xref: /freebsd/sbin/hastd/primary.c (revision 7778ab7e0cc22f0824eb1d1047a7ef8b4785267a)
1 /*-
2  * Copyright (c) 2009 The FreeBSD Foundation
3  * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net>
4  * All rights reserved.
5  *
6  * This software was developed by Pawel Jakub Dawidek under sponsorship from
7  * the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/types.h>
35 #include <sys/time.h>
36 #include <sys/bio.h>
37 #include <sys/disk.h>
38 #include <sys/refcount.h>
39 #include <sys/stat.h>
40 
41 #include <geom/gate/g_gate.h>
42 
43 #include <err.h>
44 #include <errno.h>
45 #include <fcntl.h>
46 #include <libgeom.h>
47 #include <pthread.h>
48 #include <signal.h>
49 #include <stdint.h>
50 #include <stdio.h>
51 #include <string.h>
52 #include <sysexits.h>
53 #include <unistd.h>
54 
55 #include <activemap.h>
56 #include <nv.h>
57 #include <rangelock.h>
58 
59 #include "control.h"
60 #include "event.h"
61 #include "hast.h"
62 #include "hast_proto.h"
63 #include "hastd.h"
64 #include "hooks.h"
65 #include "metadata.h"
66 #include "proto.h"
67 #include "pjdlog.h"
68 #include "subr.h"
69 #include "synch.h"
70 
71 /* The is only one remote component for now. */
72 #define	ISREMOTE(no)	((no) == 1)
73 
74 struct hio {
75 	/*
76 	 * Number of components we are still waiting for.
77 	 * When this field goes to 0, we can send the request back to the
78 	 * kernel. Each component has to decrease this counter by one
79 	 * even on failure.
80 	 */
81 	unsigned int		 hio_countdown;
82 	/*
83 	 * Each component has a place to store its own error.
84 	 * Once the request is handled by all components we can decide if the
85 	 * request overall is successful or not.
86 	 */
87 	int			*hio_errors;
88 	/*
89 	 * Structure used to communicate with GEOM Gate class.
90 	 */
91 	struct g_gate_ctl_io	 hio_ggio;
92 	TAILQ_ENTRY(hio)	*hio_next;
93 };
94 #define	hio_free_next	hio_next[0]
95 #define	hio_done_next	hio_next[0]
96 
97 /*
98  * Free list holds unused structures. When free list is empty, we have to wait
99  * until some in-progress requests are freed.
100  */
101 static TAILQ_HEAD(, hio) hio_free_list;
102 static pthread_mutex_t hio_free_list_lock;
103 static pthread_cond_t hio_free_list_cond;
104 /*
105  * There is one send list for every component. One requests is placed on all
106  * send lists - each component gets the same request, but each component is
107  * responsible for managing his own send list.
108  */
109 static TAILQ_HEAD(, hio) *hio_send_list;
110 static pthread_mutex_t *hio_send_list_lock;
111 static pthread_cond_t *hio_send_list_cond;
112 /*
113  * There is one recv list for every component, although local components don't
114  * use recv lists as local requests are done synchronously.
115  */
116 static TAILQ_HEAD(, hio) *hio_recv_list;
117 static pthread_mutex_t *hio_recv_list_lock;
118 static pthread_cond_t *hio_recv_list_cond;
119 /*
120  * Request is placed on done list by the slowest component (the one that
121  * decreased hio_countdown from 1 to 0).
122  */
123 static TAILQ_HEAD(, hio) hio_done_list;
124 static pthread_mutex_t hio_done_list_lock;
125 static pthread_cond_t hio_done_list_cond;
126 /*
127  * Structure below are for interaction with sync thread.
128  */
129 static bool sync_inprogress;
130 static pthread_mutex_t sync_lock;
131 static pthread_cond_t sync_cond;
132 /*
133  * The lock below allows to synchornize access to remote connections.
134  */
135 static pthread_rwlock_t *hio_remote_lock;
136 
137 /*
138  * Lock to synchronize metadata updates. Also synchronize access to
139  * hr_primary_localcnt and hr_primary_remotecnt fields.
140  */
141 static pthread_mutex_t metadata_lock;
142 
143 /*
144  * Maximum number of outstanding I/O requests.
145  */
146 #define	HAST_HIO_MAX	256
147 /*
148  * Number of components. At this point there are only two components: local
149  * and remote, but in the future it might be possible to use multiple local
150  * and remote components.
151  */
152 #define	HAST_NCOMPONENTS	2
153 
154 #define	ISCONNECTED(res, no)	\
155 	((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
156 
157 #define	QUEUE_INSERT1(hio, name, ncomp)	do {				\
158 	bool _wakeup;							\
159 									\
160 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
161 	_wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]);		\
162 	TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),		\
163 	    hio_next[(ncomp)]);						\
164 	mtx_unlock(&hio_##name##_list_lock[ncomp]);			\
165 	if (_wakeup)							\
166 		cv_signal(&hio_##name##_list_cond[(ncomp)]);		\
167 } while (0)
168 #define	QUEUE_INSERT2(hio, name)	do {				\
169 	bool _wakeup;							\
170 									\
171 	mtx_lock(&hio_##name##_list_lock);				\
172 	_wakeup = TAILQ_EMPTY(&hio_##name##_list);			\
173 	TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
174 	mtx_unlock(&hio_##name##_list_lock);				\
175 	if (_wakeup)							\
176 		cv_signal(&hio_##name##_list_cond);			\
177 } while (0)
178 #define	QUEUE_TAKE1(hio, name, ncomp, timeout)	do {			\
179 	bool _last;							\
180 									\
181 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
182 	_last = false;							\
183 	while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
184 		cv_timedwait(&hio_##name##_list_cond[(ncomp)],		\
185 		    &hio_##name##_list_lock[(ncomp)], (timeout));	\
186 		if ((timeout) != 0)					\
187 			_last = true;					\
188 	}								\
189 	if (hio != NULL) {						\
190 		TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),	\
191 		    hio_next[(ncomp)]);					\
192 	}								\
193 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
194 } while (0)
195 #define	QUEUE_TAKE2(hio, name)	do {					\
196 	mtx_lock(&hio_##name##_list_lock);				\
197 	while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {	\
198 		cv_wait(&hio_##name##_list_cond,			\
199 		    &hio_##name##_list_lock);				\
200 	}								\
201 	TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);	\
202 	mtx_unlock(&hio_##name##_list_lock);				\
203 } while (0)
204 
205 #define	SYNCREQ(hio)		do {					\
206 	(hio)->hio_ggio.gctl_unit = -1;					\
207 	(hio)->hio_ggio.gctl_seq = 1;					\
208 } while (0)
209 #define	ISSYNCREQ(hio)		((hio)->hio_ggio.gctl_unit == -1)
210 #define	SYNCREQDONE(hio)	do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
211 #define	ISSYNCREQDONE(hio)	((hio)->hio_ggio.gctl_unit == -2)
212 
213 static struct hast_resource *gres;
214 
215 static pthread_mutex_t range_lock;
216 static struct rangelocks *range_regular;
217 static bool range_regular_wait;
218 static pthread_cond_t range_regular_cond;
219 static struct rangelocks *range_sync;
220 static bool range_sync_wait;
221 static pthread_cond_t range_sync_cond;
222 static bool fullystarted;
223 
224 static void *ggate_recv_thread(void *arg);
225 static void *local_send_thread(void *arg);
226 static void *remote_send_thread(void *arg);
227 static void *remote_recv_thread(void *arg);
228 static void *ggate_send_thread(void *arg);
229 static void *sync_thread(void *arg);
230 static void *guard_thread(void *arg);
231 
232 static void
233 cleanup(struct hast_resource *res)
234 {
235 	int rerrno;
236 
237 	/* Remember errno. */
238 	rerrno = errno;
239 
240 	/* Destroy ggate provider if we created one. */
241 	if (res->hr_ggateunit >= 0) {
242 		struct g_gate_ctl_destroy ggiod;
243 
244 		bzero(&ggiod, sizeof(ggiod));
245 		ggiod.gctl_version = G_GATE_VERSION;
246 		ggiod.gctl_unit = res->hr_ggateunit;
247 		ggiod.gctl_force = 1;
248 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) {
249 			pjdlog_errno(LOG_WARNING,
250 			    "Unable to destroy hast/%s device",
251 			    res->hr_provname);
252 		}
253 		res->hr_ggateunit = -1;
254 	}
255 
256 	/* Restore errno. */
257 	errno = rerrno;
258 }
259 
260 static __dead2 void
261 primary_exit(int exitcode, const char *fmt, ...)
262 {
263 	va_list ap;
264 
265 	PJDLOG_ASSERT(exitcode != EX_OK);
266 	va_start(ap, fmt);
267 	pjdlogv_errno(LOG_ERR, fmt, ap);
268 	va_end(ap);
269 	cleanup(gres);
270 	exit(exitcode);
271 }
272 
273 static __dead2 void
274 primary_exitx(int exitcode, const char *fmt, ...)
275 {
276 	va_list ap;
277 
278 	va_start(ap, fmt);
279 	pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
280 	va_end(ap);
281 	cleanup(gres);
282 	exit(exitcode);
283 }
284 
285 static int
286 hast_activemap_flush(struct hast_resource *res)
287 {
288 	const unsigned char *buf;
289 	size_t size;
290 
291 	buf = activemap_bitmap(res->hr_amp, &size);
292 	PJDLOG_ASSERT(buf != NULL);
293 	PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
294 	if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
295 	    (ssize_t)size) {
296 		pjdlog_errno(LOG_ERR, "Unable to flush activemap to disk");
297 		return (-1);
298 	}
299 	if (res->hr_metaflush == 1 && g_flush(res->hr_localfd) == -1) {
300 		if (errno == EOPNOTSUPP) {
301 			pjdlog_warning("The %s provider doesn't support flushing write cache. Disabling it.",
302 			    res->hr_localpath);
303 			res->hr_metaflush = 0;
304 		} else {
305 			pjdlog_errno(LOG_ERR,
306 			    "Unable to flush disk cache on activemap update");
307 			return (-1);
308 		}
309 	}
310 	return (0);
311 }
312 
313 static bool
314 real_remote(const struct hast_resource *res)
315 {
316 
317 	return (strcmp(res->hr_remoteaddr, "none") != 0);
318 }
319 
320 static void
321 init_environment(struct hast_resource *res __unused)
322 {
323 	struct hio *hio;
324 	unsigned int ii, ncomps;
325 
326 	/*
327 	 * In the future it might be per-resource value.
328 	 */
329 	ncomps = HAST_NCOMPONENTS;
330 
331 	/*
332 	 * Allocate memory needed by lists.
333 	 */
334 	hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
335 	if (hio_send_list == NULL) {
336 		primary_exitx(EX_TEMPFAIL,
337 		    "Unable to allocate %zu bytes of memory for send lists.",
338 		    sizeof(hio_send_list[0]) * ncomps);
339 	}
340 	hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
341 	if (hio_send_list_lock == NULL) {
342 		primary_exitx(EX_TEMPFAIL,
343 		    "Unable to allocate %zu bytes of memory for send list locks.",
344 		    sizeof(hio_send_list_lock[0]) * ncomps);
345 	}
346 	hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
347 	if (hio_send_list_cond == NULL) {
348 		primary_exitx(EX_TEMPFAIL,
349 		    "Unable to allocate %zu bytes of memory for send list condition variables.",
350 		    sizeof(hio_send_list_cond[0]) * ncomps);
351 	}
352 	hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
353 	if (hio_recv_list == NULL) {
354 		primary_exitx(EX_TEMPFAIL,
355 		    "Unable to allocate %zu bytes of memory for recv lists.",
356 		    sizeof(hio_recv_list[0]) * ncomps);
357 	}
358 	hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
359 	if (hio_recv_list_lock == NULL) {
360 		primary_exitx(EX_TEMPFAIL,
361 		    "Unable to allocate %zu bytes of memory for recv list locks.",
362 		    sizeof(hio_recv_list_lock[0]) * ncomps);
363 	}
364 	hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
365 	if (hio_recv_list_cond == NULL) {
366 		primary_exitx(EX_TEMPFAIL,
367 		    "Unable to allocate %zu bytes of memory for recv list condition variables.",
368 		    sizeof(hio_recv_list_cond[0]) * ncomps);
369 	}
370 	hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
371 	if (hio_remote_lock == NULL) {
372 		primary_exitx(EX_TEMPFAIL,
373 		    "Unable to allocate %zu bytes of memory for remote connections locks.",
374 		    sizeof(hio_remote_lock[0]) * ncomps);
375 	}
376 
377 	/*
378 	 * Initialize lists, their locks and theirs condition variables.
379 	 */
380 	TAILQ_INIT(&hio_free_list);
381 	mtx_init(&hio_free_list_lock);
382 	cv_init(&hio_free_list_cond);
383 	for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
384 		TAILQ_INIT(&hio_send_list[ii]);
385 		mtx_init(&hio_send_list_lock[ii]);
386 		cv_init(&hio_send_list_cond[ii]);
387 		TAILQ_INIT(&hio_recv_list[ii]);
388 		mtx_init(&hio_recv_list_lock[ii]);
389 		cv_init(&hio_recv_list_cond[ii]);
390 		rw_init(&hio_remote_lock[ii]);
391 	}
392 	TAILQ_INIT(&hio_done_list);
393 	mtx_init(&hio_done_list_lock);
394 	cv_init(&hio_done_list_cond);
395 	mtx_init(&metadata_lock);
396 
397 	/*
398 	 * Allocate requests pool and initialize requests.
399 	 */
400 	for (ii = 0; ii < HAST_HIO_MAX; ii++) {
401 		hio = malloc(sizeof(*hio));
402 		if (hio == NULL) {
403 			primary_exitx(EX_TEMPFAIL,
404 			    "Unable to allocate %zu bytes of memory for hio request.",
405 			    sizeof(*hio));
406 		}
407 		hio->hio_countdown = 0;
408 		hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
409 		if (hio->hio_errors == NULL) {
410 			primary_exitx(EX_TEMPFAIL,
411 			    "Unable allocate %zu bytes of memory for hio errors.",
412 			    sizeof(hio->hio_errors[0]) * ncomps);
413 		}
414 		hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
415 		if (hio->hio_next == NULL) {
416 			primary_exitx(EX_TEMPFAIL,
417 			    "Unable allocate %zu bytes of memory for hio_next field.",
418 			    sizeof(hio->hio_next[0]) * ncomps);
419 		}
420 		hio->hio_ggio.gctl_version = G_GATE_VERSION;
421 		hio->hio_ggio.gctl_data = malloc(MAXPHYS);
422 		if (hio->hio_ggio.gctl_data == NULL) {
423 			primary_exitx(EX_TEMPFAIL,
424 			    "Unable to allocate %zu bytes of memory for gctl_data.",
425 			    MAXPHYS);
426 		}
427 		hio->hio_ggio.gctl_length = MAXPHYS;
428 		hio->hio_ggio.gctl_error = 0;
429 		TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
430 	}
431 }
432 
433 static bool
434 init_resuid(struct hast_resource *res)
435 {
436 
437 	mtx_lock(&metadata_lock);
438 	if (res->hr_resuid != 0) {
439 		mtx_unlock(&metadata_lock);
440 		return (false);
441 	} else {
442 		/* Initialize unique resource identifier. */
443 		arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
444 		mtx_unlock(&metadata_lock);
445 		if (metadata_write(res) < 0)
446 			exit(EX_NOINPUT);
447 		return (true);
448 	}
449 }
450 
451 static void
452 init_local(struct hast_resource *res)
453 {
454 	unsigned char *buf;
455 	size_t mapsize;
456 
457 	if (metadata_read(res, true) < 0)
458 		exit(EX_NOINPUT);
459 	mtx_init(&res->hr_amp_lock);
460 	if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
461 	    res->hr_local_sectorsize, res->hr_keepdirty) < 0) {
462 		primary_exit(EX_TEMPFAIL, "Unable to create activemap");
463 	}
464 	mtx_init(&range_lock);
465 	cv_init(&range_regular_cond);
466 	if (rangelock_init(&range_regular) < 0)
467 		primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
468 	cv_init(&range_sync_cond);
469 	if (rangelock_init(&range_sync) < 0)
470 		primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
471 	mapsize = activemap_ondisk_size(res->hr_amp);
472 	buf = calloc(1, mapsize);
473 	if (buf == NULL) {
474 		primary_exitx(EX_TEMPFAIL,
475 		    "Unable to allocate buffer for activemap.");
476 	}
477 	if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
478 	    (ssize_t)mapsize) {
479 		primary_exit(EX_NOINPUT, "Unable to read activemap");
480 	}
481 	activemap_copyin(res->hr_amp, buf, mapsize);
482 	free(buf);
483 	if (res->hr_resuid != 0)
484 		return;
485 	/*
486 	 * We're using provider for the first time. Initialize local and remote
487 	 * counters. We don't initialize resuid here, as we want to do it just
488 	 * in time. The reason for this is that we want to inform secondary
489 	 * that there were no writes yet, so there is no need to synchronize
490 	 * anything.
491 	 */
492 	res->hr_primary_localcnt = 0;
493 	res->hr_primary_remotecnt = 0;
494 	if (metadata_write(res) < 0)
495 		exit(EX_NOINPUT);
496 }
497 
498 static int
499 primary_connect(struct hast_resource *res, struct proto_conn **connp)
500 {
501 	struct proto_conn *conn;
502 	int16_t val;
503 
504 	val = 1;
505 	if (proto_send(res->hr_conn, &val, sizeof(val)) < 0) {
506 		primary_exit(EX_TEMPFAIL,
507 		    "Unable to send connection request to parent");
508 	}
509 	if (proto_recv(res->hr_conn, &val, sizeof(val)) < 0) {
510 		primary_exit(EX_TEMPFAIL,
511 		    "Unable to receive reply to connection request from parent");
512 	}
513 	if (val != 0) {
514 		errno = val;
515 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
516 		    res->hr_remoteaddr);
517 		return (-1);
518 	}
519 	if (proto_connection_recv(res->hr_conn, true, &conn) < 0) {
520 		primary_exit(EX_TEMPFAIL,
521 		    "Unable to receive connection from parent");
522 	}
523 	if (proto_connect_wait(conn, res->hr_timeout) < 0) {
524 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
525 		    res->hr_remoteaddr);
526 		proto_close(conn);
527 		return (-1);
528 	}
529 	/* Error in setting timeout is not critical, but why should it fail? */
530 	if (proto_timeout(conn, res->hr_timeout) < 0)
531 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
532 
533 	*connp = conn;
534 
535 	return (0);
536 }
537 
538 static int
539 init_remote(struct hast_resource *res, struct proto_conn **inp,
540     struct proto_conn **outp)
541 {
542 	struct proto_conn *in, *out;
543 	struct nv *nvout, *nvin;
544 	const unsigned char *token;
545 	unsigned char *map;
546 	const char *errmsg;
547 	int32_t extentsize;
548 	int64_t datasize;
549 	uint32_t mapsize;
550 	size_t size;
551 	int error;
552 
553 	PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
554 	PJDLOG_ASSERT(real_remote(res));
555 
556 	in = out = NULL;
557 	errmsg = NULL;
558 
559 	if (primary_connect(res, &out) == -1)
560 		return (ECONNREFUSED);
561 
562 	error = ECONNABORTED;
563 
564 	/*
565 	 * First handshake step.
566 	 * Setup outgoing connection with remote node.
567 	 */
568 	nvout = nv_alloc();
569 	nv_add_string(nvout, res->hr_name, "resource");
570 	if (nv_error(nvout) != 0) {
571 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
572 		    "Unable to allocate header for connection with %s",
573 		    res->hr_remoteaddr);
574 		nv_free(nvout);
575 		goto close;
576 	}
577 	if (hast_proto_send(res, out, nvout, NULL, 0) < 0) {
578 		pjdlog_errno(LOG_WARNING,
579 		    "Unable to send handshake header to %s",
580 		    res->hr_remoteaddr);
581 		nv_free(nvout);
582 		goto close;
583 	}
584 	nv_free(nvout);
585 	if (hast_proto_recv_hdr(out, &nvin) < 0) {
586 		pjdlog_errno(LOG_WARNING,
587 		    "Unable to receive handshake header from %s",
588 		    res->hr_remoteaddr);
589 		goto close;
590 	}
591 	errmsg = nv_get_string(nvin, "errmsg");
592 	if (errmsg != NULL) {
593 		pjdlog_warning("%s", errmsg);
594 		if (nv_exists(nvin, "wait"))
595 			error = EBUSY;
596 		nv_free(nvin);
597 		goto close;
598 	}
599 	token = nv_get_uint8_array(nvin, &size, "token");
600 	if (token == NULL) {
601 		pjdlog_warning("Handshake header from %s has no 'token' field.",
602 		    res->hr_remoteaddr);
603 		nv_free(nvin);
604 		goto close;
605 	}
606 	if (size != sizeof(res->hr_token)) {
607 		pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
608 		    res->hr_remoteaddr, size, sizeof(res->hr_token));
609 		nv_free(nvin);
610 		goto close;
611 	}
612 	bcopy(token, res->hr_token, sizeof(res->hr_token));
613 	nv_free(nvin);
614 
615 	/*
616 	 * Second handshake step.
617 	 * Setup incoming connection with remote node.
618 	 */
619 	if (primary_connect(res, &in) == -1)
620 		goto close;
621 
622 	nvout = nv_alloc();
623 	nv_add_string(nvout, res->hr_name, "resource");
624 	nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
625 	    "token");
626 	if (res->hr_resuid == 0) {
627 		/*
628 		 * The resuid field was not yet initialized.
629 		 * Because we do synchronization inside init_resuid(), it is
630 		 * possible that someone already initialized it, the function
631 		 * will return false then, but if we successfully initialized
632 		 * it, we will get true. True means that there were no writes
633 		 * to this resource yet and we want to inform secondary that
634 		 * synchronization is not needed by sending "virgin" argument.
635 		 */
636 		if (init_resuid(res))
637 			nv_add_int8(nvout, 1, "virgin");
638 	}
639 	nv_add_uint64(nvout, res->hr_resuid, "resuid");
640 	nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
641 	nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
642 	if (nv_error(nvout) != 0) {
643 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
644 		    "Unable to allocate header for connection with %s",
645 		    res->hr_remoteaddr);
646 		nv_free(nvout);
647 		goto close;
648 	}
649 	if (hast_proto_send(res, in, nvout, NULL, 0) < 0) {
650 		pjdlog_errno(LOG_WARNING,
651 		    "Unable to send handshake header to %s",
652 		    res->hr_remoteaddr);
653 		nv_free(nvout);
654 		goto close;
655 	}
656 	nv_free(nvout);
657 	if (hast_proto_recv_hdr(out, &nvin) < 0) {
658 		pjdlog_errno(LOG_WARNING,
659 		    "Unable to receive handshake header from %s",
660 		    res->hr_remoteaddr);
661 		goto close;
662 	}
663 	errmsg = nv_get_string(nvin, "errmsg");
664 	if (errmsg != NULL) {
665 		pjdlog_warning("%s", errmsg);
666 		nv_free(nvin);
667 		goto close;
668 	}
669 	datasize = nv_get_int64(nvin, "datasize");
670 	if (datasize != res->hr_datasize) {
671 		pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
672 		    (intmax_t)res->hr_datasize, (intmax_t)datasize);
673 		nv_free(nvin);
674 		goto close;
675 	}
676 	extentsize = nv_get_int32(nvin, "extentsize");
677 	if (extentsize != res->hr_extentsize) {
678 		pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
679 		    (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
680 		nv_free(nvin);
681 		goto close;
682 	}
683 	res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
684 	res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
685 	res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
686 	if (nv_exists(nvin, "virgin")) {
687 		/*
688 		 * Secondary was reinitialized, bump localcnt if it is 0 as
689 		 * only we have the data.
690 		 */
691 		PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY);
692 		PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
693 
694 		if (res->hr_primary_localcnt == 0) {
695 			PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0);
696 
697 			mtx_lock(&metadata_lock);
698 			res->hr_primary_localcnt++;
699 			pjdlog_debug(1, "Increasing localcnt to %ju.",
700 			    (uintmax_t)res->hr_primary_localcnt);
701 			(void)metadata_write(res);
702 			mtx_unlock(&metadata_lock);
703 		}
704 	}
705 	map = NULL;
706 	mapsize = nv_get_uint32(nvin, "mapsize");
707 	if (mapsize > 0) {
708 		map = malloc(mapsize);
709 		if (map == NULL) {
710 			pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
711 			    (uintmax_t)mapsize);
712 			nv_free(nvin);
713 			goto close;
714 		}
715 		/*
716 		 * Remote node have some dirty extents on its own, lets
717 		 * download its activemap.
718 		 */
719 		if (hast_proto_recv_data(res, out, nvin, map,
720 		    mapsize) < 0) {
721 			pjdlog_errno(LOG_ERR,
722 			    "Unable to receive remote activemap");
723 			nv_free(nvin);
724 			free(map);
725 			goto close;
726 		}
727 		/*
728 		 * Merge local and remote bitmaps.
729 		 */
730 		activemap_merge(res->hr_amp, map, mapsize);
731 		free(map);
732 		/*
733 		 * Now that we merged bitmaps from both nodes, flush it to the
734 		 * disk before we start to synchronize.
735 		 */
736 		(void)hast_activemap_flush(res);
737 	}
738 	nv_free(nvin);
739 #ifdef notyet
740 	/* Setup directions. */
741 	if (proto_send(out, NULL, 0) == -1)
742 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
743 	if (proto_recv(in, NULL, 0) == -1)
744 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
745 #endif
746 	pjdlog_info("Connected to %s.", res->hr_remoteaddr);
747 	if (inp != NULL && outp != NULL) {
748 		*inp = in;
749 		*outp = out;
750 	} else {
751 		res->hr_remotein = in;
752 		res->hr_remoteout = out;
753 	}
754 	event_send(res, EVENT_CONNECT);
755 	return (0);
756 close:
757 	if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
758 		event_send(res, EVENT_SPLITBRAIN);
759 	proto_close(out);
760 	if (in != NULL)
761 		proto_close(in);
762 	return (error);
763 }
764 
765 static void
766 sync_start(void)
767 {
768 
769 	mtx_lock(&sync_lock);
770 	sync_inprogress = true;
771 	mtx_unlock(&sync_lock);
772 	cv_signal(&sync_cond);
773 }
774 
775 static void
776 sync_stop(void)
777 {
778 
779 	mtx_lock(&sync_lock);
780 	if (sync_inprogress)
781 		sync_inprogress = false;
782 	mtx_unlock(&sync_lock);
783 }
784 
785 static void
786 init_ggate(struct hast_resource *res)
787 {
788 	struct g_gate_ctl_create ggiocreate;
789 	struct g_gate_ctl_cancel ggiocancel;
790 
791 	/*
792 	 * We communicate with ggate via /dev/ggctl. Open it.
793 	 */
794 	res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
795 	if (res->hr_ggatefd < 0)
796 		primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
797 	/*
798 	 * Create provider before trying to connect, as connection failure
799 	 * is not critical, but may take some time.
800 	 */
801 	bzero(&ggiocreate, sizeof(ggiocreate));
802 	ggiocreate.gctl_version = G_GATE_VERSION;
803 	ggiocreate.gctl_mediasize = res->hr_datasize;
804 	ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
805 	ggiocreate.gctl_flags = 0;
806 	ggiocreate.gctl_maxcount = 0;
807 	ggiocreate.gctl_timeout = 0;
808 	ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
809 	snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
810 	    res->hr_provname);
811 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
812 		pjdlog_info("Device hast/%s created.", res->hr_provname);
813 		res->hr_ggateunit = ggiocreate.gctl_unit;
814 		return;
815 	}
816 	if (errno != EEXIST) {
817 		primary_exit(EX_OSERR, "Unable to create hast/%s device",
818 		    res->hr_provname);
819 	}
820 	pjdlog_debug(1,
821 	    "Device hast/%s already exists, we will try to take it over.",
822 	    res->hr_provname);
823 	/*
824 	 * If we received EEXIST, we assume that the process who created the
825 	 * provider died and didn't clean up. In that case we will start from
826 	 * where he left of.
827 	 */
828 	bzero(&ggiocancel, sizeof(ggiocancel));
829 	ggiocancel.gctl_version = G_GATE_VERSION;
830 	ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
831 	snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
832 	    res->hr_provname);
833 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
834 		pjdlog_info("Device hast/%s recovered.", res->hr_provname);
835 		res->hr_ggateunit = ggiocancel.gctl_unit;
836 		return;
837 	}
838 	primary_exit(EX_OSERR, "Unable to take over hast/%s device",
839 	    res->hr_provname);
840 }
841 
842 void
843 hastd_primary(struct hast_resource *res)
844 {
845 	pthread_t td;
846 	pid_t pid;
847 	int error, mode, debuglevel;
848 
849 	/*
850 	 * Create communication channel for sending control commands from
851 	 * parent to child.
852 	 */
853 	if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) {
854 		/* TODO: There's no need for this to be fatal error. */
855 		KEEP_ERRNO((void)pidfile_remove(pfh));
856 		pjdlog_exit(EX_OSERR,
857 		    "Unable to create control sockets between parent and child");
858 	}
859 	/*
860 	 * Create communication channel for sending events from child to parent.
861 	 */
862 	if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) {
863 		/* TODO: There's no need for this to be fatal error. */
864 		KEEP_ERRNO((void)pidfile_remove(pfh));
865 		pjdlog_exit(EX_OSERR,
866 		    "Unable to create event sockets between child and parent");
867 	}
868 	/*
869 	 * Create communication channel for sending connection requests from
870 	 * child to parent.
871 	 */
872 	if (proto_client(NULL, "socketpair://", &res->hr_conn) < 0) {
873 		/* TODO: There's no need for this to be fatal error. */
874 		KEEP_ERRNO((void)pidfile_remove(pfh));
875 		pjdlog_exit(EX_OSERR,
876 		    "Unable to create connection sockets between child and parent");
877 	}
878 
879 	pid = fork();
880 	if (pid < 0) {
881 		/* TODO: There's no need for this to be fatal error. */
882 		KEEP_ERRNO((void)pidfile_remove(pfh));
883 		pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
884 	}
885 
886 	if (pid > 0) {
887 		/* This is parent. */
888 		/* Declare that we are receiver. */
889 		proto_recv(res->hr_event, NULL, 0);
890 		proto_recv(res->hr_conn, NULL, 0);
891 		/* Declare that we are sender. */
892 		proto_send(res->hr_ctrl, NULL, 0);
893 		res->hr_workerpid = pid;
894 		return;
895 	}
896 
897 	gres = res;
898 	mode = pjdlog_mode_get();
899 	debuglevel = pjdlog_debug_get();
900 
901 	/* Declare that we are sender. */
902 	proto_send(res->hr_event, NULL, 0);
903 	proto_send(res->hr_conn, NULL, 0);
904 	/* Declare that we are receiver. */
905 	proto_recv(res->hr_ctrl, NULL, 0);
906 	descriptors_cleanup(res);
907 
908 	descriptors_assert(res, mode);
909 
910 	pjdlog_init(mode);
911 	pjdlog_debug_set(debuglevel);
912 	pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
913 	setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
914 
915 	init_local(res);
916 	init_ggate(res);
917 	init_environment(res);
918 
919 	if (drop_privs(res) != 0) {
920 		cleanup(res);
921 		exit(EX_CONFIG);
922 	}
923 	pjdlog_info("Privileges successfully dropped.");
924 
925 	/*
926 	 * Create the guard thread first, so we can handle signals from the
927 	 * very begining.
928 	 */
929 	error = pthread_create(&td, NULL, guard_thread, res);
930 	PJDLOG_ASSERT(error == 0);
931 	/*
932 	 * Create the control thread before sending any event to the parent,
933 	 * as we can deadlock when parent sends control request to worker,
934 	 * but worker has no control thread started yet, so parent waits.
935 	 * In the meantime worker sends an event to the parent, but parent
936 	 * is unable to handle the event, because it waits for control
937 	 * request response.
938 	 */
939 	error = pthread_create(&td, NULL, ctrl_thread, res);
940 	PJDLOG_ASSERT(error == 0);
941 	if (real_remote(res)) {
942 		error = init_remote(res, NULL, NULL);
943 		if (error == 0) {
944 			sync_start();
945 		} else if (error == EBUSY) {
946 			time_t start = time(NULL);
947 
948 			pjdlog_warning("Waiting for remote node to become %s for %ds.",
949 			    role2str(HAST_ROLE_SECONDARY),
950 			    res->hr_timeout);
951 			for (;;) {
952 				sleep(1);
953 				error = init_remote(res, NULL, NULL);
954 				if (error != EBUSY)
955 					break;
956 				if (time(NULL) > start + res->hr_timeout)
957 					break;
958 			}
959 			if (error == EBUSY) {
960 				pjdlog_warning("Remote node is still %s, starting anyway.",
961 				    role2str(HAST_ROLE_PRIMARY));
962 			}
963 		}
964 	}
965 	error = pthread_create(&td, NULL, ggate_recv_thread, res);
966 	PJDLOG_ASSERT(error == 0);
967 	error = pthread_create(&td, NULL, local_send_thread, res);
968 	PJDLOG_ASSERT(error == 0);
969 	error = pthread_create(&td, NULL, remote_send_thread, res);
970 	PJDLOG_ASSERT(error == 0);
971 	error = pthread_create(&td, NULL, remote_recv_thread, res);
972 	PJDLOG_ASSERT(error == 0);
973 	error = pthread_create(&td, NULL, ggate_send_thread, res);
974 	PJDLOG_ASSERT(error == 0);
975 	fullystarted = true;
976 	(void)sync_thread(res);
977 }
978 
979 static void
980 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...)
981 {
982 	char msg[1024];
983 	va_list ap;
984 	int len;
985 
986 	va_start(ap, fmt);
987 	len = vsnprintf(msg, sizeof(msg), fmt, ap);
988 	va_end(ap);
989 	if ((size_t)len < sizeof(msg)) {
990 		switch (ggio->gctl_cmd) {
991 		case BIO_READ:
992 			(void)snprintf(msg + len, sizeof(msg) - len,
993 			    "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
994 			    (uintmax_t)ggio->gctl_length);
995 			break;
996 		case BIO_DELETE:
997 			(void)snprintf(msg + len, sizeof(msg) - len,
998 			    "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
999 			    (uintmax_t)ggio->gctl_length);
1000 			break;
1001 		case BIO_FLUSH:
1002 			(void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
1003 			break;
1004 		case BIO_WRITE:
1005 			(void)snprintf(msg + len, sizeof(msg) - len,
1006 			    "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
1007 			    (uintmax_t)ggio->gctl_length);
1008 			break;
1009 		default:
1010 			(void)snprintf(msg + len, sizeof(msg) - len,
1011 			    "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd);
1012 			break;
1013 		}
1014 	}
1015 	pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
1016 }
1017 
1018 static void
1019 remote_close(struct hast_resource *res, int ncomp)
1020 {
1021 
1022 	rw_wlock(&hio_remote_lock[ncomp]);
1023 	/*
1024 	 * A race is possible between dropping rlock and acquiring wlock -
1025 	 * another thread can close connection in-between.
1026 	 */
1027 	if (!ISCONNECTED(res, ncomp)) {
1028 		PJDLOG_ASSERT(res->hr_remotein == NULL);
1029 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
1030 		rw_unlock(&hio_remote_lock[ncomp]);
1031 		return;
1032 	}
1033 
1034 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1035 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1036 
1037 	pjdlog_debug(2, "Closing incoming connection to %s.",
1038 	    res->hr_remoteaddr);
1039 	proto_close(res->hr_remotein);
1040 	res->hr_remotein = NULL;
1041 	pjdlog_debug(2, "Closing outgoing connection to %s.",
1042 	    res->hr_remoteaddr);
1043 	proto_close(res->hr_remoteout);
1044 	res->hr_remoteout = NULL;
1045 
1046 	rw_unlock(&hio_remote_lock[ncomp]);
1047 
1048 	pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
1049 
1050 	/*
1051 	 * Stop synchronization if in-progress.
1052 	 */
1053 	sync_stop();
1054 
1055 	event_send(res, EVENT_DISCONNECT);
1056 }
1057 
1058 /*
1059  * Thread receives ggate I/O requests from the kernel and passes them to
1060  * appropriate threads:
1061  * WRITE - always goes to both local_send and remote_send threads
1062  * READ (when the block is up-to-date on local component) -
1063  *	only local_send thread
1064  * READ (when the block isn't up-to-date on local component) -
1065  *	only remote_send thread
1066  * DELETE - always goes to both local_send and remote_send threads
1067  * FLUSH - always goes to both local_send and remote_send threads
1068  */
1069 static void *
1070 ggate_recv_thread(void *arg)
1071 {
1072 	struct hast_resource *res = arg;
1073 	struct g_gate_ctl_io *ggio;
1074 	struct hio *hio;
1075 	unsigned int ii, ncomp, ncomps;
1076 	int error;
1077 
1078 	ncomps = HAST_NCOMPONENTS;
1079 
1080 	for (;;) {
1081 		pjdlog_debug(2, "ggate_recv: Taking free request.");
1082 		QUEUE_TAKE2(hio, free);
1083 		pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1084 		ggio = &hio->hio_ggio;
1085 		ggio->gctl_unit = res->hr_ggateunit;
1086 		ggio->gctl_length = MAXPHYS;
1087 		ggio->gctl_error = 0;
1088 		pjdlog_debug(2,
1089 		    "ggate_recv: (%p) Waiting for request from the kernel.",
1090 		    hio);
1091 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) {
1092 			if (sigexit_received)
1093 				pthread_exit(NULL);
1094 			primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1095 		}
1096 		error = ggio->gctl_error;
1097 		switch (error) {
1098 		case 0:
1099 			break;
1100 		case ECANCELED:
1101 			/* Exit gracefully. */
1102 			if (!sigexit_received) {
1103 				pjdlog_debug(2,
1104 				    "ggate_recv: (%p) Received cancel from the kernel.",
1105 				    hio);
1106 				pjdlog_info("Received cancel from the kernel, exiting.");
1107 			}
1108 			pthread_exit(NULL);
1109 		case ENOMEM:
1110 			/*
1111 			 * Buffer too small? Impossible, we allocate MAXPHYS
1112 			 * bytes - request can't be bigger than that.
1113 			 */
1114 			/* FALLTHROUGH */
1115 		case ENXIO:
1116 		default:
1117 			primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1118 			    strerror(error));
1119 		}
1120 		for (ii = 0; ii < ncomps; ii++)
1121 			hio->hio_errors[ii] = EINVAL;
1122 		reqlog(LOG_DEBUG, 2, ggio,
1123 		    "ggate_recv: (%p) Request received from the kernel: ",
1124 		    hio);
1125 		/*
1126 		 * Inform all components about new write request.
1127 		 * For read request prefer local component unless the given
1128 		 * range is out-of-date, then use remote component.
1129 		 */
1130 		switch (ggio->gctl_cmd) {
1131 		case BIO_READ:
1132 			res->hr_stat_read++;
1133 			pjdlog_debug(2,
1134 			    "ggate_recv: (%p) Moving request to the send queue.",
1135 			    hio);
1136 			refcount_init(&hio->hio_countdown, 1);
1137 			mtx_lock(&metadata_lock);
1138 			if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1139 			    res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1140 				/*
1141 				 * This range is up-to-date on local component,
1142 				 * so handle request locally.
1143 				 */
1144 				 /* Local component is 0 for now. */
1145 				ncomp = 0;
1146 			} else /* if (res->hr_syncsrc ==
1147 			    HAST_SYNCSRC_SECONDARY) */ {
1148 				PJDLOG_ASSERT(res->hr_syncsrc ==
1149 				    HAST_SYNCSRC_SECONDARY);
1150 				/*
1151 				 * This range is out-of-date on local component,
1152 				 * so send request to the remote node.
1153 				 */
1154 				 /* Remote component is 1 for now. */
1155 				ncomp = 1;
1156 			}
1157 			mtx_unlock(&metadata_lock);
1158 			QUEUE_INSERT1(hio, send, ncomp);
1159 			break;
1160 		case BIO_WRITE:
1161 			res->hr_stat_write++;
1162 			if (res->hr_resuid == 0) {
1163 				/*
1164 				 * This is first write, initialize localcnt and
1165 				 * resuid.
1166 				 */
1167 				res->hr_primary_localcnt = 1;
1168 				(void)init_resuid(res);
1169 			}
1170 			for (;;) {
1171 				mtx_lock(&range_lock);
1172 				if (rangelock_islocked(range_sync,
1173 				    ggio->gctl_offset, ggio->gctl_length)) {
1174 					pjdlog_debug(2,
1175 					    "regular: Range offset=%jd length=%zu locked.",
1176 					    (intmax_t)ggio->gctl_offset,
1177 					    (size_t)ggio->gctl_length);
1178 					range_regular_wait = true;
1179 					cv_wait(&range_regular_cond, &range_lock);
1180 					range_regular_wait = false;
1181 					mtx_unlock(&range_lock);
1182 					continue;
1183 				}
1184 				if (rangelock_add(range_regular,
1185 				    ggio->gctl_offset, ggio->gctl_length) < 0) {
1186 					mtx_unlock(&range_lock);
1187 					pjdlog_debug(2,
1188 					    "regular: Range offset=%jd length=%zu is already locked, waiting.",
1189 					    (intmax_t)ggio->gctl_offset,
1190 					    (size_t)ggio->gctl_length);
1191 					sleep(1);
1192 					continue;
1193 				}
1194 				mtx_unlock(&range_lock);
1195 				break;
1196 			}
1197 			mtx_lock(&res->hr_amp_lock);
1198 			if (activemap_write_start(res->hr_amp,
1199 			    ggio->gctl_offset, ggio->gctl_length)) {
1200 				res->hr_stat_activemap_update++;
1201 				(void)hast_activemap_flush(res);
1202 			}
1203 			mtx_unlock(&res->hr_amp_lock);
1204 			/* FALLTHROUGH */
1205 		case BIO_DELETE:
1206 		case BIO_FLUSH:
1207 			switch (ggio->gctl_cmd) {
1208 			case BIO_DELETE:
1209 				res->hr_stat_delete++;
1210 				break;
1211 			case BIO_FLUSH:
1212 				res->hr_stat_flush++;
1213 				break;
1214 			}
1215 			pjdlog_debug(2,
1216 			    "ggate_recv: (%p) Moving request to the send queue.",
1217 			    hio);
1218 			refcount_init(&hio->hio_countdown, ncomps);
1219 			for (ii = 0; ii < ncomps; ii++)
1220 				QUEUE_INSERT1(hio, send, ii);
1221 			break;
1222 		}
1223 	}
1224 	/* NOTREACHED */
1225 	return (NULL);
1226 }
1227 
1228 /*
1229  * Thread reads from or writes to local component.
1230  * If local read fails, it redirects it to remote_send thread.
1231  */
1232 static void *
1233 local_send_thread(void *arg)
1234 {
1235 	struct hast_resource *res = arg;
1236 	struct g_gate_ctl_io *ggio;
1237 	struct hio *hio;
1238 	unsigned int ncomp, rncomp;
1239 	ssize_t ret;
1240 
1241 	/* Local component is 0 for now. */
1242 	ncomp = 0;
1243 	/* Remote component is 1 for now. */
1244 	rncomp = 1;
1245 
1246 	for (;;) {
1247 		pjdlog_debug(2, "local_send: Taking request.");
1248 		QUEUE_TAKE1(hio, send, ncomp, 0);
1249 		pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1250 		ggio = &hio->hio_ggio;
1251 		switch (ggio->gctl_cmd) {
1252 		case BIO_READ:
1253 			ret = pread(res->hr_localfd, ggio->gctl_data,
1254 			    ggio->gctl_length,
1255 			    ggio->gctl_offset + res->hr_localoff);
1256 			if (ret == ggio->gctl_length)
1257 				hio->hio_errors[ncomp] = 0;
1258 			else if (!ISSYNCREQ(hio)) {
1259 				/*
1260 				 * If READ failed, try to read from remote node.
1261 				 */
1262 				if (ret < 0) {
1263 					reqlog(LOG_WARNING, 0, ggio,
1264 					    "Local request failed (%s), trying remote node. ",
1265 					    strerror(errno));
1266 				} else if (ret != ggio->gctl_length) {
1267 					reqlog(LOG_WARNING, 0, ggio,
1268 					    "Local request failed (%zd != %jd), trying remote node. ",
1269 					    ret, (intmax_t)ggio->gctl_length);
1270 				}
1271 				QUEUE_INSERT1(hio, send, rncomp);
1272 				continue;
1273 			}
1274 			break;
1275 		case BIO_WRITE:
1276 			ret = pwrite(res->hr_localfd, ggio->gctl_data,
1277 			    ggio->gctl_length,
1278 			    ggio->gctl_offset + res->hr_localoff);
1279 			if (ret < 0) {
1280 				hio->hio_errors[ncomp] = errno;
1281 				reqlog(LOG_WARNING, 0, ggio,
1282 				    "Local request failed (%s): ",
1283 				    strerror(errno));
1284 			} else if (ret != ggio->gctl_length) {
1285 				hio->hio_errors[ncomp] = EIO;
1286 				reqlog(LOG_WARNING, 0, ggio,
1287 				    "Local request failed (%zd != %jd): ",
1288 				    ret, (intmax_t)ggio->gctl_length);
1289 			} else {
1290 				hio->hio_errors[ncomp] = 0;
1291 			}
1292 			break;
1293 		case BIO_DELETE:
1294 			ret = g_delete(res->hr_localfd,
1295 			    ggio->gctl_offset + res->hr_localoff,
1296 			    ggio->gctl_length);
1297 			if (ret < 0) {
1298 				hio->hio_errors[ncomp] = errno;
1299 				reqlog(LOG_WARNING, 0, ggio,
1300 				    "Local request failed (%s): ",
1301 				    strerror(errno));
1302 			} else {
1303 				hio->hio_errors[ncomp] = 0;
1304 			}
1305 			break;
1306 		case BIO_FLUSH:
1307 			if (!res->hr_localflush) {
1308 				ret = -1;
1309 				errno = EOPNOTSUPP;
1310 				break;
1311 			}
1312 			ret = g_flush(res->hr_localfd);
1313 			if (ret < 0) {
1314 				if (errno == EOPNOTSUPP)
1315 					res->hr_localflush = false;
1316 				hio->hio_errors[ncomp] = errno;
1317 				reqlog(LOG_WARNING, 0, ggio,
1318 				    "Local request failed (%s): ",
1319 				    strerror(errno));
1320 			} else {
1321 				hio->hio_errors[ncomp] = 0;
1322 			}
1323 			break;
1324 		}
1325 		if (refcount_release(&hio->hio_countdown)) {
1326 			if (ISSYNCREQ(hio)) {
1327 				mtx_lock(&sync_lock);
1328 				SYNCREQDONE(hio);
1329 				mtx_unlock(&sync_lock);
1330 				cv_signal(&sync_cond);
1331 			} else {
1332 				pjdlog_debug(2,
1333 				    "local_send: (%p) Moving request to the done queue.",
1334 				    hio);
1335 				QUEUE_INSERT2(hio, done);
1336 			}
1337 		}
1338 	}
1339 	/* NOTREACHED */
1340 	return (NULL);
1341 }
1342 
1343 static void
1344 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1345 {
1346 	struct nv *nv;
1347 
1348 	rw_rlock(&hio_remote_lock[ncomp]);
1349 
1350 	if (!ISCONNECTED(res, ncomp)) {
1351 		rw_unlock(&hio_remote_lock[ncomp]);
1352 		return;
1353 	}
1354 
1355 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1356 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1357 
1358 	nv = nv_alloc();
1359 	nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1360 	if (nv_error(nv) != 0) {
1361 		rw_unlock(&hio_remote_lock[ncomp]);
1362 		nv_free(nv);
1363 		pjdlog_debug(1,
1364 		    "keepalive_send: Unable to prepare header to send.");
1365 		return;
1366 	}
1367 	if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) {
1368 		rw_unlock(&hio_remote_lock[ncomp]);
1369 		pjdlog_common(LOG_DEBUG, 1, errno,
1370 		    "keepalive_send: Unable to send request");
1371 		nv_free(nv);
1372 		remote_close(res, ncomp);
1373 		return;
1374 	}
1375 
1376 	rw_unlock(&hio_remote_lock[ncomp]);
1377 	nv_free(nv);
1378 	pjdlog_debug(2, "keepalive_send: Request sent.");
1379 }
1380 
1381 /*
1382  * Thread sends request to secondary node.
1383  */
1384 static void *
1385 remote_send_thread(void *arg)
1386 {
1387 	struct hast_resource *res = arg;
1388 	struct g_gate_ctl_io *ggio;
1389 	time_t lastcheck, now;
1390 	struct hio *hio;
1391 	struct nv *nv;
1392 	unsigned int ncomp;
1393 	bool wakeup;
1394 	uint64_t offset, length;
1395 	uint8_t cmd;
1396 	void *data;
1397 
1398 	/* Remote component is 1 for now. */
1399 	ncomp = 1;
1400 	lastcheck = time(NULL);
1401 
1402 	for (;;) {
1403 		pjdlog_debug(2, "remote_send: Taking request.");
1404 		QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE);
1405 		if (hio == NULL) {
1406 			now = time(NULL);
1407 			if (lastcheck + HAST_KEEPALIVE <= now) {
1408 				keepalive_send(res, ncomp);
1409 				lastcheck = now;
1410 			}
1411 			continue;
1412 		}
1413 		pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1414 		ggio = &hio->hio_ggio;
1415 		switch (ggio->gctl_cmd) {
1416 		case BIO_READ:
1417 			cmd = HIO_READ;
1418 			data = NULL;
1419 			offset = ggio->gctl_offset;
1420 			length = ggio->gctl_length;
1421 			break;
1422 		case BIO_WRITE:
1423 			cmd = HIO_WRITE;
1424 			data = ggio->gctl_data;
1425 			offset = ggio->gctl_offset;
1426 			length = ggio->gctl_length;
1427 			break;
1428 		case BIO_DELETE:
1429 			cmd = HIO_DELETE;
1430 			data = NULL;
1431 			offset = ggio->gctl_offset;
1432 			length = ggio->gctl_length;
1433 			break;
1434 		case BIO_FLUSH:
1435 			cmd = HIO_FLUSH;
1436 			data = NULL;
1437 			offset = 0;
1438 			length = 0;
1439 			break;
1440 		default:
1441 			PJDLOG_ABORT("invalid condition");
1442 		}
1443 		nv = nv_alloc();
1444 		nv_add_uint8(nv, cmd, "cmd");
1445 		nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1446 		nv_add_uint64(nv, offset, "offset");
1447 		nv_add_uint64(nv, length, "length");
1448 		if (nv_error(nv) != 0) {
1449 			hio->hio_errors[ncomp] = nv_error(nv);
1450 			pjdlog_debug(2,
1451 			    "remote_send: (%p) Unable to prepare header to send.",
1452 			    hio);
1453 			reqlog(LOG_ERR, 0, ggio,
1454 			    "Unable to prepare header to send (%s): ",
1455 			    strerror(nv_error(nv)));
1456 			/* Move failed request immediately to the done queue. */
1457 			goto done_queue;
1458 		}
1459 		pjdlog_debug(2,
1460 		    "remote_send: (%p) Moving request to the recv queue.",
1461 		    hio);
1462 		/*
1463 		 * Protect connection from disappearing.
1464 		 */
1465 		rw_rlock(&hio_remote_lock[ncomp]);
1466 		if (!ISCONNECTED(res, ncomp)) {
1467 			rw_unlock(&hio_remote_lock[ncomp]);
1468 			hio->hio_errors[ncomp] = ENOTCONN;
1469 			goto done_queue;
1470 		}
1471 		/*
1472 		 * Move the request to recv queue before sending it, because
1473 		 * in different order we can get reply before we move request
1474 		 * to recv queue.
1475 		 */
1476 		mtx_lock(&hio_recv_list_lock[ncomp]);
1477 		wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1478 		TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1479 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1480 		if (hast_proto_send(res, res->hr_remoteout, nv, data,
1481 		    data != NULL ? length : 0) < 0) {
1482 			hio->hio_errors[ncomp] = errno;
1483 			rw_unlock(&hio_remote_lock[ncomp]);
1484 			pjdlog_debug(2,
1485 			    "remote_send: (%p) Unable to send request.", hio);
1486 			reqlog(LOG_ERR, 0, ggio,
1487 			    "Unable to send request (%s): ",
1488 			    strerror(hio->hio_errors[ncomp]));
1489 			remote_close(res, ncomp);
1490 			/*
1491 			 * Take request back from the receive queue and move
1492 			 * it immediately to the done queue.
1493 			 */
1494 			mtx_lock(&hio_recv_list_lock[ncomp]);
1495 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1496 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1497 			goto done_queue;
1498 		}
1499 		rw_unlock(&hio_remote_lock[ncomp]);
1500 		nv_free(nv);
1501 		if (wakeup)
1502 			cv_signal(&hio_recv_list_cond[ncomp]);
1503 		continue;
1504 done_queue:
1505 		nv_free(nv);
1506 		if (ISSYNCREQ(hio)) {
1507 			if (!refcount_release(&hio->hio_countdown))
1508 				continue;
1509 			mtx_lock(&sync_lock);
1510 			SYNCREQDONE(hio);
1511 			mtx_unlock(&sync_lock);
1512 			cv_signal(&sync_cond);
1513 			continue;
1514 		}
1515 		if (ggio->gctl_cmd == BIO_WRITE) {
1516 			mtx_lock(&res->hr_amp_lock);
1517 			if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1518 			    ggio->gctl_length)) {
1519 				(void)hast_activemap_flush(res);
1520 			}
1521 			mtx_unlock(&res->hr_amp_lock);
1522 		}
1523 		if (!refcount_release(&hio->hio_countdown))
1524 			continue;
1525 		pjdlog_debug(2,
1526 		    "remote_send: (%p) Moving request to the done queue.",
1527 		    hio);
1528 		QUEUE_INSERT2(hio, done);
1529 	}
1530 	/* NOTREACHED */
1531 	return (NULL);
1532 }
1533 
1534 /*
1535  * Thread receives answer from secondary node and passes it to ggate_send
1536  * thread.
1537  */
1538 static void *
1539 remote_recv_thread(void *arg)
1540 {
1541 	struct hast_resource *res = arg;
1542 	struct g_gate_ctl_io *ggio;
1543 	struct hio *hio;
1544 	struct nv *nv;
1545 	unsigned int ncomp;
1546 	uint64_t seq;
1547 	int error;
1548 
1549 	/* Remote component is 1 for now. */
1550 	ncomp = 1;
1551 
1552 	for (;;) {
1553 		/* Wait until there is anything to receive. */
1554 		mtx_lock(&hio_recv_list_lock[ncomp]);
1555 		while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1556 			pjdlog_debug(2, "remote_recv: No requests, waiting.");
1557 			cv_wait(&hio_recv_list_cond[ncomp],
1558 			    &hio_recv_list_lock[ncomp]);
1559 		}
1560 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1561 		rw_rlock(&hio_remote_lock[ncomp]);
1562 		if (!ISCONNECTED(res, ncomp)) {
1563 			rw_unlock(&hio_remote_lock[ncomp]);
1564 			/*
1565 			 * Connection is dead, so move all pending requests to
1566 			 * the done queue (one-by-one).
1567 			 */
1568 			mtx_lock(&hio_recv_list_lock[ncomp]);
1569 			hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1570 			PJDLOG_ASSERT(hio != NULL);
1571 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1572 			    hio_next[ncomp]);
1573 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1574 			goto done_queue;
1575 		}
1576 		if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) {
1577 			pjdlog_errno(LOG_ERR,
1578 			    "Unable to receive reply header");
1579 			rw_unlock(&hio_remote_lock[ncomp]);
1580 			remote_close(res, ncomp);
1581 			continue;
1582 		}
1583 		rw_unlock(&hio_remote_lock[ncomp]);
1584 		seq = nv_get_uint64(nv, "seq");
1585 		if (seq == 0) {
1586 			pjdlog_error("Header contains no 'seq' field.");
1587 			nv_free(nv);
1588 			continue;
1589 		}
1590 		mtx_lock(&hio_recv_list_lock[ncomp]);
1591 		TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1592 			if (hio->hio_ggio.gctl_seq == seq) {
1593 				TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1594 				    hio_next[ncomp]);
1595 				break;
1596 			}
1597 		}
1598 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1599 		if (hio == NULL) {
1600 			pjdlog_error("Found no request matching received 'seq' field (%ju).",
1601 			    (uintmax_t)seq);
1602 			nv_free(nv);
1603 			continue;
1604 		}
1605 		error = nv_get_int16(nv, "error");
1606 		if (error != 0) {
1607 			/* Request failed on remote side. */
1608 			hio->hio_errors[ncomp] = error;
1609 			reqlog(LOG_WARNING, 0, &hio->hio_ggio,
1610 			    "Remote request failed (%s): ", strerror(error));
1611 			nv_free(nv);
1612 			goto done_queue;
1613 		}
1614 		ggio = &hio->hio_ggio;
1615 		switch (ggio->gctl_cmd) {
1616 		case BIO_READ:
1617 			rw_rlock(&hio_remote_lock[ncomp]);
1618 			if (!ISCONNECTED(res, ncomp)) {
1619 				rw_unlock(&hio_remote_lock[ncomp]);
1620 				nv_free(nv);
1621 				goto done_queue;
1622 			}
1623 			if (hast_proto_recv_data(res, res->hr_remotein, nv,
1624 			    ggio->gctl_data, ggio->gctl_length) < 0) {
1625 				hio->hio_errors[ncomp] = errno;
1626 				pjdlog_errno(LOG_ERR,
1627 				    "Unable to receive reply data");
1628 				rw_unlock(&hio_remote_lock[ncomp]);
1629 				nv_free(nv);
1630 				remote_close(res, ncomp);
1631 				goto done_queue;
1632 			}
1633 			rw_unlock(&hio_remote_lock[ncomp]);
1634 			break;
1635 		case BIO_WRITE:
1636 		case BIO_DELETE:
1637 		case BIO_FLUSH:
1638 			break;
1639 		default:
1640 			PJDLOG_ABORT("invalid condition");
1641 		}
1642 		hio->hio_errors[ncomp] = 0;
1643 		nv_free(nv);
1644 done_queue:
1645 		if (refcount_release(&hio->hio_countdown)) {
1646 			if (ISSYNCREQ(hio)) {
1647 				mtx_lock(&sync_lock);
1648 				SYNCREQDONE(hio);
1649 				mtx_unlock(&sync_lock);
1650 				cv_signal(&sync_cond);
1651 			} else {
1652 				pjdlog_debug(2,
1653 				    "remote_recv: (%p) Moving request to the done queue.",
1654 				    hio);
1655 				QUEUE_INSERT2(hio, done);
1656 			}
1657 		}
1658 	}
1659 	/* NOTREACHED */
1660 	return (NULL);
1661 }
1662 
1663 /*
1664  * Thread sends answer to the kernel.
1665  */
1666 static void *
1667 ggate_send_thread(void *arg)
1668 {
1669 	struct hast_resource *res = arg;
1670 	struct g_gate_ctl_io *ggio;
1671 	struct hio *hio;
1672 	unsigned int ii, ncomp, ncomps;
1673 
1674 	ncomps = HAST_NCOMPONENTS;
1675 
1676 	for (;;) {
1677 		pjdlog_debug(2, "ggate_send: Taking request.");
1678 		QUEUE_TAKE2(hio, done);
1679 		pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1680 		ggio = &hio->hio_ggio;
1681 		for (ii = 0; ii < ncomps; ii++) {
1682 			if (hio->hio_errors[ii] == 0) {
1683 				/*
1684 				 * One successful request is enough to declare
1685 				 * success.
1686 				 */
1687 				ggio->gctl_error = 0;
1688 				break;
1689 			}
1690 		}
1691 		if (ii == ncomps) {
1692 			/*
1693 			 * None of the requests were successful.
1694 			 * Use the error from local component except the
1695 			 * case when we did only remote request.
1696 			 */
1697 			if (ggio->gctl_cmd == BIO_READ &&
1698 			    res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
1699 				ggio->gctl_error = hio->hio_errors[1];
1700 			else
1701 				ggio->gctl_error = hio->hio_errors[0];
1702 		}
1703 		if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1704 			mtx_lock(&res->hr_amp_lock);
1705 			if (activemap_write_complete(res->hr_amp,
1706 			    ggio->gctl_offset, ggio->gctl_length)) {
1707 				res->hr_stat_activemap_update++;
1708 				(void)hast_activemap_flush(res);
1709 			}
1710 			mtx_unlock(&res->hr_amp_lock);
1711 		}
1712 		if (ggio->gctl_cmd == BIO_WRITE) {
1713 			/*
1714 			 * Unlock range we locked.
1715 			 */
1716 			mtx_lock(&range_lock);
1717 			rangelock_del(range_regular, ggio->gctl_offset,
1718 			    ggio->gctl_length);
1719 			if (range_sync_wait)
1720 				cv_signal(&range_sync_cond);
1721 			mtx_unlock(&range_lock);
1722 			/*
1723 			 * Bump local count if this is first write after
1724 			 * connection failure with remote node.
1725 			 */
1726 			ncomp = 1;
1727 			rw_rlock(&hio_remote_lock[ncomp]);
1728 			if (!ISCONNECTED(res, ncomp)) {
1729 				mtx_lock(&metadata_lock);
1730 				if (res->hr_primary_localcnt ==
1731 				    res->hr_secondary_remotecnt) {
1732 					res->hr_primary_localcnt++;
1733 					pjdlog_debug(1,
1734 					    "Increasing localcnt to %ju.",
1735 					    (uintmax_t)res->hr_primary_localcnt);
1736 					(void)metadata_write(res);
1737 				}
1738 				mtx_unlock(&metadata_lock);
1739 			}
1740 			rw_unlock(&hio_remote_lock[ncomp]);
1741 		}
1742 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0)
1743 			primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1744 		pjdlog_debug(2,
1745 		    "ggate_send: (%p) Moving request to the free queue.", hio);
1746 		QUEUE_INSERT2(hio, free);
1747 	}
1748 	/* NOTREACHED */
1749 	return (NULL);
1750 }
1751 
1752 /*
1753  * Thread synchronize local and remote components.
1754  */
1755 static void *
1756 sync_thread(void *arg __unused)
1757 {
1758 	struct hast_resource *res = arg;
1759 	struct hio *hio;
1760 	struct g_gate_ctl_io *ggio;
1761 	struct timeval tstart, tend, tdiff;
1762 	unsigned int ii, ncomp, ncomps;
1763 	off_t offset, length, synced;
1764 	bool dorewind;
1765 	int syncext;
1766 
1767 	ncomps = HAST_NCOMPONENTS;
1768 	dorewind = true;
1769 	synced = 0;
1770 	offset = -1;
1771 
1772 	for (;;) {
1773 		mtx_lock(&sync_lock);
1774 		if (offset >= 0 && !sync_inprogress) {
1775 			gettimeofday(&tend, NULL);
1776 			timersub(&tend, &tstart, &tdiff);
1777 			pjdlog_info("Synchronization interrupted after %#.0T. "
1778 			    "%NB synchronized so far.", &tdiff,
1779 			    (intmax_t)synced);
1780 			event_send(res, EVENT_SYNCINTR);
1781 		}
1782 		while (!sync_inprogress) {
1783 			dorewind = true;
1784 			synced = 0;
1785 			cv_wait(&sync_cond, &sync_lock);
1786 		}
1787 		mtx_unlock(&sync_lock);
1788 		/*
1789 		 * Obtain offset at which we should synchronize.
1790 		 * Rewind synchronization if needed.
1791 		 */
1792 		mtx_lock(&res->hr_amp_lock);
1793 		if (dorewind)
1794 			activemap_sync_rewind(res->hr_amp);
1795 		offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1796 		if (syncext != -1) {
1797 			/*
1798 			 * We synchronized entire syncext extent, we can mark
1799 			 * it as clean now.
1800 			 */
1801 			if (activemap_extent_complete(res->hr_amp, syncext))
1802 				(void)hast_activemap_flush(res);
1803 		}
1804 		mtx_unlock(&res->hr_amp_lock);
1805 		if (dorewind) {
1806 			dorewind = false;
1807 			if (offset < 0)
1808 				pjdlog_info("Nodes are in sync.");
1809 			else {
1810 				pjdlog_info("Synchronization started. %NB to go.",
1811 				    (intmax_t)(res->hr_extentsize *
1812 				    activemap_ndirty(res->hr_amp)));
1813 				event_send(res, EVENT_SYNCSTART);
1814 				gettimeofday(&tstart, NULL);
1815 			}
1816 		}
1817 		if (offset < 0) {
1818 			sync_stop();
1819 			pjdlog_debug(1, "Nothing to synchronize.");
1820 			/*
1821 			 * Synchronization complete, make both localcnt and
1822 			 * remotecnt equal.
1823 			 */
1824 			ncomp = 1;
1825 			rw_rlock(&hio_remote_lock[ncomp]);
1826 			if (ISCONNECTED(res, ncomp)) {
1827 				if (synced > 0) {
1828 					int64_t bps;
1829 
1830 					gettimeofday(&tend, NULL);
1831 					timersub(&tend, &tstart, &tdiff);
1832 					bps = (int64_t)((double)synced /
1833 					    ((double)tdiff.tv_sec +
1834 					    (double)tdiff.tv_usec / 1000000));
1835 					pjdlog_info("Synchronization complete. "
1836 					    "%NB synchronized in %#.0lT (%NB/sec).",
1837 					    (intmax_t)synced, &tdiff,
1838 					    (intmax_t)bps);
1839 					event_send(res, EVENT_SYNCDONE);
1840 				}
1841 				mtx_lock(&metadata_lock);
1842 				res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
1843 				res->hr_primary_localcnt =
1844 				    res->hr_secondary_remotecnt;
1845 				res->hr_primary_remotecnt =
1846 				    res->hr_secondary_localcnt;
1847 				pjdlog_debug(1,
1848 				    "Setting localcnt to %ju and remotecnt to %ju.",
1849 				    (uintmax_t)res->hr_primary_localcnt,
1850 				    (uintmax_t)res->hr_primary_remotecnt);
1851 				(void)metadata_write(res);
1852 				mtx_unlock(&metadata_lock);
1853 			}
1854 			rw_unlock(&hio_remote_lock[ncomp]);
1855 			continue;
1856 		}
1857 		pjdlog_debug(2, "sync: Taking free request.");
1858 		QUEUE_TAKE2(hio, free);
1859 		pjdlog_debug(2, "sync: (%p) Got free request.", hio);
1860 		/*
1861 		 * Lock the range we are going to synchronize. We don't want
1862 		 * race where someone writes between our read and write.
1863 		 */
1864 		for (;;) {
1865 			mtx_lock(&range_lock);
1866 			if (rangelock_islocked(range_regular, offset, length)) {
1867 				pjdlog_debug(2,
1868 				    "sync: Range offset=%jd length=%jd locked.",
1869 				    (intmax_t)offset, (intmax_t)length);
1870 				range_sync_wait = true;
1871 				cv_wait(&range_sync_cond, &range_lock);
1872 				range_sync_wait = false;
1873 				mtx_unlock(&range_lock);
1874 				continue;
1875 			}
1876 			if (rangelock_add(range_sync, offset, length) < 0) {
1877 				mtx_unlock(&range_lock);
1878 				pjdlog_debug(2,
1879 				    "sync: Range offset=%jd length=%jd is already locked, waiting.",
1880 				    (intmax_t)offset, (intmax_t)length);
1881 				sleep(1);
1882 				continue;
1883 			}
1884 			mtx_unlock(&range_lock);
1885 			break;
1886 		}
1887 		/*
1888 		 * First read the data from synchronization source.
1889 		 */
1890 		SYNCREQ(hio);
1891 		ggio = &hio->hio_ggio;
1892 		ggio->gctl_cmd = BIO_READ;
1893 		ggio->gctl_offset = offset;
1894 		ggio->gctl_length = length;
1895 		ggio->gctl_error = 0;
1896 		for (ii = 0; ii < ncomps; ii++)
1897 			hio->hio_errors[ii] = EINVAL;
1898 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1899 		    hio);
1900 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1901 		    hio);
1902 		mtx_lock(&metadata_lock);
1903 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1904 			/*
1905 			 * This range is up-to-date on local component,
1906 			 * so handle request locally.
1907 			 */
1908 			 /* Local component is 0 for now. */
1909 			ncomp = 0;
1910 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1911 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1912 			/*
1913 			 * This range is out-of-date on local component,
1914 			 * so send request to the remote node.
1915 			 */
1916 			 /* Remote component is 1 for now. */
1917 			ncomp = 1;
1918 		}
1919 		mtx_unlock(&metadata_lock);
1920 		refcount_init(&hio->hio_countdown, 1);
1921 		QUEUE_INSERT1(hio, send, ncomp);
1922 
1923 		/*
1924 		 * Let's wait for READ to finish.
1925 		 */
1926 		mtx_lock(&sync_lock);
1927 		while (!ISSYNCREQDONE(hio))
1928 			cv_wait(&sync_cond, &sync_lock);
1929 		mtx_unlock(&sync_lock);
1930 
1931 		if (hio->hio_errors[ncomp] != 0) {
1932 			pjdlog_error("Unable to read synchronization data: %s.",
1933 			    strerror(hio->hio_errors[ncomp]));
1934 			goto free_queue;
1935 		}
1936 
1937 		/*
1938 		 * We read the data from synchronization source, now write it
1939 		 * to synchronization target.
1940 		 */
1941 		SYNCREQ(hio);
1942 		ggio->gctl_cmd = BIO_WRITE;
1943 		for (ii = 0; ii < ncomps; ii++)
1944 			hio->hio_errors[ii] = EINVAL;
1945 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1946 		    hio);
1947 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1948 		    hio);
1949 		mtx_lock(&metadata_lock);
1950 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1951 			/*
1952 			 * This range is up-to-date on local component,
1953 			 * so we update remote component.
1954 			 */
1955 			 /* Remote component is 1 for now. */
1956 			ncomp = 1;
1957 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1958 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1959 			/*
1960 			 * This range is out-of-date on local component,
1961 			 * so we update it.
1962 			 */
1963 			 /* Local component is 0 for now. */
1964 			ncomp = 0;
1965 		}
1966 		mtx_unlock(&metadata_lock);
1967 
1968 		pjdlog_debug(2, "sync: (%p) Moving request to the send queues.",
1969 		    hio);
1970 		refcount_init(&hio->hio_countdown, 1);
1971 		QUEUE_INSERT1(hio, send, ncomp);
1972 
1973 		/*
1974 		 * Let's wait for WRITE to finish.
1975 		 */
1976 		mtx_lock(&sync_lock);
1977 		while (!ISSYNCREQDONE(hio))
1978 			cv_wait(&sync_cond, &sync_lock);
1979 		mtx_unlock(&sync_lock);
1980 
1981 		if (hio->hio_errors[ncomp] != 0) {
1982 			pjdlog_error("Unable to write synchronization data: %s.",
1983 			    strerror(hio->hio_errors[ncomp]));
1984 			goto free_queue;
1985 		}
1986 
1987 		synced += length;
1988 free_queue:
1989 		mtx_lock(&range_lock);
1990 		rangelock_del(range_sync, offset, length);
1991 		if (range_regular_wait)
1992 			cv_signal(&range_regular_cond);
1993 		mtx_unlock(&range_lock);
1994 		pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
1995 		    hio);
1996 		QUEUE_INSERT2(hio, free);
1997 	}
1998 	/* NOTREACHED */
1999 	return (NULL);
2000 }
2001 
2002 void
2003 primary_config_reload(struct hast_resource *res, struct nv *nv)
2004 {
2005 	unsigned int ii, ncomps;
2006 	int modified, vint;
2007 	const char *vstr;
2008 
2009 	pjdlog_info("Reloading configuration...");
2010 
2011 	PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
2012 	PJDLOG_ASSERT(gres == res);
2013 	nv_assert(nv, "remoteaddr");
2014 	nv_assert(nv, "sourceaddr");
2015 	nv_assert(nv, "replication");
2016 	nv_assert(nv, "checksum");
2017 	nv_assert(nv, "compression");
2018 	nv_assert(nv, "timeout");
2019 	nv_assert(nv, "exec");
2020 	nv_assert(nv, "metaflush");
2021 
2022 	ncomps = HAST_NCOMPONENTS;
2023 
2024 #define MODIFIED_REMOTEADDR	0x01
2025 #define MODIFIED_SOURCEADDR	0x02
2026 #define MODIFIED_REPLICATION	0x04
2027 #define MODIFIED_CHECKSUM	0x08
2028 #define MODIFIED_COMPRESSION	0x10
2029 #define MODIFIED_TIMEOUT	0x20
2030 #define MODIFIED_EXEC		0x40
2031 #define MODIFIED_METAFLUSH	0x80
2032 	modified = 0;
2033 
2034 	vstr = nv_get_string(nv, "remoteaddr");
2035 	if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
2036 		/*
2037 		 * Don't copy res->hr_remoteaddr to gres just yet.
2038 		 * We want remote_close() to log disconnect from the old
2039 		 * addresses, not from the new ones.
2040 		 */
2041 		modified |= MODIFIED_REMOTEADDR;
2042 	}
2043 	vstr = nv_get_string(nv, "sourceaddr");
2044 	if (strcmp(gres->hr_sourceaddr, vstr) != 0) {
2045 		strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr));
2046 		modified |= MODIFIED_SOURCEADDR;
2047 	}
2048 	vint = nv_get_int32(nv, "replication");
2049 	if (gres->hr_replication != vint) {
2050 		gres->hr_replication = vint;
2051 		modified |= MODIFIED_REPLICATION;
2052 	}
2053 	vint = nv_get_int32(nv, "checksum");
2054 	if (gres->hr_checksum != vint) {
2055 		gres->hr_checksum = vint;
2056 		modified |= MODIFIED_CHECKSUM;
2057 	}
2058 	vint = nv_get_int32(nv, "compression");
2059 	if (gres->hr_compression != vint) {
2060 		gres->hr_compression = vint;
2061 		modified |= MODIFIED_COMPRESSION;
2062 	}
2063 	vint = nv_get_int32(nv, "timeout");
2064 	if (gres->hr_timeout != vint) {
2065 		gres->hr_timeout = vint;
2066 		modified |= MODIFIED_TIMEOUT;
2067 	}
2068 	vstr = nv_get_string(nv, "exec");
2069 	if (strcmp(gres->hr_exec, vstr) != 0) {
2070 		strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
2071 		modified |= MODIFIED_EXEC;
2072 	}
2073 	vint = nv_get_int32(nv, "metaflush");
2074 	if (gres->hr_metaflush != vint) {
2075 		gres->hr_metaflush = vint;
2076 		modified |= MODIFIED_METAFLUSH;
2077 	}
2078 
2079 	/*
2080 	 * Change timeout for connected sockets.
2081 	 * Don't bother if we need to reconnect.
2082 	 */
2083 	if ((modified & MODIFIED_TIMEOUT) != 0 &&
2084 	    (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR |
2085 	    MODIFIED_REPLICATION)) == 0) {
2086 		for (ii = 0; ii < ncomps; ii++) {
2087 			if (!ISREMOTE(ii))
2088 				continue;
2089 			rw_rlock(&hio_remote_lock[ii]);
2090 			if (!ISCONNECTED(gres, ii)) {
2091 				rw_unlock(&hio_remote_lock[ii]);
2092 				continue;
2093 			}
2094 			rw_unlock(&hio_remote_lock[ii]);
2095 			if (proto_timeout(gres->hr_remotein,
2096 			    gres->hr_timeout) < 0) {
2097 				pjdlog_errno(LOG_WARNING,
2098 				    "Unable to set connection timeout");
2099 			}
2100 			if (proto_timeout(gres->hr_remoteout,
2101 			    gres->hr_timeout) < 0) {
2102 				pjdlog_errno(LOG_WARNING,
2103 				    "Unable to set connection timeout");
2104 			}
2105 		}
2106 	}
2107 	if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR |
2108 	    MODIFIED_REPLICATION)) != 0) {
2109 		for (ii = 0; ii < ncomps; ii++) {
2110 			if (!ISREMOTE(ii))
2111 				continue;
2112 			remote_close(gres, ii);
2113 		}
2114 		if (modified & MODIFIED_REMOTEADDR) {
2115 			vstr = nv_get_string(nv, "remoteaddr");
2116 			strlcpy(gres->hr_remoteaddr, vstr,
2117 			    sizeof(gres->hr_remoteaddr));
2118 		}
2119 	}
2120 #undef	MODIFIED_REMOTEADDR
2121 #undef	MODIFIED_SOURCEADDR
2122 #undef	MODIFIED_REPLICATION
2123 #undef	MODIFIED_CHECKSUM
2124 #undef	MODIFIED_COMPRESSION
2125 #undef	MODIFIED_TIMEOUT
2126 #undef	MODIFIED_EXEC
2127 #undef	MODIFIED_METAFLUSH
2128 
2129 	pjdlog_info("Configuration reloaded successfully.");
2130 }
2131 
2132 static void
2133 guard_one(struct hast_resource *res, unsigned int ncomp)
2134 {
2135 	struct proto_conn *in, *out;
2136 
2137 	if (!ISREMOTE(ncomp))
2138 		return;
2139 
2140 	rw_rlock(&hio_remote_lock[ncomp]);
2141 
2142 	if (!real_remote(res)) {
2143 		rw_unlock(&hio_remote_lock[ncomp]);
2144 		return;
2145 	}
2146 
2147 	if (ISCONNECTED(res, ncomp)) {
2148 		PJDLOG_ASSERT(res->hr_remotein != NULL);
2149 		PJDLOG_ASSERT(res->hr_remoteout != NULL);
2150 		rw_unlock(&hio_remote_lock[ncomp]);
2151 		pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2152 		    res->hr_remoteaddr);
2153 		return;
2154 	}
2155 
2156 	PJDLOG_ASSERT(res->hr_remotein == NULL);
2157 	PJDLOG_ASSERT(res->hr_remoteout == NULL);
2158 	/*
2159 	 * Upgrade the lock. It doesn't have to be atomic as no other thread
2160 	 * can change connection status from disconnected to connected.
2161 	 */
2162 	rw_unlock(&hio_remote_lock[ncomp]);
2163 	pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2164 	    res->hr_remoteaddr);
2165 	in = out = NULL;
2166 	if (init_remote(res, &in, &out) == 0) {
2167 		rw_wlock(&hio_remote_lock[ncomp]);
2168 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2169 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2170 		PJDLOG_ASSERT(in != NULL && out != NULL);
2171 		res->hr_remotein = in;
2172 		res->hr_remoteout = out;
2173 		rw_unlock(&hio_remote_lock[ncomp]);
2174 		pjdlog_info("Successfully reconnected to %s.",
2175 		    res->hr_remoteaddr);
2176 		sync_start();
2177 	} else {
2178 		/* Both connections should be NULL. */
2179 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2180 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2181 		PJDLOG_ASSERT(in == NULL && out == NULL);
2182 		pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2183 		    res->hr_remoteaddr);
2184 	}
2185 }
2186 
2187 /*
2188  * Thread guards remote connections and reconnects when needed, handles
2189  * signals, etc.
2190  */
2191 static void *
2192 guard_thread(void *arg)
2193 {
2194 	struct hast_resource *res = arg;
2195 	unsigned int ii, ncomps;
2196 	struct timespec timeout;
2197 	time_t lastcheck, now;
2198 	sigset_t mask;
2199 	int signo;
2200 
2201 	ncomps = HAST_NCOMPONENTS;
2202 	lastcheck = time(NULL);
2203 
2204 	PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2205 	PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2206 	PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2207 
2208 	timeout.tv_sec = HAST_KEEPALIVE;
2209 	timeout.tv_nsec = 0;
2210 	signo = -1;
2211 
2212 	for (;;) {
2213 		switch (signo) {
2214 		case SIGINT:
2215 		case SIGTERM:
2216 			sigexit_received = true;
2217 			primary_exitx(EX_OK,
2218 			    "Termination signal received, exiting.");
2219 			break;
2220 		default:
2221 			break;
2222 		}
2223 
2224 		/*
2225 		 * Don't check connections until we fully started,
2226 		 * as we may still be looping, waiting for remote node
2227 		 * to switch from primary to secondary.
2228 		 */
2229 		if (fullystarted) {
2230 			pjdlog_debug(2, "remote_guard: Checking connections.");
2231 			now = time(NULL);
2232 			if (lastcheck + HAST_KEEPALIVE <= now) {
2233 				for (ii = 0; ii < ncomps; ii++)
2234 					guard_one(res, ii);
2235 				lastcheck = now;
2236 			}
2237 		}
2238 		signo = sigtimedwait(&mask, NULL, &timeout);
2239 	}
2240 	/* NOTREACHED */
2241 	return (NULL);
2242 }
2243