xref: /freebsd/sbin/hastd/primary.c (revision aa339f1d5df9e38f36a34eb522355c4eebcae6c4)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2009 The FreeBSD Foundation
5  * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net>
6  * All rights reserved.
7  *
8  * This software was developed by Pawel Jakub Dawidek under sponsorship from
9  * the FreeBSD Foundation.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/types.h>
34 #include <sys/time.h>
35 #include <sys/bio.h>
36 #include <sys/disk.h>
37 #include <sys/stat.h>
38 
39 #include <geom/gate/g_gate.h>
40 
41 #include <err.h>
42 #include <errno.h>
43 #include <fcntl.h>
44 #include <libgeom.h>
45 #include <pthread.h>
46 #include <signal.h>
47 #include <stdint.h>
48 #include <stdio.h>
49 #include <string.h>
50 #include <sysexits.h>
51 #include <unistd.h>
52 
53 #include <activemap.h>
54 #include <nv.h>
55 #include <rangelock.h>
56 
57 #include "control.h"
58 #include "event.h"
59 #include "hast.h"
60 #include "hast_proto.h"
61 #include "hastd.h"
62 #include "hooks.h"
63 #include "metadata.h"
64 #include "proto.h"
65 #include "pjdlog.h"
66 #include "refcnt.h"
67 #include "subr.h"
68 #include "synch.h"
69 
70 /* The is only one remote component for now. */
71 #define	ISREMOTE(no)	((no) == 1)
72 
73 struct hio {
74 	/*
75 	 * Number of components we are still waiting for.
76 	 * When this field goes to 0, we can send the request back to the
77 	 * kernel. Each component has to decrease this counter by one
78 	 * even on failure.
79 	 */
80 	refcnt_t		 hio_countdown;
81 	/*
82 	 * Each component has a place to store its own error.
83 	 * Once the request is handled by all components we can decide if the
84 	 * request overall is successful or not.
85 	 */
86 	int			*hio_errors;
87 	/*
88 	 * Structure used to communicate with GEOM Gate class.
89 	 */
90 	struct g_gate_ctl_io	 hio_ggio;
91 	/*
92 	 * Request was already confirmed to GEOM Gate.
93 	 */
94 	bool			 hio_done;
95 	/*
96 	 * Number of components we are still waiting before sending write
97 	 * completion ack to GEOM Gate. Used for memsync.
98 	 */
99 	refcnt_t		 hio_writecount;
100 	/*
101 	 * Memsync request was acknowledged by remote.
102 	 */
103 	bool			 hio_memsyncacked;
104 	/*
105 	 * Remember replication from the time the request was initiated,
106 	 * so we won't get confused when replication changes on reload.
107 	 */
108 	int			 hio_replication;
109 	TAILQ_ENTRY(hio)	*hio_next;
110 };
111 #define	hio_free_next	hio_next[0]
112 #define	hio_done_next	hio_next[0]
113 
114 /*
115  * Free list holds unused structures. When free list is empty, we have to wait
116  * until some in-progress requests are freed.
117  */
118 static TAILQ_HEAD(, hio) hio_free_list;
119 static size_t hio_free_list_size;
120 static pthread_mutex_t hio_free_list_lock;
121 static pthread_cond_t hio_free_list_cond;
122 /*
123  * There is one send list for every component. One requests is placed on all
124  * send lists - each component gets the same request, but each component is
125  * responsible for managing his own send list.
126  */
127 static TAILQ_HEAD(, hio) *hio_send_list;
128 static size_t *hio_send_list_size;
129 static pthread_mutex_t *hio_send_list_lock;
130 static pthread_cond_t *hio_send_list_cond;
131 #define	hio_send_local_list_size	hio_send_list_size[0]
132 #define	hio_send_remote_list_size	hio_send_list_size[1]
133 /*
134  * There is one recv list for every component, although local components don't
135  * use recv lists as local requests are done synchronously.
136  */
137 static TAILQ_HEAD(, hio) *hio_recv_list;
138 static size_t *hio_recv_list_size;
139 static pthread_mutex_t *hio_recv_list_lock;
140 static pthread_cond_t *hio_recv_list_cond;
141 #define	hio_recv_remote_list_size	hio_recv_list_size[1]
142 /*
143  * Request is placed on done list by the slowest component (the one that
144  * decreased hio_countdown from 1 to 0).
145  */
146 static TAILQ_HEAD(, hio) hio_done_list;
147 static size_t hio_done_list_size;
148 static pthread_mutex_t hio_done_list_lock;
149 static pthread_cond_t hio_done_list_cond;
150 /*
151  * Structure below are for interaction with sync thread.
152  */
153 static bool sync_inprogress;
154 static pthread_mutex_t sync_lock;
155 static pthread_cond_t sync_cond;
156 /*
157  * The lock below allows to synchornize access to remote connections.
158  */
159 static pthread_rwlock_t *hio_remote_lock;
160 
161 /*
162  * Lock to synchronize metadata updates. Also synchronize access to
163  * hr_primary_localcnt and hr_primary_remotecnt fields.
164  */
165 static pthread_mutex_t metadata_lock;
166 
167 /*
168  * Maximum number of outstanding I/O requests.
169  */
170 #define	HAST_HIO_MAX	256
171 /*
172  * Number of components. At this point there are only two components: local
173  * and remote, but in the future it might be possible to use multiple local
174  * and remote components.
175  */
176 #define	HAST_NCOMPONENTS	2
177 
178 #define	ISCONNECTED(res, no)	\
179 	((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
180 
181 #define	QUEUE_INSERT1(hio, name, ncomp)	do {				\
182 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
183 	if (TAILQ_EMPTY(&hio_##name##_list[(ncomp)]))			\
184 		cv_broadcast(&hio_##name##_list_cond[(ncomp)]);		\
185 	TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),		\
186 	    hio_next[(ncomp)]);						\
187 	hio_##name##_list_size[(ncomp)]++;				\
188 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
189 } while (0)
190 #define	QUEUE_INSERT2(hio, name)	do {				\
191 	mtx_lock(&hio_##name##_list_lock);				\
192 	if (TAILQ_EMPTY(&hio_##name##_list))				\
193 		cv_broadcast(&hio_##name##_list_cond);			\
194 	TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
195 	hio_##name##_list_size++;					\
196 	mtx_unlock(&hio_##name##_list_lock);				\
197 } while (0)
198 #define	QUEUE_TAKE1(hio, name, ncomp, timeout)	do {			\
199 	bool _last;							\
200 									\
201 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
202 	_last = false;							\
203 	while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
204 		cv_timedwait(&hio_##name##_list_cond[(ncomp)],		\
205 		    &hio_##name##_list_lock[(ncomp)], (timeout));	\
206 		if ((timeout) != 0)					\
207 			_last = true;					\
208 	}								\
209 	if (hio != NULL) {						\
210 		PJDLOG_ASSERT(hio_##name##_list_size[(ncomp)] != 0);	\
211 		hio_##name##_list_size[(ncomp)]--;			\
212 		TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),	\
213 		    hio_next[(ncomp)]);					\
214 	}								\
215 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
216 } while (0)
217 #define	QUEUE_TAKE2(hio, name)	do {					\
218 	mtx_lock(&hio_##name##_list_lock);				\
219 	while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {	\
220 		cv_wait(&hio_##name##_list_cond,			\
221 		    &hio_##name##_list_lock);				\
222 	}								\
223 	PJDLOG_ASSERT(hio_##name##_list_size != 0);			\
224 	hio_##name##_list_size--;					\
225 	TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);	\
226 	mtx_unlock(&hio_##name##_list_lock);				\
227 } while (0)
228 
229 #define ISFULLSYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_FULLSYNC)
230 #define ISMEMSYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_MEMSYNC)
231 #define ISASYNC(hio)	((hio)->hio_replication == HAST_REPLICATION_ASYNC)
232 
233 #define	SYNCREQ(hio)		do {					\
234 	(hio)->hio_ggio.gctl_unit = -1;					\
235 	(hio)->hio_ggio.gctl_seq = 1;					\
236 } while (0)
237 #define	ISSYNCREQ(hio)		((hio)->hio_ggio.gctl_unit == -1)
238 #define	SYNCREQDONE(hio)	do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
239 #define	ISSYNCREQDONE(hio)	((hio)->hio_ggio.gctl_unit == -2)
240 
241 #define ISMEMSYNCWRITE(hio)	(ISMEMSYNC(hio) &&			\
242 	    (hio)->hio_ggio.gctl_cmd == BIO_WRITE && !ISSYNCREQ(hio))
243 
244 static struct hast_resource *gres;
245 
246 static pthread_mutex_t range_lock;
247 static struct rangelocks *range_regular;
248 static bool range_regular_wait;
249 static pthread_cond_t range_regular_cond;
250 static struct rangelocks *range_sync;
251 static bool range_sync_wait;
252 static pthread_cond_t range_sync_cond;
253 static bool fullystarted;
254 
255 static void *ggate_recv_thread(void *arg);
256 static void *local_send_thread(void *arg);
257 static void *remote_send_thread(void *arg);
258 static void *remote_recv_thread(void *arg);
259 static void *ggate_send_thread(void *arg);
260 static void *sync_thread(void *arg);
261 static void *guard_thread(void *arg);
262 
263 static void
264 output_status_aux(struct nv *nvout)
265 {
266 
267 	nv_add_uint64(nvout, (uint64_t)hio_free_list_size,
268 	    "idle_queue_size");
269 	nv_add_uint64(nvout, (uint64_t)hio_send_local_list_size,
270 	    "local_queue_size");
271 	nv_add_uint64(nvout, (uint64_t)hio_send_remote_list_size,
272 	    "send_queue_size");
273 	nv_add_uint64(nvout, (uint64_t)hio_recv_remote_list_size,
274 	    "recv_queue_size");
275 	nv_add_uint64(nvout, (uint64_t)hio_done_list_size,
276 	    "done_queue_size");
277 }
278 
279 static void
280 cleanup(struct hast_resource *res)
281 {
282 	int rerrno;
283 
284 	/* Remember errno. */
285 	rerrno = errno;
286 
287 	/* Destroy ggate provider if we created one. */
288 	if (res->hr_ggateunit >= 0) {
289 		struct g_gate_ctl_destroy ggiod;
290 
291 		bzero(&ggiod, sizeof(ggiod));
292 		ggiod.gctl_version = G_GATE_VERSION;
293 		ggiod.gctl_unit = res->hr_ggateunit;
294 		ggiod.gctl_force = 1;
295 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) == -1) {
296 			pjdlog_errno(LOG_WARNING,
297 			    "Unable to destroy hast/%s device",
298 			    res->hr_provname);
299 		}
300 		res->hr_ggateunit = -1;
301 	}
302 
303 	/* Restore errno. */
304 	errno = rerrno;
305 }
306 
307 static __dead2 void
308 primary_exit(int exitcode, const char *fmt, ...)
309 {
310 	va_list ap;
311 
312 	PJDLOG_ASSERT(exitcode != EX_OK);
313 	va_start(ap, fmt);
314 	pjdlogv_errno(LOG_ERR, fmt, ap);
315 	va_end(ap);
316 	cleanup(gres);
317 	exit(exitcode);
318 }
319 
320 static __dead2 void
321 primary_exitx(int exitcode, const char *fmt, ...)
322 {
323 	va_list ap;
324 
325 	va_start(ap, fmt);
326 	pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
327 	va_end(ap);
328 	cleanup(gres);
329 	exit(exitcode);
330 }
331 
332 static int
333 hast_activemap_flush(struct hast_resource *res) __unlocks(res->hr_amp_lock)
334 {
335 	const unsigned char *buf;
336 	size_t size;
337 	int ret;
338 
339 	mtx_lock(&res->hr_amp_diskmap_lock);
340 	buf = activemap_bitmap(res->hr_amp, &size);
341 	mtx_unlock(&res->hr_amp_lock);
342 	PJDLOG_ASSERT(buf != NULL);
343 	PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
344 	ret = 0;
345 	if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
346 	    (ssize_t)size) {
347 		pjdlog_errno(LOG_ERR, "Unable to flush activemap to disk");
348 		res->hr_stat_activemap_write_error++;
349 		ret = -1;
350 	}
351 	if (ret == 0 && res->hr_metaflush == 1 &&
352 	    g_flush(res->hr_localfd) == -1) {
353 		if (errno == EOPNOTSUPP) {
354 			pjdlog_warning("The %s provider doesn't support flushing write cache. Disabling it.",
355 			    res->hr_localpath);
356 			res->hr_metaflush = 0;
357 		} else {
358 			pjdlog_errno(LOG_ERR,
359 			    "Unable to flush disk cache on activemap update");
360 			res->hr_stat_activemap_flush_error++;
361 			ret = -1;
362 		}
363 	}
364 	mtx_unlock(&res->hr_amp_diskmap_lock);
365 	return (ret);
366 }
367 
368 static bool
369 real_remote(const struct hast_resource *res)
370 {
371 
372 	return (strcmp(res->hr_remoteaddr, "none") != 0);
373 }
374 
375 static void
376 init_environment(struct hast_resource *res __unused)
377 {
378 	struct hio *hio;
379 	unsigned int ii, ncomps;
380 
381 	/*
382 	 * In the future it might be per-resource value.
383 	 */
384 	ncomps = HAST_NCOMPONENTS;
385 
386 	/*
387 	 * Allocate memory needed by lists.
388 	 */
389 	hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
390 	if (hio_send_list == NULL) {
391 		primary_exitx(EX_TEMPFAIL,
392 		    "Unable to allocate %zu bytes of memory for send lists.",
393 		    sizeof(hio_send_list[0]) * ncomps);
394 	}
395 	hio_send_list_size = malloc(sizeof(hio_send_list_size[0]) * ncomps);
396 	if (hio_send_list_size == NULL) {
397 		primary_exitx(EX_TEMPFAIL,
398 		    "Unable to allocate %zu bytes of memory for send list counters.",
399 		    sizeof(hio_send_list_size[0]) * ncomps);
400 	}
401 	hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
402 	if (hio_send_list_lock == NULL) {
403 		primary_exitx(EX_TEMPFAIL,
404 		    "Unable to allocate %zu bytes of memory for send list locks.",
405 		    sizeof(hio_send_list_lock[0]) * ncomps);
406 	}
407 	hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
408 	if (hio_send_list_cond == NULL) {
409 		primary_exitx(EX_TEMPFAIL,
410 		    "Unable to allocate %zu bytes of memory for send list condition variables.",
411 		    sizeof(hio_send_list_cond[0]) * ncomps);
412 	}
413 	hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
414 	if (hio_recv_list == NULL) {
415 		primary_exitx(EX_TEMPFAIL,
416 		    "Unable to allocate %zu bytes of memory for recv lists.",
417 		    sizeof(hio_recv_list[0]) * ncomps);
418 	}
419 	hio_recv_list_size = malloc(sizeof(hio_recv_list_size[0]) * ncomps);
420 	if (hio_recv_list_size == NULL) {
421 		primary_exitx(EX_TEMPFAIL,
422 		    "Unable to allocate %zu bytes of memory for recv list counters.",
423 		    sizeof(hio_recv_list_size[0]) * ncomps);
424 	}
425 	hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
426 	if (hio_recv_list_lock == NULL) {
427 		primary_exitx(EX_TEMPFAIL,
428 		    "Unable to allocate %zu bytes of memory for recv list locks.",
429 		    sizeof(hio_recv_list_lock[0]) * ncomps);
430 	}
431 	hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
432 	if (hio_recv_list_cond == NULL) {
433 		primary_exitx(EX_TEMPFAIL,
434 		    "Unable to allocate %zu bytes of memory for recv list condition variables.",
435 		    sizeof(hio_recv_list_cond[0]) * ncomps);
436 	}
437 	hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
438 	if (hio_remote_lock == NULL) {
439 		primary_exitx(EX_TEMPFAIL,
440 		    "Unable to allocate %zu bytes of memory for remote connections locks.",
441 		    sizeof(hio_remote_lock[0]) * ncomps);
442 	}
443 
444 	/*
445 	 * Initialize lists, their counters, locks and condition variables.
446 	 */
447 	TAILQ_INIT(&hio_free_list);
448 	mtx_init(&hio_free_list_lock);
449 	cv_init(&hio_free_list_cond);
450 	for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
451 		TAILQ_INIT(&hio_send_list[ii]);
452 		hio_send_list_size[ii] = 0;
453 		mtx_init(&hio_send_list_lock[ii]);
454 		cv_init(&hio_send_list_cond[ii]);
455 		TAILQ_INIT(&hio_recv_list[ii]);
456 		hio_recv_list_size[ii] = 0;
457 		mtx_init(&hio_recv_list_lock[ii]);
458 		cv_init(&hio_recv_list_cond[ii]);
459 		rw_init(&hio_remote_lock[ii]);
460 	}
461 	TAILQ_INIT(&hio_done_list);
462 	mtx_init(&hio_done_list_lock);
463 	cv_init(&hio_done_list_cond);
464 	mtx_init(&metadata_lock);
465 
466 	/*
467 	 * Allocate requests pool and initialize requests.
468 	 */
469 	for (ii = 0; ii < HAST_HIO_MAX; ii++) {
470 		hio = malloc(sizeof(*hio));
471 		if (hio == NULL) {
472 			primary_exitx(EX_TEMPFAIL,
473 			    "Unable to allocate %zu bytes of memory for hio request.",
474 			    sizeof(*hio));
475 		}
476 		refcnt_init(&hio->hio_countdown, 0);
477 		hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
478 		if (hio->hio_errors == NULL) {
479 			primary_exitx(EX_TEMPFAIL,
480 			    "Unable allocate %zu bytes of memory for hio errors.",
481 			    sizeof(hio->hio_errors[0]) * ncomps);
482 		}
483 		hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
484 		if (hio->hio_next == NULL) {
485 			primary_exitx(EX_TEMPFAIL,
486 			    "Unable allocate %zu bytes of memory for hio_next field.",
487 			    sizeof(hio->hio_next[0]) * ncomps);
488 		}
489 		hio->hio_ggio.gctl_version = G_GATE_VERSION;
490 		hio->hio_ggio.gctl_data = malloc(MAXPHYS);
491 		if (hio->hio_ggio.gctl_data == NULL) {
492 			primary_exitx(EX_TEMPFAIL,
493 			    "Unable to allocate %zu bytes of memory for gctl_data.",
494 			    MAXPHYS);
495 		}
496 		hio->hio_ggio.gctl_length = MAXPHYS;
497 		hio->hio_ggio.gctl_error = 0;
498 		TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
499 		hio_free_list_size++;
500 	}
501 }
502 
503 static bool
504 init_resuid(struct hast_resource *res)
505 {
506 
507 	mtx_lock(&metadata_lock);
508 	if (res->hr_resuid != 0) {
509 		mtx_unlock(&metadata_lock);
510 		return (false);
511 	} else {
512 		/* Initialize unique resource identifier. */
513 		arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
514 		mtx_unlock(&metadata_lock);
515 		if (metadata_write(res) == -1)
516 			exit(EX_NOINPUT);
517 		return (true);
518 	}
519 }
520 
521 static void
522 init_local(struct hast_resource *res)
523 {
524 	unsigned char *buf;
525 	size_t mapsize;
526 
527 	if (metadata_read(res, true) == -1)
528 		exit(EX_NOINPUT);
529 	mtx_init(&res->hr_amp_lock);
530 	if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
531 	    res->hr_local_sectorsize, res->hr_keepdirty) == -1) {
532 		primary_exit(EX_TEMPFAIL, "Unable to create activemap");
533 	}
534 	mtx_init(&range_lock);
535 	cv_init(&range_regular_cond);
536 	if (rangelock_init(&range_regular) == -1)
537 		primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
538 	cv_init(&range_sync_cond);
539 	if (rangelock_init(&range_sync) == -1)
540 		primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
541 	mapsize = activemap_ondisk_size(res->hr_amp);
542 	buf = calloc(1, mapsize);
543 	if (buf == NULL) {
544 		primary_exitx(EX_TEMPFAIL,
545 		    "Unable to allocate buffer for activemap.");
546 	}
547 	if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
548 	    (ssize_t)mapsize) {
549 		primary_exit(EX_NOINPUT, "Unable to read activemap");
550 	}
551 	activemap_copyin(res->hr_amp, buf, mapsize);
552 	free(buf);
553 	if (res->hr_resuid != 0)
554 		return;
555 	/*
556 	 * We're using provider for the first time. Initialize local and remote
557 	 * counters. We don't initialize resuid here, as we want to do it just
558 	 * in time. The reason for this is that we want to inform secondary
559 	 * that there were no writes yet, so there is no need to synchronize
560 	 * anything.
561 	 */
562 	res->hr_primary_localcnt = 0;
563 	res->hr_primary_remotecnt = 0;
564 	if (metadata_write(res) == -1)
565 		exit(EX_NOINPUT);
566 }
567 
568 static int
569 primary_connect(struct hast_resource *res, struct proto_conn **connp)
570 {
571 	struct proto_conn *conn;
572 	int16_t val;
573 
574 	val = 1;
575 	if (proto_send(res->hr_conn, &val, sizeof(val)) == -1) {
576 		primary_exit(EX_TEMPFAIL,
577 		    "Unable to send connection request to parent");
578 	}
579 	if (proto_recv(res->hr_conn, &val, sizeof(val)) == -1) {
580 		primary_exit(EX_TEMPFAIL,
581 		    "Unable to receive reply to connection request from parent");
582 	}
583 	if (val != 0) {
584 		errno = val;
585 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
586 		    res->hr_remoteaddr);
587 		return (-1);
588 	}
589 	if (proto_connection_recv(res->hr_conn, true, &conn) == -1) {
590 		primary_exit(EX_TEMPFAIL,
591 		    "Unable to receive connection from parent");
592 	}
593 	if (proto_connect_wait(conn, res->hr_timeout) == -1) {
594 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
595 		    res->hr_remoteaddr);
596 		proto_close(conn);
597 		return (-1);
598 	}
599 	/* Error in setting timeout is not critical, but why should it fail? */
600 	if (proto_timeout(conn, res->hr_timeout) == -1)
601 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
602 
603 	*connp = conn;
604 
605 	return (0);
606 }
607 
608 /*
609  * Function instructs GEOM_GATE to handle reads directly from within the kernel.
610  */
611 static void
612 enable_direct_reads(struct hast_resource *res)
613 {
614 	struct g_gate_ctl_modify ggiomodify;
615 
616 	bzero(&ggiomodify, sizeof(ggiomodify));
617 	ggiomodify.gctl_version = G_GATE_VERSION;
618 	ggiomodify.gctl_unit = res->hr_ggateunit;
619 	ggiomodify.gctl_modify = GG_MODIFY_READPROV | GG_MODIFY_READOFFSET;
620 	strlcpy(ggiomodify.gctl_readprov, res->hr_localpath,
621 	    sizeof(ggiomodify.gctl_readprov));
622 	ggiomodify.gctl_readoffset = res->hr_localoff;
623 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_MODIFY, &ggiomodify) == 0)
624 		pjdlog_debug(1, "Direct reads enabled.");
625 	else
626 		pjdlog_errno(LOG_WARNING, "Failed to enable direct reads");
627 }
628 
629 static int
630 init_remote(struct hast_resource *res, struct proto_conn **inp,
631     struct proto_conn **outp)
632 {
633 	struct proto_conn *in, *out;
634 	struct nv *nvout, *nvin;
635 	const unsigned char *token;
636 	unsigned char *map;
637 	const char *errmsg;
638 	int32_t extentsize;
639 	int64_t datasize;
640 	uint32_t mapsize;
641 	uint8_t version;
642 	size_t size;
643 	int error;
644 
645 	PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
646 	PJDLOG_ASSERT(real_remote(res));
647 
648 	in = out = NULL;
649 	errmsg = NULL;
650 
651 	if (primary_connect(res, &out) == -1)
652 		return (ECONNREFUSED);
653 
654 	error = ECONNABORTED;
655 
656 	/*
657 	 * First handshake step.
658 	 * Setup outgoing connection with remote node.
659 	 */
660 	nvout = nv_alloc();
661 	nv_add_string(nvout, res->hr_name, "resource");
662 	nv_add_uint8(nvout, HAST_PROTO_VERSION, "version");
663 	if (nv_error(nvout) != 0) {
664 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
665 		    "Unable to allocate header for connection with %s",
666 		    res->hr_remoteaddr);
667 		nv_free(nvout);
668 		goto close;
669 	}
670 	if (hast_proto_send(res, out, nvout, NULL, 0) == -1) {
671 		pjdlog_errno(LOG_WARNING,
672 		    "Unable to send handshake header to %s",
673 		    res->hr_remoteaddr);
674 		nv_free(nvout);
675 		goto close;
676 	}
677 	nv_free(nvout);
678 	if (hast_proto_recv_hdr(out, &nvin) == -1) {
679 		pjdlog_errno(LOG_WARNING,
680 		    "Unable to receive handshake header from %s",
681 		    res->hr_remoteaddr);
682 		goto close;
683 	}
684 	errmsg = nv_get_string(nvin, "errmsg");
685 	if (errmsg != NULL) {
686 		pjdlog_warning("%s", errmsg);
687 		if (nv_exists(nvin, "wait"))
688 			error = EBUSY;
689 		nv_free(nvin);
690 		goto close;
691 	}
692 	version = nv_get_uint8(nvin, "version");
693 	if (version == 0) {
694 		/*
695 		 * If no version is sent, it means this is protocol version 1.
696 		 */
697 		version = 1;
698 	}
699 	if (version > HAST_PROTO_VERSION) {
700 		pjdlog_warning("Invalid version received (%hhu).", version);
701 		nv_free(nvin);
702 		goto close;
703 	}
704 	res->hr_version = version;
705 	pjdlog_debug(1, "Negotiated protocol version %d.", res->hr_version);
706 	token = nv_get_uint8_array(nvin, &size, "token");
707 	if (token == NULL) {
708 		pjdlog_warning("Handshake header from %s has no 'token' field.",
709 		    res->hr_remoteaddr);
710 		nv_free(nvin);
711 		goto close;
712 	}
713 	if (size != sizeof(res->hr_token)) {
714 		pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
715 		    res->hr_remoteaddr, size, sizeof(res->hr_token));
716 		nv_free(nvin);
717 		goto close;
718 	}
719 	bcopy(token, res->hr_token, sizeof(res->hr_token));
720 	nv_free(nvin);
721 
722 	/*
723 	 * Second handshake step.
724 	 * Setup incoming connection with remote node.
725 	 */
726 	if (primary_connect(res, &in) == -1)
727 		goto close;
728 
729 	nvout = nv_alloc();
730 	nv_add_string(nvout, res->hr_name, "resource");
731 	nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
732 	    "token");
733 	if (res->hr_resuid == 0) {
734 		/*
735 		 * The resuid field was not yet initialized.
736 		 * Because we do synchronization inside init_resuid(), it is
737 		 * possible that someone already initialized it, the function
738 		 * will return false then, but if we successfully initialized
739 		 * it, we will get true. True means that there were no writes
740 		 * to this resource yet and we want to inform secondary that
741 		 * synchronization is not needed by sending "virgin" argument.
742 		 */
743 		if (init_resuid(res))
744 			nv_add_int8(nvout, 1, "virgin");
745 	}
746 	nv_add_uint64(nvout, res->hr_resuid, "resuid");
747 	nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
748 	nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
749 	if (nv_error(nvout) != 0) {
750 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
751 		    "Unable to allocate header for connection with %s",
752 		    res->hr_remoteaddr);
753 		nv_free(nvout);
754 		goto close;
755 	}
756 	if (hast_proto_send(res, in, nvout, NULL, 0) == -1) {
757 		pjdlog_errno(LOG_WARNING,
758 		    "Unable to send handshake header to %s",
759 		    res->hr_remoteaddr);
760 		nv_free(nvout);
761 		goto close;
762 	}
763 	nv_free(nvout);
764 	if (hast_proto_recv_hdr(out, &nvin) == -1) {
765 		pjdlog_errno(LOG_WARNING,
766 		    "Unable to receive handshake header from %s",
767 		    res->hr_remoteaddr);
768 		goto close;
769 	}
770 	errmsg = nv_get_string(nvin, "errmsg");
771 	if (errmsg != NULL) {
772 		pjdlog_warning("%s", errmsg);
773 		nv_free(nvin);
774 		goto close;
775 	}
776 	datasize = nv_get_int64(nvin, "datasize");
777 	if (datasize != res->hr_datasize) {
778 		pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
779 		    (intmax_t)res->hr_datasize, (intmax_t)datasize);
780 		nv_free(nvin);
781 		goto close;
782 	}
783 	extentsize = nv_get_int32(nvin, "extentsize");
784 	if (extentsize != res->hr_extentsize) {
785 		pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
786 		    (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
787 		nv_free(nvin);
788 		goto close;
789 	}
790 	res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
791 	res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
792 	res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
793 	if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY)
794 		enable_direct_reads(res);
795 	if (nv_exists(nvin, "virgin")) {
796 		/*
797 		 * Secondary was reinitialized, bump localcnt if it is 0 as
798 		 * only we have the data.
799 		 */
800 		PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY);
801 		PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
802 
803 		if (res->hr_primary_localcnt == 0) {
804 			PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0);
805 
806 			mtx_lock(&metadata_lock);
807 			res->hr_primary_localcnt++;
808 			pjdlog_debug(1, "Increasing localcnt to %ju.",
809 			    (uintmax_t)res->hr_primary_localcnt);
810 			(void)metadata_write(res);
811 			mtx_unlock(&metadata_lock);
812 		}
813 	}
814 	map = NULL;
815 	mapsize = nv_get_uint32(nvin, "mapsize");
816 	if (mapsize > 0) {
817 		map = malloc(mapsize);
818 		if (map == NULL) {
819 			pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
820 			    (uintmax_t)mapsize);
821 			nv_free(nvin);
822 			goto close;
823 		}
824 		/*
825 		 * Remote node have some dirty extents on its own, lets
826 		 * download its activemap.
827 		 */
828 		if (hast_proto_recv_data(res, out, nvin, map,
829 		    mapsize) == -1) {
830 			pjdlog_errno(LOG_ERR,
831 			    "Unable to receive remote activemap");
832 			nv_free(nvin);
833 			free(map);
834 			goto close;
835 		}
836 		mtx_lock(&res->hr_amp_lock);
837 		/*
838 		 * Merge local and remote bitmaps.
839 		 */
840 		activemap_merge(res->hr_amp, map, mapsize);
841 		free(map);
842 		/*
843 		 * Now that we merged bitmaps from both nodes, flush it to the
844 		 * disk before we start to synchronize.
845 		 */
846 		(void)hast_activemap_flush(res);
847 	}
848 	nv_free(nvin);
849 #ifdef notyet
850 	/* Setup directions. */
851 	if (proto_send(out, NULL, 0) == -1)
852 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
853 	if (proto_recv(in, NULL, 0) == -1)
854 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
855 #endif
856 	pjdlog_info("Connected to %s.", res->hr_remoteaddr);
857 	if (res->hr_original_replication == HAST_REPLICATION_MEMSYNC &&
858 	    res->hr_version < 2) {
859 		pjdlog_warning("The 'memsync' replication mode is not supported by the remote node, falling back to 'fullsync' mode.");
860 		res->hr_replication = HAST_REPLICATION_FULLSYNC;
861 	} else if (res->hr_replication != res->hr_original_replication) {
862 		/*
863 		 * This is in case hastd disconnected and was upgraded.
864 		 */
865 		res->hr_replication = res->hr_original_replication;
866 	}
867 	if (inp != NULL && outp != NULL) {
868 		*inp = in;
869 		*outp = out;
870 	} else {
871 		res->hr_remotein = in;
872 		res->hr_remoteout = out;
873 	}
874 	event_send(res, EVENT_CONNECT);
875 	return (0);
876 close:
877 	if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
878 		event_send(res, EVENT_SPLITBRAIN);
879 	proto_close(out);
880 	if (in != NULL)
881 		proto_close(in);
882 	return (error);
883 }
884 
885 static void
886 sync_start(void)
887 {
888 
889 	mtx_lock(&sync_lock);
890 	sync_inprogress = true;
891 	mtx_unlock(&sync_lock);
892 	cv_signal(&sync_cond);
893 }
894 
895 static void
896 sync_stop(void)
897 {
898 
899 	mtx_lock(&sync_lock);
900 	if (sync_inprogress)
901 		sync_inprogress = false;
902 	mtx_unlock(&sync_lock);
903 }
904 
905 static void
906 init_ggate(struct hast_resource *res)
907 {
908 	struct g_gate_ctl_create ggiocreate;
909 	struct g_gate_ctl_cancel ggiocancel;
910 
911 	/*
912 	 * We communicate with ggate via /dev/ggctl. Open it.
913 	 */
914 	res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
915 	if (res->hr_ggatefd == -1)
916 		primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
917 	/*
918 	 * Create provider before trying to connect, as connection failure
919 	 * is not critical, but may take some time.
920 	 */
921 	bzero(&ggiocreate, sizeof(ggiocreate));
922 	ggiocreate.gctl_version = G_GATE_VERSION;
923 	ggiocreate.gctl_mediasize = res->hr_datasize;
924 	ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
925 	ggiocreate.gctl_flags = 0;
926 	ggiocreate.gctl_maxcount = 0;
927 	ggiocreate.gctl_timeout = 0;
928 	ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
929 	snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
930 	    res->hr_provname);
931 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
932 		pjdlog_info("Device hast/%s created.", res->hr_provname);
933 		res->hr_ggateunit = ggiocreate.gctl_unit;
934 		return;
935 	}
936 	if (errno != EEXIST) {
937 		primary_exit(EX_OSERR, "Unable to create hast/%s device",
938 		    res->hr_provname);
939 	}
940 	pjdlog_debug(1,
941 	    "Device hast/%s already exists, we will try to take it over.",
942 	    res->hr_provname);
943 	/*
944 	 * If we received EEXIST, we assume that the process who created the
945 	 * provider died and didn't clean up. In that case we will start from
946 	 * where he left of.
947 	 */
948 	bzero(&ggiocancel, sizeof(ggiocancel));
949 	ggiocancel.gctl_version = G_GATE_VERSION;
950 	ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
951 	snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
952 	    res->hr_provname);
953 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
954 		pjdlog_info("Device hast/%s recovered.", res->hr_provname);
955 		res->hr_ggateunit = ggiocancel.gctl_unit;
956 		return;
957 	}
958 	primary_exit(EX_OSERR, "Unable to take over hast/%s device",
959 	    res->hr_provname);
960 }
961 
962 void
963 hastd_primary(struct hast_resource *res)
964 {
965 	pthread_t td;
966 	pid_t pid;
967 	int error, mode, debuglevel;
968 
969 	/*
970 	 * Create communication channel for sending control commands from
971 	 * parent to child.
972 	 */
973 	if (proto_client(NULL, "socketpair://", &res->hr_ctrl) == -1) {
974 		/* TODO: There's no need for this to be fatal error. */
975 		KEEP_ERRNO((void)pidfile_remove(pfh));
976 		pjdlog_exit(EX_OSERR,
977 		    "Unable to create control sockets between parent and child");
978 	}
979 	/*
980 	 * Create communication channel for sending events from child to parent.
981 	 */
982 	if (proto_client(NULL, "socketpair://", &res->hr_event) == -1) {
983 		/* TODO: There's no need for this to be fatal error. */
984 		KEEP_ERRNO((void)pidfile_remove(pfh));
985 		pjdlog_exit(EX_OSERR,
986 		    "Unable to create event sockets between child and parent");
987 	}
988 	/*
989 	 * Create communication channel for sending connection requests from
990 	 * child to parent.
991 	 */
992 	if (proto_client(NULL, "socketpair://", &res->hr_conn) == -1) {
993 		/* TODO: There's no need for this to be fatal error. */
994 		KEEP_ERRNO((void)pidfile_remove(pfh));
995 		pjdlog_exit(EX_OSERR,
996 		    "Unable to create connection sockets between child and parent");
997 	}
998 
999 	pid = fork();
1000 	if (pid == -1) {
1001 		/* TODO: There's no need for this to be fatal error. */
1002 		KEEP_ERRNO((void)pidfile_remove(pfh));
1003 		pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
1004 	}
1005 
1006 	if (pid > 0) {
1007 		/* This is parent. */
1008 		/* Declare that we are receiver. */
1009 		proto_recv(res->hr_event, NULL, 0);
1010 		proto_recv(res->hr_conn, NULL, 0);
1011 		/* Declare that we are sender. */
1012 		proto_send(res->hr_ctrl, NULL, 0);
1013 		res->hr_workerpid = pid;
1014 		return;
1015 	}
1016 
1017 	gres = res;
1018 	res->output_status_aux = output_status_aux;
1019 	mode = pjdlog_mode_get();
1020 	debuglevel = pjdlog_debug_get();
1021 
1022 	/* Declare that we are sender. */
1023 	proto_send(res->hr_event, NULL, 0);
1024 	proto_send(res->hr_conn, NULL, 0);
1025 	/* Declare that we are receiver. */
1026 	proto_recv(res->hr_ctrl, NULL, 0);
1027 	descriptors_cleanup(res);
1028 
1029 	descriptors_assert(res, mode);
1030 
1031 	pjdlog_init(mode);
1032 	pjdlog_debug_set(debuglevel);
1033 	pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
1034 	setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
1035 
1036 	init_local(res);
1037 	init_ggate(res);
1038 	init_environment(res);
1039 
1040 	if (drop_privs(res) != 0) {
1041 		cleanup(res);
1042 		exit(EX_CONFIG);
1043 	}
1044 	pjdlog_info("Privileges successfully dropped.");
1045 
1046 	/*
1047 	 * Create the guard thread first, so we can handle signals from the
1048 	 * very beginning.
1049 	 */
1050 	error = pthread_create(&td, NULL, guard_thread, res);
1051 	PJDLOG_ASSERT(error == 0);
1052 	/*
1053 	 * Create the control thread before sending any event to the parent,
1054 	 * as we can deadlock when parent sends control request to worker,
1055 	 * but worker has no control thread started yet, so parent waits.
1056 	 * In the meantime worker sends an event to the parent, but parent
1057 	 * is unable to handle the event, because it waits for control
1058 	 * request response.
1059 	 */
1060 	error = pthread_create(&td, NULL, ctrl_thread, res);
1061 	PJDLOG_ASSERT(error == 0);
1062 	if (real_remote(res)) {
1063 		error = init_remote(res, NULL, NULL);
1064 		if (error == 0) {
1065 			sync_start();
1066 		} else if (error == EBUSY) {
1067 			time_t start = time(NULL);
1068 
1069 			pjdlog_warning("Waiting for remote node to become %s for %ds.",
1070 			    role2str(HAST_ROLE_SECONDARY),
1071 			    res->hr_timeout);
1072 			for (;;) {
1073 				sleep(1);
1074 				error = init_remote(res, NULL, NULL);
1075 				if (error != EBUSY)
1076 					break;
1077 				if (time(NULL) > start + res->hr_timeout)
1078 					break;
1079 			}
1080 			if (error == EBUSY) {
1081 				pjdlog_warning("Remote node is still %s, starting anyway.",
1082 				    role2str(HAST_ROLE_PRIMARY));
1083 			}
1084 		}
1085 	}
1086 	error = pthread_create(&td, NULL, ggate_recv_thread, res);
1087 	PJDLOG_ASSERT(error == 0);
1088 	error = pthread_create(&td, NULL, local_send_thread, res);
1089 	PJDLOG_ASSERT(error == 0);
1090 	error = pthread_create(&td, NULL, remote_send_thread, res);
1091 	PJDLOG_ASSERT(error == 0);
1092 	error = pthread_create(&td, NULL, remote_recv_thread, res);
1093 	PJDLOG_ASSERT(error == 0);
1094 	error = pthread_create(&td, NULL, ggate_send_thread, res);
1095 	PJDLOG_ASSERT(error == 0);
1096 	fullystarted = true;
1097 	(void)sync_thread(res);
1098 }
1099 
1100 static void
1101 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio,
1102     const char *fmt, ...)
1103 {
1104 	char msg[1024];
1105 	va_list ap;
1106 
1107 	va_start(ap, fmt);
1108 	(void)vsnprintf(msg, sizeof(msg), fmt, ap);
1109 	va_end(ap);
1110 	switch (ggio->gctl_cmd) {
1111 	case BIO_READ:
1112 		(void)snprlcat(msg, sizeof(msg), "READ(%ju, %ju).",
1113 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1114 		break;
1115 	case BIO_DELETE:
1116 		(void)snprlcat(msg, sizeof(msg), "DELETE(%ju, %ju).",
1117 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1118 		break;
1119 	case BIO_FLUSH:
1120 		(void)snprlcat(msg, sizeof(msg), "FLUSH.");
1121 		break;
1122 	case BIO_WRITE:
1123 		(void)snprlcat(msg, sizeof(msg), "WRITE(%ju, %ju).",
1124 		    (uintmax_t)ggio->gctl_offset, (uintmax_t)ggio->gctl_length);
1125 		break;
1126 	default:
1127 		(void)snprlcat(msg, sizeof(msg), "UNKNOWN(%u).",
1128 		    (unsigned int)ggio->gctl_cmd);
1129 		break;
1130 	}
1131 	pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
1132 }
1133 
1134 static void
1135 remote_close(struct hast_resource *res, int ncomp)
1136 {
1137 
1138 	rw_wlock(&hio_remote_lock[ncomp]);
1139 	/*
1140 	 * Check for a race between dropping rlock and acquiring wlock -
1141 	 * another thread can close connection in-between.
1142 	 */
1143 	if (!ISCONNECTED(res, ncomp)) {
1144 		PJDLOG_ASSERT(res->hr_remotein == NULL);
1145 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
1146 		rw_unlock(&hio_remote_lock[ncomp]);
1147 		return;
1148 	}
1149 
1150 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1151 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1152 
1153 	pjdlog_debug(2, "Closing incoming connection to %s.",
1154 	    res->hr_remoteaddr);
1155 	proto_close(res->hr_remotein);
1156 	res->hr_remotein = NULL;
1157 	pjdlog_debug(2, "Closing outgoing connection to %s.",
1158 	    res->hr_remoteaddr);
1159 	proto_close(res->hr_remoteout);
1160 	res->hr_remoteout = NULL;
1161 
1162 	rw_unlock(&hio_remote_lock[ncomp]);
1163 
1164 	pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
1165 
1166 	/*
1167 	 * Stop synchronization if in-progress.
1168 	 */
1169 	sync_stop();
1170 
1171 	event_send(res, EVENT_DISCONNECT);
1172 }
1173 
1174 /*
1175  * Acknowledge write completion to the kernel, but don't update activemap yet.
1176  */
1177 static void
1178 write_complete(struct hast_resource *res, struct hio *hio)
1179 {
1180 	struct g_gate_ctl_io *ggio;
1181 	unsigned int ncomp;
1182 
1183 	PJDLOG_ASSERT(!hio->hio_done);
1184 
1185 	ggio = &hio->hio_ggio;
1186 	PJDLOG_ASSERT(ggio->gctl_cmd == BIO_WRITE);
1187 
1188 	/*
1189 	 * Bump local count if this is first write after
1190 	 * connection failure with remote node.
1191 	 */
1192 	ncomp = 1;
1193 	rw_rlock(&hio_remote_lock[ncomp]);
1194 	if (!ISCONNECTED(res, ncomp)) {
1195 		mtx_lock(&metadata_lock);
1196 		if (res->hr_primary_localcnt == res->hr_secondary_remotecnt) {
1197 			res->hr_primary_localcnt++;
1198 			pjdlog_debug(1, "Increasing localcnt to %ju.",
1199 			    (uintmax_t)res->hr_primary_localcnt);
1200 			(void)metadata_write(res);
1201 		}
1202 		mtx_unlock(&metadata_lock);
1203 	}
1204 	rw_unlock(&hio_remote_lock[ncomp]);
1205 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1)
1206 		primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1207 	hio->hio_done = true;
1208 }
1209 
1210 /*
1211  * Thread receives ggate I/O requests from the kernel and passes them to
1212  * appropriate threads:
1213  * WRITE - always goes to both local_send and remote_send threads
1214  * READ (when the block is up-to-date on local component) -
1215  *	only local_send thread
1216  * READ (when the block isn't up-to-date on local component) -
1217  *	only remote_send thread
1218  * DELETE - always goes to both local_send and remote_send threads
1219  * FLUSH - always goes to both local_send and remote_send threads
1220  */
1221 static void *
1222 ggate_recv_thread(void *arg)
1223 {
1224 	struct hast_resource *res = arg;
1225 	struct g_gate_ctl_io *ggio;
1226 	struct hio *hio;
1227 	unsigned int ii, ncomp, ncomps;
1228 	int error;
1229 
1230 	for (;;) {
1231 		pjdlog_debug(2, "ggate_recv: Taking free request.");
1232 		QUEUE_TAKE2(hio, free);
1233 		pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1234 		ggio = &hio->hio_ggio;
1235 		ggio->gctl_unit = res->hr_ggateunit;
1236 		ggio->gctl_length = MAXPHYS;
1237 		ggio->gctl_error = 0;
1238 		hio->hio_done = false;
1239 		hio->hio_replication = res->hr_replication;
1240 		pjdlog_debug(2,
1241 		    "ggate_recv: (%p) Waiting for request from the kernel.",
1242 		    hio);
1243 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) == -1) {
1244 			if (sigexit_received)
1245 				pthread_exit(NULL);
1246 			primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1247 		}
1248 		error = ggio->gctl_error;
1249 		switch (error) {
1250 		case 0:
1251 			break;
1252 		case ECANCELED:
1253 			/* Exit gracefully. */
1254 			if (!sigexit_received) {
1255 				pjdlog_debug(2,
1256 				    "ggate_recv: (%p) Received cancel from the kernel.",
1257 				    hio);
1258 				pjdlog_info("Received cancel from the kernel, exiting.");
1259 			}
1260 			pthread_exit(NULL);
1261 		case ENOMEM:
1262 			/*
1263 			 * Buffer too small? Impossible, we allocate MAXPHYS
1264 			 * bytes - request can't be bigger than that.
1265 			 */
1266 			/* FALLTHROUGH */
1267 		case ENXIO:
1268 		default:
1269 			primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1270 			    strerror(error));
1271 		}
1272 
1273 		ncomp = 0;
1274 		ncomps = HAST_NCOMPONENTS;
1275 
1276 		for (ii = 0; ii < ncomps; ii++)
1277 			hio->hio_errors[ii] = EINVAL;
1278 		reqlog(LOG_DEBUG, 2, ggio,
1279 		    "ggate_recv: (%p) Request received from the kernel: ",
1280 		    hio);
1281 
1282 		/*
1283 		 * Inform all components about new write request.
1284 		 * For read request prefer local component unless the given
1285 		 * range is out-of-date, then use remote component.
1286 		 */
1287 		switch (ggio->gctl_cmd) {
1288 		case BIO_READ:
1289 			res->hr_stat_read++;
1290 			ncomps = 1;
1291 			mtx_lock(&metadata_lock);
1292 			if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1293 			    res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1294 				/*
1295 				 * This range is up-to-date on local component,
1296 				 * so handle request locally.
1297 				 */
1298 				 /* Local component is 0 for now. */
1299 				ncomp = 0;
1300 			} else /* if (res->hr_syncsrc ==
1301 			    HAST_SYNCSRC_SECONDARY) */ {
1302 				PJDLOG_ASSERT(res->hr_syncsrc ==
1303 				    HAST_SYNCSRC_SECONDARY);
1304 				/*
1305 				 * This range is out-of-date on local component,
1306 				 * so send request to the remote node.
1307 				 */
1308 				 /* Remote component is 1 for now. */
1309 				ncomp = 1;
1310 			}
1311 			mtx_unlock(&metadata_lock);
1312 			break;
1313 		case BIO_WRITE:
1314 			res->hr_stat_write++;
1315 			if (res->hr_resuid == 0 &&
1316 			    res->hr_primary_localcnt == 0) {
1317 				/* This is first write. */
1318 				res->hr_primary_localcnt = 1;
1319 			}
1320 			for (;;) {
1321 				mtx_lock(&range_lock);
1322 				if (rangelock_islocked(range_sync,
1323 				    ggio->gctl_offset, ggio->gctl_length)) {
1324 					pjdlog_debug(2,
1325 					    "regular: Range offset=%jd length=%zu locked.",
1326 					    (intmax_t)ggio->gctl_offset,
1327 					    (size_t)ggio->gctl_length);
1328 					range_regular_wait = true;
1329 					cv_wait(&range_regular_cond, &range_lock);
1330 					range_regular_wait = false;
1331 					mtx_unlock(&range_lock);
1332 					continue;
1333 				}
1334 				if (rangelock_add(range_regular,
1335 				    ggio->gctl_offset, ggio->gctl_length) == -1) {
1336 					mtx_unlock(&range_lock);
1337 					pjdlog_debug(2,
1338 					    "regular: Range offset=%jd length=%zu is already locked, waiting.",
1339 					    (intmax_t)ggio->gctl_offset,
1340 					    (size_t)ggio->gctl_length);
1341 					sleep(1);
1342 					continue;
1343 				}
1344 				mtx_unlock(&range_lock);
1345 				break;
1346 			}
1347 			mtx_lock(&res->hr_amp_lock);
1348 			if (activemap_write_start(res->hr_amp,
1349 			    ggio->gctl_offset, ggio->gctl_length)) {
1350 				res->hr_stat_activemap_update++;
1351 				(void)hast_activemap_flush(res);
1352 			} else {
1353 				mtx_unlock(&res->hr_amp_lock);
1354 			}
1355 			if (ISMEMSYNC(hio)) {
1356 				hio->hio_memsyncacked = false;
1357 				refcnt_init(&hio->hio_writecount, ncomps);
1358 			}
1359 			break;
1360 		case BIO_DELETE:
1361 			res->hr_stat_delete++;
1362 			break;
1363 		case BIO_FLUSH:
1364 			res->hr_stat_flush++;
1365 			break;
1366 		}
1367 		pjdlog_debug(2,
1368 		    "ggate_recv: (%p) Moving request to the send queues.", hio);
1369 		refcnt_init(&hio->hio_countdown, ncomps);
1370 		for (ii = ncomp; ii < ncomps; ii++)
1371 			QUEUE_INSERT1(hio, send, ii);
1372 	}
1373 	/* NOTREACHED */
1374 	return (NULL);
1375 }
1376 
1377 /*
1378  * Thread reads from or writes to local component.
1379  * If local read fails, it redirects it to remote_send thread.
1380  */
1381 static void *
1382 local_send_thread(void *arg)
1383 {
1384 	struct hast_resource *res = arg;
1385 	struct g_gate_ctl_io *ggio;
1386 	struct hio *hio;
1387 	unsigned int ncomp, rncomp;
1388 	ssize_t ret;
1389 
1390 	/* Local component is 0 for now. */
1391 	ncomp = 0;
1392 	/* Remote component is 1 for now. */
1393 	rncomp = 1;
1394 
1395 	for (;;) {
1396 		pjdlog_debug(2, "local_send: Taking request.");
1397 		QUEUE_TAKE1(hio, send, ncomp, 0);
1398 		pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1399 		ggio = &hio->hio_ggio;
1400 		switch (ggio->gctl_cmd) {
1401 		case BIO_READ:
1402 			ret = pread(res->hr_localfd, ggio->gctl_data,
1403 			    ggio->gctl_length,
1404 			    ggio->gctl_offset + res->hr_localoff);
1405 			if (ret == ggio->gctl_length)
1406 				hio->hio_errors[ncomp] = 0;
1407 			else if (!ISSYNCREQ(hio)) {
1408 				/*
1409 				 * If READ failed, try to read from remote node.
1410 				 */
1411 				if (ret == -1) {
1412 					reqlog(LOG_WARNING, 0, ggio,
1413 					    "Local request failed (%s), trying remote node. ",
1414 					    strerror(errno));
1415 				} else if (ret != ggio->gctl_length) {
1416 					reqlog(LOG_WARNING, 0, ggio,
1417 					    "Local request failed (%zd != %jd), trying remote node. ",
1418 					    ret, (intmax_t)ggio->gctl_length);
1419 				}
1420 				QUEUE_INSERT1(hio, send, rncomp);
1421 				continue;
1422 			}
1423 			break;
1424 		case BIO_WRITE:
1425 			ret = pwrite(res->hr_localfd, ggio->gctl_data,
1426 			    ggio->gctl_length,
1427 			    ggio->gctl_offset + res->hr_localoff);
1428 			if (ret == -1) {
1429 				hio->hio_errors[ncomp] = errno;
1430 				reqlog(LOG_WARNING, 0, ggio,
1431 				    "Local request failed (%s): ",
1432 				    strerror(errno));
1433 			} else if (ret != ggio->gctl_length) {
1434 				hio->hio_errors[ncomp] = EIO;
1435 				reqlog(LOG_WARNING, 0, ggio,
1436 				    "Local request failed (%zd != %jd): ",
1437 				    ret, (intmax_t)ggio->gctl_length);
1438 			} else {
1439 				hio->hio_errors[ncomp] = 0;
1440 				if (ISASYNC(hio)) {
1441 					ggio->gctl_error = 0;
1442 					write_complete(res, hio);
1443 				}
1444 			}
1445 			break;
1446 		case BIO_DELETE:
1447 			ret = g_delete(res->hr_localfd,
1448 			    ggio->gctl_offset + res->hr_localoff,
1449 			    ggio->gctl_length);
1450 			if (ret == -1) {
1451 				hio->hio_errors[ncomp] = errno;
1452 				reqlog(LOG_WARNING, 0, ggio,
1453 				    "Local request failed (%s): ",
1454 				    strerror(errno));
1455 			} else {
1456 				hio->hio_errors[ncomp] = 0;
1457 			}
1458 			break;
1459 		case BIO_FLUSH:
1460 			if (!res->hr_localflush) {
1461 				ret = -1;
1462 				errno = EOPNOTSUPP;
1463 				break;
1464 			}
1465 			ret = g_flush(res->hr_localfd);
1466 			if (ret == -1) {
1467 				if (errno == EOPNOTSUPP)
1468 					res->hr_localflush = false;
1469 				hio->hio_errors[ncomp] = errno;
1470 				reqlog(LOG_WARNING, 0, ggio,
1471 				    "Local request failed (%s): ",
1472 				    strerror(errno));
1473 			} else {
1474 				hio->hio_errors[ncomp] = 0;
1475 			}
1476 			break;
1477 		}
1478 		if (ISMEMSYNCWRITE(hio)) {
1479 			if (refcnt_release(&hio->hio_writecount) == 0) {
1480 				write_complete(res, hio);
1481 			}
1482 		}
1483 		if (refcnt_release(&hio->hio_countdown) > 0)
1484 			continue;
1485 		if (ISSYNCREQ(hio)) {
1486 			mtx_lock(&sync_lock);
1487 			SYNCREQDONE(hio);
1488 			mtx_unlock(&sync_lock);
1489 			cv_signal(&sync_cond);
1490 		} else {
1491 			pjdlog_debug(2,
1492 			    "local_send: (%p) Moving request to the done queue.",
1493 			    hio);
1494 			QUEUE_INSERT2(hio, done);
1495 		}
1496 	}
1497 	/* NOTREACHED */
1498 	return (NULL);
1499 }
1500 
1501 static void
1502 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1503 {
1504 	struct nv *nv;
1505 
1506 	rw_rlock(&hio_remote_lock[ncomp]);
1507 
1508 	if (!ISCONNECTED(res, ncomp)) {
1509 		rw_unlock(&hio_remote_lock[ncomp]);
1510 		return;
1511 	}
1512 
1513 	PJDLOG_ASSERT(res->hr_remotein != NULL);
1514 	PJDLOG_ASSERT(res->hr_remoteout != NULL);
1515 
1516 	nv = nv_alloc();
1517 	nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1518 	if (nv_error(nv) != 0) {
1519 		rw_unlock(&hio_remote_lock[ncomp]);
1520 		nv_free(nv);
1521 		pjdlog_debug(1,
1522 		    "keepalive_send: Unable to prepare header to send.");
1523 		return;
1524 	}
1525 	if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) == -1) {
1526 		rw_unlock(&hio_remote_lock[ncomp]);
1527 		pjdlog_common(LOG_DEBUG, 1, errno,
1528 		    "keepalive_send: Unable to send request");
1529 		nv_free(nv);
1530 		remote_close(res, ncomp);
1531 		return;
1532 	}
1533 
1534 	rw_unlock(&hio_remote_lock[ncomp]);
1535 	nv_free(nv);
1536 	pjdlog_debug(2, "keepalive_send: Request sent.");
1537 }
1538 
1539 /*
1540  * Thread sends request to secondary node.
1541  */
1542 static void *
1543 remote_send_thread(void *arg)
1544 {
1545 	struct hast_resource *res = arg;
1546 	struct g_gate_ctl_io *ggio;
1547 	time_t lastcheck, now;
1548 	struct hio *hio;
1549 	struct nv *nv;
1550 	unsigned int ncomp;
1551 	bool wakeup;
1552 	uint64_t offset, length;
1553 	uint8_t cmd;
1554 	void *data;
1555 
1556 	/* Remote component is 1 for now. */
1557 	ncomp = 1;
1558 	lastcheck = time(NULL);
1559 
1560 	for (;;) {
1561 		pjdlog_debug(2, "remote_send: Taking request.");
1562 		QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE);
1563 		if (hio == NULL) {
1564 			now = time(NULL);
1565 			if (lastcheck + HAST_KEEPALIVE <= now) {
1566 				keepalive_send(res, ncomp);
1567 				lastcheck = now;
1568 			}
1569 			continue;
1570 		}
1571 		pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1572 		ggio = &hio->hio_ggio;
1573 		switch (ggio->gctl_cmd) {
1574 		case BIO_READ:
1575 			cmd = HIO_READ;
1576 			data = NULL;
1577 			offset = ggio->gctl_offset;
1578 			length = ggio->gctl_length;
1579 			break;
1580 		case BIO_WRITE:
1581 			cmd = HIO_WRITE;
1582 			data = ggio->gctl_data;
1583 			offset = ggio->gctl_offset;
1584 			length = ggio->gctl_length;
1585 			break;
1586 		case BIO_DELETE:
1587 			cmd = HIO_DELETE;
1588 			data = NULL;
1589 			offset = ggio->gctl_offset;
1590 			length = ggio->gctl_length;
1591 			break;
1592 		case BIO_FLUSH:
1593 			cmd = HIO_FLUSH;
1594 			data = NULL;
1595 			offset = 0;
1596 			length = 0;
1597 			break;
1598 		default:
1599 			PJDLOG_ABORT("invalid condition");
1600 		}
1601 		nv = nv_alloc();
1602 		nv_add_uint8(nv, cmd, "cmd");
1603 		nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1604 		nv_add_uint64(nv, offset, "offset");
1605 		nv_add_uint64(nv, length, "length");
1606 		if (ISMEMSYNCWRITE(hio))
1607 			nv_add_uint8(nv, 1, "memsync");
1608 		if (nv_error(nv) != 0) {
1609 			hio->hio_errors[ncomp] = nv_error(nv);
1610 			pjdlog_debug(2,
1611 			    "remote_send: (%p) Unable to prepare header to send.",
1612 			    hio);
1613 			reqlog(LOG_ERR, 0, ggio,
1614 			    "Unable to prepare header to send (%s): ",
1615 			    strerror(nv_error(nv)));
1616 			/* Move failed request immediately to the done queue. */
1617 			goto done_queue;
1618 		}
1619 		/*
1620 		 * Protect connection from disappearing.
1621 		 */
1622 		rw_rlock(&hio_remote_lock[ncomp]);
1623 		if (!ISCONNECTED(res, ncomp)) {
1624 			rw_unlock(&hio_remote_lock[ncomp]);
1625 			hio->hio_errors[ncomp] = ENOTCONN;
1626 			goto done_queue;
1627 		}
1628 		/*
1629 		 * Move the request to recv queue before sending it, because
1630 		 * in different order we can get reply before we move request
1631 		 * to recv queue.
1632 		 */
1633 		pjdlog_debug(2,
1634 		    "remote_send: (%p) Moving request to the recv queue.",
1635 		    hio);
1636 		mtx_lock(&hio_recv_list_lock[ncomp]);
1637 		wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1638 		TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1639 		hio_recv_list_size[ncomp]++;
1640 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1641 		if (hast_proto_send(res, res->hr_remoteout, nv, data,
1642 		    data != NULL ? length : 0) == -1) {
1643 			hio->hio_errors[ncomp] = errno;
1644 			rw_unlock(&hio_remote_lock[ncomp]);
1645 			pjdlog_debug(2,
1646 			    "remote_send: (%p) Unable to send request.", hio);
1647 			reqlog(LOG_ERR, 0, ggio,
1648 			    "Unable to send request (%s): ",
1649 			    strerror(hio->hio_errors[ncomp]));
1650 			remote_close(res, ncomp);
1651 		} else {
1652 			rw_unlock(&hio_remote_lock[ncomp]);
1653 		}
1654 		nv_free(nv);
1655 		if (wakeup)
1656 			cv_signal(&hio_recv_list_cond[ncomp]);
1657 		continue;
1658 done_queue:
1659 		nv_free(nv);
1660 		if (ISSYNCREQ(hio)) {
1661 			if (refcnt_release(&hio->hio_countdown) > 0)
1662 				continue;
1663 			mtx_lock(&sync_lock);
1664 			SYNCREQDONE(hio);
1665 			mtx_unlock(&sync_lock);
1666 			cv_signal(&sync_cond);
1667 			continue;
1668 		}
1669 		if (ggio->gctl_cmd == BIO_WRITE) {
1670 			mtx_lock(&res->hr_amp_lock);
1671 			if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1672 			    ggio->gctl_length)) {
1673 				(void)hast_activemap_flush(res);
1674 			} else {
1675 				mtx_unlock(&res->hr_amp_lock);
1676 			}
1677 			if (ISMEMSYNCWRITE(hio)) {
1678 				if (refcnt_release(&hio->hio_writecount) == 0) {
1679 					if (hio->hio_errors[0] == 0)
1680 						write_complete(res, hio);
1681 				}
1682 			}
1683 		}
1684 		if (refcnt_release(&hio->hio_countdown) > 0)
1685 			continue;
1686 		pjdlog_debug(2,
1687 		    "remote_send: (%p) Moving request to the done queue.",
1688 		    hio);
1689 		QUEUE_INSERT2(hio, done);
1690 	}
1691 	/* NOTREACHED */
1692 	return (NULL);
1693 }
1694 
1695 /*
1696  * Thread receives answer from secondary node and passes it to ggate_send
1697  * thread.
1698  */
1699 static void *
1700 remote_recv_thread(void *arg)
1701 {
1702 	struct hast_resource *res = arg;
1703 	struct g_gate_ctl_io *ggio;
1704 	struct hio *hio;
1705 	struct nv *nv;
1706 	unsigned int ncomp;
1707 	uint64_t seq;
1708 	bool memsyncack;
1709 	int error;
1710 
1711 	/* Remote component is 1 for now. */
1712 	ncomp = 1;
1713 
1714 	for (;;) {
1715 		/* Wait until there is anything to receive. */
1716 		mtx_lock(&hio_recv_list_lock[ncomp]);
1717 		while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1718 			pjdlog_debug(2, "remote_recv: No requests, waiting.");
1719 			cv_wait(&hio_recv_list_cond[ncomp],
1720 			    &hio_recv_list_lock[ncomp]);
1721 		}
1722 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1723 
1724 		memsyncack = false;
1725 
1726 		rw_rlock(&hio_remote_lock[ncomp]);
1727 		if (!ISCONNECTED(res, ncomp)) {
1728 			rw_unlock(&hio_remote_lock[ncomp]);
1729 			/*
1730 			 * Connection is dead, so move all pending requests to
1731 			 * the done queue (one-by-one).
1732 			 */
1733 			mtx_lock(&hio_recv_list_lock[ncomp]);
1734 			hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1735 			PJDLOG_ASSERT(hio != NULL);
1736 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1737 			    hio_next[ncomp]);
1738 			hio_recv_list_size[ncomp]--;
1739 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1740 			hio->hio_errors[ncomp] = ENOTCONN;
1741 			goto done_queue;
1742 		}
1743 		if (hast_proto_recv_hdr(res->hr_remotein, &nv) == -1) {
1744 			pjdlog_errno(LOG_ERR,
1745 			    "Unable to receive reply header");
1746 			rw_unlock(&hio_remote_lock[ncomp]);
1747 			remote_close(res, ncomp);
1748 			continue;
1749 		}
1750 		rw_unlock(&hio_remote_lock[ncomp]);
1751 		seq = nv_get_uint64(nv, "seq");
1752 		if (seq == 0) {
1753 			pjdlog_error("Header contains no 'seq' field.");
1754 			nv_free(nv);
1755 			continue;
1756 		}
1757 		memsyncack = nv_exists(nv, "received");
1758 		mtx_lock(&hio_recv_list_lock[ncomp]);
1759 		TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1760 			if (hio->hio_ggio.gctl_seq == seq) {
1761 				TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1762 				    hio_next[ncomp]);
1763 				hio_recv_list_size[ncomp]--;
1764 				break;
1765 			}
1766 		}
1767 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1768 		if (hio == NULL) {
1769 			pjdlog_error("Found no request matching received 'seq' field (%ju).",
1770 			    (uintmax_t)seq);
1771 			nv_free(nv);
1772 			continue;
1773 		}
1774 		ggio = &hio->hio_ggio;
1775 		error = nv_get_int16(nv, "error");
1776 		if (error != 0) {
1777 			/* Request failed on remote side. */
1778 			hio->hio_errors[ncomp] = error;
1779 			reqlog(LOG_WARNING, 0, ggio,
1780 			    "Remote request failed (%s): ", strerror(error));
1781 			nv_free(nv);
1782 			goto done_queue;
1783 		}
1784 		switch (ggio->gctl_cmd) {
1785 		case BIO_READ:
1786 			rw_rlock(&hio_remote_lock[ncomp]);
1787 			if (!ISCONNECTED(res, ncomp)) {
1788 				rw_unlock(&hio_remote_lock[ncomp]);
1789 				nv_free(nv);
1790 				goto done_queue;
1791 			}
1792 			if (hast_proto_recv_data(res, res->hr_remotein, nv,
1793 			    ggio->gctl_data, ggio->gctl_length) == -1) {
1794 				hio->hio_errors[ncomp] = errno;
1795 				pjdlog_errno(LOG_ERR,
1796 				    "Unable to receive reply data");
1797 				rw_unlock(&hio_remote_lock[ncomp]);
1798 				nv_free(nv);
1799 				remote_close(res, ncomp);
1800 				goto done_queue;
1801 			}
1802 			rw_unlock(&hio_remote_lock[ncomp]);
1803 			break;
1804 		case BIO_WRITE:
1805 		case BIO_DELETE:
1806 		case BIO_FLUSH:
1807 			break;
1808 		default:
1809 			PJDLOG_ABORT("invalid condition");
1810 		}
1811 		hio->hio_errors[ncomp] = 0;
1812 		nv_free(nv);
1813 done_queue:
1814 		if (ISMEMSYNCWRITE(hio)) {
1815 			if (!hio->hio_memsyncacked) {
1816 				PJDLOG_ASSERT(memsyncack ||
1817 				    hio->hio_errors[ncomp] != 0);
1818 				/* Remote ack arrived. */
1819 				if (refcnt_release(&hio->hio_writecount) == 0) {
1820 					if (hio->hio_errors[0] == 0)
1821 						write_complete(res, hio);
1822 				}
1823 				hio->hio_memsyncacked = true;
1824 				if (hio->hio_errors[ncomp] == 0) {
1825 					pjdlog_debug(2,
1826 					    "remote_recv: (%p) Moving request "
1827 					    "back to the recv queue.", hio);
1828 					mtx_lock(&hio_recv_list_lock[ncomp]);
1829 					TAILQ_INSERT_TAIL(&hio_recv_list[ncomp],
1830 					    hio, hio_next[ncomp]);
1831 					hio_recv_list_size[ncomp]++;
1832 					mtx_unlock(&hio_recv_list_lock[ncomp]);
1833 					continue;
1834 				}
1835 			} else {
1836 				PJDLOG_ASSERT(!memsyncack);
1837 				/* Remote final reply arrived. */
1838 			}
1839 		}
1840 		if (refcnt_release(&hio->hio_countdown) > 0)
1841 			continue;
1842 		if (ISSYNCREQ(hio)) {
1843 			mtx_lock(&sync_lock);
1844 			SYNCREQDONE(hio);
1845 			mtx_unlock(&sync_lock);
1846 			cv_signal(&sync_cond);
1847 		} else {
1848 			pjdlog_debug(2,
1849 			    "remote_recv: (%p) Moving request to the done queue.",
1850 			    hio);
1851 			QUEUE_INSERT2(hio, done);
1852 		}
1853 	}
1854 	/* NOTREACHED */
1855 	return (NULL);
1856 }
1857 
1858 /*
1859  * Thread sends answer to the kernel.
1860  */
1861 static void *
1862 ggate_send_thread(void *arg)
1863 {
1864 	struct hast_resource *res = arg;
1865 	struct g_gate_ctl_io *ggio;
1866 	struct hio *hio;
1867 	unsigned int ii, ncomps;
1868 
1869 	ncomps = HAST_NCOMPONENTS;
1870 
1871 	for (;;) {
1872 		pjdlog_debug(2, "ggate_send: Taking request.");
1873 		QUEUE_TAKE2(hio, done);
1874 		pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1875 		ggio = &hio->hio_ggio;
1876 		for (ii = 0; ii < ncomps; ii++) {
1877 			if (hio->hio_errors[ii] == 0) {
1878 				/*
1879 				 * One successful request is enough to declare
1880 				 * success.
1881 				 */
1882 				ggio->gctl_error = 0;
1883 				break;
1884 			}
1885 		}
1886 		if (ii == ncomps) {
1887 			/*
1888 			 * None of the requests were successful.
1889 			 * Use the error from local component except the
1890 			 * case when we did only remote request.
1891 			 */
1892 			if (ggio->gctl_cmd == BIO_READ &&
1893 			    res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
1894 				ggio->gctl_error = hio->hio_errors[1];
1895 			else
1896 				ggio->gctl_error = hio->hio_errors[0];
1897 		}
1898 		if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1899 			mtx_lock(&res->hr_amp_lock);
1900 			if (activemap_write_complete(res->hr_amp,
1901 			    ggio->gctl_offset, ggio->gctl_length)) {
1902 				res->hr_stat_activemap_update++;
1903 				(void)hast_activemap_flush(res);
1904 			} else {
1905 				mtx_unlock(&res->hr_amp_lock);
1906 			}
1907 		}
1908 		if (ggio->gctl_cmd == BIO_WRITE) {
1909 			/*
1910 			 * Unlock range we locked.
1911 			 */
1912 			mtx_lock(&range_lock);
1913 			rangelock_del(range_regular, ggio->gctl_offset,
1914 			    ggio->gctl_length);
1915 			if (range_sync_wait)
1916 				cv_signal(&range_sync_cond);
1917 			mtx_unlock(&range_lock);
1918 			if (!hio->hio_done)
1919 				write_complete(res, hio);
1920 		} else {
1921 			if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) == -1) {
1922 				primary_exit(EX_OSERR,
1923 				    "G_GATE_CMD_DONE failed");
1924 			}
1925 		}
1926 		if (hio->hio_errors[0]) {
1927 			switch (ggio->gctl_cmd) {
1928 			case BIO_READ:
1929 				res->hr_stat_read_error++;
1930 				break;
1931 			case BIO_WRITE:
1932 				res->hr_stat_write_error++;
1933 				break;
1934 			case BIO_DELETE:
1935 				res->hr_stat_delete_error++;
1936 				break;
1937 			case BIO_FLUSH:
1938 				res->hr_stat_flush_error++;
1939 				break;
1940 			}
1941 		}
1942 		pjdlog_debug(2,
1943 		    "ggate_send: (%p) Moving request to the free queue.", hio);
1944 		QUEUE_INSERT2(hio, free);
1945 	}
1946 	/* NOTREACHED */
1947 	return (NULL);
1948 }
1949 
1950 /*
1951  * Thread synchronize local and remote components.
1952  */
1953 static void *
1954 sync_thread(void *arg __unused)
1955 {
1956 	struct hast_resource *res = arg;
1957 	struct hio *hio;
1958 	struct g_gate_ctl_io *ggio;
1959 	struct timeval tstart, tend, tdiff;
1960 	unsigned int ii, ncomp, ncomps;
1961 	off_t offset, length, synced;
1962 	bool dorewind, directreads;
1963 	int syncext;
1964 
1965 	ncomps = HAST_NCOMPONENTS;
1966 	dorewind = true;
1967 	synced = 0;
1968 	offset = -1;
1969 	directreads = false;
1970 
1971 	for (;;) {
1972 		mtx_lock(&sync_lock);
1973 		if (offset >= 0 && !sync_inprogress) {
1974 			gettimeofday(&tend, NULL);
1975 			timersub(&tend, &tstart, &tdiff);
1976 			pjdlog_info("Synchronization interrupted after %#.0T. "
1977 			    "%NB synchronized so far.", &tdiff,
1978 			    (intmax_t)synced);
1979 			event_send(res, EVENT_SYNCINTR);
1980 		}
1981 		while (!sync_inprogress) {
1982 			dorewind = true;
1983 			synced = 0;
1984 			cv_wait(&sync_cond, &sync_lock);
1985 		}
1986 		mtx_unlock(&sync_lock);
1987 		/*
1988 		 * Obtain offset at which we should synchronize.
1989 		 * Rewind synchronization if needed.
1990 		 */
1991 		mtx_lock(&res->hr_amp_lock);
1992 		if (dorewind)
1993 			activemap_sync_rewind(res->hr_amp);
1994 		offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1995 		if (syncext != -1) {
1996 			/*
1997 			 * We synchronized entire syncext extent, we can mark
1998 			 * it as clean now.
1999 			 */
2000 			if (activemap_extent_complete(res->hr_amp, syncext))
2001 				(void)hast_activemap_flush(res);
2002 			else
2003 				mtx_unlock(&res->hr_amp_lock);
2004 		} else {
2005 			mtx_unlock(&res->hr_amp_lock);
2006 		}
2007 		if (dorewind) {
2008 			dorewind = false;
2009 			if (offset == -1)
2010 				pjdlog_info("Nodes are in sync.");
2011 			else {
2012 				pjdlog_info("Synchronization started. %NB to go.",
2013 				    (intmax_t)(res->hr_extentsize *
2014 				    activemap_ndirty(res->hr_amp)));
2015 				event_send(res, EVENT_SYNCSTART);
2016 				gettimeofday(&tstart, NULL);
2017 			}
2018 		}
2019 		if (offset == -1) {
2020 			sync_stop();
2021 			pjdlog_debug(1, "Nothing to synchronize.");
2022 			/*
2023 			 * Synchronization complete, make both localcnt and
2024 			 * remotecnt equal.
2025 			 */
2026 			ncomp = 1;
2027 			rw_rlock(&hio_remote_lock[ncomp]);
2028 			if (ISCONNECTED(res, ncomp)) {
2029 				if (synced > 0) {
2030 					int64_t bps;
2031 
2032 					gettimeofday(&tend, NULL);
2033 					timersub(&tend, &tstart, &tdiff);
2034 					bps = (int64_t)((double)synced /
2035 					    ((double)tdiff.tv_sec +
2036 					    (double)tdiff.tv_usec / 1000000));
2037 					pjdlog_info("Synchronization complete. "
2038 					    "%NB synchronized in %#.0lT (%NB/sec).",
2039 					    (intmax_t)synced, &tdiff,
2040 					    (intmax_t)bps);
2041 					event_send(res, EVENT_SYNCDONE);
2042 				}
2043 				mtx_lock(&metadata_lock);
2044 				if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
2045 					directreads = true;
2046 				res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
2047 				res->hr_primary_localcnt =
2048 				    res->hr_secondary_remotecnt;
2049 				res->hr_primary_remotecnt =
2050 				    res->hr_secondary_localcnt;
2051 				pjdlog_debug(1,
2052 				    "Setting localcnt to %ju and remotecnt to %ju.",
2053 				    (uintmax_t)res->hr_primary_localcnt,
2054 				    (uintmax_t)res->hr_primary_remotecnt);
2055 				(void)metadata_write(res);
2056 				mtx_unlock(&metadata_lock);
2057 			}
2058 			rw_unlock(&hio_remote_lock[ncomp]);
2059 			if (directreads) {
2060 				directreads = false;
2061 				enable_direct_reads(res);
2062 			}
2063 			continue;
2064 		}
2065 		pjdlog_debug(2, "sync: Taking free request.");
2066 		QUEUE_TAKE2(hio, free);
2067 		pjdlog_debug(2, "sync: (%p) Got free request.", hio);
2068 		/*
2069 		 * Lock the range we are going to synchronize. We don't want
2070 		 * race where someone writes between our read and write.
2071 		 */
2072 		for (;;) {
2073 			mtx_lock(&range_lock);
2074 			if (rangelock_islocked(range_regular, offset, length)) {
2075 				pjdlog_debug(2,
2076 				    "sync: Range offset=%jd length=%jd locked.",
2077 				    (intmax_t)offset, (intmax_t)length);
2078 				range_sync_wait = true;
2079 				cv_wait(&range_sync_cond, &range_lock);
2080 				range_sync_wait = false;
2081 				mtx_unlock(&range_lock);
2082 				continue;
2083 			}
2084 			if (rangelock_add(range_sync, offset, length) == -1) {
2085 				mtx_unlock(&range_lock);
2086 				pjdlog_debug(2,
2087 				    "sync: Range offset=%jd length=%jd is already locked, waiting.",
2088 				    (intmax_t)offset, (intmax_t)length);
2089 				sleep(1);
2090 				continue;
2091 			}
2092 			mtx_unlock(&range_lock);
2093 			break;
2094 		}
2095 		/*
2096 		 * First read the data from synchronization source.
2097 		 */
2098 		SYNCREQ(hio);
2099 		ggio = &hio->hio_ggio;
2100 		ggio->gctl_cmd = BIO_READ;
2101 		ggio->gctl_offset = offset;
2102 		ggio->gctl_length = length;
2103 		ggio->gctl_error = 0;
2104 		hio->hio_done = false;
2105 		hio->hio_replication = res->hr_replication;
2106 		for (ii = 0; ii < ncomps; ii++)
2107 			hio->hio_errors[ii] = EINVAL;
2108 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2109 		    hio);
2110 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2111 		    hio);
2112 		mtx_lock(&metadata_lock);
2113 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2114 			/*
2115 			 * This range is up-to-date on local component,
2116 			 * so handle request locally.
2117 			 */
2118 			 /* Local component is 0 for now. */
2119 			ncomp = 0;
2120 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2121 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2122 			/*
2123 			 * This range is out-of-date on local component,
2124 			 * so send request to the remote node.
2125 			 */
2126 			 /* Remote component is 1 for now. */
2127 			ncomp = 1;
2128 		}
2129 		mtx_unlock(&metadata_lock);
2130 		refcnt_init(&hio->hio_countdown, 1);
2131 		QUEUE_INSERT1(hio, send, ncomp);
2132 
2133 		/*
2134 		 * Let's wait for READ to finish.
2135 		 */
2136 		mtx_lock(&sync_lock);
2137 		while (!ISSYNCREQDONE(hio))
2138 			cv_wait(&sync_cond, &sync_lock);
2139 		mtx_unlock(&sync_lock);
2140 
2141 		if (hio->hio_errors[ncomp] != 0) {
2142 			pjdlog_error("Unable to read synchronization data: %s.",
2143 			    strerror(hio->hio_errors[ncomp]));
2144 			goto free_queue;
2145 		}
2146 
2147 		/*
2148 		 * We read the data from synchronization source, now write it
2149 		 * to synchronization target.
2150 		 */
2151 		SYNCREQ(hio);
2152 		ggio->gctl_cmd = BIO_WRITE;
2153 		for (ii = 0; ii < ncomps; ii++)
2154 			hio->hio_errors[ii] = EINVAL;
2155 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
2156 		    hio);
2157 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2158 		    hio);
2159 		mtx_lock(&metadata_lock);
2160 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
2161 			/*
2162 			 * This range is up-to-date on local component,
2163 			 * so we update remote component.
2164 			 */
2165 			 /* Remote component is 1 for now. */
2166 			ncomp = 1;
2167 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
2168 			PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
2169 			/*
2170 			 * This range is out-of-date on local component,
2171 			 * so we update it.
2172 			 */
2173 			 /* Local component is 0 for now. */
2174 			ncomp = 0;
2175 		}
2176 		mtx_unlock(&metadata_lock);
2177 
2178 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
2179 		    hio);
2180 		refcnt_init(&hio->hio_countdown, 1);
2181 		QUEUE_INSERT1(hio, send, ncomp);
2182 
2183 		/*
2184 		 * Let's wait for WRITE to finish.
2185 		 */
2186 		mtx_lock(&sync_lock);
2187 		while (!ISSYNCREQDONE(hio))
2188 			cv_wait(&sync_cond, &sync_lock);
2189 		mtx_unlock(&sync_lock);
2190 
2191 		if (hio->hio_errors[ncomp] != 0) {
2192 			pjdlog_error("Unable to write synchronization data: %s.",
2193 			    strerror(hio->hio_errors[ncomp]));
2194 			goto free_queue;
2195 		}
2196 
2197 		synced += length;
2198 free_queue:
2199 		mtx_lock(&range_lock);
2200 		rangelock_del(range_sync, offset, length);
2201 		if (range_regular_wait)
2202 			cv_signal(&range_regular_cond);
2203 		mtx_unlock(&range_lock);
2204 		pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
2205 		    hio);
2206 		QUEUE_INSERT2(hio, free);
2207 	}
2208 	/* NOTREACHED */
2209 	return (NULL);
2210 }
2211 
2212 void
2213 primary_config_reload(struct hast_resource *res, struct nv *nv)
2214 {
2215 	unsigned int ii, ncomps;
2216 	int modified, vint;
2217 	const char *vstr;
2218 
2219 	pjdlog_info("Reloading configuration...");
2220 
2221 	PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
2222 	PJDLOG_ASSERT(gres == res);
2223 	nv_assert(nv, "remoteaddr");
2224 	nv_assert(nv, "sourceaddr");
2225 	nv_assert(nv, "replication");
2226 	nv_assert(nv, "checksum");
2227 	nv_assert(nv, "compression");
2228 	nv_assert(nv, "timeout");
2229 	nv_assert(nv, "exec");
2230 	nv_assert(nv, "metaflush");
2231 
2232 	ncomps = HAST_NCOMPONENTS;
2233 
2234 #define MODIFIED_REMOTEADDR	0x01
2235 #define MODIFIED_SOURCEADDR	0x02
2236 #define MODIFIED_REPLICATION	0x04
2237 #define MODIFIED_CHECKSUM	0x08
2238 #define MODIFIED_COMPRESSION	0x10
2239 #define MODIFIED_TIMEOUT	0x20
2240 #define MODIFIED_EXEC		0x40
2241 #define MODIFIED_METAFLUSH	0x80
2242 	modified = 0;
2243 
2244 	vstr = nv_get_string(nv, "remoteaddr");
2245 	if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
2246 		/*
2247 		 * Don't copy res->hr_remoteaddr to gres just yet.
2248 		 * We want remote_close() to log disconnect from the old
2249 		 * addresses, not from the new ones.
2250 		 */
2251 		modified |= MODIFIED_REMOTEADDR;
2252 	}
2253 	vstr = nv_get_string(nv, "sourceaddr");
2254 	if (strcmp(gres->hr_sourceaddr, vstr) != 0) {
2255 		strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr));
2256 		modified |= MODIFIED_SOURCEADDR;
2257 	}
2258 	vint = nv_get_int32(nv, "replication");
2259 	if (gres->hr_replication != vint) {
2260 		gres->hr_replication = vint;
2261 		modified |= MODIFIED_REPLICATION;
2262 	}
2263 	vint = nv_get_int32(nv, "checksum");
2264 	if (gres->hr_checksum != vint) {
2265 		gres->hr_checksum = vint;
2266 		modified |= MODIFIED_CHECKSUM;
2267 	}
2268 	vint = nv_get_int32(nv, "compression");
2269 	if (gres->hr_compression != vint) {
2270 		gres->hr_compression = vint;
2271 		modified |= MODIFIED_COMPRESSION;
2272 	}
2273 	vint = nv_get_int32(nv, "timeout");
2274 	if (gres->hr_timeout != vint) {
2275 		gres->hr_timeout = vint;
2276 		modified |= MODIFIED_TIMEOUT;
2277 	}
2278 	vstr = nv_get_string(nv, "exec");
2279 	if (strcmp(gres->hr_exec, vstr) != 0) {
2280 		strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
2281 		modified |= MODIFIED_EXEC;
2282 	}
2283 	vint = nv_get_int32(nv, "metaflush");
2284 	if (gres->hr_metaflush != vint) {
2285 		gres->hr_metaflush = vint;
2286 		modified |= MODIFIED_METAFLUSH;
2287 	}
2288 
2289 	/*
2290 	 * Change timeout for connected sockets.
2291 	 * Don't bother if we need to reconnect.
2292 	 */
2293 	if ((modified & MODIFIED_TIMEOUT) != 0 &&
2294 	    (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) == 0) {
2295 		for (ii = 0; ii < ncomps; ii++) {
2296 			if (!ISREMOTE(ii))
2297 				continue;
2298 			rw_rlock(&hio_remote_lock[ii]);
2299 			if (!ISCONNECTED(gres, ii)) {
2300 				rw_unlock(&hio_remote_lock[ii]);
2301 				continue;
2302 			}
2303 			rw_unlock(&hio_remote_lock[ii]);
2304 			if (proto_timeout(gres->hr_remotein,
2305 			    gres->hr_timeout) == -1) {
2306 				pjdlog_errno(LOG_WARNING,
2307 				    "Unable to set connection timeout");
2308 			}
2309 			if (proto_timeout(gres->hr_remoteout,
2310 			    gres->hr_timeout) == -1) {
2311 				pjdlog_errno(LOG_WARNING,
2312 				    "Unable to set connection timeout");
2313 			}
2314 		}
2315 	}
2316 	if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR)) != 0) {
2317 		for (ii = 0; ii < ncomps; ii++) {
2318 			if (!ISREMOTE(ii))
2319 				continue;
2320 			remote_close(gres, ii);
2321 		}
2322 		if (modified & MODIFIED_REMOTEADDR) {
2323 			vstr = nv_get_string(nv, "remoteaddr");
2324 			strlcpy(gres->hr_remoteaddr, vstr,
2325 			    sizeof(gres->hr_remoteaddr));
2326 		}
2327 	}
2328 #undef	MODIFIED_REMOTEADDR
2329 #undef	MODIFIED_SOURCEADDR
2330 #undef	MODIFIED_REPLICATION
2331 #undef	MODIFIED_CHECKSUM
2332 #undef	MODIFIED_COMPRESSION
2333 #undef	MODIFIED_TIMEOUT
2334 #undef	MODIFIED_EXEC
2335 #undef	MODIFIED_METAFLUSH
2336 
2337 	pjdlog_info("Configuration reloaded successfully.");
2338 }
2339 
2340 static void
2341 guard_one(struct hast_resource *res, unsigned int ncomp)
2342 {
2343 	struct proto_conn *in, *out;
2344 
2345 	if (!ISREMOTE(ncomp))
2346 		return;
2347 
2348 	rw_rlock(&hio_remote_lock[ncomp]);
2349 
2350 	if (!real_remote(res)) {
2351 		rw_unlock(&hio_remote_lock[ncomp]);
2352 		return;
2353 	}
2354 
2355 	if (ISCONNECTED(res, ncomp)) {
2356 		PJDLOG_ASSERT(res->hr_remotein != NULL);
2357 		PJDLOG_ASSERT(res->hr_remoteout != NULL);
2358 		rw_unlock(&hio_remote_lock[ncomp]);
2359 		pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2360 		    res->hr_remoteaddr);
2361 		return;
2362 	}
2363 
2364 	PJDLOG_ASSERT(res->hr_remotein == NULL);
2365 	PJDLOG_ASSERT(res->hr_remoteout == NULL);
2366 	/*
2367 	 * Upgrade the lock. It doesn't have to be atomic as no other thread
2368 	 * can change connection status from disconnected to connected.
2369 	 */
2370 	rw_unlock(&hio_remote_lock[ncomp]);
2371 	pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2372 	    res->hr_remoteaddr);
2373 	in = out = NULL;
2374 	if (init_remote(res, &in, &out) == 0) {
2375 		rw_wlock(&hio_remote_lock[ncomp]);
2376 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2377 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2378 		PJDLOG_ASSERT(in != NULL && out != NULL);
2379 		res->hr_remotein = in;
2380 		res->hr_remoteout = out;
2381 		rw_unlock(&hio_remote_lock[ncomp]);
2382 		pjdlog_info("Successfully reconnected to %s.",
2383 		    res->hr_remoteaddr);
2384 		sync_start();
2385 	} else {
2386 		/* Both connections should be NULL. */
2387 		PJDLOG_ASSERT(res->hr_remotein == NULL);
2388 		PJDLOG_ASSERT(res->hr_remoteout == NULL);
2389 		PJDLOG_ASSERT(in == NULL && out == NULL);
2390 		pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2391 		    res->hr_remoteaddr);
2392 	}
2393 }
2394 
2395 /*
2396  * Thread guards remote connections and reconnects when needed, handles
2397  * signals, etc.
2398  */
2399 static void *
2400 guard_thread(void *arg)
2401 {
2402 	struct hast_resource *res = arg;
2403 	unsigned int ii, ncomps;
2404 	struct timespec timeout;
2405 	time_t lastcheck, now;
2406 	sigset_t mask;
2407 	int signo;
2408 
2409 	ncomps = HAST_NCOMPONENTS;
2410 	lastcheck = time(NULL);
2411 
2412 	PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2413 	PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2414 	PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2415 
2416 	timeout.tv_sec = HAST_KEEPALIVE;
2417 	timeout.tv_nsec = 0;
2418 	signo = -1;
2419 
2420 	for (;;) {
2421 		switch (signo) {
2422 		case SIGINT:
2423 		case SIGTERM:
2424 			sigexit_received = true;
2425 			primary_exitx(EX_OK,
2426 			    "Termination signal received, exiting.");
2427 			break;
2428 		default:
2429 			break;
2430 		}
2431 
2432 		/*
2433 		 * Don't check connections until we fully started,
2434 		 * as we may still be looping, waiting for remote node
2435 		 * to switch from primary to secondary.
2436 		 */
2437 		if (fullystarted) {
2438 			pjdlog_debug(2, "remote_guard: Checking connections.");
2439 			now = time(NULL);
2440 			if (lastcheck + HAST_KEEPALIVE <= now) {
2441 				for (ii = 0; ii < ncomps; ii++)
2442 					guard_one(res, ii);
2443 				lastcheck = now;
2444 			}
2445 		}
2446 		signo = sigtimedwait(&mask, NULL, &timeout);
2447 	}
2448 	/* NOTREACHED */
2449 	return (NULL);
2450 }
2451