xref: /freebsd/sbin/hastd/primary.c (revision d8b878873e7aa8df1972cc6a642804b17eb61087)
1 /*-
2  * Copyright (c) 2009 The FreeBSD Foundation
3  * All rights reserved.
4  *
5  * This software was developed by Pawel Jakub Dawidek under sponsorship from
6  * the FreeBSD Foundation.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include <sys/types.h>
34 #include <sys/time.h>
35 #include <sys/bio.h>
36 #include <sys/disk.h>
37 #include <sys/refcount.h>
38 #include <sys/stat.h>
39 
40 #include <geom/gate/g_gate.h>
41 
42 #include <assert.h>
43 #include <err.h>
44 #include <errno.h>
45 #include <fcntl.h>
46 #include <libgeom.h>
47 #include <pthread.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 "hast.h"
60 #include "hast_proto.h"
61 #include "hastd.h"
62 #include "metadata.h"
63 #include "proto.h"
64 #include "pjdlog.h"
65 #include "subr.h"
66 #include "synch.h"
67 
68 struct hio {
69 	/*
70 	 * Number of components we are still waiting for.
71 	 * When this field goes to 0, we can send the request back to the
72 	 * kernel. Each component has to decrease this counter by one
73 	 * even on failure.
74 	 */
75 	unsigned int		 hio_countdown;
76 	/*
77 	 * Each component has a place to store its own error.
78 	 * Once the request is handled by all components we can decide if the
79 	 * request overall is successful or not.
80 	 */
81 	int			*hio_errors;
82 	/*
83 	 * Structure used to comunicate with GEOM Gate class.
84 	 */
85 	struct g_gate_ctl_io	 hio_ggio;
86 	TAILQ_ENTRY(hio)	*hio_next;
87 };
88 #define	hio_free_next	hio_next[0]
89 #define	hio_done_next	hio_next[0]
90 
91 /*
92  * Free list holds unused structures. When free list is empty, we have to wait
93  * until some in-progress requests are freed.
94  */
95 static TAILQ_HEAD(, hio) hio_free_list;
96 static pthread_mutex_t hio_free_list_lock;
97 static pthread_cond_t hio_free_list_cond;
98 /*
99  * There is one send list for every component. One requests is placed on all
100  * send lists - each component gets the same request, but each component is
101  * responsible for managing his own send list.
102  */
103 static TAILQ_HEAD(, hio) *hio_send_list;
104 static pthread_mutex_t *hio_send_list_lock;
105 static pthread_cond_t *hio_send_list_cond;
106 /*
107  * There is one recv list for every component, although local components don't
108  * use recv lists as local requests are done synchronously.
109  */
110 static TAILQ_HEAD(, hio) *hio_recv_list;
111 static pthread_mutex_t *hio_recv_list_lock;
112 static pthread_cond_t *hio_recv_list_cond;
113 /*
114  * Request is placed on done list by the slowest component (the one that
115  * decreased hio_countdown from 1 to 0).
116  */
117 static TAILQ_HEAD(, hio) hio_done_list;
118 static pthread_mutex_t hio_done_list_lock;
119 static pthread_cond_t hio_done_list_cond;
120 /*
121  * Structure below are for interaction with sync thread.
122  */
123 static bool sync_inprogress;
124 static pthread_mutex_t sync_lock;
125 static pthread_cond_t sync_cond;
126 /*
127  * The lock below allows to synchornize access to remote connections.
128  */
129 static pthread_rwlock_t *hio_remote_lock;
130 static pthread_mutex_t hio_guard_lock;
131 static pthread_cond_t hio_guard_cond;
132 
133 /*
134  * Lock to synchronize metadata updates. Also synchronize access to
135  * hr_primary_localcnt and hr_primary_remotecnt fields.
136  */
137 static pthread_mutex_t metadata_lock;
138 
139 /*
140  * Maximum number of outstanding I/O requests.
141  */
142 #define	HAST_HIO_MAX	256
143 /*
144  * Number of components. At this point there are only two components: local
145  * and remote, but in the future it might be possible to use multiple local
146  * and remote components.
147  */
148 #define	HAST_NCOMPONENTS	2
149 /*
150  * Number of seconds to sleep before next reconnect try.
151  */
152 #define	RECONNECT_SLEEP		5
153 
154 #define	ISCONNECTED(res, no)	\
155 	((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
156 
157 #define	QUEUE_INSERT1(hio, name, ncomp)	do {				\
158 	bool _wakeup;							\
159 									\
160 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
161 	_wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]);		\
162 	TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio),		\
163 	    hio_next[(ncomp)]);						\
164 	mtx_unlock(&hio_##name##_list_lock[ncomp]);			\
165 	if (_wakeup)							\
166 		cv_signal(&hio_##name##_list_cond[(ncomp)]);		\
167 } while (0)
168 #define	QUEUE_INSERT2(hio, name)	do {				\
169 	bool _wakeup;							\
170 									\
171 	mtx_lock(&hio_##name##_list_lock);				\
172 	_wakeup = TAILQ_EMPTY(&hio_##name##_list);			\
173 	TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
174 	mtx_unlock(&hio_##name##_list_lock);				\
175 	if (_wakeup)							\
176 		cv_signal(&hio_##name##_list_cond);			\
177 } while (0)
178 #define	QUEUE_TAKE1(hio, name, ncomp)	do {				\
179 	mtx_lock(&hio_##name##_list_lock[(ncomp)]);			\
180 	while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL) { \
181 		cv_wait(&hio_##name##_list_cond[(ncomp)],		\
182 		    &hio_##name##_list_lock[(ncomp)]);			\
183 	}								\
184 	TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio),		\
185 	    hio_next[(ncomp)]);						\
186 	mtx_unlock(&hio_##name##_list_lock[(ncomp)]);			\
187 } while (0)
188 #define	QUEUE_TAKE2(hio, name)	do {					\
189 	mtx_lock(&hio_##name##_list_lock);				\
190 	while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {	\
191 		cv_wait(&hio_##name##_list_cond,			\
192 		    &hio_##name##_list_lock);				\
193 	}								\
194 	TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next);	\
195 	mtx_unlock(&hio_##name##_list_lock);				\
196 } while (0)
197 
198 #define	SYNCREQ(hio)		do { (hio)->hio_ggio.gctl_unit = -1; } while (0)
199 #define	ISSYNCREQ(hio)		((hio)->hio_ggio.gctl_unit == -1)
200 #define	SYNCREQDONE(hio)	do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
201 #define	ISSYNCREQDONE(hio)	((hio)->hio_ggio.gctl_unit == -2)
202 
203 static struct hast_resource *gres;
204 
205 static pthread_mutex_t range_lock;
206 static struct rangelocks *range_regular;
207 static bool range_regular_wait;
208 static pthread_cond_t range_regular_cond;
209 static struct rangelocks *range_sync;
210 static bool range_sync_wait;
211 static pthread_cond_t range_sync_cond;
212 
213 static void *ggate_recv_thread(void *arg);
214 static void *local_send_thread(void *arg);
215 static void *remote_send_thread(void *arg);
216 static void *remote_recv_thread(void *arg);
217 static void *ggate_send_thread(void *arg);
218 static void *sync_thread(void *arg);
219 static void *guard_thread(void *arg);
220 
221 static void sighandler(int sig);
222 
223 static void
224 cleanup(struct hast_resource *res)
225 {
226 	int rerrno;
227 
228 	/* Remember errno. */
229 	rerrno = errno;
230 
231 	/*
232 	 * Close descriptor to /dev/hast/<name>
233 	 * to work-around race in the kernel.
234 	 */
235 	close(res->hr_localfd);
236 
237 	/* Destroy ggate provider if we created one. */
238 	if (res->hr_ggateunit >= 0) {
239 		struct g_gate_ctl_destroy ggiod;
240 
241 		ggiod.gctl_version = G_GATE_VERSION;
242 		ggiod.gctl_unit = res->hr_ggateunit;
243 		ggiod.gctl_force = 1;
244 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) {
245 			pjdlog_warning("Unable to destroy hast/%s device",
246 			    res->hr_provname);
247 		}
248 		res->hr_ggateunit = -1;
249 	}
250 
251 	/* Restore errno. */
252 	errno = rerrno;
253 }
254 
255 static void
256 primary_exit(int exitcode, const char *fmt, ...)
257 {
258 	va_list ap;
259 
260 	assert(exitcode != EX_OK);
261 	va_start(ap, fmt);
262 	pjdlogv_errno(LOG_ERR, fmt, ap);
263 	va_end(ap);
264 	cleanup(gres);
265 	exit(exitcode);
266 }
267 
268 static void
269 primary_exitx(int exitcode, const char *fmt, ...)
270 {
271 	va_list ap;
272 
273 	va_start(ap, fmt);
274 	pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
275 	va_end(ap);
276 	cleanup(gres);
277 	exit(exitcode);
278 }
279 
280 static int
281 hast_activemap_flush(struct hast_resource *res)
282 {
283 	const unsigned char *buf;
284 	size_t size;
285 
286 	buf = activemap_bitmap(res->hr_amp, &size);
287 	assert(buf != NULL);
288 	assert((size % res->hr_local_sectorsize) == 0);
289 	if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
290 	    (ssize_t)size) {
291 		KEEP_ERRNO(pjdlog_errno(LOG_ERR,
292 		    "Unable to flush activemap to disk"));
293 		return (-1);
294 	}
295 	return (0);
296 }
297 
298 static void
299 init_environment(struct hast_resource *res __unused)
300 {
301 	struct hio *hio;
302 	unsigned int ii, ncomps;
303 
304 	/*
305 	 * In the future it might be per-resource value.
306 	 */
307 	ncomps = HAST_NCOMPONENTS;
308 
309 	/*
310 	 * Allocate memory needed by lists.
311 	 */
312 	hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
313 	if (hio_send_list == NULL) {
314 		primary_exitx(EX_TEMPFAIL,
315 		    "Unable to allocate %zu bytes of memory for send lists.",
316 		    sizeof(hio_send_list[0]) * ncomps);
317 	}
318 	hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
319 	if (hio_send_list_lock == NULL) {
320 		primary_exitx(EX_TEMPFAIL,
321 		    "Unable to allocate %zu bytes of memory for send list locks.",
322 		    sizeof(hio_send_list_lock[0]) * ncomps);
323 	}
324 	hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
325 	if (hio_send_list_cond == NULL) {
326 		primary_exitx(EX_TEMPFAIL,
327 		    "Unable to allocate %zu bytes of memory for send list condition variables.",
328 		    sizeof(hio_send_list_cond[0]) * ncomps);
329 	}
330 	hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
331 	if (hio_recv_list == NULL) {
332 		primary_exitx(EX_TEMPFAIL,
333 		    "Unable to allocate %zu bytes of memory for recv lists.",
334 		    sizeof(hio_recv_list[0]) * ncomps);
335 	}
336 	hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
337 	if (hio_recv_list_lock == NULL) {
338 		primary_exitx(EX_TEMPFAIL,
339 		    "Unable to allocate %zu bytes of memory for recv list locks.",
340 		    sizeof(hio_recv_list_lock[0]) * ncomps);
341 	}
342 	hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
343 	if (hio_recv_list_cond == NULL) {
344 		primary_exitx(EX_TEMPFAIL,
345 		    "Unable to allocate %zu bytes of memory for recv list condition variables.",
346 		    sizeof(hio_recv_list_cond[0]) * ncomps);
347 	}
348 	hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
349 	if (hio_remote_lock == NULL) {
350 		primary_exitx(EX_TEMPFAIL,
351 		    "Unable to allocate %zu bytes of memory for remote connections locks.",
352 		    sizeof(hio_remote_lock[0]) * ncomps);
353 	}
354 
355 	/*
356 	 * Initialize lists, their locks and theirs condition variables.
357 	 */
358 	TAILQ_INIT(&hio_free_list);
359 	mtx_init(&hio_free_list_lock);
360 	cv_init(&hio_free_list_cond);
361 	for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
362 		TAILQ_INIT(&hio_send_list[ii]);
363 		mtx_init(&hio_send_list_lock[ii]);
364 		cv_init(&hio_send_list_cond[ii]);
365 		TAILQ_INIT(&hio_recv_list[ii]);
366 		mtx_init(&hio_recv_list_lock[ii]);
367 		cv_init(&hio_recv_list_cond[ii]);
368 		rw_init(&hio_remote_lock[ii]);
369 	}
370 	TAILQ_INIT(&hio_done_list);
371 	mtx_init(&hio_done_list_lock);
372 	cv_init(&hio_done_list_cond);
373 	mtx_init(&hio_guard_lock);
374 	cv_init(&hio_guard_cond);
375 	mtx_init(&metadata_lock);
376 
377 	/*
378 	 * Allocate requests pool and initialize requests.
379 	 */
380 	for (ii = 0; ii < HAST_HIO_MAX; ii++) {
381 		hio = malloc(sizeof(*hio));
382 		if (hio == NULL) {
383 			primary_exitx(EX_TEMPFAIL,
384 			    "Unable to allocate %zu bytes of memory for hio request.",
385 			    sizeof(*hio));
386 		}
387 		hio->hio_countdown = 0;
388 		hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
389 		if (hio->hio_errors == NULL) {
390 			primary_exitx(EX_TEMPFAIL,
391 			    "Unable allocate %zu bytes of memory for hio errors.",
392 			    sizeof(hio->hio_errors[0]) * ncomps);
393 		}
394 		hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
395 		if (hio->hio_next == NULL) {
396 			primary_exitx(EX_TEMPFAIL,
397 			    "Unable allocate %zu bytes of memory for hio_next field.",
398 			    sizeof(hio->hio_next[0]) * ncomps);
399 		}
400 		hio->hio_ggio.gctl_version = G_GATE_VERSION;
401 		hio->hio_ggio.gctl_data = malloc(MAXPHYS);
402 		if (hio->hio_ggio.gctl_data == NULL) {
403 			primary_exitx(EX_TEMPFAIL,
404 			    "Unable to allocate %zu bytes of memory for gctl_data.",
405 			    MAXPHYS);
406 		}
407 		hio->hio_ggio.gctl_length = MAXPHYS;
408 		hio->hio_ggio.gctl_error = 0;
409 		TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
410 	}
411 
412 	/*
413 	 * Turn on signals handling.
414 	 */
415 	signal(SIGINT, sighandler);
416 	signal(SIGTERM, sighandler);
417 }
418 
419 static void
420 init_local(struct hast_resource *res)
421 {
422 	unsigned char *buf;
423 	size_t mapsize;
424 
425 	if (metadata_read(res, true) < 0)
426 		exit(EX_NOINPUT);
427 	mtx_init(&res->hr_amp_lock);
428 	if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
429 	    res->hr_local_sectorsize, res->hr_keepdirty) < 0) {
430 		primary_exit(EX_TEMPFAIL, "Unable to create activemap");
431 	}
432 	mtx_init(&range_lock);
433 	cv_init(&range_regular_cond);
434 	if (rangelock_init(&range_regular) < 0)
435 		primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
436 	cv_init(&range_sync_cond);
437 	if (rangelock_init(&range_sync) < 0)
438 		primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
439 	mapsize = activemap_ondisk_size(res->hr_amp);
440 	buf = calloc(1, mapsize);
441 	if (buf == NULL) {
442 		primary_exitx(EX_TEMPFAIL,
443 		    "Unable to allocate buffer for activemap.");
444 	}
445 	if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
446 	    (ssize_t)mapsize) {
447 		primary_exit(EX_NOINPUT, "Unable to read activemap");
448 	}
449 	activemap_copyin(res->hr_amp, buf, mapsize);
450 	if (res->hr_resuid != 0)
451 		return;
452 	/*
453 	 * We're using provider for the first time, so we have to generate
454 	 * resource unique identifier and initialize local and remote counts.
455 	 */
456 	arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
457 	res->hr_primary_localcnt = 1;
458 	res->hr_primary_remotecnt = 0;
459 	if (metadata_write(res) < 0)
460 		exit(EX_NOINPUT);
461 }
462 
463 static bool
464 init_remote(struct hast_resource *res, struct proto_conn **inp,
465     struct proto_conn **outp)
466 {
467 	struct proto_conn *in, *out;
468 	struct nv *nvout, *nvin;
469 	const unsigned char *token;
470 	unsigned char *map;
471 	const char *errmsg;
472 	int32_t extentsize;
473 	int64_t datasize;
474 	uint32_t mapsize;
475 	size_t size;
476 
477 	assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
478 
479 	in = out = NULL;
480 
481 	/* Prepare outgoing connection with remote node. */
482 	if (proto_client(res->hr_remoteaddr, &out) < 0) {
483 		primary_exit(EX_OSERR, "Unable to create connection to %s",
484 		    res->hr_remoteaddr);
485 	}
486 	/* Try to connect, but accept failure. */
487 	if (proto_connect(out) < 0) {
488 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
489 		    res->hr_remoteaddr);
490 		goto close;
491 	}
492 	/*
493 	 * First handshake step.
494 	 * Setup outgoing connection with remote node.
495 	 */
496 	nvout = nv_alloc();
497 	nv_add_string(nvout, res->hr_name, "resource");
498 	if (nv_error(nvout) != 0) {
499 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
500 		    "Unable to allocate header for connection with %s",
501 		    res->hr_remoteaddr);
502 		nv_free(nvout);
503 		goto close;
504 	}
505 	if (hast_proto_send(res, out, nvout, NULL, 0) < 0) {
506 		pjdlog_errno(LOG_WARNING,
507 		    "Unable to send handshake header to %s",
508 		    res->hr_remoteaddr);
509 		nv_free(nvout);
510 		goto close;
511 	}
512 	nv_free(nvout);
513 	if (hast_proto_recv_hdr(out, &nvin) < 0) {
514 		pjdlog_errno(LOG_WARNING,
515 		    "Unable to receive handshake header from %s",
516 		    res->hr_remoteaddr);
517 		goto close;
518 	}
519 	errmsg = nv_get_string(nvin, "errmsg");
520 	if (errmsg != NULL) {
521 		pjdlog_warning("%s", errmsg);
522 		nv_free(nvin);
523 		goto close;
524 	}
525 	token = nv_get_uint8_array(nvin, &size, "token");
526 	if (token == NULL) {
527 		pjdlog_warning("Handshake header from %s has no 'token' field.",
528 		    res->hr_remoteaddr);
529 		nv_free(nvin);
530 		goto close;
531 	}
532 	if (size != sizeof(res->hr_token)) {
533 		pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
534 		    res->hr_remoteaddr, size, sizeof(res->hr_token));
535 		nv_free(nvin);
536 		goto close;
537 	}
538 	bcopy(token, res->hr_token, sizeof(res->hr_token));
539 	nv_free(nvin);
540 
541 	/*
542 	 * Second handshake step.
543 	 * Setup incoming connection with remote node.
544 	 */
545 	if (proto_client(res->hr_remoteaddr, &in) < 0) {
546 		pjdlog_errno(LOG_WARNING, "Unable to create connection to %s",
547 		    res->hr_remoteaddr);
548 	}
549 	/* Try to connect, but accept failure. */
550 	if (proto_connect(in) < 0) {
551 		pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
552 		    res->hr_remoteaddr);
553 		goto close;
554 	}
555 	nvout = nv_alloc();
556 	nv_add_string(nvout, res->hr_name, "resource");
557 	nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
558 	    "token");
559 	nv_add_uint64(nvout, res->hr_resuid, "resuid");
560 	nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
561 	nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
562 	if (nv_error(nvout) != 0) {
563 		pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
564 		    "Unable to allocate header for connection with %s",
565 		    res->hr_remoteaddr);
566 		nv_free(nvout);
567 		goto close;
568 	}
569 	if (hast_proto_send(res, in, nvout, NULL, 0) < 0) {
570 		pjdlog_errno(LOG_WARNING,
571 		    "Unable to send handshake header to %s",
572 		    res->hr_remoteaddr);
573 		nv_free(nvout);
574 		goto close;
575 	}
576 	nv_free(nvout);
577 	if (hast_proto_recv_hdr(out, &nvin) < 0) {
578 		pjdlog_errno(LOG_WARNING,
579 		    "Unable to receive handshake header from %s",
580 		    res->hr_remoteaddr);
581 		goto close;
582 	}
583 	errmsg = nv_get_string(nvin, "errmsg");
584 	if (errmsg != NULL) {
585 		pjdlog_warning("%s", errmsg);
586 		nv_free(nvin);
587 		goto close;
588 	}
589 	datasize = nv_get_int64(nvin, "datasize");
590 	if (datasize != res->hr_datasize) {
591 		pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
592 		    (intmax_t)res->hr_datasize, (intmax_t)datasize);
593 		nv_free(nvin);
594 		goto close;
595 	}
596 	extentsize = nv_get_int32(nvin, "extentsize");
597 	if (extentsize != res->hr_extentsize) {
598 		pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
599 		    (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
600 		nv_free(nvin);
601 		goto close;
602 	}
603 	res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
604 	res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
605 	res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
606 	map = NULL;
607 	mapsize = nv_get_uint32(nvin, "mapsize");
608 	if (mapsize > 0) {
609 		map = malloc(mapsize);
610 		if (map == NULL) {
611 			pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
612 			    (uintmax_t)mapsize);
613 			nv_free(nvin);
614 			goto close;
615 		}
616 		/*
617 		 * Remote node have some dirty extents on its own, lets
618 		 * download its activemap.
619 		 */
620 		if (hast_proto_recv_data(res, out, nvin, map,
621 		    mapsize) < 0) {
622 			pjdlog_errno(LOG_ERR,
623 			    "Unable to receive remote activemap");
624 			nv_free(nvin);
625 			free(map);
626 			goto close;
627 		}
628 		/*
629 		 * Merge local and remote bitmaps.
630 		 */
631 		activemap_merge(res->hr_amp, map, mapsize);
632 		free(map);
633 		/*
634 		 * Now that we merged bitmaps from both nodes, flush it to the
635 		 * disk before we start to synchronize.
636 		 */
637 		(void)hast_activemap_flush(res);
638 	}
639 	pjdlog_info("Connected to %s.", res->hr_remoteaddr);
640 	if (inp != NULL && outp != NULL) {
641 		*inp = in;
642 		*outp = out;
643 	} else {
644 		res->hr_remotein = in;
645 		res->hr_remoteout = out;
646 	}
647 	return (true);
648 close:
649 	proto_close(out);
650 	if (in != NULL)
651 		proto_close(in);
652 	return (false);
653 }
654 
655 static void
656 sync_start(void)
657 {
658 
659 	mtx_lock(&sync_lock);
660 	sync_inprogress = true;
661 	mtx_unlock(&sync_lock);
662 	cv_signal(&sync_cond);
663 }
664 
665 static void
666 init_ggate(struct hast_resource *res)
667 {
668 	struct g_gate_ctl_create ggiocreate;
669 	struct g_gate_ctl_cancel ggiocancel;
670 
671 	/*
672 	 * We communicate with ggate via /dev/ggctl. Open it.
673 	 */
674 	res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
675 	if (res->hr_ggatefd < 0)
676 		primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
677 	/*
678 	 * Create provider before trying to connect, as connection failure
679 	 * is not critical, but may take some time.
680 	 */
681 	ggiocreate.gctl_version = G_GATE_VERSION;
682 	ggiocreate.gctl_mediasize = res->hr_datasize;
683 	ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
684 	ggiocreate.gctl_flags = 0;
685 	ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE;
686 	ggiocreate.gctl_timeout = 0;
687 	ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
688 	snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
689 	    res->hr_provname);
690 	bzero(ggiocreate.gctl_info, sizeof(ggiocreate.gctl_info));
691 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
692 		pjdlog_info("Device hast/%s created.", res->hr_provname);
693 		res->hr_ggateunit = ggiocreate.gctl_unit;
694 		return;
695 	}
696 	if (errno != EEXIST) {
697 		primary_exit(EX_OSERR, "Unable to create hast/%s device",
698 		    res->hr_provname);
699 	}
700 	pjdlog_debug(1,
701 	    "Device hast/%s already exists, we will try to take it over.",
702 	    res->hr_provname);
703 	/*
704 	 * If we received EEXIST, we assume that the process who created the
705 	 * provider died and didn't clean up. In that case we will start from
706 	 * where he left of.
707 	 */
708 	ggiocancel.gctl_version = G_GATE_VERSION;
709 	ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
710 	snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
711 	    res->hr_provname);
712 	if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
713 		pjdlog_info("Device hast/%s recovered.", res->hr_provname);
714 		res->hr_ggateunit = ggiocancel.gctl_unit;
715 		return;
716 	}
717 	primary_exit(EX_OSERR, "Unable to take over hast/%s device",
718 	    res->hr_provname);
719 }
720 
721 void
722 hastd_primary(struct hast_resource *res)
723 {
724 	pthread_t td;
725 	pid_t pid;
726 	int error;
727 
728 	gres = res;
729 
730 	/*
731 	 * Create communication channel between parent and child.
732 	 */
733 	if (proto_client("socketpair://", &res->hr_ctrl) < 0) {
734 		KEEP_ERRNO((void)pidfile_remove(pfh));
735 		primary_exit(EX_OSERR,
736 		    "Unable to create control sockets between parent and child");
737 	}
738 
739 	pid = fork();
740 	if (pid < 0) {
741 		KEEP_ERRNO((void)pidfile_remove(pfh));
742 		primary_exit(EX_OSERR, "Unable to fork");
743 	}
744 
745 	if (pid > 0) {
746 		/* This is parent. */
747 		res->hr_workerpid = pid;
748 		return;
749 	}
750 	(void)pidfile_close(pfh);
751 
752 	setproctitle("%s (primary)", res->hr_name);
753 
754 	init_local(res);
755 	if (init_remote(res, NULL, NULL))
756 		sync_start();
757 	init_ggate(res);
758 	init_environment(res);
759 	error = pthread_create(&td, NULL, ggate_recv_thread, res);
760 	assert(error == 0);
761 	error = pthread_create(&td, NULL, local_send_thread, res);
762 	assert(error == 0);
763 	error = pthread_create(&td, NULL, remote_send_thread, res);
764 	assert(error == 0);
765 	error = pthread_create(&td, NULL, remote_recv_thread, res);
766 	assert(error == 0);
767 	error = pthread_create(&td, NULL, ggate_send_thread, res);
768 	assert(error == 0);
769 	error = pthread_create(&td, NULL, sync_thread, res);
770 	assert(error == 0);
771 	error = pthread_create(&td, NULL, ctrl_thread, res);
772 	assert(error == 0);
773 	(void)guard_thread(res);
774 }
775 
776 static void
777 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...)
778 {
779 	char msg[1024];
780 	va_list ap;
781 	int len;
782 
783 	va_start(ap, fmt);
784 	len = vsnprintf(msg, sizeof(msg), fmt, ap);
785 	va_end(ap);
786 	if ((size_t)len < sizeof(msg)) {
787 		switch (ggio->gctl_cmd) {
788 		case BIO_READ:
789 			(void)snprintf(msg + len, sizeof(msg) - len,
790 			    "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
791 			    (uintmax_t)ggio->gctl_length);
792 			break;
793 		case BIO_DELETE:
794 			(void)snprintf(msg + len, sizeof(msg) - len,
795 			    "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
796 			    (uintmax_t)ggio->gctl_length);
797 			break;
798 		case BIO_FLUSH:
799 			(void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
800 			break;
801 		case BIO_WRITE:
802 			(void)snprintf(msg + len, sizeof(msg) - len,
803 			    "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
804 			    (uintmax_t)ggio->gctl_length);
805 			break;
806 		default:
807 			(void)snprintf(msg + len, sizeof(msg) - len,
808 			    "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd);
809 			break;
810 		}
811 	}
812 	pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
813 }
814 
815 static void
816 remote_close(struct hast_resource *res, int ncomp)
817 {
818 
819 	rw_wlock(&hio_remote_lock[ncomp]);
820 	/*
821 	 * A race is possible between dropping rlock and acquiring wlock -
822 	 * another thread can close connection in-between.
823 	 */
824 	if (!ISCONNECTED(res, ncomp)) {
825 		assert(res->hr_remotein == NULL);
826 		assert(res->hr_remoteout == NULL);
827 		rw_unlock(&hio_remote_lock[ncomp]);
828 		return;
829 	}
830 
831 	assert(res->hr_remotein != NULL);
832 	assert(res->hr_remoteout != NULL);
833 
834 	pjdlog_debug(2, "Closing old incoming connection to %s.",
835 	    res->hr_remoteaddr);
836 	proto_close(res->hr_remotein);
837 	res->hr_remotein = NULL;
838 	pjdlog_debug(2, "Closing old outgoing connection to %s.",
839 	    res->hr_remoteaddr);
840 	proto_close(res->hr_remoteout);
841 	res->hr_remoteout = NULL;
842 
843 	rw_unlock(&hio_remote_lock[ncomp]);
844 
845 	/*
846 	 * Stop synchronization if in-progress.
847 	 */
848 	mtx_lock(&sync_lock);
849 	if (sync_inprogress)
850 		sync_inprogress = false;
851 	mtx_unlock(&sync_lock);
852 
853 	/*
854 	 * Wake up guard thread, so it can immediately start reconnect.
855 	 */
856 	mtx_lock(&hio_guard_lock);
857 	cv_signal(&hio_guard_cond);
858 	mtx_unlock(&hio_guard_lock);
859 }
860 
861 /*
862  * Thread receives ggate I/O requests from the kernel and passes them to
863  * appropriate threads:
864  * WRITE - always goes to both local_send and remote_send threads
865  * READ (when the block is up-to-date on local component) -
866  *	only local_send thread
867  * READ (when the block isn't up-to-date on local component) -
868  *	only remote_send thread
869  * DELETE - always goes to both local_send and remote_send threads
870  * FLUSH - always goes to both local_send and remote_send threads
871  */
872 static void *
873 ggate_recv_thread(void *arg)
874 {
875 	struct hast_resource *res = arg;
876 	struct g_gate_ctl_io *ggio;
877 	struct hio *hio;
878 	unsigned int ii, ncomp, ncomps;
879 	int error;
880 
881 	ncomps = HAST_NCOMPONENTS;
882 
883 	for (;;) {
884 		pjdlog_debug(2, "ggate_recv: Taking free request.");
885 		QUEUE_TAKE2(hio, free);
886 		pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
887 		ggio = &hio->hio_ggio;
888 		ggio->gctl_unit = res->hr_ggateunit;
889 		ggio->gctl_length = MAXPHYS;
890 		ggio->gctl_error = 0;
891 		pjdlog_debug(2,
892 		    "ggate_recv: (%p) Waiting for request from the kernel.",
893 		    hio);
894 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) {
895 			if (sigexit_received)
896 				pthread_exit(NULL);
897 			primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
898 		}
899 		error = ggio->gctl_error;
900 		switch (error) {
901 		case 0:
902 			break;
903 		case ECANCELED:
904 			/* Exit gracefully. */
905 			if (!sigexit_received) {
906 				pjdlog_debug(2,
907 				    "ggate_recv: (%p) Received cancel from the kernel.",
908 				    hio);
909 				pjdlog_info("Received cancel from the kernel, exiting.");
910 			}
911 			pthread_exit(NULL);
912 		case ENOMEM:
913 			/*
914 			 * Buffer too small? Impossible, we allocate MAXPHYS
915 			 * bytes - request can't be bigger than that.
916 			 */
917 			/* FALLTHROUGH */
918 		case ENXIO:
919 		default:
920 			primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
921 			    strerror(error));
922 		}
923 		for (ii = 0; ii < ncomps; ii++)
924 			hio->hio_errors[ii] = EINVAL;
925 		reqlog(LOG_DEBUG, 2, ggio,
926 		    "ggate_recv: (%p) Request received from the kernel: ",
927 		    hio);
928 		/*
929 		 * Inform all components about new write request.
930 		 * For read request prefer local component unless the given
931 		 * range is out-of-date, then use remote component.
932 		 */
933 		switch (ggio->gctl_cmd) {
934 		case BIO_READ:
935 			pjdlog_debug(2,
936 			    "ggate_recv: (%p) Moving request to the send queue.",
937 			    hio);
938 			refcount_init(&hio->hio_countdown, 1);
939 			mtx_lock(&metadata_lock);
940 			if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
941 			    res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
942 				/*
943 				 * This range is up-to-date on local component,
944 				 * so handle request locally.
945 				 */
946 				 /* Local component is 0 for now. */
947 				ncomp = 0;
948 			} else /* if (res->hr_syncsrc ==
949 			    HAST_SYNCSRC_SECONDARY) */ {
950 				assert(res->hr_syncsrc ==
951 				    HAST_SYNCSRC_SECONDARY);
952 				/*
953 				 * This range is out-of-date on local component,
954 				 * so send request to the remote node.
955 				 */
956 				 /* Remote component is 1 for now. */
957 				ncomp = 1;
958 			}
959 			mtx_unlock(&metadata_lock);
960 			QUEUE_INSERT1(hio, send, ncomp);
961 			break;
962 		case BIO_WRITE:
963 			for (;;) {
964 				mtx_lock(&range_lock);
965 				if (rangelock_islocked(range_sync,
966 				    ggio->gctl_offset, ggio->gctl_length)) {
967 					pjdlog_debug(2,
968 					    "regular: Range offset=%jd length=%zu locked.",
969 					    (intmax_t)ggio->gctl_offset,
970 					    (size_t)ggio->gctl_length);
971 					range_regular_wait = true;
972 					cv_wait(&range_regular_cond, &range_lock);
973 					range_regular_wait = false;
974 					mtx_unlock(&range_lock);
975 					continue;
976 				}
977 				if (rangelock_add(range_regular,
978 				    ggio->gctl_offset, ggio->gctl_length) < 0) {
979 					mtx_unlock(&range_lock);
980 					pjdlog_debug(2,
981 					    "regular: Range offset=%jd length=%zu is already locked, waiting.",
982 					    (intmax_t)ggio->gctl_offset,
983 					    (size_t)ggio->gctl_length);
984 					sleep(1);
985 					continue;
986 				}
987 				mtx_unlock(&range_lock);
988 				break;
989 			}
990 			mtx_lock(&res->hr_amp_lock);
991 			if (activemap_write_start(res->hr_amp,
992 			    ggio->gctl_offset, ggio->gctl_length)) {
993 				(void)hast_activemap_flush(res);
994 			}
995 			mtx_unlock(&res->hr_amp_lock);
996 			/* FALLTHROUGH */
997 		case BIO_DELETE:
998 		case BIO_FLUSH:
999 			pjdlog_debug(2,
1000 			    "ggate_recv: (%p) Moving request to the send queues.",
1001 			    hio);
1002 			refcount_init(&hio->hio_countdown, ncomps);
1003 			for (ii = 0; ii < ncomps; ii++)
1004 				QUEUE_INSERT1(hio, send, ii);
1005 			break;
1006 		}
1007 	}
1008 	/* NOTREACHED */
1009 	return (NULL);
1010 }
1011 
1012 /*
1013  * Thread reads from or writes to local component.
1014  * If local read fails, it redirects it to remote_send thread.
1015  */
1016 static void *
1017 local_send_thread(void *arg)
1018 {
1019 	struct hast_resource *res = arg;
1020 	struct g_gate_ctl_io *ggio;
1021 	struct hio *hio;
1022 	unsigned int ncomp, rncomp;
1023 	ssize_t ret;
1024 
1025 	/* Local component is 0 for now. */
1026 	ncomp = 0;
1027 	/* Remote component is 1 for now. */
1028 	rncomp = 1;
1029 
1030 	for (;;) {
1031 		pjdlog_debug(2, "local_send: Taking request.");
1032 		QUEUE_TAKE1(hio, send, ncomp);
1033 		pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1034 		ggio = &hio->hio_ggio;
1035 		switch (ggio->gctl_cmd) {
1036 		case BIO_READ:
1037 			ret = pread(res->hr_localfd, ggio->gctl_data,
1038 			    ggio->gctl_length,
1039 			    ggio->gctl_offset + res->hr_localoff);
1040 			if (ret == ggio->gctl_length)
1041 				hio->hio_errors[ncomp] = 0;
1042 			else {
1043 				/*
1044 				 * If READ failed, try to read from remote node.
1045 				 */
1046 				QUEUE_INSERT1(hio, send, rncomp);
1047 				continue;
1048 			}
1049 			break;
1050 		case BIO_WRITE:
1051 			ret = pwrite(res->hr_localfd, ggio->gctl_data,
1052 			    ggio->gctl_length,
1053 			    ggio->gctl_offset + res->hr_localoff);
1054 			if (ret < 0)
1055 				hio->hio_errors[ncomp] = errno;
1056 			else if (ret != ggio->gctl_length)
1057 				hio->hio_errors[ncomp] = EIO;
1058 			else
1059 				hio->hio_errors[ncomp] = 0;
1060 			break;
1061 		case BIO_DELETE:
1062 			ret = g_delete(res->hr_localfd,
1063 			    ggio->gctl_offset + res->hr_localoff,
1064 			    ggio->gctl_length);
1065 			if (ret < 0)
1066 				hio->hio_errors[ncomp] = errno;
1067 			else
1068 				hio->hio_errors[ncomp] = 0;
1069 			break;
1070 		case BIO_FLUSH:
1071 			ret = g_flush(res->hr_localfd);
1072 			if (ret < 0)
1073 				hio->hio_errors[ncomp] = errno;
1074 			else
1075 				hio->hio_errors[ncomp] = 0;
1076 			break;
1077 		}
1078 		if (refcount_release(&hio->hio_countdown)) {
1079 			if (ISSYNCREQ(hio)) {
1080 				mtx_lock(&sync_lock);
1081 				SYNCREQDONE(hio);
1082 				mtx_unlock(&sync_lock);
1083 				cv_signal(&sync_cond);
1084 			} else {
1085 				pjdlog_debug(2,
1086 				    "local_send: (%p) Moving request to the done queue.",
1087 				    hio);
1088 				QUEUE_INSERT2(hio, done);
1089 			}
1090 		}
1091 	}
1092 	/* NOTREACHED */
1093 	return (NULL);
1094 }
1095 
1096 /*
1097  * Thread sends request to secondary node.
1098  */
1099 static void *
1100 remote_send_thread(void *arg)
1101 {
1102 	struct hast_resource *res = arg;
1103 	struct g_gate_ctl_io *ggio;
1104 	struct hio *hio;
1105 	struct nv *nv;
1106 	unsigned int ncomp;
1107 	bool wakeup;
1108 	uint64_t offset, length;
1109 	uint8_t cmd;
1110 	void *data;
1111 
1112 	/* Remote component is 1 for now. */
1113 	ncomp = 1;
1114 
1115 	for (;;) {
1116 		pjdlog_debug(2, "remote_send: Taking request.");
1117 		QUEUE_TAKE1(hio, send, ncomp);
1118 		pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1119 		ggio = &hio->hio_ggio;
1120 		switch (ggio->gctl_cmd) {
1121 		case BIO_READ:
1122 			cmd = HIO_READ;
1123 			data = NULL;
1124 			offset = ggio->gctl_offset;
1125 			length = ggio->gctl_length;
1126 			break;
1127 		case BIO_WRITE:
1128 			cmd = HIO_WRITE;
1129 			data = ggio->gctl_data;
1130 			offset = ggio->gctl_offset;
1131 			length = ggio->gctl_length;
1132 			break;
1133 		case BIO_DELETE:
1134 			cmd = HIO_DELETE;
1135 			data = NULL;
1136 			offset = ggio->gctl_offset;
1137 			length = ggio->gctl_length;
1138 			break;
1139 		case BIO_FLUSH:
1140 			cmd = HIO_FLUSH;
1141 			data = NULL;
1142 			offset = 0;
1143 			length = 0;
1144 			break;
1145 		default:
1146 			assert(!"invalid condition");
1147 			abort();
1148 		}
1149 		nv = nv_alloc();
1150 		nv_add_uint8(nv, cmd, "cmd");
1151 		nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1152 		nv_add_uint64(nv, offset, "offset");
1153 		nv_add_uint64(nv, length, "length");
1154 		if (nv_error(nv) != 0) {
1155 			hio->hio_errors[ncomp] = nv_error(nv);
1156 			pjdlog_debug(2,
1157 			    "remote_send: (%p) Unable to prepare header to send.",
1158 			    hio);
1159 			reqlog(LOG_ERR, 0, ggio,
1160 			    "Unable to prepare header to send (%s): ",
1161 			    strerror(nv_error(nv)));
1162 			/* Move failed request immediately to the done queue. */
1163 			goto done_queue;
1164 		}
1165 		pjdlog_debug(2,
1166 		    "remote_send: (%p) Moving request to the recv queue.",
1167 		    hio);
1168 		/*
1169 		 * Protect connection from disappearing.
1170 		 */
1171 		rw_rlock(&hio_remote_lock[ncomp]);
1172 		if (!ISCONNECTED(res, ncomp)) {
1173 			rw_unlock(&hio_remote_lock[ncomp]);
1174 			hio->hio_errors[ncomp] = ENOTCONN;
1175 			goto done_queue;
1176 		}
1177 		/*
1178 		 * Move the request to recv queue before sending it, because
1179 		 * in different order we can get reply before we move request
1180 		 * to recv queue.
1181 		 */
1182 		mtx_lock(&hio_recv_list_lock[ncomp]);
1183 		wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1184 		TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1185 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1186 		if (hast_proto_send(res, res->hr_remoteout, nv, data,
1187 		    data != NULL ? length : 0) < 0) {
1188 			hio->hio_errors[ncomp] = errno;
1189 			rw_unlock(&hio_remote_lock[ncomp]);
1190 			remote_close(res, ncomp);
1191 			pjdlog_debug(2,
1192 			    "remote_send: (%p) Unable to send request.", hio);
1193 			reqlog(LOG_ERR, 0, ggio,
1194 			    "Unable to send request (%s): ",
1195 			    strerror(hio->hio_errors[ncomp]));
1196 			/*
1197 			 * Take request back from the receive queue and move
1198 			 * it immediately to the done queue.
1199 			 */
1200 			mtx_lock(&hio_recv_list_lock[ncomp]);
1201 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1202 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1203 			goto done_queue;
1204 		}
1205 		rw_unlock(&hio_remote_lock[ncomp]);
1206 		nv_free(nv);
1207 		if (wakeup)
1208 			cv_signal(&hio_recv_list_cond[ncomp]);
1209 		continue;
1210 done_queue:
1211 		nv_free(nv);
1212 		if (ISSYNCREQ(hio)) {
1213 			if (!refcount_release(&hio->hio_countdown))
1214 				continue;
1215 			mtx_lock(&sync_lock);
1216 			SYNCREQDONE(hio);
1217 			mtx_unlock(&sync_lock);
1218 			cv_signal(&sync_cond);
1219 			continue;
1220 		}
1221 		if (ggio->gctl_cmd == BIO_WRITE) {
1222 			mtx_lock(&res->hr_amp_lock);
1223 			if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1224 			    ggio->gctl_length)) {
1225 				(void)hast_activemap_flush(res);
1226 			}
1227 			mtx_unlock(&res->hr_amp_lock);
1228 		}
1229 		if (!refcount_release(&hio->hio_countdown))
1230 			continue;
1231 		pjdlog_debug(2,
1232 		    "remote_send: (%p) Moving request to the done queue.",
1233 		    hio);
1234 		QUEUE_INSERT2(hio, done);
1235 	}
1236 	/* NOTREACHED */
1237 	return (NULL);
1238 }
1239 
1240 /*
1241  * Thread receives answer from secondary node and passes it to ggate_send
1242  * thread.
1243  */
1244 static void *
1245 remote_recv_thread(void *arg)
1246 {
1247 	struct hast_resource *res = arg;
1248 	struct g_gate_ctl_io *ggio;
1249 	struct hio *hio;
1250 	struct nv *nv;
1251 	unsigned int ncomp;
1252 	uint64_t seq;
1253 	int error;
1254 
1255 	/* Remote component is 1 for now. */
1256 	ncomp = 1;
1257 
1258 	for (;;) {
1259 		/* Wait until there is anything to receive. */
1260 		mtx_lock(&hio_recv_list_lock[ncomp]);
1261 		while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1262 			pjdlog_debug(2, "remote_recv: No requests, waiting.");
1263 			cv_wait(&hio_recv_list_cond[ncomp],
1264 			    &hio_recv_list_lock[ncomp]);
1265 		}
1266 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1267 		rw_rlock(&hio_remote_lock[ncomp]);
1268 		if (!ISCONNECTED(res, ncomp)) {
1269 			rw_unlock(&hio_remote_lock[ncomp]);
1270 			/*
1271 			 * Connection is dead, so move all pending requests to
1272 			 * the done queue (one-by-one).
1273 			 */
1274 			mtx_lock(&hio_recv_list_lock[ncomp]);
1275 			hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1276 			assert(hio != NULL);
1277 			TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1278 			    hio_next[ncomp]);
1279 			mtx_unlock(&hio_recv_list_lock[ncomp]);
1280 			goto done_queue;
1281 		}
1282 		if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) {
1283 			pjdlog_errno(LOG_ERR,
1284 			    "Unable to receive reply header");
1285 			rw_unlock(&hio_remote_lock[ncomp]);
1286 			remote_close(res, ncomp);
1287 			continue;
1288 		}
1289 		rw_unlock(&hio_remote_lock[ncomp]);
1290 		seq = nv_get_uint64(nv, "seq");
1291 		if (seq == 0) {
1292 			pjdlog_error("Header contains no 'seq' field.");
1293 			nv_free(nv);
1294 			continue;
1295 		}
1296 		mtx_lock(&hio_recv_list_lock[ncomp]);
1297 		TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1298 			if (hio->hio_ggio.gctl_seq == seq) {
1299 				TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1300 				    hio_next[ncomp]);
1301 				break;
1302 			}
1303 		}
1304 		mtx_unlock(&hio_recv_list_lock[ncomp]);
1305 		if (hio == NULL) {
1306 			pjdlog_error("Found no request matching received 'seq' field (%ju).",
1307 			    (uintmax_t)seq);
1308 			nv_free(nv);
1309 			continue;
1310 		}
1311 		error = nv_get_int16(nv, "error");
1312 		if (error != 0) {
1313 			/* Request failed on remote side. */
1314 			hio->hio_errors[ncomp] = 0;
1315 			nv_free(nv);
1316 			goto done_queue;
1317 		}
1318 		ggio = &hio->hio_ggio;
1319 		switch (ggio->gctl_cmd) {
1320 		case BIO_READ:
1321 			rw_rlock(&hio_remote_lock[ncomp]);
1322 			if (!ISCONNECTED(res, ncomp)) {
1323 				rw_unlock(&hio_remote_lock[ncomp]);
1324 				nv_free(nv);
1325 				goto done_queue;
1326 			}
1327 			if (hast_proto_recv_data(res, res->hr_remotein, nv,
1328 			    ggio->gctl_data, ggio->gctl_length) < 0) {
1329 				hio->hio_errors[ncomp] = errno;
1330 				pjdlog_errno(LOG_ERR,
1331 				    "Unable to receive reply data");
1332 				rw_unlock(&hio_remote_lock[ncomp]);
1333 				nv_free(nv);
1334 				remote_close(res, ncomp);
1335 				goto done_queue;
1336 			}
1337 			rw_unlock(&hio_remote_lock[ncomp]);
1338 			break;
1339 		case BIO_WRITE:
1340 		case BIO_DELETE:
1341 		case BIO_FLUSH:
1342 			break;
1343 		default:
1344 			assert(!"invalid condition");
1345 			abort();
1346 		}
1347 		hio->hio_errors[ncomp] = 0;
1348 		nv_free(nv);
1349 done_queue:
1350 		if (refcount_release(&hio->hio_countdown)) {
1351 			if (ISSYNCREQ(hio)) {
1352 				mtx_lock(&sync_lock);
1353 				SYNCREQDONE(hio);
1354 				mtx_unlock(&sync_lock);
1355 				cv_signal(&sync_cond);
1356 			} else {
1357 				pjdlog_debug(2,
1358 				    "remote_recv: (%p) Moving request to the done queue.",
1359 				    hio);
1360 				QUEUE_INSERT2(hio, done);
1361 			}
1362 		}
1363 	}
1364 	/* NOTREACHED */
1365 	return (NULL);
1366 }
1367 
1368 /*
1369  * Thread sends answer to the kernel.
1370  */
1371 static void *
1372 ggate_send_thread(void *arg)
1373 {
1374 	struct hast_resource *res = arg;
1375 	struct g_gate_ctl_io *ggio;
1376 	struct hio *hio;
1377 	unsigned int ii, ncomp, ncomps;
1378 
1379 	ncomps = HAST_NCOMPONENTS;
1380 
1381 	for (;;) {
1382 		pjdlog_debug(2, "ggate_send: Taking request.");
1383 		QUEUE_TAKE2(hio, done);
1384 		pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1385 		ggio = &hio->hio_ggio;
1386 		for (ii = 0; ii < ncomps; ii++) {
1387 			if (hio->hio_errors[ii] == 0) {
1388 				/*
1389 				 * One successful request is enough to declare
1390 				 * success.
1391 				 */
1392 				ggio->gctl_error = 0;
1393 				break;
1394 			}
1395 		}
1396 		if (ii == ncomps) {
1397 			/*
1398 			 * None of the requests were successful.
1399 			 * Use first error.
1400 			 */
1401 			ggio->gctl_error = hio->hio_errors[0];
1402 		}
1403 		if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1404 			mtx_lock(&res->hr_amp_lock);
1405 			activemap_write_complete(res->hr_amp,
1406 			    ggio->gctl_offset, ggio->gctl_length);
1407 			mtx_unlock(&res->hr_amp_lock);
1408 		}
1409 		if (ggio->gctl_cmd == BIO_WRITE) {
1410 			/*
1411 			 * Unlock range we locked.
1412 			 */
1413 			mtx_lock(&range_lock);
1414 			rangelock_del(range_regular, ggio->gctl_offset,
1415 			    ggio->gctl_length);
1416 			if (range_sync_wait)
1417 				cv_signal(&range_sync_cond);
1418 			mtx_unlock(&range_lock);
1419 			/*
1420 			 * Bump local count if this is first write after
1421 			 * connection failure with remote node.
1422 			 */
1423 			ncomp = 1;
1424 			rw_rlock(&hio_remote_lock[ncomp]);
1425 			if (!ISCONNECTED(res, ncomp)) {
1426 				mtx_lock(&metadata_lock);
1427 				if (res->hr_primary_localcnt ==
1428 				    res->hr_secondary_remotecnt) {
1429 					res->hr_primary_localcnt++;
1430 					pjdlog_debug(1,
1431 					    "Increasing localcnt to %ju.",
1432 					    (uintmax_t)res->hr_primary_localcnt);
1433 					(void)metadata_write(res);
1434 				}
1435 				mtx_unlock(&metadata_lock);
1436 			}
1437 			rw_unlock(&hio_remote_lock[ncomp]);
1438 		}
1439 		if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0)
1440 			primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1441 		pjdlog_debug(2,
1442 		    "ggate_send: (%p) Moving request to the free queue.", hio);
1443 		QUEUE_INSERT2(hio, free);
1444 	}
1445 	/* NOTREACHED */
1446 	return (NULL);
1447 }
1448 
1449 /*
1450  * Thread synchronize local and remote components.
1451  */
1452 static void *
1453 sync_thread(void *arg __unused)
1454 {
1455 	struct hast_resource *res = arg;
1456 	struct hio *hio;
1457 	struct g_gate_ctl_io *ggio;
1458 	unsigned int ii, ncomp, ncomps;
1459 	off_t offset, length, synced;
1460 	bool dorewind;
1461 	int syncext;
1462 
1463 	ncomps = HAST_NCOMPONENTS;
1464 	dorewind = true;
1465 	synced = 0;
1466 
1467 	for (;;) {
1468 		mtx_lock(&sync_lock);
1469 		while (!sync_inprogress) {
1470 			dorewind = true;
1471 			synced = 0;
1472 			cv_wait(&sync_cond, &sync_lock);
1473 		}
1474 		mtx_unlock(&sync_lock);
1475 		/*
1476 		 * Obtain offset at which we should synchronize.
1477 		 * Rewind synchronization if needed.
1478 		 */
1479 		mtx_lock(&res->hr_amp_lock);
1480 		if (dorewind)
1481 			activemap_sync_rewind(res->hr_amp);
1482 		offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1483 		if (syncext != -1) {
1484 			/*
1485 			 * We synchronized entire syncext extent, we can mark
1486 			 * it as clean now.
1487 			 */
1488 			if (activemap_extent_complete(res->hr_amp, syncext))
1489 				(void)hast_activemap_flush(res);
1490 		}
1491 		mtx_unlock(&res->hr_amp_lock);
1492 		if (dorewind) {
1493 			dorewind = false;
1494 			if (offset < 0)
1495 				pjdlog_info("Nodes are in sync.");
1496 			else {
1497 				pjdlog_info("Synchronization started. %ju bytes to go.",
1498 				    (uintmax_t)(res->hr_extentsize *
1499 				    activemap_ndirty(res->hr_amp)));
1500 			}
1501 		}
1502 		if (offset < 0) {
1503 			mtx_lock(&sync_lock);
1504 			sync_inprogress = false;
1505 			mtx_unlock(&sync_lock);
1506 			pjdlog_debug(1, "Nothing to synchronize.");
1507 			/*
1508 			 * Synchronization complete, make both localcnt and
1509 			 * remotecnt equal.
1510 			 */
1511 			ncomp = 1;
1512 			rw_rlock(&hio_remote_lock[ncomp]);
1513 			if (ISCONNECTED(res, ncomp)) {
1514 				if (synced > 0) {
1515 					pjdlog_info("Synchronization complete. "
1516 					    "%jd bytes synchronized.",
1517 					    (intmax_t)synced);
1518 				}
1519 				mtx_lock(&metadata_lock);
1520 				res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
1521 				res->hr_primary_localcnt =
1522 				    res->hr_secondary_localcnt;
1523 				res->hr_primary_remotecnt =
1524 				    res->hr_secondary_remotecnt;
1525 				pjdlog_debug(1,
1526 				    "Setting localcnt to %ju and remotecnt to %ju.",
1527 				    (uintmax_t)res->hr_primary_localcnt,
1528 				    (uintmax_t)res->hr_secondary_localcnt);
1529 				(void)metadata_write(res);
1530 				mtx_unlock(&metadata_lock);
1531 			} else if (synced > 0) {
1532 				pjdlog_info("Synchronization interrupted. "
1533 				    "%jd bytes synchronized so far.",
1534 				    (intmax_t)synced);
1535 			}
1536 			rw_unlock(&hio_remote_lock[ncomp]);
1537 			continue;
1538 		}
1539 		pjdlog_debug(2, "sync: Taking free request.");
1540 		QUEUE_TAKE2(hio, free);
1541 		pjdlog_debug(2, "sync: (%p) Got free request.", hio);
1542 		/*
1543 		 * Lock the range we are going to synchronize. We don't want
1544 		 * race where someone writes between our read and write.
1545 		 */
1546 		for (;;) {
1547 			mtx_lock(&range_lock);
1548 			if (rangelock_islocked(range_regular, offset, length)) {
1549 				pjdlog_debug(2,
1550 				    "sync: Range offset=%jd length=%jd locked.",
1551 				    (intmax_t)offset, (intmax_t)length);
1552 				range_sync_wait = true;
1553 				cv_wait(&range_sync_cond, &range_lock);
1554 				range_sync_wait = false;
1555 				mtx_unlock(&range_lock);
1556 				continue;
1557 			}
1558 			if (rangelock_add(range_sync, offset, length) < 0) {
1559 				mtx_unlock(&range_lock);
1560 				pjdlog_debug(2,
1561 				    "sync: Range offset=%jd length=%jd is already locked, waiting.",
1562 				    (intmax_t)offset, (intmax_t)length);
1563 				sleep(1);
1564 				continue;
1565 			}
1566 			mtx_unlock(&range_lock);
1567 			break;
1568 		}
1569 		/*
1570 		 * First read the data from synchronization source.
1571 		 */
1572 		SYNCREQ(hio);
1573 		ggio = &hio->hio_ggio;
1574 		ggio->gctl_cmd = BIO_READ;
1575 		ggio->gctl_offset = offset;
1576 		ggio->gctl_length = length;
1577 		ggio->gctl_error = 0;
1578 		for (ii = 0; ii < ncomps; ii++)
1579 			hio->hio_errors[ii] = EINVAL;
1580 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1581 		    hio);
1582 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1583 		    hio);
1584 		mtx_lock(&metadata_lock);
1585 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1586 			/*
1587 			 * This range is up-to-date on local component,
1588 			 * so handle request locally.
1589 			 */
1590 			 /* Local component is 0 for now. */
1591 			ncomp = 0;
1592 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1593 			assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1594 			/*
1595 			 * This range is out-of-date on local component,
1596 			 * so send request to the remote node.
1597 			 */
1598 			 /* Remote component is 1 for now. */
1599 			ncomp = 1;
1600 		}
1601 		mtx_unlock(&metadata_lock);
1602 		refcount_init(&hio->hio_countdown, 1);
1603 		QUEUE_INSERT1(hio, send, ncomp);
1604 
1605 		/*
1606 		 * Let's wait for READ to finish.
1607 		 */
1608 		mtx_lock(&sync_lock);
1609 		while (!ISSYNCREQDONE(hio))
1610 			cv_wait(&sync_cond, &sync_lock);
1611 		mtx_unlock(&sync_lock);
1612 
1613 		if (hio->hio_errors[ncomp] != 0) {
1614 			pjdlog_error("Unable to read synchronization data: %s.",
1615 			    strerror(hio->hio_errors[ncomp]));
1616 			goto free_queue;
1617 		}
1618 
1619 		/*
1620 		 * We read the data from synchronization source, now write it
1621 		 * to synchronization target.
1622 		 */
1623 		SYNCREQ(hio);
1624 		ggio->gctl_cmd = BIO_WRITE;
1625 		for (ii = 0; ii < ncomps; ii++)
1626 			hio->hio_errors[ii] = EINVAL;
1627 		reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1628 		    hio);
1629 		pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1630 		    hio);
1631 		mtx_lock(&metadata_lock);
1632 		if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1633 			/*
1634 			 * This range is up-to-date on local component,
1635 			 * so we update remote component.
1636 			 */
1637 			 /* Remote component is 1 for now. */
1638 			ncomp = 1;
1639 		} else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1640 			assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1641 			/*
1642 			 * This range is out-of-date on local component,
1643 			 * so we update it.
1644 			 */
1645 			 /* Local component is 0 for now. */
1646 			ncomp = 0;
1647 		}
1648 		mtx_unlock(&metadata_lock);
1649 
1650 		pjdlog_debug(2, "sync: (%p) Moving request to the send queues.",
1651 		    hio);
1652 		refcount_init(&hio->hio_countdown, 1);
1653 		QUEUE_INSERT1(hio, send, ncomp);
1654 
1655 		/*
1656 		 * Let's wait for WRITE to finish.
1657 		 */
1658 		mtx_lock(&sync_lock);
1659 		while (!ISSYNCREQDONE(hio))
1660 			cv_wait(&sync_cond, &sync_lock);
1661 		mtx_unlock(&sync_lock);
1662 
1663 		if (hio->hio_errors[ncomp] != 0) {
1664 			pjdlog_error("Unable to write synchronization data: %s.",
1665 			    strerror(hio->hio_errors[ncomp]));
1666 			goto free_queue;
1667 		}
1668 free_queue:
1669 		mtx_lock(&range_lock);
1670 		rangelock_del(range_sync, offset, length);
1671 		if (range_regular_wait)
1672 			cv_signal(&range_regular_cond);
1673 		mtx_unlock(&range_lock);
1674 
1675 		synced += length;
1676 
1677 		pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
1678 		    hio);
1679 		QUEUE_INSERT2(hio, free);
1680 	}
1681 	/* NOTREACHED */
1682 	return (NULL);
1683 }
1684 
1685 static void
1686 sighandler(int sig)
1687 {
1688 	bool unlock;
1689 
1690 	switch (sig) {
1691 	case SIGINT:
1692 	case SIGTERM:
1693 		sigexit_received = true;
1694 		break;
1695 	default:
1696 		assert(!"invalid condition");
1697 	}
1698 	/*
1699 	 * XXX: Racy, but if we cannot obtain hio_guard_lock here, we don't
1700 	 * want to risk deadlock.
1701 	 */
1702 	unlock = mtx_trylock(&hio_guard_lock);
1703 	cv_signal(&hio_guard_cond);
1704 	if (unlock)
1705 		mtx_unlock(&hio_guard_lock);
1706 }
1707 
1708 /*
1709  * Thread guards remote connections and reconnects when needed, handles
1710  * signals, etc.
1711  */
1712 static void *
1713 guard_thread(void *arg)
1714 {
1715 	struct hast_resource *res = arg;
1716 	struct proto_conn *in, *out;
1717 	unsigned int ii, ncomps;
1718 	int timeout;
1719 
1720 	ncomps = HAST_NCOMPONENTS;
1721 	/* The is only one remote component for now. */
1722 #define	ISREMOTE(no)	((no) == 1)
1723 
1724 	for (;;) {
1725 		if (sigexit_received) {
1726 			primary_exitx(EX_OK,
1727 			    "Termination signal received, exiting.");
1728 		}
1729 		/*
1730 		 * If all the connection will be fine, we will sleep until
1731 		 * someone wakes us up.
1732 		 * If any of the connections will be broken and we won't be
1733 		 * able to connect, we will sleep only for RECONNECT_SLEEP
1734 		 * seconds so we can retry soon.
1735 		 */
1736 		timeout = 0;
1737 		pjdlog_debug(2, "remote_guard: Checking connections.");
1738 		mtx_lock(&hio_guard_lock);
1739 		for (ii = 0; ii < ncomps; ii++) {
1740 			if (!ISREMOTE(ii))
1741 				continue;
1742 			rw_rlock(&hio_remote_lock[ii]);
1743 			if (ISCONNECTED(res, ii)) {
1744 				assert(res->hr_remotein != NULL);
1745 				assert(res->hr_remoteout != NULL);
1746 				rw_unlock(&hio_remote_lock[ii]);
1747 				pjdlog_debug(2,
1748 				    "remote_guard: Connection to %s is ok.",
1749 				    res->hr_remoteaddr);
1750 			} else {
1751 				assert(res->hr_remotein == NULL);
1752 				assert(res->hr_remoteout == NULL);
1753 				/*
1754 				 * Upgrade the lock. It doesn't have to be
1755 				 * atomic as no other thread can change
1756 				 * connection status from disconnected to
1757 				 * connected.
1758 				 */
1759 				rw_unlock(&hio_remote_lock[ii]);
1760 				pjdlog_debug(2,
1761 				    "remote_guard: Reconnecting to %s.",
1762 				    res->hr_remoteaddr);
1763 				in = out = NULL;
1764 				if (init_remote(res, &in, &out)) {
1765 					rw_wlock(&hio_remote_lock[ii]);
1766 					assert(res->hr_remotein == NULL);
1767 					assert(res->hr_remoteout == NULL);
1768 					assert(in != NULL && out != NULL);
1769 					res->hr_remotein = in;
1770 					res->hr_remoteout = out;
1771 					rw_unlock(&hio_remote_lock[ii]);
1772 					pjdlog_info("Successfully reconnected to %s.",
1773 					    res->hr_remoteaddr);
1774 					sync_start();
1775 				} else {
1776 					/* Both connections should be NULL. */
1777 					assert(res->hr_remotein == NULL);
1778 					assert(res->hr_remoteout == NULL);
1779 					assert(in == NULL && out == NULL);
1780 					pjdlog_debug(2,
1781 					    "remote_guard: Reconnect to %s failed.",
1782 					    res->hr_remoteaddr);
1783 					timeout = RECONNECT_SLEEP;
1784 				}
1785 			}
1786 		}
1787 		(void)cv_timedwait(&hio_guard_cond, &hio_guard_lock, timeout);
1788 		mtx_unlock(&hio_guard_lock);
1789 	}
1790 #undef	ISREMOTE
1791 	/* NOTREACHED */
1792 	return (NULL);
1793 }
1794