xref: /freebsd/sbin/hastd/secondary.c (revision 884a2a699669ec61e2366e3e358342dbc94be24a)
1 /*-
2  * Copyright (c) 2009-2010 The FreeBSD Foundation
3  * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org>
4  * All rights reserved.
5  *
6  * This software was developed by Pawel Jakub Dawidek under sponsorship from
7  * the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/param.h>
35 #include <sys/time.h>
36 #include <sys/bio.h>
37 #include <sys/disk.h>
38 #include <sys/stat.h>
39 
40 #include <err.h>
41 #include <errno.h>
42 #include <fcntl.h>
43 #include <libgeom.h>
44 #include <pthread.h>
45 #include <signal.h>
46 #include <stdint.h>
47 #include <stdio.h>
48 #include <string.h>
49 #include <sysexits.h>
50 #include <unistd.h>
51 
52 #include <activemap.h>
53 #include <nv.h>
54 #include <pjdlog.h>
55 
56 #include "control.h"
57 #include "event.h"
58 #include "hast.h"
59 #include "hast_proto.h"
60 #include "hastd.h"
61 #include "hooks.h"
62 #include "metadata.h"
63 #include "proto.h"
64 #include "subr.h"
65 #include "synch.h"
66 
67 struct hio {
68 	uint64_t	 hio_seq;
69 	int		 hio_error;
70 	struct nv	*hio_nv;
71 	void		*hio_data;
72 	uint8_t		 hio_cmd;
73 	uint64_t	 hio_offset;
74 	uint64_t	 hio_length;
75 	TAILQ_ENTRY(hio) hio_next;
76 };
77 
78 static struct hast_resource *gres;
79 
80 /*
81  * Free list holds unused structures. When free list is empty, we have to wait
82  * until some in-progress requests are freed.
83  */
84 static TAILQ_HEAD(, hio) hio_free_list;
85 static pthread_mutex_t hio_free_list_lock;
86 static pthread_cond_t hio_free_list_cond;
87 /*
88  * Disk thread (the one that do I/O requests) takes requests from this list.
89  */
90 static TAILQ_HEAD(, hio) hio_disk_list;
91 static pthread_mutex_t hio_disk_list_lock;
92 static pthread_cond_t hio_disk_list_cond;
93 /*
94  * There is one recv list for every component, although local components don't
95  * use recv lists as local requests are done synchronously.
96  */
97 static TAILQ_HEAD(, hio) hio_send_list;
98 static pthread_mutex_t hio_send_list_lock;
99 static pthread_cond_t hio_send_list_cond;
100 
101 /*
102  * Maximum number of outstanding I/O requests.
103  */
104 #define	HAST_HIO_MAX	256
105 
106 static void *recv_thread(void *arg);
107 static void *disk_thread(void *arg);
108 static void *send_thread(void *arg);
109 
110 #define	QUEUE_INSERT(name, hio)	do {					\
111 	bool _wakeup;							\
112 									\
113 	mtx_lock(&hio_##name##_list_lock);				\
114 	_wakeup = TAILQ_EMPTY(&hio_##name##_list);			\
115 	TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_next);		\
116 	mtx_unlock(&hio_##name##_list_lock);				\
117 	if (_wakeup)							\
118 		cv_signal(&hio_##name##_list_cond);			\
119 } while (0)
120 #define	QUEUE_TAKE(name, hio)	do {					\
121 	mtx_lock(&hio_##name##_list_lock);				\
122 	while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {	\
123 		cv_wait(&hio_##name##_list_cond,			\
124 		    &hio_##name##_list_lock);				\
125 	}								\
126 	TAILQ_REMOVE(&hio_##name##_list, (hio), hio_next);		\
127 	mtx_unlock(&hio_##name##_list_lock);				\
128 } while (0)
129 
130 static void
131 init_environment(void)
132 {
133 	struct hio *hio;
134 	unsigned int ii;
135 
136 	/*
137 	 * Initialize lists, their locks and theirs condition variables.
138 	 */
139 	TAILQ_INIT(&hio_free_list);
140 	mtx_init(&hio_free_list_lock);
141 	cv_init(&hio_free_list_cond);
142 	TAILQ_INIT(&hio_disk_list);
143 	mtx_init(&hio_disk_list_lock);
144 	cv_init(&hio_disk_list_cond);
145 	TAILQ_INIT(&hio_send_list);
146 	mtx_init(&hio_send_list_lock);
147 	cv_init(&hio_send_list_cond);
148 
149 	/*
150 	 * Allocate requests pool and initialize requests.
151 	 */
152 	for (ii = 0; ii < HAST_HIO_MAX; ii++) {
153 		hio = malloc(sizeof(*hio));
154 		if (hio == NULL) {
155 			pjdlog_exitx(EX_TEMPFAIL,
156 			    "Unable to allocate memory (%zu bytes) for hio request.",
157 			    sizeof(*hio));
158 		}
159 		hio->hio_error = 0;
160 		hio->hio_data = malloc(MAXPHYS);
161 		if (hio->hio_data == NULL) {
162 			pjdlog_exitx(EX_TEMPFAIL,
163 			    "Unable to allocate memory (%zu bytes) for gctl_data.",
164 			    (size_t)MAXPHYS);
165 		}
166 		TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_next);
167 	}
168 }
169 
170 static void
171 init_local(struct hast_resource *res)
172 {
173 
174 	if (metadata_read(res, true) < 0)
175 		exit(EX_NOINPUT);
176 }
177 
178 static void
179 init_remote(struct hast_resource *res, struct nv *nvin)
180 {
181 	uint64_t resuid;
182 	struct nv *nvout;
183 	unsigned char *map;
184 	size_t mapsize;
185 
186 	/* Setup direction. */
187 	if (proto_send(res->hr_remoteout, NULL, 0) == -1)
188 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
189 
190 	map = NULL;
191 	mapsize = 0;
192 	nvout = nv_alloc();
193 	nv_add_int64(nvout, (int64_t)res->hr_datasize, "datasize");
194 	nv_add_int32(nvout, (int32_t)res->hr_extentsize, "extentsize");
195 	resuid = nv_get_uint64(nvin, "resuid");
196 	res->hr_primary_localcnt = nv_get_uint64(nvin, "localcnt");
197 	res->hr_primary_remotecnt = nv_get_uint64(nvin, "remotecnt");
198 	nv_add_uint64(nvout, res->hr_secondary_localcnt, "localcnt");
199 	nv_add_uint64(nvout, res->hr_secondary_remotecnt, "remotecnt");
200 	mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize -
201 	    METADATA_SIZE, res->hr_extentsize, res->hr_local_sectorsize);
202 	map = malloc(mapsize);
203 	if (map == NULL) {
204 		pjdlog_exitx(EX_TEMPFAIL,
205 		    "Unable to allocate memory (%zu bytes) for activemap.",
206 		    mapsize);
207 	}
208 	/*
209 	 * When we work as primary and secondary is missing we will increase
210 	 * localcnt in our metadata. When secondary is connected and synced
211 	 * we make localcnt be equal to remotecnt, which means nodes are more
212 	 * or less in sync.
213 	 * Split-brain condition is when both nodes are not able to communicate
214 	 * and are both configured as primary nodes. In turn, they can both
215 	 * make incompatible changes to the data and we have to detect that.
216 	 * Under split-brain condition we will increase our localcnt on first
217 	 * write and remote node will increase its localcnt on first write.
218 	 * When we connect we can see that primary's localcnt is greater than
219 	 * our remotecnt (primary was modified while we weren't watching) and
220 	 * our localcnt is greater than primary's remotecnt (we were modified
221 	 * while primary wasn't watching).
222 	 * There are many possible combinations which are all gathered below.
223 	 * Don't pay too much attention to exact numbers, the more important
224 	 * is to compare them. We compare secondary's local with primary's
225 	 * remote and secondary's remote with primary's local.
226 	 * Note that every case where primary's localcnt is smaller than
227 	 * secondary's remotecnt and where secondary's localcnt is smaller than
228 	 * primary's remotecnt should be impossible in practise. We will perform
229 	 * full synchronization then. Those cases are marked with an asterisk.
230 	 * Regular synchronization means that only extents marked as dirty are
231 	 * synchronized (regular synchronization).
232 	 *
233 	 * SECONDARY METADATA PRIMARY METADATA
234 	 * local=3 remote=3   local=2 remote=2*  ?! Full sync from secondary.
235 	 * local=3 remote=3   local=2 remote=3*  ?! Full sync from primary.
236 	 * local=3 remote=3   local=2 remote=4*  ?! Full sync from primary.
237 	 * local=3 remote=3   local=3 remote=2   Primary is out-of-date,
238 	 *                                       regular sync from secondary.
239 	 * local=3 remote=3   local=3 remote=3   Regular sync just in case.
240 	 * local=3 remote=3   local=3 remote=4*  ?! Full sync from primary.
241 	 * local=3 remote=3   local=4 remote=2   Split-brain condition.
242 	 * local=3 remote=3   local=4 remote=3   Secondary out-of-date,
243 	 *                                       regular sync from primary.
244 	 * local=3 remote=3   local=4 remote=4*  ?! Full sync from primary.
245 	 */
246 	if (res->hr_resuid == 0) {
247 		/*
248 		 * Provider is used for the first time. If primary node done no
249 		 * writes yet as well (we will find "virgin" argument) then
250 		 * there is no need to synchronize anything. If primary node
251 		 * done any writes already we have to synchronize everything.
252 		 */
253 		PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
254 		res->hr_resuid = resuid;
255 		if (metadata_write(res) < 0)
256 			exit(EX_NOINPUT);
257 		if (nv_exists(nvin, "virgin")) {
258 			free(map);
259 			map = NULL;
260 			mapsize = 0;
261 		} else {
262 			memset(map, 0xff, mapsize);
263 		}
264 		nv_add_int8(nvout, 1, "virgin");
265 		nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc");
266 	} else if (res->hr_resuid != resuid) {
267 		char errmsg[256];
268 
269 		(void)snprintf(errmsg, sizeof(errmsg),
270 		    "Resource unique ID mismatch (primary=%ju, secondary=%ju).",
271 		    (uintmax_t)resuid, (uintmax_t)res->hr_resuid);
272 		pjdlog_error("%s", errmsg);
273 		nv_add_string(nvout, errmsg, "errmsg");
274 		if (hast_proto_send(res, res->hr_remotein, nvout, NULL, 0) < 0) {
275 			pjdlog_exit(EX_TEMPFAIL, "Unable to send response to %s",
276 			    res->hr_remoteaddr);
277 		}
278 		nv_free(nvout);
279 		exit(EX_CONFIG);
280 	} else if (
281 	    /* Is primary is out-of-date? */
282 	    (res->hr_secondary_localcnt > res->hr_primary_remotecnt &&
283 	     res->hr_secondary_remotecnt == res->hr_primary_localcnt) ||
284 	    /* Nodes are more or less in sync? */
285 	    (res->hr_secondary_localcnt == res->hr_primary_remotecnt &&
286 	     res->hr_secondary_remotecnt == res->hr_primary_localcnt) ||
287 	    /* Is secondary is out-of-date? */
288 	    (res->hr_secondary_localcnt == res->hr_primary_remotecnt &&
289 	     res->hr_secondary_remotecnt < res->hr_primary_localcnt)) {
290 		/*
291 		 * Nodes are more or less in sync or one of the nodes is
292 		 * out-of-date.
293 		 * It doesn't matter at this point which one, we just have to
294 		 * send out local bitmap to the remote node.
295 		 */
296 		if (pread(res->hr_localfd, map, mapsize, METADATA_SIZE) !=
297 		    (ssize_t)mapsize) {
298 			pjdlog_exit(LOG_ERR, "Unable to read activemap");
299 		}
300 		if (res->hr_secondary_localcnt > res->hr_primary_remotecnt &&
301 		     res->hr_secondary_remotecnt == res->hr_primary_localcnt) {
302 			/* Primary is out-of-date, sync from secondary. */
303 			nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc");
304 		} else {
305 			/*
306 			 * Secondary is out-of-date or counts match.
307 			 * Sync from primary.
308 			 */
309 			nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc");
310 		}
311 	} else if (res->hr_secondary_localcnt > res->hr_primary_remotecnt &&
312 	     res->hr_primary_localcnt > res->hr_secondary_remotecnt) {
313 		/*
314 		 * Not good, we have split-brain condition.
315 		 */
316 		pjdlog_error("Split-brain detected, exiting.");
317 		nv_add_string(nvout, "Split-brain condition!", "errmsg");
318 		free(map);
319 		map = NULL;
320 		mapsize = 0;
321 	} else /* if (res->hr_secondary_localcnt < res->hr_primary_remotecnt ||
322 	    res->hr_primary_localcnt < res->hr_secondary_remotecnt) */ {
323 		/*
324 		 * This should never happen in practise, but we will perform
325 		 * full synchronization.
326 		 */
327 		PJDLOG_ASSERT(res->hr_secondary_localcnt < res->hr_primary_remotecnt ||
328 		    res->hr_primary_localcnt < res->hr_secondary_remotecnt);
329 		mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize -
330 		    METADATA_SIZE, res->hr_extentsize,
331 		    res->hr_local_sectorsize);
332 		memset(map, 0xff, mapsize);
333 		if (res->hr_secondary_localcnt > res->hr_primary_remotecnt) {
334 			/* In this one of five cases sync from secondary. */
335 			nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc");
336 		} else {
337 			/* For the rest four cases sync from primary. */
338 			nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc");
339 		}
340 		pjdlog_warning("This should never happen, asking for full synchronization (primary(local=%ju, remote=%ju), secondary(local=%ju, remote=%ju)).",
341 		    (uintmax_t)res->hr_primary_localcnt,
342 		    (uintmax_t)res->hr_primary_remotecnt,
343 		    (uintmax_t)res->hr_secondary_localcnt,
344 		    (uintmax_t)res->hr_secondary_remotecnt);
345 	}
346 	nv_add_uint32(nvout, (uint32_t)mapsize, "mapsize");
347 	if (hast_proto_send(res, res->hr_remotein, nvout, map, mapsize) < 0) {
348 		pjdlog_exit(EX_TEMPFAIL, "Unable to send activemap to %s",
349 		    res->hr_remoteaddr);
350 	}
351 	if (map != NULL)
352 		free(map);
353 	nv_free(nvout);
354 	/* Setup direction. */
355 	if (proto_recv(res->hr_remotein, NULL, 0) == -1)
356 		pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
357 	if (res->hr_secondary_localcnt > res->hr_primary_remotecnt &&
358 	     res->hr_primary_localcnt > res->hr_secondary_remotecnt) {
359 		/* Exit on split-brain. */
360 		event_send(res, EVENT_SPLITBRAIN);
361 		exit(EX_CONFIG);
362 	}
363 }
364 
365 void
366 hastd_secondary(struct hast_resource *res, struct nv *nvin)
367 {
368 	sigset_t mask;
369 	pthread_t td;
370 	pid_t pid;
371 	int error, mode, debuglevel;
372 
373 	/*
374 	 * Create communication channel between parent and child.
375 	 */
376 	if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) {
377 		KEEP_ERRNO((void)pidfile_remove(pfh));
378 		pjdlog_exit(EX_OSERR,
379 		    "Unable to create control sockets between parent and child");
380 	}
381 	/*
382 	 * Create communication channel between child and parent.
383 	 */
384 	if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) {
385 		KEEP_ERRNO((void)pidfile_remove(pfh));
386 		pjdlog_exit(EX_OSERR,
387 		    "Unable to create event sockets between child and parent");
388 	}
389 
390 	pid = fork();
391 	if (pid < 0) {
392 		KEEP_ERRNO((void)pidfile_remove(pfh));
393 		pjdlog_exit(EX_OSERR, "Unable to fork");
394 	}
395 
396 	if (pid > 0) {
397 		/* This is parent. */
398 		proto_close(res->hr_remotein);
399 		res->hr_remotein = NULL;
400 		proto_close(res->hr_remoteout);
401 		res->hr_remoteout = NULL;
402 		/* Declare that we are receiver. */
403 		proto_recv(res->hr_event, NULL, 0);
404 		/* Declare that we are sender. */
405 		proto_send(res->hr_ctrl, NULL, 0);
406 		res->hr_workerpid = pid;
407 		return;
408 	}
409 
410 	gres = res;
411 	mode = pjdlog_mode_get();
412 	debuglevel = pjdlog_debug_get();
413 
414 	/* Declare that we are sender. */
415 	proto_send(res->hr_event, NULL, 0);
416 	/* Declare that we are receiver. */
417 	proto_recv(res->hr_ctrl, NULL, 0);
418 	descriptors_cleanup(res);
419 
420 	descriptors_assert(res, mode);
421 
422 	pjdlog_init(mode);
423 	pjdlog_debug_set(debuglevel);
424 	pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
425 	setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
426 
427 	PJDLOG_VERIFY(sigemptyset(&mask) == 0);
428 	PJDLOG_VERIFY(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
429 
430 	/* Error in setting timeout is not critical, but why should it fail? */
431 	if (proto_timeout(res->hr_remotein, 2 * HAST_KEEPALIVE) < 0)
432 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
433 	if (proto_timeout(res->hr_remoteout, res->hr_timeout) < 0)
434 		pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
435 
436 	init_local(res);
437 	init_environment();
438 
439 	if (drop_privs(res) != 0)
440 		exit(EX_CONFIG);
441 	pjdlog_info("Privileges successfully dropped.");
442 
443 	/*
444 	 * Create the control thread before sending any event to the parent,
445 	 * as we can deadlock when parent sends control request to worker,
446 	 * but worker has no control thread started yet, so parent waits.
447 	 * In the meantime worker sends an event to the parent, but parent
448 	 * is unable to handle the event, because it waits for control
449 	 * request response.
450 	 */
451 	error = pthread_create(&td, NULL, ctrl_thread, res);
452 	PJDLOG_ASSERT(error == 0);
453 
454 	init_remote(res, nvin);
455 	event_send(res, EVENT_CONNECT);
456 
457 	error = pthread_create(&td, NULL, recv_thread, res);
458 	PJDLOG_ASSERT(error == 0);
459 	error = pthread_create(&td, NULL, disk_thread, res);
460 	PJDLOG_ASSERT(error == 0);
461 	(void)send_thread(res);
462 }
463 
464 static void
465 reqlog(int loglevel, int debuglevel, int error, struct hio *hio, const char *fmt, ...)
466 {
467 	char msg[1024];
468 	va_list ap;
469 	int len;
470 
471 	va_start(ap, fmt);
472 	len = vsnprintf(msg, sizeof(msg), fmt, ap);
473 	va_end(ap);
474 	if ((size_t)len < sizeof(msg)) {
475 		switch (hio->hio_cmd) {
476 		case HIO_READ:
477 			(void)snprintf(msg + len, sizeof(msg) - len,
478 			    "READ(%ju, %ju).", (uintmax_t)hio->hio_offset,
479 			    (uintmax_t)hio->hio_length);
480 			break;
481 		case HIO_DELETE:
482 			(void)snprintf(msg + len, sizeof(msg) - len,
483 			    "DELETE(%ju, %ju).", (uintmax_t)hio->hio_offset,
484 			    (uintmax_t)hio->hio_length);
485 			break;
486 		case HIO_FLUSH:
487 			(void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
488 			break;
489 		case HIO_WRITE:
490 			(void)snprintf(msg + len, sizeof(msg) - len,
491 			    "WRITE(%ju, %ju).", (uintmax_t)hio->hio_offset,
492 			    (uintmax_t)hio->hio_length);
493 			break;
494 		case HIO_KEEPALIVE:
495 			(void)snprintf(msg + len, sizeof(msg) - len, "KEEPALIVE.");
496 			break;
497 		default:
498 			(void)snprintf(msg + len, sizeof(msg) - len,
499 			    "UNKNOWN(%u).", (unsigned int)hio->hio_cmd);
500 			break;
501 		}
502 	}
503 	pjdlog_common(loglevel, debuglevel, error, "%s", msg);
504 }
505 
506 static int
507 requnpack(struct hast_resource *res, struct hio *hio)
508 {
509 
510 	hio->hio_cmd = nv_get_uint8(hio->hio_nv, "cmd");
511 	if (hio->hio_cmd == 0) {
512 		pjdlog_error("Header contains no 'cmd' field.");
513 		hio->hio_error = EINVAL;
514 		goto end;
515 	}
516 	switch (hio->hio_cmd) {
517 	case HIO_FLUSH:
518 	case HIO_KEEPALIVE:
519 		break;
520 	case HIO_READ:
521 	case HIO_WRITE:
522 	case HIO_DELETE:
523 		hio->hio_offset = nv_get_uint64(hio->hio_nv, "offset");
524 		if (nv_error(hio->hio_nv) != 0) {
525 			pjdlog_error("Header is missing 'offset' field.");
526 			hio->hio_error = EINVAL;
527 			goto end;
528 		}
529 		hio->hio_length = nv_get_uint64(hio->hio_nv, "length");
530 		if (nv_error(hio->hio_nv) != 0) {
531 			pjdlog_error("Header is missing 'length' field.");
532 			hio->hio_error = EINVAL;
533 			goto end;
534 		}
535 		if (hio->hio_length == 0) {
536 			pjdlog_error("Data length is zero.");
537 			hio->hio_error = EINVAL;
538 			goto end;
539 		}
540 		if (hio->hio_length > MAXPHYS) {
541 			pjdlog_error("Data length is too large (%ju > %ju).",
542 			    (uintmax_t)hio->hio_length, (uintmax_t)MAXPHYS);
543 			hio->hio_error = EINVAL;
544 			goto end;
545 		}
546 		if ((hio->hio_offset % res->hr_local_sectorsize) != 0) {
547 			pjdlog_error("Offset %ju is not multiple of sector size.",
548 			    (uintmax_t)hio->hio_offset);
549 			hio->hio_error = EINVAL;
550 			goto end;
551 		}
552 		if ((hio->hio_length % res->hr_local_sectorsize) != 0) {
553 			pjdlog_error("Length %ju is not multiple of sector size.",
554 			    (uintmax_t)hio->hio_length);
555 			hio->hio_error = EINVAL;
556 			goto end;
557 		}
558 		if (hio->hio_offset + hio->hio_length >
559 		    (uint64_t)res->hr_datasize) {
560 			pjdlog_error("Data offset is too large (%ju > %ju).",
561 			    (uintmax_t)(hio->hio_offset + hio->hio_length),
562 			    (uintmax_t)res->hr_datasize);
563 			hio->hio_error = EINVAL;
564 			goto end;
565 		}
566 		break;
567 	default:
568 		pjdlog_error("Header contains invalid 'cmd' (%hhu).",
569 		    hio->hio_cmd);
570 		hio->hio_error = EINVAL;
571 		goto end;
572 	}
573 	hio->hio_error = 0;
574 end:
575 	return (hio->hio_error);
576 }
577 
578 static __dead2 void
579 secondary_exit(int exitcode, const char *fmt, ...)
580 {
581 	va_list ap;
582 
583 	PJDLOG_ASSERT(exitcode != EX_OK);
584 	va_start(ap, fmt);
585 	pjdlogv_errno(LOG_ERR, fmt, ap);
586 	va_end(ap);
587 	event_send(gres, EVENT_DISCONNECT);
588 	exit(exitcode);
589 }
590 
591 /*
592  * Thread receives requests from the primary node.
593  */
594 static void *
595 recv_thread(void *arg)
596 {
597 	struct hast_resource *res = arg;
598 	struct hio *hio;
599 
600 	for (;;) {
601 		pjdlog_debug(2, "recv: Taking free request.");
602 		QUEUE_TAKE(free, hio);
603 		pjdlog_debug(2, "recv: (%p) Got request.", hio);
604 		if (hast_proto_recv_hdr(res->hr_remotein, &hio->hio_nv) < 0) {
605 			secondary_exit(EX_TEMPFAIL,
606 			    "Unable to receive request header");
607 		}
608 		if (requnpack(res, hio) != 0) {
609 			pjdlog_debug(2,
610 			    "recv: (%p) Moving request to the send queue.",
611 			    hio);
612 			QUEUE_INSERT(send, hio);
613 			continue;
614 		}
615 		switch (hio->hio_cmd) {
616 		case HIO_READ:
617 			res->hr_stat_read++;
618 			break;
619 		case HIO_WRITE:
620 			res->hr_stat_write++;
621 			break;
622 		case HIO_DELETE:
623 			res->hr_stat_delete++;
624 			break;
625 		case HIO_FLUSH:
626 			res->hr_stat_flush++;
627 			break;
628 		}
629 		reqlog(LOG_DEBUG, 2, -1, hio,
630 		    "recv: (%p) Got request header: ", hio);
631 		if (hio->hio_cmd == HIO_KEEPALIVE) {
632 			pjdlog_debug(2,
633 			    "recv: (%p) Moving request to the free queue.",
634 			    hio);
635 			nv_free(hio->hio_nv);
636 			QUEUE_INSERT(free, hio);
637 			continue;
638 		} else if (hio->hio_cmd == HIO_WRITE) {
639 			if (hast_proto_recv_data(res, res->hr_remotein,
640 			    hio->hio_nv, hio->hio_data, MAXPHYS) < 0) {
641 				secondary_exit(EX_TEMPFAIL,
642 				    "Unable to receive request data");
643 			}
644 		}
645 		pjdlog_debug(2, "recv: (%p) Moving request to the disk queue.",
646 		    hio);
647 		QUEUE_INSERT(disk, hio);
648 	}
649 	/* NOTREACHED */
650 	return (NULL);
651 }
652 
653 /*
654  * Thread reads from or writes to local component and also handles DELETE and
655  * FLUSH requests.
656  */
657 static void *
658 disk_thread(void *arg)
659 {
660 	struct hast_resource *res = arg;
661 	struct hio *hio;
662 	ssize_t ret;
663 	bool clear_activemap;
664 
665 	clear_activemap = true;
666 
667 	for (;;) {
668 		pjdlog_debug(2, "disk: Taking request.");
669 		QUEUE_TAKE(disk, hio);
670 		while (clear_activemap) {
671 			unsigned char *map;
672 			size_t mapsize;
673 
674 			/*
675 			 * When first request is received, it means that primary
676 			 * already received our activemap, merged it and stored
677 			 * locally. We can now safely clear our activemap.
678 			 */
679 			mapsize =
680 			    activemap_calc_ondisk_size(res->hr_local_mediasize -
681 			    METADATA_SIZE, res->hr_extentsize,
682 			    res->hr_local_sectorsize);
683 			map = calloc(1, mapsize);
684 			if (map == NULL) {
685 				pjdlog_warning("Unable to allocate memory to clear local activemap.");
686 				break;
687 			}
688 			if (pwrite(res->hr_localfd, map, mapsize,
689 			    METADATA_SIZE) != (ssize_t)mapsize) {
690 				pjdlog_errno(LOG_WARNING,
691 				    "Unable to store cleared activemap");
692 				free(map);
693 				break;
694 			}
695 			free(map);
696 			clear_activemap = false;
697 			pjdlog_debug(1, "Local activemap cleared.");
698 		}
699 		reqlog(LOG_DEBUG, 2, -1, hio, "disk: (%p) Got request: ", hio);
700 		/* Handle the actual request. */
701 		switch (hio->hio_cmd) {
702 		case HIO_READ:
703 			ret = pread(res->hr_localfd, hio->hio_data,
704 			    hio->hio_length,
705 			    hio->hio_offset + res->hr_localoff);
706 			if (ret < 0)
707 				hio->hio_error = errno;
708 			else if (ret != (int64_t)hio->hio_length)
709 				hio->hio_error = EIO;
710 			else
711 				hio->hio_error = 0;
712 			break;
713 		case HIO_WRITE:
714 			ret = pwrite(res->hr_localfd, hio->hio_data,
715 			    hio->hio_length,
716 			    hio->hio_offset + res->hr_localoff);
717 			if (ret < 0)
718 				hio->hio_error = errno;
719 			else if (ret != (int64_t)hio->hio_length)
720 				hio->hio_error = EIO;
721 			else
722 				hio->hio_error = 0;
723 			break;
724 		case HIO_DELETE:
725 			ret = g_delete(res->hr_localfd,
726 			    hio->hio_offset + res->hr_localoff,
727 			    hio->hio_length);
728 			if (ret < 0)
729 				hio->hio_error = errno;
730 			else
731 				hio->hio_error = 0;
732 			break;
733 		case HIO_FLUSH:
734 			ret = g_flush(res->hr_localfd);
735 			if (ret < 0)
736 				hio->hio_error = errno;
737 			else
738 				hio->hio_error = 0;
739 			break;
740 		}
741 		if (hio->hio_error != 0) {
742 			reqlog(LOG_ERR, 0, hio->hio_error, hio,
743 			    "Request failed: ");
744 		}
745 		pjdlog_debug(2, "disk: (%p) Moving request to the send queue.",
746 		    hio);
747 		QUEUE_INSERT(send, hio);
748 	}
749 	/* NOTREACHED */
750 	return (NULL);
751 }
752 
753 /*
754  * Thread sends requests back to primary node.
755  */
756 static void *
757 send_thread(void *arg)
758 {
759 	struct hast_resource *res = arg;
760 	struct nv *nvout;
761 	struct hio *hio;
762 	void *data;
763 	size_t length;
764 
765 	for (;;) {
766 		pjdlog_debug(2, "send: Taking request.");
767 		QUEUE_TAKE(send, hio);
768 		reqlog(LOG_DEBUG, 2, -1, hio, "send: (%p) Got request: ", hio);
769 		nvout = nv_alloc();
770 		/* Copy sequence number. */
771 		nv_add_uint64(nvout, nv_get_uint64(hio->hio_nv, "seq"), "seq");
772 		switch (hio->hio_cmd) {
773 		case HIO_READ:
774 			if (hio->hio_error == 0) {
775 				data = hio->hio_data;
776 				length = hio->hio_length;
777 				break;
778 			}
779 			/*
780 			 * We send no data in case of an error.
781 			 */
782 			/* FALLTHROUGH */
783 		case HIO_DELETE:
784 		case HIO_FLUSH:
785 		case HIO_WRITE:
786 			data = NULL;
787 			length = 0;
788 			break;
789 		default:
790 			abort();
791 			break;
792 		}
793 		if (hio->hio_error != 0)
794 			nv_add_int16(nvout, hio->hio_error, "error");
795 		if (hast_proto_send(res, res->hr_remoteout, nvout, data,
796 		    length) < 0) {
797 			secondary_exit(EX_TEMPFAIL, "Unable to send reply.");
798 		}
799 		nv_free(nvout);
800 		pjdlog_debug(2, "send: (%p) Moving request to the free queue.",
801 		    hio);
802 		nv_free(hio->hio_nv);
803 		hio->hio_error = 0;
804 		QUEUE_INSERT(free, hio);
805 	}
806 	/* NOTREACHED */
807 	return (NULL);
808 }
809