2 * Copyright (c) 2009 The FreeBSD Foundation
3 * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org>
6 * This software was developed by Pawel Jakub Dawidek under sponsorship from
7 * the FreeBSD Foundation.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
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.
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
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
34 #include <sys/types.h>
38 #include <sys/refcount.h>
41 #include <geom/gate/g_gate.h>
56 #include <activemap.h>
58 #include <rangelock.h>
63 #include "hast_proto.h"
72 /* The is only one remote component for now. */
73 #define ISREMOTE(no) ((no) == 1)
77 * Number of components we are still waiting for.
78 * When this field goes to 0, we can send the request back to the
79 * kernel. Each component has to decrease this counter by one
82 unsigned int hio_countdown;
84 * Each component has a place to store its own error.
85 * Once the request is handled by all components we can decide if the
86 * request overall is successful or not.
90 * Structure used to comunicate with GEOM Gate class.
92 struct g_gate_ctl_io hio_ggio;
93 TAILQ_ENTRY(hio) *hio_next;
95 #define hio_free_next hio_next[0]
96 #define hio_done_next hio_next[0]
99 * Free list holds unused structures. When free list is empty, we have to wait
100 * until some in-progress requests are freed.
102 static TAILQ_HEAD(, hio) hio_free_list;
103 static pthread_mutex_t hio_free_list_lock;
104 static pthread_cond_t hio_free_list_cond;
106 * There is one send list for every component. One requests is placed on all
107 * send lists - each component gets the same request, but each component is
108 * responsible for managing his own send list.
110 static TAILQ_HEAD(, hio) *hio_send_list;
111 static pthread_mutex_t *hio_send_list_lock;
112 static pthread_cond_t *hio_send_list_cond;
114 * There is one recv list for every component, although local components don't
115 * use recv lists as local requests are done synchronously.
117 static TAILQ_HEAD(, hio) *hio_recv_list;
118 static pthread_mutex_t *hio_recv_list_lock;
119 static pthread_cond_t *hio_recv_list_cond;
121 * Request is placed on done list by the slowest component (the one that
122 * decreased hio_countdown from 1 to 0).
124 static TAILQ_HEAD(, hio) hio_done_list;
125 static pthread_mutex_t hio_done_list_lock;
126 static pthread_cond_t hio_done_list_cond;
128 * Structure below are for interaction with sync thread.
130 static bool sync_inprogress;
131 static pthread_mutex_t sync_lock;
132 static pthread_cond_t sync_cond;
134 * The lock below allows to synchornize access to remote connections.
136 static pthread_rwlock_t *hio_remote_lock;
139 * Lock to synchronize metadata updates. Also synchronize access to
140 * hr_primary_localcnt and hr_primary_remotecnt fields.
142 static pthread_mutex_t metadata_lock;
145 * Maximum number of outstanding I/O requests.
147 #define HAST_HIO_MAX 256
149 * Number of components. At this point there are only two components: local
150 * and remote, but in the future it might be possible to use multiple local
151 * and remote components.
153 #define HAST_NCOMPONENTS 2
155 * Number of seconds to sleep between reconnect retries or keepalive packets.
157 #define RETRY_SLEEP 10
159 #define ISCONNECTED(res, no) \
160 ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
162 #define QUEUE_INSERT1(hio, name, ncomp) do { \
165 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \
166 _wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]); \
167 TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio), \
168 hio_next[(ncomp)]); \
169 mtx_unlock(&hio_##name##_list_lock[ncomp]); \
171 cv_signal(&hio_##name##_list_cond[(ncomp)]); \
173 #define QUEUE_INSERT2(hio, name) do { \
176 mtx_lock(&hio_##name##_list_lock); \
177 _wakeup = TAILQ_EMPTY(&hio_##name##_list); \
178 TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
179 mtx_unlock(&hio_##name##_list_lock); \
181 cv_signal(&hio_##name##_list_cond); \
183 #define QUEUE_TAKE1(hio, name, ncomp) do { \
184 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \
185 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL) { \
186 cv_wait(&hio_##name##_list_cond[(ncomp)], \
187 &hio_##name##_list_lock[(ncomp)]); \
189 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \
190 hio_next[(ncomp)]); \
191 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \
193 #define QUEUE_TAKE2(hio, name) do { \
194 mtx_lock(&hio_##name##_list_lock); \
195 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \
196 cv_wait(&hio_##name##_list_cond, \
197 &hio_##name##_list_lock); \
199 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \
200 mtx_unlock(&hio_##name##_list_lock); \
203 #define SYNCREQ(hio) do { \
204 (hio)->hio_ggio.gctl_unit = -1; \
205 (hio)->hio_ggio.gctl_seq = 1; \
207 #define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1)
208 #define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
209 #define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2)
211 static struct hast_resource *gres;
213 static pthread_mutex_t range_lock;
214 static struct rangelocks *range_regular;
215 static bool range_regular_wait;
216 static pthread_cond_t range_regular_cond;
217 static struct rangelocks *range_sync;
218 static bool range_sync_wait;
219 static pthread_cond_t range_sync_cond;
221 static void *ggate_recv_thread(void *arg);
222 static void *local_send_thread(void *arg);
223 static void *remote_send_thread(void *arg);
224 static void *remote_recv_thread(void *arg);
225 static void *ggate_send_thread(void *arg);
226 static void *sync_thread(void *arg);
227 static void *guard_thread(void *arg);
230 cleanup(struct hast_resource *res)
234 /* Remember errno. */
237 /* Destroy ggate provider if we created one. */
238 if (res->hr_ggateunit >= 0) {
239 struct g_gate_ctl_destroy ggiod;
241 bzero(&ggiod, sizeof(ggiod));
242 ggiod.gctl_version = G_GATE_VERSION;
243 ggiod.gctl_unit = res->hr_ggateunit;
244 ggiod.gctl_force = 1;
245 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) {
246 pjdlog_errno(LOG_WARNING,
247 "Unable to destroy hast/%s device",
250 res->hr_ggateunit = -1;
258 primary_exit(int exitcode, const char *fmt, ...)
262 assert(exitcode != EX_OK);
264 pjdlogv_errno(LOG_ERR, fmt, ap);
271 primary_exitx(int exitcode, const char *fmt, ...)
276 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
283 hast_activemap_flush(struct hast_resource *res)
285 const unsigned char *buf;
288 buf = activemap_bitmap(res->hr_amp, &size);
290 assert((size % res->hr_local_sectorsize) == 0);
291 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
293 KEEP_ERRNO(pjdlog_errno(LOG_ERR,
294 "Unable to flush activemap to disk"));
301 real_remote(const struct hast_resource *res)
304 return (strcmp(res->hr_remoteaddr, "none") != 0);
308 init_environment(struct hast_resource *res __unused)
311 unsigned int ii, ncomps;
314 * In the future it might be per-resource value.
316 ncomps = HAST_NCOMPONENTS;
319 * Allocate memory needed by lists.
321 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
322 if (hio_send_list == NULL) {
323 primary_exitx(EX_TEMPFAIL,
324 "Unable to allocate %zu bytes of memory for send lists.",
325 sizeof(hio_send_list[0]) * ncomps);
327 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
328 if (hio_send_list_lock == NULL) {
329 primary_exitx(EX_TEMPFAIL,
330 "Unable to allocate %zu bytes of memory for send list locks.",
331 sizeof(hio_send_list_lock[0]) * ncomps);
333 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
334 if (hio_send_list_cond == NULL) {
335 primary_exitx(EX_TEMPFAIL,
336 "Unable to allocate %zu bytes of memory for send list condition variables.",
337 sizeof(hio_send_list_cond[0]) * ncomps);
339 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
340 if (hio_recv_list == NULL) {
341 primary_exitx(EX_TEMPFAIL,
342 "Unable to allocate %zu bytes of memory for recv lists.",
343 sizeof(hio_recv_list[0]) * ncomps);
345 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
346 if (hio_recv_list_lock == NULL) {
347 primary_exitx(EX_TEMPFAIL,
348 "Unable to allocate %zu bytes of memory for recv list locks.",
349 sizeof(hio_recv_list_lock[0]) * ncomps);
351 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
352 if (hio_recv_list_cond == NULL) {
353 primary_exitx(EX_TEMPFAIL,
354 "Unable to allocate %zu bytes of memory for recv list condition variables.",
355 sizeof(hio_recv_list_cond[0]) * ncomps);
357 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
358 if (hio_remote_lock == NULL) {
359 primary_exitx(EX_TEMPFAIL,
360 "Unable to allocate %zu bytes of memory for remote connections locks.",
361 sizeof(hio_remote_lock[0]) * ncomps);
365 * Initialize lists, their locks and theirs condition variables.
367 TAILQ_INIT(&hio_free_list);
368 mtx_init(&hio_free_list_lock);
369 cv_init(&hio_free_list_cond);
370 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
371 TAILQ_INIT(&hio_send_list[ii]);
372 mtx_init(&hio_send_list_lock[ii]);
373 cv_init(&hio_send_list_cond[ii]);
374 TAILQ_INIT(&hio_recv_list[ii]);
375 mtx_init(&hio_recv_list_lock[ii]);
376 cv_init(&hio_recv_list_cond[ii]);
377 rw_init(&hio_remote_lock[ii]);
379 TAILQ_INIT(&hio_done_list);
380 mtx_init(&hio_done_list_lock);
381 cv_init(&hio_done_list_cond);
382 mtx_init(&metadata_lock);
385 * Allocate requests pool and initialize requests.
387 for (ii = 0; ii < HAST_HIO_MAX; ii++) {
388 hio = malloc(sizeof(*hio));
390 primary_exitx(EX_TEMPFAIL,
391 "Unable to allocate %zu bytes of memory for hio request.",
394 hio->hio_countdown = 0;
395 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
396 if (hio->hio_errors == NULL) {
397 primary_exitx(EX_TEMPFAIL,
398 "Unable allocate %zu bytes of memory for hio errors.",
399 sizeof(hio->hio_errors[0]) * ncomps);
401 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
402 if (hio->hio_next == NULL) {
403 primary_exitx(EX_TEMPFAIL,
404 "Unable allocate %zu bytes of memory for hio_next field.",
405 sizeof(hio->hio_next[0]) * ncomps);
407 hio->hio_ggio.gctl_version = G_GATE_VERSION;
408 hio->hio_ggio.gctl_data = malloc(MAXPHYS);
409 if (hio->hio_ggio.gctl_data == NULL) {
410 primary_exitx(EX_TEMPFAIL,
411 "Unable to allocate %zu bytes of memory for gctl_data.",
414 hio->hio_ggio.gctl_length = MAXPHYS;
415 hio->hio_ggio.gctl_error = 0;
416 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
421 init_local(struct hast_resource *res)
426 if (metadata_read(res, true) < 0)
428 mtx_init(&res->hr_amp_lock);
429 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
430 res->hr_local_sectorsize, res->hr_keepdirty) < 0) {
431 primary_exit(EX_TEMPFAIL, "Unable to create activemap");
433 mtx_init(&range_lock);
434 cv_init(&range_regular_cond);
435 if (rangelock_init(&range_regular) < 0)
436 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
437 cv_init(&range_sync_cond);
438 if (rangelock_init(&range_sync) < 0)
439 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
440 mapsize = activemap_ondisk_size(res->hr_amp);
441 buf = calloc(1, mapsize);
443 primary_exitx(EX_TEMPFAIL,
444 "Unable to allocate buffer for activemap.");
446 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
448 primary_exit(EX_NOINPUT, "Unable to read activemap");
450 activemap_copyin(res->hr_amp, buf, mapsize);
452 if (res->hr_resuid != 0)
455 * We're using provider for the first time, so we have to generate
456 * resource unique identifier and initialize local and remote counts.
458 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
459 res->hr_primary_localcnt = 1;
460 res->hr_primary_remotecnt = 0;
461 if (metadata_write(res) < 0)
466 init_remote(struct hast_resource *res, struct proto_conn **inp,
467 struct proto_conn **outp)
469 struct proto_conn *in, *out;
470 struct nv *nvout, *nvin;
471 const unsigned char *token;
479 assert((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
480 assert(real_remote(res));
485 /* Prepare outgoing connection with remote node. */
486 if (proto_client(res->hr_remoteaddr, &out) < 0) {
487 primary_exit(EX_TEMPFAIL, "Unable to create connection to %s",
490 /* Try to connect, but accept failure. */
491 if (proto_connect(out) < 0) {
492 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
496 /* Error in setting timeout is not critical, but why should it fail? */
497 if (proto_timeout(out, res->hr_timeout) < 0)
498 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
500 * First handshake step.
501 * Setup outgoing connection with remote node.
504 nv_add_string(nvout, res->hr_name, "resource");
505 if (nv_error(nvout) != 0) {
506 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
507 "Unable to allocate header for connection with %s",
512 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) {
513 pjdlog_errno(LOG_WARNING,
514 "Unable to send handshake header to %s",
520 if (hast_proto_recv_hdr(out, &nvin) < 0) {
521 pjdlog_errno(LOG_WARNING,
522 "Unable to receive handshake header from %s",
526 errmsg = nv_get_string(nvin, "errmsg");
527 if (errmsg != NULL) {
528 pjdlog_warning("%s", errmsg);
532 token = nv_get_uint8_array(nvin, &size, "token");
534 pjdlog_warning("Handshake header from %s has no 'token' field.",
539 if (size != sizeof(res->hr_token)) {
540 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
541 res->hr_remoteaddr, size, sizeof(res->hr_token));
545 bcopy(token, res->hr_token, sizeof(res->hr_token));
549 * Second handshake step.
550 * Setup incoming connection with remote node.
552 if (proto_client(res->hr_remoteaddr, &in) < 0) {
553 pjdlog_errno(LOG_WARNING, "Unable to create connection to %s",
556 /* Try to connect, but accept failure. */
557 if (proto_connect(in) < 0) {
558 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
562 /* Error in setting timeout is not critical, but why should it fail? */
563 if (proto_timeout(in, res->hr_timeout) < 0)
564 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
566 nv_add_string(nvout, res->hr_name, "resource");
567 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
569 nv_add_uint64(nvout, res->hr_resuid, "resuid");
570 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
571 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
572 if (nv_error(nvout) != 0) {
573 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
574 "Unable to allocate header for connection with %s",
579 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) {
580 pjdlog_errno(LOG_WARNING,
581 "Unable to send handshake header to %s",
587 if (hast_proto_recv_hdr(out, &nvin) < 0) {
588 pjdlog_errno(LOG_WARNING,
589 "Unable to receive handshake header from %s",
593 errmsg = nv_get_string(nvin, "errmsg");
594 if (errmsg != NULL) {
595 pjdlog_warning("%s", errmsg);
599 datasize = nv_get_int64(nvin, "datasize");
600 if (datasize != res->hr_datasize) {
601 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
602 (intmax_t)res->hr_datasize, (intmax_t)datasize);
606 extentsize = nv_get_int32(nvin, "extentsize");
607 if (extentsize != res->hr_extentsize) {
608 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
609 (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
613 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
614 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
615 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
617 mapsize = nv_get_uint32(nvin, "mapsize");
619 map = malloc(mapsize);
621 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
627 * Remote node have some dirty extents on its own, lets
628 * download its activemap.
630 if (hast_proto_recv_data(res, out, nvin, map,
632 pjdlog_errno(LOG_ERR,
633 "Unable to receive remote activemap");
639 * Merge local and remote bitmaps.
641 activemap_merge(res->hr_amp, map, mapsize);
644 * Now that we merged bitmaps from both nodes, flush it to the
645 * disk before we start to synchronize.
647 (void)hast_activemap_flush(res);
649 pjdlog_info("Connected to %s.", res->hr_remoteaddr);
650 if (inp != NULL && outp != NULL) {
654 res->hr_remotein = in;
655 res->hr_remoteout = out;
657 event_send(res, EVENT_CONNECT);
660 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
661 event_send(res, EVENT_SPLITBRAIN);
672 mtx_lock(&sync_lock);
673 sync_inprogress = true;
674 mtx_unlock(&sync_lock);
675 cv_signal(&sync_cond);
682 mtx_lock(&sync_lock);
684 sync_inprogress = false;
685 mtx_unlock(&sync_lock);
689 init_ggate(struct hast_resource *res)
691 struct g_gate_ctl_create ggiocreate;
692 struct g_gate_ctl_cancel ggiocancel;
695 * We communicate with ggate via /dev/ggctl. Open it.
697 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
698 if (res->hr_ggatefd < 0)
699 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
701 * Create provider before trying to connect, as connection failure
702 * is not critical, but may take some time.
704 bzero(&ggiocreate, sizeof(ggiocreate));
705 ggiocreate.gctl_version = G_GATE_VERSION;
706 ggiocreate.gctl_mediasize = res->hr_datasize;
707 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
708 ggiocreate.gctl_flags = 0;
709 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE;
710 ggiocreate.gctl_timeout = 0;
711 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
712 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
714 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
715 pjdlog_info("Device hast/%s created.", res->hr_provname);
716 res->hr_ggateunit = ggiocreate.gctl_unit;
719 if (errno != EEXIST) {
720 primary_exit(EX_OSERR, "Unable to create hast/%s device",
724 "Device hast/%s already exists, we will try to take it over.",
727 * If we received EEXIST, we assume that the process who created the
728 * provider died and didn't clean up. In that case we will start from
731 bzero(&ggiocancel, sizeof(ggiocancel));
732 ggiocancel.gctl_version = G_GATE_VERSION;
733 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
734 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
736 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
737 pjdlog_info("Device hast/%s recovered.", res->hr_provname);
738 res->hr_ggateunit = ggiocancel.gctl_unit;
741 primary_exit(EX_OSERR, "Unable to take over hast/%s device",
746 hastd_primary(struct hast_resource *res)
753 * Create communication channel between parent and child.
755 if (proto_client("socketpair://", &res->hr_ctrl) < 0) {
756 KEEP_ERRNO((void)pidfile_remove(pfh));
757 pjdlog_exit(EX_OSERR,
758 "Unable to create control sockets between parent and child");
761 * Create communication channel between child and parent.
763 if (proto_client("socketpair://", &res->hr_event) < 0) {
764 KEEP_ERRNO((void)pidfile_remove(pfh));
765 pjdlog_exit(EX_OSERR,
766 "Unable to create event sockets between child and parent");
771 KEEP_ERRNO((void)pidfile_remove(pfh));
772 pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
776 /* This is parent. */
777 /* Declare that we are receiver. */
778 proto_recv(res->hr_event, NULL, 0);
779 res->hr_workerpid = pid;
785 (void)pidfile_close(pfh);
788 setproctitle("%s (primary)", res->hr_name);
790 /* Declare that we are sender. */
791 proto_send(res->hr_event, NULL, 0);
795 init_environment(res);
797 * Create the guard thread first, so we can handle signals from the
800 error = pthread_create(&td, NULL, guard_thread, res);
803 * Create the control thread before sending any event to the parent,
804 * as we can deadlock when parent sends control request to worker,
805 * but worker has no control thread started yet, so parent waits.
806 * In the meantime worker sends an event to the parent, but parent
807 * is unable to handle the event, because it waits for control
810 error = pthread_create(&td, NULL, ctrl_thread, res);
812 if (real_remote(res) && init_remote(res, NULL, NULL))
814 error = pthread_create(&td, NULL, ggate_recv_thread, res);
816 error = pthread_create(&td, NULL, local_send_thread, res);
818 error = pthread_create(&td, NULL, remote_send_thread, res);
820 error = pthread_create(&td, NULL, remote_recv_thread, res);
822 error = pthread_create(&td, NULL, ggate_send_thread, res);
824 (void)sync_thread(res);
828 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...)
835 len = vsnprintf(msg, sizeof(msg), fmt, ap);
837 if ((size_t)len < sizeof(msg)) {
838 switch (ggio->gctl_cmd) {
840 (void)snprintf(msg + len, sizeof(msg) - len,
841 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
842 (uintmax_t)ggio->gctl_length);
845 (void)snprintf(msg + len, sizeof(msg) - len,
846 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
847 (uintmax_t)ggio->gctl_length);
850 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
853 (void)snprintf(msg + len, sizeof(msg) - len,
854 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
855 (uintmax_t)ggio->gctl_length);
858 (void)snprintf(msg + len, sizeof(msg) - len,
859 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd);
863 pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
867 remote_close(struct hast_resource *res, int ncomp)
870 rw_wlock(&hio_remote_lock[ncomp]);
872 * A race is possible between dropping rlock and acquiring wlock -
873 * another thread can close connection in-between.
875 if (!ISCONNECTED(res, ncomp)) {
876 assert(res->hr_remotein == NULL);
877 assert(res->hr_remoteout == NULL);
878 rw_unlock(&hio_remote_lock[ncomp]);
882 assert(res->hr_remotein != NULL);
883 assert(res->hr_remoteout != NULL);
885 pjdlog_debug(2, "Closing incoming connection to %s.",
887 proto_close(res->hr_remotein);
888 res->hr_remotein = NULL;
889 pjdlog_debug(2, "Closing outgoing connection to %s.",
891 proto_close(res->hr_remoteout);
892 res->hr_remoteout = NULL;
894 rw_unlock(&hio_remote_lock[ncomp]);
896 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
899 * Stop synchronization if in-progress.
903 event_send(res, EVENT_DISCONNECT);
907 * Thread receives ggate I/O requests from the kernel and passes them to
908 * appropriate threads:
909 * WRITE - always goes to both local_send and remote_send threads
910 * READ (when the block is up-to-date on local component) -
911 * only local_send thread
912 * READ (when the block isn't up-to-date on local component) -
913 * only remote_send thread
914 * DELETE - always goes to both local_send and remote_send threads
915 * FLUSH - always goes to both local_send and remote_send threads
918 ggate_recv_thread(void *arg)
920 struct hast_resource *res = arg;
921 struct g_gate_ctl_io *ggio;
923 unsigned int ii, ncomp, ncomps;
926 ncomps = HAST_NCOMPONENTS;
929 pjdlog_debug(2, "ggate_recv: Taking free request.");
930 QUEUE_TAKE2(hio, free);
931 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
932 ggio = &hio->hio_ggio;
933 ggio->gctl_unit = res->hr_ggateunit;
934 ggio->gctl_length = MAXPHYS;
935 ggio->gctl_error = 0;
937 "ggate_recv: (%p) Waiting for request from the kernel.",
939 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) {
940 if (sigexit_received)
942 primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
944 error = ggio->gctl_error;
949 /* Exit gracefully. */
950 if (!sigexit_received) {
952 "ggate_recv: (%p) Received cancel from the kernel.",
954 pjdlog_info("Received cancel from the kernel, exiting.");
959 * Buffer too small? Impossible, we allocate MAXPHYS
960 * bytes - request can't be bigger than that.
965 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
968 for (ii = 0; ii < ncomps; ii++)
969 hio->hio_errors[ii] = EINVAL;
970 reqlog(LOG_DEBUG, 2, ggio,
971 "ggate_recv: (%p) Request received from the kernel: ",
974 * Inform all components about new write request.
975 * For read request prefer local component unless the given
976 * range is out-of-date, then use remote component.
978 switch (ggio->gctl_cmd) {
981 "ggate_recv: (%p) Moving request to the send queue.",
983 refcount_init(&hio->hio_countdown, 1);
984 mtx_lock(&metadata_lock);
985 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
986 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
988 * This range is up-to-date on local component,
989 * so handle request locally.
991 /* Local component is 0 for now. */
993 } else /* if (res->hr_syncsrc ==
994 HAST_SYNCSRC_SECONDARY) */ {
995 assert(res->hr_syncsrc ==
996 HAST_SYNCSRC_SECONDARY);
998 * This range is out-of-date on local component,
999 * so send request to the remote node.
1001 /* Remote component is 1 for now. */
1004 mtx_unlock(&metadata_lock);
1005 QUEUE_INSERT1(hio, send, ncomp);
1009 mtx_lock(&range_lock);
1010 if (rangelock_islocked(range_sync,
1011 ggio->gctl_offset, ggio->gctl_length)) {
1013 "regular: Range offset=%jd length=%zu locked.",
1014 (intmax_t)ggio->gctl_offset,
1015 (size_t)ggio->gctl_length);
1016 range_regular_wait = true;
1017 cv_wait(&range_regular_cond, &range_lock);
1018 range_regular_wait = false;
1019 mtx_unlock(&range_lock);
1022 if (rangelock_add(range_regular,
1023 ggio->gctl_offset, ggio->gctl_length) < 0) {
1024 mtx_unlock(&range_lock);
1026 "regular: Range offset=%jd length=%zu is already locked, waiting.",
1027 (intmax_t)ggio->gctl_offset,
1028 (size_t)ggio->gctl_length);
1032 mtx_unlock(&range_lock);
1035 mtx_lock(&res->hr_amp_lock);
1036 if (activemap_write_start(res->hr_amp,
1037 ggio->gctl_offset, ggio->gctl_length)) {
1038 (void)hast_activemap_flush(res);
1040 mtx_unlock(&res->hr_amp_lock);
1045 "ggate_recv: (%p) Moving request to the send queues.",
1047 refcount_init(&hio->hio_countdown, ncomps);
1048 for (ii = 0; ii < ncomps; ii++)
1049 QUEUE_INSERT1(hio, send, ii);
1058 * Thread reads from or writes to local component.
1059 * If local read fails, it redirects it to remote_send thread.
1062 local_send_thread(void *arg)
1064 struct hast_resource *res = arg;
1065 struct g_gate_ctl_io *ggio;
1067 unsigned int ncomp, rncomp;
1070 /* Local component is 0 for now. */
1072 /* Remote component is 1 for now. */
1076 pjdlog_debug(2, "local_send: Taking request.");
1077 QUEUE_TAKE1(hio, send, ncomp);
1078 pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1079 ggio = &hio->hio_ggio;
1080 switch (ggio->gctl_cmd) {
1082 ret = pread(res->hr_localfd, ggio->gctl_data,
1084 ggio->gctl_offset + res->hr_localoff);
1085 if (ret == ggio->gctl_length)
1086 hio->hio_errors[ncomp] = 0;
1089 * If READ failed, try to read from remote node.
1091 QUEUE_INSERT1(hio, send, rncomp);
1096 ret = pwrite(res->hr_localfd, ggio->gctl_data,
1098 ggio->gctl_offset + res->hr_localoff);
1100 hio->hio_errors[ncomp] = errno;
1101 else if (ret != ggio->gctl_length)
1102 hio->hio_errors[ncomp] = EIO;
1104 hio->hio_errors[ncomp] = 0;
1107 ret = g_delete(res->hr_localfd,
1108 ggio->gctl_offset + res->hr_localoff,
1111 hio->hio_errors[ncomp] = errno;
1113 hio->hio_errors[ncomp] = 0;
1116 ret = g_flush(res->hr_localfd);
1118 hio->hio_errors[ncomp] = errno;
1120 hio->hio_errors[ncomp] = 0;
1123 if (refcount_release(&hio->hio_countdown)) {
1124 if (ISSYNCREQ(hio)) {
1125 mtx_lock(&sync_lock);
1127 mtx_unlock(&sync_lock);
1128 cv_signal(&sync_cond);
1131 "local_send: (%p) Moving request to the done queue.",
1133 QUEUE_INSERT2(hio, done);
1142 * Thread sends request to secondary node.
1145 remote_send_thread(void *arg)
1147 struct hast_resource *res = arg;
1148 struct g_gate_ctl_io *ggio;
1153 uint64_t offset, length;
1157 /* Remote component is 1 for now. */
1161 pjdlog_debug(2, "remote_send: Taking request.");
1162 QUEUE_TAKE1(hio, send, ncomp);
1163 pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1164 ggio = &hio->hio_ggio;
1165 switch (ggio->gctl_cmd) {
1169 offset = ggio->gctl_offset;
1170 length = ggio->gctl_length;
1174 data = ggio->gctl_data;
1175 offset = ggio->gctl_offset;
1176 length = ggio->gctl_length;
1181 offset = ggio->gctl_offset;
1182 length = ggio->gctl_length;
1191 assert(!"invalid condition");
1195 nv_add_uint8(nv, cmd, "cmd");
1196 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1197 nv_add_uint64(nv, offset, "offset");
1198 nv_add_uint64(nv, length, "length");
1199 if (nv_error(nv) != 0) {
1200 hio->hio_errors[ncomp] = nv_error(nv);
1202 "remote_send: (%p) Unable to prepare header to send.",
1204 reqlog(LOG_ERR, 0, ggio,
1205 "Unable to prepare header to send (%s): ",
1206 strerror(nv_error(nv)));
1207 /* Move failed request immediately to the done queue. */
1211 "remote_send: (%p) Moving request to the recv queue.",
1214 * Protect connection from disappearing.
1216 rw_rlock(&hio_remote_lock[ncomp]);
1217 if (!ISCONNECTED(res, ncomp)) {
1218 rw_unlock(&hio_remote_lock[ncomp]);
1219 hio->hio_errors[ncomp] = ENOTCONN;
1223 * Move the request to recv queue before sending it, because
1224 * in different order we can get reply before we move request
1227 mtx_lock(&hio_recv_list_lock[ncomp]);
1228 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1229 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1230 mtx_unlock(&hio_recv_list_lock[ncomp]);
1231 if (hast_proto_send(res, res->hr_remoteout, nv, data,
1232 data != NULL ? length : 0) < 0) {
1233 hio->hio_errors[ncomp] = errno;
1234 rw_unlock(&hio_remote_lock[ncomp]);
1236 "remote_send: (%p) Unable to send request.", hio);
1237 reqlog(LOG_ERR, 0, ggio,
1238 "Unable to send request (%s): ",
1239 strerror(hio->hio_errors[ncomp]));
1240 remote_close(res, ncomp);
1242 * Take request back from the receive queue and move
1243 * it immediately to the done queue.
1245 mtx_lock(&hio_recv_list_lock[ncomp]);
1246 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1247 mtx_unlock(&hio_recv_list_lock[ncomp]);
1250 rw_unlock(&hio_remote_lock[ncomp]);
1253 cv_signal(&hio_recv_list_cond[ncomp]);
1257 if (ISSYNCREQ(hio)) {
1258 if (!refcount_release(&hio->hio_countdown))
1260 mtx_lock(&sync_lock);
1262 mtx_unlock(&sync_lock);
1263 cv_signal(&sync_cond);
1266 if (ggio->gctl_cmd == BIO_WRITE) {
1267 mtx_lock(&res->hr_amp_lock);
1268 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1269 ggio->gctl_length)) {
1270 (void)hast_activemap_flush(res);
1272 mtx_unlock(&res->hr_amp_lock);
1274 if (!refcount_release(&hio->hio_countdown))
1277 "remote_send: (%p) Moving request to the done queue.",
1279 QUEUE_INSERT2(hio, done);
1286 * Thread receives answer from secondary node and passes it to ggate_send
1290 remote_recv_thread(void *arg)
1292 struct hast_resource *res = arg;
1293 struct g_gate_ctl_io *ggio;
1300 /* Remote component is 1 for now. */
1304 /* Wait until there is anything to receive. */
1305 mtx_lock(&hio_recv_list_lock[ncomp]);
1306 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1307 pjdlog_debug(2, "remote_recv: No requests, waiting.");
1308 cv_wait(&hio_recv_list_cond[ncomp],
1309 &hio_recv_list_lock[ncomp]);
1311 mtx_unlock(&hio_recv_list_lock[ncomp]);
1312 rw_rlock(&hio_remote_lock[ncomp]);
1313 if (!ISCONNECTED(res, ncomp)) {
1314 rw_unlock(&hio_remote_lock[ncomp]);
1316 * Connection is dead, so move all pending requests to
1317 * the done queue (one-by-one).
1319 mtx_lock(&hio_recv_list_lock[ncomp]);
1320 hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1321 assert(hio != NULL);
1322 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1324 mtx_unlock(&hio_recv_list_lock[ncomp]);
1327 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) {
1328 pjdlog_errno(LOG_ERR,
1329 "Unable to receive reply header");
1330 rw_unlock(&hio_remote_lock[ncomp]);
1331 remote_close(res, ncomp);
1334 rw_unlock(&hio_remote_lock[ncomp]);
1335 seq = nv_get_uint64(nv, "seq");
1337 pjdlog_error("Header contains no 'seq' field.");
1341 mtx_lock(&hio_recv_list_lock[ncomp]);
1342 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1343 if (hio->hio_ggio.gctl_seq == seq) {
1344 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1349 mtx_unlock(&hio_recv_list_lock[ncomp]);
1351 pjdlog_error("Found no request matching received 'seq' field (%ju).",
1356 error = nv_get_int16(nv, "error");
1358 /* Request failed on remote side. */
1359 hio->hio_errors[ncomp] = 0;
1363 ggio = &hio->hio_ggio;
1364 switch (ggio->gctl_cmd) {
1366 rw_rlock(&hio_remote_lock[ncomp]);
1367 if (!ISCONNECTED(res, ncomp)) {
1368 rw_unlock(&hio_remote_lock[ncomp]);
1372 if (hast_proto_recv_data(res, res->hr_remotein, nv,
1373 ggio->gctl_data, ggio->gctl_length) < 0) {
1374 hio->hio_errors[ncomp] = errno;
1375 pjdlog_errno(LOG_ERR,
1376 "Unable to receive reply data");
1377 rw_unlock(&hio_remote_lock[ncomp]);
1379 remote_close(res, ncomp);
1382 rw_unlock(&hio_remote_lock[ncomp]);
1389 assert(!"invalid condition");
1392 hio->hio_errors[ncomp] = 0;
1395 if (refcount_release(&hio->hio_countdown)) {
1396 if (ISSYNCREQ(hio)) {
1397 mtx_lock(&sync_lock);
1399 mtx_unlock(&sync_lock);
1400 cv_signal(&sync_cond);
1403 "remote_recv: (%p) Moving request to the done queue.",
1405 QUEUE_INSERT2(hio, done);
1414 * Thread sends answer to the kernel.
1417 ggate_send_thread(void *arg)
1419 struct hast_resource *res = arg;
1420 struct g_gate_ctl_io *ggio;
1422 unsigned int ii, ncomp, ncomps;
1424 ncomps = HAST_NCOMPONENTS;
1427 pjdlog_debug(2, "ggate_send: Taking request.");
1428 QUEUE_TAKE2(hio, done);
1429 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1430 ggio = &hio->hio_ggio;
1431 for (ii = 0; ii < ncomps; ii++) {
1432 if (hio->hio_errors[ii] == 0) {
1434 * One successful request is enough to declare
1437 ggio->gctl_error = 0;
1443 * None of the requests were successful.
1446 ggio->gctl_error = hio->hio_errors[0];
1448 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1449 mtx_lock(&res->hr_amp_lock);
1450 activemap_write_complete(res->hr_amp,
1451 ggio->gctl_offset, ggio->gctl_length);
1452 mtx_unlock(&res->hr_amp_lock);
1454 if (ggio->gctl_cmd == BIO_WRITE) {
1456 * Unlock range we locked.
1458 mtx_lock(&range_lock);
1459 rangelock_del(range_regular, ggio->gctl_offset,
1461 if (range_sync_wait)
1462 cv_signal(&range_sync_cond);
1463 mtx_unlock(&range_lock);
1465 * Bump local count if this is first write after
1466 * connection failure with remote node.
1469 rw_rlock(&hio_remote_lock[ncomp]);
1470 if (!ISCONNECTED(res, ncomp)) {
1471 mtx_lock(&metadata_lock);
1472 if (res->hr_primary_localcnt ==
1473 res->hr_secondary_remotecnt) {
1474 res->hr_primary_localcnt++;
1476 "Increasing localcnt to %ju.",
1477 (uintmax_t)res->hr_primary_localcnt);
1478 (void)metadata_write(res);
1480 mtx_unlock(&metadata_lock);
1482 rw_unlock(&hio_remote_lock[ncomp]);
1484 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0)
1485 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1487 "ggate_send: (%p) Moving request to the free queue.", hio);
1488 QUEUE_INSERT2(hio, free);
1495 * Thread synchronize local and remote components.
1498 sync_thread(void *arg __unused)
1500 struct hast_resource *res = arg;
1502 struct g_gate_ctl_io *ggio;
1503 unsigned int ii, ncomp, ncomps;
1504 off_t offset, length, synced;
1508 ncomps = HAST_NCOMPONENTS;
1514 mtx_lock(&sync_lock);
1515 if (offset >= 0 && !sync_inprogress) {
1516 pjdlog_info("Synchronization interrupted. "
1517 "%jd bytes synchronized so far.",
1519 event_send(res, EVENT_SYNCINTR);
1521 while (!sync_inprogress) {
1524 cv_wait(&sync_cond, &sync_lock);
1526 mtx_unlock(&sync_lock);
1528 * Obtain offset at which we should synchronize.
1529 * Rewind synchronization if needed.
1531 mtx_lock(&res->hr_amp_lock);
1533 activemap_sync_rewind(res->hr_amp);
1534 offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1535 if (syncext != -1) {
1537 * We synchronized entire syncext extent, we can mark
1540 if (activemap_extent_complete(res->hr_amp, syncext))
1541 (void)hast_activemap_flush(res);
1543 mtx_unlock(&res->hr_amp_lock);
1547 pjdlog_info("Nodes are in sync.");
1549 pjdlog_info("Synchronization started. %ju bytes to go.",
1550 (uintmax_t)(res->hr_extentsize *
1551 activemap_ndirty(res->hr_amp)));
1552 event_send(res, EVENT_SYNCSTART);
1557 pjdlog_debug(1, "Nothing to synchronize.");
1559 * Synchronization complete, make both localcnt and
1563 rw_rlock(&hio_remote_lock[ncomp]);
1564 if (ISCONNECTED(res, ncomp)) {
1566 pjdlog_info("Synchronization complete. "
1567 "%jd bytes synchronized.",
1569 event_send(res, EVENT_SYNCDONE);
1571 mtx_lock(&metadata_lock);
1572 res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
1573 res->hr_primary_localcnt =
1574 res->hr_secondary_localcnt;
1575 res->hr_primary_remotecnt =
1576 res->hr_secondary_remotecnt;
1578 "Setting localcnt to %ju and remotecnt to %ju.",
1579 (uintmax_t)res->hr_primary_localcnt,
1580 (uintmax_t)res->hr_secondary_localcnt);
1581 (void)metadata_write(res);
1582 mtx_unlock(&metadata_lock);
1584 rw_unlock(&hio_remote_lock[ncomp]);
1587 pjdlog_debug(2, "sync: Taking free request.");
1588 QUEUE_TAKE2(hio, free);
1589 pjdlog_debug(2, "sync: (%p) Got free request.", hio);
1591 * Lock the range we are going to synchronize. We don't want
1592 * race where someone writes between our read and write.
1595 mtx_lock(&range_lock);
1596 if (rangelock_islocked(range_regular, offset, length)) {
1598 "sync: Range offset=%jd length=%jd locked.",
1599 (intmax_t)offset, (intmax_t)length);
1600 range_sync_wait = true;
1601 cv_wait(&range_sync_cond, &range_lock);
1602 range_sync_wait = false;
1603 mtx_unlock(&range_lock);
1606 if (rangelock_add(range_sync, offset, length) < 0) {
1607 mtx_unlock(&range_lock);
1609 "sync: Range offset=%jd length=%jd is already locked, waiting.",
1610 (intmax_t)offset, (intmax_t)length);
1614 mtx_unlock(&range_lock);
1618 * First read the data from synchronization source.
1621 ggio = &hio->hio_ggio;
1622 ggio->gctl_cmd = BIO_READ;
1623 ggio->gctl_offset = offset;
1624 ggio->gctl_length = length;
1625 ggio->gctl_error = 0;
1626 for (ii = 0; ii < ncomps; ii++)
1627 hio->hio_errors[ii] = EINVAL;
1628 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1630 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1632 mtx_lock(&metadata_lock);
1633 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1635 * This range is up-to-date on local component,
1636 * so handle request locally.
1638 /* Local component is 0 for now. */
1640 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1641 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1643 * This range is out-of-date on local component,
1644 * so send request to the remote node.
1646 /* Remote component is 1 for now. */
1649 mtx_unlock(&metadata_lock);
1650 refcount_init(&hio->hio_countdown, 1);
1651 QUEUE_INSERT1(hio, send, ncomp);
1654 * Let's wait for READ to finish.
1656 mtx_lock(&sync_lock);
1657 while (!ISSYNCREQDONE(hio))
1658 cv_wait(&sync_cond, &sync_lock);
1659 mtx_unlock(&sync_lock);
1661 if (hio->hio_errors[ncomp] != 0) {
1662 pjdlog_error("Unable to read synchronization data: %s.",
1663 strerror(hio->hio_errors[ncomp]));
1668 * We read the data from synchronization source, now write it
1669 * to synchronization target.
1672 ggio->gctl_cmd = BIO_WRITE;
1673 for (ii = 0; ii < ncomps; ii++)
1674 hio->hio_errors[ii] = EINVAL;
1675 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1677 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1679 mtx_lock(&metadata_lock);
1680 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1682 * This range is up-to-date on local component,
1683 * so we update remote component.
1685 /* Remote component is 1 for now. */
1687 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1688 assert(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1690 * This range is out-of-date on local component,
1693 /* Local component is 0 for now. */
1696 mtx_unlock(&metadata_lock);
1698 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.",
1700 refcount_init(&hio->hio_countdown, 1);
1701 QUEUE_INSERT1(hio, send, ncomp);
1704 * Let's wait for WRITE to finish.
1706 mtx_lock(&sync_lock);
1707 while (!ISSYNCREQDONE(hio))
1708 cv_wait(&sync_cond, &sync_lock);
1709 mtx_unlock(&sync_lock);
1711 if (hio->hio_errors[ncomp] != 0) {
1712 pjdlog_error("Unable to write synchronization data: %s.",
1713 strerror(hio->hio_errors[ncomp]));
1719 mtx_lock(&range_lock);
1720 rangelock_del(range_sync, offset, length);
1721 if (range_regular_wait)
1722 cv_signal(&range_regular_cond);
1723 mtx_unlock(&range_lock);
1724 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
1726 QUEUE_INSERT2(hio, free);
1735 struct hastd_config *newcfg;
1736 struct hast_resource *res;
1737 unsigned int ii, ncomps;
1740 pjdlog_info("Reloading configuration...");
1742 ncomps = HAST_NCOMPONENTS;
1744 newcfg = yy_config_parse(cfgpath, false);
1748 TAILQ_FOREACH(res, &newcfg->hc_resources, hr_next) {
1749 if (strcmp(res->hr_name, gres->hr_name) == 0)
1753 * If resource was removed from the configuration file, resource
1754 * name, provider name or path to local component was modified we
1755 * shouldn't be here. This means that someone modified configuration
1756 * file and send SIGHUP to us instead of main hastd process.
1757 * Log advice and ignore the signal.
1759 if (res == NULL || strcmp(gres->hr_name, res->hr_name) != 0 ||
1760 strcmp(gres->hr_provname, res->hr_provname) != 0 ||
1761 strcmp(gres->hr_localpath, res->hr_localpath) != 0) {
1762 pjdlog_warning("To reload configuration send SIGHUP to the main hastd process (pid %u).",
1763 (unsigned int)getppid());
1767 #define MODIFIED_REMOTEADDR 0x1
1768 #define MODIFIED_REPLICATION 0x2
1769 #define MODIFIED_TIMEOUT 0x4
1770 #define MODIFIED_EXEC 0x8
1772 if (strcmp(gres->hr_remoteaddr, res->hr_remoteaddr) != 0) {
1774 * Don't copy res->hr_remoteaddr to gres just yet.
1775 * We want remote_close() to log disconnect from the old
1776 * addresses, not from the new ones.
1778 modified |= MODIFIED_REMOTEADDR;
1780 if (gres->hr_replication != res->hr_replication) {
1781 gres->hr_replication = res->hr_replication;
1782 modified |= MODIFIED_REPLICATION;
1784 if (gres->hr_timeout != res->hr_timeout) {
1785 gres->hr_timeout = res->hr_timeout;
1786 modified |= MODIFIED_TIMEOUT;
1788 if (strcmp(gres->hr_exec, res->hr_exec) != 0) {
1789 strlcpy(gres->hr_exec, res->hr_exec, sizeof(gres->hr_exec));
1790 modified |= MODIFIED_EXEC;
1793 * If only timeout was modified we only need to change it without
1796 if (modified == MODIFIED_TIMEOUT) {
1797 for (ii = 0; ii < ncomps; ii++) {
1800 rw_rlock(&hio_remote_lock[ii]);
1801 if (!ISCONNECTED(gres, ii)) {
1802 rw_unlock(&hio_remote_lock[ii]);
1805 rw_unlock(&hio_remote_lock[ii]);
1806 if (proto_timeout(gres->hr_remotein,
1807 gres->hr_timeout) < 0) {
1808 pjdlog_errno(LOG_WARNING,
1809 "Unable to set connection timeout");
1811 if (proto_timeout(gres->hr_remoteout,
1812 gres->hr_timeout) < 0) {
1813 pjdlog_errno(LOG_WARNING,
1814 "Unable to set connection timeout");
1817 } else if ((modified &
1818 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) {
1819 for (ii = 0; ii < ncomps; ii++) {
1822 remote_close(gres, ii);
1824 if (modified & MODIFIED_REMOTEADDR) {
1825 strlcpy(gres->hr_remoteaddr, res->hr_remoteaddr,
1826 sizeof(gres->hr_remoteaddr));
1829 #undef MODIFIED_REMOTEADDR
1830 #undef MODIFIED_REPLICATION
1831 #undef MODIFIED_TIMEOUT
1832 #undef MODIFIED_EXEC
1834 pjdlog_info("Configuration reloaded successfully.");
1837 if (newcfg != NULL) {
1838 if (newcfg->hc_controlconn != NULL)
1839 proto_close(newcfg->hc_controlconn);
1840 if (newcfg->hc_listenconn != NULL)
1841 proto_close(newcfg->hc_listenconn);
1842 yy_config_free(newcfg);
1844 pjdlog_warning("Configuration not reloaded.");
1848 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1853 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1854 if (nv_error(nv) != 0) {
1857 "keepalive_send: Unable to prepare header to send.");
1860 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) {
1861 pjdlog_common(LOG_DEBUG, 1, errno,
1862 "keepalive_send: Unable to send request");
1864 rw_unlock(&hio_remote_lock[ncomp]);
1865 remote_close(res, ncomp);
1866 rw_rlock(&hio_remote_lock[ncomp]);
1870 pjdlog_debug(2, "keepalive_send: Request sent.");
1874 guard_one(struct hast_resource *res, unsigned int ncomp)
1876 struct proto_conn *in, *out;
1878 if (!ISREMOTE(ncomp))
1881 rw_rlock(&hio_remote_lock[ncomp]);
1883 if (!real_remote(res)) {
1884 rw_unlock(&hio_remote_lock[ncomp]);
1888 if (ISCONNECTED(res, ncomp)) {
1889 assert(res->hr_remotein != NULL);
1890 assert(res->hr_remoteout != NULL);
1891 keepalive_send(res, ncomp);
1894 if (ISCONNECTED(res, ncomp)) {
1895 assert(res->hr_remotein != NULL);
1896 assert(res->hr_remoteout != NULL);
1897 rw_unlock(&hio_remote_lock[ncomp]);
1898 pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
1899 res->hr_remoteaddr);
1903 assert(res->hr_remotein == NULL);
1904 assert(res->hr_remoteout == NULL);
1906 * Upgrade the lock. It doesn't have to be atomic as no other thread
1907 * can change connection status from disconnected to connected.
1909 rw_unlock(&hio_remote_lock[ncomp]);
1910 pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
1911 res->hr_remoteaddr);
1913 if (init_remote(res, &in, &out)) {
1914 rw_wlock(&hio_remote_lock[ncomp]);
1915 assert(res->hr_remotein == NULL);
1916 assert(res->hr_remoteout == NULL);
1917 assert(in != NULL && out != NULL);
1918 res->hr_remotein = in;
1919 res->hr_remoteout = out;
1920 rw_unlock(&hio_remote_lock[ncomp]);
1921 pjdlog_info("Successfully reconnected to %s.",
1922 res->hr_remoteaddr);
1925 /* Both connections should be NULL. */
1926 assert(res->hr_remotein == NULL);
1927 assert(res->hr_remoteout == NULL);
1928 assert(in == NULL && out == NULL);
1929 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
1930 res->hr_remoteaddr);
1935 * Thread guards remote connections and reconnects when needed, handles
1939 guard_thread(void *arg)
1941 struct hast_resource *res = arg;
1942 unsigned int ii, ncomps;
1943 struct timespec timeout;
1944 time_t lastcheck, now;
1948 ncomps = HAST_NCOMPONENTS;
1949 lastcheck = time(NULL);
1951 PJDLOG_VERIFY(sigemptyset(&mask) == 0);
1952 PJDLOG_VERIFY(sigaddset(&mask, SIGHUP) == 0);
1953 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
1954 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
1956 timeout.tv_nsec = 0;
1966 sigexit_received = true;
1967 primary_exitx(EX_OK,
1968 "Termination signal received, exiting.");
1974 pjdlog_debug(2, "remote_guard: Checking connections.");
1976 if (lastcheck + RETRY_SLEEP <= now) {
1977 for (ii = 0; ii < ncomps; ii++)
1981 timeout.tv_sec = RETRY_SLEEP;
1982 signo = sigtimedwait(&mask, NULL, &timeout);