xref: /freebsd/sbin/hastd/primary.c (revision dadef94c7a762d05890e2891bc4a7d1dfe0cf758)
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)	do {				\
184 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
185 	while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL) { \
186 		cv_wait(&hio_##name##_list_cond[(ncomp)],		\
187 		    &hio_##name##_list_lock[(ncomp)]);			\
188 	}								\
189 	TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),		\
190 	    hio_next[(ncomp)]);						\
191 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
192 } while (0)
193 #define	QUEUE_TAKE2(hio, name)	do {					\
194 	mtx_lock(&hio_##name##_list_lock);				\
195 	while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {	\
196 		cv_wait(&hio_##name##_list_cond,			\
197 		    &hio_##name##_list_lock);				\
198 	}								\
199 	TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);	\
200 	mtx_unlock(&hio_##name##_list_lock);				\
201 } while (0)
202 
203 #define	SYNCREQ(hio)		do {					\
204 	(hio)->hio_ggio.gctl_unit = -1;					\
205 	(hio)->hio_ggio.gctl_seq = 1;					\
206 } while (0)
207 #define	ISSYNCREQ(hio)		((hio)->hio_ggio.gctl_unit == -1)
208 #define	SYNCREQDONE(hio)	do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
209 #define	ISSYNCREQDONE(hio)	((hio)->hio_ggio.gctl_unit == -2)
210 
211 static struct hast_resource *gres;
212 
213 static pthread_mutex_t range_lock;
214 static struct rangelocks *range_regular;
215 static bool range_regular_wait;
216 static pthread_cond_t range_regular_cond;
217 static struct rangelocks *range_sync;
218 static bool range_sync_wait;
219 static pthread_cond_t range_sync_cond;
220 
221 static void *ggate_recv_thread(void *arg);
222 static void *local_send_thread(void *arg);
223 static void *remote_send_thread(void *arg);
224 static void *remote_recv_thread(void *arg);
225 static void *ggate_send_thread(void *arg);
226 static void *sync_thread(void *arg);
227 static void *guard_thread(void *arg);
228 
229 static void
230 cleanup(struct hast_resource *res)
231 {
232 	int rerrno;
233 
234 	/* Remember errno. */
235 	rerrno = errno;
236 
237 	/* Destroy ggate provider if we created one. */
238 	if (res->hr_ggateunit >= 0) {
239 		struct g_gate_ctl_destroy ggiod;
240 
241 		bzero(&ggiod, sizeof(ggiod));
242 		ggiod.gctl_version = G_GATE_VERSION;
243 		ggiod.gctl_unit = res->hr_ggateunit;
244 		ggiod.gctl_force = 1;
245 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) {
246 			pjdlog_errno(LOG_WARNING,
247 			    "Unable to destroy hast/%s device",
248 			    res->hr_provname);
249 		}
250 		res->hr_ggateunit = -1;
251 	}
252 
253 	/* Restore errno. */
254 	errno = rerrno;
255 }
256 
257 static __dead2 void
258 primary_exit(int exitcode, const char *fmt, ...)
259 {
260 	va_list ap;
261 
262 	assert(exitcode != EX_OK);
263 	va_start(ap, fmt);
264 	pjdlogv_errno(LOG_ERR, fmt, ap);
265 	va_end(ap);
266 	cleanup(gres);
267 	exit(exitcode);
268 }
269 
270 static __dead2 void
271 primary_exitx(int exitcode, const char *fmt, ...)
272 {
273 	va_list ap;
274 
275 	va_start(ap, fmt);
276 	pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
277 	va_end(ap);
278 	cleanup(gres);
279 	exit(exitcode);
280 }
281 
282 static int
283 hast_activemap_flush(struct hast_resource *res)
284 {
285 	const unsigned char *buf;
286 	size_t size;
287 
288 	buf = activemap_bitmap(res->hr_amp, &size);
289 	assert(buf != NULL);
290 	assert((size % res->hr_local_sectorsize) == 0);
291 	if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
292 	    (ssize_t)size) {
293 		KEEP_ERRNO(pjdlog_errno(LOG_ERR,
294 		    "Unable to flush activemap to disk"));
295 		return (-1);
296 	}
297 	return (0);
298 }
299 
300 static bool
301 real_remote(const struct hast_resource *res)
302 {
303 
304 	return (strcmp(res->hr_remoteaddr, "none") != 0);
305 }
306 
307 static void
308 init_environment(struct hast_resource *res __unused)
309 {
310 	struct hio *hio;
311 	unsigned int ii, ncomps;
312 
313 	/*
314 	 * In the future it might be per-resource value.
315 	 */
316 	ncomps = HAST_NCOMPONENTS;
317 
318 	/*
319 	 * Allocate memory needed by lists.
320 	 */
321 	hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
322 	if (hio_send_list == NULL) {
323 		primary_exitx(EX_TEMPFAIL,
324 		    "Unable to allocate %zu bytes of memory for send lists.",
325 		    sizeof(hio_send_list[0]) * ncomps);
326 	}
327 	hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
328 	if (hio_send_list_lock == NULL) {
329 		primary_exitx(EX_TEMPFAIL,
330 		    "Unable to allocate %zu bytes of memory for send list locks.",
331 		    sizeof(hio_send_list_lock[0]) * ncomps);
332 	}
333 	hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
334 	if (hio_send_list_cond == NULL) {
335 		primary_exitx(EX_TEMPFAIL,
336 		    "Unable to allocate %zu bytes of memory for send list condition variables.",
337 		    sizeof(hio_send_list_cond[0]) * ncomps);
338 	}
339 	hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
340 	if (hio_recv_list == NULL) {
341 		primary_exitx(EX_TEMPFAIL,
342 		    "Unable to allocate %zu bytes of memory for recv lists.",
343 		    sizeof(hio_recv_list[0]) * ncomps);
344 	}
345 	hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
346 	if (hio_recv_list_lock == NULL) {
347 		primary_exitx(EX_TEMPFAIL,
348 		    "Unable to allocate %zu bytes of memory for recv list locks.",
349 		    sizeof(hio_recv_list_lock[0]) * ncomps);
350 	}
351 	hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
352 	if (hio_recv_list_cond == NULL) {
353 		primary_exitx(EX_TEMPFAIL,
354 		    "Unable to allocate %zu bytes of memory for recv list condition variables.",
355 		    sizeof(hio_recv_list_cond[0]) * ncomps);
356 	}
357 	hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
358 	if (hio_remote_lock == NULL) {
359 		primary_exitx(EX_TEMPFAIL,
360 		    "Unable to allocate %zu bytes of memory for remote connections locks.",
361 		    sizeof(hio_remote_lock[0]) * ncomps);
362 	}
363 
364 	/*
365 	 * Initialize lists, their locks and theirs condition variables.
366 	 */
367 	TAILQ_INIT(&hio_free_list);
368 	mtx_init(&hio_free_list_lock);
369 	cv_init(&hio_free_list_cond);
370 	for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
371 		TAILQ_INIT(&hio_send_list[ii]);
372 		mtx_init(&hio_send_list_lock[ii]);
373 		cv_init(&hio_send_list_cond[ii]);
374 		TAILQ_INIT(&hio_recv_list[ii]);
375 		mtx_init(&hio_recv_list_lock[ii]);
376 		cv_init(&hio_recv_list_cond[ii]);
377 		rw_init(&hio_remote_lock[ii]);
378 	}
379 	TAILQ_INIT(&hio_done_list);
380 	mtx_init(&hio_done_list_lock);
381 	cv_init(&hio_done_list_cond);
382 	mtx_init(&metadata_lock);
383 
384 	/*
385 	 * Allocate requests pool and initialize requests.
386 	 */
387 	for (ii = 0; ii < HAST_HIO_MAX; ii++) {
388 		hio = malloc(sizeof(*hio));
389 		if (hio == NULL) {
390 			primary_exitx(EX_TEMPFAIL,
391 			    "Unable to allocate %zu bytes of memory for hio request.",
392 			    sizeof(*hio));
393 		}
394 		hio->hio_countdown = 0;
395 		hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
396 		if (hio->hio_errors == NULL) {
397 			primary_exitx(EX_TEMPFAIL,
398 			    "Unable allocate %zu bytes of memory for hio errors.",
399 			    sizeof(hio->hio_errors[0]) * ncomps);
400 		}
401 		hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
402 		if (hio->hio_next == NULL) {
403 			primary_exitx(EX_TEMPFAIL,
404 			    "Unable allocate %zu bytes of memory for hio_next field.",
405 			    sizeof(hio->hio_next[0]) * ncomps);
406 		}
407 		hio->hio_ggio.gctl_version = G_GATE_VERSION;
408 		hio->hio_ggio.gctl_data = malloc(MAXPHYS);
409 		if (hio->hio_ggio.gctl_data == NULL) {
410 			primary_exitx(EX_TEMPFAIL,
411 			    "Unable to allocate %zu bytes of memory for gctl_data.",
412 			    MAXPHYS);
413 		}
414 		hio->hio_ggio.gctl_length = MAXPHYS;
415 		hio->hio_ggio.gctl_error = 0;
416 		TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
417 	}
418 }
419 
420 static void
421 init_local(struct hast_resource *res)
422 {
423 	unsigned char *buf;
424 	size_t mapsize;
425 
426 	if (metadata_read(res, true) < 0)
427 		exit(EX_NOINPUT);
428 	mtx_init(&res->hr_amp_lock);
429 	if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
430 	    res->hr_local_sectorsize, res->hr_keepdirty) < 0) {
431 		primary_exit(EX_TEMPFAIL, "Unable to create activemap");
432 	}
433 	mtx_init(&range_lock);
434 	cv_init(&range_regular_cond);
435 	if (rangelock_init(&range_regular) < 0)
436 		primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
437 	cv_init(&range_sync_cond);
438 	if (rangelock_init(&range_sync) < 0)
439 		primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
440 	mapsize = activemap_ondisk_size(res->hr_amp);
441 	buf = calloc(1, mapsize);
442 	if (buf == NULL) {
443 		primary_exitx(EX_TEMPFAIL,
444 		    "Unable to allocate buffer for activemap.");
445 	}
446 	if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
447 	    (ssize_t)mapsize) {
448 		primary_exit(EX_NOINPUT, "Unable to read activemap");
449 	}
450 	activemap_copyin(res->hr_amp, buf, mapsize);
451 	free(buf);
452 	if (res->hr_resuid != 0)
453 		return;
454 	/*
455 	 * We're using provider for the first time, so we have to generate
456 	 * resource unique identifier and initialize local and remote counts.
457 	 */
458 	arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
459 	res->hr_primary_localcnt = 1;
460 	res->hr_primary_remotecnt = 0;
461 	if (metadata_write(res) < 0)
462 		exit(EX_NOINPUT);
463 }
464 
465 static bool
466 init_remote(struct hast_resource *res, struct proto_conn **inp,
467     struct proto_conn **outp)
468 {
469 	struct proto_conn *in, *out;
470 	struct nv *nvout, *nvin;
471 	const unsigned char *token;
472 	unsigned char *map;
473 	const char *errmsg;
474 	int32_t extentsize;
475 	int64_t datasize;
476 	uint32_t mapsize;
477 	size_t size;
478 
479 	assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
480 	assert(real_remote(res));
481 
482 	in = out = NULL;
483 	errmsg = NULL;
484 
485 	/* Prepare outgoing connection with remote node. */
486 	if (proto_client(res->hr_remoteaddr, &out) < 0) {
487 		primary_exit(EX_TEMPFAIL, "Unable to create connection to %s",
488 		    res->hr_remoteaddr);
489 	}
490 	/* Try to connect, but accept failure. */
491 	if (proto_connect(out) < 0) {
492 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
493 		    res->hr_remoteaddr);
494 		goto close;
495 	}
496 	/* Error in setting timeout is not critical, but why should it fail? */
497 	if (proto_timeout(out, res->hr_timeout) < 0)
498 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
499 	/*
500 	 * First handshake step.
501 	 * Setup outgoing connection with remote node.
502 	 */
503 	nvout = nv_alloc();
504 	nv_add_string(nvout, res->hr_name, "resource");
505 	if (nv_error(nvout) != 0) {
506 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
507 		    "Unable to allocate header for connection with %s",
508 		    res->hr_remoteaddr);
509 		nv_free(nvout);
510 		goto close;
511 	}
512 	if (hast_proto_send(res, out, nvout, NULL, 0) < 0) {
513 		pjdlog_errno(LOG_WARNING,
514 		    "Unable to send handshake header to %s",
515 		    res->hr_remoteaddr);
516 		nv_free(nvout);
517 		goto close;
518 	}
519 	nv_free(nvout);
520 	if (hast_proto_recv_hdr(out, &nvin) < 0) {
521 		pjdlog_errno(LOG_WARNING,
522 		    "Unable to receive handshake header from %s",
523 		    res->hr_remoteaddr);
524 		goto close;
525 	}
526 	errmsg = nv_get_string(nvin, "errmsg");
527 	if (errmsg != NULL) {
528 		pjdlog_warning("%s", errmsg);
529 		nv_free(nvin);
530 		goto close;
531 	}
532 	token = nv_get_uint8_array(nvin, &size, "token");
533 	if (token == NULL) {
534 		pjdlog_warning("Handshake header from %s has no 'token' field.",
535 		    res->hr_remoteaddr);
536 		nv_free(nvin);
537 		goto close;
538 	}
539 	if (size != sizeof(res->hr_token)) {
540 		pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
541 		    res->hr_remoteaddr, size, sizeof(res->hr_token));
542 		nv_free(nvin);
543 		goto close;
544 	}
545 	bcopy(token, res->hr_token, sizeof(res->hr_token));
546 	nv_free(nvin);
547 
548 	/*
549 	 * Second handshake step.
550 	 * Setup incoming connection with remote node.
551 	 */
552 	if (proto_client(res->hr_remoteaddr, &in) < 0) {
553 		pjdlog_errno(LOG_WARNING, "Unable to create connection to %s",
554 		    res->hr_remoteaddr);
555 	}
556 	/* Try to connect, but accept failure. */
557 	if (proto_connect(in) < 0) {
558 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
559 		    res->hr_remoteaddr);
560 		goto close;
561 	}
562 	/* Error in setting timeout is not critical, but why should it fail? */
563 	if (proto_timeout(in, res->hr_timeout) < 0)
564 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
565 	nvout = nv_alloc();
566 	nv_add_string(nvout, res->hr_name, "resource");
567 	nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
568 	    "token");
569 	nv_add_uint64(nvout, res->hr_resuid, "resuid");
570 	nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
571 	nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
572 	if (nv_error(nvout) != 0) {
573 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
574 		    "Unable to allocate header for connection with %s",
575 		    res->hr_remoteaddr);
576 		nv_free(nvout);
577 		goto close;
578 	}
579 	if (hast_proto_send(res, in, nvout, NULL, 0) < 0) {
580 		pjdlog_errno(LOG_WARNING,
581 		    "Unable to send handshake header to %s",
582 		    res->hr_remoteaddr);
583 		nv_free(nvout);
584 		goto close;
585 	}
586 	nv_free(nvout);
587 	if (hast_proto_recv_hdr(out, &nvin) < 0) {
588 		pjdlog_errno(LOG_WARNING,
589 		    "Unable to receive handshake header from %s",
590 		    res->hr_remoteaddr);
591 		goto close;
592 	}
593 	errmsg = nv_get_string(nvin, "errmsg");
594 	if (errmsg != NULL) {
595 		pjdlog_warning("%s", errmsg);
596 		nv_free(nvin);
597 		goto close;
598 	}
599 	datasize = nv_get_int64(nvin, "datasize");
600 	if (datasize != res->hr_datasize) {
601 		pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
602 		    (intmax_t)res->hr_datasize, (intmax_t)datasize);
603 		nv_free(nvin);
604 		goto close;
605 	}
606 	extentsize = nv_get_int32(nvin, "extentsize");
607 	if (extentsize != res->hr_extentsize) {
608 		pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
609 		    (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
610 		nv_free(nvin);
611 		goto close;
612 	}
613 	res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
614 	res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
615 	res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
616 	map = NULL;
617 	mapsize = nv_get_uint32(nvin, "mapsize");
618 	if (mapsize > 0) {
619 		map = malloc(mapsize);
620 		if (map == NULL) {
621 			pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
622 			    (uintmax_t)mapsize);
623 			nv_free(nvin);
624 			goto close;
625 		}
626 		/*
627 		 * Remote node have some dirty extents on its own, lets
628 		 * download its activemap.
629 		 */
630 		if (hast_proto_recv_data(res, out, nvin, map,
631 		    mapsize) < 0) {
632 			pjdlog_errno(LOG_ERR,
633 			    "Unable to receive remote activemap");
634 			nv_free(nvin);
635 			free(map);
636 			goto close;
637 		}
638 		/*
639 		 * Merge local and remote bitmaps.
640 		 */
641 		activemap_merge(res->hr_amp, map, mapsize);
642 		free(map);
643 		/*
644 		 * Now that we merged bitmaps from both nodes, flush it to the
645 		 * disk before we start to synchronize.
646 		 */
647 		(void)hast_activemap_flush(res);
648 	}
649 	pjdlog_info("Connected to %s.", res->hr_remoteaddr);
650 	if (inp != NULL && outp != NULL) {
651 		*inp = in;
652 		*outp = out;
653 	} else {
654 		res->hr_remotein = in;
655 		res->hr_remoteout = out;
656 	}
657 	event_send(res, EVENT_CONNECT);
658 	return (true);
659 close:
660 	if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
661 		event_send(res, EVENT_SPLITBRAIN);
662 	proto_close(out);
663 	if (in != NULL)
664 		proto_close(in);
665 	return (false);
666 }
667 
668 static void
669 sync_start(void)
670 {
671 
672 	mtx_lock(&sync_lock);
673 	sync_inprogress = true;
674 	mtx_unlock(&sync_lock);
675 	cv_signal(&sync_cond);
676 }
677 
678 static void
679 sync_stop(void)
680 {
681 
682 	mtx_lock(&sync_lock);
683 	if (sync_inprogress)
684 		sync_inprogress = false;
685 	mtx_unlock(&sync_lock);
686 }
687 
688 static void
689 init_ggate(struct hast_resource *res)
690 {
691 	struct g_gate_ctl_create ggiocreate;
692 	struct g_gate_ctl_cancel ggiocancel;
693 
694 	/*
695 	 * We communicate with ggate via /dev/ggctl. Open it.
696 	 */
697 	res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
698 	if (res->hr_ggatefd < 0)
699 		primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
700 	/*
701 	 * Create provider before trying to connect, as connection failure
702 	 * is not critical, but may take some time.
703 	 */
704 	bzero(&ggiocreate, sizeof(ggiocreate));
705 	ggiocreate.gctl_version = G_GATE_VERSION;
706 	ggiocreate.gctl_mediasize = res->hr_datasize;
707 	ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
708 	ggiocreate.gctl_flags = 0;
709 	ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE;
710 	ggiocreate.gctl_timeout = 0;
711 	ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
712 	snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
713 	    res->hr_provname);
714 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
715 		pjdlog_info("Device hast/%s created.", res->hr_provname);
716 		res->hr_ggateunit = ggiocreate.gctl_unit;
717 		return;
718 	}
719 	if (errno != EEXIST) {
720 		primary_exit(EX_OSERR, "Unable to create hast/%s device",
721 		    res->hr_provname);
722 	}
723 	pjdlog_debug(1,
724 	    "Device hast/%s already exists, we will try to take it over.",
725 	    res->hr_provname);
726 	/*
727 	 * If we received EEXIST, we assume that the process who created the
728 	 * provider died and didn't clean up. In that case we will start from
729 	 * where he left of.
730 	 */
731 	bzero(&ggiocancel, sizeof(ggiocancel));
732 	ggiocancel.gctl_version = G_GATE_VERSION;
733 	ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
734 	snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
735 	    res->hr_provname);
736 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
737 		pjdlog_info("Device hast/%s recovered.", res->hr_provname);
738 		res->hr_ggateunit = ggiocancel.gctl_unit;
739 		return;
740 	}
741 	primary_exit(EX_OSERR, "Unable to take over hast/%s device",
742 	    res->hr_provname);
743 }
744 
745 void
746 hastd_primary(struct hast_resource *res)
747 {
748 	pthread_t td;
749 	pid_t pid;
750 	int error;
751 
752 	/*
753 	 * Create communication channel between parent and child.
754 	 */
755 	if (proto_client("socketpair://", &res->hr_ctrl) < 0) {
756 		KEEP_ERRNO((void)pidfile_remove(pfh));
757 		pjdlog_exit(EX_OSERR,
758 		    "Unable to create control sockets between parent and child");
759 	}
760 	/*
761 	 * Create communication channel between child and parent.
762 	 */
763 	if (proto_client("socketpair://", &res->hr_event) < 0) {
764 		KEEP_ERRNO((void)pidfile_remove(pfh));
765 		pjdlog_exit(EX_OSERR,
766 		    "Unable to create event sockets between child and parent");
767 	}
768 
769 	pid = fork();
770 	if (pid < 0) {
771 		KEEP_ERRNO((void)pidfile_remove(pfh));
772 		pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
773 	}
774 
775 	if (pid > 0) {
776 		/* This is parent. */
777 		/* Declare that we are receiver. */
778 		proto_recv(res->hr_event, NULL, 0);
779 		res->hr_workerpid = pid;
780 		return;
781 	}
782 
783 	gres = res;
784 
785 	(void)pidfile_close(pfh);
786 	hook_fini();
787 
788 	setproctitle("%s (primary)", res->hr_name);
789 
790 	/* Declare that we are sender. */
791 	proto_send(res->hr_event, NULL, 0);
792 
793 	init_local(res);
794 	init_ggate(res);
795 	init_environment(res);
796 	/*
797 	 * Create the guard thread first, so we can handle signals from the
798 	 * very begining.
799 	 */
800 	error = pthread_create(&td, NULL, guard_thread, res);
801 	assert(error == 0);
802 	/*
803 	 * Create the control thread before sending any event to the parent,
804 	 * as we can deadlock when parent sends control request to worker,
805 	 * but worker has no control thread started yet, so parent waits.
806 	 * In the meantime worker sends an event to the parent, but parent
807 	 * is unable to handle the event, because it waits for control
808 	 * request response.
809 	 */
810 	error = pthread_create(&td, NULL, ctrl_thread, res);
811 	assert(error == 0);
812 	if (real_remote(res) && init_remote(res, NULL, NULL))
813 		sync_start();
814 	error = pthread_create(&td, NULL, ggate_recv_thread, res);
815 	assert(error == 0);
816 	error = pthread_create(&td, NULL, local_send_thread, res);
817 	assert(error == 0);
818 	error = pthread_create(&td, NULL, remote_send_thread, res);
819 	assert(error == 0);
820 	error = pthread_create(&td, NULL, remote_recv_thread, res);
821 	assert(error == 0);
822 	error = pthread_create(&td, NULL, ggate_send_thread, res);
823 	assert(error == 0);
824 	(void)sync_thread(res);
825 }
826 
827 static void
828 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...)
829 {
830 	char msg[1024];
831 	va_list ap;
832 	int len;
833 
834 	va_start(ap, fmt);
835 	len = vsnprintf(msg, sizeof(msg), fmt, ap);
836 	va_end(ap);
837 	if ((size_t)len < sizeof(msg)) {
838 		switch (ggio->gctl_cmd) {
839 		case BIO_READ:
840 			(void)snprintf(msg + len, sizeof(msg) - len,
841 			    "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
842 			    (uintmax_t)ggio->gctl_length);
843 			break;
844 		case BIO_DELETE:
845 			(void)snprintf(msg + len, sizeof(msg) - len,
846 			    "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
847 			    (uintmax_t)ggio->gctl_length);
848 			break;
849 		case BIO_FLUSH:
850 			(void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
851 			break;
852 		case BIO_WRITE:
853 			(void)snprintf(msg + len, sizeof(msg) - len,
854 			    "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
855 			    (uintmax_t)ggio->gctl_length);
856 			break;
857 		default:
858 			(void)snprintf(msg + len, sizeof(msg) - len,
859 			    "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd);
860 			break;
861 		}
862 	}
863 	pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
864 }
865 
866 static void
867 remote_close(struct hast_resource *res, int ncomp)
868 {
869 
870 	rw_wlock(&hio_remote_lock[ncomp]);
871 	/*
872 	 * A race is possible between dropping rlock and acquiring wlock -
873 	 * another thread can close connection in-between.
874 	 */
875 	if (!ISCONNECTED(res, ncomp)) {
876 		assert(res->hr_remotein == NULL);
877 		assert(res->hr_remoteout == NULL);
878 		rw_unlock(&hio_remote_lock[ncomp]);
879 		return;
880 	}
881 
882 	assert(res->hr_remotein != NULL);
883 	assert(res->hr_remoteout != NULL);
884 
885 	pjdlog_debug(2, "Closing incoming connection to %s.",
886 	    res->hr_remoteaddr);
887 	proto_close(res->hr_remotein);
888 	res->hr_remotein = NULL;
889 	pjdlog_debug(2, "Closing outgoing connection to %s.",
890 	    res->hr_remoteaddr);
891 	proto_close(res->hr_remoteout);
892 	res->hr_remoteout = NULL;
893 
894 	rw_unlock(&hio_remote_lock[ncomp]);
895 
896 	pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
897 
898 	/*
899 	 * Stop synchronization if in-progress.
900 	 */
901 	sync_stop();
902 
903 	event_send(res, EVENT_DISCONNECT);
904 }
905 
906 /*
907  * Thread receives ggate I/O requests from the kernel and passes them to
908  * appropriate threads:
909  * WRITE - always goes to both local_send and remote_send threads
910  * READ (when the block is up-to-date on local component) -
911  *	only local_send thread
912  * READ (when the block isn't up-to-date on local component) -
913  *	only remote_send thread
914  * DELETE - always goes to both local_send and remote_send threads
915  * FLUSH - always goes to both local_send and remote_send threads
916  */
917 static void *
918 ggate_recv_thread(void *arg)
919 {
920 	struct hast_resource *res = arg;
921 	struct g_gate_ctl_io *ggio;
922 	struct hio *hio;
923 	unsigned int ii, ncomp, ncomps;
924 	int error;
925 
926 	ncomps = HAST_NCOMPONENTS;
927 
928 	for (;;) {
929 		pjdlog_debug(2, "ggate_recv: Taking free request.");
930 		QUEUE_TAKE2(hio, free);
931 		pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
932 		ggio = &hio->hio_ggio;
933 		ggio->gctl_unit = res->hr_ggateunit;
934 		ggio->gctl_length = MAXPHYS;
935 		ggio->gctl_error = 0;
936 		pjdlog_debug(2,
937 		    "ggate_recv: (%p) Waiting for request from the kernel.",
938 		    hio);
939 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) {
940 			if (sigexit_received)
941 				pthread_exit(NULL);
942 			primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
943 		}
944 		error = ggio->gctl_error;
945 		switch (error) {
946 		case 0:
947 			break;
948 		case ECANCELED:
949 			/* Exit gracefully. */
950 			if (!sigexit_received) {
951 				pjdlog_debug(2,
952 				    "ggate_recv: (%p) Received cancel from the kernel.",
953 				    hio);
954 				pjdlog_info("Received cancel from the kernel, exiting.");
955 			}
956 			pthread_exit(NULL);
957 		case ENOMEM:
958 			/*
959 			 * Buffer too small? Impossible, we allocate MAXPHYS
960 			 * bytes - request can't be bigger than that.
961 			 */
962 			/* FALLTHROUGH */
963 		case ENXIO:
964 		default:
965 			primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
966 			    strerror(error));
967 		}
968 		for (ii = 0; ii < ncomps; ii++)
969 			hio->hio_errors[ii] = EINVAL;
970 		reqlog(LOG_DEBUG, 2, ggio,
971 		    "ggate_recv: (%p) Request received from the kernel: ",
972 		    hio);
973 		/*
974 		 * Inform all components about new write request.
975 		 * For read request prefer local component unless the given
976 		 * range is out-of-date, then use remote component.
977 		 */
978 		switch (ggio->gctl_cmd) {
979 		case BIO_READ:
980 			pjdlog_debug(2,
981 			    "ggate_recv: (%p) Moving request to the send queue.",
982 			    hio);
983 			refcount_init(&hio->hio_countdown, 1);
984 			mtx_lock(&metadata_lock);
985 			if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
986 			    res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
987 				/*
988 				 * This range is up-to-date on local component,
989 				 * so handle request locally.
990 				 */
991 				 /* Local component is 0 for now. */
992 				ncomp = 0;
993 			} else /* if (res->hr_syncsrc ==
994 			    HAST_SYNCSRC_SECONDARY) */ {
995 				assert(res->hr_syncsrc ==
996 				    HAST_SYNCSRC_SECONDARY);
997 				/*
998 				 * This range is out-of-date on local component,
999 				 * so send request to the remote node.
1000 				 */
1001 				 /* Remote component is 1 for now. */
1002 				ncomp = 1;
1003 			}
1004 			mtx_unlock(&metadata_lock);
1005 			QUEUE_INSERT1(hio, send, ncomp);
1006 			break;
1007 		case BIO_WRITE:
1008 			for (;;) {
1009 				mtx_lock(&range_lock);
1010 				if (rangelock_islocked(range_sync,
1011 				    ggio->gctl_offset, ggio->gctl_length)) {
1012 					pjdlog_debug(2,
1013 					    "regular: Range offset=%jd length=%zu locked.",
1014 					    (intmax_t)ggio->gctl_offset,
1015 					    (size_t)ggio->gctl_length);
1016 					range_regular_wait = true;
1017 					cv_wait(&range_regular_cond, &range_lock);
1018 					range_regular_wait = false;
1019 					mtx_unlock(&range_lock);
1020 					continue;
1021 				}
1022 				if (rangelock_add(range_regular,
1023 				    ggio->gctl_offset, ggio->gctl_length) < 0) {
1024 					mtx_unlock(&range_lock);
1025 					pjdlog_debug(2,
1026 					    "regular: Range offset=%jd length=%zu is already locked, waiting.",
1027 					    (intmax_t)ggio->gctl_offset,
1028 					    (size_t)ggio->gctl_length);
1029 					sleep(1);
1030 					continue;
1031 				}
1032 				mtx_unlock(&range_lock);
1033 				break;
1034 			}
1035 			mtx_lock(&res->hr_amp_lock);
1036 			if (activemap_write_start(res->hr_amp,
1037 			    ggio->gctl_offset, ggio->gctl_length)) {
1038 				(void)hast_activemap_flush(res);
1039 			}
1040 			mtx_unlock(&res->hr_amp_lock);
1041 			/* FALLTHROUGH */
1042 		case BIO_DELETE:
1043 		case BIO_FLUSH:
1044 			pjdlog_debug(2,
1045 			    "ggate_recv: (%p) Moving request to the send queues.",
1046 			    hio);
1047 			refcount_init(&hio->hio_countdown, ncomps);
1048 			for (ii = 0; ii < ncomps; ii++)
1049 				QUEUE_INSERT1(hio, send, ii);
1050 			break;
1051 		}
1052 	}
1053 	/* NOTREACHED */
1054 	return (NULL);
1055 }
1056 
1057 /*
1058  * Thread reads from or writes to local component.
1059  * If local read fails, it redirects it to remote_send thread.
1060  */
1061 static void *
1062 local_send_thread(void *arg)
1063 {
1064 	struct hast_resource *res = arg;
1065 	struct g_gate_ctl_io *ggio;
1066 	struct hio *hio;
1067 	unsigned int ncomp, rncomp;
1068 	ssize_t ret;
1069 
1070 	/* Local component is 0 for now. */
1071 	ncomp = 0;
1072 	/* Remote component is 1 for now. */
1073 	rncomp = 1;
1074 
1075 	for (;;) {
1076 		pjdlog_debug(2, "local_send: Taking request.");
1077 		QUEUE_TAKE1(hio, send, ncomp);
1078 		pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1079 		ggio = &hio->hio_ggio;
1080 		switch (ggio->gctl_cmd) {
1081 		case BIO_READ:
1082 			ret = pread(res->hr_localfd, ggio->gctl_data,
1083 			    ggio->gctl_length,
1084 			    ggio->gctl_offset + res->hr_localoff);
1085 			if (ret == ggio->gctl_length)
1086 				hio->hio_errors[ncomp] = 0;
1087 			else {
1088 				/*
1089 				 * If READ failed, try to read from remote node.
1090 				 */
1091 				QUEUE_INSERT1(hio, send, rncomp);
1092 				continue;
1093 			}
1094 			break;
1095 		case BIO_WRITE:
1096 			ret = pwrite(res->hr_localfd, ggio->gctl_data,
1097 			    ggio->gctl_length,
1098 			    ggio->gctl_offset + res->hr_localoff);
1099 			if (ret < 0)
1100 				hio->hio_errors[ncomp] = errno;
1101 			else if (ret != ggio->gctl_length)
1102 				hio->hio_errors[ncomp] = EIO;
1103 			else
1104 				hio->hio_errors[ncomp] = 0;
1105 			break;
1106 		case BIO_DELETE:
1107 			ret = g_delete(res->hr_localfd,
1108 			    ggio->gctl_offset + res->hr_localoff,
1109 			    ggio->gctl_length);
1110 			if (ret < 0)
1111 				hio->hio_errors[ncomp] = errno;
1112 			else
1113 				hio->hio_errors[ncomp] = 0;
1114 			break;
1115 		case BIO_FLUSH:
1116 			ret = g_flush(res->hr_localfd);
1117 			if (ret < 0)
1118 				hio->hio_errors[ncomp] = errno;
1119 			else
1120 				hio->hio_errors[ncomp] = 0;
1121 			break;
1122 		}
1123 		if (refcount_release(&hio->hio_countdown)) {
1124 			if (ISSYNCREQ(hio)) {
1125 				mtx_lock(&sync_lock);
1126 				SYNCREQDONE(hio);
1127 				mtx_unlock(&sync_lock);
1128 				cv_signal(&sync_cond);
1129 			} else {
1130 				pjdlog_debug(2,
1131 				    "local_send: (%p) Moving request to the done queue.",
1132 				    hio);
1133 				QUEUE_INSERT2(hio, done);
1134 			}
1135 		}
1136 	}
1137 	/* NOTREACHED */
1138 	return (NULL);
1139 }
1140 
1141 /*
1142  * Thread sends request to secondary node.
1143  */
1144 static void *
1145 remote_send_thread(void *arg)
1146 {
1147 	struct hast_resource *res = arg;
1148 	struct g_gate_ctl_io *ggio;
1149 	struct hio *hio;
1150 	struct nv *nv;
1151 	unsigned int ncomp;
1152 	bool wakeup;
1153 	uint64_t offset, length;
1154 	uint8_t cmd;
1155 	void *data;
1156 
1157 	/* Remote component is 1 for now. */
1158 	ncomp = 1;
1159 
1160 	for (;;) {
1161 		pjdlog_debug(2, "remote_send: Taking request.");
1162 		QUEUE_TAKE1(hio, send, ncomp);
1163 		pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1164 		ggio = &hio->hio_ggio;
1165 		switch (ggio->gctl_cmd) {
1166 		case BIO_READ:
1167 			cmd = HIO_READ;
1168 			data = NULL;
1169 			offset = ggio->gctl_offset;
1170 			length = ggio->gctl_length;
1171 			break;
1172 		case BIO_WRITE:
1173 			cmd = HIO_WRITE;
1174 			data = ggio->gctl_data;
1175 			offset = ggio->gctl_offset;
1176 			length = ggio->gctl_length;
1177 			break;
1178 		case BIO_DELETE:
1179 			cmd = HIO_DELETE;
1180 			data = NULL;
1181 			offset = ggio->gctl_offset;
1182 			length = ggio->gctl_length;
1183 			break;
1184 		case BIO_FLUSH:
1185 			cmd = HIO_FLUSH;
1186 			data = NULL;
1187 			offset = 0;
1188 			length = 0;
1189 			break;
1190 		default:
1191 			assert(!"invalid condition");
1192 			abort();
1193 		}
1194 		nv = nv_alloc();
1195 		nv_add_uint8(nv, cmd, "cmd");
1196 		nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1197 		nv_add_uint64(nv, offset, "offset");
1198 		nv_add_uint64(nv, length, "length");
1199 		if (nv_error(nv) != 0) {
1200 			hio->hio_errors[ncomp] = nv_error(nv);
1201 			pjdlog_debug(2,
1202 			    "remote_send: (%p) Unable to prepare header to send.",
1203 			    hio);
1204 			reqlog(LOG_ERR, 0, ggio,
1205 			    "Unable to prepare header to send (%s): ",
1206 			    strerror(nv_error(nv)));
1207 			/* Move failed request immediately to the done queue. */
1208 			goto done_queue;
1209 		}
1210 		pjdlog_debug(2,
1211 		    "remote_send: (%p) Moving request to the recv queue.",
1212 		    hio);
1213 		/*
1214 		 * Protect connection from disappearing.
1215 		 */
1216 		rw_rlock(&hio_remote_lock[ncomp]);
1217 		if (!ISCONNECTED(res, ncomp)) {
1218 			rw_unlock(&hio_remote_lock[ncomp]);
1219 			hio->hio_errors[ncomp] = ENOTCONN;
1220 			goto done_queue;
1221 		}
1222 		/*
1223 		 * Move the request to recv queue before sending it, because
1224 		 * in different order we can get reply before we move request
1225 		 * to recv queue.
1226 		 */
1227 		mtx_lock(&hio_recv_list_lock[ncomp]);
1228 		wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1229 		TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1230 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1231 		if (hast_proto_send(res, res->hr_remoteout, nv, data,
1232 		    data != NULL ? length : 0) < 0) {
1233 			hio->hio_errors[ncomp] = errno;
1234 			rw_unlock(&hio_remote_lock[ncomp]);
1235 			pjdlog_debug(2,
1236 			    "remote_send: (%p) Unable to send request.", hio);
1237 			reqlog(LOG_ERR, 0, ggio,
1238 			    "Unable to send request (%s): ",
1239 			    strerror(hio->hio_errors[ncomp]));
1240 			remote_close(res, ncomp);
1241 			/*
1242 			 * Take request back from the receive queue and move
1243 			 * it immediately to the done queue.
1244 			 */
1245 			mtx_lock(&hio_recv_list_lock[ncomp]);
1246 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1247 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1248 			goto done_queue;
1249 		}
1250 		rw_unlock(&hio_remote_lock[ncomp]);
1251 		nv_free(nv);
1252 		if (wakeup)
1253 			cv_signal(&hio_recv_list_cond[ncomp]);
1254 		continue;
1255 done_queue:
1256 		nv_free(nv);
1257 		if (ISSYNCREQ(hio)) {
1258 			if (!refcount_release(&hio->hio_countdown))
1259 				continue;
1260 			mtx_lock(&sync_lock);
1261 			SYNCREQDONE(hio);
1262 			mtx_unlock(&sync_lock);
1263 			cv_signal(&sync_cond);
1264 			continue;
1265 		}
1266 		if (ggio->gctl_cmd == BIO_WRITE) {
1267 			mtx_lock(&res->hr_amp_lock);
1268 			if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1269 			    ggio->gctl_length)) {
1270 				(void)hast_activemap_flush(res);
1271 			}
1272 			mtx_unlock(&res->hr_amp_lock);
1273 		}
1274 		if (!refcount_release(&hio->hio_countdown))
1275 			continue;
1276 		pjdlog_debug(2,
1277 		    "remote_send: (%p) Moving request to the done queue.",
1278 		    hio);
1279 		QUEUE_INSERT2(hio, done);
1280 	}
1281 	/* NOTREACHED */
1282 	return (NULL);
1283 }
1284 
1285 /*
1286  * Thread receives answer from secondary node and passes it to ggate_send
1287  * thread.
1288  */
1289 static void *
1290 remote_recv_thread(void *arg)
1291 {
1292 	struct hast_resource *res = arg;
1293 	struct g_gate_ctl_io *ggio;
1294 	struct hio *hio;
1295 	struct nv *nv;
1296 	unsigned int ncomp;
1297 	uint64_t seq;
1298 	int error;
1299 
1300 	/* Remote component is 1 for now. */
1301 	ncomp = 1;
1302 
1303 	for (;;) {
1304 		/* Wait until there is anything to receive. */
1305 		mtx_lock(&hio_recv_list_lock[ncomp]);
1306 		while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1307 			pjdlog_debug(2, "remote_recv: No requests, waiting.");
1308 			cv_wait(&hio_recv_list_cond[ncomp],
1309 			    &hio_recv_list_lock[ncomp]);
1310 		}
1311 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1312 		rw_rlock(&hio_remote_lock[ncomp]);
1313 		if (!ISCONNECTED(res, ncomp)) {
1314 			rw_unlock(&hio_remote_lock[ncomp]);
1315 			/*
1316 			 * Connection is dead, so move all pending requests to
1317 			 * the done queue (one-by-one).
1318 			 */
1319 			mtx_lock(&hio_recv_list_lock[ncomp]);
1320 			hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1321 			assert(hio != NULL);
1322 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1323 			    hio_next[ncomp]);
1324 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1325 			goto done_queue;
1326 		}
1327 		if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) {
1328 			pjdlog_errno(LOG_ERR,
1329 			    "Unable to receive reply header");
1330 			rw_unlock(&hio_remote_lock[ncomp]);
1331 			remote_close(res, ncomp);
1332 			continue;
1333 		}
1334 		rw_unlock(&hio_remote_lock[ncomp]);
1335 		seq = nv_get_uint64(nv, "seq");
1336 		if (seq == 0) {
1337 			pjdlog_error("Header contains no 'seq' field.");
1338 			nv_free(nv);
1339 			continue;
1340 		}
1341 		mtx_lock(&hio_recv_list_lock[ncomp]);
1342 		TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1343 			if (hio->hio_ggio.gctl_seq == seq) {
1344 				TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1345 				    hio_next[ncomp]);
1346 				break;
1347 			}
1348 		}
1349 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1350 		if (hio == NULL) {
1351 			pjdlog_error("Found no request matching received 'seq' field (%ju).",
1352 			    (uintmax_t)seq);
1353 			nv_free(nv);
1354 			continue;
1355 		}
1356 		error = nv_get_int16(nv, "error");
1357 		if (error != 0) {
1358 			/* Request failed on remote side. */
1359 			hio->hio_errors[ncomp] = 0;
1360 			nv_free(nv);
1361 			goto done_queue;
1362 		}
1363 		ggio = &hio->hio_ggio;
1364 		switch (ggio->gctl_cmd) {
1365 		case BIO_READ:
1366 			rw_rlock(&hio_remote_lock[ncomp]);
1367 			if (!ISCONNECTED(res, ncomp)) {
1368 				rw_unlock(&hio_remote_lock[ncomp]);
1369 				nv_free(nv);
1370 				goto done_queue;
1371 			}
1372 			if (hast_proto_recv_data(res, res->hr_remotein, nv,
1373 			    ggio->gctl_data, ggio->gctl_length) < 0) {
1374 				hio->hio_errors[ncomp] = errno;
1375 				pjdlog_errno(LOG_ERR,
1376 				    "Unable to receive reply data");
1377 				rw_unlock(&hio_remote_lock[ncomp]);
1378 				nv_free(nv);
1379 				remote_close(res, ncomp);
1380 				goto done_queue;
1381 			}
1382 			rw_unlock(&hio_remote_lock[ncomp]);
1383 			break;
1384 		case BIO_WRITE:
1385 		case BIO_DELETE:
1386 		case BIO_FLUSH:
1387 			break;
1388 		default:
1389 			assert(!"invalid condition");
1390 			abort();
1391 		}
1392 		hio->hio_errors[ncomp] = 0;
1393 		nv_free(nv);
1394 done_queue:
1395 		if (refcount_release(&hio->hio_countdown)) {
1396 			if (ISSYNCREQ(hio)) {
1397 				mtx_lock(&sync_lock);
1398 				SYNCREQDONE(hio);
1399 				mtx_unlock(&sync_lock);
1400 				cv_signal(&sync_cond);
1401 			} else {
1402 				pjdlog_debug(2,
1403 				    "remote_recv: (%p) Moving request to the done queue.",
1404 				    hio);
1405 				QUEUE_INSERT2(hio, done);
1406 			}
1407 		}
1408 	}
1409 	/* NOTREACHED */
1410 	return (NULL);
1411 }
1412 
1413 /*
1414  * Thread sends answer to the kernel.
1415  */
1416 static void *
1417 ggate_send_thread(void *arg)
1418 {
1419 	struct hast_resource *res = arg;
1420 	struct g_gate_ctl_io *ggio;
1421 	struct hio *hio;
1422 	unsigned int ii, ncomp, ncomps;
1423 
1424 	ncomps = HAST_NCOMPONENTS;
1425 
1426 	for (;;) {
1427 		pjdlog_debug(2, "ggate_send: Taking request.");
1428 		QUEUE_TAKE2(hio, done);
1429 		pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1430 		ggio = &hio->hio_ggio;
1431 		for (ii = 0; ii < ncomps; ii++) {
1432 			if (hio->hio_errors[ii] == 0) {
1433 				/*
1434 				 * One successful request is enough to declare
1435 				 * success.
1436 				 */
1437 				ggio->gctl_error = 0;
1438 				break;
1439 			}
1440 		}
1441 		if (ii == ncomps) {
1442 			/*
1443 			 * None of the requests were successful.
1444 			 * Use first error.
1445 			 */
1446 			ggio->gctl_error = hio->hio_errors[0];
1447 		}
1448 		if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1449 			mtx_lock(&res->hr_amp_lock);
1450 			activemap_write_complete(res->hr_amp,
1451 			    ggio->gctl_offset, ggio->gctl_length);
1452 			mtx_unlock(&res->hr_amp_lock);
1453 		}
1454 		if (ggio->gctl_cmd == BIO_WRITE) {
1455 			/*
1456 			 * Unlock range we locked.
1457 			 */
1458 			mtx_lock(&range_lock);
1459 			rangelock_del(range_regular, ggio->gctl_offset,
1460 			    ggio->gctl_length);
1461 			if (range_sync_wait)
1462 				cv_signal(&range_sync_cond);
1463 			mtx_unlock(&range_lock);
1464 			/*
1465 			 * Bump local count if this is first write after
1466 			 * connection failure with remote node.
1467 			 */
1468 			ncomp = 1;
1469 			rw_rlock(&hio_remote_lock[ncomp]);
1470 			if (!ISCONNECTED(res, ncomp)) {
1471 				mtx_lock(&metadata_lock);
1472 				if (res->hr_primary_localcnt ==
1473 				    res->hr_secondary_remotecnt) {
1474 					res->hr_primary_localcnt++;
1475 					pjdlog_debug(1,
1476 					    "Increasing localcnt to %ju.",
1477 					    (uintmax_t)res->hr_primary_localcnt);
1478 					(void)metadata_write(res);
1479 				}
1480 				mtx_unlock(&metadata_lock);
1481 			}
1482 			rw_unlock(&hio_remote_lock[ncomp]);
1483 		}
1484 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0)
1485 			primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1486 		pjdlog_debug(2,
1487 		    "ggate_send: (%p) Moving request to the free queue.", hio);
1488 		QUEUE_INSERT2(hio, free);
1489 	}
1490 	/* NOTREACHED */
1491 	return (NULL);
1492 }
1493 
1494 /*
1495  * Thread synchronize local and remote components.
1496  */
1497 static void *
1498 sync_thread(void *arg __unused)
1499 {
1500 	struct hast_resource *res = arg;
1501 	struct hio *hio;
1502 	struct g_gate_ctl_io *ggio;
1503 	unsigned int ii, ncomp, ncomps;
1504 	off_t offset, length, synced;
1505 	bool dorewind;
1506 	int syncext;
1507 
1508 	ncomps = HAST_NCOMPONENTS;
1509 	dorewind = true;
1510 	synced = 0;
1511 	offset = -1;
1512 
1513 	for (;;) {
1514 		mtx_lock(&sync_lock);
1515 		if (offset >= 0 && !sync_inprogress) {
1516 			pjdlog_info("Synchronization interrupted. "
1517 			    "%jd bytes synchronized so far.",
1518 			    (intmax_t)synced);
1519 			event_send(res, EVENT_SYNCINTR);
1520 		}
1521 		while (!sync_inprogress) {
1522 			dorewind = true;
1523 			synced = 0;
1524 			cv_wait(&sync_cond, &sync_lock);
1525 		}
1526 		mtx_unlock(&sync_lock);
1527 		/*
1528 		 * Obtain offset at which we should synchronize.
1529 		 * Rewind synchronization if needed.
1530 		 */
1531 		mtx_lock(&res->hr_amp_lock);
1532 		if (dorewind)
1533 			activemap_sync_rewind(res->hr_amp);
1534 		offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1535 		if (syncext != -1) {
1536 			/*
1537 			 * We synchronized entire syncext extent, we can mark
1538 			 * it as clean now.
1539 			 */
1540 			if (activemap_extent_complete(res->hr_amp, syncext))
1541 				(void)hast_activemap_flush(res);
1542 		}
1543 		mtx_unlock(&res->hr_amp_lock);
1544 		if (dorewind) {
1545 			dorewind = false;
1546 			if (offset < 0)
1547 				pjdlog_info("Nodes are in sync.");
1548 			else {
1549 				pjdlog_info("Synchronization started. %ju bytes to go.",
1550 				    (uintmax_t)(res->hr_extentsize *
1551 				    activemap_ndirty(res->hr_amp)));
1552 				event_send(res, EVENT_SYNCSTART);
1553 			}
1554 		}
1555 		if (offset < 0) {
1556 			sync_stop();
1557 			pjdlog_debug(1, "Nothing to synchronize.");
1558 			/*
1559 			 * Synchronization complete, make both localcnt and
1560 			 * remotecnt equal.
1561 			 */
1562 			ncomp = 1;
1563 			rw_rlock(&hio_remote_lock[ncomp]);
1564 			if (ISCONNECTED(res, ncomp)) {
1565 				if (synced > 0) {
1566 					pjdlog_info("Synchronization complete. "
1567 					    "%jd bytes synchronized.",
1568 					    (intmax_t)synced);
1569 					event_send(res, EVENT_SYNCDONE);
1570 				}
1571 				mtx_lock(&metadata_lock);
1572 				res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
1573 				res->hr_primary_localcnt =
1574 				    res->hr_secondary_localcnt;
1575 				res->hr_primary_remotecnt =
1576 				    res->hr_secondary_remotecnt;
1577 				pjdlog_debug(1,
1578 				    "Setting localcnt to %ju and remotecnt to %ju.",
1579 				    (uintmax_t)res->hr_primary_localcnt,
1580 				    (uintmax_t)res->hr_secondary_localcnt);
1581 				(void)metadata_write(res);
1582 				mtx_unlock(&metadata_lock);
1583 			}
1584 			rw_unlock(&hio_remote_lock[ncomp]);
1585 			continue;
1586 		}
1587 		pjdlog_debug(2, "sync: Taking free request.");
1588 		QUEUE_TAKE2(hio, free);
1589 		pjdlog_debug(2, "sync: (%p) Got free request.", hio);
1590 		/*
1591 		 * Lock the range we are going to synchronize. We don't want
1592 		 * race where someone writes between our read and write.
1593 		 */
1594 		for (;;) {
1595 			mtx_lock(&range_lock);
1596 			if (rangelock_islocked(range_regular, offset, length)) {
1597 				pjdlog_debug(2,
1598 				    "sync: Range offset=%jd length=%jd locked.",
1599 				    (intmax_t)offset, (intmax_t)length);
1600 				range_sync_wait = true;
1601 				cv_wait(&range_sync_cond, &range_lock);
1602 				range_sync_wait = false;
1603 				mtx_unlock(&range_lock);
1604 				continue;
1605 			}
1606 			if (rangelock_add(range_sync, offset, length) < 0) {
1607 				mtx_unlock(&range_lock);
1608 				pjdlog_debug(2,
1609 				    "sync: Range offset=%jd length=%jd is already locked, waiting.",
1610 				    (intmax_t)offset, (intmax_t)length);
1611 				sleep(1);
1612 				continue;
1613 			}
1614 			mtx_unlock(&range_lock);
1615 			break;
1616 		}
1617 		/*
1618 		 * First read the data from synchronization source.
1619 		 */
1620 		SYNCREQ(hio);
1621 		ggio = &hio->hio_ggio;
1622 		ggio->gctl_cmd = BIO_READ;
1623 		ggio->gctl_offset = offset;
1624 		ggio->gctl_length = length;
1625 		ggio->gctl_error = 0;
1626 		for (ii = 0; ii < ncomps; ii++)
1627 			hio->hio_errors[ii] = EINVAL;
1628 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1629 		    hio);
1630 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1631 		    hio);
1632 		mtx_lock(&metadata_lock);
1633 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1634 			/*
1635 			 * This range is up-to-date on local component,
1636 			 * so handle request locally.
1637 			 */
1638 			 /* Local component is 0 for now. */
1639 			ncomp = 0;
1640 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1641 			assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1642 			/*
1643 			 * This range is out-of-date on local component,
1644 			 * so send request to the remote node.
1645 			 */
1646 			 /* Remote component is 1 for now. */
1647 			ncomp = 1;
1648 		}
1649 		mtx_unlock(&metadata_lock);
1650 		refcount_init(&hio->hio_countdown, 1);
1651 		QUEUE_INSERT1(hio, send, ncomp);
1652 
1653 		/*
1654 		 * Let's wait for READ to finish.
1655 		 */
1656 		mtx_lock(&sync_lock);
1657 		while (!ISSYNCREQDONE(hio))
1658 			cv_wait(&sync_cond, &sync_lock);
1659 		mtx_unlock(&sync_lock);
1660 
1661 		if (hio->hio_errors[ncomp] != 0) {
1662 			pjdlog_error("Unable to read synchronization data: %s.",
1663 			    strerror(hio->hio_errors[ncomp]));
1664 			goto free_queue;
1665 		}
1666 
1667 		/*
1668 		 * We read the data from synchronization source, now write it
1669 		 * to synchronization target.
1670 		 */
1671 		SYNCREQ(hio);
1672 		ggio->gctl_cmd = BIO_WRITE;
1673 		for (ii = 0; ii < ncomps; ii++)
1674 			hio->hio_errors[ii] = EINVAL;
1675 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1676 		    hio);
1677 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1678 		    hio);
1679 		mtx_lock(&metadata_lock);
1680 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1681 			/*
1682 			 * This range is up-to-date on local component,
1683 			 * so we update remote component.
1684 			 */
1685 			 /* Remote component is 1 for now. */
1686 			ncomp = 1;
1687 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1688 			assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1689 			/*
1690 			 * This range is out-of-date on local component,
1691 			 * so we update it.
1692 			 */
1693 			 /* Local component is 0 for now. */
1694 			ncomp = 0;
1695 		}
1696 		mtx_unlock(&metadata_lock);
1697 
1698 		pjdlog_debug(2, "sync: (%p) Moving request to the send queues.",
1699 		    hio);
1700 		refcount_init(&hio->hio_countdown, 1);
1701 		QUEUE_INSERT1(hio, send, ncomp);
1702 
1703 		/*
1704 		 * Let's wait for WRITE to finish.
1705 		 */
1706 		mtx_lock(&sync_lock);
1707 		while (!ISSYNCREQDONE(hio))
1708 			cv_wait(&sync_cond, &sync_lock);
1709 		mtx_unlock(&sync_lock);
1710 
1711 		if (hio->hio_errors[ncomp] != 0) {
1712 			pjdlog_error("Unable to write synchronization data: %s.",
1713 			    strerror(hio->hio_errors[ncomp]));
1714 			goto free_queue;
1715 		}
1716 
1717 		synced += length;
1718 free_queue:
1719 		mtx_lock(&range_lock);
1720 		rangelock_del(range_sync, offset, length);
1721 		if (range_regular_wait)
1722 			cv_signal(&range_regular_cond);
1723 		mtx_unlock(&range_lock);
1724 		pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
1725 		    hio);
1726 		QUEUE_INSERT2(hio, free);
1727 	}
1728 	/* NOTREACHED */
1729 	return (NULL);
1730 }
1731 
1732 static void
1733 config_reload(void)
1734 {
1735 	struct hastd_config *newcfg;
1736 	struct hast_resource *res;
1737 	unsigned int ii, ncomps;
1738 	int modified;
1739 
1740 	pjdlog_info("Reloading configuration...");
1741 
1742 	ncomps = HAST_NCOMPONENTS;
1743 
1744 	newcfg = yy_config_parse(cfgpath, false);
1745 	if (newcfg == NULL)
1746 		goto failed;
1747 
1748 	TAILQ_FOREACH(res, &newcfg->hc_resources, hr_next) {
1749 		if (strcmp(res->hr_name, gres->hr_name) == 0)
1750 			break;
1751 	}
1752 	/*
1753 	 * If resource was removed from the configuration file, resource
1754 	 * name, provider name or path to local component was modified we
1755 	 * shouldn't be here. This means that someone modified configuration
1756 	 * file and send SIGHUP to us instead of main hastd process.
1757 	 * Log advice and ignore the signal.
1758 	 */
1759 	if (res == NULL || strcmp(gres->hr_name, res->hr_name) != 0 ||
1760 	    strcmp(gres->hr_provname, res->hr_provname) != 0 ||
1761 	    strcmp(gres->hr_localpath, res->hr_localpath) != 0) {
1762 		pjdlog_warning("To reload configuration send SIGHUP to the main hastd process (pid %u).",
1763 		    (unsigned int)getppid());
1764 		goto failed;
1765 	}
1766 
1767 #define MODIFIED_REMOTEADDR	0x1
1768 #define MODIFIED_REPLICATION	0x2
1769 #define MODIFIED_TIMEOUT	0x4
1770 #define MODIFIED_EXEC		0x8
1771 	modified = 0;
1772 	if (strcmp(gres->hr_remoteaddr, res->hr_remoteaddr) != 0) {
1773 		/*
1774 		 * Don't copy res->hr_remoteaddr to gres just yet.
1775 		 * We want remote_close() to log disconnect from the old
1776 		 * addresses, not from the new ones.
1777 		 */
1778 		modified |= MODIFIED_REMOTEADDR;
1779 	}
1780 	if (gres->hr_replication != res->hr_replication) {
1781 		gres->hr_replication = res->hr_replication;
1782 		modified |= MODIFIED_REPLICATION;
1783 	}
1784 	if (gres->hr_timeout != res->hr_timeout) {
1785 		gres->hr_timeout = res->hr_timeout;
1786 		modified |= MODIFIED_TIMEOUT;
1787 	}
1788 	if (strcmp(gres->hr_exec, res->hr_exec) != 0) {
1789 		strlcpy(gres->hr_exec, res->hr_exec, sizeof(gres->hr_exec));
1790 		modified |= MODIFIED_EXEC;
1791 	}
1792 	/*
1793 	 * If only timeout was modified we only need to change it without
1794 	 * reconnecting.
1795 	 */
1796 	if (modified == MODIFIED_TIMEOUT) {
1797 		for (ii = 0; ii < ncomps; ii++) {
1798 			if (!ISREMOTE(ii))
1799 				continue;
1800 			rw_rlock(&hio_remote_lock[ii]);
1801 			if (!ISCONNECTED(gres, ii)) {
1802 				rw_unlock(&hio_remote_lock[ii]);
1803 				continue;
1804 			}
1805 			rw_unlock(&hio_remote_lock[ii]);
1806 			if (proto_timeout(gres->hr_remotein,
1807 			    gres->hr_timeout) < 0) {
1808 				pjdlog_errno(LOG_WARNING,
1809 				    "Unable to set connection timeout");
1810 			}
1811 			if (proto_timeout(gres->hr_remoteout,
1812 			    gres->hr_timeout) < 0) {
1813 				pjdlog_errno(LOG_WARNING,
1814 				    "Unable to set connection timeout");
1815 			}
1816 		}
1817 	} else if ((modified &
1818 	    (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) {
1819 		for (ii = 0; ii < ncomps; ii++) {
1820 			if (!ISREMOTE(ii))
1821 				continue;
1822 			remote_close(gres, ii);
1823 		}
1824 		if (modified & MODIFIED_REMOTEADDR) {
1825 			strlcpy(gres->hr_remoteaddr, res->hr_remoteaddr,
1826 			    sizeof(gres->hr_remoteaddr));
1827 		}
1828 	}
1829 #undef	MODIFIED_REMOTEADDR
1830 #undef	MODIFIED_REPLICATION
1831 #undef	MODIFIED_TIMEOUT
1832 #undef	MODIFIED_EXEC
1833 
1834 	pjdlog_info("Configuration reloaded successfully.");
1835 	return;
1836 failed:
1837 	if (newcfg != NULL) {
1838 		if (newcfg->hc_controlconn != NULL)
1839 			proto_close(newcfg->hc_controlconn);
1840 		if (newcfg->hc_listenconn != NULL)
1841 			proto_close(newcfg->hc_listenconn);
1842 		yy_config_free(newcfg);
1843 	}
1844 	pjdlog_warning("Configuration not reloaded.");
1845 }
1846 
1847 static void
1848 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1849 {
1850 	struct nv *nv;
1851 
1852 	nv = nv_alloc();
1853 	nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1854 	if (nv_error(nv) != 0) {
1855 		nv_free(nv);
1856 		pjdlog_debug(1,
1857 		    "keepalive_send: Unable to prepare header to send.");
1858 		return;
1859 	}
1860 	if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) {
1861 		pjdlog_common(LOG_DEBUG, 1, errno,
1862 		    "keepalive_send: Unable to send request");
1863 		nv_free(nv);
1864 		rw_unlock(&hio_remote_lock[ncomp]);
1865 		remote_close(res, ncomp);
1866 		rw_rlock(&hio_remote_lock[ncomp]);
1867 		return;
1868 	}
1869 	nv_free(nv);
1870 	pjdlog_debug(2, "keepalive_send: Request sent.");
1871 }
1872 
1873 static void
1874 guard_one(struct hast_resource *res, unsigned int ncomp)
1875 {
1876 	struct proto_conn *in, *out;
1877 
1878 	if (!ISREMOTE(ncomp))
1879 		return;
1880 
1881 	rw_rlock(&hio_remote_lock[ncomp]);
1882 
1883 	if (!real_remote(res)) {
1884 		rw_unlock(&hio_remote_lock[ncomp]);
1885 		return;
1886 	}
1887 
1888 	if (ISCONNECTED(res, ncomp)) {
1889 		assert(res->hr_remotein != NULL);
1890 		assert(res->hr_remoteout != NULL);
1891 		keepalive_send(res, ncomp);
1892 	}
1893 
1894 	if (ISCONNECTED(res, ncomp)) {
1895 		assert(res->hr_remotein != NULL);
1896 		assert(res->hr_remoteout != NULL);
1897 		rw_unlock(&hio_remote_lock[ncomp]);
1898 		pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
1899 		    res->hr_remoteaddr);
1900 		return;
1901 	}
1902 
1903 	assert(res->hr_remotein == NULL);
1904 	assert(res->hr_remoteout == NULL);
1905 	/*
1906 	 * Upgrade the lock. It doesn't have to be atomic as no other thread
1907 	 * can change connection status from disconnected to connected.
1908 	 */
1909 	rw_unlock(&hio_remote_lock[ncomp]);
1910 	pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
1911 	    res->hr_remoteaddr);
1912 	in = out = NULL;
1913 	if (init_remote(res, &in, &out)) {
1914 		rw_wlock(&hio_remote_lock[ncomp]);
1915 		assert(res->hr_remotein == NULL);
1916 		assert(res->hr_remoteout == NULL);
1917 		assert(in != NULL && out != NULL);
1918 		res->hr_remotein = in;
1919 		res->hr_remoteout = out;
1920 		rw_unlock(&hio_remote_lock[ncomp]);
1921 		pjdlog_info("Successfully reconnected to %s.",
1922 		    res->hr_remoteaddr);
1923 		sync_start();
1924 	} else {
1925 		/* Both connections should be NULL. */
1926 		assert(res->hr_remotein == NULL);
1927 		assert(res->hr_remoteout == NULL);
1928 		assert(in == NULL && out == NULL);
1929 		pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
1930 		    res->hr_remoteaddr);
1931 	}
1932 }
1933 
1934 /*
1935  * Thread guards remote connections and reconnects when needed, handles
1936  * signals, etc.
1937  */
1938 static void *
1939 guard_thread(void *arg)
1940 {
1941 	struct hast_resource *res = arg;
1942 	unsigned int ii, ncomps;
1943 	struct timespec timeout;
1944 	time_t lastcheck, now;
1945 	sigset_t mask;
1946 	int signo;
1947 
1948 	ncomps = HAST_NCOMPONENTS;
1949 	lastcheck = time(NULL);
1950 
1951 	PJDLOG_VERIFY(sigemptyset(&mask) == 0);
1952 	PJDLOG_VERIFY(sigaddset(&mask, SIGHUP) == 0);
1953 	PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
1954 	PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
1955 
1956 	timeout.tv_nsec = 0;
1957 	signo = -1;
1958 
1959 	for (;;) {
1960 		switch (signo) {
1961 		case SIGHUP:
1962 			config_reload();
1963 			break;
1964 		case SIGINT:
1965 		case SIGTERM:
1966 			sigexit_received = true;
1967 			primary_exitx(EX_OK,
1968 			    "Termination signal received, exiting.");
1969 			break;
1970 		default:
1971 			break;
1972 		}
1973 
1974 		pjdlog_debug(2, "remote_guard: Checking connections.");
1975 		now = time(NULL);
1976 		if (lastcheck + RETRY_SLEEP <= now) {
1977 			for (ii = 0; ii < ncomps; ii++)
1978 				guard_one(res, ii);
1979 			lastcheck = now;
1980 		}
1981 		timeout.tv_sec = RETRY_SLEEP;
1982 		signo = sigtimedwait(&mask, NULL, &timeout);
1983 	}
1984 	/* NOTREACHED */
1985 	return (NULL);
1986 }
1987