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/stat.h>
  39
  40 #include <geom/gate/g_gate.h>
  41
  42 #include <err.h>
  43 #include <errno.h>
  44 #include <fcntl.h>
  45 #include <libgeom.h>
  46 #include <pthread.h>
  47 #include <signal.h>
  48 #include <stdint.h>
  49 #include <stdio.h>
  50 #include <string.h>
  51 #include <sysexits.h>
  52 #include <unistd.h>
  53
  54 #include <activemap.h>
  55 #include <nv.h>
  56 #include <rangelock.h>
  57
  58 #include "control.h"
  59 #include "event.h"
  60 #include "hast.h"
  61 #include "hast_proto.h"
  62 #include "hastd.h"
  63 #include "hooks.h"
  64 #include "metadata.h"
  65 #include "proto.h"
  66 #include "pjdlog.h"
  67 #include "refcnt.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         refcnt_t                 hio_countdown;
  82         /*
  83          * Each component has a place to store its own error.
  84          * Once the request is handled by all components we can decide if the
  85          * request overall is successful or not.
  86          */
  87         int                     *hio_errors;
  88         /*
  89          * Structure used to communicate with GEOM Gate class.
  90          */
  91         struct g_gate_ctl_io     hio_ggio;
  92         /*
  93          * Request was already confirmed to GEOM Gate.
  94          */
  95         bool                     hio_done;
  96         /*
  97          * Number of components we are still waiting before sending write
  98          * completion ack to GEOM Gate. Used for memsync.
  99          */
 100         refcnt_t                 hio_writecount;
 101         /*
 102          * Memsync request was acknowleged by remote.
 103          */
 104         bool                     hio_memsyncacked;
 105         /*
 106          * Remember replication from the time the request was initiated,
 107          * so we won't get confused when replication changes on reload.
 108          */
 109         int                      hio_replication;
 110         TAILQ_ENTRY(hio)        *hio_next;
 111 };
 112 #define hio_free_next   hio_next[0]
 113 #define hio_done_next   hio_next[0]
 114
 115 /*
 116  * Free list holds unused structures. When free list is empty, we have to wait
 117  * until some in-progress requests are freed.
 118  */
 119 static TAILQ_HEAD(, hio) hio_free_list;
 120 static size_t hio_free_list_size;
 121 static pthread_mutex_t hio_free_list_lock;
 122 static pthread_cond_t hio_free_list_cond;
 123 /*
 124  * There is one send list for every component. One requests is placed on all
 125  * send lists - each component gets the same request, but each component is
 126  * responsible for managing his own send list.
 127  */
 128 static TAILQ_HEAD(, hio) *hio_send_list;
 129 static size_t *hio_send_list_size;
 130 static pthread_mutex_t *hio_send_list_lock;
 131 static pthread_cond_t *hio_send_list_cond;
 132 #define hio_send_local_list_size        hio_send_list_size[0]
 133 #define hio_send_remote_list_size       hio_send_list_size[1]
 134 /*
 135  * There is one recv list for every component, although local components don't
 136  * use recv lists as local requests are done synchronously.
 137  */
 138 static TAILQ_HEAD(, hio) *hio_recv_list;
 139 static size_t *hio_recv_list_size;
 140 static pthread_mutex_t *hio_recv_list_lock;
 141 static pthread_cond_t *hio_recv_list_cond;
 142 #define hio_recv_remote_list_size       hio_recv_list_size[1]
 143 /*
 144  * Request is placed on done list by the slowest component (the one that
 145  * decreased hio_countdown from 1 to 0).
 146  */
 147 static TAILQ_HEAD(, hio) hio_done_list;
 148 static size_t hio_done_list_size;
 149 static pthread_mutex_t hio_done_list_lock;
 150 static pthread_cond_t hio_done_list_cond;
 151 /*
 152  * Structure below are for interaction with sync thread.
 153  */
 154 static bool sync_inprogress;
 155 static pthread_mutex_t sync_lock;
 156 static pthread_cond_t sync_cond;
 157 /*
 158  * The lock below allows to synchornize access to remote connections.
 159  */
 160 static pthread_rwlock_t *hio_remote_lock;
 161
 162 /*
 163  * Lock to synchronize metadata updates. Also synchronize access to
 164  * hr_primary_localcnt and hr_primary_remotecnt fields.
 165  */
 166 static pthread_mutex_t metadata_lock;
 167
 168 /*
 169  * Maximum number of outstanding I/O requests.
 170  */
 171 #define HAST_HIO_MAX    256
 172 /*
 173  * Number of components. At this point there are only two components: local
 174  * and remote, but in the future it might be possible to use multiple local
 175  * and remote components.
 176  */
 177 #define HAST_NCOMPONENTS        2
 178
 179 #define ISCONNECTED(res, no)    \
 180         ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
 181
 182 #define QUEUE_INSERT1(hio, name, ncomp) do {                            \
 183         mtx_lock(&hio_##name##_list_lock[(ncomp)]);                     \
 184         if (TAILQ_EMPTY(&hio_##name##_list[(ncomp)]))                   \
 185                 cv_broadcast(&hio_##name##_list_cond[(ncomp)]);         \
 186         TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),           \
 187             hio_next[(ncomp)]);                                         \
 188         hio_##name##_list_size[(ncomp)]++;                              \
 189         mtx_unlock(&hio_##name##_list_lock[(ncomp)]);                   \
 190 } while (0)
 191 #define QUEUE_INSERT2(hio, name)        do {                            \
 192         mtx_lock(&hio_##name##_list_lock);                              \
 193         if (TAILQ_EMPTY(&hio_##name##_list))                            \
 194                 cv_broadcast(&hio_##name##_list_cond);                  \
 195         TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
 196         hio_##name##_list_size++;                                       \
 197         mtx_unlock(&hio_##name##_list_lock);                            \
 198 } while (0)
 199 #define QUEUE_TAKE1(hio, name, ncomp, timeout)  do {                    \
 200         bool _last;                                                     \
 201                                                                         \
 202         mtx_lock(&hio_##name##_list_lock[(ncomp)]);                     \
 203         _last = false;                                                  \
 204         while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
 205                 cv_timedwait(&hio_##name##_list_cond[(ncomp)],          \
 206                     &hio_##name##_list_lock[(ncomp)], (timeout));       \
 207                 if ((timeout) != 0)                                     \
 208                         _last = true;                                   \
 209         }                                                               \
 210         if (hio != NULL) {                                              \
 211                 PJDLOG_ASSERT(hio_##name##_list_size[(ncomp)] != 0);    \
 212                 hio_##name##_list_size[(ncomp)]--;                      \
 213                 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),        \
 214                     hio_next[(ncomp)]);                                 \
 215         }                                                               \
 216         mtx_unlock(&hio_##name##_list_lock[(ncomp)]);                   \
 217 } while (0)
 218 #define QUEUE_TAKE2(hio, name)  do {                                    \
 219         mtx_lock(&hio_##name##_list_lock);                              \
 220         while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {     \
 221                 cv_wait(&hio_##name##_list_cond,                        \
 222                     &hio_##name##_list_lock);                           \
 223         }                                                               \
 224         PJDLOG_ASSERT(hio_##name##_list_size != 0);                     \
 225         hio_##name##_list_size--;                                       \
 226         TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);     \
 227         mtx_unlock(&hio_##name##_list_lock);                            \
 228 } while (0)
 229
 230 #define ISFULLSYNC(hio) ((hio)->hio_replication == HAST_REPLICATION_FULLSYNC)
 231 #define ISMEMSYNC(hio)  ((hio)->hio_replication == HAST_REPLICATION_MEMSYNC)
 232 #define ISASYNC(hio)    ((hio)->hio_replication == HAST_REPLICATION_ASYNC)
 233
 234 #define SYNCREQ(hio)            do {                                    \
 235         (hio)->hio_ggio.gctl_unit = -1;                                 \
 236         (hio)->hio_ggio.gctl_seq = 1;                                   \
 237 } while (0)
 238 #define ISSYNCREQ(hio)          ((hio)->hio_ggio.gctl_unit == -1)
 239 #define SYNCREQDONE(hio)        do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
 240 #define ISSYNCREQDONE(hio)      ((hio)->hio_ggio.gctl_unit == -2)
 241
 242 #define ISMEMSYNCWRITE(hio)     (ISMEMSYNC(hio) &&                      \
 243             (hio)->hio_ggio.gctl_cmd == BIO_WRITE && !ISSYNCREQ(hio))
 244
 245 static struct hast_resource *gres;
 246
 247 static pthread_mutex_t range_lock;
 248 static struct rangelocks *range_regular;
 249 static bool range_regular_wait;
 250 static pthread_cond_t range_regular_cond;
 251 static struct rangelocks *range_sync;
 252 static bool range_sync_wait;
 253 static pthread_cond_t range_sync_cond;
 254 static bool fullystarted;
 255
 256 static void *ggate_recv_thread(void *arg);
 257 static void *local_send_thread(void *arg);
 258 static void *remote_send_thread(void *arg);
 259 static void *remote_recv_thread(void *arg);
 260 static void *ggate_send_thread(void *arg);
 261 static void *sync_thread(void *arg);
 262 static void *guard_thread(void *arg);
 263
 264 static void
 265 output_status_aux(struct nv *nvout)
 266 {
 267
 268         nv_add_uint64(nvout, (uint64_t)hio_free_list_size,
 269             "idle_queue_size");
 270         nv_add_uint64(nvout, (uint64_t)hio_send_local_list_size,
 271             "local_queue_size");
 272         nv_add_uint64(nvout, (uint64_t)hio_send_remote_list_size,
 273             "send_queue_size");
 274         nv_add_uint64(nvout, (uint64_t)hio_recv_remote_list_size,
 275             "recv_queue_size");
 276         nv_add_uint64(nvout, (uint64_t)hio_done_list_size,
 277             "done_queue_size");
 278 }
 279
 280 static void
 281 cleanup(struct hast_resource *res)
 282 {
 283         int rerrno;
 284
 285         /* Remember errno. */
 286         rerrno = errno;
 287
 288         /* Destroy ggate provider if we created one. */
 289         if (res->hr_ggateunit >= 0) {
 290                 struct g_gate_ctl_destroy ggiod;
 291
 292                 bzero(&ggiod, sizeof(ggiod));
 293                 ggiod.gctl_version = G_GATE_VERSION;
 294                 ggiod.gctl_unit = res->hr_ggateunit;
 295                 ggiod.gctl_force = 1;
 296                 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) == -1) {
 297                         pjdlog_errno(LOG_WARNING,
 298                             "Unable to destroy hast/%s device",
 299                             res->hr_provname);
 300                 }
 301                 res->hr_ggateunit = -1;
 302         }
 303
 304         /* Restore errno. */
 305         errno = rerrno;
 306 }
 307
 308 static __dead2 void
 309 primary_exit(int exitcode, const char *fmt, ...)
 310 {
 311         va_list ap;
 312
 313         PJDLOG_ASSERT(exitcode != EX_OK);
 314         va_start(ap, fmt);
 315         pjdlogv_errno(LOG_ERR, fmt, ap);
 316         va_end(ap);
 317         cleanup(gres);
 318         exit(exitcode);
 319 }
 320
 321 static __dead2 void
 322 primary_exitx(int exitcode, const char *fmt, ...)
 323 {
 324         va_list ap;
 325
 326         va_start(ap, fmt);
 327         pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
 328         va_end(ap);
 329         cleanup(gres);
 330         exit(exitcode);
 331 }
 332
 333 /* Expects res->hr_amp locked, returns unlocked. */
 334 static int
 335 hast_activemap_flush(struct hast_resource *res)
 336 {
 337         const unsigned char *buf;
 338         size_t size;
 339         int ret;
 340
 341         mtx_lock(&res->hr_amp_diskmap_lock);
 342         buf = activemap_bitmap(res->hr_amp, &size);
 343         mtx_unlock(&res->hr_amp_lock);
 344         PJDLOG_ASSERT(buf != NULL);
 345         PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
 346         ret = 0;
 347         if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
 348             (ssize_t)size) {
 349                 pjdlog_errno(LOG_ERR, "Unable to flush activemap to disk");
 350                 res->hr_stat_activemap_write_error++;
 351                 ret = -1;
 352         }
 353         if (ret == 0 && res->hr_metaflush == 1 &&
 354             g_flush(res->hr_localfd) == -1) {
 355                 if (errno == EOPNOTSUPP) {
 356                         pjdlog_warning("The %s provider doesn't support flushing write cache. Disabling it.",
 357                             res->hr_localpath);
 358                         res->hr_metaflush = 0;
 359                 } else {
 360                         pjdlog_errno(LOG_ERR,
 361                             "Unable to flush disk cache on activemap update");
 362                         res->hr_stat_activemap_flush_error++;
 363                         ret = -1;
 364                 }
 365         }
 366         mtx_unlock(&res->hr_amp_diskmap_lock);
 367         return (ret);
 368 }
 369
 370 static bool
 371 real_remote(const struct hast_resource *res)
 372 {
 373
 374         return (strcmp(res->hr_remoteaddr, "none") != 0);
 375 }
 376
 377 static void
 378 init_environment(struct hast_resource *res __unused)
 379 {
 380         struct hio *hio;
 381         unsigned int ii, ncomps;
 382
 383         /*
 384          * In the future it might be per-resource value.
 385          */
 386         ncomps = HAST_NCOMPONENTS;
 387
 388         /*
 389          * Allocate memory needed by lists.
 390          */
 391         hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
 392         if (hio_send_list == NULL) {
 393                 primary_exitx(EX_TEMPFAIL,
 394                     "Unable to allocate %zu bytes of memory for send lists.",
 395                     sizeof(hio_send_list[0]) * ncomps);
 396         }
 397         hio_send_list_size = malloc(sizeof(hio_send_list_size[0]) * ncomps);
 398         if (hio_send_list_size == NULL) {
 399                 primary_exitx(EX_TEMPFAIL,
 400                     "Unable to allocate %zu bytes of memory for send list counters.",
 401                     sizeof(hio_send_list_size[0]) * ncomps);
 402         }
 403         hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
 404         if (hio_send_list_lock == NULL) {
 405                 primary_exitx(EX_TEMPFAIL,
 406                     "Unable to allocate %zu bytes of memory for send list locks.",
 407                     sizeof(hio_send_list_lock[0]) * ncomps);
 408         }
 409         hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
 410         if (hio_send_list_cond == NULL) {
 411                 primary_exitx(EX_TEMPFAIL,
 412                     "Unable to allocate %zu bytes of memory for send list condition variables.",
 413                     sizeof(hio_send_list_cond[0]) * ncomps);
 414         }
 415         hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
 416         if (hio_recv_list == NULL) {
 417                 primary_exitx(EX_TEMPFAIL,
 418                     "Unable to allocate %zu bytes of memory for recv lists.",
 419                     sizeof(hio_recv_list[0]) * ncomps);
 420         }
 421         hio_recv_list_size = malloc(sizeof(hio_recv_list_size[0]) * ncomps);
 422         if (hio_recv_list_size == NULL) {
 423                 primary_exitx(EX_TEMPFAIL,
 424                     "Unable to allocate %zu bytes of memory for recv list counters.",
 425                     sizeof(hio_recv_list_size[0]) * ncomps);
 426         }
 427         hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
 428         if (hio_recv_list_lock == NULL) {
 429                 primary_exitx(EX_TEMPFAIL,
 430                     "Unable to allocate %zu bytes of memory for recv list locks.",
 431                     sizeof(hio_recv_list_lock[0]) * ncomps);
 432         }
 433         hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
 434         if (hio_recv_list_cond == NULL) {
 435                 primary_exitx(EX_TEMPFAIL,
 436                     "Unable to allocate %zu bytes of memory for recv list condition variables.",
 437                     sizeof(hio_recv_list_cond[0]) * ncomps);
 438         }
 439         hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
 440         if (hio_remote_lock == NULL) {
 441                 primary_exitx(EX_TEMPFAIL,
 442                     "Unable to allocate %zu bytes of memory for remote connections locks.",
 443                     sizeof(hio_remote_lock[0]) * ncomps);
 444         }
 445
 446         /*
 447          * Initialize lists, their counters, locks and condition variables.
 448          */
 449         TAILQ_INIT(&hio_free_list);
 450         mtx_init(&hio_free_list_lock);
 451         cv_init(&hio_free_list_cond);
 452         for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
 453                 TAILQ_INIT(&hio_send_list[ii]);
 454                 hio_send_list_size[ii] = 0;
 455                 mtx_init(&hio_send_list_lock[ii]);
 456                 cv_init(&hio_send_list_cond[ii]);
 457                 TAILQ_INIT(&hio_recv_list[ii]);
 458                 hio_recv_list_size[ii] = 0;
 459                 mtx_init(&hio_recv_list_lock[ii]);
 460                 cv_init(&hio_recv_list_cond[ii]);
 461                 rw_init(&hio_remote_lock[ii]);
 462         }
 463         TAILQ_INIT(&hio_done_list);
 464         mtx_init(&hio_done_list_lock);
 465         cv_init(&hio_done_list_cond);
 466         mtx_init(&metadata_lock);
 467
 468         /*
 469          * Allocate requests pool and initialize requests.
 470          */
 471         for (ii = 0; ii < HAST_HIO_MAX; ii++) {
 472                 hio = malloc(sizeof(*hio));
 473                 if (hio == NULL) {
 474                         primary_exitx(EX_TEMPFAIL,
 475                             "Unable to allocate %zu bytes of memory for hio request.",
 476                             sizeof(*hio));
 477                 }
 478                 refcnt_init(&hio->hio_countdown, 0);
 479                 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
 480                 if (hio->hio_errors == NULL) {
 481                         primary_exitx(EX_TEMPFAIL,
 482                             "Unable allocate %zu bytes of memory for hio errors.",
 483                             sizeof(hio->hio_errors[0]) * ncomps);
 484                 }
 485                 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
 486                 if (hio->hio_next == NULL) {
 487                         primary_exitx(EX_TEMPFAIL,
 488                             "Unable allocate %zu bytes of memory for hio_next field.",
 489                             sizeof(hio->hio_next[0]) * ncomps);
 490                 }
 491                 hio->hio_ggio.gctl_version = G_GATE_VERSION;
 492                 hio->hio_ggio.gctl_data = malloc(MAXPHYS);
 493                 if (hio->hio_ggio.gctl_data == NULL) {
 494                         primary_exitx(EX_TEMPFAIL,
 495                             "Unable to allocate %zu bytes of memory for gctl_data.",
 496                             MAXPHYS);
 497                 }
 498                 hio->hio_ggio.gctl_length = MAXPHYS;
 499                 hio->hio_ggio.gctl_error = 0;
 500                 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
 501                 hio_free_list_size++;
 502         }
 503 }
 504
 505 static bool
 506 init_resuid(struct hast_resource *res)
 507 {
 508
 509         mtx_lock(&metadata_lock);
 510         if (res->hr_resuid != 0) {
 511                 mtx_unlock(&metadata_lock);
 512                 return (false);
 513         } else {
 514                 /* Initialize unique resource identifier. */
 515                 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
 516                 mtx_unlock(&metadata_lock);
 517                 if (metadata_write(res) == -1)
 518                         exit(EX_NOINPUT);
 519                 return (true);
 520         }
 521 }
 522
 523 static void
 524 init_local(struct hast_resource *res)
 525 {
 526         unsigned char *buf;
 527         size_t mapsize;
 528
 529         if (metadata_read(res, true) == -1)
 530                 exit(EX_NOINPUT);
 531         mtx_init(&res->hr_amp_lock);
 532         if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
 533             res->hr_local_sectorsize, res->hr_keepdirty) == -1) {
 534                 primary_exit(EX_TEMPFAIL, "Unable to create activemap");
 535         }
 536         mtx_init(&range_lock);
 537         cv_init(&range_regular_cond);
 538         if (rangelock_init(&range_regular) == -1)
 539                 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
 540         cv_init(&range_sync_cond);
 541         if (rangelock_init(&range_sync) == -1)
 542                 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
 543         mapsize = activemap_ondisk_size(res->hr_amp);
 544         buf = calloc(1, mapsize);
 545         if (buf == NULL) {
 546                 primary_exitx(EX_TEMPFAIL,
 547                     "Unable to allocate buffer for activemap.");
 548         }
 549         if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
 550             (ssize_t)mapsize) {
 551                 primary_exit(EX_NOINPUT, "Unable to read activemap");
 552         }
 553         activemap_copyin(res->hr_amp, buf, mapsize);
 554         free(buf);
 555         if (res->hr_resuid != 0)
 556                 return;
 557         /*
 558          * We're using provider for the first time. Initialize local and remote
 559          * counters. We don't initialize resuid here, as we want to do it just
 560          * in time. The reason for this is that we want to inform secondary
 561          * that there were no writes yet, so there is no need to synchronize
 562          * anything.
 563          */
 564         res->hr_primary_localcnt = 0;
 565         res->hr_primary_remotecnt = 0;
 566         if (metadata_write(res) == -1)
 567                 exit(EX_NOINPUT);
 568 }
 569
 570 static int
 571 primary_connect(struct hast_resource *res, struct proto_conn **connp)
 572 {
 573         struct proto_conn *conn;
 574         int16_t val;
 575
 576         val = 1;
 577         if (proto_send(res->hr_conn, &val, sizeof(val)) == -1) {
 578                 primary_exit(EX_TEMPFAIL,
 579                     "Unable to send connection request to parent");
 580         }
 581         if (proto_recv(res->hr_conn, &val, sizeof(val)) == -1) {
 582                 primary_exit(EX_TEMPFAIL,
 583                     "Unable to receive reply to connection request from parent");
 584         }
 585         if (val != 0) {
 586                 errno = val;
 587                 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
 588                     res->hr_remoteaddr);
 589                 return (-1);
 590         }
 591         if (proto_connection_recv(res->hr_conn, true, &conn) == -1) {
 592                 primary_exit(EX_TEMPFAIL,
 593                     "Unable to receive connection from parent");
 594         }
 595         if (proto_connect_wait(conn, res->hr_timeout) == -1) {
 596                 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
 597                     res->hr_remoteaddr);
 598                 proto_close(conn);
 599                 return (-1);
 600         }
 601         /* Error in setting timeout is not critical, but why should it fail? */
 602         if (proto_timeout(conn, res->hr_timeout) == -1)
 603                 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
 604
 605         *connp = conn;
 606
 607         return (0);
 608 }
 609
 610 /*
 611  * Function instructs GEOM_GATE to handle reads directly from within the kernel.
 612  */
 613 static void
 614 enable_direct_reads(struct hast_resource *res)
 615 {
 616         struct g_gate_ctl_modify ggiomodify;
 617
 618         bzero(&ggiomodify, sizeof(ggiomodify));
 619         ggiomodify.gctl_version = G_GATE_VERSION;
 620         ggiomodify.gctl_unit = res->hr_ggateunit;
 621         ggiomodify.gctl_modify = GG_MODIFY_READPROV | GG_MODIFY_READOFFSET;
 622         strlcpy(ggiomodify.gctl_readprov, res->hr_localpath,
 623             sizeof(ggiomodify.gctl_readprov));
 624         ggiomodify.gctl_readoffset = res->hr_localoff;
 625         if (ioctl(res->hr_ggatefd, G_GATE_CMD_MODIFY, &ggiomodify) == 0)
 626                 pjdlog_debug(1, "Direct reads enabled.");
 627         else
 628                 pjdlog_errno(LOG_WARNING, "Failed to enable direct reads");
 629 }
 630
 631 static int
 632 init_remote(struct hast_resource *res, struct proto_conn **inp,
 633     struct proto_conn **outp)
 634 {
 635         struct proto_conn *in, *out;
 636         struct nv *nvout, *nvin;
 637         const unsigned char *token;
 638         unsigned char *map;
 639         const char *errmsg;
 640         int32_t extentsize;
 641         int64_t datasize;
 642         uint32_t mapsize;
 643         uint8_t version;
 644         size_t size;
 645         int error;
 646
 647         PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
 648         PJDLOG_ASSERT(real_remote(res));
 649
 650         in = out = NULL;
 651         errmsg = NULL;
 652
 653         if (primary_connect(res, &out) == -1)
 654                 return (ECONNREFUSED);
 655
 656         error = ECONNABORTED;
 657
 658         /*
 659          * First handshake step.
 660          * Setup outgoing connection with remote node.
 661          */
 662         nvout = nv_alloc();
 663         nv_add_string(nvout, res->hr_name, "resource");
 664         nv_add_uint8(nvout, HAST_PROTO_VERSION, "version");
 665         if (nv_error(nvout) != 0) {
 666                 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
 667                     "Unable to allocate header for connection with %s",
 668                     res->hr_remoteaddr);
 669                 nv_free(nvout);
 670                 goto close;
 671         }
 672         if (hast_proto_send(res, out, nvout, NULL, 0) == -1) {
 673                 pjdlog_errno(LOG_WARNING,
 674                     "Unable to send handshake header to %s",
 675                     res->hr_remoteaddr);
 676                 nv_free(nvout);
 677                 goto close;
 678         }
 679         nv_free(nvout);
 680         if (hast_proto_recv_hdr(out, &nvin) == -1) {
 681                 pjdlog_errno(LOG_WARNING,
 682                     "Unable to receive handshake header from %s",
 683                     res->hr_remoteaddr);
 684                 goto close;
 685         }
 686         errmsg = nv_get_string(nvin, "errmsg");
 687         if (errmsg != NULL) {
 688                 pjdlog_warning("%s", errmsg);
 689                 if (nv_exists(nvin, "wait"))
 690                         error = EBUSY;
 691                 nv_free(nvin);
 692                 goto close;
 693         }
 694         version = nv_get_uint8(nvin, "version");
 695         if (version == 0) {
 696                 /*
 697                  * If no version is sent, it means this is protocol version 1.
 698                  */
 699                 version = 1;
 700         }
 701         if (version > HAST_PROTO_VERSION) {
 702                 pjdlog_warning("Invalid version received (%hhu).", version);
 703                 nv_free(nvin);
 704                 goto close;
 705         }
 706         res->hr_version = version;
 707         pjdlog_debug(1, "Negotiated protocol version %d.", res->hr_version);
 708         token = nv_get_uint8_array(nvin, &size, "token");
 709         if (token == NULL) {
 710                 pjdlog_warning("Handshake header from %s has no 'token' field.",
 711                     res->hr_remoteaddr);
 712                 nv_free(nvin);
 713                 goto close;
 714         }
 715         if (size != sizeof(res->hr_token)) {
 716                 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
 717                     res->hr_remoteaddr, size, sizeof(res->hr_token));
 718                 nv_free(nvin);
 719                 goto close;
 720         }
 721         bcopy(token, res->hr_token, sizeof(res->hr_token));
 722         nv_free(nvin);
 723
 724         /*
 725          * Second handshake step.
 726          * Setup incoming connection with remote node.
 727          */
 728         if (primary_connect(res, &in) == -1)
 729                 goto close;
 730
 731         nvout = nv_alloc();
 732         nv_add_string(nvout, res->hr_name, "resource");
 733         nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
 734             "token");
 735         if (res->hr_resuid == 0) {
 736                 /*
 737                  * The resuid field was not yet initialized.
 738                  * Because we do synchronization inside init_resuid(), it is
 739                  * possible that someone already initialized it, the function
 740                  * will return false then, but if we successfully initialized
 741                  * it, we will get true. True means that there were no writes
 742                  * to this resource yet and we want to inform secondary that
 743                  * synchronization is not needed by sending "virgin" argument.
 744                  */
 745                 if (init_resuid(res))
 746                         nv_add_int8(nvout, 1, "virgin");
 747         }
 748         nv_add_uint64(nvout, res->hr_resuid, "resuid");
 749         nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
 750         nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
 751         if (nv_error(nvout) != 0) {
 752                 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
 753                     "Unable to allocate header for connection with %s",
 754                     res->hr_remoteaddr);
 755                 nv_free(nvout);
 756                 goto close;
 757         }
 758         if (hast_proto_send(res, in, nvout, NULL, 0) == -1) {
 759                 pjdlog_errno(LOG_WARNING,
 760                     "Unable to send handshake header to %s",
 761                     res->hr_remoteaddr);
 762                 nv_free(nvout);
 763                 goto close;
 764         }
 765         nv_free(nvout);
 766         if (hast_proto_recv_hdr(out, &nvin) == -1) {
 767                 pjdlog_errno(LOG_WARNING,
 768                     "Unable to receive handshake header from %s",
 769                     res->hr_remoteaddr);
 770                 goto close;
 771         }
 772         errmsg = nv_get_string(nvin, "errmsg");
 773         if (errmsg != NULL) {
 774                 pjdlog_warning("%s", errmsg);
 775                 nv_free(nvin);
 776                 goto close;
 777         }
 778         datasize = nv_get_int64(nvin, "datasize");
 779         if (datasize != res->hr_datasize) {
 780                 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
 781                     (intmax_t)res->hr_datasize, (intmax_t)datasize);
 782                 nv_free(nvin);
 783                 goto close;
 784         }
 785         extentsize = nv_get_int32(nvin, "extentsize");
 786         if (extentsize != res->hr_extentsize) {
 787                 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
 788                     (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
 789                 nv_free(nvin);
 790                 goto close;
 791         }
 792         res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
 793         res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
 794         res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
 795         if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY)
 796                 enable_direct_reads(res);
 797         if (nv_exists(nvin, "virgin")) {
 798                 /*
 799                  * Secondary was reinitialized, bump localcnt if it is 0 as
 800                  * only we have the data.
 801                  */
 802                 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY);
 803                 PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
 804
 805                 if (res->hr_primary_localcnt == 0) {
 806                         PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0);
 807
 808                         mtx_lock(&metadata_lock);
 809                         res->hr_primary_localcnt++;
 810                         pjdlog_debug(1, "Increasing localcnt to %ju.",
 811                             (uintmax_t)res->hr_primary_localcnt);
 812                         (void)metadata_write(res);
 813                         mtx_unlock(&metadata_lock);
 814                 }
 815         }
 816         map = NULL;
 817         mapsize = nv_get_uint32(nvin, "mapsize");
 818         if (mapsize > 0) {
 819                 map = malloc(mapsize);
 820                 if (map == NULL) {
 821                         pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
 822                             (uintmax_t)mapsize);
 823                         nv_free(nvin);
 824                         goto close;
 825                 }
 826                 /*
 827                  * Remote node have some dirty extents on its own, lets
 828                  * download its activemap.
 829                  */
 830                 if (hast_proto_recv_data(res, out, nvin, map,
 831                     mapsize) == -1) {
 832                         pjdlog_errno(LOG_ERR,
 833                             "Unable to receive remote activemap");
 834                         nv_free(nvin);
 835                         free(map);
 836                         goto close;
 837                 }
 838                 mtx_lock(&res->hr_amp_lock);
 839                 /*
 840                  * Merge local and remote bitmaps.
 841                  */
 842                 activemap_merge(res->hr_amp, map, mapsize);
 843                 free(map);
 844                 /*
 845                  * Now that we merged bitmaps from both nodes, flush it to the
 846                  * disk before we start to synchronize.
 847                  */
 848                 (void)hast_activemap_flush(res);
 849         }
 850         nv_free(nvin);
 851 #ifdef notyet
 852         /* Setup directions. */
 853         if (proto_send(out, NULL, 0) == -1)
 854                 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
 855         if (proto_recv(in, NULL, 0) == -1)
 856                 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
 857 #endif
 858         pjdlog_info("Connected to %s.", res->hr_remoteaddr);
 859         if (res->hr_original_replication == HAST_REPLICATION_MEMSYNC &&
 860             res->hr_version < 2) {
 861                 pjdlog_warning("The 'memsync' replication mode is not supported by the remote node, falling back to 'fullsync' mode.");
 862                 res->hr_replication = HAST_REPLICATION_FULLSYNC;
 863         } else if (res->hr_replication != res->hr_original_replication) {
 864                 /*
 865                  * This is in case hastd disconnected and was upgraded.
 866                  */
 867                 res->hr_replication = res->hr_original_replication;
 868         }
 869         if (inp != NULL && outp != NULL) {
 870                 *inp = in;
 871                 *outp = out;
 872         } else {
 873                 res->hr_remotein = in;
 874                 res->hr_remoteout = out;
 875         }
 876         event_send(res, EVENT_CONNECT);
 877         return (0);
 878 close:
 879         if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
 880                 event_send(res, EVENT_SPLITBRAIN);
 881         proto_close(out);
 882         if (in != NULL)
 883                 proto_close(in);
 884         return (error);
 885 }
 886
 887 static void
 888 sync_start(void)
 889 {
 890
 891         mtx_lock(&sync_lock);
 892         sync_inprogress = true;
 893         mtx_unlock(&sync_lock);
 894         cv_signal(&sync_cond);
 895 }
 896
 897 static void
 898 sync_stop(void)
 899 {
 900
 901         mtx_lock(&sync_lock);
 902         if (sync_inprogress)
 903                 sync_inprogress = false;
 904         mtx_unlock(&sync_lock);
 905 }
 906
 907 static void
 908 init_ggate(struct hast_resource *res)
 909 {
 910         struct g_gate_ctl_create ggiocreate;
 911         struct g_gate_ctl_cancel ggiocancel;
 912
 913         /*
 914          * We communicate with ggate via /dev/ggctl. Open it.
 915          */
 916         res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
 917         if (res->hr_ggatefd == -1)
 918                 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
 919         /*
 920          * Create provider before trying to connect, as connection failure
 921          * is not critical, but may take some time.
 922          */
 923         bzero(&ggiocreate, sizeof(ggiocreate));
 924         ggiocreate.gctl_version = G_GATE_VERSION;
 925         ggiocreate.gctl_mediasize = res->hr_datasize;
 926         ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
 927         ggiocreate.gctl_flags = 0;
 928         ggiocreate.gctl_maxcount = 0;
 929         ggiocreate.gctl_timeout = 0;
 930         ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
 931         snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
 932             res->hr_provname);
 933         if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
 934                 pjdlog_info("Device hast/%s created.", res->hr_provname);
 935                 res->hr_ggateunit = ggiocreate.gctl_unit;
 936                 return;
 937         }
 938         if (errno != EEXIST) {
 939                 primary_exit(EX_OSERR, "Unable to create hast/%s device",
 940                     res->hr_provname);
 941         }
 942         pjdlog_debug(1,
 943             "Device hast/%s already exists, we will try to take it over.",
 944             res->hr_provname);
 945         /*
 946          * If we received EEXIST, we assume that the process who created the
 947          * provider died and didn't clean up. In that case we will start from
 948          * where he left of.
 949          */
 950         bzero(&ggiocancel, sizeof(ggiocancel));
 951         ggiocancel.gctl_version = G_GATE_VERSION;
 952         ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
 953         snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
 954             res->hr_provname);
 955         if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
 956                 pjdlog_info("Device hast/%s recovered.", res->hr_provname);
 957                 res->hr_ggateunit = ggiocancel.gctl_unit;
 958                 return;
 959         }
 960         primary_exit(EX_OSERR, "Unable to take over hast/%s device",
 961             res->hr_provname);
 962 }
 963
 964 void
 965 hastd_primary(struct hast_resource *res)
 966 {
 967         pthread_t td;
 968         pid_t pid;
 969         int error, mode, debuglevel;
 970
 971         /*
 972          * Create communication channel for sending control commands from
 973          * parent to child.
 974          */
 975         if (proto_client(NULL, "socketpair://", &res->hr_ctrl) == -1) {
 976                 /* TODO: There's no need for this to be fatal error. */
 977                 KEEP_ERRNO((void)pidfile_remove(pfh));
 978                 pjdlog_exit(EX_OSERR,
 979                     "Unable to create control sockets between parent and child");
 980         }
 981         /*
 982          * Create communication channel for sending events from child to parent.
 983          */
 984         if (proto_client(NULL, "socketpair://", &res->hr_event) == -1) {
 985                 /* TODO: There's no need for this to be fatal error. */
 986                 KEEP_ERRNO((void)pidfile_remove(pfh));
 987                 pjdlog_exit(EX_OSERR,
 988                     "Unable to create event sockets between child and parent");
 989         }
 990         /*
 991          * Create communication channel for sending connection requests from
 992          * child to parent.
 993          */
 994         if (proto_client(NULL, "socketpair://", &res->hr_conn) == -1) {
 995                 /* TODO: There's no need for this to be fatal error. */
 996                 KEEP_ERRNO((void)pidfile_remove(pfh));
 997                 pjdlog_exit(EX_OSERR,
 998                     "Unable to create connection sockets between child and parent");
 999         }
1000
1001         pid = fork();
1002         if (pid == -1) {
1003                 /* TODO: There's no need for this to be fatal error. */
1004                 KEEP_ERRNO((void)pidfile_remove(pfh));
1005                 pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
1006         }
1007
1008         if (pid > 0) {
1009                 /* This is parent. */
1010                 /* Declare that we are receiver. */
1011                 proto_recv(res->hr_event, NULL, 0);
1012                 proto_recv(res->hr_conn, NULL, 0);
1013                 /* Declare that we are sender. */
1014                 proto_send(res->hr_ctrl, NULL, 0);
1015                 res->hr_workerpid = pid;
1016                 return;
1017         }
1018
1019         gres = res;
1020         res->output_status_aux = output_status_aux;
1021         mode = pjdlog_mode_get();
1022         debuglevel = pjdlog_debug_get();
1023
1024         /* Declare that we are sender. */
1025         proto_send(res->hr_event, NULL, 0);
1026         proto_send(res->hr_conn, NULL, 0);
1027         /* Declare that we are receiver. */
1028         proto_recv(res->hr_ctrl, NULL, 0);
1029         descriptors_cleanup(res);
1030
1031         descriptors_assert(res, mode);
1032
1033         pjdlog_init(mode);
1034         pjdlog_debug_set(debuglevel);
1035         pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
1036         setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
1037
1038         init_local(res);
1039         init_ggate(res);
1040         init_environment(res);
1041
1042         if (drop_privs(res) != 0) {
1043                 cleanup(res);
1044                 exit(EX_CONFIG);
1045         }
1046         pjdlog_info("Privileges successfully dropped.");
1047
1048         /*
1049          * Create the guard thread first, so we can handle signals from the
1050          * very beginning.
1051          */
1052         error = pthread_create(&td, NULL, guard_thread, res);
1053         PJDLOG_ASSERT(error == 0);
1054         /*
1055          * Create the control thread before sending any event to the parent,
1056          * as we can deadlock when parent sends control request to worker,
1057          * but worker has no control thread started yet, so parent waits.
1058          * In the meantime worker sends an event to the parent, but parent
1059          * is unable to handle the event, because it waits for control
1060          * request response.
1061          */
1062         error = pthread_create(&td, NULL, ctrl_thread, res);
1063         PJDLOG_ASSERT(error == 0);
1064         if (real_remote(res)) {
1065                 error = init_remote(res, NULL, NULL);
1066                 if (error == 0) {
1067                         sync_start();
1068                 } else if (error == EBUSY) {
1069                         time_t start = time(NULL);
1070
1071                         pjdlog_warning("Waiting for remote node to become %s for %ds.",
1072                             role2str(HAST_ROLE_SECONDARY),
1073                             res->hr_timeout);
1074                         for (;;) {
1075                                 sleep(1);
1076                                 error = init_remote(res, NULL, NULL);
1077                                 if (error != EBUSY)
1078                                         break;
1079                                 if (time(NULL) > start + res->hr_timeout)
1080                                         break;
1081                         }
1082                         if (error == EBUSY) {
1083                                 pjdlog_warning("Remote node is still %s, starting anyway.",
1084                                     role2str(HAST_ROLE_PRIMARY));
1085                         }
1086                 }
1087         }
1088         error = pthread_create(&td, NULL, ggate_recv_thread, res);
1089         PJDLOG_ASSERT(error == 0);
1090         error = pthread_create(&td, NULL, local_send_thread, res);
1091         PJDLOG_ASSERT(error == 0);
1092         error = pthread_create(&td, NULL, remote_send_thread, res);
1093         PJDLOG_ASSERT(error == 0);
1094         error = pthread_create(&td, NULL, remote_recv_thread, res);
1095         PJDLOG_ASSERT(error == 0);
1096         error = pthread_create(&td, NULL, ggate_send_thread, res);
1097         PJDLOG_ASSERT(error == 0);
1098         fullystarted = true;
1099         (void)sync_thread(res);
1100 }
1101
1102 static void
1103 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio,
1104     const char *fmt, ...)
1105 {
1106         char msg[1024];
1107         va_list ap;
1108
1109         va_start(ap, fmt);
1110         (void)vsnprintf(msg, sizeof(msg), fmt, ap);
1111         va_end(ap);
1112         switch (ggio->gctl_cmd) {
1113         case BIO_READ:
1114                 (void)snprlcat(msg, sizeof(msg), "READ(%ju, %ju).",
1115                     (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1116                 break;
1117         case BIO_DELETE:
1118                 (void)snprlcat(msg, sizeof(msg), "DELETE(%ju, %ju).",
1119                     (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1120                 break;
1121         case BIO_FLUSH:
1122                 (void)snprlcat(msg, sizeof(msg), "FLUSH.");
1123                 break;
1124         case BIO_WRITE:
1125                 (void)snprlcat(msg, sizeof(msg), "WRITE(%ju, %ju).",
1126                     (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1127                 break;
1128         default:
1129                 (void)snprlcat(msg, sizeof(msg), "UNKNOWN(%u).",
1130                     (unsigned int)ggio->gctl_cmd);
1131                 break;
1132         }
1133         pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
1134 }
1135
1136 static void
1137 remote_close(struct hast_resource *res, int ncomp)
1138 {
1139
1140         rw_wlock(&hio_remote_lock[ncomp]);
1141         /*
1142          * Check for a race between dropping rlock and acquiring wlock -
1143          * another thread can close connection in-between.
1144          */
1145         if (!ISCONNECTED(res, ncomp)) {
1146                 PJDLOG_ASSERT(res->hr_remotein == NULL);
1147                 PJDLOG_ASSERT(res->hr_remoteout == NULL);
1148                 rw_unlock(&hio_remote_lock[ncomp]);
1149                 return;
1150         }
1151
1152         PJDLOG_ASSERT(res->hr_remotein != NULL);
1153         PJDLOG_ASSERT(res->hr_remoteout != NULL);
1154
1155         pjdlog_debug(2, "Closing incoming connection to %s.",
1156             res->hr_remoteaddr);
1157         proto_close(res->hr_remotein);
1158         res->hr_remotein = NULL;
1159         pjdlog_debug(2, "Closing outgoing connection to %s.",
1160             res->hr_remoteaddr);
1161         proto_close(res->hr_remoteout);
1162         res->hr_remoteout = NULL;
1163
1164         rw_unlock(&hio_remote_lock[ncomp]);
1165
1166         pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
1167
1168         /*
1169          * Stop synchronization if in-progress.
1170          */
1171         sync_stop();
1172
1173         event_send(res, EVENT_DISCONNECT);
1174 }
1175
1176 /*
1177  * Acknowledge write completion to the kernel, but don't update activemap yet.
1178  */
1179 static void
1180 write_complete(struct hast_resource *res, struct hio *hio)
1181 {
1182         struct g_gate_ctl_io *ggio;
1183         unsigned int ncomp;
1184
1185         PJDLOG_ASSERT(!hio->hio_done);
1186
1187         ggio = &hio->hio_ggio;
1188         PJDLOG_ASSERT(ggio->gctl_cmd == BIO_WRITE);
1189
1190         /*
1191          * Bump local count if this is first write after
1192          * connection failure with remote node.
1193          */
1194         ncomp = 1;
1195         rw_rlock(&hio_remote_lock[ncomp]);
1196         if (!ISCONNECTED(res, ncomp)) {
1197                 mtx_lock(&metadata_lock);
1198                 if (res->hr_primary_localcnt == res->hr_secondary_remotecnt) {
1199                         res->hr_primary_localcnt++;
1200                         pjdlog_debug(1, "Increasing localcnt to %ju.",
1201                             (uintmax_t)res->hr_primary_localcnt);
1202                         (void)metadata_write(res);
1203                 }
1204                 mtx_unlock(&metadata_lock);
1205         }
1206         rw_unlock(&hio_remote_lock[ncomp]);
1207         if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1)
1208                 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1209         hio->hio_done = true;
1210 }
1211
1212 /*
1213  * Thread receives ggate I/O requests from the kernel and passes them to
1214  * appropriate threads:
1215  * WRITE - always goes to both local_send and remote_send threads
1216  * READ (when the block is up-to-date on local component) -
1217  *      only local_send thread
1218  * READ (when the block isn't up-to-date on local component) -
1219  *      only remote_send thread
1220  * DELETE - always goes to both local_send and remote_send threads
1221  * FLUSH - always goes to both local_send and remote_send threads
1222  */
1223 static void *
1224 ggate_recv_thread(void *arg)
1225 {
1226         struct hast_resource *res = arg;
1227         struct g_gate_ctl_io *ggio;
1228         struct hio *hio;
1229         unsigned int ii, ncomp, ncomps;
1230         int error;
1231
1232         for (;;) {
1233                 pjdlog_debug(2, "ggate_recv: Taking free request.");
1234                 QUEUE_TAKE2(hio, free);
1235                 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1236                 ggio = &hio->hio_ggio;
1237                 ggio->gctl_unit = res->hr_ggateunit;
1238                 ggio->gctl_length = MAXPHYS;
1239                 ggio->gctl_error = 0;
1240                 hio->hio_done = false;
1241                 hio->hio_replication = res->hr_replication;
1242                 pjdlog_debug(2,
1243                     "ggate_recv: (%p) Waiting for request from the kernel.",
1244                     hio);
1245                 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) == -1) {
1246                         if (sigexit_received)
1247                                 pthread_exit(NULL);
1248                         primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1249                 }
1250                 error = ggio->gctl_error;
1251                 switch (error) {
1252                 case 0:
1253                         break;
1254                 case ECANCELED:
1255                         /* Exit gracefully. */
1256                         if (!sigexit_received) {
1257                                 pjdlog_debug(2,
1258                                     "ggate_recv: (%p) Received cancel from the kernel.",
1259                                     hio);
1260                                 pjdlog_info("Received cancel from the kernel, exiting.");
1261                         }
1262                         pthread_exit(NULL);
1263                 case ENOMEM:
1264                         /*
1265                          * Buffer too small? Impossible, we allocate MAXPHYS
1266                          * bytes - request can't be bigger than that.
1267                          */
1268                         /* FALLTHROUGH */
1269                 case ENXIO:
1270                 default:
1271                         primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1272                             strerror(error));
1273                 }
1274
1275                 ncomp = 0;
1276                 ncomps = HAST_NCOMPONENTS;
1277
1278                 for (ii = 0; ii < ncomps; ii++)
1279                         hio->hio_errors[ii] = EINVAL;
1280                 reqlog(LOG_DEBUG, 2, ggio,
1281                     "ggate_recv: (%p) Request received from the kernel: ",
1282                     hio);
1283
1284                 /*
1285                  * Inform all components about new write request.
1286                  * For read request prefer local component unless the given
1287                  * range is out-of-date, then use remote component.
1288                  */
1289                 switch (ggio->gctl_cmd) {
1290                 case BIO_READ:
1291                         res->hr_stat_read++;
1292                         ncomps = 1;
1293                         mtx_lock(&metadata_lock);
1294                         if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1295                             res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1296                                 /*
1297                                  * This range is up-to-date on local component,
1298                                  * so handle request locally.
1299                                  */
1300                                  /* Local component is 0 for now. */
1301                                 ncomp = 0;
1302                         } else /* if (res->hr_syncsrc ==
1303                             HAST_SYNCSRC_SECONDARY) */ {
1304                                 PJDLOG_ASSERT(res->hr_syncsrc ==
1305                                     HAST_SYNCSRC_SECONDARY);
1306                                 /*
1307                                  * This range is out-of-date on local component,
1308                                  * so send request to the remote node.
1309                                  */
1310                                  /* Remote component is 1 for now. */
1311                                 ncomp = 1;
1312                         }
1313                         mtx_unlock(&metadata_lock);
1314                         break;
1315                 case BIO_WRITE:
1316                         res->hr_stat_write++;
1317                         if (res->hr_resuid == 0 &&
1318                             res->hr_primary_localcnt == 0) {
1319                                 /* This is first write. */
1320                                 res->hr_primary_localcnt = 1;
1321                         }
1322                         for (;;) {
1323                                 mtx_lock(&range_lock);
1324                                 if (rangelock_islocked(range_sync,
1325                                     ggio->gctl_offset, ggio->gctl_length)) {
1326                                         pjdlog_debug(2,
1327                                             "regular: Range offset=%jd length=%zu locked.",
1328                                             (intmax_t)ggio->gctl_offset,
1329                                             (size_t)ggio->gctl_length);
1330                                         range_regular_wait = true;
1331                                         cv_wait(&range_regular_cond, &range_lock);
1332                                         range_regular_wait = false;
1333                                         mtx_unlock(&range_lock);
1334                                         continue;
1335                                 }
1336                                 if (rangelock_add(range_regular,
1337                                     ggio->gctl_offset, ggio->gctl_length) == -1) {
1338                                         mtx_unlock(&range_lock);
1339                                         pjdlog_debug(2,
1340                                             "regular: Range offset=%jd length=%zu is already locked, waiting.",
1341                                             (intmax_t)ggio->gctl_offset,
1342                                             (size_t)ggio->gctl_length);
1343                                         sleep(1);
1344                                         continue;
1345                                 }
1346                                 mtx_unlock(&range_lock);
1347                                 break;
1348                         }
1349                         mtx_lock(&res->hr_amp_lock);
1350                         if (activemap_write_start(res->hr_amp,
1351                             ggio->gctl_offset, ggio->gctl_length)) {
1352                                 res->hr_stat_activemap_update++;
1353                                 (void)hast_activemap_flush(res);
1354                         } else {
1355                                 mtx_unlock(&res->hr_amp_lock);
1356                         }
1357                         if (ISMEMSYNC(hio)) {
1358                                 hio->hio_memsyncacked = false;
1359                                 refcnt_init(&hio->hio_writecount, ncomps);
1360                         }
1361                         break;
1362                 case BIO_DELETE:
1363                         res->hr_stat_delete++;
1364                         break;
1365                 case BIO_FLUSH:
1366                         res->hr_stat_flush++;
1367                         break;
1368                 }
1369                 pjdlog_debug(2,
1370                     "ggate_recv: (%p) Moving request to the send queues.", hio);
1371                 refcnt_init(&hio->hio_countdown, ncomps);
1372                 for (ii = ncomp; ii < ncomps; ii++)
1373                         QUEUE_INSERT1(hio, send, ii);
1374         }
1375         /* NOTREACHED */
1376         return (NULL);
1377 }
1378
1379 /*
1380  * Thread reads from or writes to local component.
1381  * If local read fails, it redirects it to remote_send thread.
1382  */
1383 static void *
1384 local_send_thread(void *arg)
1385 {
1386         struct hast_resource *res = arg;
1387         struct g_gate_ctl_io *ggio;
1388         struct hio *hio;
1389         unsigned int ncomp, rncomp;
1390         ssize_t ret;
1391
1392         /* Local component is 0 for now. */
1393         ncomp = 0;
1394         /* Remote component is 1 for now. */
1395         rncomp = 1;
1396
1397         for (;;) {
1398                 pjdlog_debug(2, "local_send: Taking request.");
1399                 QUEUE_TAKE1(hio, send, ncomp, 0);
1400                 pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1401                 ggio = &hio->hio_ggio;
1402                 switch (ggio->gctl_cmd) {
1403                 case BIO_READ:
1404                         ret = pread(res->hr_localfd, ggio->gctl_data,
1405                             ggio->gctl_length,
1406                             ggio->gctl_offset + res->hr_localoff);
1407                         if (ret == ggio->gctl_length)
1408                                 hio->hio_errors[ncomp] = 0;
1409                         else if (!ISSYNCREQ(hio)) {
1410                                 /*
1411                                  * If READ failed, try to read from remote node.
1412                                  */
1413                                 if (ret == -1) {
1414                                         reqlog(LOG_WARNING, 0, ggio,
1415                                             "Local request failed (%s), trying remote node. ",
1416                                             strerror(errno));
1417                                 } else if (ret != ggio->gctl_length) {
1418                                         reqlog(LOG_WARNING, 0, ggio,
1419                                             "Local request failed (%zd != %jd), trying remote node. ",
1420                                             ret, (intmax_t)ggio->gctl_length);
1421                                 }
1422                                 QUEUE_INSERT1(hio, send, rncomp);
1423                                 continue;
1424                         }
1425                         break;
1426                 case BIO_WRITE:
1427                         ret = pwrite(res->hr_localfd, ggio->gctl_data,
1428                             ggio->gctl_length,
1429                             ggio->gctl_offset + res->hr_localoff);
1430                         if (ret == -1) {
1431                                 hio->hio_errors[ncomp] = errno;
1432                                 reqlog(LOG_WARNING, 0, ggio,
1433                                     "Local request failed (%s): ",
1434                                     strerror(errno));
1435                         } else if (ret != ggio->gctl_length) {
1436                                 hio->hio_errors[ncomp] = EIO;
1437                                 reqlog(LOG_WARNING, 0, ggio,
1438                                     "Local request failed (%zd != %jd): ",
1439                                     ret, (intmax_t)ggio->gctl_length);
1440                         } else {
1441                                 hio->hio_errors[ncomp] = 0;
1442                                 if (ISASYNC(hio)) {
1443                                         ggio->gctl_error = 0;
1444                                         write_complete(res, hio);
1445                                 }
1446                         }
1447                         break;
1448                 case BIO_DELETE:
1449                         ret = g_delete(res->hr_localfd,
1450                             ggio->gctl_offset + res->hr_localoff,
1451                             ggio->gctl_length);
1452                         if (ret == -1) {
1453                                 hio->hio_errors[ncomp] = errno;
1454                                 reqlog(LOG_WARNING, 0, ggio,
1455                                     "Local request failed (%s): ",
1456                                     strerror(errno));
1457                         } else {
1458                                 hio->hio_errors[ncomp] = 0;
1459                         }
1460                         break;
1461                 case BIO_FLUSH:
1462                         if (!res->hr_localflush) {
1463                                 ret = -1;
1464                                 errno = EOPNOTSUPP;
1465                                 break;
1466                         }
1467                         ret = g_flush(res->hr_localfd);
1468                         if (ret == -1) {
1469                                 if (errno == EOPNOTSUPP)
1470                                         res->hr_localflush = false;
1471                                 hio->hio_errors[ncomp] = errno;
1472                                 reqlog(LOG_WARNING, 0, ggio,
1473                                     "Local request failed (%s): ",
1474                                     strerror(errno));
1475                         } else {
1476                                 hio->hio_errors[ncomp] = 0;
1477                         }
1478                         break;
1479                 }
1480                 if (ISMEMSYNCWRITE(hio)) {
1481                         if (refcnt_release(&hio->hio_writecount) == 0) {
1482                                 write_complete(res, hio);
1483                         }
1484                 }
1485                 if (refcnt_release(&hio->hio_countdown) > 0)
1486                         continue;
1487                 if (ISSYNCREQ(hio)) {
1488                         mtx_lock(&sync_lock);
1489                         SYNCREQDONE(hio);
1490                         mtx_unlock(&sync_lock);
1491                         cv_signal(&sync_cond);
1492                 } else {
1493                         pjdlog_debug(2,
1494                             "local_send: (%p) Moving request to the done queue.",
1495                             hio);
1496                         QUEUE_INSERT2(hio, done);
1497                 }
1498         }
1499         /* NOTREACHED */
1500         return (NULL);
1501 }
1502
1503 static void
1504 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1505 {
1506         struct nv *nv;
1507
1508         rw_rlock(&hio_remote_lock[ncomp]);
1509
1510         if (!ISCONNECTED(res, ncomp)) {
1511                 rw_unlock(&hio_remote_lock[ncomp]);
1512                 return;
1513         }
1514
1515         PJDLOG_ASSERT(res->hr_remotein != NULL);
1516         PJDLOG_ASSERT(res->hr_remoteout != NULL);
1517
1518         nv = nv_alloc();
1519         nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1520         if (nv_error(nv) != 0) {
1521                 rw_unlock(&hio_remote_lock[ncomp]);
1522                 nv_free(nv);
1523                 pjdlog_debug(1,
1524                     "keepalive_send: Unable to prepare header to send.");
1525                 return;
1526         }
1527         if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) == -1) {
1528                 rw_unlock(&hio_remote_lock[ncomp]);
1529                 pjdlog_common(LOG_DEBUG, 1, errno,
1530                     "keepalive_send: Unable to send request");
1531                 nv_free(nv);
1532                 remote_close(res, ncomp);
1533                 return;
1534         }
1535
1536         rw_unlock(&hio_remote_lock[ncomp]);
1537         nv_free(nv);
1538         pjdlog_debug(2, "keepalive_send: Request sent.");
1539 }
1540
1541 /*
1542  * Thread sends request to secondary node.
1543  */
1544 static void *
1545 remote_send_thread(void *arg)
1546 {
1547         struct hast_resource *res = arg;
1548         struct g_gate_ctl_io *ggio;
1549         time_t lastcheck, now;
1550         struct hio *hio;
1551         struct nv *nv;
1552         unsigned int ncomp;
1553         bool wakeup;
1554         uint64_t offset, length;
1555         uint8_t cmd;
1556         void *data;
1557
1558         /* Remote component is 1 for now. */
1559         ncomp = 1;
1560         lastcheck = time(NULL);
1561
1562         for (;;) {
1563                 pjdlog_debug(2, "remote_send: Taking request.");
1564                 QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE);
1565                 if (hio == NULL) {
1566                         now = time(NULL);
1567                         if (lastcheck + HAST_KEEPALIVE <= now) {
1568                                 keepalive_send(res, ncomp);
1569                                 lastcheck = now;
1570                         }
1571                         continue;
1572                 }
1573                 pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1574                 ggio = &hio->hio_ggio;
1575                 switch (ggio->gctl_cmd) {
1576                 case BIO_READ:
1577                         cmd = HIO_READ;
1578                         data = NULL;
1579                         offset = ggio->gctl_offset;
1580                         length = ggio->gctl_length;
1581                         break;
1582                 case BIO_WRITE:
1583                         cmd = HIO_WRITE;
1584                         data = ggio->gctl_data;
1585                         offset = ggio->gctl_offset;
1586                         length = ggio->gctl_length;
1587                         break;
1588                 case BIO_DELETE:
1589                         cmd = HIO_DELETE;
1590                         data = NULL;
1591                         offset = ggio->gctl_offset;
1592                         length = ggio->gctl_length;
1593                         break;
1594                 case BIO_FLUSH:
1595                         cmd = HIO_FLUSH;
1596                         data = NULL;
1597                         offset = 0;
1598                         length = 0;
1599                         break;
1600                 default:
1601                         PJDLOG_ABORT("invalid condition");
1602                 }
1603                 nv = nv_alloc();
1604                 nv_add_uint8(nv, cmd, "cmd");
1605                 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1606                 nv_add_uint64(nv, offset, "offset");
1607                 nv_add_uint64(nv, length, "length");
1608                 if (ISMEMSYNCWRITE(hio))
1609                         nv_add_uint8(nv, 1, "memsync");
1610                 if (nv_error(nv) != 0) {
1611                         hio->hio_errors[ncomp] = nv_error(nv);
1612                         pjdlog_debug(2,
1613                             "remote_send: (%p) Unable to prepare header to send.",
1614                             hio);
1615                         reqlog(LOG_ERR, 0, ggio,
1616                             "Unable to prepare header to send (%s): ",
1617                             strerror(nv_error(nv)));
1618                         /* Move failed request immediately to the done queue. */
1619                         goto done_queue;
1620                 }
1621                 /*
1622                  * Protect connection from disappearing.
1623                  */
1624                 rw_rlock(&hio_remote_lock[ncomp]);
1625                 if (!ISCONNECTED(res, ncomp)) {
1626                         rw_unlock(&hio_remote_lock[ncomp]);
1627                         hio->hio_errors[ncomp] = ENOTCONN;
1628                         goto done_queue;
1629                 }
1630                 /*
1631                  * Move the request to recv queue before sending it, because
1632                  * in different order we can get reply before we move request
1633                  * to recv queue.
1634                  */
1635                 pjdlog_debug(2,
1636                     "remote_send: (%p) Moving request to the recv queue.",
1637                     hio);
1638                 mtx_lock(&hio_recv_list_lock[ncomp]);
1639                 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1640                 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1641                 hio_recv_list_size[ncomp]++;
1642                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1643                 if (hast_proto_send(res, res->hr_remoteout, nv, data,
1644                     data != NULL ? length : 0) == -1) {
1645                         hio->hio_errors[ncomp] = errno;
1646                         rw_unlock(&hio_remote_lock[ncomp]);
1647                         pjdlog_debug(2,
1648                             "remote_send: (%p) Unable to send request.", hio);
1649                         reqlog(LOG_ERR, 0, ggio,
1650                             "Unable to send request (%s): ",
1651                             strerror(hio->hio_errors[ncomp]));
1652                         remote_close(res, ncomp);
1653                 } else {
1654                         rw_unlock(&hio_remote_lock[ncomp]);
1655                 }
1656                 nv_free(nv);
1657                 if (wakeup)
1658                         cv_signal(&hio_recv_list_cond[ncomp]);
1659                 continue;
1660 done_queue:
1661                 nv_free(nv);
1662                 if (ISSYNCREQ(hio)) {
1663                         if (refcnt_release(&hio->hio_countdown) > 0)
1664                                 continue;
1665                         mtx_lock(&sync_lock);
1666                         SYNCREQDONE(hio);
1667                         mtx_unlock(&sync_lock);
1668                         cv_signal(&sync_cond);
1669                         continue;
1670                 }
1671                 if (ggio->gctl_cmd == BIO_WRITE) {
1672                         mtx_lock(&res->hr_amp_lock);
1673                         if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1674                             ggio->gctl_length)) {
1675                                 (void)hast_activemap_flush(res);
1676                         } else {
1677                                 mtx_unlock(&res->hr_amp_lock);
1678                         }
1679                         if (ISMEMSYNCWRITE(hio)) {
1680                                 if (refcnt_release(&hio->hio_writecount) == 0) {
1681                                         if (hio->hio_errors[0] == 0)
1682                                                 write_complete(res, hio);
1683                                 }
1684                         }
1685                 }
1686                 if (refcnt_release(&hio->hio_countdown) > 0)
1687                         continue;
1688                 pjdlog_debug(2,
1689                     "remote_send: (%p) Moving request to the done queue.",
1690                     hio);
1691                 QUEUE_INSERT2(hio, done);
1692         }
1693         /* NOTREACHED */
1694         return (NULL);
1695 }
1696
1697 /*
1698  * Thread receives answer from secondary node and passes it to ggate_send
1699  * thread.
1700  */
1701 static void *
1702 remote_recv_thread(void *arg)
1703 {
1704         struct hast_resource *res = arg;
1705         struct g_gate_ctl_io *ggio;
1706         struct hio *hio;
1707         struct nv *nv;
1708         unsigned int ncomp;
1709         uint64_t seq;
1710         bool memsyncack;
1711         int error;
1712
1713         /* Remote component is 1 for now. */
1714         ncomp = 1;
1715
1716         for (;;) {
1717                 /* Wait until there is anything to receive. */
1718                 mtx_lock(&hio_recv_list_lock[ncomp]);
1719                 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1720                         pjdlog_debug(2, "remote_recv: No requests, waiting.");
1721                         cv_wait(&hio_recv_list_cond[ncomp],
1722                             &hio_recv_list_lock[ncomp]);
1723                 }
1724                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1725
1726                 memsyncack = false;
1727
1728                 rw_rlock(&hio_remote_lock[ncomp]);
1729                 if (!ISCONNECTED(res, ncomp)) {
1730                         rw_unlock(&hio_remote_lock[ncomp]);
1731                         /*
1732                          * Connection is dead, so move all pending requests to
1733                          * the done queue (one-by-one).
1734                          */
1735                         mtx_lock(&hio_recv_list_lock[ncomp]);
1736                         hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1737                         PJDLOG_ASSERT(hio != NULL);
1738                         TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1739                             hio_next[ncomp]);
1740                         hio_recv_list_size[ncomp]--;
1741                         mtx_unlock(&hio_recv_list_lock[ncomp]);
1742                         hio->hio_errors[ncomp] = ENOTCONN;
1743                         goto done_queue;
1744                 }
1745                 if (hast_proto_recv_hdr(res->hr_remotein, &nv) == -1) {
1746                         pjdlog_errno(LOG_ERR,
1747                             "Unable to receive reply header");
1748                         rw_unlock(&hio_remote_lock[ncomp]);
1749                         remote_close(res, ncomp);
1750                         continue;
1751                 }
1752                 rw_unlock(&hio_remote_lock[ncomp]);
1753                 seq = nv_get_uint64(nv, "seq");
1754                 if (seq == 0) {
1755                         pjdlog_error("Header contains no 'seq' field.");
1756                         nv_free(nv);
1757                         continue;
1758                 }
1759                 memsyncack = nv_exists(nv, "received");
1760                 mtx_lock(&hio_recv_list_lock[ncomp]);
1761                 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1762                         if (hio->hio_ggio.gctl_seq == seq) {
1763                                 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1764                                     hio_next[ncomp]);
1765                                 hio_recv_list_size[ncomp]--;
1766                                 break;
1767                         }
1768                 }
1769                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1770                 if (hio == NULL) {
1771                         pjdlog_error("Found no request matching received 'seq' field (%ju).",
1772                             (uintmax_t)seq);
1773                         nv_free(nv);
1774                         continue;
1775                 }
1776                 ggio = &hio->hio_ggio;
1777                 error = nv_get_int16(nv, "error");
1778                 if (error != 0) {
1779                         /* Request failed on remote side. */
1780                         hio->hio_errors[ncomp] = error;
1781                         reqlog(LOG_WARNING, 0, ggio,
1782                             "Remote request failed (%s): ", strerror(error));
1783                         nv_free(nv);
1784                         goto done_queue;
1785                 }
1786                 switch (ggio->gctl_cmd) {
1787                 case BIO_READ:
1788                         rw_rlock(&hio_remote_lock[ncomp]);
1789                         if (!ISCONNECTED(res, ncomp)) {
1790                                 rw_unlock(&hio_remote_lock[ncomp]);
1791                                 nv_free(nv);
1792                                 goto done_queue;
1793                         }
1794                         if (hast_proto_recv_data(res, res->hr_remotein, nv,
1795                             ggio->gctl_data, ggio->gctl_length) == -1) {
1796                                 hio->hio_errors[ncomp] = errno;
1797                                 pjdlog_errno(LOG_ERR,
1798                                     "Unable to receive reply data");
1799                                 rw_unlock(&hio_remote_lock[ncomp]);
1800                                 nv_free(nv);
1801                                 remote_close(res, ncomp);
1802                                 goto done_queue;
1803                         }
1804                         rw_unlock(&hio_remote_lock[ncomp]);
1805                         break;
1806                 case BIO_WRITE:
1807                 case BIO_DELETE:
1808                 case BIO_FLUSH:
1809                         break;
1810                 default:
1811                         PJDLOG_ABORT("invalid condition");
1812                 }
1813                 hio->hio_errors[ncomp] = 0;
1814                 nv_free(nv);
1815 done_queue:
1816                 if (ISMEMSYNCWRITE(hio)) {
1817                         if (!hio->hio_memsyncacked) {
1818                                 PJDLOG_ASSERT(memsyncack ||
1819                                     hio->hio_errors[ncomp] != 0);
1820                                 /* Remote ack arrived. */
1821                                 if (refcnt_release(&hio->hio_writecount) == 0) {
1822                                         if (hio->hio_errors[0] == 0)
1823                                                 write_complete(res, hio);
1824                                 }
1825                                 hio->hio_memsyncacked = true;
1826                                 if (hio->hio_errors[ncomp] == 0) {
1827                                         pjdlog_debug(2,
1828                                             "remote_recv: (%p) Moving request "
1829                                             "back to the recv queue.", hio);
1830                                         mtx_lock(&hio_recv_list_lock[ncomp]);
1831                                         TAILQ_INSERT_TAIL(&hio_recv_list[ncomp],
1832                                             hio, hio_next[ncomp]);
1833                                         hio_recv_list_size[ncomp]++;
1834                                         mtx_unlock(&hio_recv_list_lock[ncomp]);
1835                                         continue;
1836                                 }
1837                         } else {
1838                                 PJDLOG_ASSERT(!memsyncack);
1839                                 /* Remote final reply arrived. */
1840                         }
1841                 }
1842                 if (refcnt_release(&hio->hio_countdown) > 0)
1843                         continue;
1844                 if (ISSYNCREQ(hio)) {
1845                         mtx_lock(&sync_lock);
1846                         SYNCREQDONE(hio);
1847                         mtx_unlock(&sync_lock);
1848                         cv_signal(&sync_cond);
1849                 } else {
1850                         pjdlog_debug(2,
1851                             "remote_recv: (%p) Moving request to the done queue.",
1852                             hio);
1853                         QUEUE_INSERT2(hio, done);
1854                 }
1855         }
1856         /* NOTREACHED */
1857         return (NULL);
1858 }
1859
1860 /*
1861  * Thread sends answer to the kernel.
1862  */
1863 static void *
1864 ggate_send_thread(void *arg)
1865 {
1866         struct hast_resource *res = arg;
1867         struct g_gate_ctl_io *ggio;
1868         struct hio *hio;
1869         unsigned int ii, ncomps;
1870
1871         ncomps = HAST_NCOMPONENTS;
1872
1873         for (;;) {
1874                 pjdlog_debug(2, "ggate_send: Taking request.");
1875                 QUEUE_TAKE2(hio, done);
1876                 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1877                 ggio = &hio->hio_ggio;
1878                 for (ii = 0; ii < ncomps; ii++) {
1879                         if (hio->hio_errors[ii] == 0) {
1880                                 /*
1881                                  * One successful request is enough to declare
1882                                  * success.
1883                                  */
1884                                 ggio->gctl_error = 0;
1885                                 break;
1886                         }
1887                 }
1888                 if (ii == ncomps) {
1889                         /*
1890                          * None of the requests were successful.
1891                          * Use the error from local component except the
1892                          * case when we did only remote request.
1893                          */
1894                         if (ggio->gctl_cmd == BIO_READ &&
1895                             res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
1896                                 ggio->gctl_error = hio->hio_errors[1];
1897                         else
1898                                 ggio->gctl_error = hio->hio_errors[0];
1899                 }
1900                 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1901                         mtx_lock(&res->hr_amp_lock);
1902                         if (activemap_write_complete(res->hr_amp,
1903                             ggio->gctl_offset, ggio->gctl_length)) {
1904                                 res->hr_stat_activemap_update++;
1905                                 (void)hast_activemap_flush(res);
1906                         } else {
1907                                 mtx_unlock(&res->hr_amp_lock);
1908                         }
1909                 }
1910                 if (ggio->gctl_cmd == BIO_WRITE) {
1911                         /*
1912                          * Unlock range we locked.
1913                          */
1914                         mtx_lock(&range_lock);
1915                         rangelock_del(range_regular, ggio->gctl_offset,
1916                             ggio->gctl_length);
1917                         if (range_sync_wait)
1918                                 cv_signal(&range_sync_cond);
1919                         mtx_unlock(&range_lock);
1920                         if (!hio->hio_done)
1921                                 write_complete(res, hio);
1922                 } else {
1923                         if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1) {
1924                                 primary_exit(EX_OSERR,
1925                                     "G_GATE_CMD_DONE failed");
1926                         }
1927                 }
1928                 if (hio->hio_errors[0]) {
1929                         switch (ggio->gctl_cmd) {
1930                         case BIO_READ:
1931                                 res->hr_stat_read_error++;
1932                                 break;
1933                         case BIO_WRITE:
1934                                 res->hr_stat_write_error++;
1935                                 break;
1936                         case BIO_DELETE:
1937                                 res->hr_stat_delete_error++;
1938                                 break;
1939                         case BIO_FLUSH:
1940                                 res->hr_stat_flush_error++;
1941                                 break;
1942                         }
1943                 }
1944                 pjdlog_debug(2,
1945                     "ggate_send: (%p) Moving request to the free queue.", hio);
1946                 QUEUE_INSERT2(hio, free);
1947         }
1948         /* NOTREACHED */
1949         return (NULL);
1950 }
1951
1952 /*
1953  * Thread synchronize local and remote components.
1954  */
1955 static void *
1956 sync_thread(void *arg __unused)
1957 {
1958         struct hast_resource *res = arg;
1959         struct hio *hio;
1960         struct g_gate_ctl_io *ggio;
1961         struct timeval tstart, tend, tdiff;
1962         unsigned int ii, ncomp, ncomps;
1963         off_t offset, length, synced;
1964         bool dorewind, directreads;
1965         int syncext;
1966
1967         ncomps = HAST_NCOMPONENTS;
1968         dorewind = true;
1969         synced = 0;
1970         offset = -1;
1971         directreads = false;
1972
1973         for (;;) {
1974                 mtx_lock(&sync_lock);
1975                 if (offset >= 0 && !sync_inprogress) {
1976                         gettimeofday(&tend, NULL);
1977                         timersub(&tend, &tstart, &tdiff);
1978                         pjdlog_info("Synchronization interrupted after %#.0T. "
1979                             "%NB synchronized so far.", &tdiff,
1980                             (intmax_t)synced);
1981                         event_send(res, EVENT_SYNCINTR);
1982                 }
1983                 while (!sync_inprogress) {
1984                         dorewind = true;
1985                         synced = 0;
1986                         cv_wait(&sync_cond, &sync_lock);
1987                 }
1988                 mtx_unlock(&sync_lock);
1989                 /*
1990                  * Obtain offset at which we should synchronize.
1991                  * Rewind synchronization if needed.
1992                  */
1993                 mtx_lock(&res->hr_amp_lock);
1994                 if (dorewind)
1995                         activemap_sync_rewind(res->hr_amp);
1996                 offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1997                 if (syncext != -1) {
1998                         /*
1999                          * We synchronized entire syncext extent, we can mark
2000                          * it as clean now.
2001                          */
2002                         if (activemap_extent_complete(res->hr_amp, syncext))
2003                                 (void)hast_activemap_flush(res);
2004                         else
2005                                 mtx_unlock(&res->hr_amp_lock);
2006                 } else {
2007                         mtx_unlock(&res->hr_amp_lock);
2008                 }
2009                 if (dorewind) {
2010                         dorewind = false;
2011                         if (offset == -1)
2012                                 pjdlog_info("Nodes are in sync.");
2013                         else {
2014                                 pjdlog_info("Synchronization started. %NB to go.",
2015                                     (intmax_t)(res->hr_extentsize *
2016                                     activemap_ndirty(res->hr_amp)));
2017                                 event_send(res, EVENT_SYNCSTART);
2018                                 gettimeofday(&tstart, NULL);
2019                         }
2020                 }
2021                 if (offset == -1) {
2022                         sync_stop();
2023                         pjdlog_debug(1, "Nothing to synchronize.");
2024                         /*
2025                          * Synchronization complete, make both localcnt and
2026                          * remotecnt equal.
2027                          */
2028                         ncomp = 1;
2029                         rw_rlock(&hio_remote_lock[ncomp]);
2030                         if (ISCONNECTED(res, ncomp)) {
2031                                 if (synced > 0) {
2032                                         int64_t bps;
2033
2034                                         gettimeofday(&tend, NULL);
2035                                         timersub(&tend, &tstart, &tdiff);
2036                                         bps = (int64_t)((double)synced /
2037                                             ((double)tdiff.tv_sec +
2038                                             (double)tdiff.tv_usec / 1000000));
2039                                         pjdlog_info("Synchronization complete. "
2040                                             "%NB synchronized in %#.0lT (%NB/sec).",
2041                                             (intmax_t)synced, &tdiff,
2042                                             (intmax_t)bps);
2043                                         event_send(res, EVENT_SYNCDONE);
2044                                 }
2045                                 mtx_lock(&metadata_lock);
2046                                 if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
2047                                         directreads = true;
2048                                 res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
2049                                 res->hr_primary_localcnt =
2050                                     res->hr_secondary_remotecnt;
2051                                 res->hr_primary_remotecnt =
2052                                     res->hr_secondary_localcnt;
2053                                 pjdlog_debug(1,
2054                                     "Setting localcnt to %ju and remotecnt to %ju.",
2055                                     (uintmax_t)res->hr_primary_localcnt,
2056                                     (uintmax_t)res->hr_primary_remotecnt);
2057                                 (void)metadata_write(res);
2058                                 mtx_unlock(&metadata_lock);
2059                         }
2060                         rw_unlock(&hio_remote_lock[ncomp]);
2061                         if (directreads) {
2062                                 directreads = false;
2063                                 enable_direct_reads(res);
2064                         }
2065                         continue;
2066                 }
2067                 pjdlog_debug(2, "sync: Taking free request.");
2068                 QUEUE_TAKE2(hio, free);
2069                 pjdlog_debug(2, "sync: (%p) Got free request.", hio);
2070                 /*
2071                  * Lock the range we are going to synchronize. We don't want
2072                  * race where someone writes between our read and write.
2073                  */
2074                 for (;;) {
2075                         mtx_lock(&range_lock);
2076                         if (rangelock_islocked(range_regular, offset, length)) {
2077                                 pjdlog_debug(2,
2078                                     "sync: Range offset=%jd length=%jd locked.",
2079                                     (intmax_t)offset, (intmax_t)length);
2080                                 range_sync_wait = true;
2081                                 cv_wait(&range_sync_cond, &range_lock);
2082                                 range_sync_wait = false;
2083                                 mtx_unlock(&range_lock);
2084                                 continue;
2085                         }
2086                         if (rangelock_add(range_sync, offset, length) == -1) {
2087                                 mtx_unlock(&range_lock);
2088                                 pjdlog_debug(2,
2089                                     "sync: Range offset=%jd length=%jd is already locked, waiting.",
2090                                     (intmax_t)offset, (intmax_t)length);
2091                                 sleep(1);
2092                                 continue;
2093                         }
2094                         mtx_unlock(&range_lock);
2095                         break;
2096                 }
2097                 /*
2098                  * First read the data from synchronization source.
2099                  */
2100                 SYNCREQ(hio);
2101                 ggio = &hio->hio_ggio;
2102                 ggio->gctl_cmd = BIO_READ;
2103                 ggio->gctl_offset = offset;
2104                 ggio->gctl_length = length;
2105                 ggio->gctl_error = 0;
2106                 hio->hio_done = false;
2107                 hio->hio_replication = res->hr_replication;
2108                 for (ii = 0; ii < ncomps; ii++)
2109                         hio->hio_errors[ii] = EINVAL;
2110                 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2111                     hio);
2112                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2113                     hio);
2114                 mtx_lock(&metadata_lock);
2115                 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2116                         /*
2117                          * This range is up-to-date on local component,
2118                          * so handle request locally.
2119                          */
2120                          /* Local component is 0 for now. */
2121                         ncomp = 0;
2122                 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2123                         PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2124                         /*
2125                          * This range is out-of-date on local component,
2126                          * so send request to the remote node.
2127                          */
2128                          /* Remote component is 1 for now. */
2129                         ncomp = 1;
2130                 }
2131                 mtx_unlock(&metadata_lock);
2132                 refcnt_init(&hio->hio_countdown, 1);
2133                 QUEUE_INSERT1(hio, send, ncomp);
2134
2135                 /*
2136                  * Let's wait for READ to finish.
2137                  */
2138                 mtx_lock(&sync_lock);
2139                 while (!ISSYNCREQDONE(hio))
2140                         cv_wait(&sync_cond, &sync_lock);
2141                 mtx_unlock(&sync_lock);
2142
2143                 if (hio->hio_errors[ncomp] != 0) {
2144                         pjdlog_error("Unable to read synchronization data: %s.",
2145                             strerror(hio->hio_errors[ncomp]));
2146                         goto free_queue;
2147                 }
2148
2149                 /*
2150                  * We read the data from synchronization source, now write it
2151                  * to synchronization target.
2152                  */
2153                 SYNCREQ(hio);
2154                 ggio->gctl_cmd = BIO_WRITE;
2155                 for (ii = 0; ii < ncomps; ii++)
2156                         hio->hio_errors[ii] = EINVAL;
2157                 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2158                     hio);
2159                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2160                     hio);
2161                 mtx_lock(&metadata_lock);
2162                 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2163                         /*
2164                          * This range is up-to-date on local component,
2165                          * so we update remote component.
2166                          */
2167                          /* Remote component is 1 for now. */
2168                         ncomp = 1;
2169                 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2170                         PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2171                         /*
2172                          * This range is out-of-date on local component,
2173                          * so we update it.
2174                          */
2175                          /* Local component is 0 for now. */
2176                         ncomp = 0;
2177                 }
2178                 mtx_unlock(&metadata_lock);
2179
2180                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2181                     hio);
2182                 refcnt_init(&hio->hio_countdown, 1);
2183                 QUEUE_INSERT1(hio, send, ncomp);
2184
2185                 /*
2186                  * Let's wait for WRITE to finish.
2187                  */
2188                 mtx_lock(&sync_lock);
2189                 while (!ISSYNCREQDONE(hio))
2190                         cv_wait(&sync_cond, &sync_lock);
2191                 mtx_unlock(&sync_lock);
2192
2193                 if (hio->hio_errors[ncomp] != 0) {
2194                         pjdlog_error("Unable to write synchronization data: %s.",
2195                             strerror(hio->hio_errors[ncomp]));
2196                         goto free_queue;
2197                 }
2198
2199                 synced += length;
2200 free_queue:
2201                 mtx_lock(&range_lock);
2202                 rangelock_del(range_sync, offset, length);
2203                 if (range_regular_wait)
2204                         cv_signal(&range_regular_cond);
2205                 mtx_unlock(&range_lock);
2206                 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
2207                     hio);
2208                 QUEUE_INSERT2(hio, free);
2209         }
2210         /* NOTREACHED */
2211         return (NULL);
2212 }
2213
2214 void
2215 primary_config_reload(struct hast_resource *res, struct nv *nv)
2216 {
2217         unsigned int ii, ncomps;
2218         int modified, vint;
2219         const char *vstr;
2220
2221         pjdlog_info("Reloading configuration...");
2222
2223         PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
2224         PJDLOG_ASSERT(gres == res);
2225         nv_assert(nv, "remoteaddr");
2226         nv_assert(nv, "sourceaddr");
2227         nv_assert(nv, "replication");
2228         nv_assert(nv, "checksum");
2229         nv_assert(nv, "compression");
2230         nv_assert(nv, "timeout");
2231         nv_assert(nv, "exec");
2232         nv_assert(nv, "metaflush");
2233
2234         ncomps = HAST_NCOMPONENTS;
2235
2236 #define MODIFIED_REMOTEADDR     0x01
2237 #define MODIFIED_SOURCEADDR     0x02
2238 #define MODIFIED_REPLICATION    0x04
2239 #define MODIFIED_CHECKSUM       0x08
2240 #define MODIFIED_COMPRESSION    0x10
2241 #define MODIFIED_TIMEOUT        0x20
2242 #define MODIFIED_EXEC           0x40
2243 #define MODIFIED_METAFLUSH      0x80
2244         modified = 0;
2245
2246         vstr = nv_get_string(nv, "remoteaddr");
2247         if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
2248                 /*
2249                  * Don't copy res->hr_remoteaddr to gres just yet.
2250                  * We want remote_close() to log disconnect from the old
2251                  * addresses, not from the new ones.
2252                  */
2253                 modified |= MODIFIED_REMOTEADDR;
2254         }
2255         vstr = nv_get_string(nv, "sourceaddr");
2256         if (strcmp(gres->hr_sourceaddr, vstr) != 0) {
2257                 strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr));
2258                 modified |= MODIFIED_SOURCEADDR;
2259         }
2260         vint = nv_get_int32(nv, "replication");
2261         if (gres->hr_replication != vint) {
2262                 gres->hr_replication = vint;
2263                 modified |= MODIFIED_REPLICATION;
2264         }
2265         vint = nv_get_int32(nv, "checksum");
2266         if (gres->hr_checksum != vint) {
2267                 gres->hr_checksum = vint;
2268                 modified |= MODIFIED_CHECKSUM;
2269         }
2270         vint = nv_get_int32(nv, "compression");
2271         if (gres->hr_compression != vint) {
2272                 gres->hr_compression = vint;
2273                 modified |= MODIFIED_COMPRESSION;
2274         }
2275         vint = nv_get_int32(nv, "timeout");
2276         if (gres->hr_timeout != vint) {
2277                 gres->hr_timeout = vint;
2278                 modified |= MODIFIED_TIMEOUT;
2279         }
2280         vstr = nv_get_string(nv, "exec");
2281         if (strcmp(gres->hr_exec, vstr) != 0) {
2282                 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
2283                 modified |= MODIFIED_EXEC;
2284         }
2285         vint = nv_get_int32(nv, "metaflush");
2286         if (gres->hr_metaflush != vint) {
2287                 gres->hr_metaflush = vint;
2288                 modified |= MODIFIED_METAFLUSH;
2289         }
2290
2291         /*
2292          * Change timeout for connected sockets.
2293          * Don't bother if we need to reconnect.
2294          */
2295         if ((modified & MODIFIED_TIMEOUT) != 0 &&
2296             (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) == 0) {
2297                 for (ii = 0; ii < ncomps; ii++) {
2298                         if (!ISREMOTE(ii))
2299                                 continue;
2300                         rw_rlock(&hio_remote_lock[ii]);
2301                         if (!ISCONNECTED(gres, ii)) {
2302                                 rw_unlock(&hio_remote_lock[ii]);
2303                                 continue;
2304                         }
2305                         rw_unlock(&hio_remote_lock[ii]);
2306                         if (proto_timeout(gres->hr_remotein,
2307                             gres->hr_timeout) == -1) {
2308                                 pjdlog_errno(LOG_WARNING,
2309                                     "Unable to set connection timeout");
2310                         }
2311                         if (proto_timeout(gres->hr_remoteout,
2312                             gres->hr_timeout) == -1) {
2313                                 pjdlog_errno(LOG_WARNING,
2314                                     "Unable to set connection timeout");
2315                         }
2316                 }
2317         }
2318         if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) != 0) {
2319                 for (ii = 0; ii < ncomps; ii++) {
2320                         if (!ISREMOTE(ii))
2321                                 continue;
2322                         remote_close(gres, ii);
2323                 }
2324                 if (modified & MODIFIED_REMOTEADDR) {
2325                         vstr = nv_get_string(nv, "remoteaddr");
2326                         strlcpy(gres->hr_remoteaddr, vstr,
2327                             sizeof(gres->hr_remoteaddr));
2328                 }
2329         }
2330 #undef  MODIFIED_REMOTEADDR
2331 #undef  MODIFIED_SOURCEADDR
2332 #undef  MODIFIED_REPLICATION
2333 #undef  MODIFIED_CHECKSUM
2334 #undef  MODIFIED_COMPRESSION
2335 #undef  MODIFIED_TIMEOUT
2336 #undef  MODIFIED_EXEC
2337 #undef  MODIFIED_METAFLUSH
2338
2339         pjdlog_info("Configuration reloaded successfully.");
2340 }
2341
2342 static void
2343 guard_one(struct hast_resource *res, unsigned int ncomp)
2344 {
2345         struct proto_conn *in, *out;
2346
2347         if (!ISREMOTE(ncomp))
2348                 return;
2349
2350         rw_rlock(&hio_remote_lock[ncomp]);
2351
2352         if (!real_remote(res)) {
2353                 rw_unlock(&hio_remote_lock[ncomp]);
2354                 return;
2355         }
2356
2357         if (ISCONNECTED(res, ncomp)) {
2358                 PJDLOG_ASSERT(res->hr_remotein != NULL);
2359                 PJDLOG_ASSERT(res->hr_remoteout != NULL);
2360                 rw_unlock(&hio_remote_lock[ncomp]);
2361                 pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2362                     res->hr_remoteaddr);
2363                 return;
2364         }
2365
2366         PJDLOG_ASSERT(res->hr_remotein == NULL);
2367         PJDLOG_ASSERT(res->hr_remoteout == NULL);
2368         /*
2369          * Upgrade the lock. It doesn't have to be atomic as no other thread
2370          * can change connection status from disconnected to connected.
2371          */
2372         rw_unlock(&hio_remote_lock[ncomp]);
2373         pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2374             res->hr_remoteaddr);
2375         in = out = NULL;
2376         if (init_remote(res, &in, &out) == 0) {
2377                 rw_wlock(&hio_remote_lock[ncomp]);
2378                 PJDLOG_ASSERT(res->hr_remotein == NULL);
2379                 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2380                 PJDLOG_ASSERT(in != NULL && out != NULL);
2381                 res->hr_remotein = in;
2382                 res->hr_remoteout = out;
2383                 rw_unlock(&hio_remote_lock[ncomp]);
2384                 pjdlog_info("Successfully reconnected to %s.",
2385                     res->hr_remoteaddr);
2386                 sync_start();
2387         } else {
2388                 /* Both connections should be NULL. */
2389                 PJDLOG_ASSERT(res->hr_remotein == NULL);
2390                 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2391                 PJDLOG_ASSERT(in == NULL && out == NULL);
2392                 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2393                     res->hr_remoteaddr);
2394         }
2395 }
2396
2397 /*
2398  * Thread guards remote connections and reconnects when needed, handles
2399  * signals, etc.
2400  */
2401 static void *
2402 guard_thread(void *arg)
2403 {
2404         struct hast_resource *res = arg;
2405         unsigned int ii, ncomps;
2406         struct timespec timeout;
2407         time_t lastcheck, now;
2408         sigset_t mask;
2409         int signo;
2410
2411         ncomps = HAST_NCOMPONENTS;
2412         lastcheck = time(NULL);
2413
2414         PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2415         PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2416         PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2417
2418         timeout.tv_sec = HAST_KEEPALIVE;
2419         timeout.tv_nsec = 0;
2420         signo = -1;
2421
2422         for (;;) {
2423                 switch (signo) {
2424                 case SIGINT:
2425                 case SIGTERM:
2426                         sigexit_received = true;
2427                         primary_exitx(EX_OK,
2428                             "Termination signal received, exiting.");
2429                         break;
2430                 default:
2431                         break;
2432                 }
2433
2434                 /*
2435                  * Don't check connections until we fully started,
2436                  * as we may still be looping, waiting for remote node
2437                  * to switch from primary to secondary.
2438                  */
2439                 if (fullystarted) {
2440                         pjdlog_debug(2, "remote_guard: Checking connections.");
2441                         now = time(NULL);
2442                         if (lastcheck + HAST_KEEPALIVE <= now) {
2443                                 for (ii = 0; ii < ncomps; ii++)
2444                                         guard_one(res, ii);
2445                                 lastcheck = now;
2446                         }
2447                 }
2448                 signo = sigtimedwait(&mask, NULL, &timeout);
2449         }
2450         /* NOTREACHED */
2451         return (NULL);
2452 }