sbin/hastd/primary.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   3  *
   4  * Copyright (c) 2009 The FreeBSD Foundation
   5  * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net>
   6  * All rights reserved.
   7  *
   8  * This software was developed by Pawel Jakub Dawidek under sponsorship from
   9  * the FreeBSD Foundation.
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  * 1. Redistributions of source code must retain the above copyright
  15  *    notice, this list of conditions and the following disclaimer.
  16  * 2. Redistributions in binary form must reproduce the above copyright
  17  *    notice, this list of conditions and the following disclaimer in the
  18  *    documentation and/or other materials provided with the distribution.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  */
  32
  33 #include <sys/types.h>
  34 #include <sys/time.h>
  35 #include <sys/bio.h>
  36 #include <sys/disk.h>
  37 #include <sys/stat.h>
  38
  39 #include <geom/gate/g_gate.h>
  40
  41 #include <err.h>
  42 #include <errno.h>
  43 #include <fcntl.h>
  44 #include <libgeom.h>
  45 #include <pthread.h>
  46 #include <signal.h>
  47 #include <stdint.h>
  48 #include <stdio.h>
  49 #include <string.h>
  50 #include <sysexits.h>
  51 #include <unistd.h>
  52
  53 #include <activemap.h>
  54 #include <nv.h>
  55 #include <rangelock.h>
  56
  57 #include "control.h"
  58 #include "event.h"
  59 #include "hast.h"
  60 #include "hast_proto.h"
  61 #include "hastd.h"
  62 #include "hooks.h"
  63 #include "metadata.h"
  64 #include "proto.h"
  65 #include "pjdlog.h"
  66 #include "refcnt.h"
  67 #include "subr.h"
  68 #include "synch.h"
  69
  70 /* The is only one remote component for now. */
  71 #define ISREMOTE(no)    ((no) == 1)
  72
  73 struct hio {
  74         /*
  75          * Number of components we are still waiting for.
  76          * When this field goes to 0, we can send the request back to the
  77          * kernel. Each component has to decrease this counter by one
  78          * even on failure.
  79          */
  80         refcnt_t                 hio_countdown;
  81         /*
  82          * Each component has a place to store its own error.
  83          * Once the request is handled by all components we can decide if the
  84          * request overall is successful or not.
  85          */
  86         int                     *hio_errors;
  87         /*
  88          * Structure used to communicate with GEOM Gate class.
  89          */
  90         struct g_gate_ctl_io     hio_ggio;
  91         /*
  92          * Request was already confirmed to GEOM Gate.
  93          */
  94         bool                     hio_done;
  95         /*
  96          * Number of components we are still waiting before sending write
  97          * completion ack to GEOM Gate. Used for memsync.
  98          */
  99         refcnt_t                 hio_writecount;
 100         /*
 101          * Memsync request was acknowledged by remote.
 102          */
 103         bool                     hio_memsyncacked;
 104         /*
 105          * Remember replication from the time the request was initiated,
 106          * so we won't get confused when replication changes on reload.
 107          */
 108         int                      hio_replication;
 109         TAILQ_ENTRY(hio)        *hio_next;
 110 };
 111 #define hio_free_next   hio_next[0]
 112 #define hio_done_next   hio_next[0]
 113
 114 /*
 115  * Free list holds unused structures. When free list is empty, we have to wait
 116  * until some in-progress requests are freed.
 117  */
 118 static TAILQ_HEAD(, hio) hio_free_list;
 119 static size_t hio_free_list_size;
 120 static pthread_mutex_t hio_free_list_lock;
 121 static pthread_cond_t hio_free_list_cond;
 122 /*
 123  * There is one send list for every component. One requests is placed on all
 124  * send lists - each component gets the same request, but each component is
 125  * responsible for managing his own send list.
 126  */
 127 static TAILQ_HEAD(, hio) *hio_send_list;
 128 static size_t *hio_send_list_size;
 129 static pthread_mutex_t *hio_send_list_lock;
 130 static pthread_cond_t *hio_send_list_cond;
 131 #define hio_send_local_list_size        hio_send_list_size[0]
 132 #define hio_send_remote_list_size       hio_send_list_size[1]
 133 /*
 134  * There is one recv list for every component, although local components don't
 135  * use recv lists as local requests are done synchronously.
 136  */
 137 static TAILQ_HEAD(, hio) *hio_recv_list;
 138 static size_t *hio_recv_list_size;
 139 static pthread_mutex_t *hio_recv_list_lock;
 140 static pthread_cond_t *hio_recv_list_cond;
 141 #define hio_recv_remote_list_size       hio_recv_list_size[1]
 142 /*
 143  * Request is placed on done list by the slowest component (the one that
 144  * decreased hio_countdown from 1 to 0).
 145  */
 146 static TAILQ_HEAD(, hio) hio_done_list;
 147 static size_t hio_done_list_size;
 148 static pthread_mutex_t hio_done_list_lock;
 149 static pthread_cond_t hio_done_list_cond;
 150 /*
 151  * Structure below are for interaction with sync thread.
 152  */
 153 static bool sync_inprogress;
 154 static pthread_mutex_t sync_lock;
 155 static pthread_cond_t sync_cond;
 156 /*
 157  * The lock below allows to synchornize access to remote connections.
 158  */
 159 static pthread_rwlock_t *hio_remote_lock;
 160
 161 /*
 162  * Lock to synchronize metadata updates. Also synchronize access to
 163  * hr_primary_localcnt and hr_primary_remotecnt fields.
 164  */
 165 static pthread_mutex_t metadata_lock;
 166
 167 /*
 168  * Maximum number of outstanding I/O requests.
 169  */
 170 #define HAST_HIO_MAX    256
 171 /*
 172  * Number of components. At this point there are only two components: local
 173  * and remote, but in the future it might be possible to use multiple local
 174  * and remote components.
 175  */
 176 #define HAST_NCOMPONENTS        2
 177
 178 #define ISCONNECTED(res, no)    \
 179         ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
 180
 181 #define QUEUE_INSERT1(hio, name, ncomp) do {                            \
 182         mtx_lock(&hio_##name##_list_lock[(ncomp)]);                     \
 183         if (TAILQ_EMPTY(&hio_##name##_list[(ncomp)]))                   \
 184                 cv_broadcast(&hio_##name##_list_cond[(ncomp)]);         \
 185         TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),           \
 186             hio_next[(ncomp)]);                                         \
 187         hio_##name##_list_size[(ncomp)]++;                              \
 188         mtx_unlock(&hio_##name##_list_lock[(ncomp)]);                   \
 189 } while (0)
 190 #define QUEUE_INSERT2(hio, name)        do {                            \
 191         mtx_lock(&hio_##name##_list_lock);                              \
 192         if (TAILQ_EMPTY(&hio_##name##_list))                            \
 193                 cv_broadcast(&hio_##name##_list_cond);                  \
 194         TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
 195         hio_##name##_list_size++;                                       \
 196         mtx_unlock(&hio_##name##_list_lock);                            \
 197 } while (0)
 198 #define QUEUE_TAKE1(hio, name, ncomp, timeout)  do {                    \
 199         bool _last;                                                     \
 200                                                                         \
 201         mtx_lock(&hio_##name##_list_lock[(ncomp)]);                     \
 202         _last = false;                                                  \
 203         while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
 204                 cv_timedwait(&hio_##name##_list_cond[(ncomp)],          \
 205                     &hio_##name##_list_lock[(ncomp)], (timeout));       \
 206                 if ((timeout) != 0)                                     \
 207                         _last = true;                                   \
 208         }                                                               \
 209         if (hio != NULL) {                                              \
 210                 PJDLOG_ASSERT(hio_##name##_list_size[(ncomp)] != 0);    \
 211                 hio_##name##_list_size[(ncomp)]--;                      \
 212                 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),        \
 213                     hio_next[(ncomp)]);                                 \
 214         }                                                               \
 215         mtx_unlock(&hio_##name##_list_lock[(ncomp)]);                   \
 216 } while (0)
 217 #define QUEUE_TAKE2(hio, name)  do {                                    \
 218         mtx_lock(&hio_##name##_list_lock);                              \
 219         while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {     \
 220                 cv_wait(&hio_##name##_list_cond,                        \
 221                     &hio_##name##_list_lock);                           \
 222         }                                                               \
 223         PJDLOG_ASSERT(hio_##name##_list_size != 0);                     \
 224         hio_##name##_list_size--;                                       \
 225         TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);     \
 226         mtx_unlock(&hio_##name##_list_lock);                            \
 227 } while (0)
 228
 229 #define ISFULLSYNC(hio) ((hio)->hio_replication == HAST_REPLICATION_FULLSYNC)
 230 #define ISMEMSYNC(hio)  ((hio)->hio_replication == HAST_REPLICATION_MEMSYNC)
 231 #define ISASYNC(hio)    ((hio)->hio_replication == HAST_REPLICATION_ASYNC)
 232
 233 #define SYNCREQ(hio)            do {                                    \
 234         (hio)->hio_ggio.gctl_unit = -1;                                 \
 235         (hio)->hio_ggio.gctl_seq = 1;                                   \
 236 } while (0)
 237 #define ISSYNCREQ(hio)          ((hio)->hio_ggio.gctl_unit == -1)
 238 #define SYNCREQDONE(hio)        do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
 239 #define ISSYNCREQDONE(hio)      ((hio)->hio_ggio.gctl_unit == -2)
 240
 241 #define ISMEMSYNCWRITE(hio)     (ISMEMSYNC(hio) &&                      \
 242             (hio)->hio_ggio.gctl_cmd == BIO_WRITE && !ISSYNCREQ(hio))
 243
 244 static struct hast_resource *gres;
 245
 246 static pthread_mutex_t range_lock;
 247 static struct rangelocks *range_regular;
 248 static bool range_regular_wait;
 249 static pthread_cond_t range_regular_cond;
 250 static struct rangelocks *range_sync;
 251 static bool range_sync_wait;
 252 static pthread_cond_t range_sync_cond;
 253 static bool fullystarted;
 254
 255 static void *ggate_recv_thread(void *arg);
 256 static void *local_send_thread(void *arg);
 257 static void *remote_send_thread(void *arg);
 258 static void *remote_recv_thread(void *arg);
 259 static void *ggate_send_thread(void *arg);
 260 static void *sync_thread(void *arg);
 261 static void *guard_thread(void *arg);
 262
 263 static void
 264 output_status_aux(struct nv *nvout)
 265 {
 266
 267         nv_add_uint64(nvout, (uint64_t)hio_free_list_size,
 268             "idle_queue_size");
 269         nv_add_uint64(nvout, (uint64_t)hio_send_local_list_size,
 270             "local_queue_size");
 271         nv_add_uint64(nvout, (uint64_t)hio_send_remote_list_size,
 272             "send_queue_size");
 273         nv_add_uint64(nvout, (uint64_t)hio_recv_remote_list_size,
 274             "recv_queue_size");
 275         nv_add_uint64(nvout, (uint64_t)hio_done_list_size,
 276             "done_queue_size");
 277 }
 278
 279 static void
 280 cleanup(struct hast_resource *res)
 281 {
 282         int rerrno;
 283
 284         /* Remember errno. */
 285         rerrno = errno;
 286
 287         /* Destroy ggate provider if we created one. */
 288         if (res->hr_ggateunit >= 0) {
 289                 struct g_gate_ctl_destroy ggiod;
 290
 291                 bzero(&ggiod, sizeof(ggiod));
 292                 ggiod.gctl_version = G_GATE_VERSION;
 293                 ggiod.gctl_unit = res->hr_ggateunit;
 294                 ggiod.gctl_force = 1;
 295                 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) == -1) {
 296                         pjdlog_errno(LOG_WARNING,
 297                             "Unable to destroy hast/%s device",
 298                             res->hr_provname);
 299                 }
 300                 res->hr_ggateunit = -1;
 301         }
 302
 303         /* Restore errno. */
 304         errno = rerrno;
 305 }
 306
 307 static __dead2 void
 308 primary_exit(int exitcode, const char *fmt, ...)
 309 {
 310         va_list ap;
 311
 312         PJDLOG_ASSERT(exitcode != EX_OK);
 313         va_start(ap, fmt);
 314         pjdlogv_errno(LOG_ERR, fmt, ap);
 315         va_end(ap);
 316         cleanup(gres);
 317         exit(exitcode);
 318 }
 319
 320 static __dead2 void
 321 primary_exitx(int exitcode, const char *fmt, ...)
 322 {
 323         va_list ap;
 324
 325         va_start(ap, fmt);
 326         pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
 327         va_end(ap);
 328         cleanup(gres);
 329         exit(exitcode);
 330 }
 331
 332 static int
 333 hast_activemap_flush(struct hast_resource *res) __unlocks(res->hr_amp_lock)
 334 {
 335         const unsigned char *buf;
 336         size_t size;
 337         int ret;
 338
 339         mtx_lock(&res->hr_amp_diskmap_lock);
 340         buf = activemap_bitmap(res->hr_amp, &size);
 341         mtx_unlock(&res->hr_amp_lock);
 342         PJDLOG_ASSERT(buf != NULL);
 343         PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
 344         ret = 0;
 345         if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
 346             (ssize_t)size) {
 347                 pjdlog_errno(LOG_ERR, "Unable to flush activemap to disk");
 348                 res->hr_stat_activemap_write_error++;
 349                 ret = -1;
 350         }
 351         if (ret == 0 && res->hr_metaflush == 1 &&
 352             g_flush(res->hr_localfd) == -1) {
 353                 if (errno == EOPNOTSUPP) {
 354                         pjdlog_warning("The %s provider doesn't support flushing write cache. Disabling it.",
 355                             res->hr_localpath);
 356                         res->hr_metaflush = 0;
 357                 } else {
 358                         pjdlog_errno(LOG_ERR,
 359                             "Unable to flush disk cache on activemap update");
 360                         res->hr_stat_activemap_flush_error++;
 361                         ret = -1;
 362                 }
 363         }
 364         mtx_unlock(&res->hr_amp_diskmap_lock);
 365         return (ret);
 366 }
 367
 368 static bool
 369 real_remote(const struct hast_resource *res)
 370 {
 371
 372         return (strcmp(res->hr_remoteaddr, "none") != 0);
 373 }
 374
 375 static void
 376 init_environment(struct hast_resource *res __unused)
 377 {
 378         struct hio *hio;
 379         unsigned int ii, ncomps;
 380
 381         /*
 382          * In the future it might be per-resource value.
 383          */
 384         ncomps = HAST_NCOMPONENTS;
 385
 386         /*
 387          * Allocate memory needed by lists.
 388          */
 389         hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
 390         if (hio_send_list == NULL) {
 391                 primary_exitx(EX_TEMPFAIL,
 392                     "Unable to allocate %zu bytes of memory for send lists.",
 393                     sizeof(hio_send_list[0]) * ncomps);
 394         }
 395         hio_send_list_size = malloc(sizeof(hio_send_list_size[0]) * ncomps);
 396         if (hio_send_list_size == NULL) {
 397                 primary_exitx(EX_TEMPFAIL,
 398                     "Unable to allocate %zu bytes of memory for send list counters.",
 399                     sizeof(hio_send_list_size[0]) * ncomps);
 400         }
 401         hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
 402         if (hio_send_list_lock == NULL) {
 403                 primary_exitx(EX_TEMPFAIL,
 404                     "Unable to allocate %zu bytes of memory for send list locks.",
 405                     sizeof(hio_send_list_lock[0]) * ncomps);
 406         }
 407         hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
 408         if (hio_send_list_cond == NULL) {
 409                 primary_exitx(EX_TEMPFAIL,
 410                     "Unable to allocate %zu bytes of memory for send list condition variables.",
 411                     sizeof(hio_send_list_cond[0]) * ncomps);
 412         }
 413         hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
 414         if (hio_recv_list == NULL) {
 415                 primary_exitx(EX_TEMPFAIL,
 416                     "Unable to allocate %zu bytes of memory for recv lists.",
 417                     sizeof(hio_recv_list[0]) * ncomps);
 418         }
 419         hio_recv_list_size = malloc(sizeof(hio_recv_list_size[0]) * ncomps);
 420         if (hio_recv_list_size == NULL) {
 421                 primary_exitx(EX_TEMPFAIL,
 422                     "Unable to allocate %zu bytes of memory for recv list counters.",
 423                     sizeof(hio_recv_list_size[0]) * ncomps);
 424         }
 425         hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
 426         if (hio_recv_list_lock == NULL) {
 427                 primary_exitx(EX_TEMPFAIL,
 428                     "Unable to allocate %zu bytes of memory for recv list locks.",
 429                     sizeof(hio_recv_list_lock[0]) * ncomps);
 430         }
 431         hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
 432         if (hio_recv_list_cond == NULL) {
 433                 primary_exitx(EX_TEMPFAIL,
 434                     "Unable to allocate %zu bytes of memory for recv list condition variables.",
 435                     sizeof(hio_recv_list_cond[0]) * ncomps);
 436         }
 437         hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
 438         if (hio_remote_lock == NULL) {
 439                 primary_exitx(EX_TEMPFAIL,
 440                     "Unable to allocate %zu bytes of memory for remote connections locks.",
 441                     sizeof(hio_remote_lock[0]) * ncomps);
 442         }
 443
 444         /*
 445          * Initialize lists, their counters, locks and condition variables.
 446          */
 447         TAILQ_INIT(&hio_free_list);
 448         mtx_init(&hio_free_list_lock);
 449         cv_init(&hio_free_list_cond);
 450         for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
 451                 TAILQ_INIT(&hio_send_list[ii]);
 452                 hio_send_list_size[ii] = 0;
 453                 mtx_init(&hio_send_list_lock[ii]);
 454                 cv_init(&hio_send_list_cond[ii]);
 455                 TAILQ_INIT(&hio_recv_list[ii]);
 456                 hio_recv_list_size[ii] = 0;
 457                 mtx_init(&hio_recv_list_lock[ii]);
 458                 cv_init(&hio_recv_list_cond[ii]);
 459                 rw_init(&hio_remote_lock[ii]);
 460         }
 461         TAILQ_INIT(&hio_done_list);
 462         mtx_init(&hio_done_list_lock);
 463         cv_init(&hio_done_list_cond);
 464         mtx_init(&metadata_lock);
 465
 466         /*
 467          * Allocate requests pool and initialize requests.
 468          */
 469         for (ii = 0; ii < HAST_HIO_MAX; ii++) {
 470                 hio = malloc(sizeof(*hio));
 471                 if (hio == NULL) {
 472                         primary_exitx(EX_TEMPFAIL,
 473                             "Unable to allocate %zu bytes of memory for hio request.",
 474                             sizeof(*hio));
 475                 }
 476                 refcnt_init(&hio->hio_countdown, 0);
 477                 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
 478                 if (hio->hio_errors == NULL) {
 479                         primary_exitx(EX_TEMPFAIL,
 480                             "Unable allocate %zu bytes of memory for hio errors.",
 481                             sizeof(hio->hio_errors[0]) * ncomps);
 482                 }
 483                 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
 484                 if (hio->hio_next == NULL) {
 485                         primary_exitx(EX_TEMPFAIL,
 486                             "Unable allocate %zu bytes of memory for hio_next field.",
 487                             sizeof(hio->hio_next[0]) * ncomps);
 488                 }
 489                 hio->hio_ggio.gctl_version = G_GATE_VERSION;
 490                 hio->hio_ggio.gctl_data = malloc(MAXPHYS);
 491                 if (hio->hio_ggio.gctl_data == NULL) {
 492                         primary_exitx(EX_TEMPFAIL,
 493                             "Unable to allocate %zu bytes of memory for gctl_data.",
 494                             MAXPHYS);
 495                 }
 496                 hio->hio_ggio.gctl_length = MAXPHYS;
 497                 hio->hio_ggio.gctl_error = 0;
 498                 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
 499                 hio_free_list_size++;
 500         }
 501 }
 502
 503 static bool
 504 init_resuid(struct hast_resource *res)
 505 {
 506
 507         mtx_lock(&metadata_lock);
 508         if (res->hr_resuid != 0) {
 509                 mtx_unlock(&metadata_lock);
 510                 return (false);
 511         } else {
 512                 /* Initialize unique resource identifier. */
 513                 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
 514                 mtx_unlock(&metadata_lock);
 515                 if (metadata_write(res) == -1)
 516                         exit(EX_NOINPUT);
 517                 return (true);
 518         }
 519 }
 520
 521 static void
 522 init_local(struct hast_resource *res)
 523 {
 524         unsigned char *buf;
 525         size_t mapsize;
 526
 527         if (metadata_read(res, true) == -1)
 528                 exit(EX_NOINPUT);
 529         mtx_init(&res->hr_amp_lock);
 530         if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
 531             res->hr_local_sectorsize, res->hr_keepdirty) == -1) {
 532                 primary_exit(EX_TEMPFAIL, "Unable to create activemap");
 533         }
 534         mtx_init(&range_lock);
 535         cv_init(&range_regular_cond);
 536         if (rangelock_init(&range_regular) == -1)
 537                 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
 538         cv_init(&range_sync_cond);
 539         if (rangelock_init(&range_sync) == -1)
 540                 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
 541         mapsize = activemap_ondisk_size(res->hr_amp);
 542         buf = calloc(1, mapsize);
 543         if (buf == NULL) {
 544                 primary_exitx(EX_TEMPFAIL,
 545                     "Unable to allocate buffer for activemap.");
 546         }
 547         if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
 548             (ssize_t)mapsize) {
 549                 primary_exit(EX_NOINPUT, "Unable to read activemap");
 550         }
 551         activemap_copyin(res->hr_amp, buf, mapsize);
 552         free(buf);
 553         if (res->hr_resuid != 0)
 554                 return;
 555         /*
 556          * We're using provider for the first time. Initialize local and remote
 557          * counters. We don't initialize resuid here, as we want to do it just
 558          * in time. The reason for this is that we want to inform secondary
 559          * that there were no writes yet, so there is no need to synchronize
 560          * anything.
 561          */
 562         res->hr_primary_localcnt = 0;
 563         res->hr_primary_remotecnt = 0;
 564         if (metadata_write(res) == -1)
 565                 exit(EX_NOINPUT);
 566 }
 567
 568 static int
 569 primary_connect(struct hast_resource *res, struct proto_conn **connp)
 570 {
 571         struct proto_conn *conn;
 572         int16_t val;
 573
 574         val = 1;
 575         if (proto_send(res->hr_conn, &val, sizeof(val)) == -1) {
 576                 primary_exit(EX_TEMPFAIL,
 577                     "Unable to send connection request to parent");
 578         }
 579         if (proto_recv(res->hr_conn, &val, sizeof(val)) == -1) {
 580                 primary_exit(EX_TEMPFAIL,
 581                     "Unable to receive reply to connection request from parent");
 582         }
 583         if (val != 0) {
 584                 errno = val;
 585                 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
 586                     res->hr_remoteaddr);
 587                 return (-1);
 588         }
 589         if (proto_connection_recv(res->hr_conn, true, &conn) == -1) {
 590                 primary_exit(EX_TEMPFAIL,
 591                     "Unable to receive connection from parent");
 592         }
 593         if (proto_connect_wait(conn, res->hr_timeout) == -1) {
 594                 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
 595                     res->hr_remoteaddr);
 596                 proto_close(conn);
 597                 return (-1);
 598         }
 599         /* Error in setting timeout is not critical, but why should it fail? */
 600         if (proto_timeout(conn, res->hr_timeout) == -1)
 601                 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
 602
 603         *connp = conn;
 604
 605         return (0);
 606 }
 607
 608 /*
 609  * Function instructs GEOM_GATE to handle reads directly from within the kernel.
 610  */
 611 static void
 612 enable_direct_reads(struct hast_resource *res)
 613 {
 614         struct g_gate_ctl_modify ggiomodify;
 615
 616         bzero(&ggiomodify, sizeof(ggiomodify));
 617         ggiomodify.gctl_version = G_GATE_VERSION;
 618         ggiomodify.gctl_unit = res->hr_ggateunit;
 619         ggiomodify.gctl_modify = GG_MODIFY_READPROV | GG_MODIFY_READOFFSET;
 620         strlcpy(ggiomodify.gctl_readprov, res->hr_localpath,
 621             sizeof(ggiomodify.gctl_readprov));
 622         ggiomodify.gctl_readoffset = res->hr_localoff;
 623         if (ioctl(res->hr_ggatefd, G_GATE_CMD_MODIFY, &ggiomodify) == 0)
 624                 pjdlog_debug(1, "Direct reads enabled.");
 625         else
 626                 pjdlog_errno(LOG_WARNING, "Failed to enable direct reads");
 627 }
 628
 629 static int
 630 init_remote(struct hast_resource *res, struct proto_conn **inp,
 631     struct proto_conn **outp)
 632 {
 633         struct proto_conn *in, *out;
 634         struct nv *nvout, *nvin;
 635         const unsigned char *token;
 636         unsigned char *map;
 637         const char *errmsg;
 638         int32_t extentsize;
 639         int64_t datasize;
 640         uint32_t mapsize;
 641         uint8_t version;
 642         size_t size;
 643         int error;
 644
 645         PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
 646         PJDLOG_ASSERT(real_remote(res));
 647
 648         in = out = NULL;
 649         errmsg = NULL;
 650
 651         if (primary_connect(res, &out) == -1)
 652                 return (ECONNREFUSED);
 653
 654         error = ECONNABORTED;
 655
 656         /*
 657          * First handshake step.
 658          * Setup outgoing connection with remote node.
 659          */
 660         nvout = nv_alloc();
 661         nv_add_string(nvout, res->hr_name, "resource");
 662         nv_add_uint8(nvout, HAST_PROTO_VERSION, "version");
 663         if (nv_error(nvout) != 0) {
 664                 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
 665                     "Unable to allocate header for connection with %s",
 666                     res->hr_remoteaddr);
 667                 nv_free(nvout);
 668                 goto close;
 669         }
 670         if (hast_proto_send(res, out, nvout, NULL, 0) == -1) {
 671                 pjdlog_errno(LOG_WARNING,
 672                     "Unable to send handshake header to %s",
 673                     res->hr_remoteaddr);
 674                 nv_free(nvout);
 675                 goto close;
 676         }
 677         nv_free(nvout);
 678         if (hast_proto_recv_hdr(out, &nvin) == -1) {
 679                 pjdlog_errno(LOG_WARNING,
 680                     "Unable to receive handshake header from %s",
 681                     res->hr_remoteaddr);
 682                 goto close;
 683         }
 684         errmsg = nv_get_string(nvin, "errmsg");
 685         if (errmsg != NULL) {
 686                 pjdlog_warning("%s", errmsg);
 687                 if (nv_exists(nvin, "wait"))
 688                         error = EBUSY;
 689                 nv_free(nvin);
 690                 goto close;
 691         }
 692         version = nv_get_uint8(nvin, "version");
 693         if (version == 0) {
 694                 /*
 695                  * If no version is sent, it means this is protocol version 1.
 696                  */
 697                 version = 1;
 698         }
 699         if (version > HAST_PROTO_VERSION) {
 700                 pjdlog_warning("Invalid version received (%hhu).", version);
 701                 nv_free(nvin);
 702                 goto close;
 703         }
 704         res->hr_version = version;
 705         pjdlog_debug(1, "Negotiated protocol version %d.", res->hr_version);
 706         token = nv_get_uint8_array(nvin, &size, "token");
 707         if (token == NULL) {
 708                 pjdlog_warning("Handshake header from %s has no 'token' field.",
 709                     res->hr_remoteaddr);
 710                 nv_free(nvin);
 711                 goto close;
 712         }
 713         if (size != sizeof(res->hr_token)) {
 714                 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
 715                     res->hr_remoteaddr, size, sizeof(res->hr_token));
 716                 nv_free(nvin);
 717                 goto close;
 718         }
 719         bcopy(token, res->hr_token, sizeof(res->hr_token));
 720         nv_free(nvin);
 721
 722         /*
 723          * Second handshake step.
 724          * Setup incoming connection with remote node.
 725          */
 726         if (primary_connect(res, &in) == -1)
 727                 goto close;
 728
 729         nvout = nv_alloc();
 730         nv_add_string(nvout, res->hr_name, "resource");
 731         nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
 732             "token");
 733         if (res->hr_resuid == 0) {
 734                 /*
 735                  * The resuid field was not yet initialized.
 736                  * Because we do synchronization inside init_resuid(), it is
 737                  * possible that someone already initialized it, the function
 738                  * will return false then, but if we successfully initialized
 739                  * it, we will get true. True means that there were no writes
 740                  * to this resource yet and we want to inform secondary that
 741                  * synchronization is not needed by sending "virgin" argument.
 742                  */
 743                 if (init_resuid(res))
 744                         nv_add_int8(nvout, 1, "virgin");
 745         }
 746         nv_add_uint64(nvout, res->hr_resuid, "resuid");
 747         nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
 748         nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
 749         if (nv_error(nvout) != 0) {
 750                 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
 751                     "Unable to allocate header for connection with %s",
 752                     res->hr_remoteaddr);
 753                 nv_free(nvout);
 754                 goto close;
 755         }
 756         if (hast_proto_send(res, in, nvout, NULL, 0) == -1) {
 757                 pjdlog_errno(LOG_WARNING,
 758                     "Unable to send handshake header to %s",
 759                     res->hr_remoteaddr);
 760                 nv_free(nvout);
 761                 goto close;
 762         }
 763         nv_free(nvout);
 764         if (hast_proto_recv_hdr(out, &nvin) == -1) {
 765                 pjdlog_errno(LOG_WARNING,
 766                     "Unable to receive handshake header from %s",
 767                     res->hr_remoteaddr);
 768                 goto close;
 769         }
 770         errmsg = nv_get_string(nvin, "errmsg");
 771         if (errmsg != NULL) {
 772                 pjdlog_warning("%s", errmsg);
 773                 nv_free(nvin);
 774                 goto close;
 775         }
 776         datasize = nv_get_int64(nvin, "datasize");
 777         if (datasize != res->hr_datasize) {
 778                 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
 779                     (intmax_t)res->hr_datasize, (intmax_t)datasize);
 780                 nv_free(nvin);
 781                 goto close;
 782         }
 783         extentsize = nv_get_int32(nvin, "extentsize");
 784         if (extentsize != res->hr_extentsize) {
 785                 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
 786                     (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
 787                 nv_free(nvin);
 788                 goto close;
 789         }
 790         res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
 791         res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
 792         res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
 793         if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY)
 794                 enable_direct_reads(res);
 795         if (nv_exists(nvin, "virgin")) {
 796                 /*
 797                  * Secondary was reinitialized, bump localcnt if it is 0 as
 798                  * only we have the data.
 799                  */
 800                 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY);
 801                 PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
 802
 803                 if (res->hr_primary_localcnt == 0) {
 804                         PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0);
 805
 806                         mtx_lock(&metadata_lock);
 807                         res->hr_primary_localcnt++;
 808                         pjdlog_debug(1, "Increasing localcnt to %ju.",
 809                             (uintmax_t)res->hr_primary_localcnt);
 810                         (void)metadata_write(res);
 811                         mtx_unlock(&metadata_lock);
 812                 }
 813         }
 814         map = NULL;
 815         mapsize = nv_get_uint32(nvin, "mapsize");
 816         if (mapsize > 0) {
 817                 map = malloc(mapsize);
 818                 if (map == NULL) {
 819                         pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
 820                             (uintmax_t)mapsize);
 821                         nv_free(nvin);
 822                         goto close;
 823                 }
 824                 /*
 825                  * Remote node have some dirty extents on its own, lets
 826                  * download its activemap.
 827                  */
 828                 if (hast_proto_recv_data(res, out, nvin, map,
 829                     mapsize) == -1) {
 830                         pjdlog_errno(LOG_ERR,
 831                             "Unable to receive remote activemap");
 832                         nv_free(nvin);
 833                         free(map);
 834                         goto close;
 835                 }
 836                 mtx_lock(&res->hr_amp_lock);
 837                 /*
 838                  * Merge local and remote bitmaps.
 839                  */
 840                 activemap_merge(res->hr_amp, map, mapsize);
 841                 free(map);
 842                 /*
 843                  * Now that we merged bitmaps from both nodes, flush it to the
 844                  * disk before we start to synchronize.
 845                  */
 846                 (void)hast_activemap_flush(res);
 847         }
 848         nv_free(nvin);
 849 #ifdef notyet
 850         /* Setup directions. */
 851         if (proto_send(out, NULL, 0) == -1)
 852                 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
 853         if (proto_recv(in, NULL, 0) == -1)
 854                 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
 855 #endif
 856         pjdlog_info("Connected to %s.", res->hr_remoteaddr);
 857         if (res->hr_original_replication == HAST_REPLICATION_MEMSYNC &&
 858             res->hr_version < 2) {
 859                 pjdlog_warning("The 'memsync' replication mode is not supported by the remote node, falling back to 'fullsync' mode.");
 860                 res->hr_replication = HAST_REPLICATION_FULLSYNC;
 861         } else if (res->hr_replication != res->hr_original_replication) {
 862                 /*
 863                  * This is in case hastd disconnected and was upgraded.
 864                  */
 865                 res->hr_replication = res->hr_original_replication;
 866         }
 867         if (inp != NULL && outp != NULL) {
 868                 *inp = in;
 869                 *outp = out;
 870         } else {
 871                 res->hr_remotein = in;
 872                 res->hr_remoteout = out;
 873         }
 874         event_send(res, EVENT_CONNECT);
 875         return (0);
 876 close:
 877         if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
 878                 event_send(res, EVENT_SPLITBRAIN);
 879         proto_close(out);
 880         if (in != NULL)
 881                 proto_close(in);
 882         return (error);
 883 }
 884
 885 static void
 886 sync_start(void)
 887 {
 888
 889         mtx_lock(&sync_lock);
 890         sync_inprogress = true;
 891         mtx_unlock(&sync_lock);
 892         cv_signal(&sync_cond);
 893 }
 894
 895 static void
 896 sync_stop(void)
 897 {
 898
 899         mtx_lock(&sync_lock);
 900         if (sync_inprogress)
 901                 sync_inprogress = false;
 902         mtx_unlock(&sync_lock);
 903 }
 904
 905 static void
 906 init_ggate(struct hast_resource *res)
 907 {
 908         struct g_gate_ctl_create ggiocreate;
 909         struct g_gate_ctl_cancel ggiocancel;
 910
 911         /*
 912          * We communicate with ggate via /dev/ggctl. Open it.
 913          */
 914         res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
 915         if (res->hr_ggatefd == -1)
 916                 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
 917         /*
 918          * Create provider before trying to connect, as connection failure
 919          * is not critical, but may take some time.
 920          */
 921         bzero(&ggiocreate, sizeof(ggiocreate));
 922         ggiocreate.gctl_version = G_GATE_VERSION;
 923         ggiocreate.gctl_mediasize = res->hr_datasize;
 924         ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
 925         ggiocreate.gctl_flags = 0;
 926         ggiocreate.gctl_maxcount = 0;
 927         ggiocreate.gctl_timeout = 0;
 928         ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
 929         snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
 930             res->hr_provname);
 931         if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
 932                 pjdlog_info("Device hast/%s created.", res->hr_provname);
 933                 res->hr_ggateunit = ggiocreate.gctl_unit;
 934                 return;
 935         }
 936         if (errno != EEXIST) {
 937                 primary_exit(EX_OSERR, "Unable to create hast/%s device",
 938                     res->hr_provname);
 939         }
 940         pjdlog_debug(1,
 941             "Device hast/%s already exists, we will try to take it over.",
 942             res->hr_provname);
 943         /*
 944          * If we received EEXIST, we assume that the process who created the
 945          * provider died and didn't clean up. In that case we will start from
 946          * where he left of.
 947          */
 948         bzero(&ggiocancel, sizeof(ggiocancel));
 949         ggiocancel.gctl_version = G_GATE_VERSION;
 950         ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
 951         snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
 952             res->hr_provname);
 953         if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
 954                 pjdlog_info("Device hast/%s recovered.", res->hr_provname);
 955                 res->hr_ggateunit = ggiocancel.gctl_unit;
 956                 return;
 957         }
 958         primary_exit(EX_OSERR, "Unable to take over hast/%s device",
 959             res->hr_provname);
 960 }
 961
 962 void
 963 hastd_primary(struct hast_resource *res)
 964 {
 965         pthread_t td;
 966         pid_t pid;
 967         int error, mode, debuglevel;
 968
 969         /*
 970          * Create communication channel for sending control commands from
 971          * parent to child.
 972          */
 973         if (proto_client(NULL, "socketpair://", &res->hr_ctrl) == -1) {
 974                 /* TODO: There's no need for this to be fatal error. */
 975                 KEEP_ERRNO((void)pidfile_remove(pfh));
 976                 pjdlog_exit(EX_OSERR,
 977                     "Unable to create control sockets between parent and child");
 978         }
 979         /*
 980          * Create communication channel for sending events from child to parent.
 981          */
 982         if (proto_client(NULL, "socketpair://", &res->hr_event) == -1) {
 983                 /* TODO: There's no need for this to be fatal error. */
 984                 KEEP_ERRNO((void)pidfile_remove(pfh));
 985                 pjdlog_exit(EX_OSERR,
 986                     "Unable to create event sockets between child and parent");
 987         }
 988         /*
 989          * Create communication channel for sending connection requests from
 990          * child to parent.
 991          */
 992         if (proto_client(NULL, "socketpair://", &res->hr_conn) == -1) {
 993                 /* TODO: There's no need for this to be fatal error. */
 994                 KEEP_ERRNO((void)pidfile_remove(pfh));
 995                 pjdlog_exit(EX_OSERR,
 996                     "Unable to create connection sockets between child and parent");
 997         }
 998
 999         pid = fork();
1000         if (pid == -1) {
1001                 /* TODO: There's no need for this to be fatal error. */
1002                 KEEP_ERRNO((void)pidfile_remove(pfh));
1003                 pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
1004         }
1005
1006         if (pid > 0) {
1007                 /* This is parent. */
1008                 /* Declare that we are receiver. */
1009                 proto_recv(res->hr_event, NULL, 0);
1010                 proto_recv(res->hr_conn, NULL, 0);
1011                 /* Declare that we are sender. */
1012                 proto_send(res->hr_ctrl, NULL, 0);
1013                 res->hr_workerpid = pid;
1014                 return;
1015         }
1016
1017         gres = res;
1018         res->output_status_aux = output_status_aux;
1019         mode = pjdlog_mode_get();
1020         debuglevel = pjdlog_debug_get();
1021
1022         /* Declare that we are sender. */
1023         proto_send(res->hr_event, NULL, 0);
1024         proto_send(res->hr_conn, NULL, 0);
1025         /* Declare that we are receiver. */
1026         proto_recv(res->hr_ctrl, NULL, 0);
1027         descriptors_cleanup(res);
1028
1029         descriptors_assert(res, mode);
1030
1031         pjdlog_init(mode);
1032         pjdlog_debug_set(debuglevel);
1033         pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
1034         setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
1035
1036         init_local(res);
1037         init_ggate(res);
1038         init_environment(res);
1039
1040         if (drop_privs(res) != 0) {
1041                 cleanup(res);
1042                 exit(EX_CONFIG);
1043         }
1044         pjdlog_info("Privileges successfully dropped.");
1045
1046         /*
1047          * Create the guard thread first, so we can handle signals from the
1048          * very beginning.
1049          */
1050         error = pthread_create(&td, NULL, guard_thread, res);
1051         PJDLOG_ASSERT(error == 0);
1052         /*
1053          * Create the control thread before sending any event to the parent,
1054          * as we can deadlock when parent sends control request to worker,
1055          * but worker has no control thread started yet, so parent waits.
1056          * In the meantime worker sends an event to the parent, but parent
1057          * is unable to handle the event, because it waits for control
1058          * request response.
1059          */
1060         error = pthread_create(&td, NULL, ctrl_thread, res);
1061         PJDLOG_ASSERT(error == 0);
1062         if (real_remote(res)) {
1063                 error = init_remote(res, NULL, NULL);
1064                 if (error == 0) {
1065                         sync_start();
1066                 } else if (error == EBUSY) {
1067                         time_t start = time(NULL);
1068
1069                         pjdlog_warning("Waiting for remote node to become %s for %ds.",
1070                             role2str(HAST_ROLE_SECONDARY),
1071                             res->hr_timeout);
1072                         for (;;) {
1073                                 sleep(1);
1074                                 error = init_remote(res, NULL, NULL);
1075                                 if (error != EBUSY)
1076                                         break;
1077                                 if (time(NULL) > start + res->hr_timeout)
1078                                         break;
1079                         }
1080                         if (error == EBUSY) {
1081                                 pjdlog_warning("Remote node is still %s, starting anyway.",
1082                                     role2str(HAST_ROLE_PRIMARY));
1083                         }
1084                 }
1085         }
1086         error = pthread_create(&td, NULL, ggate_recv_thread, res);
1087         PJDLOG_ASSERT(error == 0);
1088         error = pthread_create(&td, NULL, local_send_thread, res);
1089         PJDLOG_ASSERT(error == 0);
1090         error = pthread_create(&td, NULL, remote_send_thread, res);
1091         PJDLOG_ASSERT(error == 0);
1092         error = pthread_create(&td, NULL, remote_recv_thread, res);
1093         PJDLOG_ASSERT(error == 0);
1094         error = pthread_create(&td, NULL, ggate_send_thread, res);
1095         PJDLOG_ASSERT(error == 0);
1096         fullystarted = true;
1097         (void)sync_thread(res);
1098 }
1099
1100 static void
1101 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio,
1102     const char *fmt, ...)
1103 {
1104         char msg[1024];
1105         va_list ap;
1106
1107         va_start(ap, fmt);
1108         (void)vsnprintf(msg, sizeof(msg), fmt, ap);
1109         va_end(ap);
1110         switch (ggio->gctl_cmd) {
1111         case BIO_READ:
1112                 (void)snprlcat(msg, sizeof(msg), "READ(%ju, %ju).",
1113                     (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1114                 break;
1115         case BIO_DELETE:
1116                 (void)snprlcat(msg, sizeof(msg), "DELETE(%ju, %ju).",
1117                     (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1118                 break;
1119         case BIO_FLUSH:
1120                 (void)snprlcat(msg, sizeof(msg), "FLUSH.");
1121                 break;
1122         case BIO_WRITE:
1123                 (void)snprlcat(msg, sizeof(msg), "WRITE(%ju, %ju).",
1124                     (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1125                 break;
1126         default:
1127                 (void)snprlcat(msg, sizeof(msg), "UNKNOWN(%u).",
1128                     (unsigned int)ggio->gctl_cmd);
1129                 break;
1130         }
1131         pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
1132 }
1133
1134 static void
1135 remote_close(struct hast_resource *res, int ncomp)
1136 {
1137
1138         rw_wlock(&hio_remote_lock[ncomp]);
1139         /*
1140          * Check for a race between dropping rlock and acquiring wlock -
1141          * another thread can close connection in-between.
1142          */
1143         if (!ISCONNECTED(res, ncomp)) {
1144                 PJDLOG_ASSERT(res->hr_remotein == NULL);
1145                 PJDLOG_ASSERT(res->hr_remoteout == NULL);
1146                 rw_unlock(&hio_remote_lock[ncomp]);
1147                 return;
1148         }
1149
1150         PJDLOG_ASSERT(res->hr_remotein != NULL);
1151         PJDLOG_ASSERT(res->hr_remoteout != NULL);
1152
1153         pjdlog_debug(2, "Closing incoming connection to %s.",
1154             res->hr_remoteaddr);
1155         proto_close(res->hr_remotein);
1156         res->hr_remotein = NULL;
1157         pjdlog_debug(2, "Closing outgoing connection to %s.",
1158             res->hr_remoteaddr);
1159         proto_close(res->hr_remoteout);
1160         res->hr_remoteout = NULL;
1161
1162         rw_unlock(&hio_remote_lock[ncomp]);
1163
1164         pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
1165
1166         /*
1167          * Stop synchronization if in-progress.
1168          */
1169         sync_stop();
1170
1171         event_send(res, EVENT_DISCONNECT);
1172 }
1173
1174 /*
1175  * Acknowledge write completion to the kernel, but don't update activemap yet.
1176  */
1177 static void
1178 write_complete(struct hast_resource *res, struct hio *hio)
1179 {
1180         struct g_gate_ctl_io *ggio;
1181         unsigned int ncomp;
1182
1183         PJDLOG_ASSERT(!hio->hio_done);
1184
1185         ggio = &hio->hio_ggio;
1186         PJDLOG_ASSERT(ggio->gctl_cmd == BIO_WRITE);
1187
1188         /*
1189          * Bump local count if this is first write after
1190          * connection failure with remote node.
1191          */
1192         ncomp = 1;
1193         rw_rlock(&hio_remote_lock[ncomp]);
1194         if (!ISCONNECTED(res, ncomp)) {
1195                 mtx_lock(&metadata_lock);
1196                 if (res->hr_primary_localcnt == res->hr_secondary_remotecnt) {
1197                         res->hr_primary_localcnt++;
1198                         pjdlog_debug(1, "Increasing localcnt to %ju.",
1199                             (uintmax_t)res->hr_primary_localcnt);
1200                         (void)metadata_write(res);
1201                 }
1202                 mtx_unlock(&metadata_lock);
1203         }
1204         rw_unlock(&hio_remote_lock[ncomp]);
1205         if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1)
1206                 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1207         hio->hio_done = true;
1208 }
1209
1210 /*
1211  * Thread receives ggate I/O requests from the kernel and passes them to
1212  * appropriate threads:
1213  * WRITE - always goes to both local_send and remote_send threads
1214  * READ (when the block is up-to-date on local component) -
1215  *      only local_send thread
1216  * READ (when the block isn't up-to-date on local component) -
1217  *      only remote_send thread
1218  * DELETE - always goes to both local_send and remote_send threads
1219  * FLUSH - always goes to both local_send and remote_send threads
1220  */
1221 static void *
1222 ggate_recv_thread(void *arg)
1223 {
1224         struct hast_resource *res = arg;
1225         struct g_gate_ctl_io *ggio;
1226         struct hio *hio;
1227         unsigned int ii, ncomp, ncomps;
1228         int error;
1229
1230         for (;;) {
1231                 pjdlog_debug(2, "ggate_recv: Taking free request.");
1232                 QUEUE_TAKE2(hio, free);
1233                 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1234                 ggio = &hio->hio_ggio;
1235                 ggio->gctl_unit = res->hr_ggateunit;
1236                 ggio->gctl_length = MAXPHYS;
1237                 ggio->gctl_error = 0;
1238                 hio->hio_done = false;
1239                 hio->hio_replication = res->hr_replication;
1240                 pjdlog_debug(2,
1241                     "ggate_recv: (%p) Waiting for request from the kernel.",
1242                     hio);
1243                 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) == -1) {
1244                         if (sigexit_received)
1245                                 pthread_exit(NULL);
1246                         primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1247                 }
1248                 error = ggio->gctl_error;
1249                 switch (error) {
1250                 case 0:
1251                         break;
1252                 case ECANCELED:
1253                         /* Exit gracefully. */
1254                         if (!sigexit_received) {
1255                                 pjdlog_debug(2,
1256                                     "ggate_recv: (%p) Received cancel from the kernel.",
1257                                     hio);
1258                                 pjdlog_info("Received cancel from the kernel, exiting.");
1259                         }
1260                         pthread_exit(NULL);
1261                 case ENOMEM:
1262                         /*
1263                          * Buffer too small? Impossible, we allocate MAXPHYS
1264                          * bytes - request can't be bigger than that.
1265                          */
1266                         /* FALLTHROUGH */
1267                 case ENXIO:
1268                 default:
1269                         primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1270                             strerror(error));
1271                 }
1272
1273                 ncomp = 0;
1274                 ncomps = HAST_NCOMPONENTS;
1275
1276                 for (ii = 0; ii < ncomps; ii++)
1277                         hio->hio_errors[ii] = EINVAL;
1278                 reqlog(LOG_DEBUG, 2, ggio,
1279                     "ggate_recv: (%p) Request received from the kernel: ",
1280                     hio);
1281
1282                 /*
1283                  * Inform all components about new write request.
1284                  * For read request prefer local component unless the given
1285                  * range is out-of-date, then use remote component.
1286                  */
1287                 switch (ggio->gctl_cmd) {
1288                 case BIO_READ:
1289                         res->hr_stat_read++;
1290                         ncomps = 1;
1291                         mtx_lock(&metadata_lock);
1292                         if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1293                             res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1294                                 /*
1295                                  * This range is up-to-date on local component,
1296                                  * so handle request locally.
1297                                  */
1298                                  /* Local component is 0 for now. */
1299                                 ncomp = 0;
1300                         } else /* if (res->hr_syncsrc ==
1301                             HAST_SYNCSRC_SECONDARY) */ {
1302                                 PJDLOG_ASSERT(res->hr_syncsrc ==
1303                                     HAST_SYNCSRC_SECONDARY);
1304                                 /*
1305                                  * This range is out-of-date on local component,
1306                                  * so send request to the remote node.
1307                                  */
1308                                  /* Remote component is 1 for now. */
1309                                 ncomp = 1;
1310                         }
1311                         mtx_unlock(&metadata_lock);
1312                         break;
1313                 case BIO_WRITE:
1314                         res->hr_stat_write++;
1315                         if (res->hr_resuid == 0 &&
1316                             res->hr_primary_localcnt == 0) {
1317                                 /* This is first write. */
1318                                 res->hr_primary_localcnt = 1;
1319                         }
1320                         for (;;) {
1321                                 mtx_lock(&range_lock);
1322                                 if (rangelock_islocked(range_sync,
1323                                     ggio->gctl_offset, ggio->gctl_length)) {
1324                                         pjdlog_debug(2,
1325                                             "regular: Range offset=%jd length=%zu locked.",
1326                                             (intmax_t)ggio->gctl_offset,
1327                                             (size_t)ggio->gctl_length);
1328                                         range_regular_wait = true;
1329                                         cv_wait(&range_regular_cond, &range_lock);
1330                                         range_regular_wait = false;
1331                                         mtx_unlock(&range_lock);
1332                                         continue;
1333                                 }
1334                                 if (rangelock_add(range_regular,
1335                                     ggio->gctl_offset, ggio->gctl_length) == -1) {
1336                                         mtx_unlock(&range_lock);
1337                                         pjdlog_debug(2,
1338                                             "regular: Range offset=%jd length=%zu is already locked, waiting.",
1339                                             (intmax_t)ggio->gctl_offset,
1340                                             (size_t)ggio->gctl_length);
1341                                         sleep(1);
1342                                         continue;
1343                                 }
1344                                 mtx_unlock(&range_lock);
1345                                 break;
1346                         }
1347                         mtx_lock(&res->hr_amp_lock);
1348                         if (activemap_write_start(res->hr_amp,
1349                             ggio->gctl_offset, ggio->gctl_length)) {
1350                                 res->hr_stat_activemap_update++;
1351                                 (void)hast_activemap_flush(res);
1352                         } else {
1353                                 mtx_unlock(&res->hr_amp_lock);
1354                         }
1355                         if (ISMEMSYNC(hio)) {
1356                                 hio->hio_memsyncacked = false;
1357                                 refcnt_init(&hio->hio_writecount, ncomps);
1358                         }
1359                         break;
1360                 case BIO_DELETE:
1361                         res->hr_stat_delete++;
1362                         break;
1363                 case BIO_FLUSH:
1364                         res->hr_stat_flush++;
1365                         break;
1366                 }
1367                 pjdlog_debug(2,
1368                     "ggate_recv: (%p) Moving request to the send queues.", hio);
1369                 refcnt_init(&hio->hio_countdown, ncomps);
1370                 for (ii = ncomp; ii < ncomps; ii++)
1371                         QUEUE_INSERT1(hio, send, ii);
1372         }
1373         /* NOTREACHED */
1374         return (NULL);
1375 }
1376
1377 /*
1378  * Thread reads from or writes to local component.
1379  * If local read fails, it redirects it to remote_send thread.
1380  */
1381 static void *
1382 local_send_thread(void *arg)
1383 {
1384         struct hast_resource *res = arg;
1385         struct g_gate_ctl_io *ggio;
1386         struct hio *hio;
1387         unsigned int ncomp, rncomp;
1388         ssize_t ret;
1389
1390         /* Local component is 0 for now. */
1391         ncomp = 0;
1392         /* Remote component is 1 for now. */
1393         rncomp = 1;
1394
1395         for (;;) {
1396                 pjdlog_debug(2, "local_send: Taking request.");
1397                 QUEUE_TAKE1(hio, send, ncomp, 0);
1398                 pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1399                 ggio = &hio->hio_ggio;
1400                 switch (ggio->gctl_cmd) {
1401                 case BIO_READ:
1402                         ret = pread(res->hr_localfd, ggio->gctl_data,
1403                             ggio->gctl_length,
1404                             ggio->gctl_offset + res->hr_localoff);
1405                         if (ret == ggio->gctl_length)
1406                                 hio->hio_errors[ncomp] = 0;
1407                         else if (!ISSYNCREQ(hio)) {
1408                                 /*
1409                                  * If READ failed, try to read from remote node.
1410                                  */
1411                                 if (ret == -1) {
1412                                         reqlog(LOG_WARNING, 0, ggio,
1413                                             "Local request failed (%s), trying remote node. ",
1414                                             strerror(errno));
1415                                 } else if (ret != ggio->gctl_length) {
1416                                         reqlog(LOG_WARNING, 0, ggio,
1417                                             "Local request failed (%zd != %jd), trying remote node. ",
1418                                             ret, (intmax_t)ggio->gctl_length);
1419                                 }
1420                                 QUEUE_INSERT1(hio, send, rncomp);
1421                                 continue;
1422                         }
1423                         break;
1424                 case BIO_WRITE:
1425                         ret = pwrite(res->hr_localfd, ggio->gctl_data,
1426                             ggio->gctl_length,
1427                             ggio->gctl_offset + res->hr_localoff);
1428                         if (ret == -1) {
1429                                 hio->hio_errors[ncomp] = errno;
1430                                 reqlog(LOG_WARNING, 0, ggio,
1431                                     "Local request failed (%s): ",
1432                                     strerror(errno));
1433                         } else if (ret != ggio->gctl_length) {
1434                                 hio->hio_errors[ncomp] = EIO;
1435                                 reqlog(LOG_WARNING, 0, ggio,
1436                                     "Local request failed (%zd != %jd): ",
1437                                     ret, (intmax_t)ggio->gctl_length);
1438                         } else {
1439                                 hio->hio_errors[ncomp] = 0;
1440                                 if (ISASYNC(hio)) {
1441                                         ggio->gctl_error = 0;
1442                                         write_complete(res, hio);
1443                                 }
1444                         }
1445                         break;
1446                 case BIO_DELETE:
1447                         ret = g_delete(res->hr_localfd,
1448                             ggio->gctl_offset + res->hr_localoff,
1449                             ggio->gctl_length);
1450                         if (ret == -1) {
1451                                 hio->hio_errors[ncomp] = errno;
1452                                 reqlog(LOG_WARNING, 0, ggio,
1453                                     "Local request failed (%s): ",
1454                                     strerror(errno));
1455                         } else {
1456                                 hio->hio_errors[ncomp] = 0;
1457                         }
1458                         break;
1459                 case BIO_FLUSH:
1460                         if (!res->hr_localflush) {
1461                                 ret = -1;
1462                                 errno = EOPNOTSUPP;
1463                                 break;
1464                         }
1465                         ret = g_flush(res->hr_localfd);
1466                         if (ret == -1) {
1467                                 if (errno == EOPNOTSUPP)
1468                                         res->hr_localflush = false;
1469                                 hio->hio_errors[ncomp] = errno;
1470                                 reqlog(LOG_WARNING, 0, ggio,
1471                                     "Local request failed (%s): ",
1472                                     strerror(errno));
1473                         } else {
1474                                 hio->hio_errors[ncomp] = 0;
1475                         }
1476                         break;
1477                 }
1478                 if (ISMEMSYNCWRITE(hio)) {
1479                         if (refcnt_release(&hio->hio_writecount) == 0) {
1480                                 write_complete(res, hio);
1481                         }
1482                 }
1483                 if (refcnt_release(&hio->hio_countdown) > 0)
1484                         continue;
1485                 if (ISSYNCREQ(hio)) {
1486                         mtx_lock(&sync_lock);
1487                         SYNCREQDONE(hio);
1488                         mtx_unlock(&sync_lock);
1489                         cv_signal(&sync_cond);
1490                 } else {
1491                         pjdlog_debug(2,
1492                             "local_send: (%p) Moving request to the done queue.",
1493                             hio);
1494                         QUEUE_INSERT2(hio, done);
1495                 }
1496         }
1497         /* NOTREACHED */
1498         return (NULL);
1499 }
1500
1501 static void
1502 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1503 {
1504         struct nv *nv;
1505
1506         rw_rlock(&hio_remote_lock[ncomp]);
1507
1508         if (!ISCONNECTED(res, ncomp)) {
1509                 rw_unlock(&hio_remote_lock[ncomp]);
1510                 return;
1511         }
1512
1513         PJDLOG_ASSERT(res->hr_remotein != NULL);
1514         PJDLOG_ASSERT(res->hr_remoteout != NULL);
1515
1516         nv = nv_alloc();
1517         nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1518         if (nv_error(nv) != 0) {
1519                 rw_unlock(&hio_remote_lock[ncomp]);
1520                 nv_free(nv);
1521                 pjdlog_debug(1,
1522                     "keepalive_send: Unable to prepare header to send.");
1523                 return;
1524         }
1525         if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) == -1) {
1526                 rw_unlock(&hio_remote_lock[ncomp]);
1527                 pjdlog_common(LOG_DEBUG, 1, errno,
1528                     "keepalive_send: Unable to send request");
1529                 nv_free(nv);
1530                 remote_close(res, ncomp);
1531                 return;
1532         }
1533
1534         rw_unlock(&hio_remote_lock[ncomp]);
1535         nv_free(nv);
1536         pjdlog_debug(2, "keepalive_send: Request sent.");
1537 }
1538
1539 /*
1540  * Thread sends request to secondary node.
1541  */
1542 static void *
1543 remote_send_thread(void *arg)
1544 {
1545         struct hast_resource *res = arg;
1546         struct g_gate_ctl_io *ggio;
1547         time_t lastcheck, now;
1548         struct hio *hio;
1549         struct nv *nv;
1550         unsigned int ncomp;
1551         bool wakeup;
1552         uint64_t offset, length;
1553         uint8_t cmd;
1554         void *data;
1555
1556         /* Remote component is 1 for now. */
1557         ncomp = 1;
1558         lastcheck = time(NULL);
1559
1560         for (;;) {
1561                 pjdlog_debug(2, "remote_send: Taking request.");
1562                 QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE);
1563                 if (hio == NULL) {
1564                         now = time(NULL);
1565                         if (lastcheck + HAST_KEEPALIVE <= now) {
1566                                 keepalive_send(res, ncomp);
1567                                 lastcheck = now;
1568                         }
1569                         continue;
1570                 }
1571                 pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1572                 ggio = &hio->hio_ggio;
1573                 switch (ggio->gctl_cmd) {
1574                 case BIO_READ:
1575                         cmd = HIO_READ;
1576                         data = NULL;
1577                         offset = ggio->gctl_offset;
1578                         length = ggio->gctl_length;
1579                         break;
1580                 case BIO_WRITE:
1581                         cmd = HIO_WRITE;
1582                         data = ggio->gctl_data;
1583                         offset = ggio->gctl_offset;
1584                         length = ggio->gctl_length;
1585                         break;
1586                 case BIO_DELETE:
1587                         cmd = HIO_DELETE;
1588                         data = NULL;
1589                         offset = ggio->gctl_offset;
1590                         length = ggio->gctl_length;
1591                         break;
1592                 case BIO_FLUSH:
1593                         cmd = HIO_FLUSH;
1594                         data = NULL;
1595                         offset = 0;
1596                         length = 0;
1597                         break;
1598                 default:
1599                         PJDLOG_ABORT("invalid condition");
1600                 }
1601                 nv = nv_alloc();
1602                 nv_add_uint8(nv, cmd, "cmd");
1603                 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1604                 nv_add_uint64(nv, offset, "offset");
1605                 nv_add_uint64(nv, length, "length");
1606                 if (ISMEMSYNCWRITE(hio))
1607                         nv_add_uint8(nv, 1, "memsync");
1608                 if (nv_error(nv) != 0) {
1609                         hio->hio_errors[ncomp] = nv_error(nv);
1610                         pjdlog_debug(2,
1611                             "remote_send: (%p) Unable to prepare header to send.",
1612                             hio);
1613                         reqlog(LOG_ERR, 0, ggio,
1614                             "Unable to prepare header to send (%s): ",
1615                             strerror(nv_error(nv)));
1616                         /* Move failed request immediately to the done queue. */
1617                         goto done_queue;
1618                 }
1619                 /*
1620                  * Protect connection from disappearing.
1621                  */
1622                 rw_rlock(&hio_remote_lock[ncomp]);
1623                 if (!ISCONNECTED(res, ncomp)) {
1624                         rw_unlock(&hio_remote_lock[ncomp]);
1625                         hio->hio_errors[ncomp] = ENOTCONN;
1626                         goto done_queue;
1627                 }
1628                 /*
1629                  * Move the request to recv queue before sending it, because
1630                  * in different order we can get reply before we move request
1631                  * to recv queue.
1632                  */
1633                 pjdlog_debug(2,
1634                     "remote_send: (%p) Moving request to the recv queue.",
1635                     hio);
1636                 mtx_lock(&hio_recv_list_lock[ncomp]);
1637                 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1638                 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1639                 hio_recv_list_size[ncomp]++;
1640                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1641                 if (hast_proto_send(res, res->hr_remoteout, nv, data,
1642                     data != NULL ? length : 0) == -1) {
1643                         hio->hio_errors[ncomp] = errno;
1644                         rw_unlock(&hio_remote_lock[ncomp]);
1645                         pjdlog_debug(2,
1646                             "remote_send: (%p) Unable to send request.", hio);
1647                         reqlog(LOG_ERR, 0, ggio,
1648                             "Unable to send request (%s): ",
1649                             strerror(hio->hio_errors[ncomp]));
1650                         remote_close(res, ncomp);
1651                 } else {
1652                         rw_unlock(&hio_remote_lock[ncomp]);
1653                 }
1654                 nv_free(nv);
1655                 if (wakeup)
1656                         cv_signal(&hio_recv_list_cond[ncomp]);
1657                 continue;
1658 done_queue:
1659                 nv_free(nv);
1660                 if (ISSYNCREQ(hio)) {
1661                         if (refcnt_release(&hio->hio_countdown) > 0)
1662                                 continue;
1663                         mtx_lock(&sync_lock);
1664                         SYNCREQDONE(hio);
1665                         mtx_unlock(&sync_lock);
1666                         cv_signal(&sync_cond);
1667                         continue;
1668                 }
1669                 if (ggio->gctl_cmd == BIO_WRITE) {
1670                         mtx_lock(&res->hr_amp_lock);
1671                         if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1672                             ggio->gctl_length)) {
1673                                 (void)hast_activemap_flush(res);
1674                         } else {
1675                                 mtx_unlock(&res->hr_amp_lock);
1676                         }
1677                         if (ISMEMSYNCWRITE(hio)) {
1678                                 if (refcnt_release(&hio->hio_writecount) == 0) {
1679                                         if (hio->hio_errors[0] == 0)
1680                                                 write_complete(res, hio);
1681                                 }
1682                         }
1683                 }
1684                 if (refcnt_release(&hio->hio_countdown) > 0)
1685                         continue;
1686                 pjdlog_debug(2,
1687                     "remote_send: (%p) Moving request to the done queue.",
1688                     hio);
1689                 QUEUE_INSERT2(hio, done);
1690         }
1691         /* NOTREACHED */
1692         return (NULL);
1693 }
1694
1695 /*
1696  * Thread receives answer from secondary node and passes it to ggate_send
1697  * thread.
1698  */
1699 static void *
1700 remote_recv_thread(void *arg)
1701 {
1702         struct hast_resource *res = arg;
1703         struct g_gate_ctl_io *ggio;
1704         struct hio *hio;
1705         struct nv *nv;
1706         unsigned int ncomp;
1707         uint64_t seq;
1708         bool memsyncack;
1709         int error;
1710
1711         /* Remote component is 1 for now. */
1712         ncomp = 1;
1713
1714         for (;;) {
1715                 /* Wait until there is anything to receive. */
1716                 mtx_lock(&hio_recv_list_lock[ncomp]);
1717                 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1718                         pjdlog_debug(2, "remote_recv: No requests, waiting.");
1719                         cv_wait(&hio_recv_list_cond[ncomp],
1720                             &hio_recv_list_lock[ncomp]);
1721                 }
1722                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1723
1724                 memsyncack = false;
1725
1726                 rw_rlock(&hio_remote_lock[ncomp]);
1727                 if (!ISCONNECTED(res, ncomp)) {
1728                         rw_unlock(&hio_remote_lock[ncomp]);
1729                         /*
1730                          * Connection is dead, so move all pending requests to
1731                          * the done queue (one-by-one).
1732                          */
1733                         mtx_lock(&hio_recv_list_lock[ncomp]);
1734                         hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1735                         PJDLOG_ASSERT(hio != NULL);
1736                         TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1737                             hio_next[ncomp]);
1738                         hio_recv_list_size[ncomp]--;
1739                         mtx_unlock(&hio_recv_list_lock[ncomp]);
1740                         hio->hio_errors[ncomp] = ENOTCONN;
1741                         goto done_queue;
1742                 }
1743                 if (hast_proto_recv_hdr(res->hr_remotein, &nv) == -1) {
1744                         pjdlog_errno(LOG_ERR,
1745                             "Unable to receive reply header");
1746                         rw_unlock(&hio_remote_lock[ncomp]);
1747                         remote_close(res, ncomp);
1748                         continue;
1749                 }
1750                 rw_unlock(&hio_remote_lock[ncomp]);
1751                 seq = nv_get_uint64(nv, "seq");
1752                 if (seq == 0) {
1753                         pjdlog_error("Header contains no 'seq' field.");
1754                         nv_free(nv);
1755                         continue;
1756                 }
1757                 memsyncack = nv_exists(nv, "received");
1758                 mtx_lock(&hio_recv_list_lock[ncomp]);
1759                 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1760                         if (hio->hio_ggio.gctl_seq == seq) {
1761                                 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1762                                     hio_next[ncomp]);
1763                                 hio_recv_list_size[ncomp]--;
1764                                 break;
1765                         }
1766                 }
1767                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1768                 if (hio == NULL) {
1769                         pjdlog_error("Found no request matching received 'seq' field (%ju).",
1770                             (uintmax_t)seq);
1771                         nv_free(nv);
1772                         continue;
1773                 }
1774                 ggio = &hio->hio_ggio;
1775                 error = nv_get_int16(nv, "error");
1776                 if (error != 0) {
1777                         /* Request failed on remote side. */
1778                         hio->hio_errors[ncomp] = error;
1779                         reqlog(LOG_WARNING, 0, ggio,
1780                             "Remote request failed (%s): ", strerror(error));
1781                         nv_free(nv);
1782                         goto done_queue;
1783                 }
1784                 switch (ggio->gctl_cmd) {
1785                 case BIO_READ:
1786                         rw_rlock(&hio_remote_lock[ncomp]);
1787                         if (!ISCONNECTED(res, ncomp)) {
1788                                 rw_unlock(&hio_remote_lock[ncomp]);
1789                                 nv_free(nv);
1790                                 goto done_queue;
1791                         }
1792                         if (hast_proto_recv_data(res, res->hr_remotein, nv,
1793                             ggio->gctl_data, ggio->gctl_length) == -1) {
1794                                 hio->hio_errors[ncomp] = errno;
1795                                 pjdlog_errno(LOG_ERR,
1796                                     "Unable to receive reply data");
1797                                 rw_unlock(&hio_remote_lock[ncomp]);
1798                                 nv_free(nv);
1799                                 remote_close(res, ncomp);
1800                                 goto done_queue;
1801                         }
1802                         rw_unlock(&hio_remote_lock[ncomp]);
1803                         break;
1804                 case BIO_WRITE:
1805                 case BIO_DELETE:
1806                 case BIO_FLUSH:
1807                         break;
1808                 default:
1809                         PJDLOG_ABORT("invalid condition");
1810                 }
1811                 hio->hio_errors[ncomp] = 0;
1812                 nv_free(nv);
1813 done_queue:
1814                 if (ISMEMSYNCWRITE(hio)) {
1815                         if (!hio->hio_memsyncacked) {
1816                                 PJDLOG_ASSERT(memsyncack ||
1817                                     hio->hio_errors[ncomp] != 0);
1818                                 /* Remote ack arrived. */
1819                                 if (refcnt_release(&hio->hio_writecount) == 0) {
1820                                         if (hio->hio_errors[0] == 0)
1821                                                 write_complete(res, hio);
1822                                 }
1823                                 hio->hio_memsyncacked = true;
1824                                 if (hio->hio_errors[ncomp] == 0) {
1825                                         pjdlog_debug(2,
1826                                             "remote_recv: (%p) Moving request "
1827                                             "back to the recv queue.", hio);
1828                                         mtx_lock(&hio_recv_list_lock[ncomp]);
1829                                         TAILQ_INSERT_TAIL(&hio_recv_list[ncomp],
1830                                             hio, hio_next[ncomp]);
1831                                         hio_recv_list_size[ncomp]++;
1832                                         mtx_unlock(&hio_recv_list_lock[ncomp]);
1833                                         continue;
1834                                 }
1835                         } else {
1836                                 PJDLOG_ASSERT(!memsyncack);
1837                                 /* Remote final reply arrived. */
1838                         }
1839                 }
1840                 if (refcnt_release(&hio->hio_countdown) > 0)
1841                         continue;
1842                 if (ISSYNCREQ(hio)) {
1843                         mtx_lock(&sync_lock);
1844                         SYNCREQDONE(hio);
1845                         mtx_unlock(&sync_lock);
1846                         cv_signal(&sync_cond);
1847                 } else {
1848                         pjdlog_debug(2,
1849                             "remote_recv: (%p) Moving request to the done queue.",
1850                             hio);
1851                         QUEUE_INSERT2(hio, done);
1852                 }
1853         }
1854         /* NOTREACHED */
1855         return (NULL);
1856 }
1857
1858 /*
1859  * Thread sends answer to the kernel.
1860  */
1861 static void *
1862 ggate_send_thread(void *arg)
1863 {
1864         struct hast_resource *res = arg;
1865         struct g_gate_ctl_io *ggio;
1866         struct hio *hio;
1867         unsigned int ii, ncomps;
1868
1869         ncomps = HAST_NCOMPONENTS;
1870
1871         for (;;) {
1872                 pjdlog_debug(2, "ggate_send: Taking request.");
1873                 QUEUE_TAKE2(hio, done);
1874                 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1875                 ggio = &hio->hio_ggio;
1876                 for (ii = 0; ii < ncomps; ii++) {
1877                         if (hio->hio_errors[ii] == 0) {
1878                                 /*
1879                                  * One successful request is enough to declare
1880                                  * success.
1881                                  */
1882                                 ggio->gctl_error = 0;
1883                                 break;
1884                         }
1885                 }
1886                 if (ii == ncomps) {
1887                         /*
1888                          * None of the requests were successful.
1889                          * Use the error from local component except the
1890                          * case when we did only remote request.
1891                          */
1892                         if (ggio->gctl_cmd == BIO_READ &&
1893                             res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
1894                                 ggio->gctl_error = hio->hio_errors[1];
1895                         else
1896                                 ggio->gctl_error = hio->hio_errors[0];
1897                 }
1898                 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1899                         mtx_lock(&res->hr_amp_lock);
1900                         if (activemap_write_complete(res->hr_amp,
1901                             ggio->gctl_offset, ggio->gctl_length)) {
1902                                 res->hr_stat_activemap_update++;
1903                                 (void)hast_activemap_flush(res);
1904                         } else {
1905                                 mtx_unlock(&res->hr_amp_lock);
1906                         }
1907                 }
1908                 if (ggio->gctl_cmd == BIO_WRITE) {
1909                         /*
1910                          * Unlock range we locked.
1911                          */
1912                         mtx_lock(&range_lock);
1913                         rangelock_del(range_regular, ggio->gctl_offset,
1914                             ggio->gctl_length);
1915                         if (range_sync_wait)
1916                                 cv_signal(&range_sync_cond);
1917                         mtx_unlock(&range_lock);
1918                         if (!hio->hio_done)
1919                                 write_complete(res, hio);
1920                 } else {
1921                         if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1) {
1922                                 primary_exit(EX_OSERR,
1923                                     "G_GATE_CMD_DONE failed");
1924                         }
1925                 }
1926                 if (hio->hio_errors[0]) {
1927                         switch (ggio->gctl_cmd) {
1928                         case BIO_READ:
1929                                 res->hr_stat_read_error++;
1930                                 break;
1931                         case BIO_WRITE:
1932                                 res->hr_stat_write_error++;
1933                                 break;
1934                         case BIO_DELETE:
1935                                 res->hr_stat_delete_error++;
1936                                 break;
1937                         case BIO_FLUSH:
1938                                 res->hr_stat_flush_error++;
1939                                 break;
1940                         }
1941                 }
1942                 pjdlog_debug(2,
1943                     "ggate_send: (%p) Moving request to the free queue.", hio);
1944                 QUEUE_INSERT2(hio, free);
1945         }
1946         /* NOTREACHED */
1947         return (NULL);
1948 }
1949
1950 /*
1951  * Thread synchronize local and remote components.
1952  */
1953 static void *
1954 sync_thread(void *arg __unused)
1955 {
1956         struct hast_resource *res = arg;
1957         struct hio *hio;
1958         struct g_gate_ctl_io *ggio;
1959         struct timeval tstart, tend, tdiff;
1960         unsigned int ii, ncomp, ncomps;
1961         off_t offset, length, synced;
1962         bool dorewind, directreads;
1963         int syncext;
1964
1965         ncomps = HAST_NCOMPONENTS;
1966         dorewind = true;
1967         synced = 0;
1968         offset = -1;
1969         directreads = false;
1970
1971         for (;;) {
1972                 mtx_lock(&sync_lock);
1973                 if (offset >= 0 && !sync_inprogress) {
1974                         gettimeofday(&tend, NULL);
1975                         timersub(&tend, &tstart, &tdiff);
1976                         pjdlog_info("Synchronization interrupted after %#.0T. "
1977                             "%NB synchronized so far.", &tdiff,
1978                             (intmax_t)synced);
1979                         event_send(res, EVENT_SYNCINTR);
1980                 }
1981                 while (!sync_inprogress) {
1982                         dorewind = true;
1983                         synced = 0;
1984                         cv_wait(&sync_cond, &sync_lock);
1985                 }
1986                 mtx_unlock(&sync_lock);
1987                 /*
1988                  * Obtain offset at which we should synchronize.
1989                  * Rewind synchronization if needed.
1990                  */
1991                 mtx_lock(&res->hr_amp_lock);
1992                 if (dorewind)
1993                         activemap_sync_rewind(res->hr_amp);
1994                 offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1995                 if (syncext != -1) {
1996                         /*
1997                          * We synchronized entire syncext extent, we can mark
1998                          * it as clean now.
1999                          */
2000                         if (activemap_extent_complete(res->hr_amp, syncext))
2001                                 (void)hast_activemap_flush(res);
2002                         else
2003                                 mtx_unlock(&res->hr_amp_lock);
2004                 } else {
2005                         mtx_unlock(&res->hr_amp_lock);
2006                 }
2007                 if (dorewind) {
2008                         dorewind = false;
2009                         if (offset == -1)
2010                                 pjdlog_info("Nodes are in sync.");
2011                         else {
2012                                 pjdlog_info("Synchronization started. %NB to go.",
2013                                     (intmax_t)(res->hr_extentsize *
2014                                     activemap_ndirty(res->hr_amp)));
2015                                 event_send(res, EVENT_SYNCSTART);
2016                                 gettimeofday(&tstart, NULL);
2017                         }
2018                 }
2019                 if (offset == -1) {
2020                         sync_stop();
2021                         pjdlog_debug(1, "Nothing to synchronize.");
2022                         /*
2023                          * Synchronization complete, make both localcnt and
2024                          * remotecnt equal.
2025                          */
2026                         ncomp = 1;
2027                         rw_rlock(&hio_remote_lock[ncomp]);
2028                         if (ISCONNECTED(res, ncomp)) {
2029                                 if (synced > 0) {
2030                                         int64_t bps;
2031
2032                                         gettimeofday(&tend, NULL);
2033                                         timersub(&tend, &tstart, &tdiff);
2034                                         bps = (int64_t)((double)synced /
2035                                             ((double)tdiff.tv_sec +
2036                                             (double)tdiff.tv_usec / 1000000));
2037                                         pjdlog_info("Synchronization complete. "
2038                                             "%NB synchronized in %#.0lT (%NB/sec).",
2039                                             (intmax_t)synced, &tdiff,
2040                                             (intmax_t)bps);
2041                                         event_send(res, EVENT_SYNCDONE);
2042                                 }
2043                                 mtx_lock(&metadata_lock);
2044                                 if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
2045                                         directreads = true;
2046                                 res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
2047                                 res->hr_primary_localcnt =
2048                                     res->hr_secondary_remotecnt;
2049                                 res->hr_primary_remotecnt =
2050                                     res->hr_secondary_localcnt;
2051                                 pjdlog_debug(1,
2052                                     "Setting localcnt to %ju and remotecnt to %ju.",
2053                                     (uintmax_t)res->hr_primary_localcnt,
2054                                     (uintmax_t)res->hr_primary_remotecnt);
2055                                 (void)metadata_write(res);
2056                                 mtx_unlock(&metadata_lock);
2057                         }
2058                         rw_unlock(&hio_remote_lock[ncomp]);
2059                         if (directreads) {
2060                                 directreads = false;
2061                                 enable_direct_reads(res);
2062                         }
2063                         continue;
2064                 }
2065                 pjdlog_debug(2, "sync: Taking free request.");
2066                 QUEUE_TAKE2(hio, free);
2067                 pjdlog_debug(2, "sync: (%p) Got free request.", hio);
2068                 /*
2069                  * Lock the range we are going to synchronize. We don't want
2070                  * race where someone writes between our read and write.
2071                  */
2072                 for (;;) {
2073                         mtx_lock(&range_lock);
2074                         if (rangelock_islocked(range_regular, offset, length)) {
2075                                 pjdlog_debug(2,
2076                                     "sync: Range offset=%jd length=%jd locked.",
2077                                     (intmax_t)offset, (intmax_t)length);
2078                                 range_sync_wait = true;
2079                                 cv_wait(&range_sync_cond, &range_lock);
2080                                 range_sync_wait = false;
2081                                 mtx_unlock(&range_lock);
2082                                 continue;
2083                         }
2084                         if (rangelock_add(range_sync, offset, length) == -1) {
2085                                 mtx_unlock(&range_lock);
2086                                 pjdlog_debug(2,
2087                                     "sync: Range offset=%jd length=%jd is already locked, waiting.",
2088                                     (intmax_t)offset, (intmax_t)length);
2089                                 sleep(1);
2090                                 continue;
2091                         }
2092                         mtx_unlock(&range_lock);
2093                         break;
2094                 }
2095                 /*
2096                  * First read the data from synchronization source.
2097                  */
2098                 SYNCREQ(hio);
2099                 ggio = &hio->hio_ggio;
2100                 ggio->gctl_cmd = BIO_READ;
2101                 ggio->gctl_offset = offset;
2102                 ggio->gctl_length = length;
2103                 ggio->gctl_error = 0;
2104                 hio->hio_done = false;
2105                 hio->hio_replication = res->hr_replication;
2106                 for (ii = 0; ii < ncomps; ii++)
2107                         hio->hio_errors[ii] = EINVAL;
2108                 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2109                     hio);
2110                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2111                     hio);
2112                 mtx_lock(&metadata_lock);
2113                 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2114                         /*
2115                          * This range is up-to-date on local component,
2116                          * so handle request locally.
2117                          */
2118                          /* Local component is 0 for now. */
2119                         ncomp = 0;
2120                 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2121                         PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2122                         /*
2123                          * This range is out-of-date on local component,
2124                          * so send request to the remote node.
2125                          */
2126                          /* Remote component is 1 for now. */
2127                         ncomp = 1;
2128                 }
2129                 mtx_unlock(&metadata_lock);
2130                 refcnt_init(&hio->hio_countdown, 1);
2131                 QUEUE_INSERT1(hio, send, ncomp);
2132
2133                 /*
2134                  * Let's wait for READ to finish.
2135                  */
2136                 mtx_lock(&sync_lock);
2137                 while (!ISSYNCREQDONE(hio))
2138                         cv_wait(&sync_cond, &sync_lock);
2139                 mtx_unlock(&sync_lock);
2140
2141                 if (hio->hio_errors[ncomp] != 0) {
2142                         pjdlog_error("Unable to read synchronization data: %s.",
2143                             strerror(hio->hio_errors[ncomp]));
2144                         goto free_queue;
2145                 }
2146
2147                 /*
2148                  * We read the data from synchronization source, now write it
2149                  * to synchronization target.
2150                  */
2151                 SYNCREQ(hio);
2152                 ggio->gctl_cmd = BIO_WRITE;
2153                 for (ii = 0; ii < ncomps; ii++)
2154                         hio->hio_errors[ii] = EINVAL;
2155                 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2156                     hio);
2157                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2158                     hio);
2159                 mtx_lock(&metadata_lock);
2160                 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2161                         /*
2162                          * This range is up-to-date on local component,
2163                          * so we update remote component.
2164                          */
2165                          /* Remote component is 1 for now. */
2166                         ncomp = 1;
2167                 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2168                         PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2169                         /*
2170                          * This range is out-of-date on local component,
2171                          * so we update it.
2172                          */
2173                          /* Local component is 0 for now. */
2174                         ncomp = 0;
2175                 }
2176                 mtx_unlock(&metadata_lock);
2177
2178                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2179                     hio);
2180                 refcnt_init(&hio->hio_countdown, 1);
2181                 QUEUE_INSERT1(hio, send, ncomp);
2182
2183                 /*
2184                  * Let's wait for WRITE to finish.
2185                  */
2186                 mtx_lock(&sync_lock);
2187                 while (!ISSYNCREQDONE(hio))
2188                         cv_wait(&sync_cond, &sync_lock);
2189                 mtx_unlock(&sync_lock);
2190
2191                 if (hio->hio_errors[ncomp] != 0) {
2192                         pjdlog_error("Unable to write synchronization data: %s.",
2193                             strerror(hio->hio_errors[ncomp]));
2194                         goto free_queue;
2195                 }
2196
2197                 synced += length;
2198 free_queue:
2199                 mtx_lock(&range_lock);
2200                 rangelock_del(range_sync, offset, length);
2201                 if (range_regular_wait)
2202                         cv_signal(&range_regular_cond);
2203                 mtx_unlock(&range_lock);
2204                 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
2205                     hio);
2206                 QUEUE_INSERT2(hio, free);
2207         }
2208         /* NOTREACHED */
2209         return (NULL);
2210 }
2211
2212 void
2213 primary_config_reload(struct hast_resource *res, struct nv *nv)
2214 {
2215         unsigned int ii, ncomps;
2216         int modified, vint;
2217         const char *vstr;
2218
2219         pjdlog_info("Reloading configuration...");
2220
2221         PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
2222         PJDLOG_ASSERT(gres == res);
2223         nv_assert(nv, "remoteaddr");
2224         nv_assert(nv, "sourceaddr");
2225         nv_assert(nv, "replication");
2226         nv_assert(nv, "checksum");
2227         nv_assert(nv, "compression");
2228         nv_assert(nv, "timeout");
2229         nv_assert(nv, "exec");
2230         nv_assert(nv, "metaflush");
2231
2232         ncomps = HAST_NCOMPONENTS;
2233
2234 #define MODIFIED_REMOTEADDR     0x01
2235 #define MODIFIED_SOURCEADDR     0x02
2236 #define MODIFIED_REPLICATION    0x04
2237 #define MODIFIED_CHECKSUM       0x08
2238 #define MODIFIED_COMPRESSION    0x10
2239 #define MODIFIED_TIMEOUT        0x20
2240 #define MODIFIED_EXEC           0x40
2241 #define MODIFIED_METAFLUSH      0x80
2242         modified = 0;
2243
2244         vstr = nv_get_string(nv, "remoteaddr");
2245         if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
2246                 /*
2247                  * Don't copy res->hr_remoteaddr to gres just yet.
2248                  * We want remote_close() to log disconnect from the old
2249                  * addresses, not from the new ones.
2250                  */
2251                 modified |= MODIFIED_REMOTEADDR;
2252         }
2253         vstr = nv_get_string(nv, "sourceaddr");
2254         if (strcmp(gres->hr_sourceaddr, vstr) != 0) {
2255                 strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr));
2256                 modified |= MODIFIED_SOURCEADDR;
2257         }
2258         vint = nv_get_int32(nv, "replication");
2259         if (gres->hr_replication != vint) {
2260                 gres->hr_replication = vint;
2261                 modified |= MODIFIED_REPLICATION;
2262         }
2263         vint = nv_get_int32(nv, "checksum");
2264         if (gres->hr_checksum != vint) {
2265                 gres->hr_checksum = vint;
2266                 modified |= MODIFIED_CHECKSUM;
2267         }
2268         vint = nv_get_int32(nv, "compression");
2269         if (gres->hr_compression != vint) {
2270                 gres->hr_compression = vint;
2271                 modified |= MODIFIED_COMPRESSION;
2272         }
2273         vint = nv_get_int32(nv, "timeout");
2274         if (gres->hr_timeout != vint) {
2275                 gres->hr_timeout = vint;
2276                 modified |= MODIFIED_TIMEOUT;
2277         }
2278         vstr = nv_get_string(nv, "exec");
2279         if (strcmp(gres->hr_exec, vstr) != 0) {
2280                 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
2281                 modified |= MODIFIED_EXEC;
2282         }
2283         vint = nv_get_int32(nv, "metaflush");
2284         if (gres->hr_metaflush != vint) {
2285                 gres->hr_metaflush = vint;
2286                 modified |= MODIFIED_METAFLUSH;
2287         }
2288
2289         /*
2290          * Change timeout for connected sockets.
2291          * Don't bother if we need to reconnect.
2292          */
2293         if ((modified & MODIFIED_TIMEOUT) != 0 &&
2294             (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) == 0) {
2295                 for (ii = 0; ii < ncomps; ii++) {
2296                         if (!ISREMOTE(ii))
2297                                 continue;
2298                         rw_rlock(&hio_remote_lock[ii]);
2299                         if (!ISCONNECTED(gres, ii)) {
2300                                 rw_unlock(&hio_remote_lock[ii]);
2301                                 continue;
2302                         }
2303                         rw_unlock(&hio_remote_lock[ii]);
2304                         if (proto_timeout(gres->hr_remotein,
2305                             gres->hr_timeout) == -1) {
2306                                 pjdlog_errno(LOG_WARNING,
2307                                     "Unable to set connection timeout");
2308                         }
2309                         if (proto_timeout(gres->hr_remoteout,
2310                             gres->hr_timeout) == -1) {
2311                                 pjdlog_errno(LOG_WARNING,
2312                                     "Unable to set connection timeout");
2313                         }
2314                 }
2315         }
2316         if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) != 0) {
2317                 for (ii = 0; ii < ncomps; ii++) {
2318                         if (!ISREMOTE(ii))
2319                                 continue;
2320                         remote_close(gres, ii);
2321                 }
2322                 if (modified & MODIFIED_REMOTEADDR) {
2323                         vstr = nv_get_string(nv, "remoteaddr");
2324                         strlcpy(gres->hr_remoteaddr, vstr,
2325                             sizeof(gres->hr_remoteaddr));
2326                 }
2327         }
2328 #undef  MODIFIED_REMOTEADDR
2329 #undef  MODIFIED_SOURCEADDR
2330 #undef  MODIFIED_REPLICATION
2331 #undef  MODIFIED_CHECKSUM
2332 #undef  MODIFIED_COMPRESSION
2333 #undef  MODIFIED_TIMEOUT
2334 #undef  MODIFIED_EXEC
2335 #undef  MODIFIED_METAFLUSH
2336
2337         pjdlog_info("Configuration reloaded successfully.");
2338 }
2339
2340 static void
2341 guard_one(struct hast_resource *res, unsigned int ncomp)
2342 {
2343         struct proto_conn *in, *out;
2344
2345         if (!ISREMOTE(ncomp))
2346                 return;
2347
2348         rw_rlock(&hio_remote_lock[ncomp]);
2349
2350         if (!real_remote(res)) {
2351                 rw_unlock(&hio_remote_lock[ncomp]);
2352                 return;
2353         }
2354
2355         if (ISCONNECTED(res, ncomp)) {
2356                 PJDLOG_ASSERT(res->hr_remotein != NULL);
2357                 PJDLOG_ASSERT(res->hr_remoteout != NULL);
2358                 rw_unlock(&hio_remote_lock[ncomp]);
2359                 pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2360                     res->hr_remoteaddr);
2361                 return;
2362         }
2363
2364         PJDLOG_ASSERT(res->hr_remotein == NULL);
2365         PJDLOG_ASSERT(res->hr_remoteout == NULL);
2366         /*
2367          * Upgrade the lock. It doesn't have to be atomic as no other thread
2368          * can change connection status from disconnected to connected.
2369          */
2370         rw_unlock(&hio_remote_lock[ncomp]);
2371         pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2372             res->hr_remoteaddr);
2373         in = out = NULL;
2374         if (init_remote(res, &in, &out) == 0) {
2375                 rw_wlock(&hio_remote_lock[ncomp]);
2376                 PJDLOG_ASSERT(res->hr_remotein == NULL);
2377                 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2378                 PJDLOG_ASSERT(in != NULL && out != NULL);
2379                 res->hr_remotein = in;
2380                 res->hr_remoteout = out;
2381                 rw_unlock(&hio_remote_lock[ncomp]);
2382                 pjdlog_info("Successfully reconnected to %s.",
2383                     res->hr_remoteaddr);
2384                 sync_start();
2385         } else {
2386                 /* Both connections should be NULL. */
2387                 PJDLOG_ASSERT(res->hr_remotein == NULL);
2388                 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2389                 PJDLOG_ASSERT(in == NULL && out == NULL);
2390                 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2391                     res->hr_remoteaddr);
2392         }
2393 }
2394
2395 /*
2396  * Thread guards remote connections and reconnects when needed, handles
2397  * signals, etc.
2398  */
2399 static void *
2400 guard_thread(void *arg)
2401 {
2402         struct hast_resource *res = arg;
2403         unsigned int ii, ncomps;
2404         struct timespec timeout;
2405         time_t lastcheck, now;
2406         sigset_t mask;
2407         int signo;
2408
2409         ncomps = HAST_NCOMPONENTS;
2410         lastcheck = time(NULL);
2411
2412         PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2413         PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2414         PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2415
2416         timeout.tv_sec = HAST_KEEPALIVE;
2417         timeout.tv_nsec = 0;
2418         signo = -1;
2419
2420         for (;;) {
2421                 switch (signo) {
2422                 case SIGINT:
2423                 case SIGTERM:
2424                         sigexit_received = true;
2425                         primary_exitx(EX_OK,
2426                             "Termination signal received, exiting.");
2427                         break;
2428                 default:
2429                         break;
2430                 }
2431
2432                 /*
2433                  * Don't check connections until we fully started,
2434                  * as we may still be looping, waiting for remote node
2435                  * to switch from primary to secondary.
2436                  */
2437                 if (fullystarted) {
2438                         pjdlog_debug(2, "remote_guard: Checking connections.");
2439                         now = time(NULL);
2440                         if (lastcheck + HAST_KEEPALIVE <= now) {
2441                                 for (ii = 0; ii < ncomps; ii++)
2442                                         guard_one(res, ii);
2443                                 lastcheck = now;
2444                         }
2445                 }
2446                 signo = sigtimedwait(&mask, NULL, &timeout);
2447         }
2448         /* NOTREACHED */
2449         return (NULL);
2450 }