 goto close; 510 } 511 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) { 512 pjdlog_errno(LOG_WARNING, 513 "Unable to send handshake header to %s", 514 res->hr_remoteaddr); 515 nv_free(nvout); 516 goto close; 517 } 518 nv_free(nvout); 519 if (hast_proto_recv_hdr(out, &nvin) < 0) { 520 pjdlog_errno(LOG_WARNING, 521 "Unable to receive handshake header from %s", 522 res->hr_remoteaddr); 523 goto close; 524 } 525 errmsg = nv_get_string(nvin, "errmsg"); 526 if (errmsg != NULL) { 527 pjdlog_warning("%s", errmsg); 528 nv_free(nvin); 529 goto close; 530 } 531 token = nv_get_uint8_array(nvin, &size, "token"); 532 if (token == NULL) { 533 pjdlog_warning("Handshake header from %s has no 'token' field.", 534 res->hr_remoteaddr); 535 nv_free(nvin); 536 goto close; 537 } 538 if (size != sizeof(res->hr_token)) { 539 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 540 res->hr_remoteaddr, size, sizeof(res->hr_token)); 541 nv_free(nvin); 542 goto close; 543 } 544 bcopy(token, res->hr_token, sizeof(res->hr_token)); 545 nv_free(nvin); 546 547 /* 548 * Second handshake step. 549 * Setup incoming connection with remote node. 550 */ 551 if (proto_client(res->hr_remoteaddr, &in) < 0) { 552 pjdlog_errno(LOG_WARNING, "Unable to create connection to %s", 553 res->hr_remoteaddr); 554 } 555 /* Try to connect, but accept failure. */ 556 if (proto_connect(in) < 0) { 557 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 558 res->hr_remoteaddr); 559 goto close; 560 } 561 /* Error in setting timeout is not critical, but why should it fail? */ 562 if (proto_timeout(in, res->hr_timeout) < 0) 563 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 564 nvout = nv_alloc(); 565 nv_add_string(nvout, res->hr_name, "resource"); 566 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 567 "token"); 568 nv_add_uint64(nvout, res->hr_resuid, "resuid"); 569 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 570 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 571 if (nv_error(nvout) != 0) { 572 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 573 "Unable to allocate header for connection with %s", 574 res->hr_remoteaddr); 575 nv_free(nvout); 576 goto close; 577 } 578 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) { 579 pjdlog_errno(LOG_WARNING, 580 "Unable to send handshake header to %s", 581 res->hr_remoteaddr); 582 nv_free(nvout); 583 goto close; 584 } 585 nv_free(nvout); 586 if (hast_proto_recv_hdr(out, &nvin) < 0) { 587 pjdlog_errno(LOG_WARNING, 588 "Unable to receive handshake header from %s", 589 res->hr_remoteaddr); 590 goto close; 591 } 592 errmsg = nv_get_string(nvin, "errmsg"); 593 if (errmsg != NULL) { 594 pjdlog_warning("%s", errmsg); 595 nv_free(nvin); 596 goto close; 597 } 598 datasize = nv_get_int64(nvin, "datasize"); 599 if (datasize != res->hr_datasize) { 600 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 601 (intmax_t)res->hr_datasize, (intmax_t)datasize); 602 nv_free(nvin); 603 goto close; 604 } 605 extentsize = nv_get_int32(nvin, "extentsize"); 606 if (extentsize != res->hr_extentsize) { 607 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 608 (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 609 nv_free(nvin); 610 goto close; 611 } 612 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 613 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 614 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 615 map = NULL; 616 mapsize = nv_get_uint32(nvin, "mapsize"); 617 if (mapsize > 0) { 618 map = malloc(mapsize); 619 if (map == NULL) { 620 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 621 (uintmax_t)mapsize); 622 nv_free(nvin); 623 goto close; 624 } 625 /* 626 * Remote node have some dirty extents on its own, lets 627 * download its activemap. 628 */ 629 if (hast_proto_recv_data(res, out, nvin, map, 630 mapsize) < 0) { 631 pjdlog_errno(LOG_ERR, 632 "Unable to receive remote activemap"); 633 nv_free(nvin); 634 free(map); 635 goto close; 636 } 637 /* 638 * Merge local and remote bitmaps. 639 */ 640 activemap_merge(res->hr_amp, map, mapsize); 641 free(map); 642 /* 643 * Now that we merged bitmaps from both nodes, flush it to the 644 * disk before we start to synchronize. 645 */ 646 (void)hast_activemap_flush(res); 647 } 648 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 649 if (inp != NULL && outp != NULL) { 650 *inp = in; 651 *outp = out; 652 } else { 653 res->hr_remotein = in; 654 res->hr_remoteout = out; 655 } 656 event_send(res, EVENT_CONNECT); 657 return (true); 658close: 659 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 660 event_send(res, EVENT_SPLITBRAIN); 661 proto_close(out); 662 if (in != NULL) 663 proto_close(in); 664 return (false); 665} 666 667static void 668sync_start(void) 669{ 670 671 mtx_lock(&sync_lock); 672 sync_inprogress = true; 673 mtx_unlock(&sync_lock); 674 cv_signal(&sync_cond); 675} 676 677static void 678sync_stop(void) 679{ 680 681 mtx_lock(&sync_lock); 682 if (sync_inprogress) 683 sync_inprogress = false; 684 mtx_unlock(&sync_lock); 685} 686 687static void 688init_ggate(struct hast_resource *res) 689{ 690 struct g_gate_ctl_create ggiocreate; 691 struct g_gate_ctl_cancel ggiocancel; 692 693 /* 694 * We communicate with ggate via /dev/ggctl. Open it. 695 */ 696 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 697 if (res->hr_ggatefd < 0) 698 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 699 /* 700 * Create provider before trying to connect, as connection failure 701 * is not critical, but may take some time. 702 */ 703 ggiocreate.gctl_version = G_GATE_VERSION; 704 ggiocreate.gctl_mediasize = res->hr_datasize; 705 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 706 ggiocreate.gctl_flags = 0; 707 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE; 708 ggiocreate.gctl_timeout = 0; 709 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 710 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 711 res->hr_provname); 712 bzero(ggiocreate.gctl_info, sizeof(ggiocreate.gctl_info)); 713 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 714 pjdlog_info("Device hast/%s created.", res->hr_provname); 715 res->hr_ggateunit = ggiocreate.gctl_unit; 716 return; 717 } 718 if (errno != EEXIST) { 719 primary_exit(EX_OSERR, "Unable to create hast/%s device", 720 res->hr_provname); 721 } 722 pjdlog_debug(1, 723 "Device hast/%s already exists, we will try to take it over.", 724 res->hr_provname); 725 /* 726 * If we received EEXIST, we assume that the process who created the 727 * provider died and didn't clean up. In that case we will start from 728 * where he left of. 729 */ 730 ggiocancel.gctl_version = G_GATE_VERSION; 731 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 732 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 733 res->hr_provname); 734 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 735 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 736 res->hr_ggateunit = ggiocancel.gctl_unit; 737 return; 738 } 739 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 740 res->hr_provname); 741} 742 743void 744hastd_primary(struct hast_resource *res) 745{ 746 pthread_t td; 747 pid_t pid; 748 int error; 749 750 /* 751 * Create communication channel between parent and child. 752 */ 753 if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 754 KEEP_ERRNO((void)pidfile_remove(pfh)); 755 pjdlog_exit(EX_OSERR, 756 "Unable to create control sockets between parent and child"); 757 } 758 /* 759 * Create communication channel between child and parent. 760 */ 761 if (proto_client("socketpair://", &res->hr_event) < 0) { 762 KEEP_ERRNO((void)pidfile_remove(pfh)); 763 pjdlog_exit(EX_OSERR, 764 "Unable to create event sockets between child and parent"); 765 } 766 767 pid = fork(); 768 if (pid < 0) { 769 KEEP_ERRNO((void)pidfile_remove(pfh)); 770 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 771 } 772 773 if (pid > 0) { 774 /* This is parent. */ 775 /* Declare that we are receiver. */ 776 proto_recv(res->hr_event, NULL, 0); 777 res->hr_workerpid = pid; 778 return; 779 } 780 781 gres = res; 782 783 (void)pidfile_close(pfh); 784 hook_fini(); 785 786 setproctitle("%s (primary)", res->hr_name); 787 788 /* Declare that we are sender. */ 789 proto_send(res->hr_event, NULL, 0); 790 791 init_local(res); 792 init_ggate(res); 793 init_environment(res); 794 /* 795 * Create the guard thread first, so we can handle signals from the 796 * very begining. 797 */ 798 error = pthread_create(&td, NULL, guard_thread, res); 799 assert(error == 0); 800 /* 801 * Create the control thread before sending any event to the parent, 802 * as we can deadlock when parent sends control request to worker, 803 * but worker has no control thread started yet, so parent waits. 804 * In the meantime worker sends an event to the parent, but parent 805 * is unable to handle the event, because it waits for control 806 * request response. 807 */ 808 error = pthread_create(&td, NULL, ctrl_thread, res); 809 assert(error == 0); 810 if (real_remote(res) && init_remote(res, NULL, NULL)) 811 sync_start(); 812 error = pthread_create(&td, NULL, ggate_recv_thread, res); 813 assert(error == 0); 814 error = pthread_create(&td, NULL, local_send_thread, res); 815 assert(error == 0); 816 error = pthread_create(&td, NULL, remote_send_thread, res); 817 assert(error == 0); 818 error = pthread_create(&td, NULL, remote_recv_thread, res); 819 assert(error == 0); 820 error = pthread_create(&td, NULL, ggate_send_thread, res); 821 assert(error == 0); 822 (void)sync_thread(res); 823} 824 825static void 826reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 827{ 828 char msg[1024]; 829 va_list ap; 830 int len; 831 832 va_start(ap, fmt); 833 len = vsnprintf(msg, sizeof(msg), fmt, ap); 834 va_end(ap); 835 if ((size_t)len < sizeof(msg)) { 836 switch (ggio->gctl_cmd) { 837 case BIO_READ: 838 (void)snprintf(msg + len, sizeof(msg) - len, 839 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 840 (uintmax_t)ggio->gctl_length); 841 break; 842 case BIO_DELETE: 843 (void)snprintf(msg + len, sizeof(msg) - len, 844 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 845 (uintmax_t)ggio->gctl_length); 846 break; 847 case BIO_FLUSH: 848 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 849 break; 850 case BIO_WRITE: 851 (void)snprintf(msg + len, sizeof(msg) - len, 852 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 853 (uintmax_t)ggio->gctl_length); 854 break; 855 default: 856 (void)snprintf(msg + len, sizeof(msg) - len, 857 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 858 break; 859 } 860 } 861 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 862} 863 864static void 865remote_close(struct hast_resource *res, int ncomp) 866{ 867 868 rw_wlock(&hio_remote_lock[ncomp]); 869 /* 870 * A race is possible between dropping rlock and acquiring wlock - 871 * another thread can close connection in-between. 872 */ 873 if (!ISCONNECTED(res, ncomp)) { 874 assert(res->hr_remotein == NULL); 875 assert(res->hr_remoteout == NULL); 876 rw_unlock(&hio_remote_lock[ncomp]); 877 return; 878 } 879 880 assert(res->hr_remotein != NULL); 881 assert(res->hr_remoteout != NULL); 882 883 pjdlog_debug(2, "Closing incoming connection to %s.", 884 res->hr_remoteaddr); 885 proto_close(res->hr_remotein); 886 res->hr_remotein = NULL; 887 pjdlog_debug(2, "Closing outgoing connection to %s.", 888 res->hr_remoteaddr); 889 proto_close(res->hr_remoteout); 890 res->hr_remoteout = NULL; 891 892 rw_unlock(&hio_remote_lock[ncomp]); 893 894 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 895 896 /* 897 * Stop synchronization if in-progress. 898 */ 899 sync_stop(); 900 901 event_send(res, EVENT_DISCONNECT); 902} 903 904/* 905 * Thread receives ggate I/O requests from the kernel and passes them to 906 * appropriate threads: 907 * WRITE - always goes to both local_send and remote_send threads 908 * READ (when the block is up-to-date on local component) - 909 * only local_send thread 910 * READ (when the block isn't up-to-date on local component) - 911 * only remote_send thread 912 * DELETE - always goes to both local_send and remote_send threads 913 * FLUSH - always goes to both local_send and remote_send threads 914 */ 915static void * 916ggate_recv_thread(void *arg) 917{ 918 struct hast_resource *res = arg; 919 struct g_gate_ctl_io *ggio; 920 struct hio *hio; 921 unsigned int ii, ncomp, ncomps; 922 int error; 923 924 ncomps = HAST_NCOMPONENTS; 925 926 for (;;) { 927 pjdlog_debug(2, "ggate_recv: Taking free request."); 928 QUEUE_TAKE2(hio, free); 929 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 930 ggio = &hio->hio_ggio; 931 ggio->gctl_unit = res->hr_ggateunit; 932 ggio->gctl_length = MAXPHYS; 933 ggio->gctl_error = 0; 934 pjdlog_debug(2, 935 "ggate_recv: (%p) Waiting for request from the kernel.", 936 hio); 937 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 938 if (sigexit_received) 939 pthread_exit(NULL); 940 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 941 } 942 error = ggio->gctl_error; 943 switch (error) { 944 case 0: 945 break; 946 case ECANCELED: 947 /* Exit gracefully. */ 948 if (!sigexit_received) { 949 pjdlog_debug(2, 950 "ggate_recv: (%p) Received cancel from the kernel.", 951 hio); 952 pjdlog_info("Received cancel from the kernel, exiting."); 953 } 954 pthread_exit(NULL); 955 case ENOMEM: 956 /* 957 * Buffer too small? Impossible, we allocate MAXPHYS 958 * bytes - request can't be bigger than that. 959 */ 960 /* FALLTHROUGH */ 961 case ENXIO: 962 default: 963 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 964 strerror(error)); 965 } 966 for (ii = 0; ii < ncomps; ii++) 967 hio->hio_errors[ii] = EINVAL; 968 reqlog(LOG_DEBUG, 2, ggio, 969 "ggate_recv: (%p) Request received from the kernel: ", 970 hio); 971 /* 972 * Inform all components about new write request. 973 * For read request prefer local component unless the given 974 * range is out-of-date, then use remote component. 975 */ 976 switch (ggio->gctl_cmd) { 977 case BIO_READ: 978 pjdlog_debug(2, 979 "ggate_recv: (%p) Moving request to the send queue.", 980 hio); 981 refcount_init(&hio->hio_countdown, 1); 982 mtx_lock(&metadata_lock); 983 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 984 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 985 /* 986 * This range is up-to-date on local component, 987 * so handle request locally. 988 */ 989 /* Local component is 0 for now. */ 990 ncomp = 0; 991 } else /* if (res->hr_syncsrc == 992 HAST_SYNCSRC_SECONDARY) */ { 993 assert(res->hr_syncsrc == 994 HAST_SYNCSRC_SECONDARY); 995 /* 996 * This range is out-of-date on local component, 997 * so send request to the remote node. 998 */ 999 /* Remote component is 1 for now. */ 1000 ncomp = 1; 1001 } 1002 mtx_unlock(&metadata_lock); 1003 QUEUE_INSERT1(hio, send, ncomp); 1004 break; 1005 case BIO_WRITE: 1006 for (;;) { 1007 mtx_lock(&range_lock); 1008 if (rangelock_islocked(range_sync, 1009 ggio->gctl_offset, ggio->gctl_length)) { 1010 pjdlog_debug(2, 1011 "regular: Range offset=%jd length=%zu locked.", 1012 (intmax_t)ggio->gctl_offset, 1013 (size_t)ggio->gctl_length); 1014 range_regular_wait = true; 1015 cv_wait(&range_regular_cond, &range_lock); 1016 range_regular_wait = false; 1017 mtx_unlock(&range_lock); 1018 continue; 1019 } 1020 if (rangelock_add(range_regular, 1021 ggio->gctl_offset, ggio->gctl_length) < 0) { 1022 mtx_unlock(&range_lock); 1023 pjdlog_debug(2, 1024 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1025 (intmax_t)ggio->gctl_offset, 1026 (size_t)ggio->gctl_length); 1027 sleep(1); 1028 continue; 1029 } 1030 mtx_unlock(&range_lock); 1031 break; 1032 } 1033 mtx_lock(&res->hr_amp_lock); 1034 if (activemap_write_start(res->hr_amp, 1035 ggio->gctl_offset, ggio->gctl_length)) { 1036 (void)hast_activemap_flush(res); 1037 } 1038 mtx_unlock(&res->hr_amp_lock); 1039 /* FALLTHROUGH */ 1040 case BIO_DELETE: 1041 case BIO_FLUSH: 1042 pjdlog_debug(2, 1043 "ggate_recv: (%p) Moving request to the send queues.", 1044 hio); 1045 refcount_init(&hio->hio_countdown, ncomps); 1046 for (ii = 0; ii < ncomps; ii++) 1047 QUEUE_INSERT1(hio, send, ii); 1048 break; 1049 } 1050 } 1051 /* NOTREACHED */ 1052 return (NULL); 1053} 1054 1055/* 1056 * Thread reads from or writes to local component. 1057 * If local read fails, it redirects it to remote_send thread. 1058 */ 1059static void * 1060local_send_thread(void *arg) 1061{ 1062 struct hast_resource *res = arg; 1063 struct g_gate_ctl_io *ggio; 1064 struct hio *hio; 1065 unsigned int ncomp, rncomp; 1066 ssize_t ret; 1067 1068 /* Local component is 0 for now. */ 1069 ncomp = 0; 1070 /* Remote component is 1 for now. */ 1071 rncomp = 1; 1072 1073 for (;;) { 1074 pjdlog_debug(2, "local_send: Taking request."); 1075 QUEUE_TAKE1(hio, send, ncomp); 1076 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1077 ggio = &hio->hio_ggio; 1078 switch (ggio->gctl_cmd) { 1079 case BIO_READ: 1080 ret = pread(res->hr_localfd, ggio->gctl_data, 1081 ggio->gctl_length, 1082 ggio->gctl_offset + res->hr_localoff); 1083 if (ret == ggio->gctl_length) 1084 hio->hio_errors[ncomp] = 0; 1085 else { 1086 /* 1087 * If READ failed, try to read from remote node. 1088 */ 1089 QUEUE_INSERT1(hio, send, rncomp); 1090 continue; 1091 } 1092 break; 1093 case BIO_WRITE: 1094 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1095 ggio->gctl_length, 1096 ggio->gctl_offset + res->hr_localoff); 1097 if (ret < 0) 1098 hio->hio_errors[ncomp] = errno; 1099 else if (ret != ggio->gctl_length) 1100 hio->hio_errors[ncomp] = EIO; 1101 else 1102 hio->hio_errors[ncomp] = 0; 1103 break; 1104 case BIO_DELETE: 1105 ret = g_delete(res->hr_localfd, 1106 ggio->gctl_offset + res->hr_localoff, 1107 ggio->gctl_length); 1108 if (ret < 0) 1109 hio->hio_errors[ncomp] = errno; 1110 else 1111 hio->hio_errors[ncomp] = 0; 1112 break; 1113 case BIO_FLUSH: 1114 ret = g_flush(res->hr_localfd); 1115 if (ret < 0) 1116 hio->hio_errors[ncomp] = errno; 1117 else 1118 hio->hio_errors[ncomp] = 0; 1119 break; 1120 } 1121 if (refcount_release(&hio->hio_countdown)) { 1122 if (ISSYNCREQ(hio)) { 1123 mtx_lock(&sync_lock); 1124 SYNCREQDONE(hio); 1125 mtx_unlock(&sync_lock); 1126 cv_signal(&sync_cond); 1127 } else { 1128 pjdlog_debug(2, 1129 "local_send: (%p) Moving request to the done queue.", 1130 hio); 1131 QUEUE_INSERT2(hio, done); 1132 } 1133 } 1134 } 1135 /* NOTREACHED */ 1136 return (NULL); 1137} 1138 1139/* 1140 * Thread sends request to secondary node. 1141 */ 1142static void * 1143remote_send_thread(void *arg) 1144{ 1145 struct hast_resource *res = arg; 1146 struct g_gate_ctl_io *ggio; 1147 struct hio *hio; 1148 struct nv *nv; 1149 unsigned int ncomp; 1150 bool wakeup; 1151 uint64_t offset, length; 1152 uint8_t cmd; 1153 void *data; 1154 1155 /* Remote component is 1 for now. */ 1156 ncomp = 1; 1157 1158 for (;;) { 1159 pjdlog_debug(2, "remote_send: Taking request."); 1160 QUEUE_TAKE1(hio, send, ncomp); 1161 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1162 ggio = &hio->hio_ggio; 1163 switch (ggio->gctl_cmd) { 1164 case BIO_READ: 1165 cmd = HIO_READ; 1166 data = NULL; 1167 offset = ggio->gctl_offset; 1168 length = ggio->gctl_length; 1169 break; 1170 case BIO_WRITE: 1171 cmd = HIO_WRITE; 1172 data = ggio->gctl_data; 1173 offset = ggio->gctl_offset; 1174 length = ggio->gctl_length; 1175 break; 1176 case BIO_DELETE: 1177 cmd = HIO_DELETE; 1178 data = NULL; 1179 offset = ggio->gctl_offset; 1180 length = ggio->gctl_length; 1181 break; 1182 case BIO_FLUSH: 1183 cmd = HIO_FLUSH; 1184 data = NULL; 1185 offset = 0; 1186 length = 0; 1187 break; 1188 default: 1189 assert(!"invalid condition"); 1190 abort(); 1191 } 1192 nv = nv_alloc(); 1193 nv_add_uint8(nv, cmd, "cmd"); 1194 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1195 nv_add_uint64(nv, offset, "offset"); 1196 nv_add_uint64(nv, length, "length"); 1197 if (nv_error(nv) != 0) { 1198 hio->hio_errors[ncomp] = nv_error(nv); 1199 pjdlog_debug(2, 1200 "remote_send: (%p) Unable to prepare header to send.", 1201 hio); 1202 reqlog(LOG_ERR, 0, ggio, 1203 "Unable to prepare header to send (%s): ", 1204 strerror(nv_error(nv))); 1205 /* Move failed request immediately to the done queue. */ 1206 goto done_queue; 1207 } 1208 pjdlog_debug(2, 1209 "remote_send: (%p) Moving request to the recv queue.", 1210 hio); 1211 /* 1212 * Protect connection from disappearing. 1213 */ 1214 rw_rlock(&hio_remote_lock[ncomp]); 1215 if (!ISCONNECTED(res, ncomp)) { 1216 rw_unlock(&hio_remote_lock[ncomp]); 1217 hio->hio_errors[ncomp] = ENOTCONN; 1218 goto done_queue; 1219 } 1220 /* 1221 * Move the request to recv queue before sending it, because 1222 * in different order we can get reply before we move request 1223 * to recv queue. 1224 */ 1225 mtx_lock(&hio_recv_list_lock[ncomp]); 1226 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1227 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1228 mtx_unlock(&hio_recv_list_lock[ncomp]); 1229 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1230 data != NULL ? length : 0) < 0) { 1231 hio->hio_errors[ncomp] = errno; 1232 rw_unlock(&hio_remote_lock[ncomp]); 1233 pjdlog_debug(2, 1234 "remote_send: (%p) Unable to send request.", hio); 1235 reqlog(LOG_ERR, 0, ggio, 1236 "Unable to send request (%s): ", 1237 strerror(hio->hio_errors[ncomp])); 1238 remote_close(res, ncomp); 1239 /* 1240 * Take request back from the receive queue and move 1241 * it immediately to the done queue. 1242 */ 1243 mtx_lock(&hio_recv_list_lock[ncomp]); 1244 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1245 mtx_unlock(&hio_recv_list_lock[ncomp]); 1246 goto done_queue; 1247 } 1248 rw_unlock(&hio_remote_lock[ncomp]); 1249 nv_free(nv); 1250 if (wakeup) 1251 cv_signal(&hio_recv_list_cond[ncomp]); 1252 continue; 1253done_queue: 1254 nv_free(nv); 1255 if (ISSYNCREQ(hio)) { 1256 if (!refcount_release(&hio->hio_countdown)) 1257 continue; 1258 mtx_lock(&sync_lock); 1259 SYNCREQDONE(hio); 1260 mtx_unlock(&sync_lock); 1261 cv_signal(&sync_cond); 1262 continue; 1263 } 1264 if (ggio->gctl_cmd == BIO_WRITE) { 1265 mtx_lock(&res->hr_amp_lock); 1266 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1267 ggio->gctl_length)) { 1268 (void)hast_activemap_flush(res); 1269 } 1270 mtx_unlock(&res->hr_amp_lock); 1271 } 1272 if (!refcount_release(&hio->hio_countdown)) 1273 continue; 1274 pjdlog_debug(2, 1275 "remote_send: (%p) Moving request to the done queue.", 1276 hio); 1277 QUEUE_INSERT2(hio, done); 1278 } 1279 /* NOTREACHED */ 1280 return (NULL); 1281} 1282 1283/* 1284 * Thread receives answer from secondary node and passes it to ggate_send 1285 * thread. 1286 */ 1287static void * 1288remote_recv_thread(void *arg) 1289{ 1290 struct hast_resource *res = arg; 1291 struct g_gate_ctl_io *ggio; 1292 struct hio *hio; 1293 struct nv *nv; 1294 unsigned int ncomp; 1295 uint64_t seq; 1296 int error; 1297 1298 /* Remote component is 1 for now. */ 1299 ncomp = 1; 1300 1301 for (;;) { 1302 /* Wait until there is anything to receive. */ 1303 mtx_lock(&hio_recv_list_lock[ncomp]); 1304 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1305 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1306 cv_wait(&hio_recv_list_cond[ncomp], 1307 &hio_recv_list_lock[ncomp]); 1308 } 1309 mtx_unlock(&hio_recv_list_lock[ncomp]); 1310 rw_rlock(&hio_remote_lock[ncomp]); 1311 if (!ISCONNECTED(res, ncomp)) { 1312 rw_unlock(&hio_remote_lock[ncomp]); 1313 /* 1314 * Connection is dead, so move all pending requests to 1315 * the done queue (one-by-one). 1316 */ 1317 mtx_lock(&hio_recv_list_lock[ncomp]); 1318 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1319 assert(hio != NULL); 1320 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1321 hio_next[ncomp]); 1322 mtx_unlock(&hio_recv_list_lock[ncomp]); 1323 goto done_queue; 1324 } 1325 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1326 pjdlog_errno(LOG_ERR, 1327 "Unable to receive reply header"); 1328 rw_unlock(&hio_remote_lock[ncomp]); 1329 remote_close(res, ncomp); 1330 continue; 1331 } 1332 rw_unlock(&hio_remote_lock[ncomp]); 1333 seq = nv_get_uint64(nv, "seq"); 1334 if (seq == 0) { 1335 pjdlog_error("Header contains no 'seq' field."); 1336 nv_free(nv); 1337 continue; 1338 } 1339 mtx_lock(&hio_recv_list_lock[ncomp]); 1340 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1341 if (hio->hio_ggio.gctl_seq == seq) { 1342 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1343 hio_next[ncomp]); 1344 break; 1345 } 1346 } 1347 mtx_unlock(&hio_recv_list_lock[ncomp]); 1348 if (hio == NULL) { 1349 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1350 (uintmax_t)seq); 1351 nv_free(nv); 1352 continue; 1353 } 1354 error = nv_get_int16(nv, "error"); 1355 if (error != 0) { 1356 /* Request failed on remote side. */ 1357 hio->hio_errors[ncomp] = 0; 1358 nv_free(nv); 1359 goto done_queue; 1360 } 1361 ggio = &hio->hio_ggio; 1362 switch (ggio->gctl_cmd) { 1363 case BIO_READ: 1364 rw_rlock(&hio_remote_lock[ncomp]); 1365 if (!ISCONNECTED(res, ncomp)) { 1366 rw_unlock(&hio_remote_lock[ncomp]); 1367 nv_free(nv); 1368 goto done_queue; 1369 } 1370 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1371 ggio->gctl_data, ggio->gctl_length) < 0) { 1372 hio->hio_errors[ncomp] = errno; 1373 pjdlog_errno(LOG_ERR, 1374 "Unable to receive reply data"); 1375 rw_unlock(&hio_remote_lock[ncomp]); 1376 nv_free(nv); 1377 remote_close(res, ncomp); 1378 goto done_queue; 1379 } 1380 rw_unlock(&hio_remote_lock[ncomp]); 1381 break; 1382 case BIO_WRITE: 1383 case BIO_DELETE: 1384 case BIO_FLUSH: 1385 break; 1386 default: 1387 assert(!"invalid condition"); 1388 abort(); 1389 } 1390 hio->hio_errors[ncomp] = 0; 1391 nv_free(nv); 1392done_queue: 1393 if (refcount_release(&hio->hio_countdown)) { 1394 if (ISSYNCREQ(hio)) { 1395 mtx_lock(&sync_lock); 1396 SYNCREQDONE(hio); 1397 mtx_unlock(&sync_lock); 1398 cv_signal(&sync_cond); 1399 } else { 1400 pjdlog_debug(2, 1401 "remote_recv: (%p) Moving request to the done queue.", 1402 hio); 1403 QUEUE_INSERT2(hio, done); 1404 } 1405 } 1406 } 1407 /* NOTREACHED */ 1408 return (NULL); 1409} 1410 1411/* 1412 * Thread sends answer to the kernel. 1413 */ 1414static void * 1415ggate_send_thread(void *arg) 1416{ 1417 struct hast_resource *res = arg; 1418 struct g_gate_ctl_io *ggio; 1419 struct hio *hio; 1420 unsigned int ii, ncomp, ncomps; 1421 1422 ncomps = HAST_NCOMPONENTS; 1423 1424 for (;;) { 1425 pjdlog_debug(2, "ggate_send: Taking request."); 1426 QUEUE_TAKE2(hio, done); 1427 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1428 ggio = &hio->hio_ggio; 1429 for (ii = 0; ii < ncomps; ii++) { 1430 if (hio->hio_errors[ii] == 0) { 1431 /* 1432 * One successful request is enough to declare 1433 * success. 1434 */ 1435 ggio->gctl_error = 0; 1436 break; 1437 } 1438 } 1439 if (ii == ncomps) { 1440 /* 1441 * None of the requests were successful. 1442 * Use first error. 1443 */ 1444 ggio->gctl_error = hio->hio_errors[0]; 1445 } 1446 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1447 mtx_lock(&res->hr_amp_lock); 1448 activemap_write_complete(res->hr_amp, 1449 ggio->gctl_offset, ggio->gctl_length); 1450 mtx_unlock(&res->hr_amp_lock); 1451 } 1452 if (ggio->gctl_cmd == BIO_WRITE) { 1453 /* 1454 * Unlock range we locked. 1455 */ 1456 mtx_lock(&range_lock); 1457 rangelock_del(range_regular, ggio->gctl_offset, 1458 ggio->gctl_length); 1459 if (range_sync_wait) 1460 cv_signal(&range_sync_cond); 1461 mtx_unlock(&range_lock); 1462 /* 1463 * Bump local count if this is first write after 1464 * connection failure with remote node. 1465 */ 1466 ncomp = 1; 1467 rw_rlock(&hio_remote_lock[ncomp]); 1468 if (!ISCONNECTED(res, ncomp)) { 1469 mtx_lock(&metadata_lock); 1470 if (res->hr_primary_localcnt == 1471 res->hr_secondary_remotecnt) { 1472 res->hr_primary_localcnt++; 1473 pjdlog_debug(1, 1474 "Increasing localcnt to %ju.", 1475 (uintmax_t)res->hr_primary_localcnt); 1476 (void)metadata_write(res); 1477 } 1478 mtx_unlock(&metadata_lock); 1479 } 1480 rw_unlock(&hio_remote_lock[ncomp]); 1481 } 1482 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1483 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1484 pjdlog_debug(2, 1485 "ggate_send: (%p) Moving request to the free queue.", hio); 1486 QUEUE_INSERT2(hio, free); 1487 } 1488 /* NOTREACHED */ 1489 return (NULL); 1490} 1491 1492/* 1493 * Thread synchronize local and remote components. 1494 */ 1495static void * 1496sync_thread(void *arg __unused) 1497{ 1498 struct hast_resource *res = arg; 1499 struct hio *hio; 1500 struct g_gate_ctl_io *ggio; 1501 unsigned int ii, ncomp, ncomps; 1502 off_t offset, length, synced; 1503 bool dorewind; 1504 int syncext; 1505 1506 ncomps = HAST_NCOMPONENTS; 1507 dorewind = true; 1508 synced = 0; 1509 offset = -1; 1510 1511 for (;;) { 1512 mtx_lock(&sync_lock); 1513 if (offset >= 0 && !sync_inprogress) { 1514 pjdlog_info("Synchronization interrupted. " 1515 "%jd bytes synchronized so far.", 1516 (intmax_t)synced); 1517 event_send(res, EVENT_SYNCINTR); 1518 } 1519 while (!sync_inprogress) { 1520 dorewind = true; 1521 synced = 0; 1522 cv_wait(&sync_cond, &sync_lock); 1523 } 1524 mtx_unlock(&sync_lock); 1525 /* 1526 * Obtain offset at which we should synchronize. 1527 * Rewind synchronization if needed. 1528 */ 1529 mtx_lock(&res->hr_amp_lock); 1530 if (dorewind) 1531 activemap_sync_rewind(res->hr_amp); 1532 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1533 if (syncext != -1) { 1534 /* 1535 * We synchronized entire syncext extent, we can mark 1536 * it as clean now. 1537 */ 1538 if (activemap_extent_complete(res->hr_amp, syncext)) 1539 (void)hast_activemap_flush(res); 1540 } 1541 mtx_unlock(&res->hr_amp_lock); 1542 if (dorewind) { 1543 dorewind = false; 1544 if (offset < 0) 1545 pjdlog_info("Nodes are in sync."); 1546 else { 1547 pjdlog_info("Synchronization started. %ju bytes to go.", 1548 (uintmax_t)(res->hr_extentsize * 1549 activemap_ndirty(res->hr_amp))); 1550 event_send(res, EVENT_SYNCSTART); 1551 } 1552 } 1553 if (offset < 0) { 1554 sync_stop(); 1555 pjdlog_debug(1, "Nothing to synchronize."); 1556 /* 1557 * Synchronization complete, make both localcnt and 1558 * remotecnt equal. 1559 */ 1560 ncomp = 1; 1561 rw_rlock(&hio_remote_lock[ncomp]); 1562 if (ISCONNECTED(res, ncomp)) { 1563 if (synced > 0) { 1564 pjdlog_info("Synchronization complete. " 1565 "%jd bytes synchronized.", 1566 (intmax_t)synced); 1567 event_send(res, EVENT_SYNCDONE); 1568 } 1569 mtx_lock(&metadata_lock); 1570 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1571 res->hr_primary_localcnt = 1572 res->hr_secondary_localcnt; 1573 res->hr_primary_remotecnt = 1574 res->hr_secondary_remotecnt; 1575 pjdlog_debug(1, 1576 "Setting localcnt to %ju and remotecnt to %ju.", 1577 (uintmax_t)res->hr_primary_localcnt, 1578 (uintmax_t)res->hr_secondary_localcnt); 1579 (void)metadata_write(res); 1580 mtx_unlock(&metadata_lock); 1581 } 1582 rw_unlock(&hio_remote_lock[ncomp]); 1583 continue; 1584 } 1585 pjdlog_debug(2, "sync: Taking free request."); 1586 QUEUE_TAKE2(hio, free); 1587 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1588 /* 1589 * Lock the range we are going to synchronize. We don't want 1590 * race where someone writes between our read and write. 1591 */ 1592 for (;;) { 1593 mtx_lock(&range_lock); 1594 if (rangelock_islocked(range_regular, offset, length)) { 1595 pjdlog_debug(2, 1596 "sync: Range offset=%jd length=%jd locked.", 1597 (intmax_t)offset, (intmax_t)length); 1598 range_sync_wait = true; 1599 cv_wait(&range_sync_cond, &range_lock); 1600 range_sync_wait = false; 1601 mtx_unlock(&range_lock); 1602 continue; 1603 } 1604 if (rangelock_add(range_sync, offset, length) < 0) { 1605 mtx_unlock(&range_lock); 1606 pjdlog_debug(2, 1607 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1608 (intmax_t)offset, (intmax_t)length); 1609 sleep(1); 1610 continue; 1611 } 1612 mtx_unlock(&range_lock); 1613 break; 1614 } 1615 /* 1616 * First read the data from synchronization source. 1617 */ 1618 SYNCREQ(hio); 1619 ggio = &hio->hio_ggio; 1620 ggio->gctl_cmd = BIO_READ; 1621 ggio->gctl_offset = offset; 1622 ggio->gctl_length = length; 1623 ggio->gctl_error = 0; 1624 for (ii = 0; ii < ncomps; ii++) 1625 hio->hio_errors[ii] = EINVAL; 1626 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1627 hio); 1628 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1629 hio); 1630 mtx_lock(&metadata_lock); 1631 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1632 /* 1633 * This range is up-to-date on local component, 1634 * so handle request locally. 1635 */ 1636 /* Local component is 0 for now. */ 1637 ncomp = 0; 1638 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1639 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1640 /* 1641 * This range is out-of-date on local component, 1642 * so send request to the remote node. 1643 */ 1644 /* Remote component is 1 for now. */ 1645 ncomp = 1; 1646 } 1647 mtx_unlock(&metadata_lock); 1648 refcount_init(&hio->hio_countdown, 1); 1649 QUEUE_INSERT1(hio, send, ncomp); 1650 1651 /* 1652 * Let's wait for READ to finish. 1653 */ 1654 mtx_lock(&sync_lock); 1655 while (!ISSYNCREQDONE(hio)) 1656 cv_wait(&sync_cond, &sync_lock); 1657 mtx_unlock(&sync_lock); 1658 1659 if (hio->hio_errors[ncomp] != 0) { 1660 pjdlog_error("Unable to read synchronization data: %s.", 1661 strerror(hio->hio_errors[ncomp])); 1662 goto free_queue; 1663 } 1664 1665 /* 1666 * We read the data from synchronization source, now write it 1667 * to synchronization target. 1668 */ 1669 SYNCREQ(hio); 1670 ggio->gctl_cmd = BIO_WRITE; 1671 for (ii = 0; ii < ncomps; ii++) 1672 hio->hio_errors[ii] = EINVAL; 1673 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1674 hio); 1675 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1676 hio); 1677 mtx_lock(&metadata_lock); 1678 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1679 /* 1680 * This range is up-to-date on local component, 1681 * so we update remote component. 1682 */ 1683 /* Remote component is 1 for now. */ 1684 ncomp = 1; 1685 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1686 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1687 /* 1688 * This range is out-of-date on local component, 1689 * so we update it. 1690 */ 1691 /* Local component is 0 for now. */ 1692 ncomp = 0; 1693 } 1694 mtx_unlock(&metadata_lock); 1695 1696 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1697 hio); 1698 refcount_init(&hio->hio_countdown, 1); 1699 QUEUE_INSERT1(hio, send, ncomp); 1700 1701 /* 1702 * Let's wait for WRITE to finish. 1703 */ 1704 mtx_lock(&sync_lock); 1705 while (!ISSYNCREQDONE(hio)) 1706 cv_wait(&sync_cond, &sync_lock); 1707 mtx_unlock(&sync_lock); 1708 1709 if (hio->hio_errors[ncomp] != 0) { 1710 pjdlog_error("Unable to write synchronization data: %s.", 1711 strerror(hio->hio_errors[ncomp])); 1712 goto free_queue; 1713 } 1714 1715 synced += length; 1716free_queue: 1717 mtx_lock(&range_lock); 1718 rangelock_del(range_sync, offset, length); 1719 if (range_regular_wait) 1720 cv_signal(&range_regular_cond); 1721 mtx_unlock(&range_lock); 1722 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1723 hio); 1724 QUEUE_INSERT2(hio, free); 1725 } 1726 /* NOTREACHED */ 1727 return (NULL); 1728} 1729 1730static void 1731config_reload(void) 1732{ 1733 struct hastd_config *newcfg; 1734 struct hast_resource *res; 1735 unsigned int ii, ncomps; 1736 int modified; 1737 1738 pjdlog_info("Reloading configuration..."); 1739 1740 ncomps = HAST_NCOMPONENTS; 1741 1742 newcfg = yy_config_parse(cfgpath, false); 1743 if (newcfg == NULL) 1744 goto failed; 1745 1746 TAILQ_FOREACH(res, &newcfg->hc_resources, hr_next) { 1747 if (strcmp(res->hr_name, gres->hr_name) == 0) 1748 break; 1749 } 1750 /* 1751 * If resource was removed from the configuration file, resource 1752 * name, provider name or path to local component was modified we 1753 * shouldn't be here. This means that someone modified configuration 1754 * file and send SIGHUP to us instead of main hastd process. 1755 * Log advice and ignore the signal. 1756 */ 1757 if (res == NULL || strcmp(gres->hr_name, res->hr_name) != 0 || 1758 strcmp(gres->hr_provname, res->hr_provname) != 0 || 1759 strcmp(gres->hr_localpath, res->hr_localpath) != 0) { 1760 pjdlog_warning("To reload configuration send SIGHUP to the main hastd process (pid %u).", 1761 (unsigned int)getppid()); 1762 goto failed; 1763 } 1764 1765#define MODIFIED_REMOTEADDR 0x1 1766#define MODIFIED_REPLICATION 0x2 1767#define MODIFIED_TIMEOUT 0x4 1768#define MODIFIED_EXEC 0x8 1769 modified = 0; 1770 if (strcmp(gres->hr_remoteaddr, res->hr_remoteaddr) != 0) { 1771 /* 1772 * Don't copy res->hr_remoteaddr to gres just yet. 1773 * We want remote_close() to log disconnect from the old 1774 * addresses, not from the new ones. 1775 */ 1776 modified |= MODIFIED_REMOTEADDR; 1777 } 1778 if (gres->hr_replication != res->hr_replication) { 1779 gres->hr_replication = res->hr_replication; 1780 modified |= MODIFIED_REPLICATION; 1781 } 1782 if (gres->hr_timeout != res->hr_timeout) { 1783 gres->hr_timeout = res->hr_timeout; 1784 modified |= MODIFIED_TIMEOUT; 1785 } 1786 if (strcmp(gres->hr_exec, res->hr_exec) != 0) { 1787 strlcpy(gres->hr_exec, res->hr_exec, sizeof(gres->hr_exec)); 1788 modified |= MODIFIED_EXEC; 1789 } 1790 /* 1791 * If only timeout was modified we only need to change it without 1792 * reconnecting. 1793 */ 1794 if (modified == MODIFIED_TIMEOUT) { 1795 for (ii = 0; ii < ncomps; ii++) { 1796 if (!ISREMOTE(ii)) 1797 continue; 1798 rw_rlock(&hio_remote_lock[ii]); 1799 if (!ISCONNECTED(gres, ii)) { 1800 rw_unlock(&hio_remote_lock[ii]); 1801 continue; 1802 } 1803 rw_unlock(&hio_remote_lock[ii]); 1804 if (proto_timeout(gres->hr_remotein, 1805 gres->hr_timeout) < 0) { 1806 pjdlog_errno(LOG_WARNING, 1807 "Unable to set connection timeout"); 1808 } 1809 if (proto_timeout(gres->hr_remoteout, 1810 gres->hr_timeout) < 0) { 1811 pjdlog_errno(LOG_WARNING, 1812 "Unable to set connection timeout"); 1813 } 1814 } 1815 } else if ((modified & 1816 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) { 1817 for (ii = 0; ii < ncomps; ii++) { 1818 if (!ISREMOTE(ii)) 1819 continue; 1820 remote_close(gres, ii); 1821 } 1822 if (modified & MODIFIED_REMOTEADDR) { 1823 strlcpy(gres->hr_remoteaddr, res->hr_remoteaddr, 1824 sizeof(gres->hr_remoteaddr)); 1825 } 1826 } 1827#undef MODIFIED_REMOTEADDR 1828#undef MODIFIED_REPLICATION 1829#undef MODIFIED_TIMEOUT 1830#undef MODIFIED_EXEC 1831 1832 pjdlog_info("Configuration reloaded successfully."); 1833 return; 1834failed: 1835 if (newcfg != NULL) { 1836 if (newcfg->hc_controlconn != NULL) 1837 proto_close(newcfg->hc_controlconn); 1838 if (newcfg->hc_listenconn != NULL) 1839 proto_close(newcfg->hc_listenconn); 1840 yy_config_free(newcfg); 1841 } 1842 pjdlog_warning("Configuration not reloaded."); 1843} 1844 1845static void 1846keepalive_send(struct hast_resource *res, unsigned int ncomp) 1847{ 1848 struct nv *nv; 1849 1850 nv = nv_alloc(); 1851 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1852 if (nv_error(nv) != 0) { 1853 nv_free(nv); 1854 pjdlog_debug(1, 1855 "keepalive_send: Unable to prepare header to send."); 1856 return; 1857 } 1858 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1859 pjdlog_common(LOG_DEBUG, 1, errno, 1860 "keepalive_send: Unable to send request"); 1861 nv_free(nv); 1862 rw_unlock(&hio_remote_lock[ncomp]); 1863 remote_close(res, ncomp); 1864 rw_rlock(&hio_remote_lock[ncomp]); 1865 return; 1866 } 1867 nv_free(nv); 1868 pjdlog_debug(2, "keepalive_send: Request sent."); 1869} 1870 1871static void 1872guard_one(struct hast_resource *res, unsigned int ncomp) 1873{ 1874 struct proto_conn *in, *out; 1875 1876 if (!ISREMOTE(ncomp)) 1877 return; 1878 1879 rw_rlock(&hio_remote_lock[ncomp]); 1880 1881 if (!real_remote(res)) { 1882 rw_unlock(&hio_remote_lock[ncomp]); 1883 return; 1884 } 1885 1886 if (ISCONNECTED(res, ncomp)) { 1887 assert(res->hr_remotein != NULL); 1888 assert(res->hr_remoteout != NULL); 1889 keepalive_send(res, ncomp); 1890 } 1891 1892 if (ISCONNECTED(res, ncomp)) { 1893 assert(res->hr_remotein != NULL); 1894 assert(res->hr_remoteout != NULL); 1895 rw_unlock(&hio_remote_lock[ncomp]); 1896 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 1897 res->hr_remoteaddr); 1898 return; 1899 } 1900 1901 assert(res->hr_remotein == NULL); 1902 assert(res->hr_remoteout == NULL); 1903 /* 1904 * Upgrade the lock. It doesn't have to be atomic as no other thread 1905 * can change connection status from disconnected to connected. 1906 */ 1907 rw_unlock(&hio_remote_lock[ncomp]); 1908 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 1909 res->hr_remoteaddr); 1910 in = out = NULL; 1911 if (init_remote(res, &in, &out)) { 1912 rw_wlock(&hio_remote_lock[ncomp]); 1913 assert(res->hr_remotein == NULL); 1914 assert(res->hr_remoteout == NULL); 1915 assert(in != NULL && out != NULL); 1916 res->hr_remotein = in; 1917 res->hr_remoteout = out; 1918 rw_unlock(&hio_remote_lock[ncomp]); 1919 pjdlog_info("Successfully reconnected to %s.", 1920 res->hr_remoteaddr); 1921 sync_start(); 1922 } else { 1923 /* Both connections should be NULL. */ 1924 assert(res->hr_remotein == NULL); 1925 assert(res->hr_remoteout == NULL); 1926 assert(in == NULL && out == NULL); 1927 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 1928 res->hr_remoteaddr); 1929 } 1930} 1931 1932/* 1933 * Thread guards remote connections and reconnects when needed, handles 1934 * signals, etc. 1935 */ 1936static void * 1937guard_thread(void *arg) 1938{ 1939 struct hast_resource *res = arg; 1940 unsigned int ii, ncomps; 1941 struct timespec timeout; 1942 time_t lastcheck, now; 1943 sigset_t mask; 1944 int signo; 1945 1946 ncomps = HAST_NCOMPONENTS; 1947 lastcheck = time(NULL); 1948 1949 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 1950 PJDLOG_VERIFY(sigaddset(&mask, SIGHUP) == 0); 1951 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 1952 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 1953 1954 timeout.tv_nsec = 0; 1955 signo = -1; 1956 1957 for (;;) { 1958 switch (signo) { 1959 case SIGHUP: 1960 config_reload(); 1961 break; 1962 case SIGINT: 1963 case SIGTERM: 1964 sigexit_received = true; 1965 primary_exitx(EX_OK, 1966 "Termination signal received, exiting."); 1967 break; 1968 default: 1969 break; 1970 } 1971 1972 pjdlog_debug(2, "remote_guard: Checking connections."); 1973 now = time(NULL); 1974 if (lastcheck + RETRY_SLEEP <= now) { 1975 for (ii = 0; ii < ncomps; ii++) 1976 guard_one(res, ii); 1977 lastcheck = now; 1978 } 1979 timeout.tv_sec = RETRY_SLEEP; 1980 signo = sigtimedwait(&mask, NULL, &timeout); 1981 } 1982 /* NOTREACHED */ 1983 return (NULL); 1984} 1985