xref: /freebsd/sbin/hastd/primary.c (revision dda5b39711dab90ae1c5624bdd6ff7453177df31)
1 /*-
2  * Copyright (c) 2009 The FreeBSD Foundation
3  * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net>
4  * All rights reserved.
5  *
6  * This software was developed by Pawel Jakub Dawidek under sponsorship from
7  * the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/types.h>
35 #include <sys/time.h>
36 #include <sys/bio.h>
37 #include <sys/disk.h>
38 #include <sys/stat.h>
39 
40 #include <geom/gate/g_gate.h>
41 
42 #include <err.h>
43 #include <errno.h>
44 #include <fcntl.h>
45 #include <libgeom.h>
46 #include <pthread.h>
47 #include <signal.h>
48 #include <stdint.h>
49 #include <stdio.h>
50 #include <string.h>
51 #include <sysexits.h>
52 #include <unistd.h>
53 
54 #include <activemap.h>
55 #include <nv.h>
56 #include <rangelock.h>
57 
58 #include "control.h"
59 #include "event.h"
60 #include "hast.h"
61 #include "hast_proto.h"
62 #include "hastd.h"
63 #include "hooks.h"
64 #include "metadata.h"
65 #include "proto.h"
66 #include "pjdlog.h"
67 #include "refcnt.h"
68 #include "subr.h"
69 #include "synch.h"
70 
71 /* The is only one remote component for now. */
72 #define	ISREMOTE(no)	((no) == 1)
73 
74 struct hio {
75 	/*
76 	 * Number of components we are still waiting for.
77 	 * When this field goes to 0, we can send the request back to the
78 	 * kernel. Each component has to decrease this counter by one
79 	 * even on failure.
80 	 */
81 	refcnt_t		 hio_countdown;
82 	/*
83 	 * Each component has a place to store its own error.
84 	 * Once the request is handled by all components we can decide if the
85 	 * request overall is successful or not.
86 	 */
87 	int			*hio_errors;
88 	/*
89 	 * Structure used to communicate with GEOM Gate class.
90 	 */
91 	struct g_gate_ctl_io	 hio_ggio;
92 	/*
93 	 * Request was already confirmed to GEOM Gate.
94 	 */
95 	bool			 hio_done;
96 	/*
97 	 * Number of components we are still waiting before sending write
98 	 * completion ack to GEOM Gate. Used for memsync.
99 	 */
100 	refcnt_t		 hio_writecount;
101 	/*
102 	 * Memsync request was acknowleged by remote.
103 	 */
104 	bool			 hio_memsyncacked;
105 	/*
106 	 * Remember replication from the time the request was initiated,
107 	 * so we won't get confused when replication changes on reload.
108 	 */
109 	int			 hio_replication;
110 	TAILQ_ENTRY(hio)	*hio_next;
111 };
112 #define	hio_free_next	hio_next[0]
113 #define	hio_done_next	hio_next[0]
114 
115 /*
116  * Free list holds unused structures. When free list is empty, we have to wait
117  * until some in-progress requests are freed.
118  */
119 static TAILQ_HEAD(, hio) hio_free_list;
120 static size_t hio_free_list_size;
121 static pthread_mutex_t hio_free_list_lock;
122 static pthread_cond_t hio_free_list_cond;
123 /*
124  * There is one send list for every component. One requests is placed on all
125  * send lists - each component gets the same request, but each component is
126  * responsible for managing his own send list.
127  */
128 static TAILQ_HEAD(, hio) *hio_send_list;
129 static size_t *hio_send_list_size;
130 static pthread_mutex_t *hio_send_list_lock;
131 static pthread_cond_t *hio_send_list_cond;
132 #define	hio_send_local_list_size	hio_send_list_size[0]
133 #define	hio_send_remote_list_size	hio_send_list_size[1]
134 /*
135  * There is one recv list for every component, although local components don't
136  * use recv lists as local requests are done synchronously.
137  */
138 static TAILQ_HEAD(, hio) *hio_recv_list;
139 static size_t *hio_recv_list_size;
140 static pthread_mutex_t *hio_recv_list_lock;
141 static pthread_cond_t *hio_recv_list_cond;
142 #define	hio_recv_remote_list_size	hio_recv_list_size[1]
143 /*
144  * Request is placed on done list by the slowest component (the one that
145  * decreased hio_countdown from 1 to 0).
146  */
147 static TAILQ_HEAD(, hio) hio_done_list;
148 static size_t hio_done_list_size;
149 static pthread_mutex_t hio_done_list_lock;
150 static pthread_cond_t hio_done_list_cond;
151 /*
152  * Structure below are for interaction with sync thread.
153  */
154 static bool sync_inprogress;
155 static pthread_mutex_t sync_lock;
156 static pthread_cond_t sync_cond;
157 /*
158  * The lock below allows to synchornize access to remote connections.
159  */
160 static pthread_rwlock_t *hio_remote_lock;
161 
162 /*
163  * Lock to synchronize metadata updates. Also synchronize access to
164  * hr_primary_localcnt and hr_primary_remotecnt fields.
165  */
166 static pthread_mutex_t metadata_lock;
167 
168 /*
169  * Maximum number of outstanding I/O requests.
170  */
171 #define	HAST_HIO_MAX	256
172 /*
173  * Number of components. At this point there are only two components: local
174  * and remote, but in the future it might be possible to use multiple local
175  * and remote components.
176  */
177 #define	HAST_NCOMPONENTS	2
178 
179 #define	ISCONNECTED(res, no)	\
180 	((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
181 
182 #define	QUEUE_INSERT1(hio, name, ncomp)	do {				\
183 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
184 	if (TAILQ_EMPTY(&hio_##name##_list[(ncomp)]))			\
185 		cv_broadcast(&hio_##name##_list_cond[(ncomp)]);		\
186 	TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),		\
187 	    hio_next[(ncomp)]);						\
188 	hio_##name##_list_size[(ncomp)]++;				\
189 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
190 } while (0)
191 #define	QUEUE_INSERT2(hio, name)	do {				\
192 	mtx_lock(&hio_##name##_list_lock);				\
193 	if (TAILQ_EMPTY(&hio_##name##_list))				\
194 		cv_broadcast(&hio_##name##_list_cond);			\
195 	TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
196 	hio_##name##_list_size++;					\
197 	mtx_unlock(&hio_##name##_list_lock);				\
198 } while (0)
199 #define	QUEUE_TAKE1(hio, name, ncomp, timeout)	do {			\
200 	bool _last;							\
201 									\
202 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
203 	_last = false;							\
204 	while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
205 		cv_timedwait(&hio_##name##_list_cond[(ncomp)],		\
206 		    &hio_##name##_list_lock[(ncomp)], (timeout));	\
207 		if ((timeout) != 0)					\
208 			_last = true;					\
209 	}								\
210 	if (hio != NULL) {						\
211 		PJDLOG_ASSERT(hio_##name##_list_size[(ncomp)] != 0);	\
212 		hio_##name##_list_size[(ncomp)]--;			\
213 		TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),	\
214 		    hio_next[(ncomp)]);					\
215 	}								\
216 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
217 } while (0)
218 #define	QUEUE_TAKE2(hio, name)	do {					\
219 	mtx_lock(&hio_##name##_list_lock);				\
220 	while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {	\
221 		cv_wait(&hio_##name##_list_cond,			\
222 		    &hio_##name##_list_lock);				\
223 	}								\
224 	PJDLOG_ASSERT(hio_##name##_list_size != 0);			\
225 	hio_##name##_list_size--;					\
226 	TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);	\
227 	mtx_unlock(&hio_##name##_list_lock);				\
228 } while (0)
229 
230 #define ISFULLSYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_FULLSYNC)
231 #define ISMEMSYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_MEMSYNC)
232 #define ISASYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_ASYNC)
233 
234 #define	SYNCREQ(hio)		do {					\
235 	(hio)->hio_ggio.gctl_unit = -1;					\
236 	(hio)->hio_ggio.gctl_seq = 1;					\
237 } while (0)
238 #define	ISSYNCREQ(hio)		((hio)->hio_ggio.gctl_unit == -1)
239 #define	SYNCREQDONE(hio)	do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
240 #define	ISSYNCREQDONE(hio)	((hio)->hio_ggio.gctl_unit == -2)
241 
242 #define ISMEMSYNCWRITE(hio)	(ISMEMSYNC(hio) &&			\
243 	    (hio)->hio_ggio.gctl_cmd == BIO_WRITE && !ISSYNCREQ(hio))
244 
245 static struct hast_resource *gres;
246 
247 static pthread_mutex_t range_lock;
248 static struct rangelocks *range_regular;
249 static bool range_regular_wait;
250 static pthread_cond_t range_regular_cond;
251 static struct rangelocks *range_sync;
252 static bool range_sync_wait;
253 static pthread_cond_t range_sync_cond;
254 static bool fullystarted;
255 
256 static void *ggate_recv_thread(void *arg);
257 static void *local_send_thread(void *arg);
258 static void *remote_send_thread(void *arg);
259 static void *remote_recv_thread(void *arg);
260 static void *ggate_send_thread(void *arg);
261 static void *sync_thread(void *arg);
262 static void *guard_thread(void *arg);
263 
264 static void
265 output_status_aux(struct nv *nvout)
266 {
267 
268 	nv_add_uint64(nvout, (uint64_t)hio_free_list_size,
269 	    "idle_queue_size");
270 	nv_add_uint64(nvout, (uint64_t)hio_send_local_list_size,
271 	    "local_queue_size");
272 	nv_add_uint64(nvout, (uint64_t)hio_send_remote_list_size,
273 	    "send_queue_size");
274 	nv_add_uint64(nvout, (uint64_t)hio_recv_remote_list_size,
275 	    "recv_queue_size");
276 	nv_add_uint64(nvout, (uint64_t)hio_done_list_size,
277 	    "done_queue_size");
278 }
279 
280 static void
281 cleanup(struct hast_resource *res)
282 {
283 	int rerrno;
284 
285 	/* Remember errno. */
286 	rerrno = errno;
287 
288 	/* Destroy ggate provider if we created one. */
289 	if (res->hr_ggateunit >= 0) {
290 		struct g_gate_ctl_destroy ggiod;
291 
292 		bzero(&ggiod, sizeof(ggiod));
293 		ggiod.gctl_version = G_GATE_VERSION;
294 		ggiod.gctl_unit = res->hr_ggateunit;
295 		ggiod.gctl_force = 1;
296 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) == -1) {
297 			pjdlog_errno(LOG_WARNING,
298 			    "Unable to destroy hast/%s device",
299 			    res->hr_provname);
300 		}
301 		res->hr_ggateunit = -1;
302 	}
303 
304 	/* Restore errno. */
305 	errno = rerrno;
306 }
307 
308 static __dead2 void
309 primary_exit(int exitcode, const char *fmt, ...)
310 {
311 	va_list ap;
312 
313 	PJDLOG_ASSERT(exitcode != EX_OK);
314 	va_start(ap, fmt);
315 	pjdlogv_errno(LOG_ERR, fmt, ap);
316 	va_end(ap);
317 	cleanup(gres);
318 	exit(exitcode);
319 }
320 
321 static __dead2 void
322 primary_exitx(int exitcode, const char *fmt, ...)
323 {
324 	va_list ap;
325 
326 	va_start(ap, fmt);
327 	pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
328 	va_end(ap);
329 	cleanup(gres);
330 	exit(exitcode);
331 }
332 
333 /* Expects res->hr_amp locked, returns unlocked. */
334 static int
335 hast_activemap_flush(struct hast_resource *res)
336 {
337 	const unsigned char *buf;
338 	size_t size;
339 	int ret;
340 
341 	mtx_lock(&res->hr_amp_diskmap_lock);
342 	buf = activemap_bitmap(res->hr_amp, &size);
343 	mtx_unlock(&res->hr_amp_lock);
344 	PJDLOG_ASSERT(buf != NULL);
345 	PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
346 	ret = 0;
347 	if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
348 	    (ssize_t)size) {
349 		pjdlog_errno(LOG_ERR, "Unable to flush activemap to disk");
350 		res->hr_stat_activemap_write_error++;
351 		ret = -1;
352 	}
353 	if (ret == 0 && res->hr_metaflush == 1 &&
354 	    g_flush(res->hr_localfd) == -1) {
355 		if (errno == EOPNOTSUPP) {
356 			pjdlog_warning("The %s provider doesn't support flushing write cache. Disabling it.",
357 			    res->hr_localpath);
358 			res->hr_metaflush = 0;
359 		} else {
360 			pjdlog_errno(LOG_ERR,
361 			    "Unable to flush disk cache on activemap update");
362 			res->hr_stat_activemap_flush_error++;
363 			ret = -1;
364 		}
365 	}
366 	mtx_unlock(&res->hr_amp_diskmap_lock);
367 	return (ret);
368 }
369 
370 static bool
371 real_remote(const struct hast_resource *res)
372 {
373 
374 	return (strcmp(res->hr_remoteaddr, "none") != 0);
375 }
376 
377 static void
378 init_environment(struct hast_resource *res __unused)
379 {
380 	struct hio *hio;
381 	unsigned int ii, ncomps;
382 
383 	/*
384 	 * In the future it might be per-resource value.
385 	 */
386 	ncomps = HAST_NCOMPONENTS;
387 
388 	/*
389 	 * Allocate memory needed by lists.
390 	 */
391 	hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
392 	if (hio_send_list == NULL) {
393 		primary_exitx(EX_TEMPFAIL,
394 		    "Unable to allocate %zu bytes of memory for send lists.",
395 		    sizeof(hio_send_list[0]) * ncomps);
396 	}
397 	hio_send_list_size = malloc(sizeof(hio_send_list_size[0]) * ncomps);
398 	if (hio_send_list_size == NULL) {
399 		primary_exitx(EX_TEMPFAIL,
400 		    "Unable to allocate %zu bytes of memory for send list counters.",
401 		    sizeof(hio_send_list_size[0]) * ncomps);
402 	}
403 	hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
404 	if (hio_send_list_lock == NULL) {
405 		primary_exitx(EX_TEMPFAIL,
406 		    "Unable to allocate %zu bytes of memory for send list locks.",
407 		    sizeof(hio_send_list_lock[0]) * ncomps);
408 	}
409 	hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
410 	if (hio_send_list_cond == NULL) {
411 		primary_exitx(EX_TEMPFAIL,
412 		    "Unable to allocate %zu bytes of memory for send list condition variables.",
413 		    sizeof(hio_send_list_cond[0]) * ncomps);
414 	}
415 	hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
416 	if (hio_recv_list == NULL) {
417 		primary_exitx(EX_TEMPFAIL,
418 		    "Unable to allocate %zu bytes of memory for recv lists.",
419 		    sizeof(hio_recv_list[0]) * ncomps);
420 	}
421 	hio_recv_list_size = malloc(sizeof(hio_recv_list_size[0]) * ncomps);
422 	if (hio_recv_list_size == NULL) {
423 		primary_exitx(EX_TEMPFAIL,
424 		    "Unable to allocate %zu bytes of memory for recv list counters.",
425 		    sizeof(hio_recv_list_size[0]) * ncomps);
426 	}
427 	hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
428 	if (hio_recv_list_lock == NULL) {
429 		primary_exitx(EX_TEMPFAIL,
430 		    "Unable to allocate %zu bytes of memory for recv list locks.",
431 		    sizeof(hio_recv_list_lock[0]) * ncomps);
432 	}
433 	hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
434 	if (hio_recv_list_cond == NULL) {
435 		primary_exitx(EX_TEMPFAIL,
436 		    "Unable to allocate %zu bytes of memory for recv list condition variables.",
437 		    sizeof(hio_recv_list_cond[0]) * ncomps);
438 	}
439 	hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
440 	if (hio_remote_lock == NULL) {
441 		primary_exitx(EX_TEMPFAIL,
442 		    "Unable to allocate %zu bytes of memory for remote connections locks.",
443 		    sizeof(hio_remote_lock[0]) * ncomps);
444 	}
445 
446 	/*
447 	 * Initialize lists, their counters, locks and condition variables.
448 	 */
449 	TAILQ_INIT(&hio_free_list);
450 	mtx_init(&hio_free_list_lock);
451 	cv_init(&hio_free_list_cond);
452 	for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
453 		TAILQ_INIT(&hio_send_list[ii]);
454 		hio_send_list_size[ii] = 0;
455 		mtx_init(&hio_send_list_lock[ii]);
456 		cv_init(&hio_send_list_cond[ii]);
457 		TAILQ_INIT(&hio_recv_list[ii]);
458 		hio_recv_list_size[ii] = 0;
459 		mtx_init(&hio_recv_list_lock[ii]);
460 		cv_init(&hio_recv_list_cond[ii]);
461 		rw_init(&hio_remote_lock[ii]);
462 	}
463 	TAILQ_INIT(&hio_done_list);
464 	mtx_init(&hio_done_list_lock);
465 	cv_init(&hio_done_list_cond);
466 	mtx_init(&metadata_lock);
467 
468 	/*
469 	 * Allocate requests pool and initialize requests.
470 	 */
471 	for (ii = 0; ii < HAST_HIO_MAX; ii++) {
472 		hio = malloc(sizeof(*hio));
473 		if (hio == NULL) {
474 			primary_exitx(EX_TEMPFAIL,
475 			    "Unable to allocate %zu bytes of memory for hio request.",
476 			    sizeof(*hio));
477 		}
478 		refcnt_init(&hio->hio_countdown, 0);
479 		hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
480 		if (hio->hio_errors == NULL) {
481 			primary_exitx(EX_TEMPFAIL,
482 			    "Unable allocate %zu bytes of memory for hio errors.",
483 			    sizeof(hio->hio_errors[0]) * ncomps);
484 		}
485 		hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
486 		if (hio->hio_next == NULL) {
487 			primary_exitx(EX_TEMPFAIL,
488 			    "Unable allocate %zu bytes of memory for hio_next field.",
489 			    sizeof(hio->hio_next[0]) * ncomps);
490 		}
491 		hio->hio_ggio.gctl_version = G_GATE_VERSION;
492 		hio->hio_ggio.gctl_data = malloc(MAXPHYS);
493 		if (hio->hio_ggio.gctl_data == NULL) {
494 			primary_exitx(EX_TEMPFAIL,
495 			    "Unable to allocate %zu bytes of memory for gctl_data.",
496 			    MAXPHYS);
497 		}
498 		hio->hio_ggio.gctl_length = MAXPHYS;
499 		hio->hio_ggio.gctl_error = 0;
500 		TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
501 		hio_free_list_size++;
502 	}
503 }
504 
505 static bool
506 init_resuid(struct hast_resource *res)
507 {
508 
509 	mtx_lock(&metadata_lock);
510 	if (res->hr_resuid != 0) {
511 		mtx_unlock(&metadata_lock);
512 		return (false);
513 	} else {
514 		/* Initialize unique resource identifier. */
515 		arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
516 		mtx_unlock(&metadata_lock);
517 		if (metadata_write(res) == -1)
518 			exit(EX_NOINPUT);
519 		return (true);
520 	}
521 }
522 
523 static void
524 init_local(struct hast_resource *res)
525 {
526 	unsigned char *buf;
527 	size_t mapsize;
528 
529 	if (metadata_read(res, true) == -1)
530 		exit(EX_NOINPUT);
531 	mtx_init(&res->hr_amp_lock);
532 	if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
533 	    res->hr_local_sectorsize, res->hr_keepdirty) == -1) {
534 		primary_exit(EX_TEMPFAIL, "Unable to create activemap");
535 	}
536 	mtx_init(&range_lock);
537 	cv_init(&range_regular_cond);
538 	if (rangelock_init(&range_regular) == -1)
539 		primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
540 	cv_init(&range_sync_cond);
541 	if (rangelock_init(&range_sync) == -1)
542 		primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
543 	mapsize = activemap_ondisk_size(res->hr_amp);
544 	buf = calloc(1, mapsize);
545 	if (buf == NULL) {
546 		primary_exitx(EX_TEMPFAIL,
547 		    "Unable to allocate buffer for activemap.");
548 	}
549 	if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
550 	    (ssize_t)mapsize) {
551 		primary_exit(EX_NOINPUT, "Unable to read activemap");
552 	}
553 	activemap_copyin(res->hr_amp, buf, mapsize);
554 	free(buf);
555 	if (res->hr_resuid != 0)
556 		return;
557 	/*
558 	 * We're using provider for the first time. Initialize local and remote
559 	 * counters. We don't initialize resuid here, as we want to do it just
560 	 * in time. The reason for this is that we want to inform secondary
561 	 * that there were no writes yet, so there is no need to synchronize
562 	 * anything.
563 	 */
564 	res->hr_primary_localcnt = 0;
565 	res->hr_primary_remotecnt = 0;
566 	if (metadata_write(res) == -1)
567 		exit(EX_NOINPUT);
568 }
569 
570 static int
571 primary_connect(struct hast_resource *res, struct proto_conn **connp)
572 {
573 	struct proto_conn *conn;
574 	int16_t val;
575 
576 	val = 1;
577 	if (proto_send(res->hr_conn, &val, sizeof(val)) == -1) {
578 		primary_exit(EX_TEMPFAIL,
579 		    "Unable to send connection request to parent");
580 	}
581 	if (proto_recv(res->hr_conn, &val, sizeof(val)) == -1) {
582 		primary_exit(EX_TEMPFAIL,
583 		    "Unable to receive reply to connection request from parent");
584 	}
585 	if (val != 0) {
586 		errno = val;
587 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
588 		    res->hr_remoteaddr);
589 		return (-1);
590 	}
591 	if (proto_connection_recv(res->hr_conn, true, &conn) == -1) {
592 		primary_exit(EX_TEMPFAIL,
593 		    "Unable to receive connection from parent");
594 	}
595 	if (proto_connect_wait(conn, res->hr_timeout) == -1) {
596 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
597 		    res->hr_remoteaddr);
598 		proto_close(conn);
599 		return (-1);
600 	}
601 	/* Error in setting timeout is not critical, but why should it fail? */
602 	if (proto_timeout(conn, res->hr_timeout) == -1)
603 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
604 
605 	*connp = conn;
606 
607 	return (0);
608 }
609 
610 /*
611  * Function instructs GEOM_GATE to handle reads directly from within the kernel.
612  */
613 static void
614 enable_direct_reads(struct hast_resource *res)
615 {
616 	struct g_gate_ctl_modify ggiomodify;
617 
618 	bzero(&ggiomodify, sizeof(ggiomodify));
619 	ggiomodify.gctl_version = G_GATE_VERSION;
620 	ggiomodify.gctl_unit = res->hr_ggateunit;
621 	ggiomodify.gctl_modify = GG_MODIFY_READPROV | GG_MODIFY_READOFFSET;
622 	strlcpy(ggiomodify.gctl_readprov, res->hr_localpath,
623 	    sizeof(ggiomodify.gctl_readprov));
624 	ggiomodify.gctl_readoffset = res->hr_localoff;
625 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_MODIFY, &ggiomodify) == 0)
626 		pjdlog_debug(1, "Direct reads enabled.");
627 	else
628 		pjdlog_errno(LOG_WARNING, "Failed to enable direct reads");
629 }
630 
631 static int
632 init_remote(struct hast_resource *res, struct proto_conn **inp,
633     struct proto_conn **outp)
634 {
635 	struct proto_conn *in, *out;
636 	struct nv *nvout, *nvin;
637 	const unsigned char *token;
638 	unsigned char *map;
639 	const char *errmsg;
640 	int32_t extentsize;
641 	int64_t datasize;
642 	uint32_t mapsize;
643 	uint8_t version;
644 	size_t size;
645 	int error;
646 
647 	PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
648 	PJDLOG_ASSERT(real_remote(res));
649 
650 	in = out = NULL;
651 	errmsg = NULL;
652 
653 	if (primary_connect(res, &out) == -1)
654 		return (ECONNREFUSED);
655 
656 	error = ECONNABORTED;
657 
658 	/*
659 	 * First handshake step.
660 	 * Setup outgoing connection with remote node.
661 	 */
662 	nvout = nv_alloc();
663 	nv_add_string(nvout, res->hr_name, "resource");
664 	nv_add_uint8(nvout, HAST_PROTO_VERSION, "version");
665 	if (nv_error(nvout) != 0) {
666 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
667 		    "Unable to allocate header for connection with %s",
668 		    res->hr_remoteaddr);
669 		nv_free(nvout);
670 		goto close;
671 	}
672 	if (hast_proto_send(res, out, nvout, NULL, 0) == -1) {
673 		pjdlog_errno(LOG_WARNING,
674 		    "Unable to send handshake header to %s",
675 		    res->hr_remoteaddr);
676 		nv_free(nvout);
677 		goto close;
678 	}
679 	nv_free(nvout);
680 	if (hast_proto_recv_hdr(out, &nvin) == -1) {
681 		pjdlog_errno(LOG_WARNING,
682 		    "Unable to receive handshake header from %s",
683 		    res->hr_remoteaddr);
684 		goto close;
685 	}
686 	errmsg = nv_get_string(nvin, "errmsg");
687 	if (errmsg != NULL) {
688 		pjdlog_warning("%s", errmsg);
689 		if (nv_exists(nvin, "wait"))
690 			error = EBUSY;
691 		nv_free(nvin);
692 		goto close;
693 	}
694 	version = nv_get_uint8(nvin, "version");
695 	if (version == 0) {
696 		/*
697 		 * If no version is sent, it means this is protocol version 1.
698 		 */
699 		version = 1;
700 	}
701 	if (version > HAST_PROTO_VERSION) {
702 		pjdlog_warning("Invalid version received (%hhu).", version);
703 		nv_free(nvin);
704 		goto close;
705 	}
706 	res->hr_version = version;
707 	pjdlog_debug(1, "Negotiated protocol version %d.", res->hr_version);
708 	token = nv_get_uint8_array(nvin, &size, "token");
709 	if (token == NULL) {
710 		pjdlog_warning("Handshake header from %s has no 'token' field.",
711 		    res->hr_remoteaddr);
712 		nv_free(nvin);
713 		goto close;
714 	}
715 	if (size != sizeof(res->hr_token)) {
716 		pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
717 		    res->hr_remoteaddr, size, sizeof(res->hr_token));
718 		nv_free(nvin);
719 		goto close;
720 	}
721 	bcopy(token, res->hr_token, sizeof(res->hr_token));
722 	nv_free(nvin);
723 
724 	/*
725 	 * Second handshake step.
726 	 * Setup incoming connection with remote node.
727 	 */
728 	if (primary_connect(res, &in) == -1)
729 		goto close;
730 
731 	nvout = nv_alloc();
732 	nv_add_string(nvout, res->hr_name, "resource");
733 	nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
734 	    "token");
735 	if (res->hr_resuid == 0) {
736 		/*
737 		 * The resuid field was not yet initialized.
738 		 * Because we do synchronization inside init_resuid(), it is
739 		 * possible that someone already initialized it, the function
740 		 * will return false then, but if we successfully initialized
741 		 * it, we will get true. True means that there were no writes
742 		 * to this resource yet and we want to inform secondary that
743 		 * synchronization is not needed by sending "virgin" argument.
744 		 */
745 		if (init_resuid(res))
746 			nv_add_int8(nvout, 1, "virgin");
747 	}
748 	nv_add_uint64(nvout, res->hr_resuid, "resuid");
749 	nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
750 	nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
751 	if (nv_error(nvout) != 0) {
752 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
753 		    "Unable to allocate header for connection with %s",
754 		    res->hr_remoteaddr);
755 		nv_free(nvout);
756 		goto close;
757 	}
758 	if (hast_proto_send(res, in, nvout, NULL, 0) == -1) {
759 		pjdlog_errno(LOG_WARNING,
760 		    "Unable to send handshake header to %s",
761 		    res->hr_remoteaddr);
762 		nv_free(nvout);
763 		goto close;
764 	}
765 	nv_free(nvout);
766 	if (hast_proto_recv_hdr(out, &nvin) == -1) {
767 		pjdlog_errno(LOG_WARNING,
768 		    "Unable to receive handshake header from %s",
769 		    res->hr_remoteaddr);
770 		goto close;
771 	}
772 	errmsg = nv_get_string(nvin, "errmsg");
773 	if (errmsg != NULL) {
774 		pjdlog_warning("%s", errmsg);
775 		nv_free(nvin);
776 		goto close;
777 	}
778 	datasize = nv_get_int64(nvin, "datasize");
779 	if (datasize != res->hr_datasize) {
780 		pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
781 		    (intmax_t)res->hr_datasize, (intmax_t)datasize);
782 		nv_free(nvin);
783 		goto close;
784 	}
785 	extentsize = nv_get_int32(nvin, "extentsize");
786 	if (extentsize != res->hr_extentsize) {
787 		pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
788 		    (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
789 		nv_free(nvin);
790 		goto close;
791 	}
792 	res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
793 	res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
794 	res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
795 	if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY)
796 		enable_direct_reads(res);
797 	if (nv_exists(nvin, "virgin")) {
798 		/*
799 		 * Secondary was reinitialized, bump localcnt if it is 0 as
800 		 * only we have the data.
801 		 */
802 		PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY);
803 		PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
804 
805 		if (res->hr_primary_localcnt == 0) {
806 			PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0);
807 
808 			mtx_lock(&metadata_lock);
809 			res->hr_primary_localcnt++;
810 			pjdlog_debug(1, "Increasing localcnt to %ju.",
811 			    (uintmax_t)res->hr_primary_localcnt);
812 			(void)metadata_write(res);
813 			mtx_unlock(&metadata_lock);
814 		}
815 	}
816 	map = NULL;
817 	mapsize = nv_get_uint32(nvin, "mapsize");
818 	if (mapsize > 0) {
819 		map = malloc(mapsize);
820 		if (map == NULL) {
821 			pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
822 			    (uintmax_t)mapsize);
823 			nv_free(nvin);
824 			goto close;
825 		}
826 		/*
827 		 * Remote node have some dirty extents on its own, lets
828 		 * download its activemap.
829 		 */
830 		if (hast_proto_recv_data(res, out, nvin, map,
831 		    mapsize) == -1) {
832 			pjdlog_errno(LOG_ERR,
833 			    "Unable to receive remote activemap");
834 			nv_free(nvin);
835 			free(map);
836 			goto close;
837 		}
838 		mtx_lock(&res->hr_amp_lock);
839 		/*
840 		 * Merge local and remote bitmaps.
841 		 */
842 		activemap_merge(res->hr_amp, map, mapsize);
843 		free(map);
844 		/*
845 		 * Now that we merged bitmaps from both nodes, flush it to the
846 		 * disk before we start to synchronize.
847 		 */
848 		(void)hast_activemap_flush(res);
849 	}
850 	nv_free(nvin);
851 #ifdef notyet
852 	/* Setup directions. */
853 	if (proto_send(out, NULL, 0) == -1)
854 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
855 	if (proto_recv(in, NULL, 0) == -1)
856 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
857 #endif
858 	pjdlog_info("Connected to %s.", res->hr_remoteaddr);
859 	if (res->hr_original_replication == HAST_REPLICATION_MEMSYNC &&
860 	    res->hr_version < 2) {
861 		pjdlog_warning("The 'memsync' replication mode is not supported by the remote node, falling back to 'fullsync' mode.");
862 		res->hr_replication = HAST_REPLICATION_FULLSYNC;
863 	} else if (res->hr_replication != res->hr_original_replication) {
864 		/*
865 		 * This is in case hastd disconnected and was upgraded.
866 		 */
867 		res->hr_replication = res->hr_original_replication;
868 	}
869 	if (inp != NULL && outp != NULL) {
870 		*inp = in;
871 		*outp = out;
872 	} else {
873 		res->hr_remotein = in;
874 		res->hr_remoteout = out;
875 	}
876 	event_send(res, EVENT_CONNECT);
877 	return (0);
878 close:
879 	if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
880 		event_send(res, EVENT_SPLITBRAIN);
881 	proto_close(out);
882 	if (in != NULL)
883 		proto_close(in);
884 	return (error);
885 }
886 
887 static void
888 sync_start(void)
889 {
890 
891 	mtx_lock(&sync_lock);
892 	sync_inprogress = true;
893 	mtx_unlock(&sync_lock);
894 	cv_signal(&sync_cond);
895 }
896 
897 static void
898 sync_stop(void)
899 {
900 
901 	mtx_lock(&sync_lock);
902 	if (sync_inprogress)
903 		sync_inprogress = false;
904 	mtx_unlock(&sync_lock);
905 }
906 
907 static void
908 init_ggate(struct hast_resource *res)
909 {
910 	struct g_gate_ctl_create ggiocreate;
911 	struct g_gate_ctl_cancel ggiocancel;
912 
913 	/*
914 	 * We communicate with ggate via /dev/ggctl. Open it.
915 	 */
916 	res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
917 	if (res->hr_ggatefd == -1)
918 		primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
919 	/*
920 	 * Create provider before trying to connect, as connection failure
921 	 * is not critical, but may take some time.
922 	 */
923 	bzero(&ggiocreate, sizeof(ggiocreate));
924 	ggiocreate.gctl_version = G_GATE_VERSION;
925 	ggiocreate.gctl_mediasize = res->hr_datasize;
926 	ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
927 	ggiocreate.gctl_flags = 0;
928 	ggiocreate.gctl_maxcount = 0;
929 	ggiocreate.gctl_timeout = 0;
930 	ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
931 	snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
932 	    res->hr_provname);
933 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
934 		pjdlog_info("Device hast/%s created.", res->hr_provname);
935 		res->hr_ggateunit = ggiocreate.gctl_unit;
936 		return;
937 	}
938 	if (errno != EEXIST) {
939 		primary_exit(EX_OSERR, "Unable to create hast/%s device",
940 		    res->hr_provname);
941 	}
942 	pjdlog_debug(1,
943 	    "Device hast/%s already exists, we will try to take it over.",
944 	    res->hr_provname);
945 	/*
946 	 * If we received EEXIST, we assume that the process who created the
947 	 * provider died and didn't clean up. In that case we will start from
948 	 * where he left of.
949 	 */
950 	bzero(&ggiocancel, sizeof(ggiocancel));
951 	ggiocancel.gctl_version = G_GATE_VERSION;
952 	ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
953 	snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
954 	    res->hr_provname);
955 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
956 		pjdlog_info("Device hast/%s recovered.", res->hr_provname);
957 		res->hr_ggateunit = ggiocancel.gctl_unit;
958 		return;
959 	}
960 	primary_exit(EX_OSERR, "Unable to take over hast/%s device",
961 	    res->hr_provname);
962 }
963 
964 void
965 hastd_primary(struct hast_resource *res)
966 {
967 	pthread_t td;
968 	pid_t pid;
969 	int error, mode, debuglevel;
970 
971 	/*
972 	 * Create communication channel for sending control commands from
973 	 * parent to child.
974 	 */
975 	if (proto_client(NULL, "socketpair://", &res->hr_ctrl) == -1) {
976 		/* TODO: There's no need for this to be fatal error. */
977 		KEEP_ERRNO((void)pidfile_remove(pfh));
978 		pjdlog_exit(EX_OSERR,
979 		    "Unable to create control sockets between parent and child");
980 	}
981 	/*
982 	 * Create communication channel for sending events from child to parent.
983 	 */
984 	if (proto_client(NULL, "socketpair://", &res->hr_event) == -1) {
985 		/* TODO: There's no need for this to be fatal error. */
986 		KEEP_ERRNO((void)pidfile_remove(pfh));
987 		pjdlog_exit(EX_OSERR,
988 		    "Unable to create event sockets between child and parent");
989 	}
990 	/*
991 	 * Create communication channel for sending connection requests from
992 	 * child to parent.
993 	 */
994 	if (proto_client(NULL, "socketpair://", &res->hr_conn) == -1) {
995 		/* TODO: There's no need for this to be fatal error. */
996 		KEEP_ERRNO((void)pidfile_remove(pfh));
997 		pjdlog_exit(EX_OSERR,
998 		    "Unable to create connection sockets between child and parent");
999 	}
1000 
1001 	pid = fork();
1002 	if (pid == -1) {
1003 		/* TODO: There's no need for this to be fatal error. */
1004 		KEEP_ERRNO((void)pidfile_remove(pfh));
1005 		pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
1006 	}
1007 
1008 	if (pid > 0) {
1009 		/* This is parent. */
1010 		/* Declare that we are receiver. */
1011 		proto_recv(res->hr_event, NULL, 0);
1012 		proto_recv(res->hr_conn, NULL, 0);
1013 		/* Declare that we are sender. */
1014 		proto_send(res->hr_ctrl, NULL, 0);
1015 		res->hr_workerpid = pid;
1016 		return;
1017 	}
1018 
1019 	gres = res;
1020 	res->output_status_aux = output_status_aux;
1021 	mode = pjdlog_mode_get();
1022 	debuglevel = pjdlog_debug_get();
1023 
1024 	/* Declare that we are sender. */
1025 	proto_send(res->hr_event, NULL, 0);
1026 	proto_send(res->hr_conn, NULL, 0);
1027 	/* Declare that we are receiver. */
1028 	proto_recv(res->hr_ctrl, NULL, 0);
1029 	descriptors_cleanup(res);
1030 
1031 	descriptors_assert(res, mode);
1032 
1033 	pjdlog_init(mode);
1034 	pjdlog_debug_set(debuglevel);
1035 	pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
1036 	setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
1037 
1038 	init_local(res);
1039 	init_ggate(res);
1040 	init_environment(res);
1041 
1042 	if (drop_privs(res) != 0) {
1043 		cleanup(res);
1044 		exit(EX_CONFIG);
1045 	}
1046 	pjdlog_info("Privileges successfully dropped.");
1047 
1048 	/*
1049 	 * Create the guard thread first, so we can handle signals from the
1050 	 * very beginning.
1051 	 */
1052 	error = pthread_create(&td, NULL, guard_thread, res);
1053 	PJDLOG_ASSERT(error == 0);
1054 	/*
1055 	 * Create the control thread before sending any event to the parent,
1056 	 * as we can deadlock when parent sends control request to worker,
1057 	 * but worker has no control thread started yet, so parent waits.
1058 	 * In the meantime worker sends an event to the parent, but parent
1059 	 * is unable to handle the event, because it waits for control
1060 	 * request response.
1061 	 */
1062 	error = pthread_create(&td, NULL, ctrl_thread, res);
1063 	PJDLOG_ASSERT(error == 0);
1064 	if (real_remote(res)) {
1065 		error = init_remote(res, NULL, NULL);
1066 		if (error == 0) {
1067 			sync_start();
1068 		} else if (error == EBUSY) {
1069 			time_t start = time(NULL);
1070 
1071 			pjdlog_warning("Waiting for remote node to become %s for %ds.",
1072 			    role2str(HAST_ROLE_SECONDARY),
1073 			    res->hr_timeout);
1074 			for (;;) {
1075 				sleep(1);
1076 				error = init_remote(res, NULL, NULL);
1077 				if (error != EBUSY)
1078 					break;
1079 				if (time(NULL) > start + res->hr_timeout)
1080 					break;
1081 			}
1082 			if (error == EBUSY) {
1083 				pjdlog_warning("Remote node is still %s, starting anyway.",
1084 				    role2str(HAST_ROLE_PRIMARY));
1085 			}
1086 		}
1087 	}
1088 	error = pthread_create(&td, NULL, ggate_recv_thread, res);
1089 	PJDLOG_ASSERT(error == 0);
1090 	error = pthread_create(&td, NULL, local_send_thread, res);
1091 	PJDLOG_ASSERT(error == 0);
1092 	error = pthread_create(&td, NULL, remote_send_thread, res);
1093 	PJDLOG_ASSERT(error == 0);
1094 	error = pthread_create(&td, NULL, remote_recv_thread, res);
1095 	PJDLOG_ASSERT(error == 0);
1096 	error = pthread_create(&td, NULL, ggate_send_thread, res);
1097 	PJDLOG_ASSERT(error == 0);
1098 	fullystarted = true;
1099 	(void)sync_thread(res);
1100 }
1101 
1102 static void
1103 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio,
1104     const char *fmt, ...)
1105 {
1106 	char msg[1024];
1107 	va_list ap;
1108 
1109 	va_start(ap, fmt);
1110 	(void)vsnprintf(msg, sizeof(msg), fmt, ap);
1111 	va_end(ap);
1112 	switch (ggio->gctl_cmd) {
1113 	case BIO_READ:
1114 		(void)snprlcat(msg, sizeof(msg), "READ(%ju, %ju).",
1115 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1116 		break;
1117 	case BIO_DELETE:
1118 		(void)snprlcat(msg, sizeof(msg), "DELETE(%ju, %ju).",
1119 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1120 		break;
1121 	case BIO_FLUSH:
1122 		(void)snprlcat(msg, sizeof(msg), "FLUSH.");
1123 		break;
1124 	case BIO_WRITE:
1125 		(void)snprlcat(msg, sizeof(msg), "WRITE(%ju, %ju).",
1126 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1127 		break;
1128 	default:
1129 		(void)snprlcat(msg, sizeof(msg), "UNKNOWN(%u).",
1130 		    (unsigned int)ggio->gctl_cmd);
1131 		break;
1132 	}
1133 	pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
1134 }
1135 
1136 static void
1137 remote_close(struct hast_resource *res, int ncomp)
1138 {
1139 
1140 	rw_wlock(&hio_remote_lock[ncomp]);
1141 	/*
1142 	 * Check for a race between dropping rlock and acquiring wlock -
1143 	 * another thread can close connection in-between.
1144 	 */
1145 	if (!ISCONNECTED(res, ncomp)) {
1146 		PJDLOG_ASSERT(res->hr_remotein == NULL);
1147 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
1148 		rw_unlock(&hio_remote_lock[ncomp]);
1149 		return;
1150 	}
1151 
1152 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1153 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1154 
1155 	pjdlog_debug(2, "Closing incoming connection to %s.",
1156 	    res->hr_remoteaddr);
1157 	proto_close(res->hr_remotein);
1158 	res->hr_remotein = NULL;
1159 	pjdlog_debug(2, "Closing outgoing connection to %s.",
1160 	    res->hr_remoteaddr);
1161 	proto_close(res->hr_remoteout);
1162 	res->hr_remoteout = NULL;
1163 
1164 	rw_unlock(&hio_remote_lock[ncomp]);
1165 
1166 	pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
1167 
1168 	/*
1169 	 * Stop synchronization if in-progress.
1170 	 */
1171 	sync_stop();
1172 
1173 	event_send(res, EVENT_DISCONNECT);
1174 }
1175 
1176 /*
1177  * Acknowledge write completion to the kernel, but don't update activemap yet.
1178  */
1179 static void
1180 write_complete(struct hast_resource *res, struct hio *hio)
1181 {
1182 	struct g_gate_ctl_io *ggio;
1183 	unsigned int ncomp;
1184 
1185 	PJDLOG_ASSERT(!hio->hio_done);
1186 
1187 	ggio = &hio->hio_ggio;
1188 	PJDLOG_ASSERT(ggio->gctl_cmd == BIO_WRITE);
1189 
1190 	/*
1191 	 * Bump local count if this is first write after
1192 	 * connection failure with remote node.
1193 	 */
1194 	ncomp = 1;
1195 	rw_rlock(&hio_remote_lock[ncomp]);
1196 	if (!ISCONNECTED(res, ncomp)) {
1197 		mtx_lock(&metadata_lock);
1198 		if (res->hr_primary_localcnt == res->hr_secondary_remotecnt) {
1199 			res->hr_primary_localcnt++;
1200 			pjdlog_debug(1, "Increasing localcnt to %ju.",
1201 			    (uintmax_t)res->hr_primary_localcnt);
1202 			(void)metadata_write(res);
1203 		}
1204 		mtx_unlock(&metadata_lock);
1205 	}
1206 	rw_unlock(&hio_remote_lock[ncomp]);
1207 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1)
1208 		primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1209 	hio->hio_done = true;
1210 }
1211 
1212 /*
1213  * Thread receives ggate I/O requests from the kernel and passes them to
1214  * appropriate threads:
1215  * WRITE - always goes to both local_send and remote_send threads
1216  * READ (when the block is up-to-date on local component) -
1217  *	only local_send thread
1218  * READ (when the block isn't up-to-date on local component) -
1219  *	only remote_send thread
1220  * DELETE - always goes to both local_send and remote_send threads
1221  * FLUSH - always goes to both local_send and remote_send threads
1222  */
1223 static void *
1224 ggate_recv_thread(void *arg)
1225 {
1226 	struct hast_resource *res = arg;
1227 	struct g_gate_ctl_io *ggio;
1228 	struct hio *hio;
1229 	unsigned int ii, ncomp, ncomps;
1230 	int error;
1231 
1232 	for (;;) {
1233 		pjdlog_debug(2, "ggate_recv: Taking free request.");
1234 		QUEUE_TAKE2(hio, free);
1235 		pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1236 		ggio = &hio->hio_ggio;
1237 		ggio->gctl_unit = res->hr_ggateunit;
1238 		ggio->gctl_length = MAXPHYS;
1239 		ggio->gctl_error = 0;
1240 		hio->hio_done = false;
1241 		hio->hio_replication = res->hr_replication;
1242 		pjdlog_debug(2,
1243 		    "ggate_recv: (%p) Waiting for request from the kernel.",
1244 		    hio);
1245 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) == -1) {
1246 			if (sigexit_received)
1247 				pthread_exit(NULL);
1248 			primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1249 		}
1250 		error = ggio->gctl_error;
1251 		switch (error) {
1252 		case 0:
1253 			break;
1254 		case ECANCELED:
1255 			/* Exit gracefully. */
1256 			if (!sigexit_received) {
1257 				pjdlog_debug(2,
1258 				    "ggate_recv: (%p) Received cancel from the kernel.",
1259 				    hio);
1260 				pjdlog_info("Received cancel from the kernel, exiting.");
1261 			}
1262 			pthread_exit(NULL);
1263 		case ENOMEM:
1264 			/*
1265 			 * Buffer too small? Impossible, we allocate MAXPHYS
1266 			 * bytes - request can't be bigger than that.
1267 			 */
1268 			/* FALLTHROUGH */
1269 		case ENXIO:
1270 		default:
1271 			primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1272 			    strerror(error));
1273 		}
1274 
1275 		ncomp = 0;
1276 		ncomps = HAST_NCOMPONENTS;
1277 
1278 		for (ii = 0; ii < ncomps; ii++)
1279 			hio->hio_errors[ii] = EINVAL;
1280 		reqlog(LOG_DEBUG, 2, ggio,
1281 		    "ggate_recv: (%p) Request received from the kernel: ",
1282 		    hio);
1283 
1284 		/*
1285 		 * Inform all components about new write request.
1286 		 * For read request prefer local component unless the given
1287 		 * range is out-of-date, then use remote component.
1288 		 */
1289 		switch (ggio->gctl_cmd) {
1290 		case BIO_READ:
1291 			res->hr_stat_read++;
1292 			ncomps = 1;
1293 			mtx_lock(&metadata_lock);
1294 			if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1295 			    res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1296 				/*
1297 				 * This range is up-to-date on local component,
1298 				 * so handle request locally.
1299 				 */
1300 				 /* Local component is 0 for now. */
1301 				ncomp = 0;
1302 			} else /* if (res->hr_syncsrc ==
1303 			    HAST_SYNCSRC_SECONDARY) */ {
1304 				PJDLOG_ASSERT(res->hr_syncsrc ==
1305 				    HAST_SYNCSRC_SECONDARY);
1306 				/*
1307 				 * This range is out-of-date on local component,
1308 				 * so send request to the remote node.
1309 				 */
1310 				 /* Remote component is 1 for now. */
1311 				ncomp = 1;
1312 			}
1313 			mtx_unlock(&metadata_lock);
1314 			break;
1315 		case BIO_WRITE:
1316 			res->hr_stat_write++;
1317 			if (res->hr_resuid == 0 &&
1318 			    res->hr_primary_localcnt == 0) {
1319 				/* This is first write. */
1320 				res->hr_primary_localcnt = 1;
1321 			}
1322 			for (;;) {
1323 				mtx_lock(&range_lock);
1324 				if (rangelock_islocked(range_sync,
1325 				    ggio->gctl_offset, ggio->gctl_length)) {
1326 					pjdlog_debug(2,
1327 					    "regular: Range offset=%jd length=%zu locked.",
1328 					    (intmax_t)ggio->gctl_offset,
1329 					    (size_t)ggio->gctl_length);
1330 					range_regular_wait = true;
1331 					cv_wait(&range_regular_cond, &range_lock);
1332 					range_regular_wait = false;
1333 					mtx_unlock(&range_lock);
1334 					continue;
1335 				}
1336 				if (rangelock_add(range_regular,
1337 				    ggio->gctl_offset, ggio->gctl_length) == -1) {
1338 					mtx_unlock(&range_lock);
1339 					pjdlog_debug(2,
1340 					    "regular: Range offset=%jd length=%zu is already locked, waiting.",
1341 					    (intmax_t)ggio->gctl_offset,
1342 					    (size_t)ggio->gctl_length);
1343 					sleep(1);
1344 					continue;
1345 				}
1346 				mtx_unlock(&range_lock);
1347 				break;
1348 			}
1349 			mtx_lock(&res->hr_amp_lock);
1350 			if (activemap_write_start(res->hr_amp,
1351 			    ggio->gctl_offset, ggio->gctl_length)) {
1352 				res->hr_stat_activemap_update++;
1353 				(void)hast_activemap_flush(res);
1354 			} else {
1355 				mtx_unlock(&res->hr_amp_lock);
1356 			}
1357 			if (ISMEMSYNC(hio)) {
1358 				hio->hio_memsyncacked = false;
1359 				refcnt_init(&hio->hio_writecount, ncomps);
1360 			}
1361 			break;
1362 		case BIO_DELETE:
1363 			res->hr_stat_delete++;
1364 			break;
1365 		case BIO_FLUSH:
1366 			res->hr_stat_flush++;
1367 			break;
1368 		}
1369 		pjdlog_debug(2,
1370 		    "ggate_recv: (%p) Moving request to the send queues.", hio);
1371 		refcnt_init(&hio->hio_countdown, ncomps);
1372 		for (ii = ncomp; ii < ncomps; ii++)
1373 			QUEUE_INSERT1(hio, send, ii);
1374 	}
1375 	/* NOTREACHED */
1376 	return (NULL);
1377 }
1378 
1379 /*
1380  * Thread reads from or writes to local component.
1381  * If local read fails, it redirects it to remote_send thread.
1382  */
1383 static void *
1384 local_send_thread(void *arg)
1385 {
1386 	struct hast_resource *res = arg;
1387 	struct g_gate_ctl_io *ggio;
1388 	struct hio *hio;
1389 	unsigned int ncomp, rncomp;
1390 	ssize_t ret;
1391 
1392 	/* Local component is 0 for now. */
1393 	ncomp = 0;
1394 	/* Remote component is 1 for now. */
1395 	rncomp = 1;
1396 
1397 	for (;;) {
1398 		pjdlog_debug(2, "local_send: Taking request.");
1399 		QUEUE_TAKE1(hio, send, ncomp, 0);
1400 		pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1401 		ggio = &hio->hio_ggio;
1402 		switch (ggio->gctl_cmd) {
1403 		case BIO_READ:
1404 			ret = pread(res->hr_localfd, ggio->gctl_data,
1405 			    ggio->gctl_length,
1406 			    ggio->gctl_offset + res->hr_localoff);
1407 			if (ret == ggio->gctl_length)
1408 				hio->hio_errors[ncomp] = 0;
1409 			else if (!ISSYNCREQ(hio)) {
1410 				/*
1411 				 * If READ failed, try to read from remote node.
1412 				 */
1413 				if (ret == -1) {
1414 					reqlog(LOG_WARNING, 0, ggio,
1415 					    "Local request failed (%s), trying remote node. ",
1416 					    strerror(errno));
1417 				} else if (ret != ggio->gctl_length) {
1418 					reqlog(LOG_WARNING, 0, ggio,
1419 					    "Local request failed (%zd != %jd), trying remote node. ",
1420 					    ret, (intmax_t)ggio->gctl_length);
1421 				}
1422 				QUEUE_INSERT1(hio, send, rncomp);
1423 				continue;
1424 			}
1425 			break;
1426 		case BIO_WRITE:
1427 			ret = pwrite(res->hr_localfd, ggio->gctl_data,
1428 			    ggio->gctl_length,
1429 			    ggio->gctl_offset + res->hr_localoff);
1430 			if (ret == -1) {
1431 				hio->hio_errors[ncomp] = errno;
1432 				reqlog(LOG_WARNING, 0, ggio,
1433 				    "Local request failed (%s): ",
1434 				    strerror(errno));
1435 			} else if (ret != ggio->gctl_length) {
1436 				hio->hio_errors[ncomp] = EIO;
1437 				reqlog(LOG_WARNING, 0, ggio,
1438 				    "Local request failed (%zd != %jd): ",
1439 				    ret, (intmax_t)ggio->gctl_length);
1440 			} else {
1441 				hio->hio_errors[ncomp] = 0;
1442 				if (ISASYNC(hio)) {
1443 					ggio->gctl_error = 0;
1444 					write_complete(res, hio);
1445 				}
1446 			}
1447 			break;
1448 		case BIO_DELETE:
1449 			ret = g_delete(res->hr_localfd,
1450 			    ggio->gctl_offset + res->hr_localoff,
1451 			    ggio->gctl_length);
1452 			if (ret == -1) {
1453 				hio->hio_errors[ncomp] = errno;
1454 				reqlog(LOG_WARNING, 0, ggio,
1455 				    "Local request failed (%s): ",
1456 				    strerror(errno));
1457 			} else {
1458 				hio->hio_errors[ncomp] = 0;
1459 			}
1460 			break;
1461 		case BIO_FLUSH:
1462 			if (!res->hr_localflush) {
1463 				ret = -1;
1464 				errno = EOPNOTSUPP;
1465 				break;
1466 			}
1467 			ret = g_flush(res->hr_localfd);
1468 			if (ret == -1) {
1469 				if (errno == EOPNOTSUPP)
1470 					res->hr_localflush = false;
1471 				hio->hio_errors[ncomp] = errno;
1472 				reqlog(LOG_WARNING, 0, ggio,
1473 				    "Local request failed (%s): ",
1474 				    strerror(errno));
1475 			} else {
1476 				hio->hio_errors[ncomp] = 0;
1477 			}
1478 			break;
1479 		}
1480 		if (ISMEMSYNCWRITE(hio)) {
1481 			if (refcnt_release(&hio->hio_writecount) == 0) {
1482 				write_complete(res, hio);
1483 			}
1484 		}
1485 		if (refcnt_release(&hio->hio_countdown) > 0)
1486 			continue;
1487 		if (ISSYNCREQ(hio)) {
1488 			mtx_lock(&sync_lock);
1489 			SYNCREQDONE(hio);
1490 			mtx_unlock(&sync_lock);
1491 			cv_signal(&sync_cond);
1492 		} else {
1493 			pjdlog_debug(2,
1494 			    "local_send: (%p) Moving request to the done queue.",
1495 			    hio);
1496 			QUEUE_INSERT2(hio, done);
1497 		}
1498 	}
1499 	/* NOTREACHED */
1500 	return (NULL);
1501 }
1502 
1503 static void
1504 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1505 {
1506 	struct nv *nv;
1507 
1508 	rw_rlock(&hio_remote_lock[ncomp]);
1509 
1510 	if (!ISCONNECTED(res, ncomp)) {
1511 		rw_unlock(&hio_remote_lock[ncomp]);
1512 		return;
1513 	}
1514 
1515 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1516 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1517 
1518 	nv = nv_alloc();
1519 	nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1520 	if (nv_error(nv) != 0) {
1521 		rw_unlock(&hio_remote_lock[ncomp]);
1522 		nv_free(nv);
1523 		pjdlog_debug(1,
1524 		    "keepalive_send: Unable to prepare header to send.");
1525 		return;
1526 	}
1527 	if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) == -1) {
1528 		rw_unlock(&hio_remote_lock[ncomp]);
1529 		pjdlog_common(LOG_DEBUG, 1, errno,
1530 		    "keepalive_send: Unable to send request");
1531 		nv_free(nv);
1532 		remote_close(res, ncomp);
1533 		return;
1534 	}
1535 
1536 	rw_unlock(&hio_remote_lock[ncomp]);
1537 	nv_free(nv);
1538 	pjdlog_debug(2, "keepalive_send: Request sent.");
1539 }
1540 
1541 /*
1542  * Thread sends request to secondary node.
1543  */
1544 static void *
1545 remote_send_thread(void *arg)
1546 {
1547 	struct hast_resource *res = arg;
1548 	struct g_gate_ctl_io *ggio;
1549 	time_t lastcheck, now;
1550 	struct hio *hio;
1551 	struct nv *nv;
1552 	unsigned int ncomp;
1553 	bool wakeup;
1554 	uint64_t offset, length;
1555 	uint8_t cmd;
1556 	void *data;
1557 
1558 	/* Remote component is 1 for now. */
1559 	ncomp = 1;
1560 	lastcheck = time(NULL);
1561 
1562 	for (;;) {
1563 		pjdlog_debug(2, "remote_send: Taking request.");
1564 		QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE);
1565 		if (hio == NULL) {
1566 			now = time(NULL);
1567 			if (lastcheck + HAST_KEEPALIVE <= now) {
1568 				keepalive_send(res, ncomp);
1569 				lastcheck = now;
1570 			}
1571 			continue;
1572 		}
1573 		pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1574 		ggio = &hio->hio_ggio;
1575 		switch (ggio->gctl_cmd) {
1576 		case BIO_READ:
1577 			cmd = HIO_READ;
1578 			data = NULL;
1579 			offset = ggio->gctl_offset;
1580 			length = ggio->gctl_length;
1581 			break;
1582 		case BIO_WRITE:
1583 			cmd = HIO_WRITE;
1584 			data = ggio->gctl_data;
1585 			offset = ggio->gctl_offset;
1586 			length = ggio->gctl_length;
1587 			break;
1588 		case BIO_DELETE:
1589 			cmd = HIO_DELETE;
1590 			data = NULL;
1591 			offset = ggio->gctl_offset;
1592 			length = ggio->gctl_length;
1593 			break;
1594 		case BIO_FLUSH:
1595 			cmd = HIO_FLUSH;
1596 			data = NULL;
1597 			offset = 0;
1598 			length = 0;
1599 			break;
1600 		default:
1601 			PJDLOG_ABORT("invalid condition");
1602 		}
1603 		nv = nv_alloc();
1604 		nv_add_uint8(nv, cmd, "cmd");
1605 		nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1606 		nv_add_uint64(nv, offset, "offset");
1607 		nv_add_uint64(nv, length, "length");
1608 		if (ISMEMSYNCWRITE(hio))
1609 			nv_add_uint8(nv, 1, "memsync");
1610 		if (nv_error(nv) != 0) {
1611 			hio->hio_errors[ncomp] = nv_error(nv);
1612 			pjdlog_debug(2,
1613 			    "remote_send: (%p) Unable to prepare header to send.",
1614 			    hio);
1615 			reqlog(LOG_ERR, 0, ggio,
1616 			    "Unable to prepare header to send (%s): ",
1617 			    strerror(nv_error(nv)));
1618 			/* Move failed request immediately to the done queue. */
1619 			goto done_queue;
1620 		}
1621 		/*
1622 		 * Protect connection from disappearing.
1623 		 */
1624 		rw_rlock(&hio_remote_lock[ncomp]);
1625 		if (!ISCONNECTED(res, ncomp)) {
1626 			rw_unlock(&hio_remote_lock[ncomp]);
1627 			hio->hio_errors[ncomp] = ENOTCONN;
1628 			goto done_queue;
1629 		}
1630 		/*
1631 		 * Move the request to recv queue before sending it, because
1632 		 * in different order we can get reply before we move request
1633 		 * to recv queue.
1634 		 */
1635 		pjdlog_debug(2,
1636 		    "remote_send: (%p) Moving request to the recv queue.",
1637 		    hio);
1638 		mtx_lock(&hio_recv_list_lock[ncomp]);
1639 		wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1640 		TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1641 		hio_recv_list_size[ncomp]++;
1642 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1643 		if (hast_proto_send(res, res->hr_remoteout, nv, data,
1644 		    data != NULL ? length : 0) == -1) {
1645 			hio->hio_errors[ncomp] = errno;
1646 			rw_unlock(&hio_remote_lock[ncomp]);
1647 			pjdlog_debug(2,
1648 			    "remote_send: (%p) Unable to send request.", hio);
1649 			reqlog(LOG_ERR, 0, ggio,
1650 			    "Unable to send request (%s): ",
1651 			    strerror(hio->hio_errors[ncomp]));
1652 			remote_close(res, ncomp);
1653 		} else {
1654 			rw_unlock(&hio_remote_lock[ncomp]);
1655 		}
1656 		nv_free(nv);
1657 		if (wakeup)
1658 			cv_signal(&hio_recv_list_cond[ncomp]);
1659 		continue;
1660 done_queue:
1661 		nv_free(nv);
1662 		if (ISSYNCREQ(hio)) {
1663 			if (refcnt_release(&hio->hio_countdown) > 0)
1664 				continue;
1665 			mtx_lock(&sync_lock);
1666 			SYNCREQDONE(hio);
1667 			mtx_unlock(&sync_lock);
1668 			cv_signal(&sync_cond);
1669 			continue;
1670 		}
1671 		if (ggio->gctl_cmd == BIO_WRITE) {
1672 			mtx_lock(&res->hr_amp_lock);
1673 			if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1674 			    ggio->gctl_length)) {
1675 				(void)hast_activemap_flush(res);
1676 			} else {
1677 				mtx_unlock(&res->hr_amp_lock);
1678 			}
1679 			if (ISMEMSYNCWRITE(hio)) {
1680 				if (refcnt_release(&hio->hio_writecount) == 0) {
1681 					if (hio->hio_errors[0] == 0)
1682 						write_complete(res, hio);
1683 				}
1684 			}
1685 		}
1686 		if (refcnt_release(&hio->hio_countdown) > 0)
1687 			continue;
1688 		pjdlog_debug(2,
1689 		    "remote_send: (%p) Moving request to the done queue.",
1690 		    hio);
1691 		QUEUE_INSERT2(hio, done);
1692 	}
1693 	/* NOTREACHED */
1694 	return (NULL);
1695 }
1696 
1697 /*
1698  * Thread receives answer from secondary node and passes it to ggate_send
1699  * thread.
1700  */
1701 static void *
1702 remote_recv_thread(void *arg)
1703 {
1704 	struct hast_resource *res = arg;
1705 	struct g_gate_ctl_io *ggio;
1706 	struct hio *hio;
1707 	struct nv *nv;
1708 	unsigned int ncomp;
1709 	uint64_t seq;
1710 	bool memsyncack;
1711 	int error;
1712 
1713 	/* Remote component is 1 for now. */
1714 	ncomp = 1;
1715 
1716 	for (;;) {
1717 		/* Wait until there is anything to receive. */
1718 		mtx_lock(&hio_recv_list_lock[ncomp]);
1719 		while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1720 			pjdlog_debug(2, "remote_recv: No requests, waiting.");
1721 			cv_wait(&hio_recv_list_cond[ncomp],
1722 			    &hio_recv_list_lock[ncomp]);
1723 		}
1724 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1725 
1726 		memsyncack = false;
1727 
1728 		rw_rlock(&hio_remote_lock[ncomp]);
1729 		if (!ISCONNECTED(res, ncomp)) {
1730 			rw_unlock(&hio_remote_lock[ncomp]);
1731 			/*
1732 			 * Connection is dead, so move all pending requests to
1733 			 * the done queue (one-by-one).
1734 			 */
1735 			mtx_lock(&hio_recv_list_lock[ncomp]);
1736 			hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1737 			PJDLOG_ASSERT(hio != NULL);
1738 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1739 			    hio_next[ncomp]);
1740 			hio_recv_list_size[ncomp]--;
1741 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1742 			hio->hio_errors[ncomp] = ENOTCONN;
1743 			goto done_queue;
1744 		}
1745 		if (hast_proto_recv_hdr(res->hr_remotein, &nv) == -1) {
1746 			pjdlog_errno(LOG_ERR,
1747 			    "Unable to receive reply header");
1748 			rw_unlock(&hio_remote_lock[ncomp]);
1749 			remote_close(res, ncomp);
1750 			continue;
1751 		}
1752 		rw_unlock(&hio_remote_lock[ncomp]);
1753 		seq = nv_get_uint64(nv, "seq");
1754 		if (seq == 0) {
1755 			pjdlog_error("Header contains no 'seq' field.");
1756 			nv_free(nv);
1757 			continue;
1758 		}
1759 		memsyncack = nv_exists(nv, "received");
1760 		mtx_lock(&hio_recv_list_lock[ncomp]);
1761 		TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1762 			if (hio->hio_ggio.gctl_seq == seq) {
1763 				TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1764 				    hio_next[ncomp]);
1765 				hio_recv_list_size[ncomp]--;
1766 				break;
1767 			}
1768 		}
1769 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1770 		if (hio == NULL) {
1771 			pjdlog_error("Found no request matching received 'seq' field (%ju).",
1772 			    (uintmax_t)seq);
1773 			nv_free(nv);
1774 			continue;
1775 		}
1776 		ggio = &hio->hio_ggio;
1777 		error = nv_get_int16(nv, "error");
1778 		if (error != 0) {
1779 			/* Request failed on remote side. */
1780 			hio->hio_errors[ncomp] = error;
1781 			reqlog(LOG_WARNING, 0, ggio,
1782 			    "Remote request failed (%s): ", strerror(error));
1783 			nv_free(nv);
1784 			goto done_queue;
1785 		}
1786 		switch (ggio->gctl_cmd) {
1787 		case BIO_READ:
1788 			rw_rlock(&hio_remote_lock[ncomp]);
1789 			if (!ISCONNECTED(res, ncomp)) {
1790 				rw_unlock(&hio_remote_lock[ncomp]);
1791 				nv_free(nv);
1792 				goto done_queue;
1793 			}
1794 			if (hast_proto_recv_data(res, res->hr_remotein, nv,
1795 			    ggio->gctl_data, ggio->gctl_length) == -1) {
1796 				hio->hio_errors[ncomp] = errno;
1797 				pjdlog_errno(LOG_ERR,
1798 				    "Unable to receive reply data");
1799 				rw_unlock(&hio_remote_lock[ncomp]);
1800 				nv_free(nv);
1801 				remote_close(res, ncomp);
1802 				goto done_queue;
1803 			}
1804 			rw_unlock(&hio_remote_lock[ncomp]);
1805 			break;
1806 		case BIO_WRITE:
1807 		case BIO_DELETE:
1808 		case BIO_FLUSH:
1809 			break;
1810 		default:
1811 			PJDLOG_ABORT("invalid condition");
1812 		}
1813 		hio->hio_errors[ncomp] = 0;
1814 		nv_free(nv);
1815 done_queue:
1816 		if (ISMEMSYNCWRITE(hio)) {
1817 			if (!hio->hio_memsyncacked) {
1818 				PJDLOG_ASSERT(memsyncack ||
1819 				    hio->hio_errors[ncomp] != 0);
1820 				/* Remote ack arrived. */
1821 				if (refcnt_release(&hio->hio_writecount) == 0) {
1822 					if (hio->hio_errors[0] == 0)
1823 						write_complete(res, hio);
1824 				}
1825 				hio->hio_memsyncacked = true;
1826 				if (hio->hio_errors[ncomp] == 0) {
1827 					pjdlog_debug(2,
1828 					    "remote_recv: (%p) Moving request "
1829 					    "back to the recv queue.", hio);
1830 					mtx_lock(&hio_recv_list_lock[ncomp]);
1831 					TAILQ_INSERT_TAIL(&hio_recv_list[ncomp],
1832 					    hio, hio_next[ncomp]);
1833 					hio_recv_list_size[ncomp]++;
1834 					mtx_unlock(&hio_recv_list_lock[ncomp]);
1835 					continue;
1836 				}
1837 			} else {
1838 				PJDLOG_ASSERT(!memsyncack);
1839 				/* Remote final reply arrived. */
1840 			}
1841 		}
1842 		if (refcnt_release(&hio->hio_countdown) > 0)
1843 			continue;
1844 		if (ISSYNCREQ(hio)) {
1845 			mtx_lock(&sync_lock);
1846 			SYNCREQDONE(hio);
1847 			mtx_unlock(&sync_lock);
1848 			cv_signal(&sync_cond);
1849 		} else {
1850 			pjdlog_debug(2,
1851 			    "remote_recv: (%p) Moving request to the done queue.",
1852 			    hio);
1853 			QUEUE_INSERT2(hio, done);
1854 		}
1855 	}
1856 	/* NOTREACHED */
1857 	return (NULL);
1858 }
1859 
1860 /*
1861  * Thread sends answer to the kernel.
1862  */
1863 static void *
1864 ggate_send_thread(void *arg)
1865 {
1866 	struct hast_resource *res = arg;
1867 	struct g_gate_ctl_io *ggio;
1868 	struct hio *hio;
1869 	unsigned int ii, ncomps;
1870 
1871 	ncomps = HAST_NCOMPONENTS;
1872 
1873 	for (;;) {
1874 		pjdlog_debug(2, "ggate_send: Taking request.");
1875 		QUEUE_TAKE2(hio, done);
1876 		pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1877 		ggio = &hio->hio_ggio;
1878 		for (ii = 0; ii < ncomps; ii++) {
1879 			if (hio->hio_errors[ii] == 0) {
1880 				/*
1881 				 * One successful request is enough to declare
1882 				 * success.
1883 				 */
1884 				ggio->gctl_error = 0;
1885 				break;
1886 			}
1887 		}
1888 		if (ii == ncomps) {
1889 			/*
1890 			 * None of the requests were successful.
1891 			 * Use the error from local component except the
1892 			 * case when we did only remote request.
1893 			 */
1894 			if (ggio->gctl_cmd == BIO_READ &&
1895 			    res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
1896 				ggio->gctl_error = hio->hio_errors[1];
1897 			else
1898 				ggio->gctl_error = hio->hio_errors[0];
1899 		}
1900 		if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1901 			mtx_lock(&res->hr_amp_lock);
1902 			if (activemap_write_complete(res->hr_amp,
1903 			    ggio->gctl_offset, ggio->gctl_length)) {
1904 				res->hr_stat_activemap_update++;
1905 				(void)hast_activemap_flush(res);
1906 			} else {
1907 				mtx_unlock(&res->hr_amp_lock);
1908 			}
1909 		}
1910 		if (ggio->gctl_cmd == BIO_WRITE) {
1911 			/*
1912 			 * Unlock range we locked.
1913 			 */
1914 			mtx_lock(&range_lock);
1915 			rangelock_del(range_regular, ggio->gctl_offset,
1916 			    ggio->gctl_length);
1917 			if (range_sync_wait)
1918 				cv_signal(&range_sync_cond);
1919 			mtx_unlock(&range_lock);
1920 			if (!hio->hio_done)
1921 				write_complete(res, hio);
1922 		} else {
1923 			if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1) {
1924 				primary_exit(EX_OSERR,
1925 				    "G_GATE_CMD_DONE failed");
1926 			}
1927 		}
1928 		if (hio->hio_errors[0]) {
1929 			switch (ggio->gctl_cmd) {
1930 			case BIO_READ:
1931 				res->hr_stat_read_error++;
1932 				break;
1933 			case BIO_WRITE:
1934 				res->hr_stat_write_error++;
1935 				break;
1936 			case BIO_DELETE:
1937 				res->hr_stat_delete_error++;
1938 				break;
1939 			case BIO_FLUSH:
1940 				res->hr_stat_flush_error++;
1941 				break;
1942 			}
1943 		}
1944 		pjdlog_debug(2,
1945 		    "ggate_send: (%p) Moving request to the free queue.", hio);
1946 		QUEUE_INSERT2(hio, free);
1947 	}
1948 	/* NOTREACHED */
1949 	return (NULL);
1950 }
1951 
1952 /*
1953  * Thread synchronize local and remote components.
1954  */
1955 static void *
1956 sync_thread(void *arg __unused)
1957 {
1958 	struct hast_resource *res = arg;
1959 	struct hio *hio;
1960 	struct g_gate_ctl_io *ggio;
1961 	struct timeval tstart, tend, tdiff;
1962 	unsigned int ii, ncomp, ncomps;
1963 	off_t offset, length, synced;
1964 	bool dorewind, directreads;
1965 	int syncext;
1966 
1967 	ncomps = HAST_NCOMPONENTS;
1968 	dorewind = true;
1969 	synced = 0;
1970 	offset = -1;
1971 	directreads = false;
1972 
1973 	for (;;) {
1974 		mtx_lock(&sync_lock);
1975 		if (offset >= 0 && !sync_inprogress) {
1976 			gettimeofday(&tend, NULL);
1977 			timersub(&tend, &tstart, &tdiff);
1978 			pjdlog_info("Synchronization interrupted after %#.0T. "
1979 			    "%NB synchronized so far.", &tdiff,
1980 			    (intmax_t)synced);
1981 			event_send(res, EVENT_SYNCINTR);
1982 		}
1983 		while (!sync_inprogress) {
1984 			dorewind = true;
1985 			synced = 0;
1986 			cv_wait(&sync_cond, &sync_lock);
1987 		}
1988 		mtx_unlock(&sync_lock);
1989 		/*
1990 		 * Obtain offset at which we should synchronize.
1991 		 * Rewind synchronization if needed.
1992 		 */
1993 		mtx_lock(&res->hr_amp_lock);
1994 		if (dorewind)
1995 			activemap_sync_rewind(res->hr_amp);
1996 		offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1997 		if (syncext != -1) {
1998 			/*
1999 			 * We synchronized entire syncext extent, we can mark
2000 			 * it as clean now.
2001 			 */
2002 			if (activemap_extent_complete(res->hr_amp, syncext))
2003 				(void)hast_activemap_flush(res);
2004 			else
2005 				mtx_unlock(&res->hr_amp_lock);
2006 		} else {
2007 			mtx_unlock(&res->hr_amp_lock);
2008 		}
2009 		if (dorewind) {
2010 			dorewind = false;
2011 			if (offset == -1)
2012 				pjdlog_info("Nodes are in sync.");
2013 			else {
2014 				pjdlog_info("Synchronization started. %NB to go.",
2015 				    (intmax_t)(res->hr_extentsize *
2016 				    activemap_ndirty(res->hr_amp)));
2017 				event_send(res, EVENT_SYNCSTART);
2018 				gettimeofday(&tstart, NULL);
2019 			}
2020 		}
2021 		if (offset == -1) {
2022 			sync_stop();
2023 			pjdlog_debug(1, "Nothing to synchronize.");
2024 			/*
2025 			 * Synchronization complete, make both localcnt and
2026 			 * remotecnt equal.
2027 			 */
2028 			ncomp = 1;
2029 			rw_rlock(&hio_remote_lock[ncomp]);
2030 			if (ISCONNECTED(res, ncomp)) {
2031 				if (synced > 0) {
2032 					int64_t bps;
2033 
2034 					gettimeofday(&tend, NULL);
2035 					timersub(&tend, &tstart, &tdiff);
2036 					bps = (int64_t)((double)synced /
2037 					    ((double)tdiff.tv_sec +
2038 					    (double)tdiff.tv_usec / 1000000));
2039 					pjdlog_info("Synchronization complete. "
2040 					    "%NB synchronized in %#.0lT (%NB/sec).",
2041 					    (intmax_t)synced, &tdiff,
2042 					    (intmax_t)bps);
2043 					event_send(res, EVENT_SYNCDONE);
2044 				}
2045 				mtx_lock(&metadata_lock);
2046 				if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
2047 					directreads = true;
2048 				res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
2049 				res->hr_primary_localcnt =
2050 				    res->hr_secondary_remotecnt;
2051 				res->hr_primary_remotecnt =
2052 				    res->hr_secondary_localcnt;
2053 				pjdlog_debug(1,
2054 				    "Setting localcnt to %ju and remotecnt to %ju.",
2055 				    (uintmax_t)res->hr_primary_localcnt,
2056 				    (uintmax_t)res->hr_primary_remotecnt);
2057 				(void)metadata_write(res);
2058 				mtx_unlock(&metadata_lock);
2059 			}
2060 			rw_unlock(&hio_remote_lock[ncomp]);
2061 			if (directreads) {
2062 				directreads = false;
2063 				enable_direct_reads(res);
2064 			}
2065 			continue;
2066 		}
2067 		pjdlog_debug(2, "sync: Taking free request.");
2068 		QUEUE_TAKE2(hio, free);
2069 		pjdlog_debug(2, "sync: (%p) Got free request.", hio);
2070 		/*
2071 		 * Lock the range we are going to synchronize. We don't want
2072 		 * race where someone writes between our read and write.
2073 		 */
2074 		for (;;) {
2075 			mtx_lock(&range_lock);
2076 			if (rangelock_islocked(range_regular, offset, length)) {
2077 				pjdlog_debug(2,
2078 				    "sync: Range offset=%jd length=%jd locked.",
2079 				    (intmax_t)offset, (intmax_t)length);
2080 				range_sync_wait = true;
2081 				cv_wait(&range_sync_cond, &range_lock);
2082 				range_sync_wait = false;
2083 				mtx_unlock(&range_lock);
2084 				continue;
2085 			}
2086 			if (rangelock_add(range_sync, offset, length) == -1) {
2087 				mtx_unlock(&range_lock);
2088 				pjdlog_debug(2,
2089 				    "sync: Range offset=%jd length=%jd is already locked, waiting.",
2090 				    (intmax_t)offset, (intmax_t)length);
2091 				sleep(1);
2092 				continue;
2093 			}
2094 			mtx_unlock(&range_lock);
2095 			break;
2096 		}
2097 		/*
2098 		 * First read the data from synchronization source.
2099 		 */
2100 		SYNCREQ(hio);
2101 		ggio = &hio->hio_ggio;
2102 		ggio->gctl_cmd = BIO_READ;
2103 		ggio->gctl_offset = offset;
2104 		ggio->gctl_length = length;
2105 		ggio->gctl_error = 0;
2106 		hio->hio_done = false;
2107 		hio->hio_replication = res->hr_replication;
2108 		for (ii = 0; ii < ncomps; ii++)
2109 			hio->hio_errors[ii] = EINVAL;
2110 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2111 		    hio);
2112 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2113 		    hio);
2114 		mtx_lock(&metadata_lock);
2115 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2116 			/*
2117 			 * This range is up-to-date on local component,
2118 			 * so handle request locally.
2119 			 */
2120 			 /* Local component is 0 for now. */
2121 			ncomp = 0;
2122 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2123 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2124 			/*
2125 			 * This range is out-of-date on local component,
2126 			 * so send request to the remote node.
2127 			 */
2128 			 /* Remote component is 1 for now. */
2129 			ncomp = 1;
2130 		}
2131 		mtx_unlock(&metadata_lock);
2132 		refcnt_init(&hio->hio_countdown, 1);
2133 		QUEUE_INSERT1(hio, send, ncomp);
2134 
2135 		/*
2136 		 * Let's wait for READ to finish.
2137 		 */
2138 		mtx_lock(&sync_lock);
2139 		while (!ISSYNCREQDONE(hio))
2140 			cv_wait(&sync_cond, &sync_lock);
2141 		mtx_unlock(&sync_lock);
2142 
2143 		if (hio->hio_errors[ncomp] != 0) {
2144 			pjdlog_error("Unable to read synchronization data: %s.",
2145 			    strerror(hio->hio_errors[ncomp]));
2146 			goto free_queue;
2147 		}
2148 
2149 		/*
2150 		 * We read the data from synchronization source, now write it
2151 		 * to synchronization target.
2152 		 */
2153 		SYNCREQ(hio);
2154 		ggio->gctl_cmd = BIO_WRITE;
2155 		for (ii = 0; ii < ncomps; ii++)
2156 			hio->hio_errors[ii] = EINVAL;
2157 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2158 		    hio);
2159 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2160 		    hio);
2161 		mtx_lock(&metadata_lock);
2162 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2163 			/*
2164 			 * This range is up-to-date on local component,
2165 			 * so we update remote component.
2166 			 */
2167 			 /* Remote component is 1 for now. */
2168 			ncomp = 1;
2169 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2170 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2171 			/*
2172 			 * This range is out-of-date on local component,
2173 			 * so we update it.
2174 			 */
2175 			 /* Local component is 0 for now. */
2176 			ncomp = 0;
2177 		}
2178 		mtx_unlock(&metadata_lock);
2179 
2180 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2181 		    hio);
2182 		refcnt_init(&hio->hio_countdown, 1);
2183 		QUEUE_INSERT1(hio, send, ncomp);
2184 
2185 		/*
2186 		 * Let's wait for WRITE to finish.
2187 		 */
2188 		mtx_lock(&sync_lock);
2189 		while (!ISSYNCREQDONE(hio))
2190 			cv_wait(&sync_cond, &sync_lock);
2191 		mtx_unlock(&sync_lock);
2192 
2193 		if (hio->hio_errors[ncomp] != 0) {
2194 			pjdlog_error("Unable to write synchronization data: %s.",
2195 			    strerror(hio->hio_errors[ncomp]));
2196 			goto free_queue;
2197 		}
2198 
2199 		synced += length;
2200 free_queue:
2201 		mtx_lock(&range_lock);
2202 		rangelock_del(range_sync, offset, length);
2203 		if (range_regular_wait)
2204 			cv_signal(&range_regular_cond);
2205 		mtx_unlock(&range_lock);
2206 		pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
2207 		    hio);
2208 		QUEUE_INSERT2(hio, free);
2209 	}
2210 	/* NOTREACHED */
2211 	return (NULL);
2212 }
2213 
2214 void
2215 primary_config_reload(struct hast_resource *res, struct nv *nv)
2216 {
2217 	unsigned int ii, ncomps;
2218 	int modified, vint;
2219 	const char *vstr;
2220 
2221 	pjdlog_info("Reloading configuration...");
2222 
2223 	PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
2224 	PJDLOG_ASSERT(gres == res);
2225 	nv_assert(nv, "remoteaddr");
2226 	nv_assert(nv, "sourceaddr");
2227 	nv_assert(nv, "replication");
2228 	nv_assert(nv, "checksum");
2229 	nv_assert(nv, "compression");
2230 	nv_assert(nv, "timeout");
2231 	nv_assert(nv, "exec");
2232 	nv_assert(nv, "metaflush");
2233 
2234 	ncomps = HAST_NCOMPONENTS;
2235 
2236 #define MODIFIED_REMOTEADDR	0x01
2237 #define MODIFIED_SOURCEADDR	0x02
2238 #define MODIFIED_REPLICATION	0x04
2239 #define MODIFIED_CHECKSUM	0x08
2240 #define MODIFIED_COMPRESSION	0x10
2241 #define MODIFIED_TIMEOUT	0x20
2242 #define MODIFIED_EXEC		0x40
2243 #define MODIFIED_METAFLUSH	0x80
2244 	modified = 0;
2245 
2246 	vstr = nv_get_string(nv, "remoteaddr");
2247 	if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
2248 		/*
2249 		 * Don't copy res->hr_remoteaddr to gres just yet.
2250 		 * We want remote_close() to log disconnect from the old
2251 		 * addresses, not from the new ones.
2252 		 */
2253 		modified |= MODIFIED_REMOTEADDR;
2254 	}
2255 	vstr = nv_get_string(nv, "sourceaddr");
2256 	if (strcmp(gres->hr_sourceaddr, vstr) != 0) {
2257 		strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr));
2258 		modified |= MODIFIED_SOURCEADDR;
2259 	}
2260 	vint = nv_get_int32(nv, "replication");
2261 	if (gres->hr_replication != vint) {
2262 		gres->hr_replication = vint;
2263 		modified |= MODIFIED_REPLICATION;
2264 	}
2265 	vint = nv_get_int32(nv, "checksum");
2266 	if (gres->hr_checksum != vint) {
2267 		gres->hr_checksum = vint;
2268 		modified |= MODIFIED_CHECKSUM;
2269 	}
2270 	vint = nv_get_int32(nv, "compression");
2271 	if (gres->hr_compression != vint) {
2272 		gres->hr_compression = vint;
2273 		modified |= MODIFIED_COMPRESSION;
2274 	}
2275 	vint = nv_get_int32(nv, "timeout");
2276 	if (gres->hr_timeout != vint) {
2277 		gres->hr_timeout = vint;
2278 		modified |= MODIFIED_TIMEOUT;
2279 	}
2280 	vstr = nv_get_string(nv, "exec");
2281 	if (strcmp(gres->hr_exec, vstr) != 0) {
2282 		strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
2283 		modified |= MODIFIED_EXEC;
2284 	}
2285 	vint = nv_get_int32(nv, "metaflush");
2286 	if (gres->hr_metaflush != vint) {
2287 		gres->hr_metaflush = vint;
2288 		modified |= MODIFIED_METAFLUSH;
2289 	}
2290 
2291 	/*
2292 	 * Change timeout for connected sockets.
2293 	 * Don't bother if we need to reconnect.
2294 	 */
2295 	if ((modified & MODIFIED_TIMEOUT) != 0 &&
2296 	    (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) == 0) {
2297 		for (ii = 0; ii < ncomps; ii++) {
2298 			if (!ISREMOTE(ii))
2299 				continue;
2300 			rw_rlock(&hio_remote_lock[ii]);
2301 			if (!ISCONNECTED(gres, ii)) {
2302 				rw_unlock(&hio_remote_lock[ii]);
2303 				continue;
2304 			}
2305 			rw_unlock(&hio_remote_lock[ii]);
2306 			if (proto_timeout(gres->hr_remotein,
2307 			    gres->hr_timeout) == -1) {
2308 				pjdlog_errno(LOG_WARNING,
2309 				    "Unable to set connection timeout");
2310 			}
2311 			if (proto_timeout(gres->hr_remoteout,
2312 			    gres->hr_timeout) == -1) {
2313 				pjdlog_errno(LOG_WARNING,
2314 				    "Unable to set connection timeout");
2315 			}
2316 		}
2317 	}
2318 	if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) != 0) {
2319 		for (ii = 0; ii < ncomps; ii++) {
2320 			if (!ISREMOTE(ii))
2321 				continue;
2322 			remote_close(gres, ii);
2323 		}
2324 		if (modified & MODIFIED_REMOTEADDR) {
2325 			vstr = nv_get_string(nv, "remoteaddr");
2326 			strlcpy(gres->hr_remoteaddr, vstr,
2327 			    sizeof(gres->hr_remoteaddr));
2328 		}
2329 	}
2330 #undef	MODIFIED_REMOTEADDR
2331 #undef	MODIFIED_SOURCEADDR
2332 #undef	MODIFIED_REPLICATION
2333 #undef	MODIFIED_CHECKSUM
2334 #undef	MODIFIED_COMPRESSION
2335 #undef	MODIFIED_TIMEOUT
2336 #undef	MODIFIED_EXEC
2337 #undef	MODIFIED_METAFLUSH
2338 
2339 	pjdlog_info("Configuration reloaded successfully.");
2340 }
2341 
2342 static void
2343 guard_one(struct hast_resource *res, unsigned int ncomp)
2344 {
2345 	struct proto_conn *in, *out;
2346 
2347 	if (!ISREMOTE(ncomp))
2348 		return;
2349 
2350 	rw_rlock(&hio_remote_lock[ncomp]);
2351 
2352 	if (!real_remote(res)) {
2353 		rw_unlock(&hio_remote_lock[ncomp]);
2354 		return;
2355 	}
2356 
2357 	if (ISCONNECTED(res, ncomp)) {
2358 		PJDLOG_ASSERT(res->hr_remotein != NULL);
2359 		PJDLOG_ASSERT(res->hr_remoteout != NULL);
2360 		rw_unlock(&hio_remote_lock[ncomp]);
2361 		pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2362 		    res->hr_remoteaddr);
2363 		return;
2364 	}
2365 
2366 	PJDLOG_ASSERT(res->hr_remotein == NULL);
2367 	PJDLOG_ASSERT(res->hr_remoteout == NULL);
2368 	/*
2369 	 * Upgrade the lock. It doesn't have to be atomic as no other thread
2370 	 * can change connection status from disconnected to connected.
2371 	 */
2372 	rw_unlock(&hio_remote_lock[ncomp]);
2373 	pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2374 	    res->hr_remoteaddr);
2375 	in = out = NULL;
2376 	if (init_remote(res, &in, &out) == 0) {
2377 		rw_wlock(&hio_remote_lock[ncomp]);
2378 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2379 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2380 		PJDLOG_ASSERT(in != NULL && out != NULL);
2381 		res->hr_remotein = in;
2382 		res->hr_remoteout = out;
2383 		rw_unlock(&hio_remote_lock[ncomp]);
2384 		pjdlog_info("Successfully reconnected to %s.",
2385 		    res->hr_remoteaddr);
2386 		sync_start();
2387 	} else {
2388 		/* Both connections should be NULL. */
2389 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2390 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2391 		PJDLOG_ASSERT(in == NULL && out == NULL);
2392 		pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2393 		    res->hr_remoteaddr);
2394 	}
2395 }
2396 
2397 /*
2398  * Thread guards remote connections and reconnects when needed, handles
2399  * signals, etc.
2400  */
2401 static void *
2402 guard_thread(void *arg)
2403 {
2404 	struct hast_resource *res = arg;
2405 	unsigned int ii, ncomps;
2406 	struct timespec timeout;
2407 	time_t lastcheck, now;
2408 	sigset_t mask;
2409 	int signo;
2410 
2411 	ncomps = HAST_NCOMPONENTS;
2412 	lastcheck = time(NULL);
2413 
2414 	PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2415 	PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2416 	PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2417 
2418 	timeout.tv_sec = HAST_KEEPALIVE;
2419 	timeout.tv_nsec = 0;
2420 	signo = -1;
2421 
2422 	for (;;) {
2423 		switch (signo) {
2424 		case SIGINT:
2425 		case SIGTERM:
2426 			sigexit_received = true;
2427 			primary_exitx(EX_OK,
2428 			    "Termination signal received, exiting.");
2429 			break;
2430 		default:
2431 			break;
2432 		}
2433 
2434 		/*
2435 		 * Don't check connections until we fully started,
2436 		 * as we may still be looping, waiting for remote node
2437 		 * to switch from primary to secondary.
2438 		 */
2439 		if (fullystarted) {
2440 			pjdlog_debug(2, "remote_guard: Checking connections.");
2441 			now = time(NULL);
2442 			if (lastcheck + HAST_KEEPALIVE <= now) {
2443 				for (ii = 0; ii < ncomps; ii++)
2444 					guard_one(res, ii);
2445 				lastcheck = now;
2446 			}
2447 		}
2448 		signo = sigtimedwait(&mask, NULL, &timeout);
2449 	}
2450 	/* NOTREACHED */
2451 	return (NULL);
2452 }
2453