primary.c revision 218192
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 218192 2011-02-02 15:42:00Z 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 <err.h> 44#include <errno.h> 45#include <fcntl.h> 46#include <libgeom.h> 47#include <pthread.h> 48#include <signal.h> 49#include <stdint.h> 50#include <stdio.h> 51#include <string.h> 52#include <sysexits.h> 53#include <unistd.h> 54 55#include <activemap.h> 56#include <nv.h> 57#include <rangelock.h> 58 59#include "control.h" 60#include "event.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, timeout) do { \ 183 bool _last; \ 184 \ 185 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 186 _last = false; \ 187 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \ 188 cv_timedwait(&hio_##name##_list_cond[(ncomp)], \ 189 &hio_##name##_list_lock[(ncomp)], (timeout)); \ 190 if ((timeout) != 0) \ 191 _last = true; \ 192 } \ 193 if (hio != NULL) { \ 194 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \ 195 hio_next[(ncomp)]); \ 196 } \ 197 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 198} while (0) 199#define QUEUE_TAKE2(hio, name) do { \ 200 mtx_lock(&hio_##name##_list_lock); \ 201 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 202 cv_wait(&hio_##name##_list_cond, \ 203 &hio_##name##_list_lock); \ 204 } \ 205 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \ 206 mtx_unlock(&hio_##name##_list_lock); \ 207} while (0) 208 209#define SYNCREQ(hio) do { \ 210 (hio)->hio_ggio.gctl_unit = -1; \ 211 (hio)->hio_ggio.gctl_seq = 1; \ 212} while (0) 213#define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1) 214#define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0) 215#define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2) 216 217static struct hast_resource *gres; 218 219static pthread_mutex_t range_lock; 220static struct rangelocks *range_regular; 221static bool range_regular_wait; 222static pthread_cond_t range_regular_cond; 223static struct rangelocks *range_sync; 224static bool range_sync_wait; 225static pthread_cond_t range_sync_cond; 226 227static void *ggate_recv_thread(void *arg); 228static void *local_send_thread(void *arg); 229static void *remote_send_thread(void *arg); 230static void *remote_recv_thread(void *arg); 231static void *ggate_send_thread(void *arg); 232static void *sync_thread(void *arg); 233static void *guard_thread(void *arg); 234 235static void 236cleanup(struct hast_resource *res) 237{ 238 int rerrno; 239 240 /* Remember errno. */ 241 rerrno = errno; 242 243 /* Destroy ggate provider if we created one. */ 244 if (res->hr_ggateunit >= 0) { 245 struct g_gate_ctl_destroy ggiod; 246 247 bzero(&ggiod, sizeof(ggiod)); 248 ggiod.gctl_version = G_GATE_VERSION; 249 ggiod.gctl_unit = res->hr_ggateunit; 250 ggiod.gctl_force = 1; 251 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) { 252 pjdlog_errno(LOG_WARNING, 253 "Unable to destroy hast/%s device", 254 res->hr_provname); 255 } 256 res->hr_ggateunit = -1; 257 } 258 259 /* Restore errno. */ 260 errno = rerrno; 261} 262 263static __dead2 void 264primary_exit(int exitcode, const char *fmt, ...) 265{ 266 va_list ap; 267 268 PJDLOG_ASSERT(exitcode != EX_OK); 269 va_start(ap, fmt); 270 pjdlogv_errno(LOG_ERR, fmt, ap); 271 va_end(ap); 272 cleanup(gres); 273 exit(exitcode); 274} 275 276static __dead2 void 277primary_exitx(int exitcode, const char *fmt, ...) 278{ 279 va_list ap; 280 281 va_start(ap, fmt); 282 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 283 va_end(ap); 284 cleanup(gres); 285 exit(exitcode); 286} 287 288static int 289hast_activemap_flush(struct hast_resource *res) 290{ 291 const unsigned char *buf; 292 size_t size; 293 294 buf = activemap_bitmap(res->hr_amp, &size); 295 PJDLOG_ASSERT(buf != NULL); 296 PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0); 297 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 298 (ssize_t)size) { 299 KEEP_ERRNO(pjdlog_errno(LOG_ERR, 300 "Unable to flush activemap to disk")); 301 return (-1); 302 } 303 return (0); 304} 305 306static bool 307real_remote(const struct hast_resource *res) 308{ 309 310 return (strcmp(res->hr_remoteaddr, "none") != 0); 311} 312 313static void 314init_environment(struct hast_resource *res __unused) 315{ 316 struct hio *hio; 317 unsigned int ii, ncomps; 318 319 /* 320 * In the future it might be per-resource value. 321 */ 322 ncomps = HAST_NCOMPONENTS; 323 324 /* 325 * Allocate memory needed by lists. 326 */ 327 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 328 if (hio_send_list == NULL) { 329 primary_exitx(EX_TEMPFAIL, 330 "Unable to allocate %zu bytes of memory for send lists.", 331 sizeof(hio_send_list[0]) * ncomps); 332 } 333 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 334 if (hio_send_list_lock == NULL) { 335 primary_exitx(EX_TEMPFAIL, 336 "Unable to allocate %zu bytes of memory for send list locks.", 337 sizeof(hio_send_list_lock[0]) * ncomps); 338 } 339 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 340 if (hio_send_list_cond == NULL) { 341 primary_exitx(EX_TEMPFAIL, 342 "Unable to allocate %zu bytes of memory for send list condition variables.", 343 sizeof(hio_send_list_cond[0]) * ncomps); 344 } 345 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 346 if (hio_recv_list == NULL) { 347 primary_exitx(EX_TEMPFAIL, 348 "Unable to allocate %zu bytes of memory for recv lists.", 349 sizeof(hio_recv_list[0]) * ncomps); 350 } 351 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 352 if (hio_recv_list_lock == NULL) { 353 primary_exitx(EX_TEMPFAIL, 354 "Unable to allocate %zu bytes of memory for recv list locks.", 355 sizeof(hio_recv_list_lock[0]) * ncomps); 356 } 357 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 358 if (hio_recv_list_cond == NULL) { 359 primary_exitx(EX_TEMPFAIL, 360 "Unable to allocate %zu bytes of memory for recv list condition variables.", 361 sizeof(hio_recv_list_cond[0]) * ncomps); 362 } 363 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 364 if (hio_remote_lock == NULL) { 365 primary_exitx(EX_TEMPFAIL, 366 "Unable to allocate %zu bytes of memory for remote connections locks.", 367 sizeof(hio_remote_lock[0]) * ncomps); 368 } 369 370 /* 371 * Initialize lists, their locks and theirs condition variables. 372 */ 373 TAILQ_INIT(&hio_free_list); 374 mtx_init(&hio_free_list_lock); 375 cv_init(&hio_free_list_cond); 376 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 377 TAILQ_INIT(&hio_send_list[ii]); 378 mtx_init(&hio_send_list_lock[ii]); 379 cv_init(&hio_send_list_cond[ii]); 380 TAILQ_INIT(&hio_recv_list[ii]); 381 mtx_init(&hio_recv_list_lock[ii]); 382 cv_init(&hio_recv_list_cond[ii]); 383 rw_init(&hio_remote_lock[ii]); 384 } 385 TAILQ_INIT(&hio_done_list); 386 mtx_init(&hio_done_list_lock); 387 cv_init(&hio_done_list_cond); 388 mtx_init(&metadata_lock); 389 390 /* 391 * Allocate requests pool and initialize requests. 392 */ 393 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 394 hio = malloc(sizeof(*hio)); 395 if (hio == NULL) { 396 primary_exitx(EX_TEMPFAIL, 397 "Unable to allocate %zu bytes of memory for hio request.", 398 sizeof(*hio)); 399 } 400 hio->hio_countdown = 0; 401 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 402 if (hio->hio_errors == NULL) { 403 primary_exitx(EX_TEMPFAIL, 404 "Unable allocate %zu bytes of memory for hio errors.", 405 sizeof(hio->hio_errors[0]) * ncomps); 406 } 407 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 408 if (hio->hio_next == NULL) { 409 primary_exitx(EX_TEMPFAIL, 410 "Unable allocate %zu bytes of memory for hio_next field.", 411 sizeof(hio->hio_next[0]) * ncomps); 412 } 413 hio->hio_ggio.gctl_version = G_GATE_VERSION; 414 hio->hio_ggio.gctl_data = malloc(MAXPHYS); 415 if (hio->hio_ggio.gctl_data == NULL) { 416 primary_exitx(EX_TEMPFAIL, 417 "Unable to allocate %zu bytes of memory for gctl_data.", 418 MAXPHYS); 419 } 420 hio->hio_ggio.gctl_length = MAXPHYS; 421 hio->hio_ggio.gctl_error = 0; 422 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 423 } 424} 425 426static bool 427init_resuid(struct hast_resource *res) 428{ 429 430 mtx_lock(&metadata_lock); 431 if (res->hr_resuid != 0) { 432 mtx_unlock(&metadata_lock); 433 return (false); 434 } else { 435 /* Initialize unique resource identifier. */ 436 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 437 mtx_unlock(&metadata_lock); 438 if (metadata_write(res) < 0) 439 exit(EX_NOINPUT); 440 return (true); 441 } 442} 443 444static void 445init_local(struct hast_resource *res) 446{ 447 unsigned char *buf; 448 size_t mapsize; 449 450 if (metadata_read(res, true) < 0) 451 exit(EX_NOINPUT); 452 mtx_init(&res->hr_amp_lock); 453 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 454 res->hr_local_sectorsize, res->hr_keepdirty) < 0) { 455 primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 456 } 457 mtx_init(&range_lock); 458 cv_init(&range_regular_cond); 459 if (rangelock_init(&range_regular) < 0) 460 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 461 cv_init(&range_sync_cond); 462 if (rangelock_init(&range_sync) < 0) 463 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 464 mapsize = activemap_ondisk_size(res->hr_amp); 465 buf = calloc(1, mapsize); 466 if (buf == NULL) { 467 primary_exitx(EX_TEMPFAIL, 468 "Unable to allocate buffer for activemap."); 469 } 470 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 471 (ssize_t)mapsize) { 472 primary_exit(EX_NOINPUT, "Unable to read activemap"); 473 } 474 activemap_copyin(res->hr_amp, buf, mapsize); 475 free(buf); 476 if (res->hr_resuid != 0) 477 return; 478 /* 479 * We're using provider for the first time. Initialize local and remote 480 * counters. We don't initialize resuid here, as we want to do it just 481 * in time. The reason for this is that we want to inform secondary 482 * that there were no writes yet, so there is no need to synchronize 483 * anything. 484 */ 485 res->hr_primary_localcnt = 1; 486 res->hr_primary_remotecnt = 0; 487 if (metadata_write(res) < 0) 488 exit(EX_NOINPUT); 489} 490 491static bool 492init_remote(struct hast_resource *res, struct proto_conn **inp, 493 struct proto_conn **outp) 494{ 495 struct proto_conn *in, *out; 496 struct nv *nvout, *nvin; 497 const unsigned char *token; 498 unsigned char *map; 499 const char *errmsg; 500 int32_t extentsize; 501 int64_t datasize; 502 uint32_t mapsize; 503 size_t size; 504 505 PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 506 PJDLOG_ASSERT(real_remote(res)); 507 508 in = out = NULL; 509 errmsg = NULL; 510 511 /* Prepare outgoing connection with remote node. */ 512 if (proto_client(res->hr_remoteaddr, &out) < 0) { 513 primary_exit(EX_TEMPFAIL, 514 "Unable to create outgoing connection to %s", 515 res->hr_remoteaddr); 516 } 517 /* Try to connect, but accept failure. */ 518 if (proto_connect(out, HAST_TIMEOUT) < 0) { 519 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 520 res->hr_remoteaddr); 521 goto close; 522 } 523 /* Error in setting timeout is not critical, but why should it fail? */ 524 if (proto_timeout(out, res->hr_timeout) < 0) 525 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 526 /* 527 * First handshake step. 528 * Setup outgoing connection with remote node. 529 */ 530 nvout = nv_alloc(); 531 nv_add_string(nvout, res->hr_name, "resource"); 532 if (nv_error(nvout) != 0) { 533 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 534 "Unable to allocate header for connection with %s", 535 res->hr_remoteaddr); 536 nv_free(nvout); 537 goto close; 538 } 539 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) { 540 pjdlog_errno(LOG_WARNING, 541 "Unable to send handshake header to %s", 542 res->hr_remoteaddr); 543 nv_free(nvout); 544 goto close; 545 } 546 nv_free(nvout); 547 if (hast_proto_recv_hdr(out, &nvin) < 0) { 548 pjdlog_errno(LOG_WARNING, 549 "Unable to receive handshake header from %s", 550 res->hr_remoteaddr); 551 goto close; 552 } 553 errmsg = nv_get_string(nvin, "errmsg"); 554 if (errmsg != NULL) { 555 pjdlog_warning("%s", errmsg); 556 nv_free(nvin); 557 goto close; 558 } 559 token = nv_get_uint8_array(nvin, &size, "token"); 560 if (token == NULL) { 561 pjdlog_warning("Handshake header from %s has no 'token' field.", 562 res->hr_remoteaddr); 563 nv_free(nvin); 564 goto close; 565 } 566 if (size != sizeof(res->hr_token)) { 567 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 568 res->hr_remoteaddr, size, sizeof(res->hr_token)); 569 nv_free(nvin); 570 goto close; 571 } 572 bcopy(token, res->hr_token, sizeof(res->hr_token)); 573 nv_free(nvin); 574 575 /* 576 * Second handshake step. 577 * Setup incoming connection with remote node. 578 */ 579 if (proto_client(res->hr_remoteaddr, &in) < 0) { 580 primary_exit(EX_TEMPFAIL, 581 "Unable to create incoming connection to %s", 582 res->hr_remoteaddr); 583 } 584 /* Try to connect, but accept failure. */ 585 if (proto_connect(in, HAST_TIMEOUT) < 0) { 586 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 587 res->hr_remoteaddr); 588 goto close; 589 } 590 /* Error in setting timeout is not critical, but why should it fail? */ 591 if (proto_timeout(in, res->hr_timeout) < 0) 592 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 593 nvout = nv_alloc(); 594 nv_add_string(nvout, res->hr_name, "resource"); 595 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 596 "token"); 597 if (res->hr_resuid == 0) { 598 /* 599 * The resuid field was not yet initialized. 600 * Because we do synchronization inside init_resuid(), it is 601 * possible that someone already initialized it, the function 602 * will return false then, but if we successfully initialized 603 * it, we will get true. True means that there were no writes 604 * to this resource yet and we want to inform secondary that 605 * synchronization is not needed by sending "virgin" argument. 606 */ 607 if (init_resuid(res)) 608 nv_add_int8(nvout, 1, "virgin"); 609 } 610 nv_add_uint64(nvout, res->hr_resuid, "resuid"); 611 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 612 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 613 if (nv_error(nvout) != 0) { 614 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 615 "Unable to allocate header for connection with %s", 616 res->hr_remoteaddr); 617 nv_free(nvout); 618 goto close; 619 } 620 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) { 621 pjdlog_errno(LOG_WARNING, 622 "Unable to send handshake header to %s", 623 res->hr_remoteaddr); 624 nv_free(nvout); 625 goto close; 626 } 627 nv_free(nvout); 628 if (hast_proto_recv_hdr(out, &nvin) < 0) { 629 pjdlog_errno(LOG_WARNING, 630 "Unable to receive handshake header from %s", 631 res->hr_remoteaddr); 632 goto close; 633 } 634 errmsg = nv_get_string(nvin, "errmsg"); 635 if (errmsg != NULL) { 636 pjdlog_warning("%s", errmsg); 637 nv_free(nvin); 638 goto close; 639 } 640 datasize = nv_get_int64(nvin, "datasize"); 641 if (datasize != res->hr_datasize) { 642 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 643 (intmax_t)res->hr_datasize, (intmax_t)datasize); 644 nv_free(nvin); 645 goto close; 646 } 647 extentsize = nv_get_int32(nvin, "extentsize"); 648 if (extentsize != res->hr_extentsize) { 649 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 650 (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 651 nv_free(nvin); 652 goto close; 653 } 654 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 655 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 656 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 657 map = NULL; 658 mapsize = nv_get_uint32(nvin, "mapsize"); 659 if (mapsize > 0) { 660 map = malloc(mapsize); 661 if (map == NULL) { 662 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 663 (uintmax_t)mapsize); 664 nv_free(nvin); 665 goto close; 666 } 667 /* 668 * Remote node have some dirty extents on its own, lets 669 * download its activemap. 670 */ 671 if (hast_proto_recv_data(res, out, nvin, map, 672 mapsize) < 0) { 673 pjdlog_errno(LOG_ERR, 674 "Unable to receive remote activemap"); 675 nv_free(nvin); 676 free(map); 677 goto close; 678 } 679 /* 680 * Merge local and remote bitmaps. 681 */ 682 activemap_merge(res->hr_amp, map, mapsize); 683 free(map); 684 /* 685 * Now that we merged bitmaps from both nodes, flush it to the 686 * disk before we start to synchronize. 687 */ 688 (void)hast_activemap_flush(res); 689 } 690 nv_free(nvin); 691 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 692 if (inp != NULL && outp != NULL) { 693 *inp = in; 694 *outp = out; 695 } else { 696 res->hr_remotein = in; 697 res->hr_remoteout = out; 698 } 699 event_send(res, EVENT_CONNECT); 700 return (true); 701close: 702 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 703 event_send(res, EVENT_SPLITBRAIN); 704 proto_close(out); 705 if (in != NULL) 706 proto_close(in); 707 return (false); 708} 709 710static void 711sync_start(void) 712{ 713 714 mtx_lock(&sync_lock); 715 sync_inprogress = true; 716 mtx_unlock(&sync_lock); 717 cv_signal(&sync_cond); 718} 719 720static void 721sync_stop(void) 722{ 723 724 mtx_lock(&sync_lock); 725 if (sync_inprogress) 726 sync_inprogress = false; 727 mtx_unlock(&sync_lock); 728} 729 730static void 731init_ggate(struct hast_resource *res) 732{ 733 struct g_gate_ctl_create ggiocreate; 734 struct g_gate_ctl_cancel ggiocancel; 735 736 /* 737 * We communicate with ggate via /dev/ggctl. Open it. 738 */ 739 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 740 if (res->hr_ggatefd < 0) 741 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 742 /* 743 * Create provider before trying to connect, as connection failure 744 * is not critical, but may take some time. 745 */ 746 bzero(&ggiocreate, sizeof(ggiocreate)); 747 ggiocreate.gctl_version = G_GATE_VERSION; 748 ggiocreate.gctl_mediasize = res->hr_datasize; 749 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 750 ggiocreate.gctl_flags = 0; 751 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE; 752 ggiocreate.gctl_timeout = 0; 753 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 754 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 755 res->hr_provname); 756 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 757 pjdlog_info("Device hast/%s created.", res->hr_provname); 758 res->hr_ggateunit = ggiocreate.gctl_unit; 759 return; 760 } 761 if (errno != EEXIST) { 762 primary_exit(EX_OSERR, "Unable to create hast/%s device", 763 res->hr_provname); 764 } 765 pjdlog_debug(1, 766 "Device hast/%s already exists, we will try to take it over.", 767 res->hr_provname); 768 /* 769 * If we received EEXIST, we assume that the process who created the 770 * provider died and didn't clean up. In that case we will start from 771 * where he left of. 772 */ 773 bzero(&ggiocancel, sizeof(ggiocancel)); 774 ggiocancel.gctl_version = G_GATE_VERSION; 775 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 776 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 777 res->hr_provname); 778 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 779 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 780 res->hr_ggateunit = ggiocancel.gctl_unit; 781 return; 782 } 783 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 784 res->hr_provname); 785} 786 787void 788hastd_primary(struct hast_resource *res) 789{ 790 pthread_t td; 791 pid_t pid; 792 int error, mode; 793 794 /* 795 * Create communication channel between parent and child. 796 */ 797 if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 798 /* TODO: There's no need for this to be fatal error. */ 799 KEEP_ERRNO((void)pidfile_remove(pfh)); 800 pjdlog_exit(EX_OSERR, 801 "Unable to create control sockets between parent and child"); 802 } 803 /* 804 * Create communication channel between child and parent. 805 */ 806 if (proto_client("socketpair://", &res->hr_event) < 0) { 807 /* TODO: There's no need for this to be fatal error. */ 808 KEEP_ERRNO((void)pidfile_remove(pfh)); 809 pjdlog_exit(EX_OSERR, 810 "Unable to create event sockets between child and parent"); 811 } 812 813 pid = fork(); 814 if (pid < 0) { 815 /* TODO: There's no need for this to be fatal error. */ 816 KEEP_ERRNO((void)pidfile_remove(pfh)); 817 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 818 } 819 820 if (pid > 0) { 821 /* This is parent. */ 822 /* Declare that we are receiver. */ 823 proto_recv(res->hr_event, NULL, 0); 824 /* Declare that we are sender. */ 825 proto_send(res->hr_ctrl, NULL, 0); 826 res->hr_workerpid = pid; 827 return; 828 } 829 830 gres = res; 831 mode = pjdlog_mode_get(); 832 833 /* Declare that we are sender. */ 834 proto_send(res->hr_event, NULL, 0); 835 /* Declare that we are receiver. */ 836 proto_recv(res->hr_ctrl, NULL, 0); 837 descriptors_cleanup(res); 838 839 descriptors_assert(res, mode); 840 841 pjdlog_init(mode); 842 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 843 setproctitle("%s (primary)", res->hr_name); 844 845 init_local(res); 846 init_ggate(res); 847 init_environment(res); 848 849 if (drop_privs() != 0) { 850 cleanup(res); 851 exit(EX_CONFIG); 852 } 853 854 /* 855 * Create the guard thread first, so we can handle signals from the 856 * very begining. 857 */ 858 error = pthread_create(&td, NULL, guard_thread, res); 859 PJDLOG_ASSERT(error == 0); 860 /* 861 * Create the control thread before sending any event to the parent, 862 * as we can deadlock when parent sends control request to worker, 863 * but worker has no control thread started yet, so parent waits. 864 * In the meantime worker sends an event to the parent, but parent 865 * is unable to handle the event, because it waits for control 866 * request response. 867 */ 868 error = pthread_create(&td, NULL, ctrl_thread, res); 869 PJDLOG_ASSERT(error == 0); 870 if (real_remote(res) && init_remote(res, NULL, NULL)) 871 sync_start(); 872 error = pthread_create(&td, NULL, ggate_recv_thread, res); 873 PJDLOG_ASSERT(error == 0); 874 error = pthread_create(&td, NULL, local_send_thread, res); 875 PJDLOG_ASSERT(error == 0); 876 error = pthread_create(&td, NULL, remote_send_thread, res); 877 PJDLOG_ASSERT(error == 0); 878 error = pthread_create(&td, NULL, remote_recv_thread, res); 879 PJDLOG_ASSERT(error == 0); 880 error = pthread_create(&td, NULL, ggate_send_thread, res); 881 PJDLOG_ASSERT(error == 0); 882 (void)sync_thread(res); 883} 884 885static void 886reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 887{ 888 char msg[1024]; 889 va_list ap; 890 int len; 891 892 va_start(ap, fmt); 893 len = vsnprintf(msg, sizeof(msg), fmt, ap); 894 va_end(ap); 895 if ((size_t)len < sizeof(msg)) { 896 switch (ggio->gctl_cmd) { 897 case BIO_READ: 898 (void)snprintf(msg + len, sizeof(msg) - len, 899 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 900 (uintmax_t)ggio->gctl_length); 901 break; 902 case BIO_DELETE: 903 (void)snprintf(msg + len, sizeof(msg) - len, 904 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 905 (uintmax_t)ggio->gctl_length); 906 break; 907 case BIO_FLUSH: 908 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 909 break; 910 case BIO_WRITE: 911 (void)snprintf(msg + len, sizeof(msg) - len, 912 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 913 (uintmax_t)ggio->gctl_length); 914 break; 915 default: 916 (void)snprintf(msg + len, sizeof(msg) - len, 917 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 918 break; 919 } 920 } 921 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 922} 923 924static void 925remote_close(struct hast_resource *res, int ncomp) 926{ 927 928 rw_wlock(&hio_remote_lock[ncomp]); 929 /* 930 * A race is possible between dropping rlock and acquiring wlock - 931 * another thread can close connection in-between. 932 */ 933 if (!ISCONNECTED(res, ncomp)) { 934 PJDLOG_ASSERT(res->hr_remotein == NULL); 935 PJDLOG_ASSERT(res->hr_remoteout == NULL); 936 rw_unlock(&hio_remote_lock[ncomp]); 937 return; 938 } 939 940 PJDLOG_ASSERT(res->hr_remotein != NULL); 941 PJDLOG_ASSERT(res->hr_remoteout != NULL); 942 943 pjdlog_debug(2, "Closing incoming connection to %s.", 944 res->hr_remoteaddr); 945 proto_close(res->hr_remotein); 946 res->hr_remotein = NULL; 947 pjdlog_debug(2, "Closing outgoing connection to %s.", 948 res->hr_remoteaddr); 949 proto_close(res->hr_remoteout); 950 res->hr_remoteout = NULL; 951 952 rw_unlock(&hio_remote_lock[ncomp]); 953 954 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 955 956 /* 957 * Stop synchronization if in-progress. 958 */ 959 sync_stop(); 960 961 event_send(res, EVENT_DISCONNECT); 962} 963 964/* 965 * Thread receives ggate I/O requests from the kernel and passes them to 966 * appropriate threads: 967 * WRITE - always goes to both local_send and remote_send threads 968 * READ (when the block is up-to-date on local component) - 969 * only local_send thread 970 * READ (when the block isn't up-to-date on local component) - 971 * only remote_send thread 972 * DELETE - always goes to both local_send and remote_send threads 973 * FLUSH - always goes to both local_send and remote_send threads 974 */ 975static void * 976ggate_recv_thread(void *arg) 977{ 978 struct hast_resource *res = arg; 979 struct g_gate_ctl_io *ggio; 980 struct hio *hio; 981 unsigned int ii, ncomp, ncomps; 982 int error; 983 984 ncomps = HAST_NCOMPONENTS; 985 986 for (;;) { 987 pjdlog_debug(2, "ggate_recv: Taking free request."); 988 QUEUE_TAKE2(hio, free); 989 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 990 ggio = &hio->hio_ggio; 991 ggio->gctl_unit = res->hr_ggateunit; 992 ggio->gctl_length = MAXPHYS; 993 ggio->gctl_error = 0; 994 pjdlog_debug(2, 995 "ggate_recv: (%p) Waiting for request from the kernel.", 996 hio); 997 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 998 if (sigexit_received) 999 pthread_exit(NULL); 1000 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 1001 } 1002 error = ggio->gctl_error; 1003 switch (error) { 1004 case 0: 1005 break; 1006 case ECANCELED: 1007 /* Exit gracefully. */ 1008 if (!sigexit_received) { 1009 pjdlog_debug(2, 1010 "ggate_recv: (%p) Received cancel from the kernel.", 1011 hio); 1012 pjdlog_info("Received cancel from the kernel, exiting."); 1013 } 1014 pthread_exit(NULL); 1015 case ENOMEM: 1016 /* 1017 * Buffer too small? Impossible, we allocate MAXPHYS 1018 * bytes - request can't be bigger than that. 1019 */ 1020 /* FALLTHROUGH */ 1021 case ENXIO: 1022 default: 1023 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 1024 strerror(error)); 1025 } 1026 for (ii = 0; ii < ncomps; ii++) 1027 hio->hio_errors[ii] = EINVAL; 1028 reqlog(LOG_DEBUG, 2, ggio, 1029 "ggate_recv: (%p) Request received from the kernel: ", 1030 hio); 1031 /* 1032 * Inform all components about new write request. 1033 * For read request prefer local component unless the given 1034 * range is out-of-date, then use remote component. 1035 */ 1036 switch (ggio->gctl_cmd) { 1037 case BIO_READ: 1038 pjdlog_debug(2, 1039 "ggate_recv: (%p) Moving request to the send queue.", 1040 hio); 1041 refcount_init(&hio->hio_countdown, 1); 1042 mtx_lock(&metadata_lock); 1043 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 1044 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1045 /* 1046 * This range is up-to-date on local component, 1047 * so handle request locally. 1048 */ 1049 /* Local component is 0 for now. */ 1050 ncomp = 0; 1051 } else /* if (res->hr_syncsrc == 1052 HAST_SYNCSRC_SECONDARY) */ { 1053 PJDLOG_ASSERT(res->hr_syncsrc == 1054 HAST_SYNCSRC_SECONDARY); 1055 /* 1056 * This range is out-of-date on local component, 1057 * so send request to the remote node. 1058 */ 1059 /* Remote component is 1 for now. */ 1060 ncomp = 1; 1061 } 1062 mtx_unlock(&metadata_lock); 1063 QUEUE_INSERT1(hio, send, ncomp); 1064 break; 1065 case BIO_WRITE: 1066 if (res->hr_resuid == 0) { 1067 /* This is first write, initialize resuid. */ 1068 (void)init_resuid(res); 1069 } 1070 for (;;) { 1071 mtx_lock(&range_lock); 1072 if (rangelock_islocked(range_sync, 1073 ggio->gctl_offset, ggio->gctl_length)) { 1074 pjdlog_debug(2, 1075 "regular: Range offset=%jd length=%zu locked.", 1076 (intmax_t)ggio->gctl_offset, 1077 (size_t)ggio->gctl_length); 1078 range_regular_wait = true; 1079 cv_wait(&range_regular_cond, &range_lock); 1080 range_regular_wait = false; 1081 mtx_unlock(&range_lock); 1082 continue; 1083 } 1084 if (rangelock_add(range_regular, 1085 ggio->gctl_offset, ggio->gctl_length) < 0) { 1086 mtx_unlock(&range_lock); 1087 pjdlog_debug(2, 1088 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1089 (intmax_t)ggio->gctl_offset, 1090 (size_t)ggio->gctl_length); 1091 sleep(1); 1092 continue; 1093 } 1094 mtx_unlock(&range_lock); 1095 break; 1096 } 1097 mtx_lock(&res->hr_amp_lock); 1098 if (activemap_write_start(res->hr_amp, 1099 ggio->gctl_offset, ggio->gctl_length)) { 1100 (void)hast_activemap_flush(res); 1101 } 1102 mtx_unlock(&res->hr_amp_lock); 1103 /* FALLTHROUGH */ 1104 case BIO_DELETE: 1105 case BIO_FLUSH: 1106 pjdlog_debug(2, 1107 "ggate_recv: (%p) Moving request to the send queues.", 1108 hio); 1109 refcount_init(&hio->hio_countdown, ncomps); 1110 for (ii = 0; ii < ncomps; ii++) 1111 QUEUE_INSERT1(hio, send, ii); 1112 break; 1113 } 1114 } 1115 /* NOTREACHED */ 1116 return (NULL); 1117} 1118 1119/* 1120 * Thread reads from or writes to local component. 1121 * If local read fails, it redirects it to remote_send thread. 1122 */ 1123static void * 1124local_send_thread(void *arg) 1125{ 1126 struct hast_resource *res = arg; 1127 struct g_gate_ctl_io *ggio; 1128 struct hio *hio; 1129 unsigned int ncomp, rncomp; 1130 ssize_t ret; 1131 1132 /* Local component is 0 for now. */ 1133 ncomp = 0; 1134 /* Remote component is 1 for now. */ 1135 rncomp = 1; 1136 1137 for (;;) { 1138 pjdlog_debug(2, "local_send: Taking request."); 1139 QUEUE_TAKE1(hio, send, ncomp, 0); 1140 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1141 ggio = &hio->hio_ggio; 1142 switch (ggio->gctl_cmd) { 1143 case BIO_READ: 1144 ret = pread(res->hr_localfd, ggio->gctl_data, 1145 ggio->gctl_length, 1146 ggio->gctl_offset + res->hr_localoff); 1147 if (ret == ggio->gctl_length) 1148 hio->hio_errors[ncomp] = 0; 1149 else { 1150 /* 1151 * If READ failed, try to read from remote node. 1152 */ 1153 if (ret < 0) { 1154 reqlog(LOG_WARNING, 0, ggio, 1155 "Local request failed (%s), trying remote node. ", 1156 strerror(errno)); 1157 } else if (ret != ggio->gctl_length) { 1158 reqlog(LOG_WARNING, 0, ggio, 1159 "Local request failed (%zd != %jd), trying remote node. ", 1160 ret, (intmax_t)ggio->gctl_length); 1161 } 1162 QUEUE_INSERT1(hio, send, rncomp); 1163 continue; 1164 } 1165 break; 1166 case BIO_WRITE: 1167 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1168 ggio->gctl_length, 1169 ggio->gctl_offset + res->hr_localoff); 1170 if (ret < 0) { 1171 hio->hio_errors[ncomp] = errno; 1172 reqlog(LOG_WARNING, 0, ggio, 1173 "Local request failed (%s): ", 1174 strerror(errno)); 1175 } else if (ret != ggio->gctl_length) { 1176 hio->hio_errors[ncomp] = EIO; 1177 reqlog(LOG_WARNING, 0, ggio, 1178 "Local request failed (%zd != %jd): ", 1179 ret, (intmax_t)ggio->gctl_length); 1180 } else { 1181 hio->hio_errors[ncomp] = 0; 1182 } 1183 break; 1184 case BIO_DELETE: 1185 ret = g_delete(res->hr_localfd, 1186 ggio->gctl_offset + res->hr_localoff, 1187 ggio->gctl_length); 1188 if (ret < 0) { 1189 hio->hio_errors[ncomp] = errno; 1190 reqlog(LOG_WARNING, 0, ggio, 1191 "Local request failed (%s): ", 1192 strerror(errno)); 1193 } else { 1194 hio->hio_errors[ncomp] = 0; 1195 } 1196 break; 1197 case BIO_FLUSH: 1198 ret = g_flush(res->hr_localfd); 1199 if (ret < 0) { 1200 hio->hio_errors[ncomp] = errno; 1201 reqlog(LOG_WARNING, 0, ggio, 1202 "Local request failed (%s): ", 1203 strerror(errno)); 1204 } else { 1205 hio->hio_errors[ncomp] = 0; 1206 } 1207 break; 1208 } 1209 if (refcount_release(&hio->hio_countdown)) { 1210 if (ISSYNCREQ(hio)) { 1211 mtx_lock(&sync_lock); 1212 SYNCREQDONE(hio); 1213 mtx_unlock(&sync_lock); 1214 cv_signal(&sync_cond); 1215 } else { 1216 pjdlog_debug(2, 1217 "local_send: (%p) Moving request to the done queue.", 1218 hio); 1219 QUEUE_INSERT2(hio, done); 1220 } 1221 } 1222 } 1223 /* NOTREACHED */ 1224 return (NULL); 1225} 1226 1227static void 1228keepalive_send(struct hast_resource *res, unsigned int ncomp) 1229{ 1230 struct nv *nv; 1231 1232 if (!ISCONNECTED(res, ncomp)) 1233 return; 1234 1235 PJDLOG_ASSERT(res->hr_remotein != NULL); 1236 PJDLOG_ASSERT(res->hr_remoteout != NULL); 1237 1238 nv = nv_alloc(); 1239 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1240 if (nv_error(nv) != 0) { 1241 nv_free(nv); 1242 pjdlog_debug(1, 1243 "keepalive_send: Unable to prepare header to send."); 1244 return; 1245 } 1246 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1247 pjdlog_common(LOG_DEBUG, 1, errno, 1248 "keepalive_send: Unable to send request"); 1249 nv_free(nv); 1250 rw_unlock(&hio_remote_lock[ncomp]); 1251 remote_close(res, ncomp); 1252 rw_rlock(&hio_remote_lock[ncomp]); 1253 return; 1254 } 1255 nv_free(nv); 1256 pjdlog_debug(2, "keepalive_send: Request sent."); 1257} 1258 1259/* 1260 * Thread sends request to secondary node. 1261 */ 1262static void * 1263remote_send_thread(void *arg) 1264{ 1265 struct hast_resource *res = arg; 1266 struct g_gate_ctl_io *ggio; 1267 time_t lastcheck, now; 1268 struct hio *hio; 1269 struct nv *nv; 1270 unsigned int ncomp; 1271 bool wakeup; 1272 uint64_t offset, length; 1273 uint8_t cmd; 1274 void *data; 1275 1276 /* Remote component is 1 for now. */ 1277 ncomp = 1; 1278 lastcheck = time(NULL); 1279 1280 for (;;) { 1281 pjdlog_debug(2, "remote_send: Taking request."); 1282 QUEUE_TAKE1(hio, send, ncomp, RETRY_SLEEP); 1283 if (hio == NULL) { 1284 now = time(NULL); 1285 if (lastcheck + RETRY_SLEEP <= now) { 1286 keepalive_send(res, ncomp); 1287 lastcheck = now; 1288 } 1289 continue; 1290 } 1291 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1292 ggio = &hio->hio_ggio; 1293 switch (ggio->gctl_cmd) { 1294 case BIO_READ: 1295 cmd = HIO_READ; 1296 data = NULL; 1297 offset = ggio->gctl_offset; 1298 length = ggio->gctl_length; 1299 break; 1300 case BIO_WRITE: 1301 cmd = HIO_WRITE; 1302 data = ggio->gctl_data; 1303 offset = ggio->gctl_offset; 1304 length = ggio->gctl_length; 1305 break; 1306 case BIO_DELETE: 1307 cmd = HIO_DELETE; 1308 data = NULL; 1309 offset = ggio->gctl_offset; 1310 length = ggio->gctl_length; 1311 break; 1312 case BIO_FLUSH: 1313 cmd = HIO_FLUSH; 1314 data = NULL; 1315 offset = 0; 1316 length = 0; 1317 break; 1318 default: 1319 PJDLOG_ASSERT(!"invalid condition"); 1320 abort(); 1321 } 1322 nv = nv_alloc(); 1323 nv_add_uint8(nv, cmd, "cmd"); 1324 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1325 nv_add_uint64(nv, offset, "offset"); 1326 nv_add_uint64(nv, length, "length"); 1327 if (nv_error(nv) != 0) { 1328 hio->hio_errors[ncomp] = nv_error(nv); 1329 pjdlog_debug(2, 1330 "remote_send: (%p) Unable to prepare header to send.", 1331 hio); 1332 reqlog(LOG_ERR, 0, ggio, 1333 "Unable to prepare header to send (%s): ", 1334 strerror(nv_error(nv))); 1335 /* Move failed request immediately to the done queue. */ 1336 goto done_queue; 1337 } 1338 pjdlog_debug(2, 1339 "remote_send: (%p) Moving request to the recv queue.", 1340 hio); 1341 /* 1342 * Protect connection from disappearing. 1343 */ 1344 rw_rlock(&hio_remote_lock[ncomp]); 1345 if (!ISCONNECTED(res, ncomp)) { 1346 rw_unlock(&hio_remote_lock[ncomp]); 1347 hio->hio_errors[ncomp] = ENOTCONN; 1348 goto done_queue; 1349 } 1350 /* 1351 * Move the request to recv queue before sending it, because 1352 * in different order we can get reply before we move request 1353 * to recv queue. 1354 */ 1355 mtx_lock(&hio_recv_list_lock[ncomp]); 1356 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1357 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1358 mtx_unlock(&hio_recv_list_lock[ncomp]); 1359 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1360 data != NULL ? length : 0) < 0) { 1361 hio->hio_errors[ncomp] = errno; 1362 rw_unlock(&hio_remote_lock[ncomp]); 1363 pjdlog_debug(2, 1364 "remote_send: (%p) Unable to send request.", hio); 1365 reqlog(LOG_ERR, 0, ggio, 1366 "Unable to send request (%s): ", 1367 strerror(hio->hio_errors[ncomp])); 1368 remote_close(res, ncomp); 1369 /* 1370 * Take request back from the receive queue and move 1371 * it immediately to the done queue. 1372 */ 1373 mtx_lock(&hio_recv_list_lock[ncomp]); 1374 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1375 mtx_unlock(&hio_recv_list_lock[ncomp]); 1376 goto done_queue; 1377 } 1378 rw_unlock(&hio_remote_lock[ncomp]); 1379 nv_free(nv); 1380 if (wakeup) 1381 cv_signal(&hio_recv_list_cond[ncomp]); 1382 continue; 1383done_queue: 1384 nv_free(nv); 1385 if (ISSYNCREQ(hio)) { 1386 if (!refcount_release(&hio->hio_countdown)) 1387 continue; 1388 mtx_lock(&sync_lock); 1389 SYNCREQDONE(hio); 1390 mtx_unlock(&sync_lock); 1391 cv_signal(&sync_cond); 1392 continue; 1393 } 1394 if (ggio->gctl_cmd == BIO_WRITE) { 1395 mtx_lock(&res->hr_amp_lock); 1396 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1397 ggio->gctl_length)) { 1398 (void)hast_activemap_flush(res); 1399 } 1400 mtx_unlock(&res->hr_amp_lock); 1401 } 1402 if (!refcount_release(&hio->hio_countdown)) 1403 continue; 1404 pjdlog_debug(2, 1405 "remote_send: (%p) Moving request to the done queue.", 1406 hio); 1407 QUEUE_INSERT2(hio, done); 1408 } 1409 /* NOTREACHED */ 1410 return (NULL); 1411} 1412 1413/* 1414 * Thread receives answer from secondary node and passes it to ggate_send 1415 * thread. 1416 */ 1417static void * 1418remote_recv_thread(void *arg) 1419{ 1420 struct hast_resource *res = arg; 1421 struct g_gate_ctl_io *ggio; 1422 struct hio *hio; 1423 struct nv *nv; 1424 unsigned int ncomp; 1425 uint64_t seq; 1426 int error; 1427 1428 /* Remote component is 1 for now. */ 1429 ncomp = 1; 1430 1431 for (;;) { 1432 /* Wait until there is anything to receive. */ 1433 mtx_lock(&hio_recv_list_lock[ncomp]); 1434 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1435 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1436 cv_wait(&hio_recv_list_cond[ncomp], 1437 &hio_recv_list_lock[ncomp]); 1438 } 1439 mtx_unlock(&hio_recv_list_lock[ncomp]); 1440 rw_rlock(&hio_remote_lock[ncomp]); 1441 if (!ISCONNECTED(res, ncomp)) { 1442 rw_unlock(&hio_remote_lock[ncomp]); 1443 /* 1444 * Connection is dead, so move all pending requests to 1445 * the done queue (one-by-one). 1446 */ 1447 mtx_lock(&hio_recv_list_lock[ncomp]); 1448 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1449 PJDLOG_ASSERT(hio != NULL); 1450 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1451 hio_next[ncomp]); 1452 mtx_unlock(&hio_recv_list_lock[ncomp]); 1453 goto done_queue; 1454 } 1455 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1456 pjdlog_errno(LOG_ERR, 1457 "Unable to receive reply header"); 1458 rw_unlock(&hio_remote_lock[ncomp]); 1459 remote_close(res, ncomp); 1460 continue; 1461 } 1462 rw_unlock(&hio_remote_lock[ncomp]); 1463 seq = nv_get_uint64(nv, "seq"); 1464 if (seq == 0) { 1465 pjdlog_error("Header contains no 'seq' field."); 1466 nv_free(nv); 1467 continue; 1468 } 1469 mtx_lock(&hio_recv_list_lock[ncomp]); 1470 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1471 if (hio->hio_ggio.gctl_seq == seq) { 1472 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1473 hio_next[ncomp]); 1474 break; 1475 } 1476 } 1477 mtx_unlock(&hio_recv_list_lock[ncomp]); 1478 if (hio == NULL) { 1479 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1480 (uintmax_t)seq); 1481 nv_free(nv); 1482 continue; 1483 } 1484 error = nv_get_int16(nv, "error"); 1485 if (error != 0) { 1486 /* Request failed on remote side. */ 1487 hio->hio_errors[ncomp] = error; 1488 reqlog(LOG_WARNING, 0, &hio->hio_ggio, 1489 "Remote request failed (%s): ", strerror(error)); 1490 nv_free(nv); 1491 goto done_queue; 1492 } 1493 ggio = &hio->hio_ggio; 1494 switch (ggio->gctl_cmd) { 1495 case BIO_READ: 1496 rw_rlock(&hio_remote_lock[ncomp]); 1497 if (!ISCONNECTED(res, ncomp)) { 1498 rw_unlock(&hio_remote_lock[ncomp]); 1499 nv_free(nv); 1500 goto done_queue; 1501 } 1502 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1503 ggio->gctl_data, ggio->gctl_length) < 0) { 1504 hio->hio_errors[ncomp] = errno; 1505 pjdlog_errno(LOG_ERR, 1506 "Unable to receive reply data"); 1507 rw_unlock(&hio_remote_lock[ncomp]); 1508 nv_free(nv); 1509 remote_close(res, ncomp); 1510 goto done_queue; 1511 } 1512 rw_unlock(&hio_remote_lock[ncomp]); 1513 break; 1514 case BIO_WRITE: 1515 case BIO_DELETE: 1516 case BIO_FLUSH: 1517 break; 1518 default: 1519 PJDLOG_ASSERT(!"invalid condition"); 1520 abort(); 1521 } 1522 hio->hio_errors[ncomp] = 0; 1523 nv_free(nv); 1524done_queue: 1525 if (refcount_release(&hio->hio_countdown)) { 1526 if (ISSYNCREQ(hio)) { 1527 mtx_lock(&sync_lock); 1528 SYNCREQDONE(hio); 1529 mtx_unlock(&sync_lock); 1530 cv_signal(&sync_cond); 1531 } else { 1532 pjdlog_debug(2, 1533 "remote_recv: (%p) Moving request to the done queue.", 1534 hio); 1535 QUEUE_INSERT2(hio, done); 1536 } 1537 } 1538 } 1539 /* NOTREACHED */ 1540 return (NULL); 1541} 1542 1543/* 1544 * Thread sends answer to the kernel. 1545 */ 1546static void * 1547ggate_send_thread(void *arg) 1548{ 1549 struct hast_resource *res = arg; 1550 struct g_gate_ctl_io *ggio; 1551 struct hio *hio; 1552 unsigned int ii, ncomp, ncomps; 1553 1554 ncomps = HAST_NCOMPONENTS; 1555 1556 for (;;) { 1557 pjdlog_debug(2, "ggate_send: Taking request."); 1558 QUEUE_TAKE2(hio, done); 1559 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1560 ggio = &hio->hio_ggio; 1561 for (ii = 0; ii < ncomps; ii++) { 1562 if (hio->hio_errors[ii] == 0) { 1563 /* 1564 * One successful request is enough to declare 1565 * success. 1566 */ 1567 ggio->gctl_error = 0; 1568 break; 1569 } 1570 } 1571 if (ii == ncomps) { 1572 /* 1573 * None of the requests were successful. 1574 * Use first error. 1575 */ 1576 ggio->gctl_error = hio->hio_errors[0]; 1577 } 1578 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1579 mtx_lock(&res->hr_amp_lock); 1580 activemap_write_complete(res->hr_amp, 1581 ggio->gctl_offset, ggio->gctl_length); 1582 mtx_unlock(&res->hr_amp_lock); 1583 } 1584 if (ggio->gctl_cmd == BIO_WRITE) { 1585 /* 1586 * Unlock range we locked. 1587 */ 1588 mtx_lock(&range_lock); 1589 rangelock_del(range_regular, ggio->gctl_offset, 1590 ggio->gctl_length); 1591 if (range_sync_wait) 1592 cv_signal(&range_sync_cond); 1593 mtx_unlock(&range_lock); 1594 /* 1595 * Bump local count if this is first write after 1596 * connection failure with remote node. 1597 */ 1598 ncomp = 1; 1599 rw_rlock(&hio_remote_lock[ncomp]); 1600 if (!ISCONNECTED(res, ncomp)) { 1601 mtx_lock(&metadata_lock); 1602 if (res->hr_primary_localcnt == 1603 res->hr_secondary_remotecnt) { 1604 res->hr_primary_localcnt++; 1605 pjdlog_debug(1, 1606 "Increasing localcnt to %ju.", 1607 (uintmax_t)res->hr_primary_localcnt); 1608 (void)metadata_write(res); 1609 } 1610 mtx_unlock(&metadata_lock); 1611 } 1612 rw_unlock(&hio_remote_lock[ncomp]); 1613 } 1614 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1615 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1616 pjdlog_debug(2, 1617 "ggate_send: (%p) Moving request to the free queue.", hio); 1618 QUEUE_INSERT2(hio, free); 1619 } 1620 /* NOTREACHED */ 1621 return (NULL); 1622} 1623 1624/* 1625 * Thread synchronize local and remote components. 1626 */ 1627static void * 1628sync_thread(void *arg __unused) 1629{ 1630 struct hast_resource *res = arg; 1631 struct hio *hio; 1632 struct g_gate_ctl_io *ggio; 1633 unsigned int ii, ncomp, ncomps; 1634 off_t offset, length, synced; 1635 bool dorewind; 1636 int syncext; 1637 1638 ncomps = HAST_NCOMPONENTS; 1639 dorewind = true; 1640 synced = 0; 1641 offset = -1; 1642 1643 for (;;) { 1644 mtx_lock(&sync_lock); 1645 if (offset >= 0 && !sync_inprogress) { 1646 pjdlog_info("Synchronization interrupted. " 1647 "%jd bytes synchronized so far.", 1648 (intmax_t)synced); 1649 event_send(res, EVENT_SYNCINTR); 1650 } 1651 while (!sync_inprogress) { 1652 dorewind = true; 1653 synced = 0; 1654 cv_wait(&sync_cond, &sync_lock); 1655 } 1656 mtx_unlock(&sync_lock); 1657 /* 1658 * Obtain offset at which we should synchronize. 1659 * Rewind synchronization if needed. 1660 */ 1661 mtx_lock(&res->hr_amp_lock); 1662 if (dorewind) 1663 activemap_sync_rewind(res->hr_amp); 1664 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1665 if (syncext != -1) { 1666 /* 1667 * We synchronized entire syncext extent, we can mark 1668 * it as clean now. 1669 */ 1670 if (activemap_extent_complete(res->hr_amp, syncext)) 1671 (void)hast_activemap_flush(res); 1672 } 1673 mtx_unlock(&res->hr_amp_lock); 1674 if (dorewind) { 1675 dorewind = false; 1676 if (offset < 0) 1677 pjdlog_info("Nodes are in sync."); 1678 else { 1679 pjdlog_info("Synchronization started. %ju bytes to go.", 1680 (uintmax_t)(res->hr_extentsize * 1681 activemap_ndirty(res->hr_amp))); 1682 event_send(res, EVENT_SYNCSTART); 1683 } 1684 } 1685 if (offset < 0) { 1686 sync_stop(); 1687 pjdlog_debug(1, "Nothing to synchronize."); 1688 /* 1689 * Synchronization complete, make both localcnt and 1690 * remotecnt equal. 1691 */ 1692 ncomp = 1; 1693 rw_rlock(&hio_remote_lock[ncomp]); 1694 if (ISCONNECTED(res, ncomp)) { 1695 if (synced > 0) { 1696 pjdlog_info("Synchronization complete. " 1697 "%jd bytes synchronized.", 1698 (intmax_t)synced); 1699 event_send(res, EVENT_SYNCDONE); 1700 } 1701 mtx_lock(&metadata_lock); 1702 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1703 res->hr_primary_localcnt = 1704 res->hr_secondary_localcnt; 1705 res->hr_primary_remotecnt = 1706 res->hr_secondary_remotecnt; 1707 pjdlog_debug(1, 1708 "Setting localcnt to %ju and remotecnt to %ju.", 1709 (uintmax_t)res->hr_primary_localcnt, 1710 (uintmax_t)res->hr_secondary_localcnt); 1711 (void)metadata_write(res); 1712 mtx_unlock(&metadata_lock); 1713 } 1714 rw_unlock(&hio_remote_lock[ncomp]); 1715 continue; 1716 } 1717 pjdlog_debug(2, "sync: Taking free request."); 1718 QUEUE_TAKE2(hio, free); 1719 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1720 /* 1721 * Lock the range we are going to synchronize. We don't want 1722 * race where someone writes between our read and write. 1723 */ 1724 for (;;) { 1725 mtx_lock(&range_lock); 1726 if (rangelock_islocked(range_regular, offset, length)) { 1727 pjdlog_debug(2, 1728 "sync: Range offset=%jd length=%jd locked.", 1729 (intmax_t)offset, (intmax_t)length); 1730 range_sync_wait = true; 1731 cv_wait(&range_sync_cond, &range_lock); 1732 range_sync_wait = false; 1733 mtx_unlock(&range_lock); 1734 continue; 1735 } 1736 if (rangelock_add(range_sync, offset, length) < 0) { 1737 mtx_unlock(&range_lock); 1738 pjdlog_debug(2, 1739 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1740 (intmax_t)offset, (intmax_t)length); 1741 sleep(1); 1742 continue; 1743 } 1744 mtx_unlock(&range_lock); 1745 break; 1746 } 1747 /* 1748 * First read the data from synchronization source. 1749 */ 1750 SYNCREQ(hio); 1751 ggio = &hio->hio_ggio; 1752 ggio->gctl_cmd = BIO_READ; 1753 ggio->gctl_offset = offset; 1754 ggio->gctl_length = length; 1755 ggio->gctl_error = 0; 1756 for (ii = 0; ii < ncomps; ii++) 1757 hio->hio_errors[ii] = EINVAL; 1758 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1759 hio); 1760 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1761 hio); 1762 mtx_lock(&metadata_lock); 1763 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1764 /* 1765 * This range is up-to-date on local component, 1766 * so handle request locally. 1767 */ 1768 /* Local component is 0 for now. */ 1769 ncomp = 0; 1770 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1771 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1772 /* 1773 * This range is out-of-date on local component, 1774 * so send request to the remote node. 1775 */ 1776 /* Remote component is 1 for now. */ 1777 ncomp = 1; 1778 } 1779 mtx_unlock(&metadata_lock); 1780 refcount_init(&hio->hio_countdown, 1); 1781 QUEUE_INSERT1(hio, send, ncomp); 1782 1783 /* 1784 * Let's wait for READ to finish. 1785 */ 1786 mtx_lock(&sync_lock); 1787 while (!ISSYNCREQDONE(hio)) 1788 cv_wait(&sync_cond, &sync_lock); 1789 mtx_unlock(&sync_lock); 1790 1791 if (hio->hio_errors[ncomp] != 0) { 1792 pjdlog_error("Unable to read synchronization data: %s.", 1793 strerror(hio->hio_errors[ncomp])); 1794 goto free_queue; 1795 } 1796 1797 /* 1798 * We read the data from synchronization source, now write it 1799 * to synchronization target. 1800 */ 1801 SYNCREQ(hio); 1802 ggio->gctl_cmd = BIO_WRITE; 1803 for (ii = 0; ii < ncomps; ii++) 1804 hio->hio_errors[ii] = EINVAL; 1805 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1806 hio); 1807 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1808 hio); 1809 mtx_lock(&metadata_lock); 1810 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1811 /* 1812 * This range is up-to-date on local component, 1813 * so we update remote component. 1814 */ 1815 /* Remote component is 1 for now. */ 1816 ncomp = 1; 1817 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1818 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1819 /* 1820 * This range is out-of-date on local component, 1821 * so we update it. 1822 */ 1823 /* Local component is 0 for now. */ 1824 ncomp = 0; 1825 } 1826 mtx_unlock(&metadata_lock); 1827 1828 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1829 hio); 1830 refcount_init(&hio->hio_countdown, 1); 1831 QUEUE_INSERT1(hio, send, ncomp); 1832 1833 /* 1834 * Let's wait for WRITE to finish. 1835 */ 1836 mtx_lock(&sync_lock); 1837 while (!ISSYNCREQDONE(hio)) 1838 cv_wait(&sync_cond, &sync_lock); 1839 mtx_unlock(&sync_lock); 1840 1841 if (hio->hio_errors[ncomp] != 0) { 1842 pjdlog_error("Unable to write synchronization data: %s.", 1843 strerror(hio->hio_errors[ncomp])); 1844 goto free_queue; 1845 } 1846 1847 synced += length; 1848free_queue: 1849 mtx_lock(&range_lock); 1850 rangelock_del(range_sync, offset, length); 1851 if (range_regular_wait) 1852 cv_signal(&range_regular_cond); 1853 mtx_unlock(&range_lock); 1854 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1855 hio); 1856 QUEUE_INSERT2(hio, free); 1857 } 1858 /* NOTREACHED */ 1859 return (NULL); 1860} 1861 1862void 1863primary_config_reload(struct hast_resource *res, struct nv *nv) 1864{ 1865 unsigned int ii, ncomps; 1866 int modified, vint; 1867 const char *vstr; 1868 1869 pjdlog_info("Reloading configuration..."); 1870 1871 PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY); 1872 PJDLOG_ASSERT(gres == res); 1873 nv_assert(nv, "remoteaddr"); 1874 nv_assert(nv, "replication"); 1875 nv_assert(nv, "timeout"); 1876 nv_assert(nv, "exec"); 1877 1878 ncomps = HAST_NCOMPONENTS; 1879 1880#define MODIFIED_REMOTEADDR 0x1 1881#define MODIFIED_REPLICATION 0x2 1882#define MODIFIED_TIMEOUT 0x4 1883#define MODIFIED_EXEC 0x8 1884 modified = 0; 1885 1886 vstr = nv_get_string(nv, "remoteaddr"); 1887 if (strcmp(gres->hr_remoteaddr, vstr) != 0) { 1888 /* 1889 * Don't copy res->hr_remoteaddr to gres just yet. 1890 * We want remote_close() to log disconnect from the old 1891 * addresses, not from the new ones. 1892 */ 1893 modified |= MODIFIED_REMOTEADDR; 1894 } 1895 vint = nv_get_int32(nv, "replication"); 1896 if (gres->hr_replication != vint) { 1897 gres->hr_replication = vint; 1898 modified |= MODIFIED_REPLICATION; 1899 } 1900 vint = nv_get_int32(nv, "timeout"); 1901 if (gres->hr_timeout != vint) { 1902 gres->hr_timeout = vint; 1903 modified |= MODIFIED_TIMEOUT; 1904 } 1905 vstr = nv_get_string(nv, "exec"); 1906 if (strcmp(gres->hr_exec, vstr) != 0) { 1907 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec)); 1908 modified |= MODIFIED_EXEC; 1909 } 1910 1911 /* 1912 * If only timeout was modified we only need to change it without 1913 * reconnecting. 1914 */ 1915 if (modified == MODIFIED_TIMEOUT) { 1916 for (ii = 0; ii < ncomps; ii++) { 1917 if (!ISREMOTE(ii)) 1918 continue; 1919 rw_rlock(&hio_remote_lock[ii]); 1920 if (!ISCONNECTED(gres, ii)) { 1921 rw_unlock(&hio_remote_lock[ii]); 1922 continue; 1923 } 1924 rw_unlock(&hio_remote_lock[ii]); 1925 if (proto_timeout(gres->hr_remotein, 1926 gres->hr_timeout) < 0) { 1927 pjdlog_errno(LOG_WARNING, 1928 "Unable to set connection timeout"); 1929 } 1930 if (proto_timeout(gres->hr_remoteout, 1931 gres->hr_timeout) < 0) { 1932 pjdlog_errno(LOG_WARNING, 1933 "Unable to set connection timeout"); 1934 } 1935 } 1936 } else if ((modified & 1937 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) { 1938 for (ii = 0; ii < ncomps; ii++) { 1939 if (!ISREMOTE(ii)) 1940 continue; 1941 remote_close(gres, ii); 1942 } 1943 if (modified & MODIFIED_REMOTEADDR) { 1944 vstr = nv_get_string(nv, "remoteaddr"); 1945 strlcpy(gres->hr_remoteaddr, vstr, 1946 sizeof(gres->hr_remoteaddr)); 1947 } 1948 } 1949#undef MODIFIED_REMOTEADDR 1950#undef MODIFIED_REPLICATION 1951#undef MODIFIED_TIMEOUT 1952#undef MODIFIED_EXEC 1953 1954 pjdlog_info("Configuration reloaded successfully."); 1955} 1956 1957static void 1958guard_one(struct hast_resource *res, unsigned int ncomp) 1959{ 1960 struct proto_conn *in, *out; 1961 1962 if (!ISREMOTE(ncomp)) 1963 return; 1964 1965 rw_rlock(&hio_remote_lock[ncomp]); 1966 1967 if (!real_remote(res)) { 1968 rw_unlock(&hio_remote_lock[ncomp]); 1969 return; 1970 } 1971 1972 if (ISCONNECTED(res, ncomp)) { 1973 PJDLOG_ASSERT(res->hr_remotein != NULL); 1974 PJDLOG_ASSERT(res->hr_remoteout != NULL); 1975 rw_unlock(&hio_remote_lock[ncomp]); 1976 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 1977 res->hr_remoteaddr); 1978 return; 1979 } 1980 1981 PJDLOG_ASSERT(res->hr_remotein == NULL); 1982 PJDLOG_ASSERT(res->hr_remoteout == NULL); 1983 /* 1984 * Upgrade the lock. It doesn't have to be atomic as no other thread 1985 * can change connection status from disconnected to connected. 1986 */ 1987 rw_unlock(&hio_remote_lock[ncomp]); 1988 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 1989 res->hr_remoteaddr); 1990 in = out = NULL; 1991 if (init_remote(res, &in, &out)) { 1992 rw_wlock(&hio_remote_lock[ncomp]); 1993 PJDLOG_ASSERT(res->hr_remotein == NULL); 1994 PJDLOG_ASSERT(res->hr_remoteout == NULL); 1995 PJDLOG_ASSERT(in != NULL && out != NULL); 1996 res->hr_remotein = in; 1997 res->hr_remoteout = out; 1998 rw_unlock(&hio_remote_lock[ncomp]); 1999 pjdlog_info("Successfully reconnected to %s.", 2000 res->hr_remoteaddr); 2001 sync_start(); 2002 } else { 2003 /* Both connections should be NULL. */ 2004 PJDLOG_ASSERT(res->hr_remotein == NULL); 2005 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2006 PJDLOG_ASSERT(in == NULL && out == NULL); 2007 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 2008 res->hr_remoteaddr); 2009 } 2010} 2011 2012/* 2013 * Thread guards remote connections and reconnects when needed, handles 2014 * signals, etc. 2015 */ 2016static void * 2017guard_thread(void *arg) 2018{ 2019 struct hast_resource *res = arg; 2020 unsigned int ii, ncomps; 2021 struct timespec timeout; 2022 time_t lastcheck, now; 2023 sigset_t mask; 2024 int signo; 2025 2026 ncomps = HAST_NCOMPONENTS; 2027 lastcheck = time(NULL); 2028 2029 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 2030 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 2031 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 2032 2033 timeout.tv_sec = RETRY_SLEEP; 2034 timeout.tv_nsec = 0; 2035 signo = -1; 2036 2037 for (;;) { 2038 switch (signo) { 2039 case SIGINT: 2040 case SIGTERM: 2041 sigexit_received = true; 2042 primary_exitx(EX_OK, 2043 "Termination signal received, exiting."); 2044 break; 2045 default: 2046 break; 2047 } 2048 2049 pjdlog_debug(2, "remote_guard: Checking connections."); 2050 now = time(NULL); 2051 if (lastcheck + RETRY_SLEEP <= now) { 2052 for (ii = 0; ii < ncomps; ii++) 2053 guard_one(res, ii); 2054 lastcheck = now; 2055 } 2056 signo = sigtimedwait(&mask, NULL, &timeout); 2057 } 2058 /* NOTREACHED */ 2059 return (NULL); 2060} 2061