primary.c revision 220266
1/*- 2 * Copyright (c) 2009 The FreeBSD Foundation 3 * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net> 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 220266 2011-04-02 07:01:09Z 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 communicate 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#define ISCONNECTED(res, no) \ 155 ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL) 156 157#define QUEUE_INSERT1(hio, name, ncomp) do { \ 158 bool _wakeup; \ 159 \ 160 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 161 _wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]); \ 162 TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio), \ 163 hio_next[(ncomp)]); \ 164 mtx_unlock(&hio_##name##_list_lock[ncomp]); \ 165 if (_wakeup) \ 166 cv_signal(&hio_##name##_list_cond[(ncomp)]); \ 167} while (0) 168#define QUEUE_INSERT2(hio, name) do { \ 169 bool _wakeup; \ 170 \ 171 mtx_lock(&hio_##name##_list_lock); \ 172 _wakeup = TAILQ_EMPTY(&hio_##name##_list); \ 173 TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\ 174 mtx_unlock(&hio_##name##_list_lock); \ 175 if (_wakeup) \ 176 cv_signal(&hio_##name##_list_cond); \ 177} while (0) 178#define QUEUE_TAKE1(hio, name, ncomp, timeout) do { \ 179 bool _last; \ 180 \ 181 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \ 182 _last = false; \ 183 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \ 184 cv_timedwait(&hio_##name##_list_cond[(ncomp)], \ 185 &hio_##name##_list_lock[(ncomp)], (timeout)); \ 186 if ((timeout) != 0) \ 187 _last = true; \ 188 } \ 189 if (hio != NULL) { \ 190 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \ 191 hio_next[(ncomp)]); \ 192 } \ 193 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \ 194} while (0) 195#define QUEUE_TAKE2(hio, name) do { \ 196 mtx_lock(&hio_##name##_list_lock); \ 197 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 198 cv_wait(&hio_##name##_list_cond, \ 199 &hio_##name##_list_lock); \ 200 } \ 201 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \ 202 mtx_unlock(&hio_##name##_list_lock); \ 203} while (0) 204 205#define SYNCREQ(hio) do { \ 206 (hio)->hio_ggio.gctl_unit = -1; \ 207 (hio)->hio_ggio.gctl_seq = 1; \ 208} while (0) 209#define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1) 210#define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0) 211#define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2) 212 213static struct hast_resource *gres; 214 215static pthread_mutex_t range_lock; 216static struct rangelocks *range_regular; 217static bool range_regular_wait; 218static pthread_cond_t range_regular_cond; 219static struct rangelocks *range_sync; 220static bool range_sync_wait; 221static pthread_cond_t range_sync_cond; 222 223static void *ggate_recv_thread(void *arg); 224static void *local_send_thread(void *arg); 225static void *remote_send_thread(void *arg); 226static void *remote_recv_thread(void *arg); 227static void *ggate_send_thread(void *arg); 228static void *sync_thread(void *arg); 229static void *guard_thread(void *arg); 230 231static void 232cleanup(struct hast_resource *res) 233{ 234 int rerrno; 235 236 /* Remember errno. */ 237 rerrno = errno; 238 239 /* Destroy ggate provider if we created one. */ 240 if (res->hr_ggateunit >= 0) { 241 struct g_gate_ctl_destroy ggiod; 242 243 bzero(&ggiod, sizeof(ggiod)); 244 ggiod.gctl_version = G_GATE_VERSION; 245 ggiod.gctl_unit = res->hr_ggateunit; 246 ggiod.gctl_force = 1; 247 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) { 248 pjdlog_errno(LOG_WARNING, 249 "Unable to destroy hast/%s device", 250 res->hr_provname); 251 } 252 res->hr_ggateunit = -1; 253 } 254 255 /* Restore errno. */ 256 errno = rerrno; 257} 258 259static __dead2 void 260primary_exit(int exitcode, const char *fmt, ...) 261{ 262 va_list ap; 263 264 PJDLOG_ASSERT(exitcode != EX_OK); 265 va_start(ap, fmt); 266 pjdlogv_errno(LOG_ERR, fmt, ap); 267 va_end(ap); 268 cleanup(gres); 269 exit(exitcode); 270} 271 272static __dead2 void 273primary_exitx(int exitcode, const char *fmt, ...) 274{ 275 va_list ap; 276 277 va_start(ap, fmt); 278 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap); 279 va_end(ap); 280 cleanup(gres); 281 exit(exitcode); 282} 283 284static int 285hast_activemap_flush(struct hast_resource *res) 286{ 287 const unsigned char *buf; 288 size_t size; 289 290 buf = activemap_bitmap(res->hr_amp, &size); 291 PJDLOG_ASSERT(buf != NULL); 292 PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0); 293 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) != 294 (ssize_t)size) { 295 KEEP_ERRNO(pjdlog_errno(LOG_ERR, 296 "Unable to flush activemap to disk")); 297 return (-1); 298 } 299 return (0); 300} 301 302static bool 303real_remote(const struct hast_resource *res) 304{ 305 306 return (strcmp(res->hr_remoteaddr, "none") != 0); 307} 308 309static void 310init_environment(struct hast_resource *res __unused) 311{ 312 struct hio *hio; 313 unsigned int ii, ncomps; 314 315 /* 316 * In the future it might be per-resource value. 317 */ 318 ncomps = HAST_NCOMPONENTS; 319 320 /* 321 * Allocate memory needed by lists. 322 */ 323 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps); 324 if (hio_send_list == NULL) { 325 primary_exitx(EX_TEMPFAIL, 326 "Unable to allocate %zu bytes of memory for send lists.", 327 sizeof(hio_send_list[0]) * ncomps); 328 } 329 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps); 330 if (hio_send_list_lock == NULL) { 331 primary_exitx(EX_TEMPFAIL, 332 "Unable to allocate %zu bytes of memory for send list locks.", 333 sizeof(hio_send_list_lock[0]) * ncomps); 334 } 335 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps); 336 if (hio_send_list_cond == NULL) { 337 primary_exitx(EX_TEMPFAIL, 338 "Unable to allocate %zu bytes of memory for send list condition variables.", 339 sizeof(hio_send_list_cond[0]) * ncomps); 340 } 341 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps); 342 if (hio_recv_list == NULL) { 343 primary_exitx(EX_TEMPFAIL, 344 "Unable to allocate %zu bytes of memory for recv lists.", 345 sizeof(hio_recv_list[0]) * ncomps); 346 } 347 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps); 348 if (hio_recv_list_lock == NULL) { 349 primary_exitx(EX_TEMPFAIL, 350 "Unable to allocate %zu bytes of memory for recv list locks.", 351 sizeof(hio_recv_list_lock[0]) * ncomps); 352 } 353 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps); 354 if (hio_recv_list_cond == NULL) { 355 primary_exitx(EX_TEMPFAIL, 356 "Unable to allocate %zu bytes of memory for recv list condition variables.", 357 sizeof(hio_recv_list_cond[0]) * ncomps); 358 } 359 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps); 360 if (hio_remote_lock == NULL) { 361 primary_exitx(EX_TEMPFAIL, 362 "Unable to allocate %zu bytes of memory for remote connections locks.", 363 sizeof(hio_remote_lock[0]) * ncomps); 364 } 365 366 /* 367 * Initialize lists, their locks and theirs condition variables. 368 */ 369 TAILQ_INIT(&hio_free_list); 370 mtx_init(&hio_free_list_lock); 371 cv_init(&hio_free_list_cond); 372 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) { 373 TAILQ_INIT(&hio_send_list[ii]); 374 mtx_init(&hio_send_list_lock[ii]); 375 cv_init(&hio_send_list_cond[ii]); 376 TAILQ_INIT(&hio_recv_list[ii]); 377 mtx_init(&hio_recv_list_lock[ii]); 378 cv_init(&hio_recv_list_cond[ii]); 379 rw_init(&hio_remote_lock[ii]); 380 } 381 TAILQ_INIT(&hio_done_list); 382 mtx_init(&hio_done_list_lock); 383 cv_init(&hio_done_list_cond); 384 mtx_init(&metadata_lock); 385 386 /* 387 * Allocate requests pool and initialize requests. 388 */ 389 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 390 hio = malloc(sizeof(*hio)); 391 if (hio == NULL) { 392 primary_exitx(EX_TEMPFAIL, 393 "Unable to allocate %zu bytes of memory for hio request.", 394 sizeof(*hio)); 395 } 396 hio->hio_countdown = 0; 397 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps); 398 if (hio->hio_errors == NULL) { 399 primary_exitx(EX_TEMPFAIL, 400 "Unable allocate %zu bytes of memory for hio errors.", 401 sizeof(hio->hio_errors[0]) * ncomps); 402 } 403 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps); 404 if (hio->hio_next == NULL) { 405 primary_exitx(EX_TEMPFAIL, 406 "Unable allocate %zu bytes of memory for hio_next field.", 407 sizeof(hio->hio_next[0]) * ncomps); 408 } 409 hio->hio_ggio.gctl_version = G_GATE_VERSION; 410 hio->hio_ggio.gctl_data = malloc(MAXPHYS); 411 if (hio->hio_ggio.gctl_data == NULL) { 412 primary_exitx(EX_TEMPFAIL, 413 "Unable to allocate %zu bytes of memory for gctl_data.", 414 MAXPHYS); 415 } 416 hio->hio_ggio.gctl_length = MAXPHYS; 417 hio->hio_ggio.gctl_error = 0; 418 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next); 419 } 420} 421 422static bool 423init_resuid(struct hast_resource *res) 424{ 425 426 mtx_lock(&metadata_lock); 427 if (res->hr_resuid != 0) { 428 mtx_unlock(&metadata_lock); 429 return (false); 430 } else { 431 /* Initialize unique resource identifier. */ 432 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid)); 433 mtx_unlock(&metadata_lock); 434 if (metadata_write(res) < 0) 435 exit(EX_NOINPUT); 436 return (true); 437 } 438} 439 440static void 441init_local(struct hast_resource *res) 442{ 443 unsigned char *buf; 444 size_t mapsize; 445 446 if (metadata_read(res, true) < 0) 447 exit(EX_NOINPUT); 448 mtx_init(&res->hr_amp_lock); 449 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize, 450 res->hr_local_sectorsize, res->hr_keepdirty) < 0) { 451 primary_exit(EX_TEMPFAIL, "Unable to create activemap"); 452 } 453 mtx_init(&range_lock); 454 cv_init(&range_regular_cond); 455 if (rangelock_init(&range_regular) < 0) 456 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock"); 457 cv_init(&range_sync_cond); 458 if (rangelock_init(&range_sync) < 0) 459 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock"); 460 mapsize = activemap_ondisk_size(res->hr_amp); 461 buf = calloc(1, mapsize); 462 if (buf == NULL) { 463 primary_exitx(EX_TEMPFAIL, 464 "Unable to allocate buffer for activemap."); 465 } 466 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) != 467 (ssize_t)mapsize) { 468 primary_exit(EX_NOINPUT, "Unable to read activemap"); 469 } 470 activemap_copyin(res->hr_amp, buf, mapsize); 471 free(buf); 472 if (res->hr_resuid != 0) 473 return; 474 /* 475 * We're using provider for the first time. Initialize local and remote 476 * counters. We don't initialize resuid here, as we want to do it just 477 * in time. The reason for this is that we want to inform secondary 478 * that there were no writes yet, so there is no need to synchronize 479 * anything. 480 */ 481 res->hr_primary_localcnt = 0; 482 res->hr_primary_remotecnt = 0; 483 if (metadata_write(res) < 0) 484 exit(EX_NOINPUT); 485} 486 487static int 488primary_connect(struct hast_resource *res, struct proto_conn **connp) 489{ 490 struct proto_conn *conn; 491 int16_t val; 492 493 val = 1; 494 if (proto_send(res->hr_conn, &val, sizeof(val)) < 0) { 495 primary_exit(EX_TEMPFAIL, 496 "Unable to send connection request to parent"); 497 } 498 if (proto_recv(res->hr_conn, &val, sizeof(val)) < 0) { 499 primary_exit(EX_TEMPFAIL, 500 "Unable to receive reply to connection request from parent"); 501 } 502 if (val != 0) { 503 errno = val; 504 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 505 res->hr_remoteaddr); 506 return (-1); 507 } 508 if (proto_connection_recv(res->hr_conn, true, &conn) < 0) { 509 primary_exit(EX_TEMPFAIL, 510 "Unable to receive connection from parent"); 511 } 512 if (proto_connect_wait(conn, res->hr_timeout) < 0) { 513 pjdlog_errno(LOG_WARNING, "Unable to connect to %s", 514 res->hr_remoteaddr); 515 proto_close(conn); 516 return (-1); 517 } 518 /* Error in setting timeout is not critical, but why should it fail? */ 519 if (proto_timeout(conn, res->hr_timeout) < 0) 520 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 521 522 *connp = conn; 523 524 return (0); 525} 526 527static bool 528init_remote(struct hast_resource *res, struct proto_conn **inp, 529 struct proto_conn **outp) 530{ 531 struct proto_conn *in, *out; 532 struct nv *nvout, *nvin; 533 const unsigned char *token; 534 unsigned char *map; 535 const char *errmsg; 536 int32_t extentsize; 537 int64_t datasize; 538 uint32_t mapsize; 539 size_t size; 540 541 PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL)); 542 PJDLOG_ASSERT(real_remote(res)); 543 544 in = out = NULL; 545 errmsg = NULL; 546 547 if (primary_connect(res, &out) == -1) 548 return (false); 549 550 /* 551 * First handshake step. 552 * Setup outgoing connection with remote node. 553 */ 554 nvout = nv_alloc(); 555 nv_add_string(nvout, res->hr_name, "resource"); 556 if (nv_error(nvout) != 0) { 557 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 558 "Unable to allocate header for connection with %s", 559 res->hr_remoteaddr); 560 nv_free(nvout); 561 goto close; 562 } 563 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) { 564 pjdlog_errno(LOG_WARNING, 565 "Unable to send handshake header to %s", 566 res->hr_remoteaddr); 567 nv_free(nvout); 568 goto close; 569 } 570 nv_free(nvout); 571 if (hast_proto_recv_hdr(out, &nvin) < 0) { 572 pjdlog_errno(LOG_WARNING, 573 "Unable to receive handshake header from %s", 574 res->hr_remoteaddr); 575 goto close; 576 } 577 errmsg = nv_get_string(nvin, "errmsg"); 578 if (errmsg != NULL) { 579 pjdlog_warning("%s", errmsg); 580 nv_free(nvin); 581 goto close; 582 } 583 token = nv_get_uint8_array(nvin, &size, "token"); 584 if (token == NULL) { 585 pjdlog_warning("Handshake header from %s has no 'token' field.", 586 res->hr_remoteaddr); 587 nv_free(nvin); 588 goto close; 589 } 590 if (size != sizeof(res->hr_token)) { 591 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).", 592 res->hr_remoteaddr, size, sizeof(res->hr_token)); 593 nv_free(nvin); 594 goto close; 595 } 596 bcopy(token, res->hr_token, sizeof(res->hr_token)); 597 nv_free(nvin); 598 599 /* 600 * Second handshake step. 601 * Setup incoming connection with remote node. 602 */ 603 if (primary_connect(res, &in) == -1) 604 goto close; 605 606 nvout = nv_alloc(); 607 nv_add_string(nvout, res->hr_name, "resource"); 608 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token), 609 "token"); 610 if (res->hr_resuid == 0) { 611 /* 612 * The resuid field was not yet initialized. 613 * Because we do synchronization inside init_resuid(), it is 614 * possible that someone already initialized it, the function 615 * will return false then, but if we successfully initialized 616 * it, we will get true. True means that there were no writes 617 * to this resource yet and we want to inform secondary that 618 * synchronization is not needed by sending "virgin" argument. 619 */ 620 if (init_resuid(res)) 621 nv_add_int8(nvout, 1, "virgin"); 622 } 623 nv_add_uint64(nvout, res->hr_resuid, "resuid"); 624 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt"); 625 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt"); 626 if (nv_error(nvout) != 0) { 627 pjdlog_common(LOG_WARNING, 0, nv_error(nvout), 628 "Unable to allocate header for connection with %s", 629 res->hr_remoteaddr); 630 nv_free(nvout); 631 goto close; 632 } 633 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) { 634 pjdlog_errno(LOG_WARNING, 635 "Unable to send handshake header to %s", 636 res->hr_remoteaddr); 637 nv_free(nvout); 638 goto close; 639 } 640 nv_free(nvout); 641 if (hast_proto_recv_hdr(out, &nvin) < 0) { 642 pjdlog_errno(LOG_WARNING, 643 "Unable to receive handshake header from %s", 644 res->hr_remoteaddr); 645 goto close; 646 } 647 errmsg = nv_get_string(nvin, "errmsg"); 648 if (errmsg != NULL) { 649 pjdlog_warning("%s", errmsg); 650 nv_free(nvin); 651 goto close; 652 } 653 datasize = nv_get_int64(nvin, "datasize"); 654 if (datasize != res->hr_datasize) { 655 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).", 656 (intmax_t)res->hr_datasize, (intmax_t)datasize); 657 nv_free(nvin); 658 goto close; 659 } 660 extentsize = nv_get_int32(nvin, "extentsize"); 661 if (extentsize != res->hr_extentsize) { 662 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).", 663 (ssize_t)res->hr_extentsize, (ssize_t)extentsize); 664 nv_free(nvin); 665 goto close; 666 } 667 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt"); 668 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 669 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc"); 670 map = NULL; 671 mapsize = nv_get_uint32(nvin, "mapsize"); 672 if (mapsize > 0) { 673 map = malloc(mapsize); 674 if (map == NULL) { 675 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).", 676 (uintmax_t)mapsize); 677 nv_free(nvin); 678 goto close; 679 } 680 /* 681 * Remote node have some dirty extents on its own, lets 682 * download its activemap. 683 */ 684 if (hast_proto_recv_data(res, out, nvin, map, 685 mapsize) < 0) { 686 pjdlog_errno(LOG_ERR, 687 "Unable to receive remote activemap"); 688 nv_free(nvin); 689 free(map); 690 goto close; 691 } 692 /* 693 * Merge local and remote bitmaps. 694 */ 695 activemap_merge(res->hr_amp, map, mapsize); 696 free(map); 697 /* 698 * Now that we merged bitmaps from both nodes, flush it to the 699 * disk before we start to synchronize. 700 */ 701 (void)hast_activemap_flush(res); 702 } 703 nv_free(nvin); 704 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 705 if (inp != NULL && outp != NULL) { 706 *inp = in; 707 *outp = out; 708 } else { 709 res->hr_remotein = in; 710 res->hr_remoteout = out; 711 } 712 event_send(res, EVENT_CONNECT); 713 return (true); 714close: 715 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 716 event_send(res, EVENT_SPLITBRAIN); 717 proto_close(out); 718 if (in != NULL) 719 proto_close(in); 720 return (false); 721} 722 723static void 724sync_start(void) 725{ 726 727 mtx_lock(&sync_lock); 728 sync_inprogress = true; 729 mtx_unlock(&sync_lock); 730 cv_signal(&sync_cond); 731} 732 733static void 734sync_stop(void) 735{ 736 737 mtx_lock(&sync_lock); 738 if (sync_inprogress) 739 sync_inprogress = false; 740 mtx_unlock(&sync_lock); 741} 742 743static void 744init_ggate(struct hast_resource *res) 745{ 746 struct g_gate_ctl_create ggiocreate; 747 struct g_gate_ctl_cancel ggiocancel; 748 749 /* 750 * We communicate with ggate via /dev/ggctl. Open it. 751 */ 752 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 753 if (res->hr_ggatefd < 0) 754 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 755 /* 756 * Create provider before trying to connect, as connection failure 757 * is not critical, but may take some time. 758 */ 759 bzero(&ggiocreate, sizeof(ggiocreate)); 760 ggiocreate.gctl_version = G_GATE_VERSION; 761 ggiocreate.gctl_mediasize = res->hr_datasize; 762 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 763 ggiocreate.gctl_flags = 0; 764 ggiocreate.gctl_maxcount = 0; 765 ggiocreate.gctl_timeout = 0; 766 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 767 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 768 res->hr_provname); 769 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 770 pjdlog_info("Device hast/%s created.", res->hr_provname); 771 res->hr_ggateunit = ggiocreate.gctl_unit; 772 return; 773 } 774 if (errno != EEXIST) { 775 primary_exit(EX_OSERR, "Unable to create hast/%s device", 776 res->hr_provname); 777 } 778 pjdlog_debug(1, 779 "Device hast/%s already exists, we will try to take it over.", 780 res->hr_provname); 781 /* 782 * If we received EEXIST, we assume that the process who created the 783 * provider died and didn't clean up. In that case we will start from 784 * where he left of. 785 */ 786 bzero(&ggiocancel, sizeof(ggiocancel)); 787 ggiocancel.gctl_version = G_GATE_VERSION; 788 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 789 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 790 res->hr_provname); 791 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 792 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 793 res->hr_ggateunit = ggiocancel.gctl_unit; 794 return; 795 } 796 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 797 res->hr_provname); 798} 799 800void 801hastd_primary(struct hast_resource *res) 802{ 803 pthread_t td; 804 pid_t pid; 805 int error, mode, debuglevel; 806 807 /* 808 * Create communication channel for sending control commands from 809 * parent to child. 810 */ 811 if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) { 812 /* TODO: There's no need for this to be fatal error. */ 813 KEEP_ERRNO((void)pidfile_remove(pfh)); 814 pjdlog_exit(EX_OSERR, 815 "Unable to create control sockets between parent and child"); 816 } 817 /* 818 * Create communication channel for sending events from child to parent. 819 */ 820 if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) { 821 /* TODO: There's no need for this to be fatal error. */ 822 KEEP_ERRNO((void)pidfile_remove(pfh)); 823 pjdlog_exit(EX_OSERR, 824 "Unable to create event sockets between child and parent"); 825 } 826 /* 827 * Create communication channel for sending connection requests from 828 * child to parent. 829 */ 830 if (proto_client(NULL, "socketpair://", &res->hr_conn) < 0) { 831 /* TODO: There's no need for this to be fatal error. */ 832 KEEP_ERRNO((void)pidfile_remove(pfh)); 833 pjdlog_exit(EX_OSERR, 834 "Unable to create connection sockets between child and parent"); 835 } 836 837 pid = fork(); 838 if (pid < 0) { 839 /* TODO: There's no need for this to be fatal error. */ 840 KEEP_ERRNO((void)pidfile_remove(pfh)); 841 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 842 } 843 844 if (pid > 0) { 845 /* This is parent. */ 846 /* Declare that we are receiver. */ 847 proto_recv(res->hr_event, NULL, 0); 848 proto_recv(res->hr_conn, NULL, 0); 849 /* Declare that we are sender. */ 850 proto_send(res->hr_ctrl, NULL, 0); 851 res->hr_workerpid = pid; 852 return; 853 } 854 855 gres = res; 856 mode = pjdlog_mode_get(); 857 debuglevel = pjdlog_debug_get(); 858 859 /* Declare that we are sender. */ 860 proto_send(res->hr_event, NULL, 0); 861 proto_send(res->hr_conn, NULL, 0); 862 /* Declare that we are receiver. */ 863 proto_recv(res->hr_ctrl, NULL, 0); 864 descriptors_cleanup(res); 865 866 descriptors_assert(res, mode); 867 868 pjdlog_init(mode); 869 pjdlog_debug_set(debuglevel); 870 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 871 setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role)); 872 873 init_local(res); 874 init_ggate(res); 875 init_environment(res); 876 877 if (drop_privs(true) != 0) { 878 cleanup(res); 879 exit(EX_CONFIG); 880 } 881 pjdlog_info("Privileges successfully dropped."); 882 883 /* 884 * Create the guard thread first, so we can handle signals from the 885 * very begining. 886 */ 887 error = pthread_create(&td, NULL, guard_thread, res); 888 PJDLOG_ASSERT(error == 0); 889 /* 890 * Create the control thread before sending any event to the parent, 891 * as we can deadlock when parent sends control request to worker, 892 * but worker has no control thread started yet, so parent waits. 893 * In the meantime worker sends an event to the parent, but parent 894 * is unable to handle the event, because it waits for control 895 * request response. 896 */ 897 error = pthread_create(&td, NULL, ctrl_thread, res); 898 PJDLOG_ASSERT(error == 0); 899 if (real_remote(res) && init_remote(res, NULL, NULL)) 900 sync_start(); 901 error = pthread_create(&td, NULL, ggate_recv_thread, res); 902 PJDLOG_ASSERT(error == 0); 903 error = pthread_create(&td, NULL, local_send_thread, res); 904 PJDLOG_ASSERT(error == 0); 905 error = pthread_create(&td, NULL, remote_send_thread, res); 906 PJDLOG_ASSERT(error == 0); 907 error = pthread_create(&td, NULL, remote_recv_thread, res); 908 PJDLOG_ASSERT(error == 0); 909 error = pthread_create(&td, NULL, ggate_send_thread, res); 910 PJDLOG_ASSERT(error == 0); 911 (void)sync_thread(res); 912} 913 914static void 915reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 916{ 917 char msg[1024]; 918 va_list ap; 919 int len; 920 921 va_start(ap, fmt); 922 len = vsnprintf(msg, sizeof(msg), fmt, ap); 923 va_end(ap); 924 if ((size_t)len < sizeof(msg)) { 925 switch (ggio->gctl_cmd) { 926 case BIO_READ: 927 (void)snprintf(msg + len, sizeof(msg) - len, 928 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 929 (uintmax_t)ggio->gctl_length); 930 break; 931 case BIO_DELETE: 932 (void)snprintf(msg + len, sizeof(msg) - len, 933 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 934 (uintmax_t)ggio->gctl_length); 935 break; 936 case BIO_FLUSH: 937 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 938 break; 939 case BIO_WRITE: 940 (void)snprintf(msg + len, sizeof(msg) - len, 941 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 942 (uintmax_t)ggio->gctl_length); 943 break; 944 default: 945 (void)snprintf(msg + len, sizeof(msg) - len, 946 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 947 break; 948 } 949 } 950 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 951} 952 953static void 954remote_close(struct hast_resource *res, int ncomp) 955{ 956 957 rw_wlock(&hio_remote_lock[ncomp]); 958 /* 959 * A race is possible between dropping rlock and acquiring wlock - 960 * another thread can close connection in-between. 961 */ 962 if (!ISCONNECTED(res, ncomp)) { 963 PJDLOG_ASSERT(res->hr_remotein == NULL); 964 PJDLOG_ASSERT(res->hr_remoteout == NULL); 965 rw_unlock(&hio_remote_lock[ncomp]); 966 return; 967 } 968 969 PJDLOG_ASSERT(res->hr_remotein != NULL); 970 PJDLOG_ASSERT(res->hr_remoteout != NULL); 971 972 pjdlog_debug(2, "Closing incoming connection to %s.", 973 res->hr_remoteaddr); 974 proto_close(res->hr_remotein); 975 res->hr_remotein = NULL; 976 pjdlog_debug(2, "Closing outgoing connection to %s.", 977 res->hr_remoteaddr); 978 proto_close(res->hr_remoteout); 979 res->hr_remoteout = NULL; 980 981 rw_unlock(&hio_remote_lock[ncomp]); 982 983 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 984 985 /* 986 * Stop synchronization if in-progress. 987 */ 988 sync_stop(); 989 990 event_send(res, EVENT_DISCONNECT); 991} 992 993/* 994 * Thread receives ggate I/O requests from the kernel and passes them to 995 * appropriate threads: 996 * WRITE - always goes to both local_send and remote_send threads 997 * READ (when the block is up-to-date on local component) - 998 * only local_send thread 999 * READ (when the block isn't up-to-date on local component) - 1000 * only remote_send thread 1001 * DELETE - always goes to both local_send and remote_send threads 1002 * FLUSH - always goes to both local_send and remote_send threads 1003 */ 1004static void * 1005ggate_recv_thread(void *arg) 1006{ 1007 struct hast_resource *res = arg; 1008 struct g_gate_ctl_io *ggio; 1009 struct hio *hio; 1010 unsigned int ii, ncomp, ncomps; 1011 int error; 1012 1013 ncomps = HAST_NCOMPONENTS; 1014 1015 for (;;) { 1016 pjdlog_debug(2, "ggate_recv: Taking free request."); 1017 QUEUE_TAKE2(hio, free); 1018 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 1019 ggio = &hio->hio_ggio; 1020 ggio->gctl_unit = res->hr_ggateunit; 1021 ggio->gctl_length = MAXPHYS; 1022 ggio->gctl_error = 0; 1023 pjdlog_debug(2, 1024 "ggate_recv: (%p) Waiting for request from the kernel.", 1025 hio); 1026 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 1027 if (sigexit_received) 1028 pthread_exit(NULL); 1029 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 1030 } 1031 error = ggio->gctl_error; 1032 switch (error) { 1033 case 0: 1034 break; 1035 case ECANCELED: 1036 /* Exit gracefully. */ 1037 if (!sigexit_received) { 1038 pjdlog_debug(2, 1039 "ggate_recv: (%p) Received cancel from the kernel.", 1040 hio); 1041 pjdlog_info("Received cancel from the kernel, exiting."); 1042 } 1043 pthread_exit(NULL); 1044 case ENOMEM: 1045 /* 1046 * Buffer too small? Impossible, we allocate MAXPHYS 1047 * bytes - request can't be bigger than that. 1048 */ 1049 /* FALLTHROUGH */ 1050 case ENXIO: 1051 default: 1052 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 1053 strerror(error)); 1054 } 1055 for (ii = 0; ii < ncomps; ii++) 1056 hio->hio_errors[ii] = EINVAL; 1057 reqlog(LOG_DEBUG, 2, ggio, 1058 "ggate_recv: (%p) Request received from the kernel: ", 1059 hio); 1060 /* 1061 * Inform all components about new write request. 1062 * For read request prefer local component unless the given 1063 * range is out-of-date, then use remote component. 1064 */ 1065 switch (ggio->gctl_cmd) { 1066 case BIO_READ: 1067 pjdlog_debug(2, 1068 "ggate_recv: (%p) Moving request to the send queue.", 1069 hio); 1070 refcount_init(&hio->hio_countdown, 1); 1071 mtx_lock(&metadata_lock); 1072 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 1073 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1074 /* 1075 * This range is up-to-date on local component, 1076 * so handle request locally. 1077 */ 1078 /* Local component is 0 for now. */ 1079 ncomp = 0; 1080 } else /* if (res->hr_syncsrc == 1081 HAST_SYNCSRC_SECONDARY) */ { 1082 PJDLOG_ASSERT(res->hr_syncsrc == 1083 HAST_SYNCSRC_SECONDARY); 1084 /* 1085 * This range is out-of-date on local component, 1086 * so send request to the remote node. 1087 */ 1088 /* Remote component is 1 for now. */ 1089 ncomp = 1; 1090 } 1091 mtx_unlock(&metadata_lock); 1092 QUEUE_INSERT1(hio, send, ncomp); 1093 break; 1094 case BIO_WRITE: 1095 if (res->hr_resuid == 0) { 1096 /* 1097 * This is first write, initialize localcnt and 1098 * resuid. 1099 */ 1100 res->hr_primary_localcnt = 1; 1101 (void)init_resuid(res); 1102 } 1103 for (;;) { 1104 mtx_lock(&range_lock); 1105 if (rangelock_islocked(range_sync, 1106 ggio->gctl_offset, ggio->gctl_length)) { 1107 pjdlog_debug(2, 1108 "regular: Range offset=%jd length=%zu locked.", 1109 (intmax_t)ggio->gctl_offset, 1110 (size_t)ggio->gctl_length); 1111 range_regular_wait = true; 1112 cv_wait(&range_regular_cond, &range_lock); 1113 range_regular_wait = false; 1114 mtx_unlock(&range_lock); 1115 continue; 1116 } 1117 if (rangelock_add(range_regular, 1118 ggio->gctl_offset, ggio->gctl_length) < 0) { 1119 mtx_unlock(&range_lock); 1120 pjdlog_debug(2, 1121 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1122 (intmax_t)ggio->gctl_offset, 1123 (size_t)ggio->gctl_length); 1124 sleep(1); 1125 continue; 1126 } 1127 mtx_unlock(&range_lock); 1128 break; 1129 } 1130 mtx_lock(&res->hr_amp_lock); 1131 if (activemap_write_start(res->hr_amp, 1132 ggio->gctl_offset, ggio->gctl_length)) { 1133 (void)hast_activemap_flush(res); 1134 } 1135 mtx_unlock(&res->hr_amp_lock); 1136 /* FALLTHROUGH */ 1137 case BIO_DELETE: 1138 case BIO_FLUSH: 1139 pjdlog_debug(2, 1140 "ggate_recv: (%p) Moving request to the send queues.", 1141 hio); 1142 refcount_init(&hio->hio_countdown, ncomps); 1143 for (ii = 0; ii < ncomps; ii++) 1144 QUEUE_INSERT1(hio, send, ii); 1145 break; 1146 } 1147 } 1148 /* NOTREACHED */ 1149 return (NULL); 1150} 1151 1152/* 1153 * Thread reads from or writes to local component. 1154 * If local read fails, it redirects it to remote_send thread. 1155 */ 1156static void * 1157local_send_thread(void *arg) 1158{ 1159 struct hast_resource *res = arg; 1160 struct g_gate_ctl_io *ggio; 1161 struct hio *hio; 1162 unsigned int ncomp, rncomp; 1163 ssize_t ret; 1164 1165 /* Local component is 0 for now. */ 1166 ncomp = 0; 1167 /* Remote component is 1 for now. */ 1168 rncomp = 1; 1169 1170 for (;;) { 1171 pjdlog_debug(2, "local_send: Taking request."); 1172 QUEUE_TAKE1(hio, send, ncomp, 0); 1173 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1174 ggio = &hio->hio_ggio; 1175 switch (ggio->gctl_cmd) { 1176 case BIO_READ: 1177 ret = pread(res->hr_localfd, ggio->gctl_data, 1178 ggio->gctl_length, 1179 ggio->gctl_offset + res->hr_localoff); 1180 if (ret == ggio->gctl_length) 1181 hio->hio_errors[ncomp] = 0; 1182 else { 1183 /* 1184 * If READ failed, try to read from remote node. 1185 */ 1186 if (ret < 0) { 1187 reqlog(LOG_WARNING, 0, ggio, 1188 "Local request failed (%s), trying remote node. ", 1189 strerror(errno)); 1190 } else if (ret != ggio->gctl_length) { 1191 reqlog(LOG_WARNING, 0, ggio, 1192 "Local request failed (%zd != %jd), trying remote node. ", 1193 ret, (intmax_t)ggio->gctl_length); 1194 } 1195 QUEUE_INSERT1(hio, send, rncomp); 1196 continue; 1197 } 1198 break; 1199 case BIO_WRITE: 1200 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1201 ggio->gctl_length, 1202 ggio->gctl_offset + res->hr_localoff); 1203 if (ret < 0) { 1204 hio->hio_errors[ncomp] = errno; 1205 reqlog(LOG_WARNING, 0, ggio, 1206 "Local request failed (%s): ", 1207 strerror(errno)); 1208 } else if (ret != ggio->gctl_length) { 1209 hio->hio_errors[ncomp] = EIO; 1210 reqlog(LOG_WARNING, 0, ggio, 1211 "Local request failed (%zd != %jd): ", 1212 ret, (intmax_t)ggio->gctl_length); 1213 } else { 1214 hio->hio_errors[ncomp] = 0; 1215 } 1216 break; 1217 case BIO_DELETE: 1218 ret = g_delete(res->hr_localfd, 1219 ggio->gctl_offset + res->hr_localoff, 1220 ggio->gctl_length); 1221 if (ret < 0) { 1222 hio->hio_errors[ncomp] = errno; 1223 reqlog(LOG_WARNING, 0, ggio, 1224 "Local request failed (%s): ", 1225 strerror(errno)); 1226 } else { 1227 hio->hio_errors[ncomp] = 0; 1228 } 1229 break; 1230 case BIO_FLUSH: 1231 ret = g_flush(res->hr_localfd); 1232 if (ret < 0) { 1233 hio->hio_errors[ncomp] = errno; 1234 reqlog(LOG_WARNING, 0, ggio, 1235 "Local request failed (%s): ", 1236 strerror(errno)); 1237 } else { 1238 hio->hio_errors[ncomp] = 0; 1239 } 1240 break; 1241 } 1242 if (refcount_release(&hio->hio_countdown)) { 1243 if (ISSYNCREQ(hio)) { 1244 mtx_lock(&sync_lock); 1245 SYNCREQDONE(hio); 1246 mtx_unlock(&sync_lock); 1247 cv_signal(&sync_cond); 1248 } else { 1249 pjdlog_debug(2, 1250 "local_send: (%p) Moving request to the done queue.", 1251 hio); 1252 QUEUE_INSERT2(hio, done); 1253 } 1254 } 1255 } 1256 /* NOTREACHED */ 1257 return (NULL); 1258} 1259 1260static void 1261keepalive_send(struct hast_resource *res, unsigned int ncomp) 1262{ 1263 struct nv *nv; 1264 1265 rw_rlock(&hio_remote_lock[ncomp]); 1266 1267 if (!ISCONNECTED(res, ncomp)) { 1268 rw_unlock(&hio_remote_lock[ncomp]); 1269 return; 1270 } 1271 1272 PJDLOG_ASSERT(res->hr_remotein != NULL); 1273 PJDLOG_ASSERT(res->hr_remoteout != NULL); 1274 1275 nv = nv_alloc(); 1276 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1277 if (nv_error(nv) != 0) { 1278 rw_unlock(&hio_remote_lock[ncomp]); 1279 nv_free(nv); 1280 pjdlog_debug(1, 1281 "keepalive_send: Unable to prepare header to send."); 1282 return; 1283 } 1284 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1285 rw_unlock(&hio_remote_lock[ncomp]); 1286 pjdlog_common(LOG_DEBUG, 1, errno, 1287 "keepalive_send: Unable to send request"); 1288 nv_free(nv); 1289 remote_close(res, ncomp); 1290 return; 1291 } 1292 1293 rw_unlock(&hio_remote_lock[ncomp]); 1294 nv_free(nv); 1295 pjdlog_debug(2, "keepalive_send: Request sent."); 1296} 1297 1298/* 1299 * Thread sends request to secondary node. 1300 */ 1301static void * 1302remote_send_thread(void *arg) 1303{ 1304 struct hast_resource *res = arg; 1305 struct g_gate_ctl_io *ggio; 1306 time_t lastcheck, now; 1307 struct hio *hio; 1308 struct nv *nv; 1309 unsigned int ncomp; 1310 bool wakeup; 1311 uint64_t offset, length; 1312 uint8_t cmd; 1313 void *data; 1314 1315 /* Remote component is 1 for now. */ 1316 ncomp = 1; 1317 lastcheck = time(NULL); 1318 1319 for (;;) { 1320 pjdlog_debug(2, "remote_send: Taking request."); 1321 QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE); 1322 if (hio == NULL) { 1323 now = time(NULL); 1324 if (lastcheck + HAST_KEEPALIVE <= now) { 1325 keepalive_send(res, ncomp); 1326 lastcheck = now; 1327 } 1328 continue; 1329 } 1330 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1331 ggio = &hio->hio_ggio; 1332 switch (ggio->gctl_cmd) { 1333 case BIO_READ: 1334 cmd = HIO_READ; 1335 data = NULL; 1336 offset = ggio->gctl_offset; 1337 length = ggio->gctl_length; 1338 break; 1339 case BIO_WRITE: 1340 cmd = HIO_WRITE; 1341 data = ggio->gctl_data; 1342 offset = ggio->gctl_offset; 1343 length = ggio->gctl_length; 1344 break; 1345 case BIO_DELETE: 1346 cmd = HIO_DELETE; 1347 data = NULL; 1348 offset = ggio->gctl_offset; 1349 length = ggio->gctl_length; 1350 break; 1351 case BIO_FLUSH: 1352 cmd = HIO_FLUSH; 1353 data = NULL; 1354 offset = 0; 1355 length = 0; 1356 break; 1357 default: 1358 PJDLOG_ASSERT(!"invalid condition"); 1359 abort(); 1360 } 1361 nv = nv_alloc(); 1362 nv_add_uint8(nv, cmd, "cmd"); 1363 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1364 nv_add_uint64(nv, offset, "offset"); 1365 nv_add_uint64(nv, length, "length"); 1366 if (nv_error(nv) != 0) { 1367 hio->hio_errors[ncomp] = nv_error(nv); 1368 pjdlog_debug(2, 1369 "remote_send: (%p) Unable to prepare header to send.", 1370 hio); 1371 reqlog(LOG_ERR, 0, ggio, 1372 "Unable to prepare header to send (%s): ", 1373 strerror(nv_error(nv))); 1374 /* Move failed request immediately to the done queue. */ 1375 goto done_queue; 1376 } 1377 pjdlog_debug(2, 1378 "remote_send: (%p) Moving request to the recv queue.", 1379 hio); 1380 /* 1381 * Protect connection from disappearing. 1382 */ 1383 rw_rlock(&hio_remote_lock[ncomp]); 1384 if (!ISCONNECTED(res, ncomp)) { 1385 rw_unlock(&hio_remote_lock[ncomp]); 1386 hio->hio_errors[ncomp] = ENOTCONN; 1387 goto done_queue; 1388 } 1389 /* 1390 * Move the request to recv queue before sending it, because 1391 * in different order we can get reply before we move request 1392 * to recv queue. 1393 */ 1394 mtx_lock(&hio_recv_list_lock[ncomp]); 1395 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1396 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1397 mtx_unlock(&hio_recv_list_lock[ncomp]); 1398 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1399 data != NULL ? length : 0) < 0) { 1400 hio->hio_errors[ncomp] = errno; 1401 rw_unlock(&hio_remote_lock[ncomp]); 1402 pjdlog_debug(2, 1403 "remote_send: (%p) Unable to send request.", hio); 1404 reqlog(LOG_ERR, 0, ggio, 1405 "Unable to send request (%s): ", 1406 strerror(hio->hio_errors[ncomp])); 1407 remote_close(res, ncomp); 1408 /* 1409 * Take request back from the receive queue and move 1410 * it immediately to the done queue. 1411 */ 1412 mtx_lock(&hio_recv_list_lock[ncomp]); 1413 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1414 mtx_unlock(&hio_recv_list_lock[ncomp]); 1415 goto done_queue; 1416 } 1417 rw_unlock(&hio_remote_lock[ncomp]); 1418 nv_free(nv); 1419 if (wakeup) 1420 cv_signal(&hio_recv_list_cond[ncomp]); 1421 continue; 1422done_queue: 1423 nv_free(nv); 1424 if (ISSYNCREQ(hio)) { 1425 if (!refcount_release(&hio->hio_countdown)) 1426 continue; 1427 mtx_lock(&sync_lock); 1428 SYNCREQDONE(hio); 1429 mtx_unlock(&sync_lock); 1430 cv_signal(&sync_cond); 1431 continue; 1432 } 1433 if (ggio->gctl_cmd == BIO_WRITE) { 1434 mtx_lock(&res->hr_amp_lock); 1435 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1436 ggio->gctl_length)) { 1437 (void)hast_activemap_flush(res); 1438 } 1439 mtx_unlock(&res->hr_amp_lock); 1440 } 1441 if (!refcount_release(&hio->hio_countdown)) 1442 continue; 1443 pjdlog_debug(2, 1444 "remote_send: (%p) Moving request to the done queue.", 1445 hio); 1446 QUEUE_INSERT2(hio, done); 1447 } 1448 /* NOTREACHED */ 1449 return (NULL); 1450} 1451 1452/* 1453 * Thread receives answer from secondary node and passes it to ggate_send 1454 * thread. 1455 */ 1456static void * 1457remote_recv_thread(void *arg) 1458{ 1459 struct hast_resource *res = arg; 1460 struct g_gate_ctl_io *ggio; 1461 struct hio *hio; 1462 struct nv *nv; 1463 unsigned int ncomp; 1464 uint64_t seq; 1465 int error; 1466 1467 /* Remote component is 1 for now. */ 1468 ncomp = 1; 1469 1470 for (;;) { 1471 /* Wait until there is anything to receive. */ 1472 mtx_lock(&hio_recv_list_lock[ncomp]); 1473 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1474 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1475 cv_wait(&hio_recv_list_cond[ncomp], 1476 &hio_recv_list_lock[ncomp]); 1477 } 1478 mtx_unlock(&hio_recv_list_lock[ncomp]); 1479 rw_rlock(&hio_remote_lock[ncomp]); 1480 if (!ISCONNECTED(res, ncomp)) { 1481 rw_unlock(&hio_remote_lock[ncomp]); 1482 /* 1483 * Connection is dead, so move all pending requests to 1484 * the done queue (one-by-one). 1485 */ 1486 mtx_lock(&hio_recv_list_lock[ncomp]); 1487 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1488 PJDLOG_ASSERT(hio != NULL); 1489 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1490 hio_next[ncomp]); 1491 mtx_unlock(&hio_recv_list_lock[ncomp]); 1492 goto done_queue; 1493 } 1494 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1495 pjdlog_errno(LOG_ERR, 1496 "Unable to receive reply header"); 1497 rw_unlock(&hio_remote_lock[ncomp]); 1498 remote_close(res, ncomp); 1499 continue; 1500 } 1501 rw_unlock(&hio_remote_lock[ncomp]); 1502 seq = nv_get_uint64(nv, "seq"); 1503 if (seq == 0) { 1504 pjdlog_error("Header contains no 'seq' field."); 1505 nv_free(nv); 1506 continue; 1507 } 1508 mtx_lock(&hio_recv_list_lock[ncomp]); 1509 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1510 if (hio->hio_ggio.gctl_seq == seq) { 1511 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1512 hio_next[ncomp]); 1513 break; 1514 } 1515 } 1516 mtx_unlock(&hio_recv_list_lock[ncomp]); 1517 if (hio == NULL) { 1518 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1519 (uintmax_t)seq); 1520 nv_free(nv); 1521 continue; 1522 } 1523 error = nv_get_int16(nv, "error"); 1524 if (error != 0) { 1525 /* Request failed on remote side. */ 1526 hio->hio_errors[ncomp] = error; 1527 reqlog(LOG_WARNING, 0, &hio->hio_ggio, 1528 "Remote request failed (%s): ", strerror(error)); 1529 nv_free(nv); 1530 goto done_queue; 1531 } 1532 ggio = &hio->hio_ggio; 1533 switch (ggio->gctl_cmd) { 1534 case BIO_READ: 1535 rw_rlock(&hio_remote_lock[ncomp]); 1536 if (!ISCONNECTED(res, ncomp)) { 1537 rw_unlock(&hio_remote_lock[ncomp]); 1538 nv_free(nv); 1539 goto done_queue; 1540 } 1541 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1542 ggio->gctl_data, ggio->gctl_length) < 0) { 1543 hio->hio_errors[ncomp] = errno; 1544 pjdlog_errno(LOG_ERR, 1545 "Unable to receive reply data"); 1546 rw_unlock(&hio_remote_lock[ncomp]); 1547 nv_free(nv); 1548 remote_close(res, ncomp); 1549 goto done_queue; 1550 } 1551 rw_unlock(&hio_remote_lock[ncomp]); 1552 break; 1553 case BIO_WRITE: 1554 case BIO_DELETE: 1555 case BIO_FLUSH: 1556 break; 1557 default: 1558 PJDLOG_ASSERT(!"invalid condition"); 1559 abort(); 1560 } 1561 hio->hio_errors[ncomp] = 0; 1562 nv_free(nv); 1563done_queue: 1564 if (refcount_release(&hio->hio_countdown)) { 1565 if (ISSYNCREQ(hio)) { 1566 mtx_lock(&sync_lock); 1567 SYNCREQDONE(hio); 1568 mtx_unlock(&sync_lock); 1569 cv_signal(&sync_cond); 1570 } else { 1571 pjdlog_debug(2, 1572 "remote_recv: (%p) Moving request to the done queue.", 1573 hio); 1574 QUEUE_INSERT2(hio, done); 1575 } 1576 } 1577 } 1578 /* NOTREACHED */ 1579 return (NULL); 1580} 1581 1582/* 1583 * Thread sends answer to the kernel. 1584 */ 1585static void * 1586ggate_send_thread(void *arg) 1587{ 1588 struct hast_resource *res = arg; 1589 struct g_gate_ctl_io *ggio; 1590 struct hio *hio; 1591 unsigned int ii, ncomp, ncomps; 1592 1593 ncomps = HAST_NCOMPONENTS; 1594 1595 for (;;) { 1596 pjdlog_debug(2, "ggate_send: Taking request."); 1597 QUEUE_TAKE2(hio, done); 1598 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1599 ggio = &hio->hio_ggio; 1600 for (ii = 0; ii < ncomps; ii++) { 1601 if (hio->hio_errors[ii] == 0) { 1602 /* 1603 * One successful request is enough to declare 1604 * success. 1605 */ 1606 ggio->gctl_error = 0; 1607 break; 1608 } 1609 } 1610 if (ii == ncomps) { 1611 /* 1612 * None of the requests were successful. 1613 * Use the error from local component except the 1614 * case when we did only remote request. 1615 */ 1616 if (ggio->gctl_cmd == BIO_READ && 1617 res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) 1618 ggio->gctl_error = hio->hio_errors[1]; 1619 else 1620 ggio->gctl_error = hio->hio_errors[0]; 1621 } 1622 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1623 mtx_lock(&res->hr_amp_lock); 1624 activemap_write_complete(res->hr_amp, 1625 ggio->gctl_offset, ggio->gctl_length); 1626 mtx_unlock(&res->hr_amp_lock); 1627 } 1628 if (ggio->gctl_cmd == BIO_WRITE) { 1629 /* 1630 * Unlock range we locked. 1631 */ 1632 mtx_lock(&range_lock); 1633 rangelock_del(range_regular, ggio->gctl_offset, 1634 ggio->gctl_length); 1635 if (range_sync_wait) 1636 cv_signal(&range_sync_cond); 1637 mtx_unlock(&range_lock); 1638 /* 1639 * Bump local count if this is first write after 1640 * connection failure with remote node. 1641 */ 1642 ncomp = 1; 1643 rw_rlock(&hio_remote_lock[ncomp]); 1644 if (!ISCONNECTED(res, ncomp)) { 1645 mtx_lock(&metadata_lock); 1646 if (res->hr_primary_localcnt == 1647 res->hr_secondary_remotecnt) { 1648 res->hr_primary_localcnt++; 1649 pjdlog_debug(1, 1650 "Increasing localcnt to %ju.", 1651 (uintmax_t)res->hr_primary_localcnt); 1652 (void)metadata_write(res); 1653 } 1654 mtx_unlock(&metadata_lock); 1655 } 1656 rw_unlock(&hio_remote_lock[ncomp]); 1657 } 1658 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1659 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1660 pjdlog_debug(2, 1661 "ggate_send: (%p) Moving request to the free queue.", hio); 1662 QUEUE_INSERT2(hio, free); 1663 } 1664 /* NOTREACHED */ 1665 return (NULL); 1666} 1667 1668/* 1669 * Thread synchronize local and remote components. 1670 */ 1671static void * 1672sync_thread(void *arg __unused) 1673{ 1674 struct hast_resource *res = arg; 1675 struct hio *hio; 1676 struct g_gate_ctl_io *ggio; 1677 struct timeval tstart, tend, tdiff; 1678 unsigned int ii, ncomp, ncomps; 1679 off_t offset, length, synced; 1680 bool dorewind; 1681 int syncext; 1682 1683 ncomps = HAST_NCOMPONENTS; 1684 dorewind = true; 1685 synced = 0; 1686 offset = -1; 1687 1688 for (;;) { 1689 mtx_lock(&sync_lock); 1690 if (offset >= 0 && !sync_inprogress) { 1691 gettimeofday(&tend, NULL); 1692 timersub(&tend, &tstart, &tdiff); 1693 pjdlog_info("Synchronization interrupted after %#.0T. " 1694 "%NB synchronized so far.", &tdiff, 1695 (intmax_t)synced); 1696 event_send(res, EVENT_SYNCINTR); 1697 } 1698 while (!sync_inprogress) { 1699 dorewind = true; 1700 synced = 0; 1701 cv_wait(&sync_cond, &sync_lock); 1702 } 1703 mtx_unlock(&sync_lock); 1704 /* 1705 * Obtain offset at which we should synchronize. 1706 * Rewind synchronization if needed. 1707 */ 1708 mtx_lock(&res->hr_amp_lock); 1709 if (dorewind) 1710 activemap_sync_rewind(res->hr_amp); 1711 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1712 if (syncext != -1) { 1713 /* 1714 * We synchronized entire syncext extent, we can mark 1715 * it as clean now. 1716 */ 1717 if (activemap_extent_complete(res->hr_amp, syncext)) 1718 (void)hast_activemap_flush(res); 1719 } 1720 mtx_unlock(&res->hr_amp_lock); 1721 if (dorewind) { 1722 dorewind = false; 1723 if (offset < 0) 1724 pjdlog_info("Nodes are in sync."); 1725 else { 1726 pjdlog_info("Synchronization started. %NB to go.", 1727 (intmax_t)(res->hr_extentsize * 1728 activemap_ndirty(res->hr_amp))); 1729 event_send(res, EVENT_SYNCSTART); 1730 gettimeofday(&tstart, NULL); 1731 } 1732 } 1733 if (offset < 0) { 1734 sync_stop(); 1735 pjdlog_debug(1, "Nothing to synchronize."); 1736 /* 1737 * Synchronization complete, make both localcnt and 1738 * remotecnt equal. 1739 */ 1740 ncomp = 1; 1741 rw_rlock(&hio_remote_lock[ncomp]); 1742 if (ISCONNECTED(res, ncomp)) { 1743 if (synced > 0) { 1744 int64_t bps; 1745 1746 gettimeofday(&tend, NULL); 1747 timersub(&tend, &tstart, &tdiff); 1748 bps = (int64_t)((double)synced / 1749 ((double)tdiff.tv_sec + 1750 (double)tdiff.tv_usec / 1000000)); 1751 pjdlog_info("Synchronization complete. " 1752 "%NB synchronized in %#.0lT (%NB/sec).", 1753 (intmax_t)synced, &tdiff, 1754 (intmax_t)bps); 1755 event_send(res, EVENT_SYNCDONE); 1756 } 1757 mtx_lock(&metadata_lock); 1758 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1759 res->hr_primary_localcnt = 1760 res->hr_secondary_remotecnt; 1761 res->hr_primary_remotecnt = 1762 res->hr_secondary_localcnt; 1763 pjdlog_debug(1, 1764 "Setting localcnt to %ju and remotecnt to %ju.", 1765 (uintmax_t)res->hr_primary_localcnt, 1766 (uintmax_t)res->hr_primary_remotecnt); 1767 (void)metadata_write(res); 1768 mtx_unlock(&metadata_lock); 1769 } 1770 rw_unlock(&hio_remote_lock[ncomp]); 1771 continue; 1772 } 1773 pjdlog_debug(2, "sync: Taking free request."); 1774 QUEUE_TAKE2(hio, free); 1775 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1776 /* 1777 * Lock the range we are going to synchronize. We don't want 1778 * race where someone writes between our read and write. 1779 */ 1780 for (;;) { 1781 mtx_lock(&range_lock); 1782 if (rangelock_islocked(range_regular, offset, length)) { 1783 pjdlog_debug(2, 1784 "sync: Range offset=%jd length=%jd locked.", 1785 (intmax_t)offset, (intmax_t)length); 1786 range_sync_wait = true; 1787 cv_wait(&range_sync_cond, &range_lock); 1788 range_sync_wait = false; 1789 mtx_unlock(&range_lock); 1790 continue; 1791 } 1792 if (rangelock_add(range_sync, offset, length) < 0) { 1793 mtx_unlock(&range_lock); 1794 pjdlog_debug(2, 1795 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1796 (intmax_t)offset, (intmax_t)length); 1797 sleep(1); 1798 continue; 1799 } 1800 mtx_unlock(&range_lock); 1801 break; 1802 } 1803 /* 1804 * First read the data from synchronization source. 1805 */ 1806 SYNCREQ(hio); 1807 ggio = &hio->hio_ggio; 1808 ggio->gctl_cmd = BIO_READ; 1809 ggio->gctl_offset = offset; 1810 ggio->gctl_length = length; 1811 ggio->gctl_error = 0; 1812 for (ii = 0; ii < ncomps; ii++) 1813 hio->hio_errors[ii] = EINVAL; 1814 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1815 hio); 1816 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1817 hio); 1818 mtx_lock(&metadata_lock); 1819 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1820 /* 1821 * This range is up-to-date on local component, 1822 * so handle request locally. 1823 */ 1824 /* Local component is 0 for now. */ 1825 ncomp = 0; 1826 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1827 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1828 /* 1829 * This range is out-of-date on local component, 1830 * so send request to the remote node. 1831 */ 1832 /* Remote component is 1 for now. */ 1833 ncomp = 1; 1834 } 1835 mtx_unlock(&metadata_lock); 1836 refcount_init(&hio->hio_countdown, 1); 1837 QUEUE_INSERT1(hio, send, ncomp); 1838 1839 /* 1840 * Let's wait for READ to finish. 1841 */ 1842 mtx_lock(&sync_lock); 1843 while (!ISSYNCREQDONE(hio)) 1844 cv_wait(&sync_cond, &sync_lock); 1845 mtx_unlock(&sync_lock); 1846 1847 if (hio->hio_errors[ncomp] != 0) { 1848 pjdlog_error("Unable to read synchronization data: %s.", 1849 strerror(hio->hio_errors[ncomp])); 1850 goto free_queue; 1851 } 1852 1853 /* 1854 * We read the data from synchronization source, now write it 1855 * to synchronization target. 1856 */ 1857 SYNCREQ(hio); 1858 ggio->gctl_cmd = BIO_WRITE; 1859 for (ii = 0; ii < ncomps; ii++) 1860 hio->hio_errors[ii] = EINVAL; 1861 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1862 hio); 1863 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1864 hio); 1865 mtx_lock(&metadata_lock); 1866 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1867 /* 1868 * This range is up-to-date on local component, 1869 * so we update remote component. 1870 */ 1871 /* Remote component is 1 for now. */ 1872 ncomp = 1; 1873 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1874 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1875 /* 1876 * This range is out-of-date on local component, 1877 * so we update it. 1878 */ 1879 /* Local component is 0 for now. */ 1880 ncomp = 0; 1881 } 1882 mtx_unlock(&metadata_lock); 1883 1884 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1885 hio); 1886 refcount_init(&hio->hio_countdown, 1); 1887 QUEUE_INSERT1(hio, send, ncomp); 1888 1889 /* 1890 * Let's wait for WRITE to finish. 1891 */ 1892 mtx_lock(&sync_lock); 1893 while (!ISSYNCREQDONE(hio)) 1894 cv_wait(&sync_cond, &sync_lock); 1895 mtx_unlock(&sync_lock); 1896 1897 if (hio->hio_errors[ncomp] != 0) { 1898 pjdlog_error("Unable to write synchronization data: %s.", 1899 strerror(hio->hio_errors[ncomp])); 1900 goto free_queue; 1901 } 1902 1903 synced += length; 1904free_queue: 1905 mtx_lock(&range_lock); 1906 rangelock_del(range_sync, offset, length); 1907 if (range_regular_wait) 1908 cv_signal(&range_regular_cond); 1909 mtx_unlock(&range_lock); 1910 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1911 hio); 1912 QUEUE_INSERT2(hio, free); 1913 } 1914 /* NOTREACHED */ 1915 return (NULL); 1916} 1917 1918void 1919primary_config_reload(struct hast_resource *res, struct nv *nv) 1920{ 1921 unsigned int ii, ncomps; 1922 int modified, vint; 1923 const char *vstr; 1924 1925 pjdlog_info("Reloading configuration..."); 1926 1927 PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY); 1928 PJDLOG_ASSERT(gres == res); 1929 nv_assert(nv, "remoteaddr"); 1930 nv_assert(nv, "sourceaddr"); 1931 nv_assert(nv, "replication"); 1932 nv_assert(nv, "checksum"); 1933 nv_assert(nv, "compression"); 1934 nv_assert(nv, "timeout"); 1935 nv_assert(nv, "exec"); 1936 1937 ncomps = HAST_NCOMPONENTS; 1938 1939#define MODIFIED_REMOTEADDR 0x01 1940#define MODIFIED_SOURCEADDR 0x02 1941#define MODIFIED_REPLICATION 0x04 1942#define MODIFIED_CHECKSUM 0x08 1943#define MODIFIED_COMPRESSION 0x10 1944#define MODIFIED_TIMEOUT 0x20 1945#define MODIFIED_EXEC 0x40 1946 modified = 0; 1947 1948 vstr = nv_get_string(nv, "remoteaddr"); 1949 if (strcmp(gres->hr_remoteaddr, vstr) != 0) { 1950 /* 1951 * Don't copy res->hr_remoteaddr to gres just yet. 1952 * We want remote_close() to log disconnect from the old 1953 * addresses, not from the new ones. 1954 */ 1955 modified |= MODIFIED_REMOTEADDR; 1956 } 1957 vstr = nv_get_string(nv, "sourceaddr"); 1958 if (strcmp(gres->hr_sourceaddr, vstr) != 0) { 1959 strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr)); 1960 modified |= MODIFIED_SOURCEADDR; 1961 } 1962 vint = nv_get_int32(nv, "replication"); 1963 if (gres->hr_replication != vint) { 1964 gres->hr_replication = vint; 1965 modified |= MODIFIED_REPLICATION; 1966 } 1967 vint = nv_get_int32(nv, "checksum"); 1968 if (gres->hr_checksum != vint) { 1969 gres->hr_checksum = vint; 1970 modified |= MODIFIED_CHECKSUM; 1971 } 1972 vint = nv_get_int32(nv, "compression"); 1973 if (gres->hr_compression != vint) { 1974 gres->hr_compression = vint; 1975 modified |= MODIFIED_COMPRESSION; 1976 } 1977 vint = nv_get_int32(nv, "timeout"); 1978 if (gres->hr_timeout != vint) { 1979 gres->hr_timeout = vint; 1980 modified |= MODIFIED_TIMEOUT; 1981 } 1982 vstr = nv_get_string(nv, "exec"); 1983 if (strcmp(gres->hr_exec, vstr) != 0) { 1984 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec)); 1985 modified |= MODIFIED_EXEC; 1986 } 1987 1988 /* 1989 * Change timeout for connected sockets. 1990 * Don't bother if we need to reconnect. 1991 */ 1992 if ((modified & MODIFIED_TIMEOUT) != 0 && 1993 (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR | 1994 MODIFIED_REPLICATION)) == 0) { 1995 for (ii = 0; ii < ncomps; ii++) { 1996 if (!ISREMOTE(ii)) 1997 continue; 1998 rw_rlock(&hio_remote_lock[ii]); 1999 if (!ISCONNECTED(gres, ii)) { 2000 rw_unlock(&hio_remote_lock[ii]); 2001 continue; 2002 } 2003 rw_unlock(&hio_remote_lock[ii]); 2004 if (proto_timeout(gres->hr_remotein, 2005 gres->hr_timeout) < 0) { 2006 pjdlog_errno(LOG_WARNING, 2007 "Unable to set connection timeout"); 2008 } 2009 if (proto_timeout(gres->hr_remoteout, 2010 gres->hr_timeout) < 0) { 2011 pjdlog_errno(LOG_WARNING, 2012 "Unable to set connection timeout"); 2013 } 2014 } 2015 } 2016 if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR | 2017 MODIFIED_REPLICATION)) != 0) { 2018 for (ii = 0; ii < ncomps; ii++) { 2019 if (!ISREMOTE(ii)) 2020 continue; 2021 remote_close(gres, ii); 2022 } 2023 if (modified & MODIFIED_REMOTEADDR) { 2024 vstr = nv_get_string(nv, "remoteaddr"); 2025 strlcpy(gres->hr_remoteaddr, vstr, 2026 sizeof(gres->hr_remoteaddr)); 2027 } 2028 } 2029#undef MODIFIED_REMOTEADDR 2030#undef MODIFIED_SOURCEADDR 2031#undef MODIFIED_REPLICATION 2032#undef MODIFIED_CHECKSUM 2033#undef MODIFIED_COMPRESSION 2034#undef MODIFIED_TIMEOUT 2035#undef MODIFIED_EXEC 2036 2037 pjdlog_info("Configuration reloaded successfully."); 2038} 2039 2040static void 2041guard_one(struct hast_resource *res, unsigned int ncomp) 2042{ 2043 struct proto_conn *in, *out; 2044 2045 if (!ISREMOTE(ncomp)) 2046 return; 2047 2048 rw_rlock(&hio_remote_lock[ncomp]); 2049 2050 if (!real_remote(res)) { 2051 rw_unlock(&hio_remote_lock[ncomp]); 2052 return; 2053 } 2054 2055 if (ISCONNECTED(res, ncomp)) { 2056 PJDLOG_ASSERT(res->hr_remotein != NULL); 2057 PJDLOG_ASSERT(res->hr_remoteout != NULL); 2058 rw_unlock(&hio_remote_lock[ncomp]); 2059 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 2060 res->hr_remoteaddr); 2061 return; 2062 } 2063 2064 PJDLOG_ASSERT(res->hr_remotein == NULL); 2065 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2066 /* 2067 * Upgrade the lock. It doesn't have to be atomic as no other thread 2068 * can change connection status from disconnected to connected. 2069 */ 2070 rw_unlock(&hio_remote_lock[ncomp]); 2071 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 2072 res->hr_remoteaddr); 2073 in = out = NULL; 2074 if (init_remote(res, &in, &out)) { 2075 rw_wlock(&hio_remote_lock[ncomp]); 2076 PJDLOG_ASSERT(res->hr_remotein == NULL); 2077 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2078 PJDLOG_ASSERT(in != NULL && out != NULL); 2079 res->hr_remotein = in; 2080 res->hr_remoteout = out; 2081 rw_unlock(&hio_remote_lock[ncomp]); 2082 pjdlog_info("Successfully reconnected to %s.", 2083 res->hr_remoteaddr); 2084 sync_start(); 2085 } else { 2086 /* Both connections should be NULL. */ 2087 PJDLOG_ASSERT(res->hr_remotein == NULL); 2088 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2089 PJDLOG_ASSERT(in == NULL && out == NULL); 2090 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 2091 res->hr_remoteaddr); 2092 } 2093} 2094 2095/* 2096 * Thread guards remote connections and reconnects when needed, handles 2097 * signals, etc. 2098 */ 2099static void * 2100guard_thread(void *arg) 2101{ 2102 struct hast_resource *res = arg; 2103 unsigned int ii, ncomps; 2104 struct timespec timeout; 2105 time_t lastcheck, now; 2106 sigset_t mask; 2107 int signo; 2108 2109 ncomps = HAST_NCOMPONENTS; 2110 lastcheck = time(NULL); 2111 2112 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 2113 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 2114 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 2115 2116 timeout.tv_sec = HAST_KEEPALIVE; 2117 timeout.tv_nsec = 0; 2118 signo = -1; 2119 2120 for (;;) { 2121 switch (signo) { 2122 case SIGINT: 2123 case SIGTERM: 2124 sigexit_received = true; 2125 primary_exitx(EX_OK, 2126 "Termination signal received, exiting."); 2127 break; 2128 default: 2129 break; 2130 } 2131 2132 pjdlog_debug(2, "remote_guard: Checking connections."); 2133 now = time(NULL); 2134 if (lastcheck + HAST_KEEPALIVE <= now) { 2135 for (ii = 0; ii < ncomps; ii++) 2136 guard_one(res, ii); 2137 lastcheck = now; 2138 } 2139 signo = sigtimedwait(&mask, NULL, &timeout); 2140 } 2141 /* NOTREACHED */ 2142 return (NULL); 2143} 2144