primary.c revision 211983
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 211983 2010-08-30 00:12:10Z 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 "hast.h" 62#include "hast_proto.h" 63#include "hastd.h" 64#include "hooks.h" 65#include "metadata.h" 66#include "proto.h" 67#include "pjdlog.h" 68#include "subr.h" 69#include "synch.h" 70 71/* The is only one remote component for now. */ 72#define ISREMOTE(no) ((no) == 1) 73 74struct hio { 75 /* 76 * Number of components we are still waiting for. 77 * When this field goes to 0, we can send the request back to the 78 * kernel. Each component has to decrease this counter by one 79 * even on failure. 80 */ 81 unsigned int hio_countdown; 82 /* 83 * Each component has a place to store its own error. 84 * Once the request is handled by all components we can decide if the 85 * request overall is successful or not. 86 */ 87 int *hio_errors; 88 /* 89 * Structure used to comunicate with GEOM Gate class. 90 */ 91 struct g_gate_ctl_io hio_ggio; 92 TAILQ_ENTRY(hio) *hio_next; 93}; 94#define hio_free_next hio_next[0] 95#define hio_done_next hio_next[0] 96 97/* 98 * Free list holds unused structures. When free list is empty, we have to wait 99 * until some in-progress requests are freed. 100 */ 101static TAILQ_HEAD(, hio) hio_free_list; 102static pthread_mutex_t hio_free_list_lock; 103static pthread_cond_t hio_free_list_cond; 104/* 105 * There is one send list for every component. One requests is placed on all 106 * send lists - each component gets the same request, but each component is 107 * responsible for managing his own send list. 108 */ 109static TAILQ_HEAD(, hio) *hio_send_list; 110static pthread_mutex_t *hio_send_list_lock; 111static pthread_cond_t *hio_send_list_cond; 112/* 113 * There is one recv list for every component, although local components don't 114 * use recv lists as local requests are done synchronously. 115 */ 116static TAILQ_HEAD(, hio) *hio_recv_list; 117static pthread_mutex_t *hio_recv_list_lock; 118static pthread_cond_t *hio_recv_list_cond; 119/* 120 * Request is placed on done list by the slowest component (the one that 121 * decreased hio_countdown from 1 to 0). 122 */ 123static TAILQ_HEAD(, hio) hio_done_list; 124static pthread_mutex_t hio_done_list_lock; 125static pthread_cond_t hio_done_list_cond; 126/* 127 * Structure below are for interaction with sync thread. 128 */ 129static bool sync_inprogress; 130static pthread_mutex_t sync_lock; 131static pthread_cond_t sync_cond; 132/* 133 * The lock below allows to synchornize access to remote connections. 134 */ 135static pthread_rwlock_t *hio_remote_lock; 136 137/* 138 * Lock to synchronize metadata updates. Also synchronize access to 139 * hr_primary_localcnt and hr_primary_remotecnt fields. 140 */ 141static pthread_mutex_t metadata_lock; 142 143/* 144 * Maximum number of outstanding I/O requests. 145 */ 146#define HAST_HIO_MAX 256 147/* 148 * Number of components. At this point there are only two components: local 149 * and remote, but in the future it might be possible to use multiple local 150 * and remote components. 151 */ 152#define HAST_NCOMPONENTS 2 153/* 154 * Number of seconds to sleep between reconnect retries or keepalive packets. 155 */ 156#define RETRY_SLEEP 10 157 158#define ISCONNECTED(res, no) \ 159 ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL) 160 161#define QUEUE_INSERT1(hio, name, ncomp) do { \ 162 bool _wakeup; \ 163 \ 164 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 165 _wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]); \ 166 TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio), \ 167 hio_next[(ncomp)]); \ 168 mtx_unlock(&hio_##name##_list_lock[ncomp]); \ 169 if (_wakeup) \ 170 cv_signal(&hio_##name##_list_cond[(ncomp)]); \ 171} while (0) 172#define QUEUE_INSERT2(hio, name) do { \ 173 bool _wakeup; \ 174 \ 175 mtx_lock(&hio_##name##_list_lock); \ 176 _wakeup = TAILQ_EMPTY(&hio_##name##_list); \ 177 TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\ 178 mtx_unlock(&hio_##name##_list_lock); \ 179 if (_wakeup) \ 180 cv_signal(&hio_##name##_list_cond); \ 181} while (0) 182#define QUEUE_TAKE1(hio, name, ncomp) do { \ 183 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 184 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL) { \ 185 cv_wait(&hio_##name##_list_cond[(ncomp)], \ 186 &hio_##name##_list_lock[(ncomp)]); \ 187 } \ 188 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \ 189 hio_next[(ncomp)]); \ 190 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 191} while (0) 192#define QUEUE_TAKE2(hio, name) do { \ 193 mtx_lock(&hio_##name##_list_lock); \ 194 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 195 cv_wait(&hio_##name##_list_cond, \ 196 &hio_##name##_list_lock); \ 197 } \ 198 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \ 199 mtx_unlock(&hio_##name##_list_lock); \ 200} while (0) 201 202#define SYNCREQ(hio) do { \ 203 (hio)->hio_ggio.gctl_unit = -1; \ 204 (hio)->hio_ggio.gctl_seq = 1; \ 205} while (0) 206#define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1) 207#define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0) 208#define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2) 209 210static struct hast_resource *gres; 211 212static pthread_mutex_t range_lock; 213static struct rangelocks *range_regular; 214static bool range_regular_wait; 215static pthread_cond_t range_regular_cond; 216static struct rangelocks *range_sync; 217static bool range_sync_wait; 218static pthread_cond_t range_sync_cond; 219 220static void *ggate_recv_thread(void *arg); 221static void *local_send_thread(void *arg); 222static void *remote_send_thread(void *arg); 223static void *remote_recv_thread(void *arg); 224static void *ggate_send_thread(void *arg); 225static void *sync_thread(void *arg); 226static void *guard_thread(void *arg); 227 228static void 229dummy_sighandler(int sig __unused) 230{ 231 /* Nothing to do. */ 232} 233 234static void 235cleanup(struct hast_resource *res) 236{ 237 int rerrno; 238 239 /* Remember errno. */ 240 rerrno = errno; 241 242 /* 243 * Close descriptor to /dev/hast/<name> 244 * to work-around race in the kernel. 245 */ 246 close(res->hr_localfd); 247 248 /* Destroy ggate provider if we created one. */ 249 if (res->hr_ggateunit >= 0) { 250 struct g_gate_ctl_destroy ggiod; 251 252 ggiod.gctl_version = G_GATE_VERSION; 253 ggiod.gctl_unit = res->hr_ggateunit; 254 ggiod.gctl_force = 1; 255 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) { 256 pjdlog_warning("Unable to destroy hast/%s device", 257 res->hr_provname); 258 } 259 res->hr_ggateunit = -1; 260 } 261 262 /* Restore errno. */ 263 errno = rerrno; 264} 265 266static void 267primary_exit(int exitcode, const char *fmt, ...) 268{ 269 va_list ap; 270 271 assert(exitcode != EX_OK); 272 va_start(ap, fmt); 273 pjdlogv_errno(LOG_ERR, fmt, ap); 274 va_end(ap); 275 cleanup(gres); 276 exit(exitcode); 277} 278 279static void 280primary_exitx(int exitcode, const char *fmt, ...) 281{ 282 va_list ap; 283 284 va_start(ap, fmt); 285 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 286 va_end(ap); 287 cleanup(gres); 288 exit(exitcode); 289} 290 291static int 292hast_activemap_flush(struct hast_resource *res) 293{ 294 const unsigned char *buf; 295 size_t size; 296 297 buf = activemap_bitmap(res->hr_amp, &size); 298 assert(buf != NULL); 299 assert((size % res->hr_local_sectorsize) == 0); 300 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 301 (ssize_t)size) { 302 KEEP_ERRNO(pjdlog_errno(LOG_ERR, 303 "Unable to flush activemap to disk")); 304 return (-1); 305 } 306 return (0); 307} 308 309static bool 310real_remote(const struct hast_resource *res) 311{ 312 313 return (strcmp(res->hr_remoteaddr, "none") != 0); 314} 315 316static void 317init_environment(struct hast_resource *res __unused) 318{ 319 struct hio *hio; 320 unsigned int ii, ncomps; 321 sigset_t mask; 322 323 /* 324 * In the future it might be per-resource value. 325 */ 326 ncomps = HAST_NCOMPONENTS; 327 328 /* 329 * Allocate memory needed by lists. 330 */ 331 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 332 if (hio_send_list == NULL) { 333 primary_exitx(EX_TEMPFAIL, 334 "Unable to allocate %zu bytes of memory for send lists.", 335 sizeof(hio_send_list[0]) * ncomps); 336 } 337 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 338 if (hio_send_list_lock == NULL) { 339 primary_exitx(EX_TEMPFAIL, 340 "Unable to allocate %zu bytes of memory for send list locks.", 341 sizeof(hio_send_list_lock[0]) * ncomps); 342 } 343 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 344 if (hio_send_list_cond == NULL) { 345 primary_exitx(EX_TEMPFAIL, 346 "Unable to allocate %zu bytes of memory for send list condition variables.", 347 sizeof(hio_send_list_cond[0]) * ncomps); 348 } 349 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 350 if (hio_recv_list == NULL) { 351 primary_exitx(EX_TEMPFAIL, 352 "Unable to allocate %zu bytes of memory for recv lists.", 353 sizeof(hio_recv_list[0]) * ncomps); 354 } 355 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 356 if (hio_recv_list_lock == NULL) { 357 primary_exitx(EX_TEMPFAIL, 358 "Unable to allocate %zu bytes of memory for recv list locks.", 359 sizeof(hio_recv_list_lock[0]) * ncomps); 360 } 361 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 362 if (hio_recv_list_cond == NULL) { 363 primary_exitx(EX_TEMPFAIL, 364 "Unable to allocate %zu bytes of memory for recv list condition variables.", 365 sizeof(hio_recv_list_cond[0]) * ncomps); 366 } 367 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 368 if (hio_remote_lock == NULL) { 369 primary_exitx(EX_TEMPFAIL, 370 "Unable to allocate %zu bytes of memory for remote connections locks.", 371 sizeof(hio_remote_lock[0]) * ncomps); 372 } 373 374 /* 375 * Initialize lists, their locks and theirs condition variables. 376 */ 377 TAILQ_INIT(&hio_free_list); 378 mtx_init(&hio_free_list_lock); 379 cv_init(&hio_free_list_cond); 380 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 381 TAILQ_INIT(&hio_send_list[ii]); 382 mtx_init(&hio_send_list_lock[ii]); 383 cv_init(&hio_send_list_cond[ii]); 384 TAILQ_INIT(&hio_recv_list[ii]); 385 mtx_init(&hio_recv_list_lock[ii]); 386 cv_init(&hio_recv_list_cond[ii]); 387 rw_init(&hio_remote_lock[ii]); 388 } 389 TAILQ_INIT(&hio_done_list); 390 mtx_init(&hio_done_list_lock); 391 cv_init(&hio_done_list_cond); 392 mtx_init(&metadata_lock); 393 394 /* 395 * Allocate requests pool and initialize requests. 396 */ 397 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 398 hio = malloc(sizeof(*hio)); 399 if (hio == NULL) { 400 primary_exitx(EX_TEMPFAIL, 401 "Unable to allocate %zu bytes of memory for hio request.", 402 sizeof(*hio)); 403 } 404 hio->hio_countdown = 0; 405 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 406 if (hio->hio_errors == NULL) { 407 primary_exitx(EX_TEMPFAIL, 408 "Unable allocate %zu bytes of memory for hio errors.", 409 sizeof(hio->hio_errors[0]) * ncomps); 410 } 411 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 412 if (hio->hio_next == NULL) { 413 primary_exitx(EX_TEMPFAIL, 414 "Unable allocate %zu bytes of memory for hio_next field.", 415 sizeof(hio->hio_next[0]) * ncomps); 416 } 417 hio->hio_ggio.gctl_version = G_GATE_VERSION; 418 hio->hio_ggio.gctl_data = malloc(MAXPHYS); 419 if (hio->hio_ggio.gctl_data == NULL) { 420 primary_exitx(EX_TEMPFAIL, 421 "Unable to allocate %zu bytes of memory for gctl_data.", 422 MAXPHYS); 423 } 424 hio->hio_ggio.gctl_length = MAXPHYS; 425 hio->hio_ggio.gctl_error = 0; 426 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 427 } 428 429 /* 430 * Turn on signals handling. 431 */ 432 /* Because SIGCHLD is ignored by default, setup dummy handler for it. */ 433 PJDLOG_VERIFY(signal(SIGCHLD, dummy_sighandler) != SIG_ERR); 434 PJDLOG_VERIFY(sigfillset(&mask) == 0); 435 PJDLOG_VERIFY(sigprocmask(SIG_SETMASK, &mask, NULL) == 0); 436} 437 438static void 439init_local(struct hast_resource *res) 440{ 441 unsigned char *buf; 442 size_t mapsize; 443 444 if (metadata_read(res, true) < 0) 445 exit(EX_NOINPUT); 446 mtx_init(&res->hr_amp_lock); 447 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 448 res->hr_local_sectorsize, res->hr_keepdirty) < 0) { 449 primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 450 } 451 mtx_init(&range_lock); 452 cv_init(&range_regular_cond); 453 if (rangelock_init(&range_regular) < 0) 454 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 455 cv_init(&range_sync_cond); 456 if (rangelock_init(&range_sync) < 0) 457 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 458 mapsize = activemap_ondisk_size(res->hr_amp); 459 buf = calloc(1, mapsize); 460 if (buf == NULL) { 461 primary_exitx(EX_TEMPFAIL, 462 "Unable to allocate buffer for activemap."); 463 } 464 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 465 (ssize_t)mapsize) { 466 primary_exit(EX_NOINPUT, "Unable to read activemap"); 467 } 468 activemap_copyin(res->hr_amp, buf, mapsize); 469 free(buf); 470 if (res->hr_resuid != 0) 471 return; 472 /* 473 * We're using provider for the first time, so we have to generate 474 * resource unique identifier and initialize local and remote counts. 475 */ 476 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 477 res->hr_primary_localcnt = 1; 478 res->hr_primary_remotecnt = 0; 479 if (metadata_write(res) < 0) 480 exit(EX_NOINPUT); 481} 482 483static bool 484init_remote(struct hast_resource *res, struct proto_conn **inp, 485 struct proto_conn **outp) 486{ 487 struct proto_conn *in, *out; 488 struct nv *nvout, *nvin; 489 const unsigned char *token; 490 unsigned char *map; 491 const char *errmsg; 492 int32_t extentsize; 493 int64_t datasize; 494 uint32_t mapsize; 495 size_t size; 496 497 assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 498 assert(real_remote(res)); 499 500 in = out = NULL; 501 errmsg = NULL; 502 503 /* Prepare outgoing connection with remote node. */ 504 if (proto_client(res->hr_remoteaddr, &out) < 0) { 505 primary_exit(EX_TEMPFAIL, "Unable to create connection to %s", 506 res->hr_remoteaddr); 507 } 508 /* Try to connect, but accept failure. */ 509 if (proto_connect(out) < 0) { 510 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 511 res->hr_remoteaddr); 512 goto close; 513 } 514 /* Error in setting timeout is not critical, but why should it fail? */ 515 if (proto_timeout(out, res->hr_timeout) < 0) 516 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 517 /* 518 * First handshake step. 519 * Setup outgoing connection with remote node. 520 */ 521 nvout = nv_alloc(); 522 nv_add_string(nvout, res->hr_name, "resource"); 523 if (nv_error(nvout) != 0) { 524 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 525 "Unable to allocate header for connection with %s", 526 res->hr_remoteaddr); 527 nv_free(nvout); 528 goto close; 529 } 530 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) { 531 pjdlog_errno(LOG_WARNING, 532 "Unable to send handshake header to %s", 533 res->hr_remoteaddr); 534 nv_free(nvout); 535 goto close; 536 } 537 nv_free(nvout); 538 if (hast_proto_recv_hdr(out, &nvin) < 0) { 539 pjdlog_errno(LOG_WARNING, 540 "Unable to receive handshake header from %s", 541 res->hr_remoteaddr); 542 goto close; 543 } 544 errmsg = nv_get_string(nvin, "errmsg"); 545 if (errmsg != NULL) { 546 pjdlog_warning("%s", errmsg); 547 nv_free(nvin); 548 goto close; 549 } 550 token = nv_get_uint8_array(nvin, &size, "token"); 551 if (token == NULL) { 552 pjdlog_warning("Handshake header from %s has no 'token' field.", 553 res->hr_remoteaddr); 554 nv_free(nvin); 555 goto close; 556 } 557 if (size != sizeof(res->hr_token)) { 558 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 559 res->hr_remoteaddr, size, sizeof(res->hr_token)); 560 nv_free(nvin); 561 goto close; 562 } 563 bcopy(token, res->hr_token, sizeof(res->hr_token)); 564 nv_free(nvin); 565 566 /* 567 * Second handshake step. 568 * Setup incoming connection with remote node. 569 */ 570 if (proto_client(res->hr_remoteaddr, &in) < 0) { 571 pjdlog_errno(LOG_WARNING, "Unable to create connection to %s", 572 res->hr_remoteaddr); 573 } 574 /* Try to connect, but accept failure. */ 575 if (proto_connect(in) < 0) { 576 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 577 res->hr_remoteaddr); 578 goto close; 579 } 580 /* Error in setting timeout is not critical, but why should it fail? */ 581 if (proto_timeout(in, res->hr_timeout) < 0) 582 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 583 nvout = nv_alloc(); 584 nv_add_string(nvout, res->hr_name, "resource"); 585 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 586 "token"); 587 nv_add_uint64(nvout, res->hr_resuid, "resuid"); 588 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 589 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 590 if (nv_error(nvout) != 0) { 591 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 592 "Unable to allocate header for connection with %s", 593 res->hr_remoteaddr); 594 nv_free(nvout); 595 goto close; 596 } 597 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) { 598 pjdlog_errno(LOG_WARNING, 599 "Unable to send handshake header to %s", 600 res->hr_remoteaddr); 601 nv_free(nvout); 602 goto close; 603 } 604 nv_free(nvout); 605 if (hast_proto_recv_hdr(out, &nvin) < 0) { 606 pjdlog_errno(LOG_WARNING, 607 "Unable to receive handshake header from %s", 608 res->hr_remoteaddr); 609 goto close; 610 } 611 errmsg = nv_get_string(nvin, "errmsg"); 612 if (errmsg != NULL) { 613 pjdlog_warning("%s", errmsg); 614 nv_free(nvin); 615 goto close; 616 } 617 datasize = nv_get_int64(nvin, "datasize"); 618 if (datasize != res->hr_datasize) { 619 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 620 (intmax_t)res->hr_datasize, (intmax_t)datasize); 621 nv_free(nvin); 622 goto close; 623 } 624 extentsize = nv_get_int32(nvin, "extentsize"); 625 if (extentsize != res->hr_extentsize) { 626 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 627 (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 628 nv_free(nvin); 629 goto close; 630 } 631 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 632 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 633 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 634 map = NULL; 635 mapsize = nv_get_uint32(nvin, "mapsize"); 636 if (mapsize > 0) { 637 map = malloc(mapsize); 638 if (map == NULL) { 639 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 640 (uintmax_t)mapsize); 641 nv_free(nvin); 642 goto close; 643 } 644 /* 645 * Remote node have some dirty extents on its own, lets 646 * download its activemap. 647 */ 648 if (hast_proto_recv_data(res, out, nvin, map, 649 mapsize) < 0) { 650 pjdlog_errno(LOG_ERR, 651 "Unable to receive remote activemap"); 652 nv_free(nvin); 653 free(map); 654 goto close; 655 } 656 /* 657 * Merge local and remote bitmaps. 658 */ 659 activemap_merge(res->hr_amp, map, mapsize); 660 free(map); 661 /* 662 * Now that we merged bitmaps from both nodes, flush it to the 663 * disk before we start to synchronize. 664 */ 665 (void)hast_activemap_flush(res); 666 } 667 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 668 if (inp != NULL && outp != NULL) { 669 *inp = in; 670 *outp = out; 671 } else { 672 res->hr_remotein = in; 673 res->hr_remoteout = out; 674 } 675 return (true); 676close: 677 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 678 hook_exec(res->hr_exec, "split-brain", res->hr_name, NULL); 679 proto_close(out); 680 if (in != NULL) 681 proto_close(in); 682 return (false); 683} 684 685static void 686sync_start(void) 687{ 688 689 mtx_lock(&sync_lock); 690 sync_inprogress = true; 691 mtx_unlock(&sync_lock); 692 cv_signal(&sync_cond); 693} 694 695static void 696sync_stop(void) 697{ 698 699 mtx_lock(&sync_lock); 700 if (sync_inprogress) 701 sync_inprogress = false; 702 mtx_unlock(&sync_lock); 703} 704 705static void 706init_ggate(struct hast_resource *res) 707{ 708 struct g_gate_ctl_create ggiocreate; 709 struct g_gate_ctl_cancel ggiocancel; 710 711 /* 712 * We communicate with ggate via /dev/ggctl. Open it. 713 */ 714 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 715 if (res->hr_ggatefd < 0) 716 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 717 /* 718 * Create provider before trying to connect, as connection failure 719 * is not critical, but may take some time. 720 */ 721 ggiocreate.gctl_version = G_GATE_VERSION; 722 ggiocreate.gctl_mediasize = res->hr_datasize; 723 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 724 ggiocreate.gctl_flags = 0; 725 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE; 726 ggiocreate.gctl_timeout = 0; 727 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 728 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 729 res->hr_provname); 730 bzero(ggiocreate.gctl_info, sizeof(ggiocreate.gctl_info)); 731 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 732 pjdlog_info("Device hast/%s created.", res->hr_provname); 733 res->hr_ggateunit = ggiocreate.gctl_unit; 734 return; 735 } 736 if (errno != EEXIST) { 737 primary_exit(EX_OSERR, "Unable to create hast/%s device", 738 res->hr_provname); 739 } 740 pjdlog_debug(1, 741 "Device hast/%s already exists, we will try to take it over.", 742 res->hr_provname); 743 /* 744 * If we received EEXIST, we assume that the process who created the 745 * provider died and didn't clean up. In that case we will start from 746 * where he left of. 747 */ 748 ggiocancel.gctl_version = G_GATE_VERSION; 749 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 750 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 751 res->hr_provname); 752 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 753 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 754 res->hr_ggateunit = ggiocancel.gctl_unit; 755 return; 756 } 757 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 758 res->hr_provname); 759} 760 761void 762hastd_primary(struct hast_resource *res) 763{ 764 pthread_t td; 765 pid_t pid; 766 int error; 767 768 gres = res; 769 770 /* 771 * Create communication channel between parent and child. 772 */ 773 if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 774 KEEP_ERRNO((void)pidfile_remove(pfh)); 775 primary_exit(EX_OSERR, 776 "Unable to create control sockets between parent and child"); 777 } 778 779 pid = fork(); 780 if (pid < 0) { 781 KEEP_ERRNO((void)pidfile_remove(pfh)); 782 primary_exit(EX_TEMPFAIL, "Unable to fork"); 783 } 784 785 if (pid > 0) { 786 /* This is parent. */ 787 res->hr_workerpid = pid; 788 return; 789 } 790 791 (void)pidfile_close(pfh); 792 hook_fini(); 793 794 setproctitle("%s (primary)", res->hr_name); 795 796 signal(SIGHUP, SIG_DFL); 797 signal(SIGCHLD, SIG_DFL); 798 799 hook_init(); 800 init_local(res); 801 if (real_remote(res) && init_remote(res, NULL, NULL)) 802 sync_start(); 803 init_ggate(res); 804 init_environment(res); 805 error = pthread_create(&td, NULL, ggate_recv_thread, res); 806 assert(error == 0); 807 error = pthread_create(&td, NULL, local_send_thread, res); 808 assert(error == 0); 809 error = pthread_create(&td, NULL, remote_send_thread, res); 810 assert(error == 0); 811 error = pthread_create(&td, NULL, remote_recv_thread, res); 812 assert(error == 0); 813 error = pthread_create(&td, NULL, ggate_send_thread, res); 814 assert(error == 0); 815 error = pthread_create(&td, NULL, sync_thread, res); 816 assert(error == 0); 817 error = pthread_create(&td, NULL, ctrl_thread, res); 818 assert(error == 0); 819 (void)guard_thread(res); 820} 821 822static void 823reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 824{ 825 char msg[1024]; 826 va_list ap; 827 int len; 828 829 va_start(ap, fmt); 830 len = vsnprintf(msg, sizeof(msg), fmt, ap); 831 va_end(ap); 832 if ((size_t)len < sizeof(msg)) { 833 switch (ggio->gctl_cmd) { 834 case BIO_READ: 835 (void)snprintf(msg + len, sizeof(msg) - len, 836 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 837 (uintmax_t)ggio->gctl_length); 838 break; 839 case BIO_DELETE: 840 (void)snprintf(msg + len, sizeof(msg) - len, 841 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 842 (uintmax_t)ggio->gctl_length); 843 break; 844 case BIO_FLUSH: 845 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 846 break; 847 case BIO_WRITE: 848 (void)snprintf(msg + len, sizeof(msg) - len, 849 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 850 (uintmax_t)ggio->gctl_length); 851 break; 852 default: 853 (void)snprintf(msg + len, sizeof(msg) - len, 854 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 855 break; 856 } 857 } 858 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 859} 860 861static void 862remote_close(struct hast_resource *res, int ncomp) 863{ 864 865 rw_wlock(&hio_remote_lock[ncomp]); 866 /* 867 * A race is possible between dropping rlock and acquiring wlock - 868 * another thread can close connection in-between. 869 */ 870 if (!ISCONNECTED(res, ncomp)) { 871 assert(res->hr_remotein == NULL); 872 assert(res->hr_remoteout == NULL); 873 rw_unlock(&hio_remote_lock[ncomp]); 874 return; 875 } 876 877 assert(res->hr_remotein != NULL); 878 assert(res->hr_remoteout != NULL); 879 880 pjdlog_debug(2, "Closing incoming connection to %s.", 881 res->hr_remoteaddr); 882 proto_close(res->hr_remotein); 883 res->hr_remotein = NULL; 884 pjdlog_debug(2, "Closing outgoing connection to %s.", 885 res->hr_remoteaddr); 886 proto_close(res->hr_remoteout); 887 res->hr_remoteout = NULL; 888 889 rw_unlock(&hio_remote_lock[ncomp]); 890 891 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 892 893 /* 894 * Stop synchronization if in-progress. 895 */ 896 sync_stop(); 897} 898 899/* 900 * Thread receives ggate I/O requests from the kernel and passes them to 901 * appropriate threads: 902 * WRITE - always goes to both local_send and remote_send threads 903 * READ (when the block is up-to-date on local component) - 904 * only local_send thread 905 * READ (when the block isn't up-to-date on local component) - 906 * only remote_send thread 907 * DELETE - always goes to both local_send and remote_send threads 908 * FLUSH - always goes to both local_send and remote_send threads 909 */ 910static void * 911ggate_recv_thread(void *arg) 912{ 913 struct hast_resource *res = arg; 914 struct g_gate_ctl_io *ggio; 915 struct hio *hio; 916 unsigned int ii, ncomp, ncomps; 917 int error; 918 919 ncomps = HAST_NCOMPONENTS; 920 921 for (;;) { 922 pjdlog_debug(2, "ggate_recv: Taking free request."); 923 QUEUE_TAKE2(hio, free); 924 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 925 ggio = &hio->hio_ggio; 926 ggio->gctl_unit = res->hr_ggateunit; 927 ggio->gctl_length = MAXPHYS; 928 ggio->gctl_error = 0; 929 pjdlog_debug(2, 930 "ggate_recv: (%p) Waiting for request from the kernel.", 931 hio); 932 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 933 if (sigexit_received) 934 pthread_exit(NULL); 935 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 936 } 937 error = ggio->gctl_error; 938 switch (error) { 939 case 0: 940 break; 941 case ECANCELED: 942 /* Exit gracefully. */ 943 if (!sigexit_received) { 944 pjdlog_debug(2, 945 "ggate_recv: (%p) Received cancel from the kernel.", 946 hio); 947 pjdlog_info("Received cancel from the kernel, exiting."); 948 } 949 pthread_exit(NULL); 950 case ENOMEM: 951 /* 952 * Buffer too small? Impossible, we allocate MAXPHYS 953 * bytes - request can't be bigger than that. 954 */ 955 /* FALLTHROUGH */ 956 case ENXIO: 957 default: 958 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 959 strerror(error)); 960 } 961 for (ii = 0; ii < ncomps; ii++) 962 hio->hio_errors[ii] = EINVAL; 963 reqlog(LOG_DEBUG, 2, ggio, 964 "ggate_recv: (%p) Request received from the kernel: ", 965 hio); 966 /* 967 * Inform all components about new write request. 968 * For read request prefer local component unless the given 969 * range is out-of-date, then use remote component. 970 */ 971 switch (ggio->gctl_cmd) { 972 case BIO_READ: 973 pjdlog_debug(2, 974 "ggate_recv: (%p) Moving request to the send queue.", 975 hio); 976 refcount_init(&hio->hio_countdown, 1); 977 mtx_lock(&metadata_lock); 978 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 979 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 980 /* 981 * This range is up-to-date on local component, 982 * so handle request locally. 983 */ 984 /* Local component is 0 for now. */ 985 ncomp = 0; 986 } else /* if (res->hr_syncsrc == 987 HAST_SYNCSRC_SECONDARY) */ { 988 assert(res->hr_syncsrc == 989 HAST_SYNCSRC_SECONDARY); 990 /* 991 * This range is out-of-date on local component, 992 * so send request to the remote node. 993 */ 994 /* Remote component is 1 for now. */ 995 ncomp = 1; 996 } 997 mtx_unlock(&metadata_lock); 998 QUEUE_INSERT1(hio, send, ncomp); 999 break; 1000 case BIO_WRITE: 1001 for (;;) { 1002 mtx_lock(&range_lock); 1003 if (rangelock_islocked(range_sync, 1004 ggio->gctl_offset, ggio->gctl_length)) { 1005 pjdlog_debug(2, 1006 "regular: Range offset=%jd length=%zu locked.", 1007 (intmax_t)ggio->gctl_offset, 1008 (size_t)ggio->gctl_length); 1009 range_regular_wait = true; 1010 cv_wait(&range_regular_cond, &range_lock); 1011 range_regular_wait = false; 1012 mtx_unlock(&range_lock); 1013 continue; 1014 } 1015 if (rangelock_add(range_regular, 1016 ggio->gctl_offset, ggio->gctl_length) < 0) { 1017 mtx_unlock(&range_lock); 1018 pjdlog_debug(2, 1019 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1020 (intmax_t)ggio->gctl_offset, 1021 (size_t)ggio->gctl_length); 1022 sleep(1); 1023 continue; 1024 } 1025 mtx_unlock(&range_lock); 1026 break; 1027 } 1028 mtx_lock(&res->hr_amp_lock); 1029 if (activemap_write_start(res->hr_amp, 1030 ggio->gctl_offset, ggio->gctl_length)) { 1031 (void)hast_activemap_flush(res); 1032 } 1033 mtx_unlock(&res->hr_amp_lock); 1034 /* FALLTHROUGH */ 1035 case BIO_DELETE: 1036 case BIO_FLUSH: 1037 pjdlog_debug(2, 1038 "ggate_recv: (%p) Moving request to the send queues.", 1039 hio); 1040 refcount_init(&hio->hio_countdown, ncomps); 1041 for (ii = 0; ii < ncomps; ii++) 1042 QUEUE_INSERT1(hio, send, ii); 1043 break; 1044 } 1045 } 1046 /* NOTREACHED */ 1047 return (NULL); 1048} 1049 1050/* 1051 * Thread reads from or writes to local component. 1052 * If local read fails, it redirects it to remote_send thread. 1053 */ 1054static void * 1055local_send_thread(void *arg) 1056{ 1057 struct hast_resource *res = arg; 1058 struct g_gate_ctl_io *ggio; 1059 struct hio *hio; 1060 unsigned int ncomp, rncomp; 1061 ssize_t ret; 1062 1063 /* Local component is 0 for now. */ 1064 ncomp = 0; 1065 /* Remote component is 1 for now. */ 1066 rncomp = 1; 1067 1068 for (;;) { 1069 pjdlog_debug(2, "local_send: Taking request."); 1070 QUEUE_TAKE1(hio, send, ncomp); 1071 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1072 ggio = &hio->hio_ggio; 1073 switch (ggio->gctl_cmd) { 1074 case BIO_READ: 1075 ret = pread(res->hr_localfd, ggio->gctl_data, 1076 ggio->gctl_length, 1077 ggio->gctl_offset + res->hr_localoff); 1078 if (ret == ggio->gctl_length) 1079 hio->hio_errors[ncomp] = 0; 1080 else { 1081 /* 1082 * If READ failed, try to read from remote node. 1083 */ 1084 QUEUE_INSERT1(hio, send, rncomp); 1085 continue; 1086 } 1087 break; 1088 case BIO_WRITE: 1089 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1090 ggio->gctl_length, 1091 ggio->gctl_offset + res->hr_localoff); 1092 if (ret < 0) 1093 hio->hio_errors[ncomp] = errno; 1094 else if (ret != ggio->gctl_length) 1095 hio->hio_errors[ncomp] = EIO; 1096 else 1097 hio->hio_errors[ncomp] = 0; 1098 break; 1099 case BIO_DELETE: 1100 ret = g_delete(res->hr_localfd, 1101 ggio->gctl_offset + res->hr_localoff, 1102 ggio->gctl_length); 1103 if (ret < 0) 1104 hio->hio_errors[ncomp] = errno; 1105 else 1106 hio->hio_errors[ncomp] = 0; 1107 break; 1108 case BIO_FLUSH: 1109 ret = g_flush(res->hr_localfd); 1110 if (ret < 0) 1111 hio->hio_errors[ncomp] = errno; 1112 else 1113 hio->hio_errors[ncomp] = 0; 1114 break; 1115 } 1116 if (refcount_release(&hio->hio_countdown)) { 1117 if (ISSYNCREQ(hio)) { 1118 mtx_lock(&sync_lock); 1119 SYNCREQDONE(hio); 1120 mtx_unlock(&sync_lock); 1121 cv_signal(&sync_cond); 1122 } else { 1123 pjdlog_debug(2, 1124 "local_send: (%p) Moving request to the done queue.", 1125 hio); 1126 QUEUE_INSERT2(hio, done); 1127 } 1128 } 1129 } 1130 /* NOTREACHED */ 1131 return (NULL); 1132} 1133 1134/* 1135 * Thread sends request to secondary node. 1136 */ 1137static void * 1138remote_send_thread(void *arg) 1139{ 1140 struct hast_resource *res = arg; 1141 struct g_gate_ctl_io *ggio; 1142 struct hio *hio; 1143 struct nv *nv; 1144 unsigned int ncomp; 1145 bool wakeup; 1146 uint64_t offset, length; 1147 uint8_t cmd; 1148 void *data; 1149 1150 /* Remote component is 1 for now. */ 1151 ncomp = 1; 1152 1153 for (;;) { 1154 pjdlog_debug(2, "remote_send: Taking request."); 1155 QUEUE_TAKE1(hio, send, ncomp); 1156 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1157 ggio = &hio->hio_ggio; 1158 switch (ggio->gctl_cmd) { 1159 case BIO_READ: 1160 cmd = HIO_READ; 1161 data = NULL; 1162 offset = ggio->gctl_offset; 1163 length = ggio->gctl_length; 1164 break; 1165 case BIO_WRITE: 1166 cmd = HIO_WRITE; 1167 data = ggio->gctl_data; 1168 offset = ggio->gctl_offset; 1169 length = ggio->gctl_length; 1170 break; 1171 case BIO_DELETE: 1172 cmd = HIO_DELETE; 1173 data = NULL; 1174 offset = ggio->gctl_offset; 1175 length = ggio->gctl_length; 1176 break; 1177 case BIO_FLUSH: 1178 cmd = HIO_FLUSH; 1179 data = NULL; 1180 offset = 0; 1181 length = 0; 1182 break; 1183 default: 1184 assert(!"invalid condition"); 1185 abort(); 1186 } 1187 nv = nv_alloc(); 1188 nv_add_uint8(nv, cmd, "cmd"); 1189 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1190 nv_add_uint64(nv, offset, "offset"); 1191 nv_add_uint64(nv, length, "length"); 1192 if (nv_error(nv) != 0) { 1193 hio->hio_errors[ncomp] = nv_error(nv); 1194 pjdlog_debug(2, 1195 "remote_send: (%p) Unable to prepare header to send.", 1196 hio); 1197 reqlog(LOG_ERR, 0, ggio, 1198 "Unable to prepare header to send (%s): ", 1199 strerror(nv_error(nv))); 1200 /* Move failed request immediately to the done queue. */ 1201 goto done_queue; 1202 } 1203 pjdlog_debug(2, 1204 "remote_send: (%p) Moving request to the recv queue.", 1205 hio); 1206 /* 1207 * Protect connection from disappearing. 1208 */ 1209 rw_rlock(&hio_remote_lock[ncomp]); 1210 if (!ISCONNECTED(res, ncomp)) { 1211 rw_unlock(&hio_remote_lock[ncomp]); 1212 hio->hio_errors[ncomp] = ENOTCONN; 1213 goto done_queue; 1214 } 1215 /* 1216 * Move the request to recv queue before sending it, because 1217 * in different order we can get reply before we move request 1218 * to recv queue. 1219 */ 1220 mtx_lock(&hio_recv_list_lock[ncomp]); 1221 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1222 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1223 mtx_unlock(&hio_recv_list_lock[ncomp]); 1224 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1225 data != NULL ? length : 0) < 0) { 1226 hio->hio_errors[ncomp] = errno; 1227 rw_unlock(&hio_remote_lock[ncomp]); 1228 pjdlog_debug(2, 1229 "remote_send: (%p) Unable to send request.", hio); 1230 reqlog(LOG_ERR, 0, ggio, 1231 "Unable to send request (%s): ", 1232 strerror(hio->hio_errors[ncomp])); 1233 remote_close(res, ncomp); 1234 /* 1235 * Take request back from the receive queue and move 1236 * it immediately to the done queue. 1237 */ 1238 mtx_lock(&hio_recv_list_lock[ncomp]); 1239 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1240 mtx_unlock(&hio_recv_list_lock[ncomp]); 1241 goto done_queue; 1242 } 1243 rw_unlock(&hio_remote_lock[ncomp]); 1244 nv_free(nv); 1245 if (wakeup) 1246 cv_signal(&hio_recv_list_cond[ncomp]); 1247 continue; 1248done_queue: 1249 nv_free(nv); 1250 if (ISSYNCREQ(hio)) { 1251 if (!refcount_release(&hio->hio_countdown)) 1252 continue; 1253 mtx_lock(&sync_lock); 1254 SYNCREQDONE(hio); 1255 mtx_unlock(&sync_lock); 1256 cv_signal(&sync_cond); 1257 continue; 1258 } 1259 if (ggio->gctl_cmd == BIO_WRITE) { 1260 mtx_lock(&res->hr_amp_lock); 1261 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1262 ggio->gctl_length)) { 1263 (void)hast_activemap_flush(res); 1264 } 1265 mtx_unlock(&res->hr_amp_lock); 1266 } 1267 if (!refcount_release(&hio->hio_countdown)) 1268 continue; 1269 pjdlog_debug(2, 1270 "remote_send: (%p) Moving request to the done queue.", 1271 hio); 1272 QUEUE_INSERT2(hio, done); 1273 } 1274 /* NOTREACHED */ 1275 return (NULL); 1276} 1277 1278/* 1279 * Thread receives answer from secondary node and passes it to ggate_send 1280 * thread. 1281 */ 1282static void * 1283remote_recv_thread(void *arg) 1284{ 1285 struct hast_resource *res = arg; 1286 struct g_gate_ctl_io *ggio; 1287 struct hio *hio; 1288 struct nv *nv; 1289 unsigned int ncomp; 1290 uint64_t seq; 1291 int error; 1292 1293 /* Remote component is 1 for now. */ 1294 ncomp = 1; 1295 1296 for (;;) { 1297 /* Wait until there is anything to receive. */ 1298 mtx_lock(&hio_recv_list_lock[ncomp]); 1299 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1300 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1301 cv_wait(&hio_recv_list_cond[ncomp], 1302 &hio_recv_list_lock[ncomp]); 1303 } 1304 mtx_unlock(&hio_recv_list_lock[ncomp]); 1305 rw_rlock(&hio_remote_lock[ncomp]); 1306 if (!ISCONNECTED(res, ncomp)) { 1307 rw_unlock(&hio_remote_lock[ncomp]); 1308 /* 1309 * Connection is dead, so move all pending requests to 1310 * the done queue (one-by-one). 1311 */ 1312 mtx_lock(&hio_recv_list_lock[ncomp]); 1313 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1314 assert(hio != NULL); 1315 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1316 hio_next[ncomp]); 1317 mtx_unlock(&hio_recv_list_lock[ncomp]); 1318 goto done_queue; 1319 } 1320 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1321 pjdlog_errno(LOG_ERR, 1322 "Unable to receive reply header"); 1323 rw_unlock(&hio_remote_lock[ncomp]); 1324 remote_close(res, ncomp); 1325 continue; 1326 } 1327 rw_unlock(&hio_remote_lock[ncomp]); 1328 seq = nv_get_uint64(nv, "seq"); 1329 if (seq == 0) { 1330 pjdlog_error("Header contains no 'seq' field."); 1331 nv_free(nv); 1332 continue; 1333 } 1334 mtx_lock(&hio_recv_list_lock[ncomp]); 1335 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1336 if (hio->hio_ggio.gctl_seq == seq) { 1337 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1338 hio_next[ncomp]); 1339 break; 1340 } 1341 } 1342 mtx_unlock(&hio_recv_list_lock[ncomp]); 1343 if (hio == NULL) { 1344 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1345 (uintmax_t)seq); 1346 nv_free(nv); 1347 continue; 1348 } 1349 error = nv_get_int16(nv, "error"); 1350 if (error != 0) { 1351 /* Request failed on remote side. */ 1352 hio->hio_errors[ncomp] = 0; 1353 nv_free(nv); 1354 goto done_queue; 1355 } 1356 ggio = &hio->hio_ggio; 1357 switch (ggio->gctl_cmd) { 1358 case BIO_READ: 1359 rw_rlock(&hio_remote_lock[ncomp]); 1360 if (!ISCONNECTED(res, ncomp)) { 1361 rw_unlock(&hio_remote_lock[ncomp]); 1362 nv_free(nv); 1363 goto done_queue; 1364 } 1365 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1366 ggio->gctl_data, ggio->gctl_length) < 0) { 1367 hio->hio_errors[ncomp] = errno; 1368 pjdlog_errno(LOG_ERR, 1369 "Unable to receive reply data"); 1370 rw_unlock(&hio_remote_lock[ncomp]); 1371 nv_free(nv); 1372 remote_close(res, ncomp); 1373 goto done_queue; 1374 } 1375 rw_unlock(&hio_remote_lock[ncomp]); 1376 break; 1377 case BIO_WRITE: 1378 case BIO_DELETE: 1379 case BIO_FLUSH: 1380 break; 1381 default: 1382 assert(!"invalid condition"); 1383 abort(); 1384 } 1385 hio->hio_errors[ncomp] = 0; 1386 nv_free(nv); 1387done_queue: 1388 if (refcount_release(&hio->hio_countdown)) { 1389 if (ISSYNCREQ(hio)) { 1390 mtx_lock(&sync_lock); 1391 SYNCREQDONE(hio); 1392 mtx_unlock(&sync_lock); 1393 cv_signal(&sync_cond); 1394 } else { 1395 pjdlog_debug(2, 1396 "remote_recv: (%p) Moving request to the done queue.", 1397 hio); 1398 QUEUE_INSERT2(hio, done); 1399 } 1400 } 1401 } 1402 /* NOTREACHED */ 1403 return (NULL); 1404} 1405 1406/* 1407 * Thread sends answer to the kernel. 1408 */ 1409static void * 1410ggate_send_thread(void *arg) 1411{ 1412 struct hast_resource *res = arg; 1413 struct g_gate_ctl_io *ggio; 1414 struct hio *hio; 1415 unsigned int ii, ncomp, ncomps; 1416 1417 ncomps = HAST_NCOMPONENTS; 1418 1419 for (;;) { 1420 pjdlog_debug(2, "ggate_send: Taking request."); 1421 QUEUE_TAKE2(hio, done); 1422 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1423 ggio = &hio->hio_ggio; 1424 for (ii = 0; ii < ncomps; ii++) { 1425 if (hio->hio_errors[ii] == 0) { 1426 /* 1427 * One successful request is enough to declare 1428 * success. 1429 */ 1430 ggio->gctl_error = 0; 1431 break; 1432 } 1433 } 1434 if (ii == ncomps) { 1435 /* 1436 * None of the requests were successful. 1437 * Use first error. 1438 */ 1439 ggio->gctl_error = hio->hio_errors[0]; 1440 } 1441 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1442 mtx_lock(&res->hr_amp_lock); 1443 activemap_write_complete(res->hr_amp, 1444 ggio->gctl_offset, ggio->gctl_length); 1445 mtx_unlock(&res->hr_amp_lock); 1446 } 1447 if (ggio->gctl_cmd == BIO_WRITE) { 1448 /* 1449 * Unlock range we locked. 1450 */ 1451 mtx_lock(&range_lock); 1452 rangelock_del(range_regular, ggio->gctl_offset, 1453 ggio->gctl_length); 1454 if (range_sync_wait) 1455 cv_signal(&range_sync_cond); 1456 mtx_unlock(&range_lock); 1457 /* 1458 * Bump local count if this is first write after 1459 * connection failure with remote node. 1460 */ 1461 ncomp = 1; 1462 rw_rlock(&hio_remote_lock[ncomp]); 1463 if (!ISCONNECTED(res, ncomp)) { 1464 mtx_lock(&metadata_lock); 1465 if (res->hr_primary_localcnt == 1466 res->hr_secondary_remotecnt) { 1467 res->hr_primary_localcnt++; 1468 pjdlog_debug(1, 1469 "Increasing localcnt to %ju.", 1470 (uintmax_t)res->hr_primary_localcnt); 1471 (void)metadata_write(res); 1472 } 1473 mtx_unlock(&metadata_lock); 1474 } 1475 rw_unlock(&hio_remote_lock[ncomp]); 1476 } 1477 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1478 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1479 pjdlog_debug(2, 1480 "ggate_send: (%p) Moving request to the free queue.", hio); 1481 QUEUE_INSERT2(hio, free); 1482 } 1483 /* NOTREACHED */ 1484 return (NULL); 1485} 1486 1487/* 1488 * Thread synchronize local and remote components. 1489 */ 1490static void * 1491sync_thread(void *arg __unused) 1492{ 1493 struct hast_resource *res = arg; 1494 struct hio *hio; 1495 struct g_gate_ctl_io *ggio; 1496 unsigned int ii, ncomp, ncomps; 1497 off_t offset, length, synced; 1498 bool dorewind; 1499 int syncext; 1500 1501 ncomps = HAST_NCOMPONENTS; 1502 dorewind = true; 1503 synced = 0; 1504 offset = -1; 1505 1506 for (;;) { 1507 mtx_lock(&sync_lock); 1508 if (offset >= 0 && !sync_inprogress) { 1509 pjdlog_info("Synchronization interrupted. " 1510 "%jd bytes synchronized so far.", 1511 (intmax_t)synced); 1512 hook_exec(res->hr_exec, "syncintr", res->hr_name, NULL); 1513 } 1514 while (!sync_inprogress) { 1515 dorewind = true; 1516 synced = 0; 1517 cv_wait(&sync_cond, &sync_lock); 1518 } 1519 mtx_unlock(&sync_lock); 1520 /* 1521 * Obtain offset at which we should synchronize. 1522 * Rewind synchronization if needed. 1523 */ 1524 mtx_lock(&res->hr_amp_lock); 1525 if (dorewind) 1526 activemap_sync_rewind(res->hr_amp); 1527 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1528 if (syncext != -1) { 1529 /* 1530 * We synchronized entire syncext extent, we can mark 1531 * it as clean now. 1532 */ 1533 if (activemap_extent_complete(res->hr_amp, syncext)) 1534 (void)hast_activemap_flush(res); 1535 } 1536 mtx_unlock(&res->hr_amp_lock); 1537 if (dorewind) { 1538 dorewind = false; 1539 if (offset < 0) 1540 pjdlog_info("Nodes are in sync."); 1541 else { 1542 pjdlog_info("Synchronization started. %ju bytes to go.", 1543 (uintmax_t)(res->hr_extentsize * 1544 activemap_ndirty(res->hr_amp))); 1545 hook_exec(res->hr_exec, "syncstart", 1546 res->hr_name, NULL); 1547 } 1548 } 1549 if (offset < 0) { 1550 sync_stop(); 1551 pjdlog_debug(1, "Nothing to synchronize."); 1552 /* 1553 * Synchronization complete, make both localcnt and 1554 * remotecnt equal. 1555 */ 1556 ncomp = 1; 1557 rw_rlock(&hio_remote_lock[ncomp]); 1558 if (ISCONNECTED(res, ncomp)) { 1559 if (synced > 0) { 1560 pjdlog_info("Synchronization complete. " 1561 "%jd bytes synchronized.", 1562 (intmax_t)synced); 1563 hook_exec(res->hr_exec, "syncdone", 1564 res->hr_name, NULL); 1565 } 1566 mtx_lock(&metadata_lock); 1567 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1568 res->hr_primary_localcnt = 1569 res->hr_secondary_localcnt; 1570 res->hr_primary_remotecnt = 1571 res->hr_secondary_remotecnt; 1572 pjdlog_debug(1, 1573 "Setting localcnt to %ju and remotecnt to %ju.", 1574 (uintmax_t)res->hr_primary_localcnt, 1575 (uintmax_t)res->hr_secondary_localcnt); 1576 (void)metadata_write(res); 1577 mtx_unlock(&metadata_lock); 1578 } 1579 rw_unlock(&hio_remote_lock[ncomp]); 1580 continue; 1581 } 1582 pjdlog_debug(2, "sync: Taking free request."); 1583 QUEUE_TAKE2(hio, free); 1584 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1585 /* 1586 * Lock the range we are going to synchronize. We don't want 1587 * race where someone writes between our read and write. 1588 */ 1589 for (;;) { 1590 mtx_lock(&range_lock); 1591 if (rangelock_islocked(range_regular, offset, length)) { 1592 pjdlog_debug(2, 1593 "sync: Range offset=%jd length=%jd locked.", 1594 (intmax_t)offset, (intmax_t)length); 1595 range_sync_wait = true; 1596 cv_wait(&range_sync_cond, &range_lock); 1597 range_sync_wait = false; 1598 mtx_unlock(&range_lock); 1599 continue; 1600 } 1601 if (rangelock_add(range_sync, offset, length) < 0) { 1602 mtx_unlock(&range_lock); 1603 pjdlog_debug(2, 1604 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1605 (intmax_t)offset, (intmax_t)length); 1606 sleep(1); 1607 continue; 1608 } 1609 mtx_unlock(&range_lock); 1610 break; 1611 } 1612 /* 1613 * First read the data from synchronization source. 1614 */ 1615 SYNCREQ(hio); 1616 ggio = &hio->hio_ggio; 1617 ggio->gctl_cmd = BIO_READ; 1618 ggio->gctl_offset = offset; 1619 ggio->gctl_length = length; 1620 ggio->gctl_error = 0; 1621 for (ii = 0; ii < ncomps; ii++) 1622 hio->hio_errors[ii] = EINVAL; 1623 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1624 hio); 1625 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1626 hio); 1627 mtx_lock(&metadata_lock); 1628 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1629 /* 1630 * This range is up-to-date on local component, 1631 * so handle request locally. 1632 */ 1633 /* Local component is 0 for now. */ 1634 ncomp = 0; 1635 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1636 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1637 /* 1638 * This range is out-of-date on local component, 1639 * so send request to the remote node. 1640 */ 1641 /* Remote component is 1 for now. */ 1642 ncomp = 1; 1643 } 1644 mtx_unlock(&metadata_lock); 1645 refcount_init(&hio->hio_countdown, 1); 1646 QUEUE_INSERT1(hio, send, ncomp); 1647 1648 /* 1649 * Let's wait for READ to finish. 1650 */ 1651 mtx_lock(&sync_lock); 1652 while (!ISSYNCREQDONE(hio)) 1653 cv_wait(&sync_cond, &sync_lock); 1654 mtx_unlock(&sync_lock); 1655 1656 if (hio->hio_errors[ncomp] != 0) { 1657 pjdlog_error("Unable to read synchronization data: %s.", 1658 strerror(hio->hio_errors[ncomp])); 1659 goto free_queue; 1660 } 1661 1662 /* 1663 * We read the data from synchronization source, now write it 1664 * to synchronization target. 1665 */ 1666 SYNCREQ(hio); 1667 ggio->gctl_cmd = BIO_WRITE; 1668 for (ii = 0; ii < ncomps; ii++) 1669 hio->hio_errors[ii] = EINVAL; 1670 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1671 hio); 1672 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1673 hio); 1674 mtx_lock(&metadata_lock); 1675 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1676 /* 1677 * This range is up-to-date on local component, 1678 * so we update remote component. 1679 */ 1680 /* Remote component is 1 for now. */ 1681 ncomp = 1; 1682 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1683 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1684 /* 1685 * This range is out-of-date on local component, 1686 * so we update it. 1687 */ 1688 /* Local component is 0 for now. */ 1689 ncomp = 0; 1690 } 1691 mtx_unlock(&metadata_lock); 1692 1693 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1694 hio); 1695 refcount_init(&hio->hio_countdown, 1); 1696 QUEUE_INSERT1(hio, send, ncomp); 1697 1698 /* 1699 * Let's wait for WRITE to finish. 1700 */ 1701 mtx_lock(&sync_lock); 1702 while (!ISSYNCREQDONE(hio)) 1703 cv_wait(&sync_cond, &sync_lock); 1704 mtx_unlock(&sync_lock); 1705 1706 if (hio->hio_errors[ncomp] != 0) { 1707 pjdlog_error("Unable to write synchronization data: %s.", 1708 strerror(hio->hio_errors[ncomp])); 1709 goto free_queue; 1710 } 1711 1712 synced += length; 1713free_queue: 1714 mtx_lock(&range_lock); 1715 rangelock_del(range_sync, offset, length); 1716 if (range_regular_wait) 1717 cv_signal(&range_regular_cond); 1718 mtx_unlock(&range_lock); 1719 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1720 hio); 1721 QUEUE_INSERT2(hio, free); 1722 } 1723 /* NOTREACHED */ 1724 return (NULL); 1725} 1726 1727static void 1728config_reload(void) 1729{ 1730 struct hastd_config *newcfg; 1731 struct hast_resource *res; 1732 unsigned int ii, ncomps; 1733 int modified; 1734 1735 pjdlog_info("Reloading configuration..."); 1736 1737 ncomps = HAST_NCOMPONENTS; 1738 1739 newcfg = yy_config_parse(cfgpath, false); 1740 if (newcfg == NULL) 1741 goto failed; 1742 1743 TAILQ_FOREACH(res, &newcfg->hc_resources, hr_next) { 1744 if (strcmp(res->hr_name, gres->hr_name) == 0) 1745 break; 1746 } 1747 /* 1748 * If resource was removed from the configuration file, resource 1749 * name, provider name or path to local component was modified we 1750 * shouldn't be here. This means that someone modified configuration 1751 * file and send SIGHUP to us instead of main hastd process. 1752 * Log advice and ignore the signal. 1753 */ 1754 if (res == NULL || strcmp(gres->hr_name, res->hr_name) != 0 || 1755 strcmp(gres->hr_provname, res->hr_provname) != 0 || 1756 strcmp(gres->hr_localpath, res->hr_localpath) != 0) { 1757 pjdlog_warning("To reload configuration send SIGHUP to the main hastd process (pid %u).", 1758 (unsigned int)getppid()); 1759 goto failed; 1760 } 1761 1762#define MODIFIED_REMOTEADDR 0x1 1763#define MODIFIED_REPLICATION 0x2 1764#define MODIFIED_TIMEOUT 0x4 1765#define MODIFIED_EXEC 0x8 1766 modified = 0; 1767 if (strcmp(gres->hr_remoteaddr, res->hr_remoteaddr) != 0) { 1768 /* 1769 * Don't copy res->hr_remoteaddr to gres just yet. 1770 * We want remote_close() to log disconnect from the old 1771 * addresses, not from the new ones. 1772 */ 1773 modified |= MODIFIED_REMOTEADDR; 1774 } 1775 if (gres->hr_replication != res->hr_replication) { 1776 gres->hr_replication = res->hr_replication; 1777 modified |= MODIFIED_REPLICATION; 1778 } 1779 if (gres->hr_timeout != res->hr_timeout) { 1780 gres->hr_timeout = res->hr_timeout; 1781 modified |= MODIFIED_TIMEOUT; 1782 } 1783 if (strcmp(gres->hr_exec, res->hr_exec) != 0) { 1784 strlcpy(gres->hr_exec, res->hr_exec, sizeof(gres->hr_exec)); 1785 modified |= MODIFIED_EXEC; 1786 } 1787 /* 1788 * If only timeout was modified we only need to change it without 1789 * reconnecting. 1790 */ 1791 if (modified == MODIFIED_TIMEOUT) { 1792 for (ii = 0; ii < ncomps; ii++) { 1793 if (!ISREMOTE(ii)) 1794 continue; 1795 rw_rlock(&hio_remote_lock[ii]); 1796 if (!ISCONNECTED(gres, ii)) { 1797 rw_unlock(&hio_remote_lock[ii]); 1798 continue; 1799 } 1800 rw_unlock(&hio_remote_lock[ii]); 1801 if (proto_timeout(gres->hr_remotein, 1802 gres->hr_timeout) < 0) { 1803 pjdlog_errno(LOG_WARNING, 1804 "Unable to set connection timeout"); 1805 } 1806 if (proto_timeout(gres->hr_remoteout, 1807 gres->hr_timeout) < 0) { 1808 pjdlog_errno(LOG_WARNING, 1809 "Unable to set connection timeout"); 1810 } 1811 } 1812 } else if ((modified & 1813 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) { 1814 for (ii = 0; ii < ncomps; ii++) { 1815 if (!ISREMOTE(ii)) 1816 continue; 1817 remote_close(gres, ii); 1818 } 1819 if (modified & MODIFIED_REMOTEADDR) { 1820 strlcpy(gres->hr_remoteaddr, res->hr_remoteaddr, 1821 sizeof(gres->hr_remoteaddr)); 1822 } 1823 } 1824#undef MODIFIED_REMOTEADDR 1825#undef MODIFIED_REPLICATION 1826#undef MODIFIED_TIMEOUT 1827#undef MODIFIED_EXEC 1828 1829 pjdlog_info("Configuration reloaded successfully."); 1830 return; 1831failed: 1832 if (newcfg != NULL) { 1833 if (newcfg->hc_controlconn != NULL) 1834 proto_close(newcfg->hc_controlconn); 1835 if (newcfg->hc_listenconn != NULL) 1836 proto_close(newcfg->hc_listenconn); 1837 yy_config_free(newcfg); 1838 } 1839 pjdlog_warning("Configuration not reloaded."); 1840} 1841 1842static void 1843keepalive_send(struct hast_resource *res, unsigned int ncomp) 1844{ 1845 struct nv *nv; 1846 1847 nv = nv_alloc(); 1848 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1849 if (nv_error(nv) != 0) { 1850 nv_free(nv); 1851 pjdlog_debug(1, 1852 "keepalive_send: Unable to prepare header to send."); 1853 return; 1854 } 1855 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1856 pjdlog_common(LOG_DEBUG, 1, errno, 1857 "keepalive_send: Unable to send request"); 1858 nv_free(nv); 1859 rw_unlock(&hio_remote_lock[ncomp]); 1860 remote_close(res, ncomp); 1861 rw_rlock(&hio_remote_lock[ncomp]); 1862 return; 1863 } 1864 nv_free(nv); 1865 pjdlog_debug(2, "keepalive_send: Request sent."); 1866} 1867 1868static void 1869guard_one(struct hast_resource *res, unsigned int ncomp) 1870{ 1871 struct proto_conn *in, *out; 1872 1873 if (!ISREMOTE(ncomp)) 1874 return; 1875 1876 rw_rlock(&hio_remote_lock[ncomp]); 1877 1878 if (!real_remote(res)) { 1879 rw_unlock(&hio_remote_lock[ncomp]); 1880 return; 1881 } 1882 1883 if (ISCONNECTED(res, ncomp)) { 1884 assert(res->hr_remotein != NULL); 1885 assert(res->hr_remoteout != NULL); 1886 keepalive_send(res, ncomp); 1887 } 1888 1889 if (ISCONNECTED(res, ncomp)) { 1890 assert(res->hr_remotein != NULL); 1891 assert(res->hr_remoteout != NULL); 1892 rw_unlock(&hio_remote_lock[ncomp]); 1893 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 1894 res->hr_remoteaddr); 1895 return; 1896 } 1897 1898 assert(res->hr_remotein == NULL); 1899 assert(res->hr_remoteout == NULL); 1900 /* 1901 * Upgrade the lock. It doesn't have to be atomic as no other thread 1902 * can change connection status from disconnected to connected. 1903 */ 1904 rw_unlock(&hio_remote_lock[ncomp]); 1905 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 1906 res->hr_remoteaddr); 1907 in = out = NULL; 1908 if (init_remote(res, &in, &out)) { 1909 rw_wlock(&hio_remote_lock[ncomp]); 1910 assert(res->hr_remotein == NULL); 1911 assert(res->hr_remoteout == NULL); 1912 assert(in != NULL && out != NULL); 1913 res->hr_remotein = in; 1914 res->hr_remoteout = out; 1915 rw_unlock(&hio_remote_lock[ncomp]); 1916 pjdlog_info("Successfully reconnected to %s.", 1917 res->hr_remoteaddr); 1918 sync_start(); 1919 } else { 1920 /* Both connections should be NULL. */ 1921 assert(res->hr_remotein == NULL); 1922 assert(res->hr_remoteout == NULL); 1923 assert(in == NULL && out == NULL); 1924 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 1925 res->hr_remoteaddr); 1926 } 1927} 1928 1929/* 1930 * Thread guards remote connections and reconnects when needed, handles 1931 * signals, etc. 1932 */ 1933static void * 1934guard_thread(void *arg) 1935{ 1936 struct hast_resource *res = arg; 1937 unsigned int ii, ncomps; 1938 struct timespec timeout; 1939 time_t lastcheck, now; 1940 sigset_t mask; 1941 int signo; 1942 1943 ncomps = HAST_NCOMPONENTS; 1944 lastcheck = time(NULL); 1945 1946 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 1947 PJDLOG_VERIFY(sigaddset(&mask, SIGHUP) == 0); 1948 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 1949 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 1950 PJDLOG_VERIFY(sigaddset(&mask, SIGCHLD) == 0); 1951 1952 timeout.tv_nsec = 0; 1953 signo = -1; 1954 1955 for (;;) { 1956 switch (signo) { 1957 case SIGHUP: 1958 config_reload(); 1959 break; 1960 case SIGINT: 1961 case SIGTERM: 1962 sigexit_received = true; 1963 primary_exitx(EX_OK, 1964 "Termination signal received, exiting."); 1965 break; 1966 default: 1967 break; 1968 } 1969 hook_check(signo == SIGCHLD); 1970 1971 pjdlog_debug(2, "remote_guard: Checking connections."); 1972 now = time(NULL); 1973 if (lastcheck + RETRY_SLEEP <= now) { 1974 for (ii = 0; ii < ncomps; ii++) 1975 guard_one(res, ii); 1976 lastcheck = now; 1977 } 1978 timeout.tv_sec = RETRY_SLEEP; 1979 signo = sigtimedwait(&mask, NULL, &timeout); 1980 } 1981 /* NOTREACHED */ 1982 return (NULL); 1983} 1984