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