primary.c revision 220271
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 220271 2011-04-02 09:25:13Z 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 /* Setup directions. */ 705 if (proto_send(out, NULL, 0) == -1) 706 pjdlog_errno(LOG_WARNING, "Unable to set connection direction"); 707 if (proto_recv(in, NULL, 0) == -1) 708 pjdlog_errno(LOG_WARNING, "Unable to set connection direction"); 709 pjdlog_info("Connected to %s.", res->hr_remoteaddr); 710 if (inp != NULL && outp != NULL) { 711 *inp = in; 712 *outp = out; 713 } else { 714 res->hr_remotein = in; 715 res->hr_remoteout = out; 716 } 717 event_send(res, EVENT_CONNECT); 718 return (true); 719close: 720 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0) 721 event_send(res, EVENT_SPLITBRAIN); 722 proto_close(out); 723 if (in != NULL) 724 proto_close(in); 725 return (false); 726} 727 728static void 729sync_start(void) 730{ 731 732 mtx_lock(&sync_lock); 733 sync_inprogress = true; 734 mtx_unlock(&sync_lock); 735 cv_signal(&sync_cond); 736} 737 738static void 739sync_stop(void) 740{ 741 742 mtx_lock(&sync_lock); 743 if (sync_inprogress) 744 sync_inprogress = false; 745 mtx_unlock(&sync_lock); 746} 747 748static void 749init_ggate(struct hast_resource *res) 750{ 751 struct g_gate_ctl_create ggiocreate; 752 struct g_gate_ctl_cancel ggiocancel; 753 754 /* 755 * We communicate with ggate via /dev/ggctl. Open it. 756 */ 757 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR); 758 if (res->hr_ggatefd < 0) 759 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME); 760 /* 761 * Create provider before trying to connect, as connection failure 762 * is not critical, but may take some time. 763 */ 764 bzero(&ggiocreate, sizeof(ggiocreate)); 765 ggiocreate.gctl_version = G_GATE_VERSION; 766 ggiocreate.gctl_mediasize = res->hr_datasize; 767 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize; 768 ggiocreate.gctl_flags = 0; 769 ggiocreate.gctl_maxcount = 0; 770 ggiocreate.gctl_timeout = 0; 771 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN; 772 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s", 773 res->hr_provname); 774 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) { 775 pjdlog_info("Device hast/%s created.", res->hr_provname); 776 res->hr_ggateunit = ggiocreate.gctl_unit; 777 return; 778 } 779 if (errno != EEXIST) { 780 primary_exit(EX_OSERR, "Unable to create hast/%s device", 781 res->hr_provname); 782 } 783 pjdlog_debug(1, 784 "Device hast/%s already exists, we will try to take it over.", 785 res->hr_provname); 786 /* 787 * If we received EEXIST, we assume that the process who created the 788 * provider died and didn't clean up. In that case we will start from 789 * where he left of. 790 */ 791 bzero(&ggiocancel, sizeof(ggiocancel)); 792 ggiocancel.gctl_version = G_GATE_VERSION; 793 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN; 794 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s", 795 res->hr_provname); 796 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) { 797 pjdlog_info("Device hast/%s recovered.", res->hr_provname); 798 res->hr_ggateunit = ggiocancel.gctl_unit; 799 return; 800 } 801 primary_exit(EX_OSERR, "Unable to take over hast/%s device", 802 res->hr_provname); 803} 804 805void 806hastd_primary(struct hast_resource *res) 807{ 808 pthread_t td; 809 pid_t pid; 810 int error, mode, debuglevel; 811 812 /* 813 * Create communication channel for sending control commands from 814 * parent to child. 815 */ 816 if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) { 817 /* TODO: There's no need for this to be fatal error. */ 818 KEEP_ERRNO((void)pidfile_remove(pfh)); 819 pjdlog_exit(EX_OSERR, 820 "Unable to create control sockets between parent and child"); 821 } 822 /* 823 * Create communication channel for sending events from child to parent. 824 */ 825 if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) { 826 /* TODO: There's no need for this to be fatal error. */ 827 KEEP_ERRNO((void)pidfile_remove(pfh)); 828 pjdlog_exit(EX_OSERR, 829 "Unable to create event sockets between child and parent"); 830 } 831 /* 832 * Create communication channel for sending connection requests from 833 * child to parent. 834 */ 835 if (proto_client(NULL, "socketpair://", &res->hr_conn) < 0) { 836 /* TODO: There's no need for this to be fatal error. */ 837 KEEP_ERRNO((void)pidfile_remove(pfh)); 838 pjdlog_exit(EX_OSERR, 839 "Unable to create connection sockets between child and parent"); 840 } 841 842 pid = fork(); 843 if (pid < 0) { 844 /* TODO: There's no need for this to be fatal error. */ 845 KEEP_ERRNO((void)pidfile_remove(pfh)); 846 pjdlog_exit(EX_TEMPFAIL, "Unable to fork"); 847 } 848 849 if (pid > 0) { 850 /* This is parent. */ 851 /* Declare that we are receiver. */ 852 proto_recv(res->hr_event, NULL, 0); 853 proto_recv(res->hr_conn, NULL, 0); 854 /* Declare that we are sender. */ 855 proto_send(res->hr_ctrl, NULL, 0); 856 res->hr_workerpid = pid; 857 return; 858 } 859 860 gres = res; 861 mode = pjdlog_mode_get(); 862 debuglevel = pjdlog_debug_get(); 863 864 /* Declare that we are sender. */ 865 proto_send(res->hr_event, NULL, 0); 866 proto_send(res->hr_conn, NULL, 0); 867 /* Declare that we are receiver. */ 868 proto_recv(res->hr_ctrl, NULL, 0); 869 descriptors_cleanup(res); 870 871 descriptors_assert(res, mode); 872 873 pjdlog_init(mode); 874 pjdlog_debug_set(debuglevel); 875 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 876 setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role)); 877 878 init_local(res); 879 init_ggate(res); 880 init_environment(res); 881 882 if (drop_privs(true) != 0) { 883 cleanup(res); 884 exit(EX_CONFIG); 885 } 886 pjdlog_info("Privileges successfully dropped."); 887 888 /* 889 * Create the guard thread first, so we can handle signals from the 890 * very begining. 891 */ 892 error = pthread_create(&td, NULL, guard_thread, res); 893 PJDLOG_ASSERT(error == 0); 894 /* 895 * Create the control thread before sending any event to the parent, 896 * as we can deadlock when parent sends control request to worker, 897 * but worker has no control thread started yet, so parent waits. 898 * In the meantime worker sends an event to the parent, but parent 899 * is unable to handle the event, because it waits for control 900 * request response. 901 */ 902 error = pthread_create(&td, NULL, ctrl_thread, res); 903 PJDLOG_ASSERT(error == 0); 904 if (real_remote(res) && init_remote(res, NULL, NULL)) 905 sync_start(); 906 error = pthread_create(&td, NULL, ggate_recv_thread, res); 907 PJDLOG_ASSERT(error == 0); 908 error = pthread_create(&td, NULL, local_send_thread, res); 909 PJDLOG_ASSERT(error == 0); 910 error = pthread_create(&td, NULL, remote_send_thread, res); 911 PJDLOG_ASSERT(error == 0); 912 error = pthread_create(&td, NULL, remote_recv_thread, res); 913 PJDLOG_ASSERT(error == 0); 914 error = pthread_create(&td, NULL, ggate_send_thread, res); 915 PJDLOG_ASSERT(error == 0); 916 (void)sync_thread(res); 917} 918 919static void 920reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...) 921{ 922 char msg[1024]; 923 va_list ap; 924 int len; 925 926 va_start(ap, fmt); 927 len = vsnprintf(msg, sizeof(msg), fmt, ap); 928 va_end(ap); 929 if ((size_t)len < sizeof(msg)) { 930 switch (ggio->gctl_cmd) { 931 case BIO_READ: 932 (void)snprintf(msg + len, sizeof(msg) - len, 933 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 934 (uintmax_t)ggio->gctl_length); 935 break; 936 case BIO_DELETE: 937 (void)snprintf(msg + len, sizeof(msg) - len, 938 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 939 (uintmax_t)ggio->gctl_length); 940 break; 941 case BIO_FLUSH: 942 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 943 break; 944 case BIO_WRITE: 945 (void)snprintf(msg + len, sizeof(msg) - len, 946 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset, 947 (uintmax_t)ggio->gctl_length); 948 break; 949 default: 950 (void)snprintf(msg + len, sizeof(msg) - len, 951 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd); 952 break; 953 } 954 } 955 pjdlog_common(loglevel, debuglevel, -1, "%s", msg); 956} 957 958static void 959remote_close(struct hast_resource *res, int ncomp) 960{ 961 962 rw_wlock(&hio_remote_lock[ncomp]); 963 /* 964 * A race is possible between dropping rlock and acquiring wlock - 965 * another thread can close connection in-between. 966 */ 967 if (!ISCONNECTED(res, ncomp)) { 968 PJDLOG_ASSERT(res->hr_remotein == NULL); 969 PJDLOG_ASSERT(res->hr_remoteout == NULL); 970 rw_unlock(&hio_remote_lock[ncomp]); 971 return; 972 } 973 974 PJDLOG_ASSERT(res->hr_remotein != NULL); 975 PJDLOG_ASSERT(res->hr_remoteout != NULL); 976 977 pjdlog_debug(2, "Closing incoming connection to %s.", 978 res->hr_remoteaddr); 979 proto_close(res->hr_remotein); 980 res->hr_remotein = NULL; 981 pjdlog_debug(2, "Closing outgoing connection to %s.", 982 res->hr_remoteaddr); 983 proto_close(res->hr_remoteout); 984 res->hr_remoteout = NULL; 985 986 rw_unlock(&hio_remote_lock[ncomp]); 987 988 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr); 989 990 /* 991 * Stop synchronization if in-progress. 992 */ 993 sync_stop(); 994 995 event_send(res, EVENT_DISCONNECT); 996} 997 998/* 999 * Thread receives ggate I/O requests from the kernel and passes them to 1000 * appropriate threads: 1001 * WRITE - always goes to both local_send and remote_send threads 1002 * READ (when the block is up-to-date on local component) - 1003 * only local_send thread 1004 * READ (when the block isn't up-to-date on local component) - 1005 * only remote_send thread 1006 * DELETE - always goes to both local_send and remote_send threads 1007 * FLUSH - always goes to both local_send and remote_send threads 1008 */ 1009static void * 1010ggate_recv_thread(void *arg) 1011{ 1012 struct hast_resource *res = arg; 1013 struct g_gate_ctl_io *ggio; 1014 struct hio *hio; 1015 unsigned int ii, ncomp, ncomps; 1016 int error; 1017 1018 ncomps = HAST_NCOMPONENTS; 1019 1020 for (;;) { 1021 pjdlog_debug(2, "ggate_recv: Taking free request."); 1022 QUEUE_TAKE2(hio, free); 1023 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio); 1024 ggio = &hio->hio_ggio; 1025 ggio->gctl_unit = res->hr_ggateunit; 1026 ggio->gctl_length = MAXPHYS; 1027 ggio->gctl_error = 0; 1028 pjdlog_debug(2, 1029 "ggate_recv: (%p) Waiting for request from the kernel.", 1030 hio); 1031 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) { 1032 if (sigexit_received) 1033 pthread_exit(NULL); 1034 primary_exit(EX_OSERR, "G_GATE_CMD_START failed"); 1035 } 1036 error = ggio->gctl_error; 1037 switch (error) { 1038 case 0: 1039 break; 1040 case ECANCELED: 1041 /* Exit gracefully. */ 1042 if (!sigexit_received) { 1043 pjdlog_debug(2, 1044 "ggate_recv: (%p) Received cancel from the kernel.", 1045 hio); 1046 pjdlog_info("Received cancel from the kernel, exiting."); 1047 } 1048 pthread_exit(NULL); 1049 case ENOMEM: 1050 /* 1051 * Buffer too small? Impossible, we allocate MAXPHYS 1052 * bytes - request can't be bigger than that. 1053 */ 1054 /* FALLTHROUGH */ 1055 case ENXIO: 1056 default: 1057 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.", 1058 strerror(error)); 1059 } 1060 for (ii = 0; ii < ncomps; ii++) 1061 hio->hio_errors[ii] = EINVAL; 1062 reqlog(LOG_DEBUG, 2, ggio, 1063 "ggate_recv: (%p) Request received from the kernel: ", 1064 hio); 1065 /* 1066 * Inform all components about new write request. 1067 * For read request prefer local component unless the given 1068 * range is out-of-date, then use remote component. 1069 */ 1070 switch (ggio->gctl_cmd) { 1071 case BIO_READ: 1072 pjdlog_debug(2, 1073 "ggate_recv: (%p) Moving request to the send queue.", 1074 hio); 1075 refcount_init(&hio->hio_countdown, 1); 1076 mtx_lock(&metadata_lock); 1077 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF || 1078 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1079 /* 1080 * This range is up-to-date on local component, 1081 * so handle request locally. 1082 */ 1083 /* Local component is 0 for now. */ 1084 ncomp = 0; 1085 } else /* if (res->hr_syncsrc == 1086 HAST_SYNCSRC_SECONDARY) */ { 1087 PJDLOG_ASSERT(res->hr_syncsrc == 1088 HAST_SYNCSRC_SECONDARY); 1089 /* 1090 * This range is out-of-date on local component, 1091 * so send request to the remote node. 1092 */ 1093 /* Remote component is 1 for now. */ 1094 ncomp = 1; 1095 } 1096 mtx_unlock(&metadata_lock); 1097 QUEUE_INSERT1(hio, send, ncomp); 1098 break; 1099 case BIO_WRITE: 1100 if (res->hr_resuid == 0) { 1101 /* 1102 * This is first write, initialize localcnt and 1103 * resuid. 1104 */ 1105 res->hr_primary_localcnt = 1; 1106 (void)init_resuid(res); 1107 } 1108 for (;;) { 1109 mtx_lock(&range_lock); 1110 if (rangelock_islocked(range_sync, 1111 ggio->gctl_offset, ggio->gctl_length)) { 1112 pjdlog_debug(2, 1113 "regular: Range offset=%jd length=%zu locked.", 1114 (intmax_t)ggio->gctl_offset, 1115 (size_t)ggio->gctl_length); 1116 range_regular_wait = true; 1117 cv_wait(&range_regular_cond, &range_lock); 1118 range_regular_wait = false; 1119 mtx_unlock(&range_lock); 1120 continue; 1121 } 1122 if (rangelock_add(range_regular, 1123 ggio->gctl_offset, ggio->gctl_length) < 0) { 1124 mtx_unlock(&range_lock); 1125 pjdlog_debug(2, 1126 "regular: Range offset=%jd length=%zu is already locked, waiting.", 1127 (intmax_t)ggio->gctl_offset, 1128 (size_t)ggio->gctl_length); 1129 sleep(1); 1130 continue; 1131 } 1132 mtx_unlock(&range_lock); 1133 break; 1134 } 1135 mtx_lock(&res->hr_amp_lock); 1136 if (activemap_write_start(res->hr_amp, 1137 ggio->gctl_offset, ggio->gctl_length)) { 1138 (void)hast_activemap_flush(res); 1139 } 1140 mtx_unlock(&res->hr_amp_lock); 1141 /* FALLTHROUGH */ 1142 case BIO_DELETE: 1143 case BIO_FLUSH: 1144 pjdlog_debug(2, 1145 "ggate_recv: (%p) Moving request to the send queues.", 1146 hio); 1147 refcount_init(&hio->hio_countdown, ncomps); 1148 for (ii = 0; ii < ncomps; ii++) 1149 QUEUE_INSERT1(hio, send, ii); 1150 break; 1151 } 1152 } 1153 /* NOTREACHED */ 1154 return (NULL); 1155} 1156 1157/* 1158 * Thread reads from or writes to local component. 1159 * If local read fails, it redirects it to remote_send thread. 1160 */ 1161static void * 1162local_send_thread(void *arg) 1163{ 1164 struct hast_resource *res = arg; 1165 struct g_gate_ctl_io *ggio; 1166 struct hio *hio; 1167 unsigned int ncomp, rncomp; 1168 ssize_t ret; 1169 1170 /* Local component is 0 for now. */ 1171 ncomp = 0; 1172 /* Remote component is 1 for now. */ 1173 rncomp = 1; 1174 1175 for (;;) { 1176 pjdlog_debug(2, "local_send: Taking request."); 1177 QUEUE_TAKE1(hio, send, ncomp, 0); 1178 pjdlog_debug(2, "local_send: (%p) Got request.", hio); 1179 ggio = &hio->hio_ggio; 1180 switch (ggio->gctl_cmd) { 1181 case BIO_READ: 1182 ret = pread(res->hr_localfd, ggio->gctl_data, 1183 ggio->gctl_length, 1184 ggio->gctl_offset + res->hr_localoff); 1185 if (ret == ggio->gctl_length) 1186 hio->hio_errors[ncomp] = 0; 1187 else { 1188 /* 1189 * If READ failed, try to read from remote node. 1190 */ 1191 if (ret < 0) { 1192 reqlog(LOG_WARNING, 0, ggio, 1193 "Local request failed (%s), trying remote node. ", 1194 strerror(errno)); 1195 } else if (ret != ggio->gctl_length) { 1196 reqlog(LOG_WARNING, 0, ggio, 1197 "Local request failed (%zd != %jd), trying remote node. ", 1198 ret, (intmax_t)ggio->gctl_length); 1199 } 1200 QUEUE_INSERT1(hio, send, rncomp); 1201 continue; 1202 } 1203 break; 1204 case BIO_WRITE: 1205 ret = pwrite(res->hr_localfd, ggio->gctl_data, 1206 ggio->gctl_length, 1207 ggio->gctl_offset + res->hr_localoff); 1208 if (ret < 0) { 1209 hio->hio_errors[ncomp] = errno; 1210 reqlog(LOG_WARNING, 0, ggio, 1211 "Local request failed (%s): ", 1212 strerror(errno)); 1213 } else if (ret != ggio->gctl_length) { 1214 hio->hio_errors[ncomp] = EIO; 1215 reqlog(LOG_WARNING, 0, ggio, 1216 "Local request failed (%zd != %jd): ", 1217 ret, (intmax_t)ggio->gctl_length); 1218 } else { 1219 hio->hio_errors[ncomp] = 0; 1220 } 1221 break; 1222 case BIO_DELETE: 1223 ret = g_delete(res->hr_localfd, 1224 ggio->gctl_offset + res->hr_localoff, 1225 ggio->gctl_length); 1226 if (ret < 0) { 1227 hio->hio_errors[ncomp] = errno; 1228 reqlog(LOG_WARNING, 0, ggio, 1229 "Local request failed (%s): ", 1230 strerror(errno)); 1231 } else { 1232 hio->hio_errors[ncomp] = 0; 1233 } 1234 break; 1235 case BIO_FLUSH: 1236 ret = g_flush(res->hr_localfd); 1237 if (ret < 0) { 1238 hio->hio_errors[ncomp] = errno; 1239 reqlog(LOG_WARNING, 0, ggio, 1240 "Local request failed (%s): ", 1241 strerror(errno)); 1242 } else { 1243 hio->hio_errors[ncomp] = 0; 1244 } 1245 break; 1246 } 1247 if (refcount_release(&hio->hio_countdown)) { 1248 if (ISSYNCREQ(hio)) { 1249 mtx_lock(&sync_lock); 1250 SYNCREQDONE(hio); 1251 mtx_unlock(&sync_lock); 1252 cv_signal(&sync_cond); 1253 } else { 1254 pjdlog_debug(2, 1255 "local_send: (%p) Moving request to the done queue.", 1256 hio); 1257 QUEUE_INSERT2(hio, done); 1258 } 1259 } 1260 } 1261 /* NOTREACHED */ 1262 return (NULL); 1263} 1264 1265static void 1266keepalive_send(struct hast_resource *res, unsigned int ncomp) 1267{ 1268 struct nv *nv; 1269 1270 rw_rlock(&hio_remote_lock[ncomp]); 1271 1272 if (!ISCONNECTED(res, ncomp)) { 1273 rw_unlock(&hio_remote_lock[ncomp]); 1274 return; 1275 } 1276 1277 PJDLOG_ASSERT(res->hr_remotein != NULL); 1278 PJDLOG_ASSERT(res->hr_remoteout != NULL); 1279 1280 nv = nv_alloc(); 1281 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd"); 1282 if (nv_error(nv) != 0) { 1283 rw_unlock(&hio_remote_lock[ncomp]); 1284 nv_free(nv); 1285 pjdlog_debug(1, 1286 "keepalive_send: Unable to prepare header to send."); 1287 return; 1288 } 1289 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) { 1290 rw_unlock(&hio_remote_lock[ncomp]); 1291 pjdlog_common(LOG_DEBUG, 1, errno, 1292 "keepalive_send: Unable to send request"); 1293 nv_free(nv); 1294 remote_close(res, ncomp); 1295 return; 1296 } 1297 1298 rw_unlock(&hio_remote_lock[ncomp]); 1299 nv_free(nv); 1300 pjdlog_debug(2, "keepalive_send: Request sent."); 1301} 1302 1303/* 1304 * Thread sends request to secondary node. 1305 */ 1306static void * 1307remote_send_thread(void *arg) 1308{ 1309 struct hast_resource *res = arg; 1310 struct g_gate_ctl_io *ggio; 1311 time_t lastcheck, now; 1312 struct hio *hio; 1313 struct nv *nv; 1314 unsigned int ncomp; 1315 bool wakeup; 1316 uint64_t offset, length; 1317 uint8_t cmd; 1318 void *data; 1319 1320 /* Remote component is 1 for now. */ 1321 ncomp = 1; 1322 lastcheck = time(NULL); 1323 1324 for (;;) { 1325 pjdlog_debug(2, "remote_send: Taking request."); 1326 QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE); 1327 if (hio == NULL) { 1328 now = time(NULL); 1329 if (lastcheck + HAST_KEEPALIVE <= now) { 1330 keepalive_send(res, ncomp); 1331 lastcheck = now; 1332 } 1333 continue; 1334 } 1335 pjdlog_debug(2, "remote_send: (%p) Got request.", hio); 1336 ggio = &hio->hio_ggio; 1337 switch (ggio->gctl_cmd) { 1338 case BIO_READ: 1339 cmd = HIO_READ; 1340 data = NULL; 1341 offset = ggio->gctl_offset; 1342 length = ggio->gctl_length; 1343 break; 1344 case BIO_WRITE: 1345 cmd = HIO_WRITE; 1346 data = ggio->gctl_data; 1347 offset = ggio->gctl_offset; 1348 length = ggio->gctl_length; 1349 break; 1350 case BIO_DELETE: 1351 cmd = HIO_DELETE; 1352 data = NULL; 1353 offset = ggio->gctl_offset; 1354 length = ggio->gctl_length; 1355 break; 1356 case BIO_FLUSH: 1357 cmd = HIO_FLUSH; 1358 data = NULL; 1359 offset = 0; 1360 length = 0; 1361 break; 1362 default: 1363 PJDLOG_ASSERT(!"invalid condition"); 1364 abort(); 1365 } 1366 nv = nv_alloc(); 1367 nv_add_uint8(nv, cmd, "cmd"); 1368 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq"); 1369 nv_add_uint64(nv, offset, "offset"); 1370 nv_add_uint64(nv, length, "length"); 1371 if (nv_error(nv) != 0) { 1372 hio->hio_errors[ncomp] = nv_error(nv); 1373 pjdlog_debug(2, 1374 "remote_send: (%p) Unable to prepare header to send.", 1375 hio); 1376 reqlog(LOG_ERR, 0, ggio, 1377 "Unable to prepare header to send (%s): ", 1378 strerror(nv_error(nv))); 1379 /* Move failed request immediately to the done queue. */ 1380 goto done_queue; 1381 } 1382 pjdlog_debug(2, 1383 "remote_send: (%p) Moving request to the recv queue.", 1384 hio); 1385 /* 1386 * Protect connection from disappearing. 1387 */ 1388 rw_rlock(&hio_remote_lock[ncomp]); 1389 if (!ISCONNECTED(res, ncomp)) { 1390 rw_unlock(&hio_remote_lock[ncomp]); 1391 hio->hio_errors[ncomp] = ENOTCONN; 1392 goto done_queue; 1393 } 1394 /* 1395 * Move the request to recv queue before sending it, because 1396 * in different order we can get reply before we move request 1397 * to recv queue. 1398 */ 1399 mtx_lock(&hio_recv_list_lock[ncomp]); 1400 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]); 1401 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1402 mtx_unlock(&hio_recv_list_lock[ncomp]); 1403 if (hast_proto_send(res, res->hr_remoteout, nv, data, 1404 data != NULL ? length : 0) < 0) { 1405 hio->hio_errors[ncomp] = errno; 1406 rw_unlock(&hio_remote_lock[ncomp]); 1407 pjdlog_debug(2, 1408 "remote_send: (%p) Unable to send request.", hio); 1409 reqlog(LOG_ERR, 0, ggio, 1410 "Unable to send request (%s): ", 1411 strerror(hio->hio_errors[ncomp])); 1412 remote_close(res, ncomp); 1413 /* 1414 * Take request back from the receive queue and move 1415 * it immediately to the done queue. 1416 */ 1417 mtx_lock(&hio_recv_list_lock[ncomp]); 1418 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]); 1419 mtx_unlock(&hio_recv_list_lock[ncomp]); 1420 goto done_queue; 1421 } 1422 rw_unlock(&hio_remote_lock[ncomp]); 1423 nv_free(nv); 1424 if (wakeup) 1425 cv_signal(&hio_recv_list_cond[ncomp]); 1426 continue; 1427done_queue: 1428 nv_free(nv); 1429 if (ISSYNCREQ(hio)) { 1430 if (!refcount_release(&hio->hio_countdown)) 1431 continue; 1432 mtx_lock(&sync_lock); 1433 SYNCREQDONE(hio); 1434 mtx_unlock(&sync_lock); 1435 cv_signal(&sync_cond); 1436 continue; 1437 } 1438 if (ggio->gctl_cmd == BIO_WRITE) { 1439 mtx_lock(&res->hr_amp_lock); 1440 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset, 1441 ggio->gctl_length)) { 1442 (void)hast_activemap_flush(res); 1443 } 1444 mtx_unlock(&res->hr_amp_lock); 1445 } 1446 if (!refcount_release(&hio->hio_countdown)) 1447 continue; 1448 pjdlog_debug(2, 1449 "remote_send: (%p) Moving request to the done queue.", 1450 hio); 1451 QUEUE_INSERT2(hio, done); 1452 } 1453 /* NOTREACHED */ 1454 return (NULL); 1455} 1456 1457/* 1458 * Thread receives answer from secondary node and passes it to ggate_send 1459 * thread. 1460 */ 1461static void * 1462remote_recv_thread(void *arg) 1463{ 1464 struct hast_resource *res = arg; 1465 struct g_gate_ctl_io *ggio; 1466 struct hio *hio; 1467 struct nv *nv; 1468 unsigned int ncomp; 1469 uint64_t seq; 1470 int error; 1471 1472 /* Remote component is 1 for now. */ 1473 ncomp = 1; 1474 1475 for (;;) { 1476 /* Wait until there is anything to receive. */ 1477 mtx_lock(&hio_recv_list_lock[ncomp]); 1478 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) { 1479 pjdlog_debug(2, "remote_recv: No requests, waiting."); 1480 cv_wait(&hio_recv_list_cond[ncomp], 1481 &hio_recv_list_lock[ncomp]); 1482 } 1483 mtx_unlock(&hio_recv_list_lock[ncomp]); 1484 rw_rlock(&hio_remote_lock[ncomp]); 1485 if (!ISCONNECTED(res, ncomp)) { 1486 rw_unlock(&hio_remote_lock[ncomp]); 1487 /* 1488 * Connection is dead, so move all pending requests to 1489 * the done queue (one-by-one). 1490 */ 1491 mtx_lock(&hio_recv_list_lock[ncomp]); 1492 hio = TAILQ_FIRST(&hio_recv_list[ncomp]); 1493 PJDLOG_ASSERT(hio != NULL); 1494 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1495 hio_next[ncomp]); 1496 mtx_unlock(&hio_recv_list_lock[ncomp]); 1497 goto done_queue; 1498 } 1499 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) { 1500 pjdlog_errno(LOG_ERR, 1501 "Unable to receive reply header"); 1502 rw_unlock(&hio_remote_lock[ncomp]); 1503 remote_close(res, ncomp); 1504 continue; 1505 } 1506 rw_unlock(&hio_remote_lock[ncomp]); 1507 seq = nv_get_uint64(nv, "seq"); 1508 if (seq == 0) { 1509 pjdlog_error("Header contains no 'seq' field."); 1510 nv_free(nv); 1511 continue; 1512 } 1513 mtx_lock(&hio_recv_list_lock[ncomp]); 1514 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) { 1515 if (hio->hio_ggio.gctl_seq == seq) { 1516 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, 1517 hio_next[ncomp]); 1518 break; 1519 } 1520 } 1521 mtx_unlock(&hio_recv_list_lock[ncomp]); 1522 if (hio == NULL) { 1523 pjdlog_error("Found no request matching received 'seq' field (%ju).", 1524 (uintmax_t)seq); 1525 nv_free(nv); 1526 continue; 1527 } 1528 error = nv_get_int16(nv, "error"); 1529 if (error != 0) { 1530 /* Request failed on remote side. */ 1531 hio->hio_errors[ncomp] = error; 1532 reqlog(LOG_WARNING, 0, &hio->hio_ggio, 1533 "Remote request failed (%s): ", strerror(error)); 1534 nv_free(nv); 1535 goto done_queue; 1536 } 1537 ggio = &hio->hio_ggio; 1538 switch (ggio->gctl_cmd) { 1539 case BIO_READ: 1540 rw_rlock(&hio_remote_lock[ncomp]); 1541 if (!ISCONNECTED(res, ncomp)) { 1542 rw_unlock(&hio_remote_lock[ncomp]); 1543 nv_free(nv); 1544 goto done_queue; 1545 } 1546 if (hast_proto_recv_data(res, res->hr_remotein, nv, 1547 ggio->gctl_data, ggio->gctl_length) < 0) { 1548 hio->hio_errors[ncomp] = errno; 1549 pjdlog_errno(LOG_ERR, 1550 "Unable to receive reply data"); 1551 rw_unlock(&hio_remote_lock[ncomp]); 1552 nv_free(nv); 1553 remote_close(res, ncomp); 1554 goto done_queue; 1555 } 1556 rw_unlock(&hio_remote_lock[ncomp]); 1557 break; 1558 case BIO_WRITE: 1559 case BIO_DELETE: 1560 case BIO_FLUSH: 1561 break; 1562 default: 1563 PJDLOG_ASSERT(!"invalid condition"); 1564 abort(); 1565 } 1566 hio->hio_errors[ncomp] = 0; 1567 nv_free(nv); 1568done_queue: 1569 if (refcount_release(&hio->hio_countdown)) { 1570 if (ISSYNCREQ(hio)) { 1571 mtx_lock(&sync_lock); 1572 SYNCREQDONE(hio); 1573 mtx_unlock(&sync_lock); 1574 cv_signal(&sync_cond); 1575 } else { 1576 pjdlog_debug(2, 1577 "remote_recv: (%p) Moving request to the done queue.", 1578 hio); 1579 QUEUE_INSERT2(hio, done); 1580 } 1581 } 1582 } 1583 /* NOTREACHED */ 1584 return (NULL); 1585} 1586 1587/* 1588 * Thread sends answer to the kernel. 1589 */ 1590static void * 1591ggate_send_thread(void *arg) 1592{ 1593 struct hast_resource *res = arg; 1594 struct g_gate_ctl_io *ggio; 1595 struct hio *hio; 1596 unsigned int ii, ncomp, ncomps; 1597 1598 ncomps = HAST_NCOMPONENTS; 1599 1600 for (;;) { 1601 pjdlog_debug(2, "ggate_send: Taking request."); 1602 QUEUE_TAKE2(hio, done); 1603 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio); 1604 ggio = &hio->hio_ggio; 1605 for (ii = 0; ii < ncomps; ii++) { 1606 if (hio->hio_errors[ii] == 0) { 1607 /* 1608 * One successful request is enough to declare 1609 * success. 1610 */ 1611 ggio->gctl_error = 0; 1612 break; 1613 } 1614 } 1615 if (ii == ncomps) { 1616 /* 1617 * None of the requests were successful. 1618 * Use the error from local component except the 1619 * case when we did only remote request. 1620 */ 1621 if (ggio->gctl_cmd == BIO_READ && 1622 res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) 1623 ggio->gctl_error = hio->hio_errors[1]; 1624 else 1625 ggio->gctl_error = hio->hio_errors[0]; 1626 } 1627 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) { 1628 mtx_lock(&res->hr_amp_lock); 1629 activemap_write_complete(res->hr_amp, 1630 ggio->gctl_offset, ggio->gctl_length); 1631 mtx_unlock(&res->hr_amp_lock); 1632 } 1633 if (ggio->gctl_cmd == BIO_WRITE) { 1634 /* 1635 * Unlock range we locked. 1636 */ 1637 mtx_lock(&range_lock); 1638 rangelock_del(range_regular, ggio->gctl_offset, 1639 ggio->gctl_length); 1640 if (range_sync_wait) 1641 cv_signal(&range_sync_cond); 1642 mtx_unlock(&range_lock); 1643 /* 1644 * Bump local count if this is first write after 1645 * connection failure with remote node. 1646 */ 1647 ncomp = 1; 1648 rw_rlock(&hio_remote_lock[ncomp]); 1649 if (!ISCONNECTED(res, ncomp)) { 1650 mtx_lock(&metadata_lock); 1651 if (res->hr_primary_localcnt == 1652 res->hr_secondary_remotecnt) { 1653 res->hr_primary_localcnt++; 1654 pjdlog_debug(1, 1655 "Increasing localcnt to %ju.", 1656 (uintmax_t)res->hr_primary_localcnt); 1657 (void)metadata_write(res); 1658 } 1659 mtx_unlock(&metadata_lock); 1660 } 1661 rw_unlock(&hio_remote_lock[ncomp]); 1662 } 1663 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0) 1664 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed"); 1665 pjdlog_debug(2, 1666 "ggate_send: (%p) Moving request to the free queue.", hio); 1667 QUEUE_INSERT2(hio, free); 1668 } 1669 /* NOTREACHED */ 1670 return (NULL); 1671} 1672 1673/* 1674 * Thread synchronize local and remote components. 1675 */ 1676static void * 1677sync_thread(void *arg __unused) 1678{ 1679 struct hast_resource *res = arg; 1680 struct hio *hio; 1681 struct g_gate_ctl_io *ggio; 1682 struct timeval tstart, tend, tdiff; 1683 unsigned int ii, ncomp, ncomps; 1684 off_t offset, length, synced; 1685 bool dorewind; 1686 int syncext; 1687 1688 ncomps = HAST_NCOMPONENTS; 1689 dorewind = true; 1690 synced = 0; 1691 offset = -1; 1692 1693 for (;;) { 1694 mtx_lock(&sync_lock); 1695 if (offset >= 0 && !sync_inprogress) { 1696 gettimeofday(&tend, NULL); 1697 timersub(&tend, &tstart, &tdiff); 1698 pjdlog_info("Synchronization interrupted after %#.0T. " 1699 "%NB synchronized so far.", &tdiff, 1700 (intmax_t)synced); 1701 event_send(res, EVENT_SYNCINTR); 1702 } 1703 while (!sync_inprogress) { 1704 dorewind = true; 1705 synced = 0; 1706 cv_wait(&sync_cond, &sync_lock); 1707 } 1708 mtx_unlock(&sync_lock); 1709 /* 1710 * Obtain offset at which we should synchronize. 1711 * Rewind synchronization if needed. 1712 */ 1713 mtx_lock(&res->hr_amp_lock); 1714 if (dorewind) 1715 activemap_sync_rewind(res->hr_amp); 1716 offset = activemap_sync_offset(res->hr_amp, &length, &syncext); 1717 if (syncext != -1) { 1718 /* 1719 * We synchronized entire syncext extent, we can mark 1720 * it as clean now. 1721 */ 1722 if (activemap_extent_complete(res->hr_amp, syncext)) 1723 (void)hast_activemap_flush(res); 1724 } 1725 mtx_unlock(&res->hr_amp_lock); 1726 if (dorewind) { 1727 dorewind = false; 1728 if (offset < 0) 1729 pjdlog_info("Nodes are in sync."); 1730 else { 1731 pjdlog_info("Synchronization started. %NB to go.", 1732 (intmax_t)(res->hr_extentsize * 1733 activemap_ndirty(res->hr_amp))); 1734 event_send(res, EVENT_SYNCSTART); 1735 gettimeofday(&tstart, NULL); 1736 } 1737 } 1738 if (offset < 0) { 1739 sync_stop(); 1740 pjdlog_debug(1, "Nothing to synchronize."); 1741 /* 1742 * Synchronization complete, make both localcnt and 1743 * remotecnt equal. 1744 */ 1745 ncomp = 1; 1746 rw_rlock(&hio_remote_lock[ncomp]); 1747 if (ISCONNECTED(res, ncomp)) { 1748 if (synced > 0) { 1749 int64_t bps; 1750 1751 gettimeofday(&tend, NULL); 1752 timersub(&tend, &tstart, &tdiff); 1753 bps = (int64_t)((double)synced / 1754 ((double)tdiff.tv_sec + 1755 (double)tdiff.tv_usec / 1000000)); 1756 pjdlog_info("Synchronization complete. " 1757 "%NB synchronized in %#.0lT (%NB/sec).", 1758 (intmax_t)synced, &tdiff, 1759 (intmax_t)bps); 1760 event_send(res, EVENT_SYNCDONE); 1761 } 1762 mtx_lock(&metadata_lock); 1763 res->hr_syncsrc = HAST_SYNCSRC_UNDEF; 1764 res->hr_primary_localcnt = 1765 res->hr_secondary_remotecnt; 1766 res->hr_primary_remotecnt = 1767 res->hr_secondary_localcnt; 1768 pjdlog_debug(1, 1769 "Setting localcnt to %ju and remotecnt to %ju.", 1770 (uintmax_t)res->hr_primary_localcnt, 1771 (uintmax_t)res->hr_primary_remotecnt); 1772 (void)metadata_write(res); 1773 mtx_unlock(&metadata_lock); 1774 } 1775 rw_unlock(&hio_remote_lock[ncomp]); 1776 continue; 1777 } 1778 pjdlog_debug(2, "sync: Taking free request."); 1779 QUEUE_TAKE2(hio, free); 1780 pjdlog_debug(2, "sync: (%p) Got free request.", hio); 1781 /* 1782 * Lock the range we are going to synchronize. We don't want 1783 * race where someone writes between our read and write. 1784 */ 1785 for (;;) { 1786 mtx_lock(&range_lock); 1787 if (rangelock_islocked(range_regular, offset, length)) { 1788 pjdlog_debug(2, 1789 "sync: Range offset=%jd length=%jd locked.", 1790 (intmax_t)offset, (intmax_t)length); 1791 range_sync_wait = true; 1792 cv_wait(&range_sync_cond, &range_lock); 1793 range_sync_wait = false; 1794 mtx_unlock(&range_lock); 1795 continue; 1796 } 1797 if (rangelock_add(range_sync, offset, length) < 0) { 1798 mtx_unlock(&range_lock); 1799 pjdlog_debug(2, 1800 "sync: Range offset=%jd length=%jd is already locked, waiting.", 1801 (intmax_t)offset, (intmax_t)length); 1802 sleep(1); 1803 continue; 1804 } 1805 mtx_unlock(&range_lock); 1806 break; 1807 } 1808 /* 1809 * First read the data from synchronization source. 1810 */ 1811 SYNCREQ(hio); 1812 ggio = &hio->hio_ggio; 1813 ggio->gctl_cmd = BIO_READ; 1814 ggio->gctl_offset = offset; 1815 ggio->gctl_length = length; 1816 ggio->gctl_error = 0; 1817 for (ii = 0; ii < ncomps; ii++) 1818 hio->hio_errors[ii] = EINVAL; 1819 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1820 hio); 1821 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1822 hio); 1823 mtx_lock(&metadata_lock); 1824 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1825 /* 1826 * This range is up-to-date on local component, 1827 * so handle request locally. 1828 */ 1829 /* Local component is 0 for now. */ 1830 ncomp = 0; 1831 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1832 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1833 /* 1834 * This range is out-of-date on local component, 1835 * so send request to the remote node. 1836 */ 1837 /* Remote component is 1 for now. */ 1838 ncomp = 1; 1839 } 1840 mtx_unlock(&metadata_lock); 1841 refcount_init(&hio->hio_countdown, 1); 1842 QUEUE_INSERT1(hio, send, ncomp); 1843 1844 /* 1845 * Let's wait for READ to finish. 1846 */ 1847 mtx_lock(&sync_lock); 1848 while (!ISSYNCREQDONE(hio)) 1849 cv_wait(&sync_cond, &sync_lock); 1850 mtx_unlock(&sync_lock); 1851 1852 if (hio->hio_errors[ncomp] != 0) { 1853 pjdlog_error("Unable to read synchronization data: %s.", 1854 strerror(hio->hio_errors[ncomp])); 1855 goto free_queue; 1856 } 1857 1858 /* 1859 * We read the data from synchronization source, now write it 1860 * to synchronization target. 1861 */ 1862 SYNCREQ(hio); 1863 ggio->gctl_cmd = BIO_WRITE; 1864 for (ii = 0; ii < ncomps; ii++) 1865 hio->hio_errors[ii] = EINVAL; 1866 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ", 1867 hio); 1868 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.", 1869 hio); 1870 mtx_lock(&metadata_lock); 1871 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) { 1872 /* 1873 * This range is up-to-date on local component, 1874 * so we update remote component. 1875 */ 1876 /* Remote component is 1 for now. */ 1877 ncomp = 1; 1878 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ { 1879 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY); 1880 /* 1881 * This range is out-of-date on local component, 1882 * so we update it. 1883 */ 1884 /* Local component is 0 for now. */ 1885 ncomp = 0; 1886 } 1887 mtx_unlock(&metadata_lock); 1888 1889 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.", 1890 hio); 1891 refcount_init(&hio->hio_countdown, 1); 1892 QUEUE_INSERT1(hio, send, ncomp); 1893 1894 /* 1895 * Let's wait for WRITE to finish. 1896 */ 1897 mtx_lock(&sync_lock); 1898 while (!ISSYNCREQDONE(hio)) 1899 cv_wait(&sync_cond, &sync_lock); 1900 mtx_unlock(&sync_lock); 1901 1902 if (hio->hio_errors[ncomp] != 0) { 1903 pjdlog_error("Unable to write synchronization data: %s.", 1904 strerror(hio->hio_errors[ncomp])); 1905 goto free_queue; 1906 } 1907 1908 synced += length; 1909free_queue: 1910 mtx_lock(&range_lock); 1911 rangelock_del(range_sync, offset, length); 1912 if (range_regular_wait) 1913 cv_signal(&range_regular_cond); 1914 mtx_unlock(&range_lock); 1915 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.", 1916 hio); 1917 QUEUE_INSERT2(hio, free); 1918 } 1919 /* NOTREACHED */ 1920 return (NULL); 1921} 1922 1923void 1924primary_config_reload(struct hast_resource *res, struct nv *nv) 1925{ 1926 unsigned int ii, ncomps; 1927 int modified, vint; 1928 const char *vstr; 1929 1930 pjdlog_info("Reloading configuration..."); 1931 1932 PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY); 1933 PJDLOG_ASSERT(gres == res); 1934 nv_assert(nv, "remoteaddr"); 1935 nv_assert(nv, "sourceaddr"); 1936 nv_assert(nv, "replication"); 1937 nv_assert(nv, "checksum"); 1938 nv_assert(nv, "compression"); 1939 nv_assert(nv, "timeout"); 1940 nv_assert(nv, "exec"); 1941 1942 ncomps = HAST_NCOMPONENTS; 1943 1944#define MODIFIED_REMOTEADDR 0x01 1945#define MODIFIED_SOURCEADDR 0x02 1946#define MODIFIED_REPLICATION 0x04 1947#define MODIFIED_CHECKSUM 0x08 1948#define MODIFIED_COMPRESSION 0x10 1949#define MODIFIED_TIMEOUT 0x20 1950#define MODIFIED_EXEC 0x40 1951 modified = 0; 1952 1953 vstr = nv_get_string(nv, "remoteaddr"); 1954 if (strcmp(gres->hr_remoteaddr, vstr) != 0) { 1955 /* 1956 * Don't copy res->hr_remoteaddr to gres just yet. 1957 * We want remote_close() to log disconnect from the old 1958 * addresses, not from the new ones. 1959 */ 1960 modified |= MODIFIED_REMOTEADDR; 1961 } 1962 vstr = nv_get_string(nv, "sourceaddr"); 1963 if (strcmp(gres->hr_sourceaddr, vstr) != 0) { 1964 strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr)); 1965 modified |= MODIFIED_SOURCEADDR; 1966 } 1967 vint = nv_get_int32(nv, "replication"); 1968 if (gres->hr_replication != vint) { 1969 gres->hr_replication = vint; 1970 modified |= MODIFIED_REPLICATION; 1971 } 1972 vint = nv_get_int32(nv, "checksum"); 1973 if (gres->hr_checksum != vint) { 1974 gres->hr_checksum = vint; 1975 modified |= MODIFIED_CHECKSUM; 1976 } 1977 vint = nv_get_int32(nv, "compression"); 1978 if (gres->hr_compression != vint) { 1979 gres->hr_compression = vint; 1980 modified |= MODIFIED_COMPRESSION; 1981 } 1982 vint = nv_get_int32(nv, "timeout"); 1983 if (gres->hr_timeout != vint) { 1984 gres->hr_timeout = vint; 1985 modified |= MODIFIED_TIMEOUT; 1986 } 1987 vstr = nv_get_string(nv, "exec"); 1988 if (strcmp(gres->hr_exec, vstr) != 0) { 1989 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec)); 1990 modified |= MODIFIED_EXEC; 1991 } 1992 1993 /* 1994 * Change timeout for connected sockets. 1995 * Don't bother if we need to reconnect. 1996 */ 1997 if ((modified & MODIFIED_TIMEOUT) != 0 && 1998 (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR | 1999 MODIFIED_REPLICATION)) == 0) { 2000 for (ii = 0; ii < ncomps; ii++) { 2001 if (!ISREMOTE(ii)) 2002 continue; 2003 rw_rlock(&hio_remote_lock[ii]); 2004 if (!ISCONNECTED(gres, ii)) { 2005 rw_unlock(&hio_remote_lock[ii]); 2006 continue; 2007 } 2008 rw_unlock(&hio_remote_lock[ii]); 2009 if (proto_timeout(gres->hr_remotein, 2010 gres->hr_timeout) < 0) { 2011 pjdlog_errno(LOG_WARNING, 2012 "Unable to set connection timeout"); 2013 } 2014 if (proto_timeout(gres->hr_remoteout, 2015 gres->hr_timeout) < 0) { 2016 pjdlog_errno(LOG_WARNING, 2017 "Unable to set connection timeout"); 2018 } 2019 } 2020 } 2021 if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR | 2022 MODIFIED_REPLICATION)) != 0) { 2023 for (ii = 0; ii < ncomps; ii++) { 2024 if (!ISREMOTE(ii)) 2025 continue; 2026 remote_close(gres, ii); 2027 } 2028 if (modified & MODIFIED_REMOTEADDR) { 2029 vstr = nv_get_string(nv, "remoteaddr"); 2030 strlcpy(gres->hr_remoteaddr, vstr, 2031 sizeof(gres->hr_remoteaddr)); 2032 } 2033 } 2034#undef MODIFIED_REMOTEADDR 2035#undef MODIFIED_SOURCEADDR 2036#undef MODIFIED_REPLICATION 2037#undef MODIFIED_CHECKSUM 2038#undef MODIFIED_COMPRESSION 2039#undef MODIFIED_TIMEOUT 2040#undef MODIFIED_EXEC 2041 2042 pjdlog_info("Configuration reloaded successfully."); 2043} 2044 2045static void 2046guard_one(struct hast_resource *res, unsigned int ncomp) 2047{ 2048 struct proto_conn *in, *out; 2049 2050 if (!ISREMOTE(ncomp)) 2051 return; 2052 2053 rw_rlock(&hio_remote_lock[ncomp]); 2054 2055 if (!real_remote(res)) { 2056 rw_unlock(&hio_remote_lock[ncomp]); 2057 return; 2058 } 2059 2060 if (ISCONNECTED(res, ncomp)) { 2061 PJDLOG_ASSERT(res->hr_remotein != NULL); 2062 PJDLOG_ASSERT(res->hr_remoteout != NULL); 2063 rw_unlock(&hio_remote_lock[ncomp]); 2064 pjdlog_debug(2, "remote_guard: Connection to %s is ok.", 2065 res->hr_remoteaddr); 2066 return; 2067 } 2068 2069 PJDLOG_ASSERT(res->hr_remotein == NULL); 2070 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2071 /* 2072 * Upgrade the lock. It doesn't have to be atomic as no other thread 2073 * can change connection status from disconnected to connected. 2074 */ 2075 rw_unlock(&hio_remote_lock[ncomp]); 2076 pjdlog_debug(2, "remote_guard: Reconnecting to %s.", 2077 res->hr_remoteaddr); 2078 in = out = NULL; 2079 if (init_remote(res, &in, &out)) { 2080 rw_wlock(&hio_remote_lock[ncomp]); 2081 PJDLOG_ASSERT(res->hr_remotein == NULL); 2082 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2083 PJDLOG_ASSERT(in != NULL && out != NULL); 2084 res->hr_remotein = in; 2085 res->hr_remoteout = out; 2086 rw_unlock(&hio_remote_lock[ncomp]); 2087 pjdlog_info("Successfully reconnected to %s.", 2088 res->hr_remoteaddr); 2089 sync_start(); 2090 } else { 2091 /* Both connections should be NULL. */ 2092 PJDLOG_ASSERT(res->hr_remotein == NULL); 2093 PJDLOG_ASSERT(res->hr_remoteout == NULL); 2094 PJDLOG_ASSERT(in == NULL && out == NULL); 2095 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.", 2096 res->hr_remoteaddr); 2097 } 2098} 2099 2100/* 2101 * Thread guards remote connections and reconnects when needed, handles 2102 * signals, etc. 2103 */ 2104static void * 2105guard_thread(void *arg) 2106{ 2107 struct hast_resource *res = arg; 2108 unsigned int ii, ncomps; 2109 struct timespec timeout; 2110 time_t lastcheck, now; 2111 sigset_t mask; 2112 int signo; 2113 2114 ncomps = HAST_NCOMPONENTS; 2115 lastcheck = time(NULL); 2116 2117 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 2118 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0); 2119 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0); 2120 2121 timeout.tv_sec = HAST_KEEPALIVE; 2122 timeout.tv_nsec = 0; 2123 signo = -1; 2124 2125 for (;;) { 2126 switch (signo) { 2127 case SIGINT: 2128 case SIGTERM: 2129 sigexit_received = true; 2130 primary_exitx(EX_OK, 2131 "Termination signal received, exiting."); 2132 break; 2133 default: 2134 break; 2135 } 2136 2137 pjdlog_debug(2, "remote_guard: Checking connections."); 2138 now = time(NULL); 2139 if (lastcheck + HAST_KEEPALIVE <= now) { 2140 for (ii = 0; ii < ncomps; ii++) 2141 guard_one(res, ii); 2142 lastcheck = now; 2143 } 2144 signo = sigtimedwait(&mask, NULL, &timeout); 2145 } 2146 /* NOTREACHED */ 2147 return (NULL); 2148} 2149