sbin/hastd/primary.c

   1 /*-
   2  * Copyright (c) 2009 The FreeBSD Foundation
   3  * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org>
   4  * All rights reserved.
   5  *
   6  * This software was developed by Pawel Jakub Dawidek under sponsorship from
   7  * the FreeBSD Foundation.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include <sys/types.h>
  35 #include <sys/time.h>
  36 #include <sys/bio.h>
  37 #include <sys/disk.h>
  38 #include <sys/refcount.h>
  39 #include <sys/stat.h>
  40
  41 #include <geom/gate/g_gate.h>
  42
  43 #include <assert.h>
  44 #include <err.h>
  45 #include <errno.h>
  46 #include <fcntl.h>
  47 #include <libgeom.h>
  48 #include <pthread.h>
  49 #include <signal.h>
  50 #include <stdint.h>
  51 #include <stdio.h>
  52 #include <string.h>
  53 #include <sysexits.h>
  54 #include <unistd.h>
  55
  56 #include <activemap.h>
  57 #include <nv.h>
  58 #include <rangelock.h>
  59
  60 #include "control.h"
  61 #include "event.h"
  62 #include "hast.h"
  63 #include "hast_proto.h"
  64 #include "hastd.h"
  65 #include "hooks.h"
  66 #include "metadata.h"
  67 #include "proto.h"
  68 #include "pjdlog.h"
  69 #include "subr.h"
  70 #include "synch.h"
  71
  72 /* The is only one remote component for now. */
  73 #define ISREMOTE(no)    ((no) == 1)
  74
  75 struct hio {
  76         /*
  77          * Number of components we are still waiting for.
  78          * When this field goes to 0, we can send the request back to the
  79          * kernel. Each component has to decrease this counter by one
  80          * even on failure.
  81          */
  82         unsigned int             hio_countdown;
  83         /*
  84          * Each component has a place to store its own error.
  85          * Once the request is handled by all components we can decide if the
  86          * request overall is successful or not.
  87          */
  88         int                     *hio_errors;
  89         /*
  90          * Structure used to comunicate with GEOM Gate class.
  91          */
  92         struct g_gate_ctl_io     hio_ggio;
  93         TAILQ_ENTRY(hio)        *hio_next;
  94 };
  95 #define hio_free_next   hio_next[0]
  96 #define hio_done_next   hio_next[0]
  97
  98 /*
  99  * Free list holds unused structures. When free list is empty, we have to wait
 100  * until some in-progress requests are freed.
 101  */
 102 static TAILQ_HEAD(, hio) hio_free_list;
 103 static pthread_mutex_t hio_free_list_lock;
 104 static pthread_cond_t hio_free_list_cond;
 105 /*
 106  * There is one send list for every component. One requests is placed on all
 107  * send lists - each component gets the same request, but each component is
 108  * responsible for managing his own send list.
 109  */
 110 static TAILQ_HEAD(, hio) *hio_send_list;
 111 static pthread_mutex_t *hio_send_list_lock;
 112 static pthread_cond_t *hio_send_list_cond;
 113 /*
 114  * There is one recv list for every component, although local components don't
 115  * use recv lists as local requests are done synchronously.
 116  */
 117 static TAILQ_HEAD(, hio) *hio_recv_list;
 118 static pthread_mutex_t *hio_recv_list_lock;
 119 static pthread_cond_t *hio_recv_list_cond;
 120 /*
 121  * Request is placed on done list by the slowest component (the one that
 122  * decreased hio_countdown from 1 to 0).
 123  */
 124 static TAILQ_HEAD(, hio) hio_done_list;
 125 static pthread_mutex_t hio_done_list_lock;
 126 static pthread_cond_t hio_done_list_cond;
 127 /*
 128  * Structure below are for interaction with sync thread.
 129  */
 130 static bool sync_inprogress;
 131 static pthread_mutex_t sync_lock;
 132 static pthread_cond_t sync_cond;
 133 /*
 134  * The lock below allows to synchornize access to remote connections.
 135  */
 136 static pthread_rwlock_t *hio_remote_lock;
 137
 138 /*
 139  * Lock to synchronize metadata updates. Also synchronize access to
 140  * hr_primary_localcnt and hr_primary_remotecnt fields.
 141  */
 142 static pthread_mutex_t metadata_lock;
 143
 144 /*
 145  * Maximum number of outstanding I/O requests.
 146  */
 147 #define HAST_HIO_MAX    256
 148 /*
 149  * Number of components. At this point there are only two components: local
 150  * and remote, but in the future it might be possible to use multiple local
 151  * and remote components.
 152  */
 153 #define HAST_NCOMPONENTS        2
 154 /*
 155  * Number of seconds to sleep between reconnect retries or keepalive packets.
 156  */
 157 #define RETRY_SLEEP             10
 158
 159 #define ISCONNECTED(res, no)    \
 160         ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
 161
 162 #define QUEUE_INSERT1(hio, name, ncomp) do {                            \
 163         bool _wakeup;                                                   \
 164                                                                         \
 165         mtx_lock(&hio_##name##_list_lock[(ncomp)]);                     \
 166         _wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]);             \
 167         TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),           \
 168             hio_next[(ncomp)]);                                         \
 169         mtx_unlock(&hio_##name##_list_lock[ncomp]);                     \
 170         if (_wakeup)                                                    \
 171                 cv_signal(&hio_##name##_list_cond[(ncomp)]);            \
 172 } while (0)
 173 #define QUEUE_INSERT2(hio, name)        do {                            \
 174         bool _wakeup;                                                   \
 175                                                                         \
 176         mtx_lock(&hio_##name##_list_lock);                              \
 177         _wakeup = TAILQ_EMPTY(&hio_##name##_list);                      \
 178         TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
 179         mtx_unlock(&hio_##name##_list_lock);                            \
 180         if (_wakeup)                                                    \
 181                 cv_signal(&hio_##name##_list_cond);                     \
 182 } while (0)
 183 #define QUEUE_TAKE1(hio, name, ncomp, timeout)  do {                    \
 184         bool _last;                                                     \
 185                                                                         \
 186         mtx_lock(&hio_##name##_list_lock[(ncomp)]);                     \
 187         _last = false;                                                  \
 188         while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
 189                 cv_timedwait(&hio_##name##_list_cond[(ncomp)],          \
 190                     &hio_##name##_list_lock[(ncomp)], (timeout));       \
 191                 if ((timeout) != 0)                                     \
 192                         _last = true;                                   \
 193         }                                                               \
 194         if (hio != NULL) {                                              \
 195                 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),        \
 196                     hio_next[(ncomp)]);                                 \
 197         }                                                               \
 198         mtx_unlock(&hio_##name##_list_lock[(ncomp)]);                   \
 199 } while (0)
 200 #define QUEUE_TAKE2(hio, name)  do {                                    \
 201         mtx_lock(&hio_##name##_list_lock);                              \
 202         while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {     \
 203                 cv_wait(&hio_##name##_list_cond,                        \
 204                     &hio_##name##_list_lock);                           \
 205         }                                                               \
 206         TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);     \
 207         mtx_unlock(&hio_##name##_list_lock);                            \
 208 } while (0)
 209
 210 #define SYNCREQ(hio)            do {                                    \
 211         (hio)->hio_ggio.gctl_unit = -1;                                 \
 212         (hio)->hio_ggio.gctl_seq = 1;                                   \
 213 } while (0)
 214 #define ISSYNCREQ(hio)          ((hio)->hio_ggio.gctl_unit == -1)
 215 #define SYNCREQDONE(hio)        do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
 216 #define ISSYNCREQDONE(hio)      ((hio)->hio_ggio.gctl_unit == -2)
 217
 218 static struct hast_resource *gres;
 219
 220 static pthread_mutex_t range_lock;
 221 static struct rangelocks *range_regular;
 222 static bool range_regular_wait;
 223 static pthread_cond_t range_regular_cond;
 224 static struct rangelocks *range_sync;
 225 static bool range_sync_wait;
 226 static pthread_cond_t range_sync_cond;
 227
 228 static void *ggate_recv_thread(void *arg);
 229 static void *local_send_thread(void *arg);
 230 static void *remote_send_thread(void *arg);
 231 static void *remote_recv_thread(void *arg);
 232 static void *ggate_send_thread(void *arg);
 233 static void *sync_thread(void *arg);
 234 static void *guard_thread(void *arg);
 235
 236 static void
 237 cleanup(struct hast_resource *res)
 238 {
 239         int rerrno;
 240
 241         /* Remember errno. */
 242         rerrno = errno;
 243
 244         /* Destroy ggate provider if we created one. */
 245         if (res->hr_ggateunit >= 0) {
 246                 struct g_gate_ctl_destroy ggiod;
 247
 248                 bzero(&ggiod, sizeof(ggiod));
 249                 ggiod.gctl_version = G_GATE_VERSION;
 250                 ggiod.gctl_unit = res->hr_ggateunit;
 251                 ggiod.gctl_force = 1;
 252                 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) {
 253                         pjdlog_errno(LOG_WARNING,
 254                             "Unable to destroy hast/%s device",
 255                             res->hr_provname);
 256                 }
 257                 res->hr_ggateunit = -1;
 258         }
 259
 260         /* Restore errno. */
 261         errno = rerrno;
 262 }
 263
 264 static __dead2 void
 265 primary_exit(int exitcode, const char *fmt, ...)
 266 {
 267         va_list ap;
 268
 269         assert(exitcode != EX_OK);
 270         va_start(ap, fmt);
 271         pjdlogv_errno(LOG_ERR, fmt, ap);
 272         va_end(ap);
 273         cleanup(gres);
 274         exit(exitcode);
 275 }
 276
 277 static __dead2 void
 278 primary_exitx(int exitcode, const char *fmt, ...)
 279 {
 280         va_list ap;
 281
 282         va_start(ap, fmt);
 283         pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
 284         va_end(ap);
 285         cleanup(gres);
 286         exit(exitcode);
 287 }
 288
 289 static int
 290 hast_activemap_flush(struct hast_resource *res)
 291 {
 292         const unsigned char *buf;
 293         size_t size;
 294
 295         buf = activemap_bitmap(res->hr_amp, &size);
 296         assert(buf != NULL);
 297         assert((size % res->hr_local_sectorsize) == 0);
 298         if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
 299             (ssize_t)size) {
 300                 KEEP_ERRNO(pjdlog_errno(LOG_ERR,
 301                     "Unable to flush activemap to disk"));
 302                 return (-1);
 303         }
 304         return (0);
 305 }
 306
 307 static bool
 308 real_remote(const struct hast_resource *res)
 309 {
 310
 311         return (strcmp(res->hr_remoteaddr, "none") != 0);
 312 }
 313
 314 static void
 315 init_environment(struct hast_resource *res __unused)
 316 {
 317         struct hio *hio;
 318         unsigned int ii, ncomps;
 319
 320         /*
 321          * In the future it might be per-resource value.
 322          */
 323         ncomps = HAST_NCOMPONENTS;
 324
 325         /*
 326          * Allocate memory needed by lists.
 327          */
 328         hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
 329         if (hio_send_list == NULL) {
 330                 primary_exitx(EX_TEMPFAIL,
 331                     "Unable to allocate %zu bytes of memory for send lists.",
 332                     sizeof(hio_send_list[0]) * ncomps);
 333         }
 334         hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
 335         if (hio_send_list_lock == NULL) {
 336                 primary_exitx(EX_TEMPFAIL,
 337                     "Unable to allocate %zu bytes of memory for send list locks.",
 338                     sizeof(hio_send_list_lock[0]) * ncomps);
 339         }
 340         hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
 341         if (hio_send_list_cond == NULL) {
 342                 primary_exitx(EX_TEMPFAIL,
 343                     "Unable to allocate %zu bytes of memory for send list condition variables.",
 344                     sizeof(hio_send_list_cond[0]) * ncomps);
 345         }
 346         hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
 347         if (hio_recv_list == NULL) {
 348                 primary_exitx(EX_TEMPFAIL,
 349                     "Unable to allocate %zu bytes of memory for recv lists.",
 350                     sizeof(hio_recv_list[0]) * ncomps);
 351         }
 352         hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
 353         if (hio_recv_list_lock == NULL) {
 354                 primary_exitx(EX_TEMPFAIL,
 355                     "Unable to allocate %zu bytes of memory for recv list locks.",
 356                     sizeof(hio_recv_list_lock[0]) * ncomps);
 357         }
 358         hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
 359         if (hio_recv_list_cond == NULL) {
 360                 primary_exitx(EX_TEMPFAIL,
 361                     "Unable to allocate %zu bytes of memory for recv list condition variables.",
 362                     sizeof(hio_recv_list_cond[0]) * ncomps);
 363         }
 364         hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
 365         if (hio_remote_lock == NULL) {
 366                 primary_exitx(EX_TEMPFAIL,
 367                     "Unable to allocate %zu bytes of memory for remote connections locks.",
 368                     sizeof(hio_remote_lock[0]) * ncomps);
 369         }
 370
 371         /*
 372          * Initialize lists, their locks and theirs condition variables.
 373          */
 374         TAILQ_INIT(&hio_free_list);
 375         mtx_init(&hio_free_list_lock);
 376         cv_init(&hio_free_list_cond);
 377         for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
 378                 TAILQ_INIT(&hio_send_list[ii]);
 379                 mtx_init(&hio_send_list_lock[ii]);
 380                 cv_init(&hio_send_list_cond[ii]);
 381                 TAILQ_INIT(&hio_recv_list[ii]);
 382                 mtx_init(&hio_recv_list_lock[ii]);
 383                 cv_init(&hio_recv_list_cond[ii]);
 384                 rw_init(&hio_remote_lock[ii]);
 385         }
 386         TAILQ_INIT(&hio_done_list);
 387         mtx_init(&hio_done_list_lock);
 388         cv_init(&hio_done_list_cond);
 389         mtx_init(&metadata_lock);
 390
 391         /*
 392          * Allocate requests pool and initialize requests.
 393          */
 394         for (ii = 0; ii < HAST_HIO_MAX; ii++) {
 395                 hio = malloc(sizeof(*hio));
 396                 if (hio == NULL) {
 397                         primary_exitx(EX_TEMPFAIL,
 398                             "Unable to allocate %zu bytes of memory for hio request.",
 399                             sizeof(*hio));
 400                 }
 401                 hio->hio_countdown = 0;
 402                 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
 403                 if (hio->hio_errors == NULL) {
 404                         primary_exitx(EX_TEMPFAIL,
 405                             "Unable allocate %zu bytes of memory for hio errors.",
 406                             sizeof(hio->hio_errors[0]) * ncomps);
 407                 }
 408                 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
 409                 if (hio->hio_next == NULL) {
 410                         primary_exitx(EX_TEMPFAIL,
 411                             "Unable allocate %zu bytes of memory for hio_next field.",
 412                             sizeof(hio->hio_next[0]) * ncomps);
 413                 }
 414                 hio->hio_ggio.gctl_version = G_GATE_VERSION;
 415                 hio->hio_ggio.gctl_data = malloc(MAXPHYS);
 416                 if (hio->hio_ggio.gctl_data == NULL) {
 417                         primary_exitx(EX_TEMPFAIL,
 418                             "Unable to allocate %zu bytes of memory for gctl_data.",
 419                             MAXPHYS);
 420                 }
 421                 hio->hio_ggio.gctl_length = MAXPHYS;
 422                 hio->hio_ggio.gctl_error = 0;
 423                 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
 424         }
 425 }
 426
 427 static bool
 428 init_resuid(struct hast_resource *res)
 429 {
 430
 431         mtx_lock(&metadata_lock);
 432         if (res->hr_resuid != 0) {
 433                 mtx_unlock(&metadata_lock);
 434                 return (false);
 435         } else {
 436                 /* Initialize unique resource identifier. */
 437                 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
 438                 mtx_unlock(&metadata_lock);
 439                 if (metadata_write(res) < 0)
 440                         exit(EX_NOINPUT);
 441                 return (true);
 442         }
 443 }
 444
 445 static void
 446 init_local(struct hast_resource *res)
 447 {
 448         unsigned char *buf;
 449         size_t mapsize;
 450
 451         if (metadata_read(res, true) < 0)
 452                 exit(EX_NOINPUT);
 453         mtx_init(&res->hr_amp_lock);
 454         if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
 455             res->hr_local_sectorsize, res->hr_keepdirty) < 0) {
 456                 primary_exit(EX_TEMPFAIL, "Unable to create activemap");
 457         }
 458         mtx_init(&range_lock);
 459         cv_init(&range_regular_cond);
 460         if (rangelock_init(&range_regular) < 0)
 461                 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
 462         cv_init(&range_sync_cond);
 463         if (rangelock_init(&range_sync) < 0)
 464                 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
 465         mapsize = activemap_ondisk_size(res->hr_amp);
 466         buf = calloc(1, mapsize);
 467         if (buf == NULL) {
 468                 primary_exitx(EX_TEMPFAIL,
 469                     "Unable to allocate buffer for activemap.");
 470         }
 471         if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
 472             (ssize_t)mapsize) {
 473                 primary_exit(EX_NOINPUT, "Unable to read activemap");
 474         }
 475         activemap_copyin(res->hr_amp, buf, mapsize);
 476         free(buf);
 477         if (res->hr_resuid != 0)
 478                 return;
 479         /*
 480          * We're using provider for the first time. Initialize local and remote
 481          * counters. We don't initialize resuid here, as we want to do it just
 482          * in time. The reason for this is that we want to inform secondary
 483          * that there were no writes yet, so there is no need to synchronize
 484          * anything.
 485          */
 486         res->hr_primary_localcnt = 1;
 487         res->hr_primary_remotecnt = 0;
 488         if (metadata_write(res) < 0)
 489                 exit(EX_NOINPUT);
 490 }
 491
 492 static bool
 493 init_remote(struct hast_resource *res, struct proto_conn **inp,
 494     struct proto_conn **outp)
 495 {
 496         struct proto_conn *in, *out;
 497         struct nv *nvout, *nvin;
 498         const unsigned char *token;
 499         unsigned char *map;
 500         const char *errmsg;
 501         int32_t extentsize;
 502         int64_t datasize;
 503         uint32_t mapsize;
 504         size_t size;
 505
 506         assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
 507         assert(real_remote(res));
 508
 509         in = out = NULL;
 510         errmsg = NULL;
 511
 512         /* Prepare outgoing connection with remote node. */
 513         if (proto_client(res->hr_remoteaddr, &out) < 0) {
 514                 primary_exit(EX_TEMPFAIL, "Unable to create connection to %s",
 515                     res->hr_remoteaddr);
 516         }
 517         /* Try to connect, but accept failure. */
 518         if (proto_connect(out) < 0) {
 519                 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
 520                     res->hr_remoteaddr);
 521                 goto close;
 522         }
 523         /* Error in setting timeout is not critical, but why should it fail? */
 524         if (proto_timeout(out, res->hr_timeout) < 0)
 525                 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
 526         /*
 527          * First handshake step.
 528          * Setup outgoing connection with remote node.
 529          */
 530         nvout = nv_alloc();
 531         nv_add_string(nvout, res->hr_name, "resource");
 532         if (nv_error(nvout) != 0) {
 533                 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
 534                     "Unable to allocate header for connection with %s",
 535                     res->hr_remoteaddr);
 536                 nv_free(nvout);
 537                 goto close;
 538         }
 539         if (hast_proto_send(res, out, nvout, NULL, 0) < 0) {
 540                 pjdlog_errno(LOG_WARNING,
 541                     "Unable to send handshake header to %s",
 542                     res->hr_remoteaddr);
 543                 nv_free(nvout);
 544                 goto close;
 545         }
 546         nv_free(nvout);
 547         if (hast_proto_recv_hdr(out, &nvin) < 0) {
 548                 pjdlog_errno(LOG_WARNING,
 549                     "Unable to receive handshake header from %s",
 550                     res->hr_remoteaddr);
 551                 goto close;
 552         }
 553         errmsg = nv_get_string(nvin, "errmsg");
 554         if (errmsg != NULL) {
 555                 pjdlog_warning("%s", errmsg);
 556                 nv_free(nvin);
 557                 goto close;
 558         }
 559         token = nv_get_uint8_array(nvin, &size, "token");
 560         if (token == NULL) {
 561                 pjdlog_warning("Handshake header from %s has no 'token' field.",
 562                     res->hr_remoteaddr);
 563                 nv_free(nvin);
 564                 goto close;
 565         }
 566         if (size != sizeof(res->hr_token)) {
 567                 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
 568                     res->hr_remoteaddr, size, sizeof(res->hr_token));
 569                 nv_free(nvin);
 570                 goto close;
 571         }
 572         bcopy(token, res->hr_token, sizeof(res->hr_token));
 573         nv_free(nvin);
 574
 575         /*
 576          * Second handshake step.
 577          * Setup incoming connection with remote node.
 578          */
 579         if (proto_client(res->hr_remoteaddr, &in) < 0) {
 580                 pjdlog_errno(LOG_WARNING, "Unable to create connection to %s",
 581                     res->hr_remoteaddr);
 582         }
 583         /* Try to connect, but accept failure. */
 584         if (proto_connect(in) < 0) {
 585                 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
 586                     res->hr_remoteaddr);
 587                 goto close;
 588         }
 589         /* Error in setting timeout is not critical, but why should it fail? */
 590         if (proto_timeout(in, res->hr_timeout) < 0)
 591                 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
 592         nvout = nv_alloc();
 593         nv_add_string(nvout, res->hr_name, "resource");
 594         nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
 595             "token");
 596         if (res->hr_resuid == 0) {
 597                 /*
 598                  * The resuid field was not yet initialized.
 599                  * Because we do synchronization inside init_resuid(), it is
 600                  * possible that someone already initialized it, the function
 601                  * will return false then, but if we successfully initialized
 602                  * it, we will get true. True means that there were no writes
 603                  * to this resource yet and we want to inform secondary that
 604                  * synchronization is not needed by sending "virgin" argument.
 605                  */
 606                 if (init_resuid(res))
 607                         nv_add_int8(nvout, 1, "virgin");
 608         }
 609         nv_add_uint64(nvout, res->hr_resuid, "resuid");
 610         nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
 611         nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
 612         if (nv_error(nvout) != 0) {
 613                 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
 614                     "Unable to allocate header for connection with %s",
 615                     res->hr_remoteaddr);
 616                 nv_free(nvout);
 617                 goto close;
 618         }
 619         if (hast_proto_send(res, in, nvout, NULL, 0) < 0) {
 620                 pjdlog_errno(LOG_WARNING,
 621                     "Unable to send handshake header to %s",
 622                     res->hr_remoteaddr);
 623                 nv_free(nvout);
 624                 goto close;
 625         }
 626         nv_free(nvout);
 627         if (hast_proto_recv_hdr(out, &nvin) < 0) {
 628                 pjdlog_errno(LOG_WARNING,
 629                     "Unable to receive handshake header from %s",
 630                     res->hr_remoteaddr);
 631                 goto close;
 632         }
 633         errmsg = nv_get_string(nvin, "errmsg");
 634         if (errmsg != NULL) {
 635                 pjdlog_warning("%s", errmsg);
 636                 nv_free(nvin);
 637                 goto close;
 638         }
 639         datasize = nv_get_int64(nvin, "datasize");
 640         if (datasize != res->hr_datasize) {
 641                 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
 642                     (intmax_t)res->hr_datasize, (intmax_t)datasize);
 643                 nv_free(nvin);
 644                 goto close;
 645         }
 646         extentsize = nv_get_int32(nvin, "extentsize");
 647         if (extentsize != res->hr_extentsize) {
 648                 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
 649                     (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
 650                 nv_free(nvin);
 651                 goto close;
 652         }
 653         res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
 654         res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
 655         res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
 656         map = NULL;
 657         mapsize = nv_get_uint32(nvin, "mapsize");
 658         if (mapsize > 0) {
 659                 map = malloc(mapsize);
 660                 if (map == NULL) {
 661                         pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
 662                             (uintmax_t)mapsize);
 663                         nv_free(nvin);
 664                         goto close;
 665                 }
 666                 /*
 667                  * Remote node have some dirty extents on its own, lets
 668                  * download its activemap.
 669                  */
 670                 if (hast_proto_recv_data(res, out, nvin, map,
 671                     mapsize) < 0) {
 672                         pjdlog_errno(LOG_ERR,
 673                             "Unable to receive remote activemap");
 674                         nv_free(nvin);
 675                         free(map);
 676                         goto close;
 677                 }
 678                 /*
 679                  * Merge local and remote bitmaps.
 680                  */
 681                 activemap_merge(res->hr_amp, map, mapsize);
 682                 free(map);
 683                 /*
 684                  * Now that we merged bitmaps from both nodes, flush it to the
 685                  * disk before we start to synchronize.
 686                  */
 687                 (void)hast_activemap_flush(res);
 688         }
 689         nv_free(nvin);
 690         pjdlog_info("Connected to %s.", res->hr_remoteaddr);
 691         if (inp != NULL && outp != NULL) {
 692                 *inp = in;
 693                 *outp = out;
 694         } else {
 695                 res->hr_remotein = in;
 696                 res->hr_remoteout = out;
 697         }
 698         event_send(res, EVENT_CONNECT);
 699         return (true);
 700 close:
 701         if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
 702                 event_send(res, EVENT_SPLITBRAIN);
 703         proto_close(out);
 704         if (in != NULL)
 705                 proto_close(in);
 706         return (false);
 707 }
 708
 709 static void
 710 sync_start(void)
 711 {
 712
 713         mtx_lock(&sync_lock);
 714         sync_inprogress = true;
 715         mtx_unlock(&sync_lock);
 716         cv_signal(&sync_cond);
 717 }
 718
 719 static void
 720 sync_stop(void)
 721 {
 722
 723         mtx_lock(&sync_lock);
 724         if (sync_inprogress)
 725                 sync_inprogress = false;
 726         mtx_unlock(&sync_lock);
 727 }
 728
 729 static void
 730 init_ggate(struct hast_resource *res)
 731 {
 732         struct g_gate_ctl_create ggiocreate;
 733         struct g_gate_ctl_cancel ggiocancel;
 734
 735         /*
 736          * We communicate with ggate via /dev/ggctl. Open it.
 737          */
 738         res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
 739         if (res->hr_ggatefd < 0)
 740                 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
 741         /*
 742          * Create provider before trying to connect, as connection failure
 743          * is not critical, but may take some time.
 744          */
 745         bzero(&ggiocreate, sizeof(ggiocreate));
 746         ggiocreate.gctl_version = G_GATE_VERSION;
 747         ggiocreate.gctl_mediasize = res->hr_datasize;
 748         ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
 749         ggiocreate.gctl_flags = 0;
 750         ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE;
 751         ggiocreate.gctl_timeout = 0;
 752         ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
 753         snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
 754             res->hr_provname);
 755         if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
 756                 pjdlog_info("Device hast/%s created.", res->hr_provname);
 757                 res->hr_ggateunit = ggiocreate.gctl_unit;
 758                 return;
 759         }
 760         if (errno != EEXIST) {
 761                 primary_exit(EX_OSERR, "Unable to create hast/%s device",
 762                     res->hr_provname);
 763         }
 764         pjdlog_debug(1,
 765             "Device hast/%s already exists, we will try to take it over.",
 766             res->hr_provname);
 767         /*
 768          * If we received EEXIST, we assume that the process who created the
 769          * provider died and didn't clean up. In that case we will start from
 770          * where he left of.
 771          */
 772         bzero(&ggiocancel, sizeof(ggiocancel));
 773         ggiocancel.gctl_version = G_GATE_VERSION;
 774         ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
 775         snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
 776             res->hr_provname);
 777         if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
 778                 pjdlog_info("Device hast/%s recovered.", res->hr_provname);
 779                 res->hr_ggateunit = ggiocancel.gctl_unit;
 780                 return;
 781         }
 782         primary_exit(EX_OSERR, "Unable to take over hast/%s device",
 783             res->hr_provname);
 784 }
 785
 786 void
 787 hastd_primary(struct hast_resource *res)
 788 {
 789         pthread_t td;
 790         pid_t pid;
 791         int error;
 792
 793         /*
 794          * Create communication channel between parent and child.
 795          */
 796         if (proto_client("socketpair://", &res->hr_ctrl) < 0) {
 797                 KEEP_ERRNO((void)pidfile_remove(pfh));
 798                 pjdlog_exit(EX_OSERR,
 799                     "Unable to create control sockets between parent and child");
 800         }
 801         /*
 802          * Create communication channel between child and parent.
 803          */
 804         if (proto_client("socketpair://", &res->hr_event) < 0) {
 805                 KEEP_ERRNO((void)pidfile_remove(pfh));
 806                 pjdlog_exit(EX_OSERR,
 807                     "Unable to create event sockets between child and parent");
 808         }
 809
 810         pid = fork();
 811         if (pid < 0) {
 812                 KEEP_ERRNO((void)pidfile_remove(pfh));
 813                 pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
 814         }
 815
 816         if (pid > 0) {
 817                 /* This is parent. */
 818                 /* Declare that we are receiver. */
 819                 proto_recv(res->hr_event, NULL, 0);
 820                 res->hr_workerpid = pid;
 821                 return;
 822         }
 823
 824         gres = res;
 825
 826         (void)pidfile_close(pfh);
 827         hook_fini();
 828
 829         setproctitle("%s (primary)", res->hr_name);
 830
 831         /* Declare that we are sender. */
 832         proto_send(res->hr_event, NULL, 0);
 833
 834         init_local(res);
 835         init_ggate(res);
 836         init_environment(res);
 837         /*
 838          * Create the guard thread first, so we can handle signals from the
 839          * very begining.
 840          */
 841         error = pthread_create(&td, NULL, guard_thread, res);
 842         assert(error == 0);
 843         /*
 844          * Create the control thread before sending any event to the parent,
 845          * as we can deadlock when parent sends control request to worker,
 846          * but worker has no control thread started yet, so parent waits.
 847          * In the meantime worker sends an event to the parent, but parent
 848          * is unable to handle the event, because it waits for control
 849          * request response.
 850          */
 851         error = pthread_create(&td, NULL, ctrl_thread, res);
 852         assert(error == 0);
 853         if (real_remote(res) && init_remote(res, NULL, NULL))
 854                 sync_start();
 855         error = pthread_create(&td, NULL, ggate_recv_thread, res);
 856         assert(error == 0);
 857         error = pthread_create(&td, NULL, local_send_thread, res);
 858         assert(error == 0);
 859         error = pthread_create(&td, NULL, remote_send_thread, res);
 860         assert(error == 0);
 861         error = pthread_create(&td, NULL, remote_recv_thread, res);
 862         assert(error == 0);
 863         error = pthread_create(&td, NULL, ggate_send_thread, res);
 864         assert(error == 0);
 865         (void)sync_thread(res);
 866 }
 867
 868 static void
 869 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...)
 870 {
 871         char msg[1024];
 872         va_list ap;
 873         int len;
 874
 875         va_start(ap, fmt);
 876         len = vsnprintf(msg, sizeof(msg), fmt, ap);
 877         va_end(ap);
 878         if ((size_t)len < sizeof(msg)) {
 879                 switch (ggio->gctl_cmd) {
 880                 case BIO_READ:
 881                         (void)snprintf(msg + len, sizeof(msg) - len,
 882                             "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
 883                             (uintmax_t)ggio->gctl_length);
 884                         break;
 885                 case BIO_DELETE:
 886                         (void)snprintf(msg + len, sizeof(msg) - len,
 887                             "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
 888                             (uintmax_t)ggio->gctl_length);
 889                         break;
 890                 case BIO_FLUSH:
 891                         (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
 892                         break;
 893                 case BIO_WRITE:
 894                         (void)snprintf(msg + len, sizeof(msg) - len,
 895                             "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
 896                             (uintmax_t)ggio->gctl_length);
 897                         break;
 898                 default:
 899                         (void)snprintf(msg + len, sizeof(msg) - len,
 900                             "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd);
 901                         break;
 902                 }
 903         }
 904         pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
 905 }
 906
 907 static void
 908 remote_close(struct hast_resource *res, int ncomp)
 909 {
 910
 911         rw_wlock(&hio_remote_lock[ncomp]);
 912         /*
 913          * A race is possible between dropping rlock and acquiring wlock -
 914          * another thread can close connection in-between.
 915          */
 916         if (!ISCONNECTED(res, ncomp)) {
 917                 assert(res->hr_remotein == NULL);
 918                 assert(res->hr_remoteout == NULL);
 919                 rw_unlock(&hio_remote_lock[ncomp]);
 920                 return;
 921         }
 922
 923         assert(res->hr_remotein != NULL);
 924         assert(res->hr_remoteout != NULL);
 925
 926         pjdlog_debug(2, "Closing incoming connection to %s.",
 927             res->hr_remoteaddr);
 928         proto_close(res->hr_remotein);
 929         res->hr_remotein = NULL;
 930         pjdlog_debug(2, "Closing outgoing connection to %s.",
 931             res->hr_remoteaddr);
 932         proto_close(res->hr_remoteout);
 933         res->hr_remoteout = NULL;
 934
 935         rw_unlock(&hio_remote_lock[ncomp]);
 936
 937         pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
 938
 939         /*
 940          * Stop synchronization if in-progress.
 941          */
 942         sync_stop();
 943
 944         event_send(res, EVENT_DISCONNECT);
 945 }
 946
 947 /*
 948  * Thread receives ggate I/O requests from the kernel and passes them to
 949  * appropriate threads:
 950  * WRITE - always goes to both local_send and remote_send threads
 951  * READ (when the block is up-to-date on local component) -
 952  *      only local_send thread
 953  * READ (when the block isn't up-to-date on local component) -
 954  *      only remote_send thread
 955  * DELETE - always goes to both local_send and remote_send threads
 956  * FLUSH - always goes to both local_send and remote_send threads
 957  */
 958 static void *
 959 ggate_recv_thread(void *arg)
 960 {
 961         struct hast_resource *res = arg;
 962         struct g_gate_ctl_io *ggio;
 963         struct hio *hio;
 964         unsigned int ii, ncomp, ncomps;
 965         int error;
 966
 967         ncomps = HAST_NCOMPONENTS;
 968
 969         for (;;) {
 970                 pjdlog_debug(2, "ggate_recv: Taking free request.");
 971                 QUEUE_TAKE2(hio, free);
 972                 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
 973                 ggio = &hio->hio_ggio;
 974                 ggio->gctl_unit = res->hr_ggateunit;
 975                 ggio->gctl_length = MAXPHYS;
 976                 ggio->gctl_error = 0;
 977                 pjdlog_debug(2,
 978                     "ggate_recv: (%p) Waiting for request from the kernel.",
 979                     hio);
 980                 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) {
 981                         if (sigexit_received)
 982                                 pthread_exit(NULL);
 983                         primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
 984                 }
 985                 error = ggio->gctl_error;
 986                 switch (error) {
 987                 case 0:
 988                         break;
 989                 case ECANCELED:
 990                         /* Exit gracefully. */
 991                         if (!sigexit_received) {
 992                                 pjdlog_debug(2,
 993                                     "ggate_recv: (%p) Received cancel from the kernel.",
 994                                     hio);
 995                                 pjdlog_info("Received cancel from the kernel, exiting.");
 996                         }
 997                         pthread_exit(NULL);
 998                 case ENOMEM:
 999                         /*
1000                          * Buffer too small? Impossible, we allocate MAXPHYS
1001                          * bytes - request can't be bigger than that.
1002                          */
1003                         /* FALLTHROUGH */
1004                 case ENXIO:
1005                 default:
1006                         primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1007                             strerror(error));
1008                 }
1009                 for (ii = 0; ii < ncomps; ii++)
1010                         hio->hio_errors[ii] = EINVAL;
1011                 reqlog(LOG_DEBUG, 2, ggio,
1012                     "ggate_recv: (%p) Request received from the kernel: ",
1013                     hio);
1014                 /*
1015                  * Inform all components about new write request.
1016                  * For read request prefer local component unless the given
1017                  * range is out-of-date, then use remote component.
1018                  */
1019                 switch (ggio->gctl_cmd) {
1020                 case BIO_READ:
1021                         pjdlog_debug(2,
1022                             "ggate_recv: (%p) Moving request to the send queue.",
1023                             hio);
1024                         refcount_init(&hio->hio_countdown, 1);
1025                         mtx_lock(&metadata_lock);
1026                         if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1027                             res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1028                                 /*
1029                                  * This range is up-to-date on local component,
1030                                  * so handle request locally.
1031                                  */
1032                                  /* Local component is 0 for now. */
1033                                 ncomp = 0;
1034                         } else /* if (res->hr_syncsrc ==
1035                             HAST_SYNCSRC_SECONDARY) */ {
1036                                 assert(res->hr_syncsrc ==
1037                                     HAST_SYNCSRC_SECONDARY);
1038                                 /*
1039                                  * This range is out-of-date on local component,
1040                                  * so send request to the remote node.
1041                                  */
1042                                  /* Remote component is 1 for now. */
1043                                 ncomp = 1;
1044                         }
1045                         mtx_unlock(&metadata_lock);
1046                         QUEUE_INSERT1(hio, send, ncomp);
1047                         break;
1048                 case BIO_WRITE:
1049                         if (res->hr_resuid == 0) {
1050                                 /* This is first write, initialize resuid. */
1051                                 (void)init_resuid(res);
1052                         }
1053                         for (;;) {
1054                                 mtx_lock(&range_lock);
1055                                 if (rangelock_islocked(range_sync,
1056                                     ggio->gctl_offset, ggio->gctl_length)) {
1057                                         pjdlog_debug(2,
1058                                             "regular: Range offset=%jd length=%zu locked.",
1059                                             (intmax_t)ggio->gctl_offset,
1060                                             (size_t)ggio->gctl_length);
1061                                         range_regular_wait = true;
1062                                         cv_wait(&range_regular_cond, &range_lock);
1063                                         range_regular_wait = false;
1064                                         mtx_unlock(&range_lock);
1065                                         continue;
1066                                 }
1067                                 if (rangelock_add(range_regular,
1068                                     ggio->gctl_offset, ggio->gctl_length) < 0) {
1069                                         mtx_unlock(&range_lock);
1070                                         pjdlog_debug(2,
1071                                             "regular: Range offset=%jd length=%zu is already locked, waiting.",
1072                                             (intmax_t)ggio->gctl_offset,
1073                                             (size_t)ggio->gctl_length);
1074                                         sleep(1);
1075                                         continue;
1076                                 }
1077                                 mtx_unlock(&range_lock);
1078                                 break;
1079                         }
1080                         mtx_lock(&res->hr_amp_lock);
1081                         if (activemap_write_start(res->hr_amp,
1082                             ggio->gctl_offset, ggio->gctl_length)) {
1083                                 (void)hast_activemap_flush(res);
1084                         }
1085                         mtx_unlock(&res->hr_amp_lock);
1086                         /* FALLTHROUGH */
1087                 case BIO_DELETE:
1088                 case BIO_FLUSH:
1089                         pjdlog_debug(2,
1090                             "ggate_recv: (%p) Moving request to the send queues.",
1091                             hio);
1092                         refcount_init(&hio->hio_countdown, ncomps);
1093                         for (ii = 0; ii < ncomps; ii++)
1094                                 QUEUE_INSERT1(hio, send, ii);
1095                         break;
1096                 }
1097         }
1098         /* NOTREACHED */
1099         return (NULL);
1100 }
1101
1102 /*
1103  * Thread reads from or writes to local component.
1104  * If local read fails, it redirects it to remote_send thread.
1105  */
1106 static void *
1107 local_send_thread(void *arg)
1108 {
1109         struct hast_resource *res = arg;
1110         struct g_gate_ctl_io *ggio;
1111         struct hio *hio;
1112         unsigned int ncomp, rncomp;
1113         ssize_t ret;
1114
1115         /* Local component is 0 for now. */
1116         ncomp = 0;
1117         /* Remote component is 1 for now. */
1118         rncomp = 1;
1119
1120         for (;;) {
1121                 pjdlog_debug(2, "local_send: Taking request.");
1122                 QUEUE_TAKE1(hio, send, ncomp, 0);
1123                 pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1124                 ggio = &hio->hio_ggio;
1125                 switch (ggio->gctl_cmd) {
1126                 case BIO_READ:
1127                         ret = pread(res->hr_localfd, ggio->gctl_data,
1128                             ggio->gctl_length,
1129                             ggio->gctl_offset + res->hr_localoff);
1130                         if (ret == ggio->gctl_length)
1131                                 hio->hio_errors[ncomp] = 0;
1132                         else {
1133                                 /*
1134                                  * If READ failed, try to read from remote node.
1135                                  */
1136                                 QUEUE_INSERT1(hio, send, rncomp);
1137                                 continue;
1138                         }
1139                         break;
1140                 case BIO_WRITE:
1141                         ret = pwrite(res->hr_localfd, ggio->gctl_data,
1142                             ggio->gctl_length,
1143                             ggio->gctl_offset + res->hr_localoff);
1144                         if (ret < 0)
1145                                 hio->hio_errors[ncomp] = errno;
1146                         else if (ret != ggio->gctl_length)
1147                                 hio->hio_errors[ncomp] = EIO;
1148                         else
1149                                 hio->hio_errors[ncomp] = 0;
1150                         break;
1151                 case BIO_DELETE:
1152                         ret = g_delete(res->hr_localfd,
1153                             ggio->gctl_offset + res->hr_localoff,
1154                             ggio->gctl_length);
1155                         if (ret < 0)
1156                                 hio->hio_errors[ncomp] = errno;
1157                         else
1158                                 hio->hio_errors[ncomp] = 0;
1159                         break;
1160                 case BIO_FLUSH:
1161                         ret = g_flush(res->hr_localfd);
1162                         if (ret < 0)
1163                                 hio->hio_errors[ncomp] = errno;
1164                         else
1165                                 hio->hio_errors[ncomp] = 0;
1166                         break;
1167                 }
1168                 if (refcount_release(&hio->hio_countdown)) {
1169                         if (ISSYNCREQ(hio)) {
1170                                 mtx_lock(&sync_lock);
1171                                 SYNCREQDONE(hio);
1172                                 mtx_unlock(&sync_lock);
1173                                 cv_signal(&sync_cond);
1174                         } else {
1175                                 pjdlog_debug(2,
1176                                     "local_send: (%p) Moving request to the done queue.",
1177                                     hio);
1178                                 QUEUE_INSERT2(hio, done);
1179                         }
1180                 }
1181         }
1182         /* NOTREACHED */
1183         return (NULL);
1184 }
1185
1186 static void
1187 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1188 {
1189         struct nv *nv;
1190
1191         if (!ISCONNECTED(res, ncomp))
1192                 return;
1193
1194         assert(res->hr_remotein != NULL);
1195         assert(res->hr_remoteout != NULL);
1196
1197         nv = nv_alloc();
1198         nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1199         if (nv_error(nv) != 0) {
1200                 nv_free(nv);
1201                 pjdlog_debug(1,
1202                     "keepalive_send: Unable to prepare header to send.");
1203                 return;
1204         }
1205         if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) {
1206                 pjdlog_common(LOG_DEBUG, 1, errno,
1207                     "keepalive_send: Unable to send request");
1208                 nv_free(nv);
1209                 rw_unlock(&hio_remote_lock[ncomp]);
1210                 remote_close(res, ncomp);
1211                 rw_rlock(&hio_remote_lock[ncomp]);
1212                 return;
1213         }
1214         nv_free(nv);
1215         pjdlog_debug(2, "keepalive_send: Request sent.");
1216 }
1217
1218 /*
1219  * Thread sends request to secondary node.
1220  */
1221 static void *
1222 remote_send_thread(void *arg)
1223 {
1224         struct hast_resource *res = arg;
1225         struct g_gate_ctl_io *ggio;
1226         time_t lastcheck, now;
1227         struct hio *hio;
1228         struct nv *nv;
1229         unsigned int ncomp;
1230         bool wakeup;
1231         uint64_t offset, length;
1232         uint8_t cmd;
1233         void *data;
1234
1235         /* Remote component is 1 for now. */
1236         ncomp = 1;
1237         lastcheck = time(NULL);
1238
1239         for (;;) {
1240                 pjdlog_debug(2, "remote_send: Taking request.");
1241                 QUEUE_TAKE1(hio, send, ncomp, RETRY_SLEEP);
1242                 if (hio == NULL) {
1243                         now = time(NULL);
1244                         if (lastcheck + RETRY_SLEEP <= now) {
1245                                 keepalive_send(res, ncomp);
1246                                 lastcheck = now;
1247                         }
1248                         continue;
1249                 }
1250                 pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1251                 ggio = &hio->hio_ggio;
1252                 switch (ggio->gctl_cmd) {
1253                 case BIO_READ:
1254                         cmd = HIO_READ;
1255                         data = NULL;
1256                         offset = ggio->gctl_offset;
1257                         length = ggio->gctl_length;
1258                         break;
1259                 case BIO_WRITE:
1260                         cmd = HIO_WRITE;
1261                         data = ggio->gctl_data;
1262                         offset = ggio->gctl_offset;
1263                         length = ggio->gctl_length;
1264                         break;
1265                 case BIO_DELETE:
1266                         cmd = HIO_DELETE;
1267                         data = NULL;
1268                         offset = ggio->gctl_offset;
1269                         length = ggio->gctl_length;
1270                         break;
1271                 case BIO_FLUSH:
1272                         cmd = HIO_FLUSH;
1273                         data = NULL;
1274                         offset = 0;
1275                         length = 0;
1276                         break;
1277                 default:
1278                         assert(!"invalid condition");
1279                         abort();
1280                 }
1281                 nv = nv_alloc();
1282                 nv_add_uint8(nv, cmd, "cmd");
1283                 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1284                 nv_add_uint64(nv, offset, "offset");
1285                 nv_add_uint64(nv, length, "length");
1286                 if (nv_error(nv) != 0) {
1287                         hio->hio_errors[ncomp] = nv_error(nv);
1288                         pjdlog_debug(2,
1289                             "remote_send: (%p) Unable to prepare header to send.",
1290                             hio);
1291                         reqlog(LOG_ERR, 0, ggio,
1292                             "Unable to prepare header to send (%s): ",
1293                             strerror(nv_error(nv)));
1294                         /* Move failed request immediately to the done queue. */
1295                         goto done_queue;
1296                 }
1297                 pjdlog_debug(2,
1298                     "remote_send: (%p) Moving request to the recv queue.",
1299                     hio);
1300                 /*
1301                  * Protect connection from disappearing.
1302                  */
1303                 rw_rlock(&hio_remote_lock[ncomp]);
1304                 if (!ISCONNECTED(res, ncomp)) {
1305                         rw_unlock(&hio_remote_lock[ncomp]);
1306                         hio->hio_errors[ncomp] = ENOTCONN;
1307                         goto done_queue;
1308                 }
1309                 /*
1310                  * Move the request to recv queue before sending it, because
1311                  * in different order we can get reply before we move request
1312                  * to recv queue.
1313                  */
1314                 mtx_lock(&hio_recv_list_lock[ncomp]);
1315                 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1316                 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1317                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1318                 if (hast_proto_send(res, res->hr_remoteout, nv, data,
1319                     data != NULL ? length : 0) < 0) {
1320                         hio->hio_errors[ncomp] = errno;
1321                         rw_unlock(&hio_remote_lock[ncomp]);
1322                         pjdlog_debug(2,
1323                             "remote_send: (%p) Unable to send request.", hio);
1324                         reqlog(LOG_ERR, 0, ggio,
1325                             "Unable to send request (%s): ",
1326                             strerror(hio->hio_errors[ncomp]));
1327                         remote_close(res, ncomp);
1328                         /*
1329                          * Take request back from the receive queue and move
1330                          * it immediately to the done queue.
1331                          */
1332                         mtx_lock(&hio_recv_list_lock[ncomp]);
1333                         TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1334                         mtx_unlock(&hio_recv_list_lock[ncomp]);
1335                         goto done_queue;
1336                 }
1337                 rw_unlock(&hio_remote_lock[ncomp]);
1338                 nv_free(nv);
1339                 if (wakeup)
1340                         cv_signal(&hio_recv_list_cond[ncomp]);
1341                 continue;
1342 done_queue:
1343                 nv_free(nv);
1344                 if (ISSYNCREQ(hio)) {
1345                         if (!refcount_release(&hio->hio_countdown))
1346                                 continue;
1347                         mtx_lock(&sync_lock);
1348                         SYNCREQDONE(hio);
1349                         mtx_unlock(&sync_lock);
1350                         cv_signal(&sync_cond);
1351                         continue;
1352                 }
1353                 if (ggio->gctl_cmd == BIO_WRITE) {
1354                         mtx_lock(&res->hr_amp_lock);
1355                         if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1356                             ggio->gctl_length)) {
1357                                 (void)hast_activemap_flush(res);
1358                         }
1359                         mtx_unlock(&res->hr_amp_lock);
1360                 }
1361                 if (!refcount_release(&hio->hio_countdown))
1362                         continue;
1363                 pjdlog_debug(2,
1364                     "remote_send: (%p) Moving request to the done queue.",
1365                     hio);
1366                 QUEUE_INSERT2(hio, done);
1367         }
1368         /* NOTREACHED */
1369         return (NULL);
1370 }
1371
1372 /*
1373  * Thread receives answer from secondary node and passes it to ggate_send
1374  * thread.
1375  */
1376 static void *
1377 remote_recv_thread(void *arg)
1378 {
1379         struct hast_resource *res = arg;
1380         struct g_gate_ctl_io *ggio;
1381         struct hio *hio;
1382         struct nv *nv;
1383         unsigned int ncomp;
1384         uint64_t seq;
1385         int error;
1386
1387         /* Remote component is 1 for now. */
1388         ncomp = 1;
1389
1390         for (;;) {
1391                 /* Wait until there is anything to receive. */
1392                 mtx_lock(&hio_recv_list_lock[ncomp]);
1393                 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1394                         pjdlog_debug(2, "remote_recv: No requests, waiting.");
1395                         cv_wait(&hio_recv_list_cond[ncomp],
1396                             &hio_recv_list_lock[ncomp]);
1397                 }
1398                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1399                 rw_rlock(&hio_remote_lock[ncomp]);
1400                 if (!ISCONNECTED(res, ncomp)) {
1401                         rw_unlock(&hio_remote_lock[ncomp]);
1402                         /*
1403                          * Connection is dead, so move all pending requests to
1404                          * the done queue (one-by-one).
1405                          */
1406                         mtx_lock(&hio_recv_list_lock[ncomp]);
1407                         hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1408                         assert(hio != NULL);
1409                         TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1410                             hio_next[ncomp]);
1411                         mtx_unlock(&hio_recv_list_lock[ncomp]);
1412                         goto done_queue;
1413                 }
1414                 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) {
1415                         pjdlog_errno(LOG_ERR,
1416                             "Unable to receive reply header");
1417                         rw_unlock(&hio_remote_lock[ncomp]);
1418                         remote_close(res, ncomp);
1419                         continue;
1420                 }
1421                 rw_unlock(&hio_remote_lock[ncomp]);
1422                 seq = nv_get_uint64(nv, "seq");
1423                 if (seq == 0) {
1424                         pjdlog_error("Header contains no 'seq' field.");
1425                         nv_free(nv);
1426                         continue;
1427                 }
1428                 mtx_lock(&hio_recv_list_lock[ncomp]);
1429                 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1430                         if (hio->hio_ggio.gctl_seq == seq) {
1431                                 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1432                                     hio_next[ncomp]);
1433                                 break;
1434                         }
1435                 }
1436                 mtx_unlock(&hio_recv_list_lock[ncomp]);
1437                 if (hio == NULL) {
1438                         pjdlog_error("Found no request matching received 'seq' field (%ju).",
1439                             (uintmax_t)seq);
1440                         nv_free(nv);
1441                         continue;
1442                 }
1443                 error = nv_get_int16(nv, "error");
1444                 if (error != 0) {
1445                         /* Request failed on remote side. */
1446                         hio->hio_errors[ncomp] = 0;
1447                         nv_free(nv);
1448                         goto done_queue;
1449                 }
1450                 ggio = &hio->hio_ggio;
1451                 switch (ggio->gctl_cmd) {
1452                 case BIO_READ:
1453                         rw_rlock(&hio_remote_lock[ncomp]);
1454                         if (!ISCONNECTED(res, ncomp)) {
1455                                 rw_unlock(&hio_remote_lock[ncomp]);
1456                                 nv_free(nv);
1457                                 goto done_queue;
1458                         }
1459                         if (hast_proto_recv_data(res, res->hr_remotein, nv,
1460                             ggio->gctl_data, ggio->gctl_length) < 0) {
1461                                 hio->hio_errors[ncomp] = errno;
1462                                 pjdlog_errno(LOG_ERR,
1463                                     "Unable to receive reply data");
1464                                 rw_unlock(&hio_remote_lock[ncomp]);
1465                                 nv_free(nv);
1466                                 remote_close(res, ncomp);
1467                                 goto done_queue;
1468                         }
1469                         rw_unlock(&hio_remote_lock[ncomp]);
1470                         break;
1471                 case BIO_WRITE:
1472                 case BIO_DELETE:
1473                 case BIO_FLUSH:
1474                         break;
1475                 default:
1476                         assert(!"invalid condition");
1477                         abort();
1478                 }
1479                 hio->hio_errors[ncomp] = 0;
1480                 nv_free(nv);
1481 done_queue:
1482                 if (refcount_release(&hio->hio_countdown)) {
1483                         if (ISSYNCREQ(hio)) {
1484                                 mtx_lock(&sync_lock);
1485                                 SYNCREQDONE(hio);
1486                                 mtx_unlock(&sync_lock);
1487                                 cv_signal(&sync_cond);
1488                         } else {
1489                                 pjdlog_debug(2,
1490                                     "remote_recv: (%p) Moving request to the done queue.",
1491                                     hio);
1492                                 QUEUE_INSERT2(hio, done);
1493                         }
1494                 }
1495         }
1496         /* NOTREACHED */
1497         return (NULL);
1498 }
1499
1500 /*
1501  * Thread sends answer to the kernel.
1502  */
1503 static void *
1504 ggate_send_thread(void *arg)
1505 {
1506         struct hast_resource *res = arg;
1507         struct g_gate_ctl_io *ggio;
1508         struct hio *hio;
1509         unsigned int ii, ncomp, ncomps;
1510
1511         ncomps = HAST_NCOMPONENTS;
1512
1513         for (;;) {
1514                 pjdlog_debug(2, "ggate_send: Taking request.");
1515                 QUEUE_TAKE2(hio, done);
1516                 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1517                 ggio = &hio->hio_ggio;
1518                 for (ii = 0; ii < ncomps; ii++) {
1519                         if (hio->hio_errors[ii] == 0) {
1520                                 /*
1521                                  * One successful request is enough to declare
1522                                  * success.
1523                                  */
1524                                 ggio->gctl_error = 0;
1525                                 break;
1526                         }
1527                 }
1528                 if (ii == ncomps) {
1529                         /*
1530                          * None of the requests were successful.
1531                          * Use first error.
1532                          */
1533                         ggio->gctl_error = hio->hio_errors[0];
1534                 }
1535                 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1536                         mtx_lock(&res->hr_amp_lock);
1537                         activemap_write_complete(res->hr_amp,
1538                             ggio->gctl_offset, ggio->gctl_length);
1539                         mtx_unlock(&res->hr_amp_lock);
1540                 }
1541                 if (ggio->gctl_cmd == BIO_WRITE) {
1542                         /*
1543                          * Unlock range we locked.
1544                          */
1545                         mtx_lock(&range_lock);
1546                         rangelock_del(range_regular, ggio->gctl_offset,
1547                             ggio->gctl_length);
1548                         if (range_sync_wait)
1549                                 cv_signal(&range_sync_cond);
1550                         mtx_unlock(&range_lock);
1551                         /*
1552                          * Bump local count if this is first write after
1553                          * connection failure with remote node.
1554                          */
1555                         ncomp = 1;
1556                         rw_rlock(&hio_remote_lock[ncomp]);
1557                         if (!ISCONNECTED(res, ncomp)) {
1558                                 mtx_lock(&metadata_lock);
1559                                 if (res->hr_primary_localcnt ==
1560                                     res->hr_secondary_remotecnt) {
1561                                         res->hr_primary_localcnt++;
1562                                         pjdlog_debug(1,
1563                                             "Increasing localcnt to %ju.",
1564                                             (uintmax_t)res->hr_primary_localcnt);
1565                                         (void)metadata_write(res);
1566                                 }
1567                                 mtx_unlock(&metadata_lock);
1568                         }
1569                         rw_unlock(&hio_remote_lock[ncomp]);
1570                 }
1571                 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0)
1572                         primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1573                 pjdlog_debug(2,
1574                     "ggate_send: (%p) Moving request to the free queue.", hio);
1575                 QUEUE_INSERT2(hio, free);
1576         }
1577         /* NOTREACHED */
1578         return (NULL);
1579 }
1580
1581 /*
1582  * Thread synchronize local and remote components.
1583  */
1584 static void *
1585 sync_thread(void *arg __unused)
1586 {
1587         struct hast_resource *res = arg;
1588         struct hio *hio;
1589         struct g_gate_ctl_io *ggio;
1590         unsigned int ii, ncomp, ncomps;
1591         off_t offset, length, synced;
1592         bool dorewind;
1593         int syncext;
1594
1595         ncomps = HAST_NCOMPONENTS;
1596         dorewind = true;
1597         synced = 0;
1598         offset = -1;
1599
1600         for (;;) {
1601                 mtx_lock(&sync_lock);
1602                 if (offset >= 0 && !sync_inprogress) {
1603                         pjdlog_info("Synchronization interrupted. "
1604                             "%jd bytes synchronized so far.",
1605                             (intmax_t)synced);
1606                         event_send(res, EVENT_SYNCINTR);
1607                 }
1608                 while (!sync_inprogress) {
1609                         dorewind = true;
1610                         synced = 0;
1611                         cv_wait(&sync_cond, &sync_lock);
1612                 }
1613                 mtx_unlock(&sync_lock);
1614                 /*
1615                  * Obtain offset at which we should synchronize.
1616                  * Rewind synchronization if needed.
1617                  */
1618                 mtx_lock(&res->hr_amp_lock);
1619                 if (dorewind)
1620                         activemap_sync_rewind(res->hr_amp);
1621                 offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1622                 if (syncext != -1) {
1623                         /*
1624                          * We synchronized entire syncext extent, we can mark
1625                          * it as clean now.
1626                          */
1627                         if (activemap_extent_complete(res->hr_amp, syncext))
1628                                 (void)hast_activemap_flush(res);
1629                 }
1630                 mtx_unlock(&res->hr_amp_lock);
1631                 if (dorewind) {
1632                         dorewind = false;
1633                         if (offset < 0)
1634                                 pjdlog_info("Nodes are in sync.");
1635                         else {
1636                                 pjdlog_info("Synchronization started. %ju bytes to go.",
1637                                     (uintmax_t)(res->hr_extentsize *
1638                                     activemap_ndirty(res->hr_amp)));
1639                                 event_send(res, EVENT_SYNCSTART);
1640                         }
1641                 }
1642                 if (offset < 0) {
1643                         sync_stop();
1644                         pjdlog_debug(1, "Nothing to synchronize.");
1645                         /*
1646                          * Synchronization complete, make both localcnt and
1647                          * remotecnt equal.
1648                          */
1649                         ncomp = 1;
1650                         rw_rlock(&hio_remote_lock[ncomp]);
1651                         if (ISCONNECTED(res, ncomp)) {
1652                                 if (synced > 0) {
1653                                         pjdlog_info("Synchronization complete. "
1654                                             "%jd bytes synchronized.",
1655                                             (intmax_t)synced);
1656                                         event_send(res, EVENT_SYNCDONE);
1657                                 }
1658                                 mtx_lock(&metadata_lock);
1659                                 res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
1660                                 res->hr_primary_localcnt =
1661                                     res->hr_secondary_localcnt;
1662                                 res->hr_primary_remotecnt =
1663                                     res->hr_secondary_remotecnt;
1664                                 pjdlog_debug(1,
1665                                     "Setting localcnt to %ju and remotecnt to %ju.",
1666                                     (uintmax_t)res->hr_primary_localcnt,
1667                                     (uintmax_t)res->hr_secondary_localcnt);
1668                                 (void)metadata_write(res);
1669                                 mtx_unlock(&metadata_lock);
1670                         }
1671                         rw_unlock(&hio_remote_lock[ncomp]);
1672                         continue;
1673                 }
1674                 pjdlog_debug(2, "sync: Taking free request.");
1675                 QUEUE_TAKE2(hio, free);
1676                 pjdlog_debug(2, "sync: (%p) Got free request.", hio);
1677                 /*
1678                  * Lock the range we are going to synchronize. We don't want
1679                  * race where someone writes between our read and write.
1680                  */
1681                 for (;;) {
1682                         mtx_lock(&range_lock);
1683                         if (rangelock_islocked(range_regular, offset, length)) {
1684                                 pjdlog_debug(2,
1685                                     "sync: Range offset=%jd length=%jd locked.",
1686                                     (intmax_t)offset, (intmax_t)length);
1687                                 range_sync_wait = true;
1688                                 cv_wait(&range_sync_cond, &range_lock);
1689                                 range_sync_wait = false;
1690                                 mtx_unlock(&range_lock);
1691                                 continue;
1692                         }
1693                         if (rangelock_add(range_sync, offset, length) < 0) {
1694                                 mtx_unlock(&range_lock);
1695                                 pjdlog_debug(2,
1696                                     "sync: Range offset=%jd length=%jd is already locked, waiting.",
1697                                     (intmax_t)offset, (intmax_t)length);
1698                                 sleep(1);
1699                                 continue;
1700                         }
1701                         mtx_unlock(&range_lock);
1702                         break;
1703                 }
1704                 /*
1705                  * First read the data from synchronization source.
1706                  */
1707                 SYNCREQ(hio);
1708                 ggio = &hio->hio_ggio;
1709                 ggio->gctl_cmd = BIO_READ;
1710                 ggio->gctl_offset = offset;
1711                 ggio->gctl_length = length;
1712                 ggio->gctl_error = 0;
1713                 for (ii = 0; ii < ncomps; ii++)
1714                         hio->hio_errors[ii] = EINVAL;
1715                 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1716                     hio);
1717                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1718                     hio);
1719                 mtx_lock(&metadata_lock);
1720                 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1721                         /*
1722                          * This range is up-to-date on local component,
1723                          * so handle request locally.
1724                          */
1725                          /* Local component is 0 for now. */
1726                         ncomp = 0;
1727                 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1728                         assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1729                         /*
1730                          * This range is out-of-date on local component,
1731                          * so send request to the remote node.
1732                          */
1733                          /* Remote component is 1 for now. */
1734                         ncomp = 1;
1735                 }
1736                 mtx_unlock(&metadata_lock);
1737                 refcount_init(&hio->hio_countdown, 1);
1738                 QUEUE_INSERT1(hio, send, ncomp);
1739
1740                 /*
1741                  * Let's wait for READ to finish.
1742                  */
1743                 mtx_lock(&sync_lock);
1744                 while (!ISSYNCREQDONE(hio))
1745                         cv_wait(&sync_cond, &sync_lock);
1746                 mtx_unlock(&sync_lock);
1747
1748                 if (hio->hio_errors[ncomp] != 0) {
1749                         pjdlog_error("Unable to read synchronization data: %s.",
1750                             strerror(hio->hio_errors[ncomp]));
1751                         goto free_queue;
1752                 }
1753
1754                 /*
1755                  * We read the data from synchronization source, now write it
1756                  * to synchronization target.
1757                  */
1758                 SYNCREQ(hio);
1759                 ggio->gctl_cmd = BIO_WRITE;
1760                 for (ii = 0; ii < ncomps; ii++)
1761                         hio->hio_errors[ii] = EINVAL;
1762                 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1763                     hio);
1764                 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1765                     hio);
1766                 mtx_lock(&metadata_lock);
1767                 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1768                         /*
1769                          * This range is up-to-date on local component,
1770                          * so we update remote component.
1771                          */
1772                          /* Remote component is 1 for now. */
1773                         ncomp = 1;
1774                 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1775                         assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1776                         /*
1777                          * This range is out-of-date on local component,
1778                          * so we update it.
1779                          */
1780                          /* Local component is 0 for now. */
1781                         ncomp = 0;
1782                 }
1783                 mtx_unlock(&metadata_lock);
1784
1785                 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.",
1786                     hio);
1787                 refcount_init(&hio->hio_countdown, 1);
1788                 QUEUE_INSERT1(hio, send, ncomp);
1789
1790                 /*
1791                  * Let's wait for WRITE to finish.
1792                  */
1793                 mtx_lock(&sync_lock);
1794                 while (!ISSYNCREQDONE(hio))
1795                         cv_wait(&sync_cond, &sync_lock);
1796                 mtx_unlock(&sync_lock);
1797
1798                 if (hio->hio_errors[ncomp] != 0) {
1799                         pjdlog_error("Unable to write synchronization data: %s.",
1800                             strerror(hio->hio_errors[ncomp]));
1801                         goto free_queue;
1802                 }
1803
1804                 synced += length;
1805 free_queue:
1806                 mtx_lock(&range_lock);
1807                 rangelock_del(range_sync, offset, length);
1808                 if (range_regular_wait)
1809                         cv_signal(&range_regular_cond);
1810                 mtx_unlock(&range_lock);
1811                 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
1812                     hio);
1813                 QUEUE_INSERT2(hio, free);
1814         }
1815         /* NOTREACHED */
1816         return (NULL);
1817 }
1818
1819 static void
1820 config_reload(void)
1821 {
1822         struct hastd_config *newcfg;
1823         struct hast_resource *res;
1824         unsigned int ii, ncomps;
1825         int modified;
1826
1827         pjdlog_info("Reloading configuration...");
1828
1829         ncomps = HAST_NCOMPONENTS;
1830
1831         newcfg = yy_config_parse(cfgpath, false);
1832         if (newcfg == NULL)
1833                 goto failed;
1834
1835         TAILQ_FOREACH(res, &newcfg->hc_resources, hr_next) {
1836                 if (strcmp(res->hr_name, gres->hr_name) == 0)
1837                         break;
1838         }
1839         /*
1840          * If resource was removed from the configuration file, resource
1841          * name, provider name or path to local component was modified we
1842          * shouldn't be here. This means that someone modified configuration
1843          * file and send SIGHUP to us instead of main hastd process.
1844          * Log advice and ignore the signal.
1845          */
1846         if (res == NULL || strcmp(gres->hr_name, res->hr_name) != 0 ||
1847             strcmp(gres->hr_provname, res->hr_provname) != 0 ||
1848             strcmp(gres->hr_localpath, res->hr_localpath) != 0) {
1849                 pjdlog_warning("To reload configuration send SIGHUP to the main hastd process (pid %u).",
1850                     (unsigned int)getppid());
1851                 goto failed;
1852         }
1853
1854 #define MODIFIED_REMOTEADDR     0x1
1855 #define MODIFIED_REPLICATION    0x2
1856 #define MODIFIED_TIMEOUT        0x4
1857 #define MODIFIED_EXEC           0x8
1858         modified = 0;
1859         if (strcmp(gres->hr_remoteaddr, res->hr_remoteaddr) != 0) {
1860                 /*
1861                  * Don't copy res->hr_remoteaddr to gres just yet.
1862                  * We want remote_close() to log disconnect from the old
1863                  * addresses, not from the new ones.
1864                  */
1865                 modified |= MODIFIED_REMOTEADDR;
1866         }
1867         if (gres->hr_replication != res->hr_replication) {
1868                 gres->hr_replication = res->hr_replication;
1869                 modified |= MODIFIED_REPLICATION;
1870         }
1871         if (gres->hr_timeout != res->hr_timeout) {
1872                 gres->hr_timeout = res->hr_timeout;
1873                 modified |= MODIFIED_TIMEOUT;
1874         }
1875         if (strcmp(gres->hr_exec, res->hr_exec) != 0) {
1876                 strlcpy(gres->hr_exec, res->hr_exec, sizeof(gres->hr_exec));
1877                 modified |= MODIFIED_EXEC;
1878         }
1879         /*
1880          * If only timeout was modified we only need to change it without
1881          * reconnecting.
1882          */
1883         if (modified == MODIFIED_TIMEOUT) {
1884                 for (ii = 0; ii < ncomps; ii++) {
1885                         if (!ISREMOTE(ii))
1886                                 continue;
1887                         rw_rlock(&hio_remote_lock[ii]);
1888                         if (!ISCONNECTED(gres, ii)) {
1889                                 rw_unlock(&hio_remote_lock[ii]);
1890                                 continue;
1891                         }
1892                         rw_unlock(&hio_remote_lock[ii]);
1893                         if (proto_timeout(gres->hr_remotein,
1894                             gres->hr_timeout) < 0) {
1895                                 pjdlog_errno(LOG_WARNING,
1896                                     "Unable to set connection timeout");
1897                         }
1898                         if (proto_timeout(gres->hr_remoteout,
1899                             gres->hr_timeout) < 0) {
1900                                 pjdlog_errno(LOG_WARNING,
1901                                     "Unable to set connection timeout");
1902                         }
1903                 }
1904         } else if ((modified &
1905             (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) {
1906                 for (ii = 0; ii < ncomps; ii++) {
1907                         if (!ISREMOTE(ii))
1908                                 continue;
1909                         remote_close(gres, ii);
1910                 }
1911                 if (modified & MODIFIED_REMOTEADDR) {
1912                         strlcpy(gres->hr_remoteaddr, res->hr_remoteaddr,
1913                             sizeof(gres->hr_remoteaddr));
1914                 }
1915         }
1916 #undef  MODIFIED_REMOTEADDR
1917 #undef  MODIFIED_REPLICATION
1918 #undef  MODIFIED_TIMEOUT
1919 #undef  MODIFIED_EXEC
1920
1921         pjdlog_info("Configuration reloaded successfully.");
1922         return;
1923 failed:
1924         if (newcfg != NULL) {
1925                 if (newcfg->hc_controlconn != NULL)
1926                         proto_close(newcfg->hc_controlconn);
1927                 if (newcfg->hc_listenconn != NULL)
1928                         proto_close(newcfg->hc_listenconn);
1929                 yy_config_free(newcfg);
1930         }
1931         pjdlog_warning("Configuration not reloaded.");
1932 }
1933
1934 static void
1935 guard_one(struct hast_resource *res, unsigned int ncomp)
1936 {
1937         struct proto_conn *in, *out;
1938
1939         if (!ISREMOTE(ncomp))
1940                 return;
1941
1942         rw_rlock(&hio_remote_lock[ncomp]);
1943
1944         if (!real_remote(res)) {
1945                 rw_unlock(&hio_remote_lock[ncomp]);
1946                 return;
1947         }
1948
1949         if (ISCONNECTED(res, ncomp)) {
1950                 assert(res->hr_remotein != NULL);
1951                 assert(res->hr_remoteout != NULL);
1952                 rw_unlock(&hio_remote_lock[ncomp]);
1953                 pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
1954                     res->hr_remoteaddr);
1955                 return;
1956         }
1957
1958         assert(res->hr_remotein == NULL);
1959         assert(res->hr_remoteout == NULL);
1960         /*
1961          * Upgrade the lock. It doesn't have to be atomic as no other thread
1962          * can change connection status from disconnected to connected.
1963          */
1964         rw_unlock(&hio_remote_lock[ncomp]);
1965         pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
1966             res->hr_remoteaddr);
1967         in = out = NULL;
1968         if (init_remote(res, &in, &out)) {
1969                 rw_wlock(&hio_remote_lock[ncomp]);
1970                 assert(res->hr_remotein == NULL);
1971                 assert(res->hr_remoteout == NULL);
1972                 assert(in != NULL && out != NULL);
1973                 res->hr_remotein = in;
1974                 res->hr_remoteout = out;
1975                 rw_unlock(&hio_remote_lock[ncomp]);
1976                 pjdlog_info("Successfully reconnected to %s.",
1977                     res->hr_remoteaddr);
1978                 sync_start();
1979         } else {
1980                 /* Both connections should be NULL. */
1981                 assert(res->hr_remotein == NULL);
1982                 assert(res->hr_remoteout == NULL);
1983                 assert(in == NULL && out == NULL);
1984                 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
1985                     res->hr_remoteaddr);
1986         }
1987 }
1988
1989 /*
1990  * Thread guards remote connections and reconnects when needed, handles
1991  * signals, etc.
1992  */
1993 static void *
1994 guard_thread(void *arg)
1995 {
1996         struct hast_resource *res = arg;
1997         unsigned int ii, ncomps;
1998         struct timespec timeout;
1999         time_t lastcheck, now;
2000         sigset_t mask;
2001         int signo;
2002
2003         ncomps = HAST_NCOMPONENTS;
2004         lastcheck = time(NULL);
2005
2006         PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2007         PJDLOG_VERIFY(sigaddset(&mask, SIGHUP) == 0);
2008         PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2009         PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2010
2011         timeout.tv_nsec = 0;
2012         signo = -1;
2013
2014         for (;;) {
2015                 switch (signo) {
2016                 case SIGHUP:
2017                         config_reload();
2018                         break;
2019                 case SIGINT:
2020                 case SIGTERM:
2021                         sigexit_received = true;
2022                         primary_exitx(EX_OK,
2023                             "Termination signal received, exiting.");
2024                         break;
2025                 default:
2026                         break;
2027                 }
2028
2029                 pjdlog_debug(2, "remote_guard: Checking connections.");
2030                 now = time(NULL);
2031                 if (lastcheck + RETRY_SLEEP <= now) {
2032                         for (ii = 0; ii < ncomps; ii++)
2033                                 guard_one(res, ii);
2034                         lastcheck = now;
2035                 }
2036                 timeout.tv_sec = RETRY_SLEEP;
2037                 signo = sigtimedwait(&mask, NULL, &timeout);
2038         }
2039         /* NOTREACHED */
2040         return (NULL);
2041 }