sbin/hastd/primary.c

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