xref: /freebsd/sbin/hastd/primary.c (revision 405c3050f102b8c74782f0366c8ead927bd07b68)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2009 The FreeBSD Foundation
5  * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net>
6  * All rights reserved.
7  *
8  * This software was developed by Pawel Jakub Dawidek under sponsorship from
9  * the FreeBSD Foundation.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35 
36 #include <sys/types.h>
37 #include <sys/time.h>
38 #include <sys/bio.h>
39 #include <sys/disk.h>
40 #include <sys/stat.h>
41 
42 #include <geom/gate/g_gate.h>
43 
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 "refcnt.h"
70 #include "subr.h"
71 #include "synch.h"
72 
73 /* The is only one remote component for now. */
74 #define	ISREMOTE(no)	((no) == 1)
75 
76 struct hio {
77 	/*
78 	 * Number of components we are still waiting for.
79 	 * When this field goes to 0, we can send the request back to the
80 	 * kernel. Each component has to decrease this counter by one
81 	 * even on failure.
82 	 */
83 	refcnt_t		 hio_countdown;
84 	/*
85 	 * Each component has a place to store its own error.
86 	 * Once the request is handled by all components we can decide if the
87 	 * request overall is successful or not.
88 	 */
89 	int			*hio_errors;
90 	/*
91 	 * Structure used to communicate with GEOM Gate class.
92 	 */
93 	struct g_gate_ctl_io	 hio_ggio;
94 	/*
95 	 * Request was already confirmed to GEOM Gate.
96 	 */
97 	bool			 hio_done;
98 	/*
99 	 * Number of components we are still waiting before sending write
100 	 * completion ack to GEOM Gate. Used for memsync.
101 	 */
102 	refcnt_t		 hio_writecount;
103 	/*
104 	 * Memsync request was acknowleged by remote.
105 	 */
106 	bool			 hio_memsyncacked;
107 	/*
108 	 * Remember replication from the time the request was initiated,
109 	 * so we won't get confused when replication changes on reload.
110 	 */
111 	int			 hio_replication;
112 	TAILQ_ENTRY(hio)	*hio_next;
113 };
114 #define	hio_free_next	hio_next[0]
115 #define	hio_done_next	hio_next[0]
116 
117 /*
118  * Free list holds unused structures. When free list is empty, we have to wait
119  * until some in-progress requests are freed.
120  */
121 static TAILQ_HEAD(, hio) hio_free_list;
122 static size_t hio_free_list_size;
123 static pthread_mutex_t hio_free_list_lock;
124 static pthread_cond_t hio_free_list_cond;
125 /*
126  * There is one send list for every component. One requests is placed on all
127  * send lists - each component gets the same request, but each component is
128  * responsible for managing his own send list.
129  */
130 static TAILQ_HEAD(, hio) *hio_send_list;
131 static size_t *hio_send_list_size;
132 static pthread_mutex_t *hio_send_list_lock;
133 static pthread_cond_t *hio_send_list_cond;
134 #define	hio_send_local_list_size	hio_send_list_size[0]
135 #define	hio_send_remote_list_size	hio_send_list_size[1]
136 /*
137  * There is one recv list for every component, although local components don't
138  * use recv lists as local requests are done synchronously.
139  */
140 static TAILQ_HEAD(, hio) *hio_recv_list;
141 static size_t *hio_recv_list_size;
142 static pthread_mutex_t *hio_recv_list_lock;
143 static pthread_cond_t *hio_recv_list_cond;
144 #define	hio_recv_remote_list_size	hio_recv_list_size[1]
145 /*
146  * Request is placed on done list by the slowest component (the one that
147  * decreased hio_countdown from 1 to 0).
148  */
149 static TAILQ_HEAD(, hio) hio_done_list;
150 static size_t hio_done_list_size;
151 static pthread_mutex_t hio_done_list_lock;
152 static pthread_cond_t hio_done_list_cond;
153 /*
154  * Structure below are for interaction with sync thread.
155  */
156 static bool sync_inprogress;
157 static pthread_mutex_t sync_lock;
158 static pthread_cond_t sync_cond;
159 /*
160  * The lock below allows to synchornize access to remote connections.
161  */
162 static pthread_rwlock_t *hio_remote_lock;
163 
164 /*
165  * Lock to synchronize metadata updates. Also synchronize access to
166  * hr_primary_localcnt and hr_primary_remotecnt fields.
167  */
168 static pthread_mutex_t metadata_lock;
169 
170 /*
171  * Maximum number of outstanding I/O requests.
172  */
173 #define	HAST_HIO_MAX	256
174 /*
175  * Number of components. At this point there are only two components: local
176  * and remote, but in the future it might be possible to use multiple local
177  * and remote components.
178  */
179 #define	HAST_NCOMPONENTS	2
180 
181 #define	ISCONNECTED(res, no)	\
182 	((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
183 
184 #define	QUEUE_INSERT1(hio, name, ncomp)	do {				\
185 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
186 	if (TAILQ_EMPTY(&hio_##name##_list[(ncomp)]))			\
187 		cv_broadcast(&hio_##name##_list_cond[(ncomp)]);		\
188 	TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),		\
189 	    hio_next[(ncomp)]);						\
190 	hio_##name##_list_size[(ncomp)]++;				\
191 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
192 } while (0)
193 #define	QUEUE_INSERT2(hio, name)	do {				\
194 	mtx_lock(&hio_##name##_list_lock);				\
195 	if (TAILQ_EMPTY(&hio_##name##_list))				\
196 		cv_broadcast(&hio_##name##_list_cond);			\
197 	TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
198 	hio_##name##_list_size++;					\
199 	mtx_unlock(&hio_##name##_list_lock);				\
200 } while (0)
201 #define	QUEUE_TAKE1(hio, name, ncomp, timeout)	do {			\
202 	bool _last;							\
203 									\
204 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
205 	_last = false;							\
206 	while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
207 		cv_timedwait(&hio_##name##_list_cond[(ncomp)],		\
208 		    &hio_##name##_list_lock[(ncomp)], (timeout));	\
209 		if ((timeout) != 0)					\
210 			_last = true;					\
211 	}								\
212 	if (hio != NULL) {						\
213 		PJDLOG_ASSERT(hio_##name##_list_size[(ncomp)] != 0);	\
214 		hio_##name##_list_size[(ncomp)]--;			\
215 		TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),	\
216 		    hio_next[(ncomp)]);					\
217 	}								\
218 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
219 } while (0)
220 #define	QUEUE_TAKE2(hio, name)	do {					\
221 	mtx_lock(&hio_##name##_list_lock);				\
222 	while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {	\
223 		cv_wait(&hio_##name##_list_cond,			\
224 		    &hio_##name##_list_lock);				\
225 	}								\
226 	PJDLOG_ASSERT(hio_##name##_list_size != 0);			\
227 	hio_##name##_list_size--;					\
228 	TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);	\
229 	mtx_unlock(&hio_##name##_list_lock);				\
230 } while (0)
231 
232 #define ISFULLSYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_FULLSYNC)
233 #define ISMEMSYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_MEMSYNC)
234 #define ISASYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_ASYNC)
235 
236 #define	SYNCREQ(hio)		do {					\
237 	(hio)->hio_ggio.gctl_unit = -1;					\
238 	(hio)->hio_ggio.gctl_seq = 1;					\
239 } while (0)
240 #define	ISSYNCREQ(hio)		((hio)->hio_ggio.gctl_unit == -1)
241 #define	SYNCREQDONE(hio)	do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
242 #define	ISSYNCREQDONE(hio)	((hio)->hio_ggio.gctl_unit == -2)
243 
244 #define ISMEMSYNCWRITE(hio)	(ISMEMSYNC(hio) &&			\
245 	    (hio)->hio_ggio.gctl_cmd == BIO_WRITE && !ISSYNCREQ(hio))
246 
247 static struct hast_resource *gres;
248 
249 static pthread_mutex_t range_lock;
250 static struct rangelocks *range_regular;
251 static bool range_regular_wait;
252 static pthread_cond_t range_regular_cond;
253 static struct rangelocks *range_sync;
254 static bool range_sync_wait;
255 static pthread_cond_t range_sync_cond;
256 static bool fullystarted;
257 
258 static void *ggate_recv_thread(void *arg);
259 static void *local_send_thread(void *arg);
260 static void *remote_send_thread(void *arg);
261 static void *remote_recv_thread(void *arg);
262 static void *ggate_send_thread(void *arg);
263 static void *sync_thread(void *arg);
264 static void *guard_thread(void *arg);
265 
266 static void
267 output_status_aux(struct nv *nvout)
268 {
269 
270 	nv_add_uint64(nvout, (uint64_t)hio_free_list_size,
271 	    "idle_queue_size");
272 	nv_add_uint64(nvout, (uint64_t)hio_send_local_list_size,
273 	    "local_queue_size");
274 	nv_add_uint64(nvout, (uint64_t)hio_send_remote_list_size,
275 	    "send_queue_size");
276 	nv_add_uint64(nvout, (uint64_t)hio_recv_remote_list_size,
277 	    "recv_queue_size");
278 	nv_add_uint64(nvout, (uint64_t)hio_done_list_size,
279 	    "done_queue_size");
280 }
281 
282 static void
283 cleanup(struct hast_resource *res)
284 {
285 	int rerrno;
286 
287 	/* Remember errno. */
288 	rerrno = errno;
289 
290 	/* Destroy ggate provider if we created one. */
291 	if (res->hr_ggateunit >= 0) {
292 		struct g_gate_ctl_destroy ggiod;
293 
294 		bzero(&ggiod, sizeof(ggiod));
295 		ggiod.gctl_version = G_GATE_VERSION;
296 		ggiod.gctl_unit = res->hr_ggateunit;
297 		ggiod.gctl_force = 1;
298 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) == -1) {
299 			pjdlog_errno(LOG_WARNING,
300 			    "Unable to destroy hast/%s device",
301 			    res->hr_provname);
302 		}
303 		res->hr_ggateunit = -1;
304 	}
305 
306 	/* Restore errno. */
307 	errno = rerrno;
308 }
309 
310 static __dead2 void
311 primary_exit(int exitcode, const char *fmt, ...)
312 {
313 	va_list ap;
314 
315 	PJDLOG_ASSERT(exitcode != EX_OK);
316 	va_start(ap, fmt);
317 	pjdlogv_errno(LOG_ERR, fmt, ap);
318 	va_end(ap);
319 	cleanup(gres);
320 	exit(exitcode);
321 }
322 
323 static __dead2 void
324 primary_exitx(int exitcode, const char *fmt, ...)
325 {
326 	va_list ap;
327 
328 	va_start(ap, fmt);
329 	pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
330 	va_end(ap);
331 	cleanup(gres);
332 	exit(exitcode);
333 }
334 
335 static int
336 hast_activemap_flush(struct hast_resource *res) __unlocks(res->hr_amp_lock)
337 {
338 	const unsigned char *buf;
339 	size_t size;
340 	int ret;
341 
342 	mtx_lock(&res->hr_amp_diskmap_lock);
343 	buf = activemap_bitmap(res->hr_amp, &size);
344 	mtx_unlock(&res->hr_amp_lock);
345 	PJDLOG_ASSERT(buf != NULL);
346 	PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
347 	ret = 0;
348 	if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
349 	    (ssize_t)size) {
350 		pjdlog_errno(LOG_ERR, "Unable to flush activemap to disk");
351 		res->hr_stat_activemap_write_error++;
352 		ret = -1;
353 	}
354 	if (ret == 0 && res->hr_metaflush == 1 &&
355 	    g_flush(res->hr_localfd) == -1) {
356 		if (errno == EOPNOTSUPP) {
357 			pjdlog_warning("The %s provider doesn't support flushing write cache. Disabling it.",
358 			    res->hr_localpath);
359 			res->hr_metaflush = 0;
360 		} else {
361 			pjdlog_errno(LOG_ERR,
362 			    "Unable to flush disk cache on activemap update");
363 			res->hr_stat_activemap_flush_error++;
364 			ret = -1;
365 		}
366 	}
367 	mtx_unlock(&res->hr_amp_diskmap_lock);
368 	return (ret);
369 }
370 
371 static bool
372 real_remote(const struct hast_resource *res)
373 {
374 
375 	return (strcmp(res->hr_remoteaddr, "none") != 0);
376 }
377 
378 static void
379 init_environment(struct hast_resource *res __unused)
380 {
381 	struct hio *hio;
382 	unsigned int ii, ncomps;
383 
384 	/*
385 	 * In the future it might be per-resource value.
386 	 */
387 	ncomps = HAST_NCOMPONENTS;
388 
389 	/*
390 	 * Allocate memory needed by lists.
391 	 */
392 	hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
393 	if (hio_send_list == NULL) {
394 		primary_exitx(EX_TEMPFAIL,
395 		    "Unable to allocate %zu bytes of memory for send lists.",
396 		    sizeof(hio_send_list[0]) * ncomps);
397 	}
398 	hio_send_list_size = malloc(sizeof(hio_send_list_size[0]) * ncomps);
399 	if (hio_send_list_size == NULL) {
400 		primary_exitx(EX_TEMPFAIL,
401 		    "Unable to allocate %zu bytes of memory for send list counters.",
402 		    sizeof(hio_send_list_size[0]) * ncomps);
403 	}
404 	hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
405 	if (hio_send_list_lock == NULL) {
406 		primary_exitx(EX_TEMPFAIL,
407 		    "Unable to allocate %zu bytes of memory for send list locks.",
408 		    sizeof(hio_send_list_lock[0]) * ncomps);
409 	}
410 	hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
411 	if (hio_send_list_cond == NULL) {
412 		primary_exitx(EX_TEMPFAIL,
413 		    "Unable to allocate %zu bytes of memory for send list condition variables.",
414 		    sizeof(hio_send_list_cond[0]) * ncomps);
415 	}
416 	hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
417 	if (hio_recv_list == NULL) {
418 		primary_exitx(EX_TEMPFAIL,
419 		    "Unable to allocate %zu bytes of memory for recv lists.",
420 		    sizeof(hio_recv_list[0]) * ncomps);
421 	}
422 	hio_recv_list_size = malloc(sizeof(hio_recv_list_size[0]) * ncomps);
423 	if (hio_recv_list_size == NULL) {
424 		primary_exitx(EX_TEMPFAIL,
425 		    "Unable to allocate %zu bytes of memory for recv list counters.",
426 		    sizeof(hio_recv_list_size[0]) * ncomps);
427 	}
428 	hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
429 	if (hio_recv_list_lock == NULL) {
430 		primary_exitx(EX_TEMPFAIL,
431 		    "Unable to allocate %zu bytes of memory for recv list locks.",
432 		    sizeof(hio_recv_list_lock[0]) * ncomps);
433 	}
434 	hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
435 	if (hio_recv_list_cond == NULL) {
436 		primary_exitx(EX_TEMPFAIL,
437 		    "Unable to allocate %zu bytes of memory for recv list condition variables.",
438 		    sizeof(hio_recv_list_cond[0]) * ncomps);
439 	}
440 	hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
441 	if (hio_remote_lock == NULL) {
442 		primary_exitx(EX_TEMPFAIL,
443 		    "Unable to allocate %zu bytes of memory for remote connections locks.",
444 		    sizeof(hio_remote_lock[0]) * ncomps);
445 	}
446 
447 	/*
448 	 * Initialize lists, their counters, locks and condition variables.
449 	 */
450 	TAILQ_INIT(&hio_free_list);
451 	mtx_init(&hio_free_list_lock);
452 	cv_init(&hio_free_list_cond);
453 	for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
454 		TAILQ_INIT(&hio_send_list[ii]);
455 		hio_send_list_size[ii] = 0;
456 		mtx_init(&hio_send_list_lock[ii]);
457 		cv_init(&hio_send_list_cond[ii]);
458 		TAILQ_INIT(&hio_recv_list[ii]);
459 		hio_recv_list_size[ii] = 0;
460 		mtx_init(&hio_recv_list_lock[ii]);
461 		cv_init(&hio_recv_list_cond[ii]);
462 		rw_init(&hio_remote_lock[ii]);
463 	}
464 	TAILQ_INIT(&hio_done_list);
465 	mtx_init(&hio_done_list_lock);
466 	cv_init(&hio_done_list_cond);
467 	mtx_init(&metadata_lock);
468 
469 	/*
470 	 * Allocate requests pool and initialize requests.
471 	 */
472 	for (ii = 0; ii < HAST_HIO_MAX; ii++) {
473 		hio = malloc(sizeof(*hio));
474 		if (hio == NULL) {
475 			primary_exitx(EX_TEMPFAIL,
476 			    "Unable to allocate %zu bytes of memory for hio request.",
477 			    sizeof(*hio));
478 		}
479 		refcnt_init(&hio->hio_countdown, 0);
480 		hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
481 		if (hio->hio_errors == NULL) {
482 			primary_exitx(EX_TEMPFAIL,
483 			    "Unable allocate %zu bytes of memory for hio errors.",
484 			    sizeof(hio->hio_errors[0]) * ncomps);
485 		}
486 		hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
487 		if (hio->hio_next == NULL) {
488 			primary_exitx(EX_TEMPFAIL,
489 			    "Unable allocate %zu bytes of memory for hio_next field.",
490 			    sizeof(hio->hio_next[0]) * ncomps);
491 		}
492 		hio->hio_ggio.gctl_version = G_GATE_VERSION;
493 		hio->hio_ggio.gctl_data = malloc(MAXPHYS);
494 		if (hio->hio_ggio.gctl_data == NULL) {
495 			primary_exitx(EX_TEMPFAIL,
496 			    "Unable to allocate %zu bytes of memory for gctl_data.",
497 			    MAXPHYS);
498 		}
499 		hio->hio_ggio.gctl_length = MAXPHYS;
500 		hio->hio_ggio.gctl_error = 0;
501 		TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
502 		hio_free_list_size++;
503 	}
504 }
505 
506 static bool
507 init_resuid(struct hast_resource *res)
508 {
509 
510 	mtx_lock(&metadata_lock);
511 	if (res->hr_resuid != 0) {
512 		mtx_unlock(&metadata_lock);
513 		return (false);
514 	} else {
515 		/* Initialize unique resource identifier. */
516 		arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
517 		mtx_unlock(&metadata_lock);
518 		if (metadata_write(res) == -1)
519 			exit(EX_NOINPUT);
520 		return (true);
521 	}
522 }
523 
524 static void
525 init_local(struct hast_resource *res)
526 {
527 	unsigned char *buf;
528 	size_t mapsize;
529 
530 	if (metadata_read(res, true) == -1)
531 		exit(EX_NOINPUT);
532 	mtx_init(&res->hr_amp_lock);
533 	if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
534 	    res->hr_local_sectorsize, res->hr_keepdirty) == -1) {
535 		primary_exit(EX_TEMPFAIL, "Unable to create activemap");
536 	}
537 	mtx_init(&range_lock);
538 	cv_init(&range_regular_cond);
539 	if (rangelock_init(&range_regular) == -1)
540 		primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
541 	cv_init(&range_sync_cond);
542 	if (rangelock_init(&range_sync) == -1)
543 		primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
544 	mapsize = activemap_ondisk_size(res->hr_amp);
545 	buf = calloc(1, mapsize);
546 	if (buf == NULL) {
547 		primary_exitx(EX_TEMPFAIL,
548 		    "Unable to allocate buffer for activemap.");
549 	}
550 	if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
551 	    (ssize_t)mapsize) {
552 		primary_exit(EX_NOINPUT, "Unable to read activemap");
553 	}
554 	activemap_copyin(res->hr_amp, buf, mapsize);
555 	free(buf);
556 	if (res->hr_resuid != 0)
557 		return;
558 	/*
559 	 * We're using provider for the first time. Initialize local and remote
560 	 * counters. We don't initialize resuid here, as we want to do it just
561 	 * in time. The reason for this is that we want to inform secondary
562 	 * that there were no writes yet, so there is no need to synchronize
563 	 * anything.
564 	 */
565 	res->hr_primary_localcnt = 0;
566 	res->hr_primary_remotecnt = 0;
567 	if (metadata_write(res) == -1)
568 		exit(EX_NOINPUT);
569 }
570 
571 static int
572 primary_connect(struct hast_resource *res, struct proto_conn **connp)
573 {
574 	struct proto_conn *conn;
575 	int16_t val;
576 
577 	val = 1;
578 	if (proto_send(res->hr_conn, &val, sizeof(val)) == -1) {
579 		primary_exit(EX_TEMPFAIL,
580 		    "Unable to send connection request to parent");
581 	}
582 	if (proto_recv(res->hr_conn, &val, sizeof(val)) == -1) {
583 		primary_exit(EX_TEMPFAIL,
584 		    "Unable to receive reply to connection request from parent");
585 	}
586 	if (val != 0) {
587 		errno = val;
588 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
589 		    res->hr_remoteaddr);
590 		return (-1);
591 	}
592 	if (proto_connection_recv(res->hr_conn, true, &conn) == -1) {
593 		primary_exit(EX_TEMPFAIL,
594 		    "Unable to receive connection from parent");
595 	}
596 	if (proto_connect_wait(conn, res->hr_timeout) == -1) {
597 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
598 		    res->hr_remoteaddr);
599 		proto_close(conn);
600 		return (-1);
601 	}
602 	/* Error in setting timeout is not critical, but why should it fail? */
603 	if (proto_timeout(conn, res->hr_timeout) == -1)
604 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
605 
606 	*connp = conn;
607 
608 	return (0);
609 }
610 
611 /*
612  * Function instructs GEOM_GATE to handle reads directly from within the kernel.
613  */
614 static void
615 enable_direct_reads(struct hast_resource *res)
616 {
617 	struct g_gate_ctl_modify ggiomodify;
618 
619 	bzero(&ggiomodify, sizeof(ggiomodify));
620 	ggiomodify.gctl_version = G_GATE_VERSION;
621 	ggiomodify.gctl_unit = res->hr_ggateunit;
622 	ggiomodify.gctl_modify = GG_MODIFY_READPROV | GG_MODIFY_READOFFSET;
623 	strlcpy(ggiomodify.gctl_readprov, res->hr_localpath,
624 	    sizeof(ggiomodify.gctl_readprov));
625 	ggiomodify.gctl_readoffset = res->hr_localoff;
626 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_MODIFY, &ggiomodify) == 0)
627 		pjdlog_debug(1, "Direct reads enabled.");
628 	else
629 		pjdlog_errno(LOG_WARNING, "Failed to enable direct reads");
630 }
631 
632 static int
633 init_remote(struct hast_resource *res, struct proto_conn **inp,
634     struct proto_conn **outp)
635 {
636 	struct proto_conn *in, *out;
637 	struct nv *nvout, *nvin;
638 	const unsigned char *token;
639 	unsigned char *map;
640 	const char *errmsg;
641 	int32_t extentsize;
642 	int64_t datasize;
643 	uint32_t mapsize;
644 	uint8_t version;
645 	size_t size;
646 	int error;
647 
648 	PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
649 	PJDLOG_ASSERT(real_remote(res));
650 
651 	in = out = NULL;
652 	errmsg = NULL;
653 
654 	if (primary_connect(res, &out) == -1)
655 		return (ECONNREFUSED);
656 
657 	error = ECONNABORTED;
658 
659 	/*
660 	 * First handshake step.
661 	 * Setup outgoing connection with remote node.
662 	 */
663 	nvout = nv_alloc();
664 	nv_add_string(nvout, res->hr_name, "resource");
665 	nv_add_uint8(nvout, HAST_PROTO_VERSION, "version");
666 	if (nv_error(nvout) != 0) {
667 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
668 		    "Unable to allocate header for connection with %s",
669 		    res->hr_remoteaddr);
670 		nv_free(nvout);
671 		goto close;
672 	}
673 	if (hast_proto_send(res, out, nvout, NULL, 0) == -1) {
674 		pjdlog_errno(LOG_WARNING,
675 		    "Unable to send handshake header to %s",
676 		    res->hr_remoteaddr);
677 		nv_free(nvout);
678 		goto close;
679 	}
680 	nv_free(nvout);
681 	if (hast_proto_recv_hdr(out, &nvin) == -1) {
682 		pjdlog_errno(LOG_WARNING,
683 		    "Unable to receive handshake header from %s",
684 		    res->hr_remoteaddr);
685 		goto close;
686 	}
687 	errmsg = nv_get_string(nvin, "errmsg");
688 	if (errmsg != NULL) {
689 		pjdlog_warning("%s", errmsg);
690 		if (nv_exists(nvin, "wait"))
691 			error = EBUSY;
692 		nv_free(nvin);
693 		goto close;
694 	}
695 	version = nv_get_uint8(nvin, "version");
696 	if (version == 0) {
697 		/*
698 		 * If no version is sent, it means this is protocol version 1.
699 		 */
700 		version = 1;
701 	}
702 	if (version > HAST_PROTO_VERSION) {
703 		pjdlog_warning("Invalid version received (%hhu).", version);
704 		nv_free(nvin);
705 		goto close;
706 	}
707 	res->hr_version = version;
708 	pjdlog_debug(1, "Negotiated protocol version %d.", res->hr_version);
709 	token = nv_get_uint8_array(nvin, &size, "token");
710 	if (token == NULL) {
711 		pjdlog_warning("Handshake header from %s has no 'token' field.",
712 		    res->hr_remoteaddr);
713 		nv_free(nvin);
714 		goto close;
715 	}
716 	if (size != sizeof(res->hr_token)) {
717 		pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
718 		    res->hr_remoteaddr, size, sizeof(res->hr_token));
719 		nv_free(nvin);
720 		goto close;
721 	}
722 	bcopy(token, res->hr_token, sizeof(res->hr_token));
723 	nv_free(nvin);
724 
725 	/*
726 	 * Second handshake step.
727 	 * Setup incoming connection with remote node.
728 	 */
729 	if (primary_connect(res, &in) == -1)
730 		goto close;
731 
732 	nvout = nv_alloc();
733 	nv_add_string(nvout, res->hr_name, "resource");
734 	nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
735 	    "token");
736 	if (res->hr_resuid == 0) {
737 		/*
738 		 * The resuid field was not yet initialized.
739 		 * Because we do synchronization inside init_resuid(), it is
740 		 * possible that someone already initialized it, the function
741 		 * will return false then, but if we successfully initialized
742 		 * it, we will get true. True means that there were no writes
743 		 * to this resource yet and we want to inform secondary that
744 		 * synchronization is not needed by sending "virgin" argument.
745 		 */
746 		if (init_resuid(res))
747 			nv_add_int8(nvout, 1, "virgin");
748 	}
749 	nv_add_uint64(nvout, res->hr_resuid, "resuid");
750 	nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
751 	nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
752 	if (nv_error(nvout) != 0) {
753 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
754 		    "Unable to allocate header for connection with %s",
755 		    res->hr_remoteaddr);
756 		nv_free(nvout);
757 		goto close;
758 	}
759 	if (hast_proto_send(res, in, nvout, NULL, 0) == -1) {
760 		pjdlog_errno(LOG_WARNING,
761 		    "Unable to send handshake header to %s",
762 		    res->hr_remoteaddr);
763 		nv_free(nvout);
764 		goto close;
765 	}
766 	nv_free(nvout);
767 	if (hast_proto_recv_hdr(out, &nvin) == -1) {
768 		pjdlog_errno(LOG_WARNING,
769 		    "Unable to receive handshake header from %s",
770 		    res->hr_remoteaddr);
771 		goto close;
772 	}
773 	errmsg = nv_get_string(nvin, "errmsg");
774 	if (errmsg != NULL) {
775 		pjdlog_warning("%s", errmsg);
776 		nv_free(nvin);
777 		goto close;
778 	}
779 	datasize = nv_get_int64(nvin, "datasize");
780 	if (datasize != res->hr_datasize) {
781 		pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
782 		    (intmax_t)res->hr_datasize, (intmax_t)datasize);
783 		nv_free(nvin);
784 		goto close;
785 	}
786 	extentsize = nv_get_int32(nvin, "extentsize");
787 	if (extentsize != res->hr_extentsize) {
788 		pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
789 		    (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
790 		nv_free(nvin);
791 		goto close;
792 	}
793 	res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
794 	res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
795 	res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
796 	if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY)
797 		enable_direct_reads(res);
798 	if (nv_exists(nvin, "virgin")) {
799 		/*
800 		 * Secondary was reinitialized, bump localcnt if it is 0 as
801 		 * only we have the data.
802 		 */
803 		PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY);
804 		PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
805 
806 		if (res->hr_primary_localcnt == 0) {
807 			PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0);
808 
809 			mtx_lock(&metadata_lock);
810 			res->hr_primary_localcnt++;
811 			pjdlog_debug(1, "Increasing localcnt to %ju.",
812 			    (uintmax_t)res->hr_primary_localcnt);
813 			(void)metadata_write(res);
814 			mtx_unlock(&metadata_lock);
815 		}
816 	}
817 	map = NULL;
818 	mapsize = nv_get_uint32(nvin, "mapsize");
819 	if (mapsize > 0) {
820 		map = malloc(mapsize);
821 		if (map == NULL) {
822 			pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
823 			    (uintmax_t)mapsize);
824 			nv_free(nvin);
825 			goto close;
826 		}
827 		/*
828 		 * Remote node have some dirty extents on its own, lets
829 		 * download its activemap.
830 		 */
831 		if (hast_proto_recv_data(res, out, nvin, map,
832 		    mapsize) == -1) {
833 			pjdlog_errno(LOG_ERR,
834 			    "Unable to receive remote activemap");
835 			nv_free(nvin);
836 			free(map);
837 			goto close;
838 		}
839 		mtx_lock(&res->hr_amp_lock);
840 		/*
841 		 * Merge local and remote bitmaps.
842 		 */
843 		activemap_merge(res->hr_amp, map, mapsize);
844 		free(map);
845 		/*
846 		 * Now that we merged bitmaps from both nodes, flush it to the
847 		 * disk before we start to synchronize.
848 		 */
849 		(void)hast_activemap_flush(res);
850 	}
851 	nv_free(nvin);
852 #ifdef notyet
853 	/* Setup directions. */
854 	if (proto_send(out, NULL, 0) == -1)
855 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
856 	if (proto_recv(in, NULL, 0) == -1)
857 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
858 #endif
859 	pjdlog_info("Connected to %s.", res->hr_remoteaddr);
860 	if (res->hr_original_replication == HAST_REPLICATION_MEMSYNC &&
861 	    res->hr_version < 2) {
862 		pjdlog_warning("The 'memsync' replication mode is not supported by the remote node, falling back to 'fullsync' mode.");
863 		res->hr_replication = HAST_REPLICATION_FULLSYNC;
864 	} else if (res->hr_replication != res->hr_original_replication) {
865 		/*
866 		 * This is in case hastd disconnected and was upgraded.
867 		 */
868 		res->hr_replication = res->hr_original_replication;
869 	}
870 	if (inp != NULL && outp != NULL) {
871 		*inp = in;
872 		*outp = out;
873 	} else {
874 		res->hr_remotein = in;
875 		res->hr_remoteout = out;
876 	}
877 	event_send(res, EVENT_CONNECT);
878 	return (0);
879 close:
880 	if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
881 		event_send(res, EVENT_SPLITBRAIN);
882 	proto_close(out);
883 	if (in != NULL)
884 		proto_close(in);
885 	return (error);
886 }
887 
888 static void
889 sync_start(void)
890 {
891 
892 	mtx_lock(&sync_lock);
893 	sync_inprogress = true;
894 	mtx_unlock(&sync_lock);
895 	cv_signal(&sync_cond);
896 }
897 
898 static void
899 sync_stop(void)
900 {
901 
902 	mtx_lock(&sync_lock);
903 	if (sync_inprogress)
904 		sync_inprogress = false;
905 	mtx_unlock(&sync_lock);
906 }
907 
908 static void
909 init_ggate(struct hast_resource *res)
910 {
911 	struct g_gate_ctl_create ggiocreate;
912 	struct g_gate_ctl_cancel ggiocancel;
913 
914 	/*
915 	 * We communicate with ggate via /dev/ggctl. Open it.
916 	 */
917 	res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
918 	if (res->hr_ggatefd == -1)
919 		primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
920 	/*
921 	 * Create provider before trying to connect, as connection failure
922 	 * is not critical, but may take some time.
923 	 */
924 	bzero(&ggiocreate, sizeof(ggiocreate));
925 	ggiocreate.gctl_version = G_GATE_VERSION;
926 	ggiocreate.gctl_mediasize = res->hr_datasize;
927 	ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
928 	ggiocreate.gctl_flags = 0;
929 	ggiocreate.gctl_maxcount = 0;
930 	ggiocreate.gctl_timeout = 0;
931 	ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
932 	snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
933 	    res->hr_provname);
934 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
935 		pjdlog_info("Device hast/%s created.", res->hr_provname);
936 		res->hr_ggateunit = ggiocreate.gctl_unit;
937 		return;
938 	}
939 	if (errno != EEXIST) {
940 		primary_exit(EX_OSERR, "Unable to create hast/%s device",
941 		    res->hr_provname);
942 	}
943 	pjdlog_debug(1,
944 	    "Device hast/%s already exists, we will try to take it over.",
945 	    res->hr_provname);
946 	/*
947 	 * If we received EEXIST, we assume that the process who created the
948 	 * provider died and didn't clean up. In that case we will start from
949 	 * where he left of.
950 	 */
951 	bzero(&ggiocancel, sizeof(ggiocancel));
952 	ggiocancel.gctl_version = G_GATE_VERSION;
953 	ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
954 	snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
955 	    res->hr_provname);
956 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
957 		pjdlog_info("Device hast/%s recovered.", res->hr_provname);
958 		res->hr_ggateunit = ggiocancel.gctl_unit;
959 		return;
960 	}
961 	primary_exit(EX_OSERR, "Unable to take over hast/%s device",
962 	    res->hr_provname);
963 }
964 
965 void
966 hastd_primary(struct hast_resource *res)
967 {
968 	pthread_t td;
969 	pid_t pid;
970 	int error, mode, debuglevel;
971 
972 	/*
973 	 * Create communication channel for sending control commands from
974 	 * parent to child.
975 	 */
976 	if (proto_client(NULL, "socketpair://", &res->hr_ctrl) == -1) {
977 		/* TODO: There's no need for this to be fatal error. */
978 		KEEP_ERRNO((void)pidfile_remove(pfh));
979 		pjdlog_exit(EX_OSERR,
980 		    "Unable to create control sockets between parent and child");
981 	}
982 	/*
983 	 * Create communication channel for sending events from child to parent.
984 	 */
985 	if (proto_client(NULL, "socketpair://", &res->hr_event) == -1) {
986 		/* TODO: There's no need for this to be fatal error. */
987 		KEEP_ERRNO((void)pidfile_remove(pfh));
988 		pjdlog_exit(EX_OSERR,
989 		    "Unable to create event sockets between child and parent");
990 	}
991 	/*
992 	 * Create communication channel for sending connection requests from
993 	 * child to parent.
994 	 */
995 	if (proto_client(NULL, "socketpair://", &res->hr_conn) == -1) {
996 		/* TODO: There's no need for this to be fatal error. */
997 		KEEP_ERRNO((void)pidfile_remove(pfh));
998 		pjdlog_exit(EX_OSERR,
999 		    "Unable to create connection sockets between child and parent");
1000 	}
1001 
1002 	pid = fork();
1003 	if (pid == -1) {
1004 		/* TODO: There's no need for this to be fatal error. */
1005 		KEEP_ERRNO((void)pidfile_remove(pfh));
1006 		pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
1007 	}
1008 
1009 	if (pid > 0) {
1010 		/* This is parent. */
1011 		/* Declare that we are receiver. */
1012 		proto_recv(res->hr_event, NULL, 0);
1013 		proto_recv(res->hr_conn, NULL, 0);
1014 		/* Declare that we are sender. */
1015 		proto_send(res->hr_ctrl, NULL, 0);
1016 		res->hr_workerpid = pid;
1017 		return;
1018 	}
1019 
1020 	gres = res;
1021 	res->output_status_aux = output_status_aux;
1022 	mode = pjdlog_mode_get();
1023 	debuglevel = pjdlog_debug_get();
1024 
1025 	/* Declare that we are sender. */
1026 	proto_send(res->hr_event, NULL, 0);
1027 	proto_send(res->hr_conn, NULL, 0);
1028 	/* Declare that we are receiver. */
1029 	proto_recv(res->hr_ctrl, NULL, 0);
1030 	descriptors_cleanup(res);
1031 
1032 	descriptors_assert(res, mode);
1033 
1034 	pjdlog_init(mode);
1035 	pjdlog_debug_set(debuglevel);
1036 	pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
1037 	setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
1038 
1039 	init_local(res);
1040 	init_ggate(res);
1041 	init_environment(res);
1042 
1043 	if (drop_privs(res) != 0) {
1044 		cleanup(res);
1045 		exit(EX_CONFIG);
1046 	}
1047 	pjdlog_info("Privileges successfully dropped.");
1048 
1049 	/*
1050 	 * Create the guard thread first, so we can handle signals from the
1051 	 * very beginning.
1052 	 */
1053 	error = pthread_create(&td, NULL, guard_thread, res);
1054 	PJDLOG_ASSERT(error == 0);
1055 	/*
1056 	 * Create the control thread before sending any event to the parent,
1057 	 * as we can deadlock when parent sends control request to worker,
1058 	 * but worker has no control thread started yet, so parent waits.
1059 	 * In the meantime worker sends an event to the parent, but parent
1060 	 * is unable to handle the event, because it waits for control
1061 	 * request response.
1062 	 */
1063 	error = pthread_create(&td, NULL, ctrl_thread, res);
1064 	PJDLOG_ASSERT(error == 0);
1065 	if (real_remote(res)) {
1066 		error = init_remote(res, NULL, NULL);
1067 		if (error == 0) {
1068 			sync_start();
1069 		} else if (error == EBUSY) {
1070 			time_t start = time(NULL);
1071 
1072 			pjdlog_warning("Waiting for remote node to become %s for %ds.",
1073 			    role2str(HAST_ROLE_SECONDARY),
1074 			    res->hr_timeout);
1075 			for (;;) {
1076 				sleep(1);
1077 				error = init_remote(res, NULL, NULL);
1078 				if (error != EBUSY)
1079 					break;
1080 				if (time(NULL) > start + res->hr_timeout)
1081 					break;
1082 			}
1083 			if (error == EBUSY) {
1084 				pjdlog_warning("Remote node is still %s, starting anyway.",
1085 				    role2str(HAST_ROLE_PRIMARY));
1086 			}
1087 		}
1088 	}
1089 	error = pthread_create(&td, NULL, ggate_recv_thread, res);
1090 	PJDLOG_ASSERT(error == 0);
1091 	error = pthread_create(&td, NULL, local_send_thread, res);
1092 	PJDLOG_ASSERT(error == 0);
1093 	error = pthread_create(&td, NULL, remote_send_thread, res);
1094 	PJDLOG_ASSERT(error == 0);
1095 	error = pthread_create(&td, NULL, remote_recv_thread, res);
1096 	PJDLOG_ASSERT(error == 0);
1097 	error = pthread_create(&td, NULL, ggate_send_thread, res);
1098 	PJDLOG_ASSERT(error == 0);
1099 	fullystarted = true;
1100 	(void)sync_thread(res);
1101 }
1102 
1103 static void
1104 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio,
1105     const char *fmt, ...)
1106 {
1107 	char msg[1024];
1108 	va_list ap;
1109 
1110 	va_start(ap, fmt);
1111 	(void)vsnprintf(msg, sizeof(msg), fmt, ap);
1112 	va_end(ap);
1113 	switch (ggio->gctl_cmd) {
1114 	case BIO_READ:
1115 		(void)snprlcat(msg, sizeof(msg), "READ(%ju, %ju).",
1116 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1117 		break;
1118 	case BIO_DELETE:
1119 		(void)snprlcat(msg, sizeof(msg), "DELETE(%ju, %ju).",
1120 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1121 		break;
1122 	case BIO_FLUSH:
1123 		(void)snprlcat(msg, sizeof(msg), "FLUSH.");
1124 		break;
1125 	case BIO_WRITE:
1126 		(void)snprlcat(msg, sizeof(msg), "WRITE(%ju, %ju).",
1127 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1128 		break;
1129 	default:
1130 		(void)snprlcat(msg, sizeof(msg), "UNKNOWN(%u).",
1131 		    (unsigned int)ggio->gctl_cmd);
1132 		break;
1133 	}
1134 	pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
1135 }
1136 
1137 static void
1138 remote_close(struct hast_resource *res, int ncomp)
1139 {
1140 
1141 	rw_wlock(&hio_remote_lock[ncomp]);
1142 	/*
1143 	 * Check for a race between dropping rlock and acquiring wlock -
1144 	 * another thread can close connection in-between.
1145 	 */
1146 	if (!ISCONNECTED(res, ncomp)) {
1147 		PJDLOG_ASSERT(res->hr_remotein == NULL);
1148 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
1149 		rw_unlock(&hio_remote_lock[ncomp]);
1150 		return;
1151 	}
1152 
1153 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1154 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1155 
1156 	pjdlog_debug(2, "Closing incoming connection to %s.",
1157 	    res->hr_remoteaddr);
1158 	proto_close(res->hr_remotein);
1159 	res->hr_remotein = NULL;
1160 	pjdlog_debug(2, "Closing outgoing connection to %s.",
1161 	    res->hr_remoteaddr);
1162 	proto_close(res->hr_remoteout);
1163 	res->hr_remoteout = NULL;
1164 
1165 	rw_unlock(&hio_remote_lock[ncomp]);
1166 
1167 	pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
1168 
1169 	/*
1170 	 * Stop synchronization if in-progress.
1171 	 */
1172 	sync_stop();
1173 
1174 	event_send(res, EVENT_DISCONNECT);
1175 }
1176 
1177 /*
1178  * Acknowledge write completion to the kernel, but don't update activemap yet.
1179  */
1180 static void
1181 write_complete(struct hast_resource *res, struct hio *hio)
1182 {
1183 	struct g_gate_ctl_io *ggio;
1184 	unsigned int ncomp;
1185 
1186 	PJDLOG_ASSERT(!hio->hio_done);
1187 
1188 	ggio = &hio->hio_ggio;
1189 	PJDLOG_ASSERT(ggio->gctl_cmd == BIO_WRITE);
1190 
1191 	/*
1192 	 * Bump local count if this is first write after
1193 	 * connection failure with remote node.
1194 	 */
1195 	ncomp = 1;
1196 	rw_rlock(&hio_remote_lock[ncomp]);
1197 	if (!ISCONNECTED(res, ncomp)) {
1198 		mtx_lock(&metadata_lock);
1199 		if (res->hr_primary_localcnt == res->hr_secondary_remotecnt) {
1200 			res->hr_primary_localcnt++;
1201 			pjdlog_debug(1, "Increasing localcnt to %ju.",
1202 			    (uintmax_t)res->hr_primary_localcnt);
1203 			(void)metadata_write(res);
1204 		}
1205 		mtx_unlock(&metadata_lock);
1206 	}
1207 	rw_unlock(&hio_remote_lock[ncomp]);
1208 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1)
1209 		primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1210 	hio->hio_done = true;
1211 }
1212 
1213 /*
1214  * Thread receives ggate I/O requests from the kernel and passes them to
1215  * appropriate threads:
1216  * WRITE - always goes to both local_send and remote_send threads
1217  * READ (when the block is up-to-date on local component) -
1218  *	only local_send thread
1219  * READ (when the block isn't up-to-date on local component) -
1220  *	only remote_send thread
1221  * DELETE - always goes to both local_send and remote_send threads
1222  * FLUSH - always goes to both local_send and remote_send threads
1223  */
1224 static void *
1225 ggate_recv_thread(void *arg)
1226 {
1227 	struct hast_resource *res = arg;
1228 	struct g_gate_ctl_io *ggio;
1229 	struct hio *hio;
1230 	unsigned int ii, ncomp, ncomps;
1231 	int error;
1232 
1233 	for (;;) {
1234 		pjdlog_debug(2, "ggate_recv: Taking free request.");
1235 		QUEUE_TAKE2(hio, free);
1236 		pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1237 		ggio = &hio->hio_ggio;
1238 		ggio->gctl_unit = res->hr_ggateunit;
1239 		ggio->gctl_length = MAXPHYS;
1240 		ggio->gctl_error = 0;
1241 		hio->hio_done = false;
1242 		hio->hio_replication = res->hr_replication;
1243 		pjdlog_debug(2,
1244 		    "ggate_recv: (%p) Waiting for request from the kernel.",
1245 		    hio);
1246 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) == -1) {
1247 			if (sigexit_received)
1248 				pthread_exit(NULL);
1249 			primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1250 		}
1251 		error = ggio->gctl_error;
1252 		switch (error) {
1253 		case 0:
1254 			break;
1255 		case ECANCELED:
1256 			/* Exit gracefully. */
1257 			if (!sigexit_received) {
1258 				pjdlog_debug(2,
1259 				    "ggate_recv: (%p) Received cancel from the kernel.",
1260 				    hio);
1261 				pjdlog_info("Received cancel from the kernel, exiting.");
1262 			}
1263 			pthread_exit(NULL);
1264 		case ENOMEM:
1265 			/*
1266 			 * Buffer too small? Impossible, we allocate MAXPHYS
1267 			 * bytes - request can't be bigger than that.
1268 			 */
1269 			/* FALLTHROUGH */
1270 		case ENXIO:
1271 		default:
1272 			primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1273 			    strerror(error));
1274 		}
1275 
1276 		ncomp = 0;
1277 		ncomps = HAST_NCOMPONENTS;
1278 
1279 		for (ii = 0; ii < ncomps; ii++)
1280 			hio->hio_errors[ii] = EINVAL;
1281 		reqlog(LOG_DEBUG, 2, ggio,
1282 		    "ggate_recv: (%p) Request received from the kernel: ",
1283 		    hio);
1284 
1285 		/*
1286 		 * Inform all components about new write request.
1287 		 * For read request prefer local component unless the given
1288 		 * range is out-of-date, then use remote component.
1289 		 */
1290 		switch (ggio->gctl_cmd) {
1291 		case BIO_READ:
1292 			res->hr_stat_read++;
1293 			ncomps = 1;
1294 			mtx_lock(&metadata_lock);
1295 			if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1296 			    res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1297 				/*
1298 				 * This range is up-to-date on local component,
1299 				 * so handle request locally.
1300 				 */
1301 				 /* Local component is 0 for now. */
1302 				ncomp = 0;
1303 			} else /* if (res->hr_syncsrc ==
1304 			    HAST_SYNCSRC_SECONDARY) */ {
1305 				PJDLOG_ASSERT(res->hr_syncsrc ==
1306 				    HAST_SYNCSRC_SECONDARY);
1307 				/*
1308 				 * This range is out-of-date on local component,
1309 				 * so send request to the remote node.
1310 				 */
1311 				 /* Remote component is 1 for now. */
1312 				ncomp = 1;
1313 			}
1314 			mtx_unlock(&metadata_lock);
1315 			break;
1316 		case BIO_WRITE:
1317 			res->hr_stat_write++;
1318 			if (res->hr_resuid == 0 &&
1319 			    res->hr_primary_localcnt == 0) {
1320 				/* This is first write. */
1321 				res->hr_primary_localcnt = 1;
1322 			}
1323 			for (;;) {
1324 				mtx_lock(&range_lock);
1325 				if (rangelock_islocked(range_sync,
1326 				    ggio->gctl_offset, ggio->gctl_length)) {
1327 					pjdlog_debug(2,
1328 					    "regular: Range offset=%jd length=%zu locked.",
1329 					    (intmax_t)ggio->gctl_offset,
1330 					    (size_t)ggio->gctl_length);
1331 					range_regular_wait = true;
1332 					cv_wait(&range_regular_cond, &range_lock);
1333 					range_regular_wait = false;
1334 					mtx_unlock(&range_lock);
1335 					continue;
1336 				}
1337 				if (rangelock_add(range_regular,
1338 				    ggio->gctl_offset, ggio->gctl_length) == -1) {
1339 					mtx_unlock(&range_lock);
1340 					pjdlog_debug(2,
1341 					    "regular: Range offset=%jd length=%zu is already locked, waiting.",
1342 					    (intmax_t)ggio->gctl_offset,
1343 					    (size_t)ggio->gctl_length);
1344 					sleep(1);
1345 					continue;
1346 				}
1347 				mtx_unlock(&range_lock);
1348 				break;
1349 			}
1350 			mtx_lock(&res->hr_amp_lock);
1351 			if (activemap_write_start(res->hr_amp,
1352 			    ggio->gctl_offset, ggio->gctl_length)) {
1353 				res->hr_stat_activemap_update++;
1354 				(void)hast_activemap_flush(res);
1355 			} else {
1356 				mtx_unlock(&res->hr_amp_lock);
1357 			}
1358 			if (ISMEMSYNC(hio)) {
1359 				hio->hio_memsyncacked = false;
1360 				refcnt_init(&hio->hio_writecount, ncomps);
1361 			}
1362 			break;
1363 		case BIO_DELETE:
1364 			res->hr_stat_delete++;
1365 			break;
1366 		case BIO_FLUSH:
1367 			res->hr_stat_flush++;
1368 			break;
1369 		}
1370 		pjdlog_debug(2,
1371 		    "ggate_recv: (%p) Moving request to the send queues.", hio);
1372 		refcnt_init(&hio->hio_countdown, ncomps);
1373 		for (ii = ncomp; ii < ncomps; ii++)
1374 			QUEUE_INSERT1(hio, send, ii);
1375 	}
1376 	/* NOTREACHED */
1377 	return (NULL);
1378 }
1379 
1380 /*
1381  * Thread reads from or writes to local component.
1382  * If local read fails, it redirects it to remote_send thread.
1383  */
1384 static void *
1385 local_send_thread(void *arg)
1386 {
1387 	struct hast_resource *res = arg;
1388 	struct g_gate_ctl_io *ggio;
1389 	struct hio *hio;
1390 	unsigned int ncomp, rncomp;
1391 	ssize_t ret;
1392 
1393 	/* Local component is 0 for now. */
1394 	ncomp = 0;
1395 	/* Remote component is 1 for now. */
1396 	rncomp = 1;
1397 
1398 	for (;;) {
1399 		pjdlog_debug(2, "local_send: Taking request.");
1400 		QUEUE_TAKE1(hio, send, ncomp, 0);
1401 		pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1402 		ggio = &hio->hio_ggio;
1403 		switch (ggio->gctl_cmd) {
1404 		case BIO_READ:
1405 			ret = pread(res->hr_localfd, ggio->gctl_data,
1406 			    ggio->gctl_length,
1407 			    ggio->gctl_offset + res->hr_localoff);
1408 			if (ret == ggio->gctl_length)
1409 				hio->hio_errors[ncomp] = 0;
1410 			else if (!ISSYNCREQ(hio)) {
1411 				/*
1412 				 * If READ failed, try to read from remote node.
1413 				 */
1414 				if (ret == -1) {
1415 					reqlog(LOG_WARNING, 0, ggio,
1416 					    "Local request failed (%s), trying remote node. ",
1417 					    strerror(errno));
1418 				} else if (ret != ggio->gctl_length) {
1419 					reqlog(LOG_WARNING, 0, ggio,
1420 					    "Local request failed (%zd != %jd), trying remote node. ",
1421 					    ret, (intmax_t)ggio->gctl_length);
1422 				}
1423 				QUEUE_INSERT1(hio, send, rncomp);
1424 				continue;
1425 			}
1426 			break;
1427 		case BIO_WRITE:
1428 			ret = pwrite(res->hr_localfd, ggio->gctl_data,
1429 			    ggio->gctl_length,
1430 			    ggio->gctl_offset + res->hr_localoff);
1431 			if (ret == -1) {
1432 				hio->hio_errors[ncomp] = errno;
1433 				reqlog(LOG_WARNING, 0, ggio,
1434 				    "Local request failed (%s): ",
1435 				    strerror(errno));
1436 			} else if (ret != ggio->gctl_length) {
1437 				hio->hio_errors[ncomp] = EIO;
1438 				reqlog(LOG_WARNING, 0, ggio,
1439 				    "Local request failed (%zd != %jd): ",
1440 				    ret, (intmax_t)ggio->gctl_length);
1441 			} else {
1442 				hio->hio_errors[ncomp] = 0;
1443 				if (ISASYNC(hio)) {
1444 					ggio->gctl_error = 0;
1445 					write_complete(res, hio);
1446 				}
1447 			}
1448 			break;
1449 		case BIO_DELETE:
1450 			ret = g_delete(res->hr_localfd,
1451 			    ggio->gctl_offset + res->hr_localoff,
1452 			    ggio->gctl_length);
1453 			if (ret == -1) {
1454 				hio->hio_errors[ncomp] = errno;
1455 				reqlog(LOG_WARNING, 0, ggio,
1456 				    "Local request failed (%s): ",
1457 				    strerror(errno));
1458 			} else {
1459 				hio->hio_errors[ncomp] = 0;
1460 			}
1461 			break;
1462 		case BIO_FLUSH:
1463 			if (!res->hr_localflush) {
1464 				ret = -1;
1465 				errno = EOPNOTSUPP;
1466 				break;
1467 			}
1468 			ret = g_flush(res->hr_localfd);
1469 			if (ret == -1) {
1470 				if (errno == EOPNOTSUPP)
1471 					res->hr_localflush = false;
1472 				hio->hio_errors[ncomp] = errno;
1473 				reqlog(LOG_WARNING, 0, ggio,
1474 				    "Local request failed (%s): ",
1475 				    strerror(errno));
1476 			} else {
1477 				hio->hio_errors[ncomp] = 0;
1478 			}
1479 			break;
1480 		}
1481 		if (ISMEMSYNCWRITE(hio)) {
1482 			if (refcnt_release(&hio->hio_writecount) == 0) {
1483 				write_complete(res, hio);
1484 			}
1485 		}
1486 		if (refcnt_release(&hio->hio_countdown) > 0)
1487 			continue;
1488 		if (ISSYNCREQ(hio)) {
1489 			mtx_lock(&sync_lock);
1490 			SYNCREQDONE(hio);
1491 			mtx_unlock(&sync_lock);
1492 			cv_signal(&sync_cond);
1493 		} else {
1494 			pjdlog_debug(2,
1495 			    "local_send: (%p) Moving request to the done queue.",
1496 			    hio);
1497 			QUEUE_INSERT2(hio, done);
1498 		}
1499 	}
1500 	/* NOTREACHED */
1501 	return (NULL);
1502 }
1503 
1504 static void
1505 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1506 {
1507 	struct nv *nv;
1508 
1509 	rw_rlock(&hio_remote_lock[ncomp]);
1510 
1511 	if (!ISCONNECTED(res, ncomp)) {
1512 		rw_unlock(&hio_remote_lock[ncomp]);
1513 		return;
1514 	}
1515 
1516 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1517 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1518 
1519 	nv = nv_alloc();
1520 	nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1521 	if (nv_error(nv) != 0) {
1522 		rw_unlock(&hio_remote_lock[ncomp]);
1523 		nv_free(nv);
1524 		pjdlog_debug(1,
1525 		    "keepalive_send: Unable to prepare header to send.");
1526 		return;
1527 	}
1528 	if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) == -1) {
1529 		rw_unlock(&hio_remote_lock[ncomp]);
1530 		pjdlog_common(LOG_DEBUG, 1, errno,
1531 		    "keepalive_send: Unable to send request");
1532 		nv_free(nv);
1533 		remote_close(res, ncomp);
1534 		return;
1535 	}
1536 
1537 	rw_unlock(&hio_remote_lock[ncomp]);
1538 	nv_free(nv);
1539 	pjdlog_debug(2, "keepalive_send: Request sent.");
1540 }
1541 
1542 /*
1543  * Thread sends request to secondary node.
1544  */
1545 static void *
1546 remote_send_thread(void *arg)
1547 {
1548 	struct hast_resource *res = arg;
1549 	struct g_gate_ctl_io *ggio;
1550 	time_t lastcheck, now;
1551 	struct hio *hio;
1552 	struct nv *nv;
1553 	unsigned int ncomp;
1554 	bool wakeup;
1555 	uint64_t offset, length;
1556 	uint8_t cmd;
1557 	void *data;
1558 
1559 	/* Remote component is 1 for now. */
1560 	ncomp = 1;
1561 	lastcheck = time(NULL);
1562 
1563 	for (;;) {
1564 		pjdlog_debug(2, "remote_send: Taking request.");
1565 		QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE);
1566 		if (hio == NULL) {
1567 			now = time(NULL);
1568 			if (lastcheck + HAST_KEEPALIVE <= now) {
1569 				keepalive_send(res, ncomp);
1570 				lastcheck = now;
1571 			}
1572 			continue;
1573 		}
1574 		pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1575 		ggio = &hio->hio_ggio;
1576 		switch (ggio->gctl_cmd) {
1577 		case BIO_READ:
1578 			cmd = HIO_READ;
1579 			data = NULL;
1580 			offset = ggio->gctl_offset;
1581 			length = ggio->gctl_length;
1582 			break;
1583 		case BIO_WRITE:
1584 			cmd = HIO_WRITE;
1585 			data = ggio->gctl_data;
1586 			offset = ggio->gctl_offset;
1587 			length = ggio->gctl_length;
1588 			break;
1589 		case BIO_DELETE:
1590 			cmd = HIO_DELETE;
1591 			data = NULL;
1592 			offset = ggio->gctl_offset;
1593 			length = ggio->gctl_length;
1594 			break;
1595 		case BIO_FLUSH:
1596 			cmd = HIO_FLUSH;
1597 			data = NULL;
1598 			offset = 0;
1599 			length = 0;
1600 			break;
1601 		default:
1602 			PJDLOG_ABORT("invalid condition");
1603 		}
1604 		nv = nv_alloc();
1605 		nv_add_uint8(nv, cmd, "cmd");
1606 		nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1607 		nv_add_uint64(nv, offset, "offset");
1608 		nv_add_uint64(nv, length, "length");
1609 		if (ISMEMSYNCWRITE(hio))
1610 			nv_add_uint8(nv, 1, "memsync");
1611 		if (nv_error(nv) != 0) {
1612 			hio->hio_errors[ncomp] = nv_error(nv);
1613 			pjdlog_debug(2,
1614 			    "remote_send: (%p) Unable to prepare header to send.",
1615 			    hio);
1616 			reqlog(LOG_ERR, 0, ggio,
1617 			    "Unable to prepare header to send (%s): ",
1618 			    strerror(nv_error(nv)));
1619 			/* Move failed request immediately to the done queue. */
1620 			goto done_queue;
1621 		}
1622 		/*
1623 		 * Protect connection from disappearing.
1624 		 */
1625 		rw_rlock(&hio_remote_lock[ncomp]);
1626 		if (!ISCONNECTED(res, ncomp)) {
1627 			rw_unlock(&hio_remote_lock[ncomp]);
1628 			hio->hio_errors[ncomp] = ENOTCONN;
1629 			goto done_queue;
1630 		}
1631 		/*
1632 		 * Move the request to recv queue before sending it, because
1633 		 * in different order we can get reply before we move request
1634 		 * to recv queue.
1635 		 */
1636 		pjdlog_debug(2,
1637 		    "remote_send: (%p) Moving request to the recv queue.",
1638 		    hio);
1639 		mtx_lock(&hio_recv_list_lock[ncomp]);
1640 		wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1641 		TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1642 		hio_recv_list_size[ncomp]++;
1643 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1644 		if (hast_proto_send(res, res->hr_remoteout, nv, data,
1645 		    data != NULL ? length : 0) == -1) {
1646 			hio->hio_errors[ncomp] = errno;
1647 			rw_unlock(&hio_remote_lock[ncomp]);
1648 			pjdlog_debug(2,
1649 			    "remote_send: (%p) Unable to send request.", hio);
1650 			reqlog(LOG_ERR, 0, ggio,
1651 			    "Unable to send request (%s): ",
1652 			    strerror(hio->hio_errors[ncomp]));
1653 			remote_close(res, ncomp);
1654 		} else {
1655 			rw_unlock(&hio_remote_lock[ncomp]);
1656 		}
1657 		nv_free(nv);
1658 		if (wakeup)
1659 			cv_signal(&hio_recv_list_cond[ncomp]);
1660 		continue;
1661 done_queue:
1662 		nv_free(nv);
1663 		if (ISSYNCREQ(hio)) {
1664 			if (refcnt_release(&hio->hio_countdown) > 0)
1665 				continue;
1666 			mtx_lock(&sync_lock);
1667 			SYNCREQDONE(hio);
1668 			mtx_unlock(&sync_lock);
1669 			cv_signal(&sync_cond);
1670 			continue;
1671 		}
1672 		if (ggio->gctl_cmd == BIO_WRITE) {
1673 			mtx_lock(&res->hr_amp_lock);
1674 			if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1675 			    ggio->gctl_length)) {
1676 				(void)hast_activemap_flush(res);
1677 			} else {
1678 				mtx_unlock(&res->hr_amp_lock);
1679 			}
1680 			if (ISMEMSYNCWRITE(hio)) {
1681 				if (refcnt_release(&hio->hio_writecount) == 0) {
1682 					if (hio->hio_errors[0] == 0)
1683 						write_complete(res, hio);
1684 				}
1685 			}
1686 		}
1687 		if (refcnt_release(&hio->hio_countdown) > 0)
1688 			continue;
1689 		pjdlog_debug(2,
1690 		    "remote_send: (%p) Moving request to the done queue.",
1691 		    hio);
1692 		QUEUE_INSERT2(hio, done);
1693 	}
1694 	/* NOTREACHED */
1695 	return (NULL);
1696 }
1697 
1698 /*
1699  * Thread receives answer from secondary node and passes it to ggate_send
1700  * thread.
1701  */
1702 static void *
1703 remote_recv_thread(void *arg)
1704 {
1705 	struct hast_resource *res = arg;
1706 	struct g_gate_ctl_io *ggio;
1707 	struct hio *hio;
1708 	struct nv *nv;
1709 	unsigned int ncomp;
1710 	uint64_t seq;
1711 	bool memsyncack;
1712 	int error;
1713 
1714 	/* Remote component is 1 for now. */
1715 	ncomp = 1;
1716 
1717 	for (;;) {
1718 		/* Wait until there is anything to receive. */
1719 		mtx_lock(&hio_recv_list_lock[ncomp]);
1720 		while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1721 			pjdlog_debug(2, "remote_recv: No requests, waiting.");
1722 			cv_wait(&hio_recv_list_cond[ncomp],
1723 			    &hio_recv_list_lock[ncomp]);
1724 		}
1725 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1726 
1727 		memsyncack = false;
1728 
1729 		rw_rlock(&hio_remote_lock[ncomp]);
1730 		if (!ISCONNECTED(res, ncomp)) {
1731 			rw_unlock(&hio_remote_lock[ncomp]);
1732 			/*
1733 			 * Connection is dead, so move all pending requests to
1734 			 * the done queue (one-by-one).
1735 			 */
1736 			mtx_lock(&hio_recv_list_lock[ncomp]);
1737 			hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1738 			PJDLOG_ASSERT(hio != NULL);
1739 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1740 			    hio_next[ncomp]);
1741 			hio_recv_list_size[ncomp]--;
1742 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1743 			hio->hio_errors[ncomp] = ENOTCONN;
1744 			goto done_queue;
1745 		}
1746 		if (hast_proto_recv_hdr(res->hr_remotein, &nv) == -1) {
1747 			pjdlog_errno(LOG_ERR,
1748 			    "Unable to receive reply header");
1749 			rw_unlock(&hio_remote_lock[ncomp]);
1750 			remote_close(res, ncomp);
1751 			continue;
1752 		}
1753 		rw_unlock(&hio_remote_lock[ncomp]);
1754 		seq = nv_get_uint64(nv, "seq");
1755 		if (seq == 0) {
1756 			pjdlog_error("Header contains no 'seq' field.");
1757 			nv_free(nv);
1758 			continue;
1759 		}
1760 		memsyncack = nv_exists(nv, "received");
1761 		mtx_lock(&hio_recv_list_lock[ncomp]);
1762 		TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1763 			if (hio->hio_ggio.gctl_seq == seq) {
1764 				TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1765 				    hio_next[ncomp]);
1766 				hio_recv_list_size[ncomp]--;
1767 				break;
1768 			}
1769 		}
1770 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1771 		if (hio == NULL) {
1772 			pjdlog_error("Found no request matching received 'seq' field (%ju).",
1773 			    (uintmax_t)seq);
1774 			nv_free(nv);
1775 			continue;
1776 		}
1777 		ggio = &hio->hio_ggio;
1778 		error = nv_get_int16(nv, "error");
1779 		if (error != 0) {
1780 			/* Request failed on remote side. */
1781 			hio->hio_errors[ncomp] = error;
1782 			reqlog(LOG_WARNING, 0, ggio,
1783 			    "Remote request failed (%s): ", strerror(error));
1784 			nv_free(nv);
1785 			goto done_queue;
1786 		}
1787 		switch (ggio->gctl_cmd) {
1788 		case BIO_READ:
1789 			rw_rlock(&hio_remote_lock[ncomp]);
1790 			if (!ISCONNECTED(res, ncomp)) {
1791 				rw_unlock(&hio_remote_lock[ncomp]);
1792 				nv_free(nv);
1793 				goto done_queue;
1794 			}
1795 			if (hast_proto_recv_data(res, res->hr_remotein, nv,
1796 			    ggio->gctl_data, ggio->gctl_length) == -1) {
1797 				hio->hio_errors[ncomp] = errno;
1798 				pjdlog_errno(LOG_ERR,
1799 				    "Unable to receive reply data");
1800 				rw_unlock(&hio_remote_lock[ncomp]);
1801 				nv_free(nv);
1802 				remote_close(res, ncomp);
1803 				goto done_queue;
1804 			}
1805 			rw_unlock(&hio_remote_lock[ncomp]);
1806 			break;
1807 		case BIO_WRITE:
1808 		case BIO_DELETE:
1809 		case BIO_FLUSH:
1810 			break;
1811 		default:
1812 			PJDLOG_ABORT("invalid condition");
1813 		}
1814 		hio->hio_errors[ncomp] = 0;
1815 		nv_free(nv);
1816 done_queue:
1817 		if (ISMEMSYNCWRITE(hio)) {
1818 			if (!hio->hio_memsyncacked) {
1819 				PJDLOG_ASSERT(memsyncack ||
1820 				    hio->hio_errors[ncomp] != 0);
1821 				/* Remote ack arrived. */
1822 				if (refcnt_release(&hio->hio_writecount) == 0) {
1823 					if (hio->hio_errors[0] == 0)
1824 						write_complete(res, hio);
1825 				}
1826 				hio->hio_memsyncacked = true;
1827 				if (hio->hio_errors[ncomp] == 0) {
1828 					pjdlog_debug(2,
1829 					    "remote_recv: (%p) Moving request "
1830 					    "back to the recv queue.", hio);
1831 					mtx_lock(&hio_recv_list_lock[ncomp]);
1832 					TAILQ_INSERT_TAIL(&hio_recv_list[ncomp],
1833 					    hio, hio_next[ncomp]);
1834 					hio_recv_list_size[ncomp]++;
1835 					mtx_unlock(&hio_recv_list_lock[ncomp]);
1836 					continue;
1837 				}
1838 			} else {
1839 				PJDLOG_ASSERT(!memsyncack);
1840 				/* Remote final reply arrived. */
1841 			}
1842 		}
1843 		if (refcnt_release(&hio->hio_countdown) > 0)
1844 			continue;
1845 		if (ISSYNCREQ(hio)) {
1846 			mtx_lock(&sync_lock);
1847 			SYNCREQDONE(hio);
1848 			mtx_unlock(&sync_lock);
1849 			cv_signal(&sync_cond);
1850 		} else {
1851 			pjdlog_debug(2,
1852 			    "remote_recv: (%p) Moving request to the done queue.",
1853 			    hio);
1854 			QUEUE_INSERT2(hio, done);
1855 		}
1856 	}
1857 	/* NOTREACHED */
1858 	return (NULL);
1859 }
1860 
1861 /*
1862  * Thread sends answer to the kernel.
1863  */
1864 static void *
1865 ggate_send_thread(void *arg)
1866 {
1867 	struct hast_resource *res = arg;
1868 	struct g_gate_ctl_io *ggio;
1869 	struct hio *hio;
1870 	unsigned int ii, ncomps;
1871 
1872 	ncomps = HAST_NCOMPONENTS;
1873 
1874 	for (;;) {
1875 		pjdlog_debug(2, "ggate_send: Taking request.");
1876 		QUEUE_TAKE2(hio, done);
1877 		pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1878 		ggio = &hio->hio_ggio;
1879 		for (ii = 0; ii < ncomps; ii++) {
1880 			if (hio->hio_errors[ii] == 0) {
1881 				/*
1882 				 * One successful request is enough to declare
1883 				 * success.
1884 				 */
1885 				ggio->gctl_error = 0;
1886 				break;
1887 			}
1888 		}
1889 		if (ii == ncomps) {
1890 			/*
1891 			 * None of the requests were successful.
1892 			 * Use the error from local component except the
1893 			 * case when we did only remote request.
1894 			 */
1895 			if (ggio->gctl_cmd == BIO_READ &&
1896 			    res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
1897 				ggio->gctl_error = hio->hio_errors[1];
1898 			else
1899 				ggio->gctl_error = hio->hio_errors[0];
1900 		}
1901 		if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1902 			mtx_lock(&res->hr_amp_lock);
1903 			if (activemap_write_complete(res->hr_amp,
1904 			    ggio->gctl_offset, ggio->gctl_length)) {
1905 				res->hr_stat_activemap_update++;
1906 				(void)hast_activemap_flush(res);
1907 			} else {
1908 				mtx_unlock(&res->hr_amp_lock);
1909 			}
1910 		}
1911 		if (ggio->gctl_cmd == BIO_WRITE) {
1912 			/*
1913 			 * Unlock range we locked.
1914 			 */
1915 			mtx_lock(&range_lock);
1916 			rangelock_del(range_regular, ggio->gctl_offset,
1917 			    ggio->gctl_length);
1918 			if (range_sync_wait)
1919 				cv_signal(&range_sync_cond);
1920 			mtx_unlock(&range_lock);
1921 			if (!hio->hio_done)
1922 				write_complete(res, hio);
1923 		} else {
1924 			if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1) {
1925 				primary_exit(EX_OSERR,
1926 				    "G_GATE_CMD_DONE failed");
1927 			}
1928 		}
1929 		if (hio->hio_errors[0]) {
1930 			switch (ggio->gctl_cmd) {
1931 			case BIO_READ:
1932 				res->hr_stat_read_error++;
1933 				break;
1934 			case BIO_WRITE:
1935 				res->hr_stat_write_error++;
1936 				break;
1937 			case BIO_DELETE:
1938 				res->hr_stat_delete_error++;
1939 				break;
1940 			case BIO_FLUSH:
1941 				res->hr_stat_flush_error++;
1942 				break;
1943 			}
1944 		}
1945 		pjdlog_debug(2,
1946 		    "ggate_send: (%p) Moving request to the free queue.", hio);
1947 		QUEUE_INSERT2(hio, free);
1948 	}
1949 	/* NOTREACHED */
1950 	return (NULL);
1951 }
1952 
1953 /*
1954  * Thread synchronize local and remote components.
1955  */
1956 static void *
1957 sync_thread(void *arg __unused)
1958 {
1959 	struct hast_resource *res = arg;
1960 	struct hio *hio;
1961 	struct g_gate_ctl_io *ggio;
1962 	struct timeval tstart, tend, tdiff;
1963 	unsigned int ii, ncomp, ncomps;
1964 	off_t offset, length, synced;
1965 	bool dorewind, directreads;
1966 	int syncext;
1967 
1968 	ncomps = HAST_NCOMPONENTS;
1969 	dorewind = true;
1970 	synced = 0;
1971 	offset = -1;
1972 	directreads = false;
1973 
1974 	for (;;) {
1975 		mtx_lock(&sync_lock);
1976 		if (offset >= 0 && !sync_inprogress) {
1977 			gettimeofday(&tend, NULL);
1978 			timersub(&tend, &tstart, &tdiff);
1979 			pjdlog_info("Synchronization interrupted after %#.0T. "
1980 			    "%NB synchronized so far.", &tdiff,
1981 			    (intmax_t)synced);
1982 			event_send(res, EVENT_SYNCINTR);
1983 		}
1984 		while (!sync_inprogress) {
1985 			dorewind = true;
1986 			synced = 0;
1987 			cv_wait(&sync_cond, &sync_lock);
1988 		}
1989 		mtx_unlock(&sync_lock);
1990 		/*
1991 		 * Obtain offset at which we should synchronize.
1992 		 * Rewind synchronization if needed.
1993 		 */
1994 		mtx_lock(&res->hr_amp_lock);
1995 		if (dorewind)
1996 			activemap_sync_rewind(res->hr_amp);
1997 		offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1998 		if (syncext != -1) {
1999 			/*
2000 			 * We synchronized entire syncext extent, we can mark
2001 			 * it as clean now.
2002 			 */
2003 			if (activemap_extent_complete(res->hr_amp, syncext))
2004 				(void)hast_activemap_flush(res);
2005 			else
2006 				mtx_unlock(&res->hr_amp_lock);
2007 		} else {
2008 			mtx_unlock(&res->hr_amp_lock);
2009 		}
2010 		if (dorewind) {
2011 			dorewind = false;
2012 			if (offset == -1)
2013 				pjdlog_info("Nodes are in sync.");
2014 			else {
2015 				pjdlog_info("Synchronization started. %NB to go.",
2016 				    (intmax_t)(res->hr_extentsize *
2017 				    activemap_ndirty(res->hr_amp)));
2018 				event_send(res, EVENT_SYNCSTART);
2019 				gettimeofday(&tstart, NULL);
2020 			}
2021 		}
2022 		if (offset == -1) {
2023 			sync_stop();
2024 			pjdlog_debug(1, "Nothing to synchronize.");
2025 			/*
2026 			 * Synchronization complete, make both localcnt and
2027 			 * remotecnt equal.
2028 			 */
2029 			ncomp = 1;
2030 			rw_rlock(&hio_remote_lock[ncomp]);
2031 			if (ISCONNECTED(res, ncomp)) {
2032 				if (synced > 0) {
2033 					int64_t bps;
2034 
2035 					gettimeofday(&tend, NULL);
2036 					timersub(&tend, &tstart, &tdiff);
2037 					bps = (int64_t)((double)synced /
2038 					    ((double)tdiff.tv_sec +
2039 					    (double)tdiff.tv_usec / 1000000));
2040 					pjdlog_info("Synchronization complete. "
2041 					    "%NB synchronized in %#.0lT (%NB/sec).",
2042 					    (intmax_t)synced, &tdiff,
2043 					    (intmax_t)bps);
2044 					event_send(res, EVENT_SYNCDONE);
2045 				}
2046 				mtx_lock(&metadata_lock);
2047 				if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
2048 					directreads = true;
2049 				res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
2050 				res->hr_primary_localcnt =
2051 				    res->hr_secondary_remotecnt;
2052 				res->hr_primary_remotecnt =
2053 				    res->hr_secondary_localcnt;
2054 				pjdlog_debug(1,
2055 				    "Setting localcnt to %ju and remotecnt to %ju.",
2056 				    (uintmax_t)res->hr_primary_localcnt,
2057 				    (uintmax_t)res->hr_primary_remotecnt);
2058 				(void)metadata_write(res);
2059 				mtx_unlock(&metadata_lock);
2060 			}
2061 			rw_unlock(&hio_remote_lock[ncomp]);
2062 			if (directreads) {
2063 				directreads = false;
2064 				enable_direct_reads(res);
2065 			}
2066 			continue;
2067 		}
2068 		pjdlog_debug(2, "sync: Taking free request.");
2069 		QUEUE_TAKE2(hio, free);
2070 		pjdlog_debug(2, "sync: (%p) Got free request.", hio);
2071 		/*
2072 		 * Lock the range we are going to synchronize. We don't want
2073 		 * race where someone writes between our read and write.
2074 		 */
2075 		for (;;) {
2076 			mtx_lock(&range_lock);
2077 			if (rangelock_islocked(range_regular, offset, length)) {
2078 				pjdlog_debug(2,
2079 				    "sync: Range offset=%jd length=%jd locked.",
2080 				    (intmax_t)offset, (intmax_t)length);
2081 				range_sync_wait = true;
2082 				cv_wait(&range_sync_cond, &range_lock);
2083 				range_sync_wait = false;
2084 				mtx_unlock(&range_lock);
2085 				continue;
2086 			}
2087 			if (rangelock_add(range_sync, offset, length) == -1) {
2088 				mtx_unlock(&range_lock);
2089 				pjdlog_debug(2,
2090 				    "sync: Range offset=%jd length=%jd is already locked, waiting.",
2091 				    (intmax_t)offset, (intmax_t)length);
2092 				sleep(1);
2093 				continue;
2094 			}
2095 			mtx_unlock(&range_lock);
2096 			break;
2097 		}
2098 		/*
2099 		 * First read the data from synchronization source.
2100 		 */
2101 		SYNCREQ(hio);
2102 		ggio = &hio->hio_ggio;
2103 		ggio->gctl_cmd = BIO_READ;
2104 		ggio->gctl_offset = offset;
2105 		ggio->gctl_length = length;
2106 		ggio->gctl_error = 0;
2107 		hio->hio_done = false;
2108 		hio->hio_replication = res->hr_replication;
2109 		for (ii = 0; ii < ncomps; ii++)
2110 			hio->hio_errors[ii] = EINVAL;
2111 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2112 		    hio);
2113 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2114 		    hio);
2115 		mtx_lock(&metadata_lock);
2116 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2117 			/*
2118 			 * This range is up-to-date on local component,
2119 			 * so handle request locally.
2120 			 */
2121 			 /* Local component is 0 for now. */
2122 			ncomp = 0;
2123 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2124 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2125 			/*
2126 			 * This range is out-of-date on local component,
2127 			 * so send request to the remote node.
2128 			 */
2129 			 /* Remote component is 1 for now. */
2130 			ncomp = 1;
2131 		}
2132 		mtx_unlock(&metadata_lock);
2133 		refcnt_init(&hio->hio_countdown, 1);
2134 		QUEUE_INSERT1(hio, send, ncomp);
2135 
2136 		/*
2137 		 * Let's wait for READ to finish.
2138 		 */
2139 		mtx_lock(&sync_lock);
2140 		while (!ISSYNCREQDONE(hio))
2141 			cv_wait(&sync_cond, &sync_lock);
2142 		mtx_unlock(&sync_lock);
2143 
2144 		if (hio->hio_errors[ncomp] != 0) {
2145 			pjdlog_error("Unable to read synchronization data: %s.",
2146 			    strerror(hio->hio_errors[ncomp]));
2147 			goto free_queue;
2148 		}
2149 
2150 		/*
2151 		 * We read the data from synchronization source, now write it
2152 		 * to synchronization target.
2153 		 */
2154 		SYNCREQ(hio);
2155 		ggio->gctl_cmd = BIO_WRITE;
2156 		for (ii = 0; ii < ncomps; ii++)
2157 			hio->hio_errors[ii] = EINVAL;
2158 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2159 		    hio);
2160 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2161 		    hio);
2162 		mtx_lock(&metadata_lock);
2163 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2164 			/*
2165 			 * This range is up-to-date on local component,
2166 			 * so we update remote component.
2167 			 */
2168 			 /* Remote component is 1 for now. */
2169 			ncomp = 1;
2170 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2171 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2172 			/*
2173 			 * This range is out-of-date on local component,
2174 			 * so we update it.
2175 			 */
2176 			 /* Local component is 0 for now. */
2177 			ncomp = 0;
2178 		}
2179 		mtx_unlock(&metadata_lock);
2180 
2181 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2182 		    hio);
2183 		refcnt_init(&hio->hio_countdown, 1);
2184 		QUEUE_INSERT1(hio, send, ncomp);
2185 
2186 		/*
2187 		 * Let's wait for WRITE to finish.
2188 		 */
2189 		mtx_lock(&sync_lock);
2190 		while (!ISSYNCREQDONE(hio))
2191 			cv_wait(&sync_cond, &sync_lock);
2192 		mtx_unlock(&sync_lock);
2193 
2194 		if (hio->hio_errors[ncomp] != 0) {
2195 			pjdlog_error("Unable to write synchronization data: %s.",
2196 			    strerror(hio->hio_errors[ncomp]));
2197 			goto free_queue;
2198 		}
2199 
2200 		synced += length;
2201 free_queue:
2202 		mtx_lock(&range_lock);
2203 		rangelock_del(range_sync, offset, length);
2204 		if (range_regular_wait)
2205 			cv_signal(&range_regular_cond);
2206 		mtx_unlock(&range_lock);
2207 		pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
2208 		    hio);
2209 		QUEUE_INSERT2(hio, free);
2210 	}
2211 	/* NOTREACHED */
2212 	return (NULL);
2213 }
2214 
2215 void
2216 primary_config_reload(struct hast_resource *res, struct nv *nv)
2217 {
2218 	unsigned int ii, ncomps;
2219 	int modified, vint;
2220 	const char *vstr;
2221 
2222 	pjdlog_info("Reloading configuration...");
2223 
2224 	PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
2225 	PJDLOG_ASSERT(gres == res);
2226 	nv_assert(nv, "remoteaddr");
2227 	nv_assert(nv, "sourceaddr");
2228 	nv_assert(nv, "replication");
2229 	nv_assert(nv, "checksum");
2230 	nv_assert(nv, "compression");
2231 	nv_assert(nv, "timeout");
2232 	nv_assert(nv, "exec");
2233 	nv_assert(nv, "metaflush");
2234 
2235 	ncomps = HAST_NCOMPONENTS;
2236 
2237 #define MODIFIED_REMOTEADDR	0x01
2238 #define MODIFIED_SOURCEADDR	0x02
2239 #define MODIFIED_REPLICATION	0x04
2240 #define MODIFIED_CHECKSUM	0x08
2241 #define MODIFIED_COMPRESSION	0x10
2242 #define MODIFIED_TIMEOUT	0x20
2243 #define MODIFIED_EXEC		0x40
2244 #define MODIFIED_METAFLUSH	0x80
2245 	modified = 0;
2246 
2247 	vstr = nv_get_string(nv, "remoteaddr");
2248 	if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
2249 		/*
2250 		 * Don't copy res->hr_remoteaddr to gres just yet.
2251 		 * We want remote_close() to log disconnect from the old
2252 		 * addresses, not from the new ones.
2253 		 */
2254 		modified |= MODIFIED_REMOTEADDR;
2255 	}
2256 	vstr = nv_get_string(nv, "sourceaddr");
2257 	if (strcmp(gres->hr_sourceaddr, vstr) != 0) {
2258 		strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr));
2259 		modified |= MODIFIED_SOURCEADDR;
2260 	}
2261 	vint = nv_get_int32(nv, "replication");
2262 	if (gres->hr_replication != vint) {
2263 		gres->hr_replication = vint;
2264 		modified |= MODIFIED_REPLICATION;
2265 	}
2266 	vint = nv_get_int32(nv, "checksum");
2267 	if (gres->hr_checksum != vint) {
2268 		gres->hr_checksum = vint;
2269 		modified |= MODIFIED_CHECKSUM;
2270 	}
2271 	vint = nv_get_int32(nv, "compression");
2272 	if (gres->hr_compression != vint) {
2273 		gres->hr_compression = vint;
2274 		modified |= MODIFIED_COMPRESSION;
2275 	}
2276 	vint = nv_get_int32(nv, "timeout");
2277 	if (gres->hr_timeout != vint) {
2278 		gres->hr_timeout = vint;
2279 		modified |= MODIFIED_TIMEOUT;
2280 	}
2281 	vstr = nv_get_string(nv, "exec");
2282 	if (strcmp(gres->hr_exec, vstr) != 0) {
2283 		strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
2284 		modified |= MODIFIED_EXEC;
2285 	}
2286 	vint = nv_get_int32(nv, "metaflush");
2287 	if (gres->hr_metaflush != vint) {
2288 		gres->hr_metaflush = vint;
2289 		modified |= MODIFIED_METAFLUSH;
2290 	}
2291 
2292 	/*
2293 	 * Change timeout for connected sockets.
2294 	 * Don't bother if we need to reconnect.
2295 	 */
2296 	if ((modified & MODIFIED_TIMEOUT) != 0 &&
2297 	    (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) == 0) {
2298 		for (ii = 0; ii < ncomps; ii++) {
2299 			if (!ISREMOTE(ii))
2300 				continue;
2301 			rw_rlock(&hio_remote_lock[ii]);
2302 			if (!ISCONNECTED(gres, ii)) {
2303 				rw_unlock(&hio_remote_lock[ii]);
2304 				continue;
2305 			}
2306 			rw_unlock(&hio_remote_lock[ii]);
2307 			if (proto_timeout(gres->hr_remotein,
2308 			    gres->hr_timeout) == -1) {
2309 				pjdlog_errno(LOG_WARNING,
2310 				    "Unable to set connection timeout");
2311 			}
2312 			if (proto_timeout(gres->hr_remoteout,
2313 			    gres->hr_timeout) == -1) {
2314 				pjdlog_errno(LOG_WARNING,
2315 				    "Unable to set connection timeout");
2316 			}
2317 		}
2318 	}
2319 	if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) != 0) {
2320 		for (ii = 0; ii < ncomps; ii++) {
2321 			if (!ISREMOTE(ii))
2322 				continue;
2323 			remote_close(gres, ii);
2324 		}
2325 		if (modified & MODIFIED_REMOTEADDR) {
2326 			vstr = nv_get_string(nv, "remoteaddr");
2327 			strlcpy(gres->hr_remoteaddr, vstr,
2328 			    sizeof(gres->hr_remoteaddr));
2329 		}
2330 	}
2331 #undef	MODIFIED_REMOTEADDR
2332 #undef	MODIFIED_SOURCEADDR
2333 #undef	MODIFIED_REPLICATION
2334 #undef	MODIFIED_CHECKSUM
2335 #undef	MODIFIED_COMPRESSION
2336 #undef	MODIFIED_TIMEOUT
2337 #undef	MODIFIED_EXEC
2338 #undef	MODIFIED_METAFLUSH
2339 
2340 	pjdlog_info("Configuration reloaded successfully.");
2341 }
2342 
2343 static void
2344 guard_one(struct hast_resource *res, unsigned int ncomp)
2345 {
2346 	struct proto_conn *in, *out;
2347 
2348 	if (!ISREMOTE(ncomp))
2349 		return;
2350 
2351 	rw_rlock(&hio_remote_lock[ncomp]);
2352 
2353 	if (!real_remote(res)) {
2354 		rw_unlock(&hio_remote_lock[ncomp]);
2355 		return;
2356 	}
2357 
2358 	if (ISCONNECTED(res, ncomp)) {
2359 		PJDLOG_ASSERT(res->hr_remotein != NULL);
2360 		PJDLOG_ASSERT(res->hr_remoteout != NULL);
2361 		rw_unlock(&hio_remote_lock[ncomp]);
2362 		pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2363 		    res->hr_remoteaddr);
2364 		return;
2365 	}
2366 
2367 	PJDLOG_ASSERT(res->hr_remotein == NULL);
2368 	PJDLOG_ASSERT(res->hr_remoteout == NULL);
2369 	/*
2370 	 * Upgrade the lock. It doesn't have to be atomic as no other thread
2371 	 * can change connection status from disconnected to connected.
2372 	 */
2373 	rw_unlock(&hio_remote_lock[ncomp]);
2374 	pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2375 	    res->hr_remoteaddr);
2376 	in = out = NULL;
2377 	if (init_remote(res, &in, &out) == 0) {
2378 		rw_wlock(&hio_remote_lock[ncomp]);
2379 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2380 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2381 		PJDLOG_ASSERT(in != NULL && out != NULL);
2382 		res->hr_remotein = in;
2383 		res->hr_remoteout = out;
2384 		rw_unlock(&hio_remote_lock[ncomp]);
2385 		pjdlog_info("Successfully reconnected to %s.",
2386 		    res->hr_remoteaddr);
2387 		sync_start();
2388 	} else {
2389 		/* Both connections should be NULL. */
2390 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2391 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2392 		PJDLOG_ASSERT(in == NULL && out == NULL);
2393 		pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2394 		    res->hr_remoteaddr);
2395 	}
2396 }
2397 
2398 /*
2399  * Thread guards remote connections and reconnects when needed, handles
2400  * signals, etc.
2401  */
2402 static void *
2403 guard_thread(void *arg)
2404 {
2405 	struct hast_resource *res = arg;
2406 	unsigned int ii, ncomps;
2407 	struct timespec timeout;
2408 	time_t lastcheck, now;
2409 	sigset_t mask;
2410 	int signo;
2411 
2412 	ncomps = HAST_NCOMPONENTS;
2413 	lastcheck = time(NULL);
2414 
2415 	PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2416 	PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2417 	PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2418 
2419 	timeout.tv_sec = HAST_KEEPALIVE;
2420 	timeout.tv_nsec = 0;
2421 	signo = -1;
2422 
2423 	for (;;) {
2424 		switch (signo) {
2425 		case SIGINT:
2426 		case SIGTERM:
2427 			sigexit_received = true;
2428 			primary_exitx(EX_OK,
2429 			    "Termination signal received, exiting.");
2430 			break;
2431 		default:
2432 			break;
2433 		}
2434 
2435 		/*
2436 		 * Don't check connections until we fully started,
2437 		 * as we may still be looping, waiting for remote node
2438 		 * to switch from primary to secondary.
2439 		 */
2440 		if (fullystarted) {
2441 			pjdlog_debug(2, "remote_guard: Checking connections.");
2442 			now = time(NULL);
2443 			if (lastcheck + HAST_KEEPALIVE <= now) {
2444 				for (ii = 0; ii < ncomps; ii++)
2445 					guard_one(res, ii);
2446 				lastcheck = now;
2447 			}
2448 		}
2449 		signo = sigtimedwait(&mask, NULL, &timeout);
2450 	}
2451 	/* NOTREACHED */
2452 	return (NULL);
2453 }
2454