2 * Copyright (c) 2009 The FreeBSD Foundation
3 * Copyright (c) 2010-2011 Pawel Jakub Dawidek <pawel@dawidek.net>
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>
55 #include <activemap.h>
57 #include <rangelock.h>
62 #include "hast_proto.h"
71 /* The is only one remote component for now. */
72 #define ISREMOTE(no) ((no) == 1)
76 * Number of components we are still waiting for.
77 * When this field goes to 0, we can send the request back to the
78 * kernel. Each component has to decrease this counter by one
81 unsigned int hio_countdown;
83 * Each component has a place to store its own error.
84 * Once the request is handled by all components we can decide if the
85 * request overall is successful or not.
89 * Structure used to communicate with GEOM Gate class.
91 struct g_gate_ctl_io hio_ggio;
92 TAILQ_ENTRY(hio) *hio_next;
94 #define hio_free_next hio_next[0]
95 #define hio_done_next hio_next[0]
98 * Free list holds unused structures. When free list is empty, we have to wait
99 * until some in-progress requests are freed.
101 static TAILQ_HEAD(, hio) hio_free_list;
102 static pthread_mutex_t hio_free_list_lock;
103 static pthread_cond_t hio_free_list_cond;
105 * There is one send list for every component. One requests is placed on all
106 * send lists - each component gets the same request, but each component is
107 * responsible for managing his own send list.
109 static TAILQ_HEAD(, hio) *hio_send_list;
110 static pthread_mutex_t *hio_send_list_lock;
111 static pthread_cond_t *hio_send_list_cond;
113 * There is one recv list for every component, although local components don't
114 * use recv lists as local requests are done synchronously.
116 static TAILQ_HEAD(, hio) *hio_recv_list;
117 static pthread_mutex_t *hio_recv_list_lock;
118 static pthread_cond_t *hio_recv_list_cond;
120 * Request is placed on done list by the slowest component (the one that
121 * decreased hio_countdown from 1 to 0).
123 static TAILQ_HEAD(, hio) hio_done_list;
124 static pthread_mutex_t hio_done_list_lock;
125 static pthread_cond_t hio_done_list_cond;
127 * Structure below are for interaction with sync thread.
129 static bool sync_inprogress;
130 static pthread_mutex_t sync_lock;
131 static pthread_cond_t sync_cond;
133 * The lock below allows to synchornize access to remote connections.
135 static pthread_rwlock_t *hio_remote_lock;
138 * Lock to synchronize metadata updates. Also synchronize access to
139 * hr_primary_localcnt and hr_primary_remotecnt fields.
141 static pthread_mutex_t metadata_lock;
144 * Maximum number of outstanding I/O requests.
146 #define HAST_HIO_MAX 256
148 * Number of components. At this point there are only two components: local
149 * and remote, but in the future it might be possible to use multiple local
150 * and remote components.
152 #define HAST_NCOMPONENTS 2
154 #define ISCONNECTED(res, no) \
155 ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
157 #define QUEUE_INSERT1(hio, name, ncomp) do { \
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]); \
166 cv_signal(&hio_##name##_list_cond[(ncomp)]); \
168 #define QUEUE_INSERT2(hio, name) do { \
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); \
176 cv_signal(&hio_##name##_list_cond); \
178 #define QUEUE_TAKE1(hio, name, ncomp, timeout) do { \
181 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \
183 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
184 cv_timedwait(&hio_##name##_list_cond[(ncomp)], \
185 &hio_##name##_list_lock[(ncomp)], (timeout)); \
186 if ((timeout) != 0) \
190 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \
191 hio_next[(ncomp)]); \
193 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \
195 #define QUEUE_TAKE2(hio, name) do { \
196 mtx_lock(&hio_##name##_list_lock); \
197 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \
198 cv_wait(&hio_##name##_list_cond, \
199 &hio_##name##_list_lock); \
201 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \
202 mtx_unlock(&hio_##name##_list_lock); \
205 #define SYNCREQ(hio) do { \
206 (hio)->hio_ggio.gctl_unit = -1; \
207 (hio)->hio_ggio.gctl_seq = 1; \
209 #define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1)
210 #define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
211 #define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2)
213 static struct hast_resource *gres;
215 static pthread_mutex_t range_lock;
216 static struct rangelocks *range_regular;
217 static bool range_regular_wait;
218 static pthread_cond_t range_regular_cond;
219 static struct rangelocks *range_sync;
220 static bool range_sync_wait;
221 static pthread_cond_t range_sync_cond;
222 static bool fullystarted;
224 static void *ggate_recv_thread(void *arg);
225 static void *local_send_thread(void *arg);
226 static void *remote_send_thread(void *arg);
227 static void *remote_recv_thread(void *arg);
228 static void *ggate_send_thread(void *arg);
229 static void *sync_thread(void *arg);
230 static void *guard_thread(void *arg);
233 cleanup(struct hast_resource *res)
237 /* Remember errno. */
240 /* Destroy ggate provider if we created one. */
241 if (res->hr_ggateunit >= 0) {
242 struct g_gate_ctl_destroy ggiod;
244 bzero(&ggiod, sizeof(ggiod));
245 ggiod.gctl_version = G_GATE_VERSION;
246 ggiod.gctl_unit = res->hr_ggateunit;
247 ggiod.gctl_force = 1;
248 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) {
249 pjdlog_errno(LOG_WARNING,
250 "Unable to destroy hast/%s device",
253 res->hr_ggateunit = -1;
261 primary_exit(int exitcode, const char *fmt, ...)
265 PJDLOG_ASSERT(exitcode != EX_OK);
267 pjdlogv_errno(LOG_ERR, fmt, ap);
274 primary_exitx(int exitcode, const char *fmt, ...)
279 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
286 hast_activemap_flush(struct hast_resource *res)
288 const unsigned char *buf;
291 buf = activemap_bitmap(res->hr_amp, &size);
292 PJDLOG_ASSERT(buf != NULL);
293 PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
294 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
296 KEEP_ERRNO(pjdlog_errno(LOG_ERR,
297 "Unable to flush activemap to disk"));
304 real_remote(const struct hast_resource *res)
307 return (strcmp(res->hr_remoteaddr, "none") != 0);
311 init_environment(struct hast_resource *res __unused)
314 unsigned int ii, ncomps;
317 * In the future it might be per-resource value.
319 ncomps = HAST_NCOMPONENTS;
322 * Allocate memory needed by lists.
324 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
325 if (hio_send_list == NULL) {
326 primary_exitx(EX_TEMPFAIL,
327 "Unable to allocate %zu bytes of memory for send lists.",
328 sizeof(hio_send_list[0]) * ncomps);
330 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
331 if (hio_send_list_lock == NULL) {
332 primary_exitx(EX_TEMPFAIL,
333 "Unable to allocate %zu bytes of memory for send list locks.",
334 sizeof(hio_send_list_lock[0]) * ncomps);
336 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
337 if (hio_send_list_cond == NULL) {
338 primary_exitx(EX_TEMPFAIL,
339 "Unable to allocate %zu bytes of memory for send list condition variables.",
340 sizeof(hio_send_list_cond[0]) * ncomps);
342 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
343 if (hio_recv_list == NULL) {
344 primary_exitx(EX_TEMPFAIL,
345 "Unable to allocate %zu bytes of memory for recv lists.",
346 sizeof(hio_recv_list[0]) * ncomps);
348 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
349 if (hio_recv_list_lock == NULL) {
350 primary_exitx(EX_TEMPFAIL,
351 "Unable to allocate %zu bytes of memory for recv list locks.",
352 sizeof(hio_recv_list_lock[0]) * ncomps);
354 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
355 if (hio_recv_list_cond == NULL) {
356 primary_exitx(EX_TEMPFAIL,
357 "Unable to allocate %zu bytes of memory for recv list condition variables.",
358 sizeof(hio_recv_list_cond[0]) * ncomps);
360 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
361 if (hio_remote_lock == NULL) {
362 primary_exitx(EX_TEMPFAIL,
363 "Unable to allocate %zu bytes of memory for remote connections locks.",
364 sizeof(hio_remote_lock[0]) * ncomps);
368 * Initialize lists, their locks and theirs condition variables.
370 TAILQ_INIT(&hio_free_list);
371 mtx_init(&hio_free_list_lock);
372 cv_init(&hio_free_list_cond);
373 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
374 TAILQ_INIT(&hio_send_list[ii]);
375 mtx_init(&hio_send_list_lock[ii]);
376 cv_init(&hio_send_list_cond[ii]);
377 TAILQ_INIT(&hio_recv_list[ii]);
378 mtx_init(&hio_recv_list_lock[ii]);
379 cv_init(&hio_recv_list_cond[ii]);
380 rw_init(&hio_remote_lock[ii]);
382 TAILQ_INIT(&hio_done_list);
383 mtx_init(&hio_done_list_lock);
384 cv_init(&hio_done_list_cond);
385 mtx_init(&metadata_lock);
388 * Allocate requests pool and initialize requests.
390 for (ii = 0; ii < HAST_HIO_MAX; ii++) {
391 hio = malloc(sizeof(*hio));
393 primary_exitx(EX_TEMPFAIL,
394 "Unable to allocate %zu bytes of memory for hio request.",
397 hio->hio_countdown = 0;
398 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
399 if (hio->hio_errors == NULL) {
400 primary_exitx(EX_TEMPFAIL,
401 "Unable allocate %zu bytes of memory for hio errors.",
402 sizeof(hio->hio_errors[0]) * ncomps);
404 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
405 if (hio->hio_next == NULL) {
406 primary_exitx(EX_TEMPFAIL,
407 "Unable allocate %zu bytes of memory for hio_next field.",
408 sizeof(hio->hio_next[0]) * ncomps);
410 hio->hio_ggio.gctl_version = G_GATE_VERSION;
411 hio->hio_ggio.gctl_data = malloc(MAXPHYS);
412 if (hio->hio_ggio.gctl_data == NULL) {
413 primary_exitx(EX_TEMPFAIL,
414 "Unable to allocate %zu bytes of memory for gctl_data.",
417 hio->hio_ggio.gctl_length = MAXPHYS;
418 hio->hio_ggio.gctl_error = 0;
419 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
424 init_resuid(struct hast_resource *res)
427 mtx_lock(&metadata_lock);
428 if (res->hr_resuid != 0) {
429 mtx_unlock(&metadata_lock);
432 /* Initialize unique resource identifier. */
433 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
434 mtx_unlock(&metadata_lock);
435 if (metadata_write(res) < 0)
442 init_local(struct hast_resource *res)
447 if (metadata_read(res, true) < 0)
449 mtx_init(&res->hr_amp_lock);
450 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
451 res->hr_local_sectorsize, res->hr_keepdirty) < 0) {
452 primary_exit(EX_TEMPFAIL, "Unable to create activemap");
454 mtx_init(&range_lock);
455 cv_init(&range_regular_cond);
456 if (rangelock_init(&range_regular) < 0)
457 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
458 cv_init(&range_sync_cond);
459 if (rangelock_init(&range_sync) < 0)
460 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
461 mapsize = activemap_ondisk_size(res->hr_amp);
462 buf = calloc(1, mapsize);
464 primary_exitx(EX_TEMPFAIL,
465 "Unable to allocate buffer for activemap.");
467 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
469 primary_exit(EX_NOINPUT, "Unable to read activemap");
471 activemap_copyin(res->hr_amp, buf, mapsize);
473 if (res->hr_resuid != 0)
476 * We're using provider for the first time. Initialize local and remote
477 * counters. We don't initialize resuid here, as we want to do it just
478 * in time. The reason for this is that we want to inform secondary
479 * that there were no writes yet, so there is no need to synchronize
482 res->hr_primary_localcnt = 0;
483 res->hr_primary_remotecnt = 0;
484 if (metadata_write(res) < 0)
489 primary_connect(struct hast_resource *res, struct proto_conn **connp)
491 struct proto_conn *conn;
495 if (proto_send(res->hr_conn, &val, sizeof(val)) < 0) {
496 primary_exit(EX_TEMPFAIL,
497 "Unable to send connection request to parent");
499 if (proto_recv(res->hr_conn, &val, sizeof(val)) < 0) {
500 primary_exit(EX_TEMPFAIL,
501 "Unable to receive reply to connection request from parent");
505 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
509 if (proto_connection_recv(res->hr_conn, true, &conn) < 0) {
510 primary_exit(EX_TEMPFAIL,
511 "Unable to receive connection from parent");
513 if (proto_connect_wait(conn, res->hr_timeout) < 0) {
514 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
519 /* Error in setting timeout is not critical, but why should it fail? */
520 if (proto_timeout(conn, res->hr_timeout) < 0)
521 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
529 init_remote(struct hast_resource *res, struct proto_conn **inp,
530 struct proto_conn **outp)
532 struct proto_conn *in, *out;
533 struct nv *nvout, *nvin;
534 const unsigned char *token;
543 PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
544 PJDLOG_ASSERT(real_remote(res));
549 if (primary_connect(res, &out) == -1)
550 return (ECONNREFUSED);
552 error = ECONNABORTED;
555 * First handshake step.
556 * Setup outgoing connection with remote node.
559 nv_add_string(nvout, res->hr_name, "resource");
560 if (nv_error(nvout) != 0) {
561 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
562 "Unable to allocate header for connection with %s",
567 if (hast_proto_send(res, out, nvout, NULL, 0) < 0) {
568 pjdlog_errno(LOG_WARNING,
569 "Unable to send handshake header to %s",
575 if (hast_proto_recv_hdr(out, &nvin) < 0) {
576 pjdlog_errno(LOG_WARNING,
577 "Unable to receive handshake header from %s",
581 errmsg = nv_get_string(nvin, "errmsg");
582 if (errmsg != NULL) {
583 pjdlog_warning("%s", errmsg);
584 if (nv_exists(nvin, "wait"))
589 token = nv_get_uint8_array(nvin, &size, "token");
591 pjdlog_warning("Handshake header from %s has no 'token' field.",
596 if (size != sizeof(res->hr_token)) {
597 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
598 res->hr_remoteaddr, size, sizeof(res->hr_token));
602 bcopy(token, res->hr_token, sizeof(res->hr_token));
606 * Second handshake step.
607 * Setup incoming connection with remote node.
609 if (primary_connect(res, &in) == -1)
613 nv_add_string(nvout, res->hr_name, "resource");
614 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
616 if (res->hr_resuid == 0) {
618 * The resuid field was not yet initialized.
619 * Because we do synchronization inside init_resuid(), it is
620 * possible that someone already initialized it, the function
621 * will return false then, but if we successfully initialized
622 * it, we will get true. True means that there were no writes
623 * to this resource yet and we want to inform secondary that
624 * synchronization is not needed by sending "virgin" argument.
626 if (init_resuid(res))
627 nv_add_int8(nvout, 1, "virgin");
629 nv_add_uint64(nvout, res->hr_resuid, "resuid");
630 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
631 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
632 if (nv_error(nvout) != 0) {
633 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
634 "Unable to allocate header for connection with %s",
639 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) {
640 pjdlog_errno(LOG_WARNING,
641 "Unable to send handshake header to %s",
647 if (hast_proto_recv_hdr(out, &nvin) < 0) {
648 pjdlog_errno(LOG_WARNING,
649 "Unable to receive handshake header from %s",
653 errmsg = nv_get_string(nvin, "errmsg");
654 if (errmsg != NULL) {
655 pjdlog_warning("%s", errmsg);
659 datasize = nv_get_int64(nvin, "datasize");
660 if (datasize != res->hr_datasize) {
661 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
662 (intmax_t)res->hr_datasize, (intmax_t)datasize);
666 extentsize = nv_get_int32(nvin, "extentsize");
667 if (extentsize != res->hr_extentsize) {
668 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
669 (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
673 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
674 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
675 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
676 if (nv_exists(nvin, "virgin")) {
678 * Secondary was reinitialized, bump localcnt if it is 0 as
679 * only we have the data.
681 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_PRIMARY);
682 PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
684 if (res->hr_primary_localcnt == 0) {
685 PJDLOG_ASSERT(res->hr_secondary_remotecnt == 0);
687 mtx_lock(&metadata_lock);
688 res->hr_primary_localcnt++;
689 pjdlog_debug(1, "Increasing localcnt to %ju.",
690 (uintmax_t)res->hr_primary_localcnt);
691 (void)metadata_write(res);
692 mtx_unlock(&metadata_lock);
696 mapsize = nv_get_uint32(nvin, "mapsize");
698 map = malloc(mapsize);
700 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
706 * Remote node have some dirty extents on its own, lets
707 * download its activemap.
709 if (hast_proto_recv_data(res, out, nvin, map,
711 pjdlog_errno(LOG_ERR,
712 "Unable to receive remote activemap");
718 * Merge local and remote bitmaps.
720 activemap_merge(res->hr_amp, map, mapsize);
723 * Now that we merged bitmaps from both nodes, flush it to the
724 * disk before we start to synchronize.
726 (void)hast_activemap_flush(res);
730 /* Setup directions. */
731 if (proto_send(out, NULL, 0) == -1)
732 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
733 if (proto_recv(in, NULL, 0) == -1)
734 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
736 pjdlog_info("Connected to %s.", res->hr_remoteaddr);
737 if (inp != NULL && outp != NULL) {
741 res->hr_remotein = in;
742 res->hr_remoteout = out;
744 event_send(res, EVENT_CONNECT);
747 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
748 event_send(res, EVENT_SPLITBRAIN);
759 mtx_lock(&sync_lock);
760 sync_inprogress = true;
761 mtx_unlock(&sync_lock);
762 cv_signal(&sync_cond);
769 mtx_lock(&sync_lock);
771 sync_inprogress = false;
772 mtx_unlock(&sync_lock);
776 init_ggate(struct hast_resource *res)
778 struct g_gate_ctl_create ggiocreate;
779 struct g_gate_ctl_cancel ggiocancel;
782 * We communicate with ggate via /dev/ggctl. Open it.
784 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
785 if (res->hr_ggatefd < 0)
786 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
788 * Create provider before trying to connect, as connection failure
789 * is not critical, but may take some time.
791 bzero(&ggiocreate, sizeof(ggiocreate));
792 ggiocreate.gctl_version = G_GATE_VERSION;
793 ggiocreate.gctl_mediasize = res->hr_datasize;
794 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
795 ggiocreate.gctl_flags = 0;
796 ggiocreate.gctl_maxcount = 0;
797 ggiocreate.gctl_timeout = 0;
798 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
799 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
801 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
802 pjdlog_info("Device hast/%s created.", res->hr_provname);
803 res->hr_ggateunit = ggiocreate.gctl_unit;
806 if (errno != EEXIST) {
807 primary_exit(EX_OSERR, "Unable to create hast/%s device",
811 "Device hast/%s already exists, we will try to take it over.",
814 * If we received EEXIST, we assume that the process who created the
815 * provider died and didn't clean up. In that case we will start from
818 bzero(&ggiocancel, sizeof(ggiocancel));
819 ggiocancel.gctl_version = G_GATE_VERSION;
820 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
821 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
823 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
824 pjdlog_info("Device hast/%s recovered.", res->hr_provname);
825 res->hr_ggateunit = ggiocancel.gctl_unit;
828 primary_exit(EX_OSERR, "Unable to take over hast/%s device",
833 hastd_primary(struct hast_resource *res)
837 int error, mode, debuglevel;
840 * Create communication channel for sending control commands from
843 if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) {
844 /* TODO: There's no need for this to be fatal error. */
845 KEEP_ERRNO((void)pidfile_remove(pfh));
846 pjdlog_exit(EX_OSERR,
847 "Unable to create control sockets between parent and child");
850 * Create communication channel for sending events from child to parent.
852 if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) {
853 /* TODO: There's no need for this to be fatal error. */
854 KEEP_ERRNO((void)pidfile_remove(pfh));
855 pjdlog_exit(EX_OSERR,
856 "Unable to create event sockets between child and parent");
859 * Create communication channel for sending connection requests from
862 if (proto_client(NULL, "socketpair://", &res->hr_conn) < 0) {
863 /* TODO: There's no need for this to be fatal error. */
864 KEEP_ERRNO((void)pidfile_remove(pfh));
865 pjdlog_exit(EX_OSERR,
866 "Unable to create connection sockets between child and parent");
871 /* TODO: There's no need for this to be fatal error. */
872 KEEP_ERRNO((void)pidfile_remove(pfh));
873 pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
877 /* This is parent. */
878 /* Declare that we are receiver. */
879 proto_recv(res->hr_event, NULL, 0);
880 proto_recv(res->hr_conn, NULL, 0);
881 /* Declare that we are sender. */
882 proto_send(res->hr_ctrl, NULL, 0);
883 res->hr_workerpid = pid;
888 mode = pjdlog_mode_get();
889 debuglevel = pjdlog_debug_get();
891 /* Declare that we are sender. */
892 proto_send(res->hr_event, NULL, 0);
893 proto_send(res->hr_conn, NULL, 0);
894 /* Declare that we are receiver. */
895 proto_recv(res->hr_ctrl, NULL, 0);
896 descriptors_cleanup(res);
898 descriptors_assert(res, mode);
901 pjdlog_debug_set(debuglevel);
902 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
903 setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
907 init_environment(res);
909 if (drop_privs(res) != 0) {
913 pjdlog_info("Privileges successfully dropped.");
916 * Create the guard thread first, so we can handle signals from the
919 error = pthread_create(&td, NULL, guard_thread, res);
920 PJDLOG_ASSERT(error == 0);
922 * Create the control thread before sending any event to the parent,
923 * as we can deadlock when parent sends control request to worker,
924 * but worker has no control thread started yet, so parent waits.
925 * In the meantime worker sends an event to the parent, but parent
926 * is unable to handle the event, because it waits for control
929 error = pthread_create(&td, NULL, ctrl_thread, res);
930 PJDLOG_ASSERT(error == 0);
931 if (real_remote(res)) {
932 error = init_remote(res, NULL, NULL);
935 } else if (error == EBUSY) {
936 time_t start = time(NULL);
938 pjdlog_warning("Waiting for remote node to become %s for %ds.",
939 role2str(HAST_ROLE_SECONDARY),
943 error = init_remote(res, NULL, NULL);
946 if (time(NULL) > start + res->hr_timeout)
949 if (error == EBUSY) {
950 pjdlog_warning("Remote node is still %s, starting anyway.",
951 role2str(HAST_ROLE_PRIMARY));
955 error = pthread_create(&td, NULL, ggate_recv_thread, res);
956 PJDLOG_ASSERT(error == 0);
957 error = pthread_create(&td, NULL, local_send_thread, res);
958 PJDLOG_ASSERT(error == 0);
959 error = pthread_create(&td, NULL, remote_send_thread, res);
960 PJDLOG_ASSERT(error == 0);
961 error = pthread_create(&td, NULL, remote_recv_thread, res);
962 PJDLOG_ASSERT(error == 0);
963 error = pthread_create(&td, NULL, ggate_send_thread, res);
964 PJDLOG_ASSERT(error == 0);
966 (void)sync_thread(res);
970 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...)
977 len = vsnprintf(msg, sizeof(msg), fmt, ap);
979 if ((size_t)len < sizeof(msg)) {
980 switch (ggio->gctl_cmd) {
982 (void)snprintf(msg + len, sizeof(msg) - len,
983 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
984 (uintmax_t)ggio->gctl_length);
987 (void)snprintf(msg + len, sizeof(msg) - len,
988 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
989 (uintmax_t)ggio->gctl_length);
992 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
995 (void)snprintf(msg + len, sizeof(msg) - len,
996 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
997 (uintmax_t)ggio->gctl_length);
1000 (void)snprintf(msg + len, sizeof(msg) - len,
1001 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd);
1005 pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
1009 remote_close(struct hast_resource *res, int ncomp)
1012 rw_wlock(&hio_remote_lock[ncomp]);
1014 * A race is possible between dropping rlock and acquiring wlock -
1015 * another thread can close connection in-between.
1017 if (!ISCONNECTED(res, ncomp)) {
1018 PJDLOG_ASSERT(res->hr_remotein == NULL);
1019 PJDLOG_ASSERT(res->hr_remoteout == NULL);
1020 rw_unlock(&hio_remote_lock[ncomp]);
1024 PJDLOG_ASSERT(res->hr_remotein != NULL);
1025 PJDLOG_ASSERT(res->hr_remoteout != NULL);
1027 pjdlog_debug(2, "Closing incoming connection to %s.",
1028 res->hr_remoteaddr);
1029 proto_close(res->hr_remotein);
1030 res->hr_remotein = NULL;
1031 pjdlog_debug(2, "Closing outgoing connection to %s.",
1032 res->hr_remoteaddr);
1033 proto_close(res->hr_remoteout);
1034 res->hr_remoteout = NULL;
1036 rw_unlock(&hio_remote_lock[ncomp]);
1038 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
1041 * Stop synchronization if in-progress.
1045 event_send(res, EVENT_DISCONNECT);
1049 * Thread receives ggate I/O requests from the kernel and passes them to
1050 * appropriate threads:
1051 * WRITE - always goes to both local_send and remote_send threads
1052 * READ (when the block is up-to-date on local component) -
1053 * only local_send thread
1054 * READ (when the block isn't up-to-date on local component) -
1055 * only remote_send thread
1056 * DELETE - always goes to both local_send and remote_send threads
1057 * FLUSH - always goes to both local_send and remote_send threads
1060 ggate_recv_thread(void *arg)
1062 struct hast_resource *res = arg;
1063 struct g_gate_ctl_io *ggio;
1065 unsigned int ii, ncomp, ncomps;
1068 ncomps = HAST_NCOMPONENTS;
1071 pjdlog_debug(2, "ggate_recv: Taking free request.");
1072 QUEUE_TAKE2(hio, free);
1073 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1074 ggio = &hio->hio_ggio;
1075 ggio->gctl_unit = res->hr_ggateunit;
1076 ggio->gctl_length = MAXPHYS;
1077 ggio->gctl_error = 0;
1079 "ggate_recv: (%p) Waiting for request from the kernel.",
1081 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) {
1082 if (sigexit_received)
1084 primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1086 error = ggio->gctl_error;
1091 /* Exit gracefully. */
1092 if (!sigexit_received) {
1094 "ggate_recv: (%p) Received cancel from the kernel.",
1096 pjdlog_info("Received cancel from the kernel, exiting.");
1101 * Buffer too small? Impossible, we allocate MAXPHYS
1102 * bytes - request can't be bigger than that.
1107 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1110 for (ii = 0; ii < ncomps; ii++)
1111 hio->hio_errors[ii] = EINVAL;
1112 reqlog(LOG_DEBUG, 2, ggio,
1113 "ggate_recv: (%p) Request received from the kernel: ",
1116 * Inform all components about new write request.
1117 * For read request prefer local component unless the given
1118 * range is out-of-date, then use remote component.
1120 switch (ggio->gctl_cmd) {
1122 res->hr_stat_read++;
1124 "ggate_recv: (%p) Moving request to the send queue.",
1126 refcount_init(&hio->hio_countdown, 1);
1127 mtx_lock(&metadata_lock);
1128 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1129 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1131 * This range is up-to-date on local component,
1132 * so handle request locally.
1134 /* Local component is 0 for now. */
1136 } else /* if (res->hr_syncsrc ==
1137 HAST_SYNCSRC_SECONDARY) */ {
1138 PJDLOG_ASSERT(res->hr_syncsrc ==
1139 HAST_SYNCSRC_SECONDARY);
1141 * This range is out-of-date on local component,
1142 * so send request to the remote node.
1144 /* Remote component is 1 for now. */
1147 mtx_unlock(&metadata_lock);
1148 QUEUE_INSERT1(hio, send, ncomp);
1151 res->hr_stat_write++;
1152 if (res->hr_resuid == 0) {
1154 * This is first write, initialize localcnt and
1157 res->hr_primary_localcnt = 1;
1158 (void)init_resuid(res);
1161 mtx_lock(&range_lock);
1162 if (rangelock_islocked(range_sync,
1163 ggio->gctl_offset, ggio->gctl_length)) {
1165 "regular: Range offset=%jd length=%zu locked.",
1166 (intmax_t)ggio->gctl_offset,
1167 (size_t)ggio->gctl_length);
1168 range_regular_wait = true;
1169 cv_wait(&range_regular_cond, &range_lock);
1170 range_regular_wait = false;
1171 mtx_unlock(&range_lock);
1174 if (rangelock_add(range_regular,
1175 ggio->gctl_offset, ggio->gctl_length) < 0) {
1176 mtx_unlock(&range_lock);
1178 "regular: Range offset=%jd length=%zu is already locked, waiting.",
1179 (intmax_t)ggio->gctl_offset,
1180 (size_t)ggio->gctl_length);
1184 mtx_unlock(&range_lock);
1187 mtx_lock(&res->hr_amp_lock);
1188 if (activemap_write_start(res->hr_amp,
1189 ggio->gctl_offset, ggio->gctl_length)) {
1190 res->hr_stat_activemap_update++;
1191 (void)hast_activemap_flush(res);
1193 mtx_unlock(&res->hr_amp_lock);
1197 switch (ggio->gctl_cmd) {
1199 res->hr_stat_delete++;
1202 res->hr_stat_flush++;
1206 "ggate_recv: (%p) Moving request to the send queues.",
1208 refcount_init(&hio->hio_countdown, ncomps);
1209 for (ii = 0; ii < ncomps; ii++)
1210 QUEUE_INSERT1(hio, send, ii);
1219 * Thread reads from or writes to local component.
1220 * If local read fails, it redirects it to remote_send thread.
1223 local_send_thread(void *arg)
1225 struct hast_resource *res = arg;
1226 struct g_gate_ctl_io *ggio;
1228 unsigned int ncomp, rncomp;
1231 /* Local component is 0 for now. */
1233 /* Remote component is 1 for now. */
1237 pjdlog_debug(2, "local_send: Taking request.");
1238 QUEUE_TAKE1(hio, send, ncomp, 0);
1239 pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1240 ggio = &hio->hio_ggio;
1241 switch (ggio->gctl_cmd) {
1243 ret = pread(res->hr_localfd, ggio->gctl_data,
1245 ggio->gctl_offset + res->hr_localoff);
1246 if (ret == ggio->gctl_length)
1247 hio->hio_errors[ncomp] = 0;
1248 else if (!ISSYNCREQ(hio)) {
1250 * If READ failed, try to read from remote node.
1253 reqlog(LOG_WARNING, 0, ggio,
1254 "Local request failed (%s), trying remote node. ",
1256 } else if (ret != ggio->gctl_length) {
1257 reqlog(LOG_WARNING, 0, ggio,
1258 "Local request failed (%zd != %jd), trying remote node. ",
1259 ret, (intmax_t)ggio->gctl_length);
1261 QUEUE_INSERT1(hio, send, rncomp);
1266 ret = pwrite(res->hr_localfd, ggio->gctl_data,
1268 ggio->gctl_offset + res->hr_localoff);
1270 hio->hio_errors[ncomp] = errno;
1271 reqlog(LOG_WARNING, 0, ggio,
1272 "Local request failed (%s): ",
1274 } else if (ret != ggio->gctl_length) {
1275 hio->hio_errors[ncomp] = EIO;
1276 reqlog(LOG_WARNING, 0, ggio,
1277 "Local request failed (%zd != %jd): ",
1278 ret, (intmax_t)ggio->gctl_length);
1280 hio->hio_errors[ncomp] = 0;
1284 ret = g_delete(res->hr_localfd,
1285 ggio->gctl_offset + res->hr_localoff,
1288 hio->hio_errors[ncomp] = errno;
1289 reqlog(LOG_WARNING, 0, ggio,
1290 "Local request failed (%s): ",
1293 hio->hio_errors[ncomp] = 0;
1297 ret = g_flush(res->hr_localfd);
1299 hio->hio_errors[ncomp] = errno;
1300 reqlog(LOG_WARNING, 0, ggio,
1301 "Local request failed (%s): ",
1304 hio->hio_errors[ncomp] = 0;
1308 if (refcount_release(&hio->hio_countdown)) {
1309 if (ISSYNCREQ(hio)) {
1310 mtx_lock(&sync_lock);
1312 mtx_unlock(&sync_lock);
1313 cv_signal(&sync_cond);
1316 "local_send: (%p) Moving request to the done queue.",
1318 QUEUE_INSERT2(hio, done);
1327 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1331 rw_rlock(&hio_remote_lock[ncomp]);
1333 if (!ISCONNECTED(res, ncomp)) {
1334 rw_unlock(&hio_remote_lock[ncomp]);
1338 PJDLOG_ASSERT(res->hr_remotein != NULL);
1339 PJDLOG_ASSERT(res->hr_remoteout != NULL);
1342 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1343 if (nv_error(nv) != 0) {
1344 rw_unlock(&hio_remote_lock[ncomp]);
1347 "keepalive_send: Unable to prepare header to send.");
1350 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) {
1351 rw_unlock(&hio_remote_lock[ncomp]);
1352 pjdlog_common(LOG_DEBUG, 1, errno,
1353 "keepalive_send: Unable to send request");
1355 remote_close(res, ncomp);
1359 rw_unlock(&hio_remote_lock[ncomp]);
1361 pjdlog_debug(2, "keepalive_send: Request sent.");
1365 * Thread sends request to secondary node.
1368 remote_send_thread(void *arg)
1370 struct hast_resource *res = arg;
1371 struct g_gate_ctl_io *ggio;
1372 time_t lastcheck, now;
1377 uint64_t offset, length;
1381 /* Remote component is 1 for now. */
1383 lastcheck = time(NULL);
1386 pjdlog_debug(2, "remote_send: Taking request.");
1387 QUEUE_TAKE1(hio, send, ncomp, HAST_KEEPALIVE);
1390 if (lastcheck + HAST_KEEPALIVE <= now) {
1391 keepalive_send(res, ncomp);
1396 pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1397 ggio = &hio->hio_ggio;
1398 switch (ggio->gctl_cmd) {
1402 offset = ggio->gctl_offset;
1403 length = ggio->gctl_length;
1407 data = ggio->gctl_data;
1408 offset = ggio->gctl_offset;
1409 length = ggio->gctl_length;
1414 offset = ggio->gctl_offset;
1415 length = ggio->gctl_length;
1424 PJDLOG_ASSERT(!"invalid condition");
1428 nv_add_uint8(nv, cmd, "cmd");
1429 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1430 nv_add_uint64(nv, offset, "offset");
1431 nv_add_uint64(nv, length, "length");
1432 if (nv_error(nv) != 0) {
1433 hio->hio_errors[ncomp] = nv_error(nv);
1435 "remote_send: (%p) Unable to prepare header to send.",
1437 reqlog(LOG_ERR, 0, ggio,
1438 "Unable to prepare header to send (%s): ",
1439 strerror(nv_error(nv)));
1440 /* Move failed request immediately to the done queue. */
1444 "remote_send: (%p) Moving request to the recv queue.",
1447 * Protect connection from disappearing.
1449 rw_rlock(&hio_remote_lock[ncomp]);
1450 if (!ISCONNECTED(res, ncomp)) {
1451 rw_unlock(&hio_remote_lock[ncomp]);
1452 hio->hio_errors[ncomp] = ENOTCONN;
1456 * Move the request to recv queue before sending it, because
1457 * in different order we can get reply before we move request
1460 mtx_lock(&hio_recv_list_lock[ncomp]);
1461 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1462 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1463 mtx_unlock(&hio_recv_list_lock[ncomp]);
1464 if (hast_proto_send(res, res->hr_remoteout, nv, data,
1465 data != NULL ? length : 0) < 0) {
1466 hio->hio_errors[ncomp] = errno;
1467 rw_unlock(&hio_remote_lock[ncomp]);
1469 "remote_send: (%p) Unable to send request.", hio);
1470 reqlog(LOG_ERR, 0, ggio,
1471 "Unable to send request (%s): ",
1472 strerror(hio->hio_errors[ncomp]));
1473 remote_close(res, ncomp);
1475 * Take request back from the receive queue and move
1476 * it immediately to the done queue.
1478 mtx_lock(&hio_recv_list_lock[ncomp]);
1479 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1480 mtx_unlock(&hio_recv_list_lock[ncomp]);
1483 rw_unlock(&hio_remote_lock[ncomp]);
1486 cv_signal(&hio_recv_list_cond[ncomp]);
1490 if (ISSYNCREQ(hio)) {
1491 if (!refcount_release(&hio->hio_countdown))
1493 mtx_lock(&sync_lock);
1495 mtx_unlock(&sync_lock);
1496 cv_signal(&sync_cond);
1499 if (ggio->gctl_cmd == BIO_WRITE) {
1500 mtx_lock(&res->hr_amp_lock);
1501 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1502 ggio->gctl_length)) {
1503 (void)hast_activemap_flush(res);
1505 mtx_unlock(&res->hr_amp_lock);
1507 if (!refcount_release(&hio->hio_countdown))
1510 "remote_send: (%p) Moving request to the done queue.",
1512 QUEUE_INSERT2(hio, done);
1519 * Thread receives answer from secondary node and passes it to ggate_send
1523 remote_recv_thread(void *arg)
1525 struct hast_resource *res = arg;
1526 struct g_gate_ctl_io *ggio;
1533 /* Remote component is 1 for now. */
1537 /* Wait until there is anything to receive. */
1538 mtx_lock(&hio_recv_list_lock[ncomp]);
1539 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1540 pjdlog_debug(2, "remote_recv: No requests, waiting.");
1541 cv_wait(&hio_recv_list_cond[ncomp],
1542 &hio_recv_list_lock[ncomp]);
1544 mtx_unlock(&hio_recv_list_lock[ncomp]);
1545 rw_rlock(&hio_remote_lock[ncomp]);
1546 if (!ISCONNECTED(res, ncomp)) {
1547 rw_unlock(&hio_remote_lock[ncomp]);
1549 * Connection is dead, so move all pending requests to
1550 * the done queue (one-by-one).
1552 mtx_lock(&hio_recv_list_lock[ncomp]);
1553 hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1554 PJDLOG_ASSERT(hio != NULL);
1555 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1557 mtx_unlock(&hio_recv_list_lock[ncomp]);
1560 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) {
1561 pjdlog_errno(LOG_ERR,
1562 "Unable to receive reply header");
1563 rw_unlock(&hio_remote_lock[ncomp]);
1564 remote_close(res, ncomp);
1567 rw_unlock(&hio_remote_lock[ncomp]);
1568 seq = nv_get_uint64(nv, "seq");
1570 pjdlog_error("Header contains no 'seq' field.");
1574 mtx_lock(&hio_recv_list_lock[ncomp]);
1575 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1576 if (hio->hio_ggio.gctl_seq == seq) {
1577 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1582 mtx_unlock(&hio_recv_list_lock[ncomp]);
1584 pjdlog_error("Found no request matching received 'seq' field (%ju).",
1589 error = nv_get_int16(nv, "error");
1591 /* Request failed on remote side. */
1592 hio->hio_errors[ncomp] = error;
1593 reqlog(LOG_WARNING, 0, &hio->hio_ggio,
1594 "Remote request failed (%s): ", strerror(error));
1598 ggio = &hio->hio_ggio;
1599 switch (ggio->gctl_cmd) {
1601 rw_rlock(&hio_remote_lock[ncomp]);
1602 if (!ISCONNECTED(res, ncomp)) {
1603 rw_unlock(&hio_remote_lock[ncomp]);
1607 if (hast_proto_recv_data(res, res->hr_remotein, nv,
1608 ggio->gctl_data, ggio->gctl_length) < 0) {
1609 hio->hio_errors[ncomp] = errno;
1610 pjdlog_errno(LOG_ERR,
1611 "Unable to receive reply data");
1612 rw_unlock(&hio_remote_lock[ncomp]);
1614 remote_close(res, ncomp);
1617 rw_unlock(&hio_remote_lock[ncomp]);
1624 PJDLOG_ASSERT(!"invalid condition");
1627 hio->hio_errors[ncomp] = 0;
1630 if (refcount_release(&hio->hio_countdown)) {
1631 if (ISSYNCREQ(hio)) {
1632 mtx_lock(&sync_lock);
1634 mtx_unlock(&sync_lock);
1635 cv_signal(&sync_cond);
1638 "remote_recv: (%p) Moving request to the done queue.",
1640 QUEUE_INSERT2(hio, done);
1649 * Thread sends answer to the kernel.
1652 ggate_send_thread(void *arg)
1654 struct hast_resource *res = arg;
1655 struct g_gate_ctl_io *ggio;
1657 unsigned int ii, ncomp, ncomps;
1659 ncomps = HAST_NCOMPONENTS;
1662 pjdlog_debug(2, "ggate_send: Taking request.");
1663 QUEUE_TAKE2(hio, done);
1664 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1665 ggio = &hio->hio_ggio;
1666 for (ii = 0; ii < ncomps; ii++) {
1667 if (hio->hio_errors[ii] == 0) {
1669 * One successful request is enough to declare
1672 ggio->gctl_error = 0;
1678 * None of the requests were successful.
1679 * Use the error from local component except the
1680 * case when we did only remote request.
1682 if (ggio->gctl_cmd == BIO_READ &&
1683 res->hr_syncsrc == HAST_SYNCSRC_SECONDARY)
1684 ggio->gctl_error = hio->hio_errors[1];
1686 ggio->gctl_error = hio->hio_errors[0];
1688 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1689 mtx_lock(&res->hr_amp_lock);
1690 activemap_write_complete(res->hr_amp,
1691 ggio->gctl_offset, ggio->gctl_length);
1692 mtx_unlock(&res->hr_amp_lock);
1694 if (ggio->gctl_cmd == BIO_WRITE) {
1696 * Unlock range we locked.
1698 mtx_lock(&range_lock);
1699 rangelock_del(range_regular, ggio->gctl_offset,
1701 if (range_sync_wait)
1702 cv_signal(&range_sync_cond);
1703 mtx_unlock(&range_lock);
1705 * Bump local count if this is first write after
1706 * connection failure with remote node.
1709 rw_rlock(&hio_remote_lock[ncomp]);
1710 if (!ISCONNECTED(res, ncomp)) {
1711 mtx_lock(&metadata_lock);
1712 if (res->hr_primary_localcnt ==
1713 res->hr_secondary_remotecnt) {
1714 res->hr_primary_localcnt++;
1716 "Increasing localcnt to %ju.",
1717 (uintmax_t)res->hr_primary_localcnt);
1718 (void)metadata_write(res);
1720 mtx_unlock(&metadata_lock);
1722 rw_unlock(&hio_remote_lock[ncomp]);
1724 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0)
1725 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1727 "ggate_send: (%p) Moving request to the free queue.", hio);
1728 QUEUE_INSERT2(hio, free);
1735 * Thread synchronize local and remote components.
1738 sync_thread(void *arg __unused)
1740 struct hast_resource *res = arg;
1742 struct g_gate_ctl_io *ggio;
1743 struct timeval tstart, tend, tdiff;
1744 unsigned int ii, ncomp, ncomps;
1745 off_t offset, length, synced;
1749 ncomps = HAST_NCOMPONENTS;
1755 mtx_lock(&sync_lock);
1756 if (offset >= 0 && !sync_inprogress) {
1757 gettimeofday(&tend, NULL);
1758 timersub(&tend, &tstart, &tdiff);
1759 pjdlog_info("Synchronization interrupted after %#.0T. "
1760 "%NB synchronized so far.", &tdiff,
1762 event_send(res, EVENT_SYNCINTR);
1764 while (!sync_inprogress) {
1767 cv_wait(&sync_cond, &sync_lock);
1769 mtx_unlock(&sync_lock);
1771 * Obtain offset at which we should synchronize.
1772 * Rewind synchronization if needed.
1774 mtx_lock(&res->hr_amp_lock);
1776 activemap_sync_rewind(res->hr_amp);
1777 offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1778 if (syncext != -1) {
1780 * We synchronized entire syncext extent, we can mark
1783 if (activemap_extent_complete(res->hr_amp, syncext))
1784 (void)hast_activemap_flush(res);
1786 mtx_unlock(&res->hr_amp_lock);
1790 pjdlog_info("Nodes are in sync.");
1792 pjdlog_info("Synchronization started. %NB to go.",
1793 (intmax_t)(res->hr_extentsize *
1794 activemap_ndirty(res->hr_amp)));
1795 event_send(res, EVENT_SYNCSTART);
1796 gettimeofday(&tstart, NULL);
1801 pjdlog_debug(1, "Nothing to synchronize.");
1803 * Synchronization complete, make both localcnt and
1807 rw_rlock(&hio_remote_lock[ncomp]);
1808 if (ISCONNECTED(res, ncomp)) {
1812 gettimeofday(&tend, NULL);
1813 timersub(&tend, &tstart, &tdiff);
1814 bps = (int64_t)((double)synced /
1815 ((double)tdiff.tv_sec +
1816 (double)tdiff.tv_usec / 1000000));
1817 pjdlog_info("Synchronization complete. "
1818 "%NB synchronized in %#.0lT (%NB/sec).",
1819 (intmax_t)synced, &tdiff,
1821 event_send(res, EVENT_SYNCDONE);
1823 mtx_lock(&metadata_lock);
1824 res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
1825 res->hr_primary_localcnt =
1826 res->hr_secondary_remotecnt;
1827 res->hr_primary_remotecnt =
1828 res->hr_secondary_localcnt;
1830 "Setting localcnt to %ju and remotecnt to %ju.",
1831 (uintmax_t)res->hr_primary_localcnt,
1832 (uintmax_t)res->hr_primary_remotecnt);
1833 (void)metadata_write(res);
1834 mtx_unlock(&metadata_lock);
1836 rw_unlock(&hio_remote_lock[ncomp]);
1839 pjdlog_debug(2, "sync: Taking free request.");
1840 QUEUE_TAKE2(hio, free);
1841 pjdlog_debug(2, "sync: (%p) Got free request.", hio);
1843 * Lock the range we are going to synchronize. We don't want
1844 * race where someone writes between our read and write.
1847 mtx_lock(&range_lock);
1848 if (rangelock_islocked(range_regular, offset, length)) {
1850 "sync: Range offset=%jd length=%jd locked.",
1851 (intmax_t)offset, (intmax_t)length);
1852 range_sync_wait = true;
1853 cv_wait(&range_sync_cond, &range_lock);
1854 range_sync_wait = false;
1855 mtx_unlock(&range_lock);
1858 if (rangelock_add(range_sync, offset, length) < 0) {
1859 mtx_unlock(&range_lock);
1861 "sync: Range offset=%jd length=%jd is already locked, waiting.",
1862 (intmax_t)offset, (intmax_t)length);
1866 mtx_unlock(&range_lock);
1870 * First read the data from synchronization source.
1873 ggio = &hio->hio_ggio;
1874 ggio->gctl_cmd = BIO_READ;
1875 ggio->gctl_offset = offset;
1876 ggio->gctl_length = length;
1877 ggio->gctl_error = 0;
1878 for (ii = 0; ii < ncomps; ii++)
1879 hio->hio_errors[ii] = EINVAL;
1880 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1882 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1884 mtx_lock(&metadata_lock);
1885 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1887 * This range is up-to-date on local component,
1888 * so handle request locally.
1890 /* Local component is 0 for now. */
1892 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1893 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1895 * This range is out-of-date on local component,
1896 * so send request to the remote node.
1898 /* Remote component is 1 for now. */
1901 mtx_unlock(&metadata_lock);
1902 refcount_init(&hio->hio_countdown, 1);
1903 QUEUE_INSERT1(hio, send, ncomp);
1906 * Let's wait for READ to finish.
1908 mtx_lock(&sync_lock);
1909 while (!ISSYNCREQDONE(hio))
1910 cv_wait(&sync_cond, &sync_lock);
1911 mtx_unlock(&sync_lock);
1913 if (hio->hio_errors[ncomp] != 0) {
1914 pjdlog_error("Unable to read synchronization data: %s.",
1915 strerror(hio->hio_errors[ncomp]));
1920 * We read the data from synchronization source, now write it
1921 * to synchronization target.
1924 ggio->gctl_cmd = BIO_WRITE;
1925 for (ii = 0; ii < ncomps; ii++)
1926 hio->hio_errors[ii] = EINVAL;
1927 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1929 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1931 mtx_lock(&metadata_lock);
1932 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1934 * This range is up-to-date on local component,
1935 * so we update remote component.
1937 /* Remote component is 1 for now. */
1939 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1940 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1942 * This range is out-of-date on local component,
1945 /* Local component is 0 for now. */
1948 mtx_unlock(&metadata_lock);
1950 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.",
1952 refcount_init(&hio->hio_countdown, 1);
1953 QUEUE_INSERT1(hio, send, ncomp);
1956 * Let's wait for WRITE to finish.
1958 mtx_lock(&sync_lock);
1959 while (!ISSYNCREQDONE(hio))
1960 cv_wait(&sync_cond, &sync_lock);
1961 mtx_unlock(&sync_lock);
1963 if (hio->hio_errors[ncomp] != 0) {
1964 pjdlog_error("Unable to write synchronization data: %s.",
1965 strerror(hio->hio_errors[ncomp]));
1971 mtx_lock(&range_lock);
1972 rangelock_del(range_sync, offset, length);
1973 if (range_regular_wait)
1974 cv_signal(&range_regular_cond);
1975 mtx_unlock(&range_lock);
1976 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
1978 QUEUE_INSERT2(hio, free);
1985 primary_config_reload(struct hast_resource *res, struct nv *nv)
1987 unsigned int ii, ncomps;
1991 pjdlog_info("Reloading configuration...");
1993 PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
1994 PJDLOG_ASSERT(gres == res);
1995 nv_assert(nv, "remoteaddr");
1996 nv_assert(nv, "sourceaddr");
1997 nv_assert(nv, "replication");
1998 nv_assert(nv, "checksum");
1999 nv_assert(nv, "compression");
2000 nv_assert(nv, "timeout");
2001 nv_assert(nv, "exec");
2003 ncomps = HAST_NCOMPONENTS;
2005 #define MODIFIED_REMOTEADDR 0x01
2006 #define MODIFIED_SOURCEADDR 0x02
2007 #define MODIFIED_REPLICATION 0x04
2008 #define MODIFIED_CHECKSUM 0x08
2009 #define MODIFIED_COMPRESSION 0x10
2010 #define MODIFIED_TIMEOUT 0x20
2011 #define MODIFIED_EXEC 0x40
2014 vstr = nv_get_string(nv, "remoteaddr");
2015 if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
2017 * Don't copy res->hr_remoteaddr to gres just yet.
2018 * We want remote_close() to log disconnect from the old
2019 * addresses, not from the new ones.
2021 modified |= MODIFIED_REMOTEADDR;
2023 vstr = nv_get_string(nv, "sourceaddr");
2024 if (strcmp(gres->hr_sourceaddr, vstr) != 0) {
2025 strlcpy(gres->hr_sourceaddr, vstr, sizeof(gres->hr_sourceaddr));
2026 modified |= MODIFIED_SOURCEADDR;
2028 vint = nv_get_int32(nv, "replication");
2029 if (gres->hr_replication != vint) {
2030 gres->hr_replication = vint;
2031 modified |= MODIFIED_REPLICATION;
2033 vint = nv_get_int32(nv, "checksum");
2034 if (gres->hr_checksum != vint) {
2035 gres->hr_checksum = vint;
2036 modified |= MODIFIED_CHECKSUM;
2038 vint = nv_get_int32(nv, "compression");
2039 if (gres->hr_compression != vint) {
2040 gres->hr_compression = vint;
2041 modified |= MODIFIED_COMPRESSION;
2043 vint = nv_get_int32(nv, "timeout");
2044 if (gres->hr_timeout != vint) {
2045 gres->hr_timeout = vint;
2046 modified |= MODIFIED_TIMEOUT;
2048 vstr = nv_get_string(nv, "exec");
2049 if (strcmp(gres->hr_exec, vstr) != 0) {
2050 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
2051 modified |= MODIFIED_EXEC;
2055 * Change timeout for connected sockets.
2056 * Don't bother if we need to reconnect.
2058 if ((modified & MODIFIED_TIMEOUT) != 0 &&
2059 (modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR |
2060 MODIFIED_REPLICATION)) == 0) {
2061 for (ii = 0; ii < ncomps; ii++) {
2064 rw_rlock(&hio_remote_lock[ii]);
2065 if (!ISCONNECTED(gres, ii)) {
2066 rw_unlock(&hio_remote_lock[ii]);
2069 rw_unlock(&hio_remote_lock[ii]);
2070 if (proto_timeout(gres->hr_remotein,
2071 gres->hr_timeout) < 0) {
2072 pjdlog_errno(LOG_WARNING,
2073 "Unable to set connection timeout");
2075 if (proto_timeout(gres->hr_remoteout,
2076 gres->hr_timeout) < 0) {
2077 pjdlog_errno(LOG_WARNING,
2078 "Unable to set connection timeout");
2082 if ((modified & (MODIFIED_REMOTEADDR | MODIFIED_SOURCEADDR |
2083 MODIFIED_REPLICATION)) != 0) {
2084 for (ii = 0; ii < ncomps; ii++) {
2087 remote_close(gres, ii);
2089 if (modified & MODIFIED_REMOTEADDR) {
2090 vstr = nv_get_string(nv, "remoteaddr");
2091 strlcpy(gres->hr_remoteaddr, vstr,
2092 sizeof(gres->hr_remoteaddr));
2095 #undef MODIFIED_REMOTEADDR
2096 #undef MODIFIED_SOURCEADDR
2097 #undef MODIFIED_REPLICATION
2098 #undef MODIFIED_CHECKSUM
2099 #undef MODIFIED_COMPRESSION
2100 #undef MODIFIED_TIMEOUT
2101 #undef MODIFIED_EXEC
2103 pjdlog_info("Configuration reloaded successfully.");
2107 guard_one(struct hast_resource *res, unsigned int ncomp)
2109 struct proto_conn *in, *out;
2111 if (!ISREMOTE(ncomp))
2114 rw_rlock(&hio_remote_lock[ncomp]);
2116 if (!real_remote(res)) {
2117 rw_unlock(&hio_remote_lock[ncomp]);
2121 if (ISCONNECTED(res, ncomp)) {
2122 PJDLOG_ASSERT(res->hr_remotein != NULL);
2123 PJDLOG_ASSERT(res->hr_remoteout != NULL);
2124 rw_unlock(&hio_remote_lock[ncomp]);
2125 pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2126 res->hr_remoteaddr);
2130 PJDLOG_ASSERT(res->hr_remotein == NULL);
2131 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2133 * Upgrade the lock. It doesn't have to be atomic as no other thread
2134 * can change connection status from disconnected to connected.
2136 rw_unlock(&hio_remote_lock[ncomp]);
2137 pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2138 res->hr_remoteaddr);
2140 if (init_remote(res, &in, &out) == 0) {
2141 rw_wlock(&hio_remote_lock[ncomp]);
2142 PJDLOG_ASSERT(res->hr_remotein == NULL);
2143 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2144 PJDLOG_ASSERT(in != NULL && out != NULL);
2145 res->hr_remotein = in;
2146 res->hr_remoteout = out;
2147 rw_unlock(&hio_remote_lock[ncomp]);
2148 pjdlog_info("Successfully reconnected to %s.",
2149 res->hr_remoteaddr);
2152 /* Both connections should be NULL. */
2153 PJDLOG_ASSERT(res->hr_remotein == NULL);
2154 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2155 PJDLOG_ASSERT(in == NULL && out == NULL);
2156 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2157 res->hr_remoteaddr);
2162 * Thread guards remote connections and reconnects when needed, handles
2166 guard_thread(void *arg)
2168 struct hast_resource *res = arg;
2169 unsigned int ii, ncomps;
2170 struct timespec timeout;
2171 time_t lastcheck, now;
2175 ncomps = HAST_NCOMPONENTS;
2176 lastcheck = time(NULL);
2178 PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2179 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2180 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2182 timeout.tv_sec = HAST_KEEPALIVE;
2183 timeout.tv_nsec = 0;
2190 sigexit_received = true;
2191 primary_exitx(EX_OK,
2192 "Termination signal received, exiting.");
2199 * Don't check connections until we fully started,
2200 * as we may still be looping, waiting for remote node
2201 * to switch from primary to secondary.
2204 pjdlog_debug(2, "remote_guard: Checking connections.");
2206 if (lastcheck + HAST_KEEPALIVE <= now) {
2207 for (ii = 0; ii < ncomps; ii++)
2212 signo = sigtimedwait(&mask, NULL, &timeout);