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>
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 comunicate 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 * Number of seconds to sleep between reconnect retries or keepalive packets.
156 #define RETRY_SLEEP 10
158 #define ISCONNECTED(res, no) \
159 ((res)->hr_remotein != NULL && (res)->hr_remoteout != NULL)
161 #define QUEUE_INSERT1(hio, name, ncomp) do { \
164 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \
165 _wakeup = TAILQ_EMPTY(&hio_##name##_list[(ncomp)]); \
166 TAILQ_INSERT_TAIL(&hio_##name##_list[(ncomp)], (hio), \
167 hio_next[(ncomp)]); \
168 mtx_unlock(&hio_##name##_list_lock[ncomp]); \
170 cv_signal(&hio_##name##_list_cond[(ncomp)]); \
172 #define QUEUE_INSERT2(hio, name) do { \
175 mtx_lock(&hio_##name##_list_lock); \
176 _wakeup = TAILQ_EMPTY(&hio_##name##_list); \
177 TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_##name##_next);\
178 mtx_unlock(&hio_##name##_list_lock); \
180 cv_signal(&hio_##name##_list_cond); \
182 #define QUEUE_TAKE1(hio, name, ncomp, timeout) do { \
185 mtx_lock(&hio_##name##_list_lock[(ncomp)]); \
187 while (((hio) = TAILQ_FIRST(&hio_##name##_list[(ncomp)])) == NULL && !_last) { \
188 cv_timedwait(&hio_##name##_list_cond[(ncomp)], \
189 &hio_##name##_list_lock[(ncomp)], (timeout)); \
190 if ((timeout) != 0) \
194 TAILQ_REMOVE(&hio_##name##_list[(ncomp)], (hio), \
195 hio_next[(ncomp)]); \
197 mtx_unlock(&hio_##name##_list_lock[(ncomp)]); \
199 #define QUEUE_TAKE2(hio, name) do { \
200 mtx_lock(&hio_##name##_list_lock); \
201 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \
202 cv_wait(&hio_##name##_list_cond, \
203 &hio_##name##_list_lock); \
205 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_##name##_next); \
206 mtx_unlock(&hio_##name##_list_lock); \
209 #define SYNCREQ(hio) do { \
210 (hio)->hio_ggio.gctl_unit = -1; \
211 (hio)->hio_ggio.gctl_seq = 1; \
213 #define ISSYNCREQ(hio) ((hio)->hio_ggio.gctl_unit == -1)
214 #define SYNCREQDONE(hio) do { (hio)->hio_ggio.gctl_unit = -2; } while (0)
215 #define ISSYNCREQDONE(hio) ((hio)->hio_ggio.gctl_unit == -2)
217 static struct hast_resource *gres;
219 static pthread_mutex_t range_lock;
220 static struct rangelocks *range_regular;
221 static bool range_regular_wait;
222 static pthread_cond_t range_regular_cond;
223 static struct rangelocks *range_sync;
224 static bool range_sync_wait;
225 static pthread_cond_t range_sync_cond;
227 static void *ggate_recv_thread(void *arg);
228 static void *local_send_thread(void *arg);
229 static void *remote_send_thread(void *arg);
230 static void *remote_recv_thread(void *arg);
231 static void *ggate_send_thread(void *arg);
232 static void *sync_thread(void *arg);
233 static void *guard_thread(void *arg);
236 cleanup(struct hast_resource *res)
240 /* Remember errno. */
243 /* Destroy ggate provider if we created one. */
244 if (res->hr_ggateunit >= 0) {
245 struct g_gate_ctl_destroy ggiod;
247 bzero(&ggiod, sizeof(ggiod));
248 ggiod.gctl_version = G_GATE_VERSION;
249 ggiod.gctl_unit = res->hr_ggateunit;
250 ggiod.gctl_force = 1;
251 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DESTROY, &ggiod) < 0) {
252 pjdlog_errno(LOG_WARNING,
253 "Unable to destroy hast/%s device",
256 res->hr_ggateunit = -1;
264 primary_exit(int exitcode, const char *fmt, ...)
268 PJDLOG_ASSERT(exitcode != EX_OK);
270 pjdlogv_errno(LOG_ERR, fmt, ap);
277 primary_exitx(int exitcode, const char *fmt, ...)
282 pjdlogv(exitcode == EX_OK ? LOG_INFO : LOG_ERR, fmt, ap);
289 hast_activemap_flush(struct hast_resource *res)
291 const unsigned char *buf;
294 buf = activemap_bitmap(res->hr_amp, &size);
295 PJDLOG_ASSERT(buf != NULL);
296 PJDLOG_ASSERT((size % res->hr_local_sectorsize) == 0);
297 if (pwrite(res->hr_localfd, buf, size, METADATA_SIZE) !=
299 KEEP_ERRNO(pjdlog_errno(LOG_ERR,
300 "Unable to flush activemap to disk"));
307 real_remote(const struct hast_resource *res)
310 return (strcmp(res->hr_remoteaddr, "none") != 0);
314 init_environment(struct hast_resource *res __unused)
317 unsigned int ii, ncomps;
320 * In the future it might be per-resource value.
322 ncomps = HAST_NCOMPONENTS;
325 * Allocate memory needed by lists.
327 hio_send_list = malloc(sizeof(hio_send_list[0]) * ncomps);
328 if (hio_send_list == NULL) {
329 primary_exitx(EX_TEMPFAIL,
330 "Unable to allocate %zu bytes of memory for send lists.",
331 sizeof(hio_send_list[0]) * ncomps);
333 hio_send_list_lock = malloc(sizeof(hio_send_list_lock[0]) * ncomps);
334 if (hio_send_list_lock == NULL) {
335 primary_exitx(EX_TEMPFAIL,
336 "Unable to allocate %zu bytes of memory for send list locks.",
337 sizeof(hio_send_list_lock[0]) * ncomps);
339 hio_send_list_cond = malloc(sizeof(hio_send_list_cond[0]) * ncomps);
340 if (hio_send_list_cond == NULL) {
341 primary_exitx(EX_TEMPFAIL,
342 "Unable to allocate %zu bytes of memory for send list condition variables.",
343 sizeof(hio_send_list_cond[0]) * ncomps);
345 hio_recv_list = malloc(sizeof(hio_recv_list[0]) * ncomps);
346 if (hio_recv_list == NULL) {
347 primary_exitx(EX_TEMPFAIL,
348 "Unable to allocate %zu bytes of memory for recv lists.",
349 sizeof(hio_recv_list[0]) * ncomps);
351 hio_recv_list_lock = malloc(sizeof(hio_recv_list_lock[0]) * ncomps);
352 if (hio_recv_list_lock == NULL) {
353 primary_exitx(EX_TEMPFAIL,
354 "Unable to allocate %zu bytes of memory for recv list locks.",
355 sizeof(hio_recv_list_lock[0]) * ncomps);
357 hio_recv_list_cond = malloc(sizeof(hio_recv_list_cond[0]) * ncomps);
358 if (hio_recv_list_cond == NULL) {
359 primary_exitx(EX_TEMPFAIL,
360 "Unable to allocate %zu bytes of memory for recv list condition variables.",
361 sizeof(hio_recv_list_cond[0]) * ncomps);
363 hio_remote_lock = malloc(sizeof(hio_remote_lock[0]) * ncomps);
364 if (hio_remote_lock == NULL) {
365 primary_exitx(EX_TEMPFAIL,
366 "Unable to allocate %zu bytes of memory for remote connections locks.",
367 sizeof(hio_remote_lock[0]) * ncomps);
371 * Initialize lists, their locks and theirs condition variables.
373 TAILQ_INIT(&hio_free_list);
374 mtx_init(&hio_free_list_lock);
375 cv_init(&hio_free_list_cond);
376 for (ii = 0; ii < HAST_NCOMPONENTS; ii++) {
377 TAILQ_INIT(&hio_send_list[ii]);
378 mtx_init(&hio_send_list_lock[ii]);
379 cv_init(&hio_send_list_cond[ii]);
380 TAILQ_INIT(&hio_recv_list[ii]);
381 mtx_init(&hio_recv_list_lock[ii]);
382 cv_init(&hio_recv_list_cond[ii]);
383 rw_init(&hio_remote_lock[ii]);
385 TAILQ_INIT(&hio_done_list);
386 mtx_init(&hio_done_list_lock);
387 cv_init(&hio_done_list_cond);
388 mtx_init(&metadata_lock);
391 * Allocate requests pool and initialize requests.
393 for (ii = 0; ii < HAST_HIO_MAX; ii++) {
394 hio = malloc(sizeof(*hio));
396 primary_exitx(EX_TEMPFAIL,
397 "Unable to allocate %zu bytes of memory for hio request.",
400 hio->hio_countdown = 0;
401 hio->hio_errors = malloc(sizeof(hio->hio_errors[0]) * ncomps);
402 if (hio->hio_errors == NULL) {
403 primary_exitx(EX_TEMPFAIL,
404 "Unable allocate %zu bytes of memory for hio errors.",
405 sizeof(hio->hio_errors[0]) * ncomps);
407 hio->hio_next = malloc(sizeof(hio->hio_next[0]) * ncomps);
408 if (hio->hio_next == NULL) {
409 primary_exitx(EX_TEMPFAIL,
410 "Unable allocate %zu bytes of memory for hio_next field.",
411 sizeof(hio->hio_next[0]) * ncomps);
413 hio->hio_ggio.gctl_version = G_GATE_VERSION;
414 hio->hio_ggio.gctl_data = malloc(MAXPHYS);
415 if (hio->hio_ggio.gctl_data == NULL) {
416 primary_exitx(EX_TEMPFAIL,
417 "Unable to allocate %zu bytes of memory for gctl_data.",
420 hio->hio_ggio.gctl_length = MAXPHYS;
421 hio->hio_ggio.gctl_error = 0;
422 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_free_next);
427 init_resuid(struct hast_resource *res)
430 mtx_lock(&metadata_lock);
431 if (res->hr_resuid != 0) {
432 mtx_unlock(&metadata_lock);
435 /* Initialize unique resource identifier. */
436 arc4random_buf(&res->hr_resuid, sizeof(res->hr_resuid));
437 mtx_unlock(&metadata_lock);
438 if (metadata_write(res) < 0)
445 init_local(struct hast_resource *res)
450 if (metadata_read(res, true) < 0)
452 mtx_init(&res->hr_amp_lock);
453 if (activemap_init(&res->hr_amp, res->hr_datasize, res->hr_extentsize,
454 res->hr_local_sectorsize, res->hr_keepdirty) < 0) {
455 primary_exit(EX_TEMPFAIL, "Unable to create activemap");
457 mtx_init(&range_lock);
458 cv_init(&range_regular_cond);
459 if (rangelock_init(&range_regular) < 0)
460 primary_exit(EX_TEMPFAIL, "Unable to create regular range lock");
461 cv_init(&range_sync_cond);
462 if (rangelock_init(&range_sync) < 0)
463 primary_exit(EX_TEMPFAIL, "Unable to create sync range lock");
464 mapsize = activemap_ondisk_size(res->hr_amp);
465 buf = calloc(1, mapsize);
467 primary_exitx(EX_TEMPFAIL,
468 "Unable to allocate buffer for activemap.");
470 if (pread(res->hr_localfd, buf, mapsize, METADATA_SIZE) !=
472 primary_exit(EX_NOINPUT, "Unable to read activemap");
474 activemap_copyin(res->hr_amp, buf, mapsize);
476 if (res->hr_resuid != 0)
479 * We're using provider for the first time. Initialize local and remote
480 * counters. We don't initialize resuid here, as we want to do it just
481 * in time. The reason for this is that we want to inform secondary
482 * that there were no writes yet, so there is no need to synchronize
485 res->hr_primary_localcnt = 1;
486 res->hr_primary_remotecnt = 0;
487 if (metadata_write(res) < 0)
492 primary_connect(struct hast_resource *res, struct proto_conn **connp)
494 struct proto_conn *conn;
498 if (proto_send(res->hr_conn, &val, sizeof(val)) < 0) {
499 primary_exit(EX_TEMPFAIL,
500 "Unable to send connection request to parent");
502 if (proto_recv(res->hr_conn, &val, sizeof(val)) < 0) {
503 primary_exit(EX_TEMPFAIL,
504 "Unable to receive reply to connection request from parent");
508 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
512 if (proto_connection_recv(res->hr_conn, true, &conn) < 0) {
513 primary_exit(EX_TEMPFAIL,
514 "Unable to receive connection from parent");
516 if (proto_connect_wait(conn, HAST_TIMEOUT) < 0) {
517 pjdlog_errno(LOG_WARNING, "Unable to connect to %s",
522 /* Error in setting timeout is not critical, but why should it fail? */
523 if (proto_timeout(conn, res->hr_timeout) < 0)
524 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
532 init_remote(struct hast_resource *res, struct proto_conn **inp,
533 struct proto_conn **outp)
535 struct proto_conn *in, *out;
536 struct nv *nvout, *nvin;
537 const unsigned char *token;
545 PJDLOG_ASSERT((inp == NULL && outp == NULL) || (inp != NULL && outp != NULL));
546 PJDLOG_ASSERT(real_remote(res));
551 if (primary_connect(res, &out) == -1)
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);
587 token = nv_get_uint8_array(nvin, &size, "token");
589 pjdlog_warning("Handshake header from %s has no 'token' field.",
594 if (size != sizeof(res->hr_token)) {
595 pjdlog_warning("Handshake header from %s contains 'token' of wrong size (got %zu, expected %zu).",
596 res->hr_remoteaddr, size, sizeof(res->hr_token));
600 bcopy(token, res->hr_token, sizeof(res->hr_token));
604 * Second handshake step.
605 * Setup incoming connection with remote node.
607 if (primary_connect(res, &in) == -1)
611 nv_add_string(nvout, res->hr_name, "resource");
612 nv_add_uint8_array(nvout, res->hr_token, sizeof(res->hr_token),
614 if (res->hr_resuid == 0) {
616 * The resuid field was not yet initialized.
617 * Because we do synchronization inside init_resuid(), it is
618 * possible that someone already initialized it, the function
619 * will return false then, but if we successfully initialized
620 * it, we will get true. True means that there were no writes
621 * to this resource yet and we want to inform secondary that
622 * synchronization is not needed by sending "virgin" argument.
624 if (init_resuid(res))
625 nv_add_int8(nvout, 1, "virgin");
627 nv_add_uint64(nvout, res->hr_resuid, "resuid");
628 nv_add_uint64(nvout, res->hr_primary_localcnt, "localcnt");
629 nv_add_uint64(nvout, res->hr_primary_remotecnt, "remotecnt");
630 if (nv_error(nvout) != 0) {
631 pjdlog_common(LOG_WARNING, 0, nv_error(nvout),
632 "Unable to allocate header for connection with %s",
637 if (hast_proto_send(res, in, nvout, NULL, 0) < 0) {
638 pjdlog_errno(LOG_WARNING,
639 "Unable to send handshake header to %s",
645 if (hast_proto_recv_hdr(out, &nvin) < 0) {
646 pjdlog_errno(LOG_WARNING,
647 "Unable to receive handshake header from %s",
651 errmsg = nv_get_string(nvin, "errmsg");
652 if (errmsg != NULL) {
653 pjdlog_warning("%s", errmsg);
657 datasize = nv_get_int64(nvin, "datasize");
658 if (datasize != res->hr_datasize) {
659 pjdlog_warning("Data size differs between nodes (local=%jd, remote=%jd).",
660 (intmax_t)res->hr_datasize, (intmax_t)datasize);
664 extentsize = nv_get_int32(nvin, "extentsize");
665 if (extentsize != res->hr_extentsize) {
666 pjdlog_warning("Extent size differs between nodes (local=%zd, remote=%zd).",
667 (ssize_t)res->hr_extentsize, (ssize_t)extentsize);
671 res->hr_secondary_localcnt = nv_get_uint64(nvin, "localcnt");
672 res->hr_secondary_remotecnt = nv_get_uint64(nvin, "remotecnt");
673 res->hr_syncsrc = nv_get_uint8(nvin, "syncsrc");
675 mapsize = nv_get_uint32(nvin, "mapsize");
677 map = malloc(mapsize);
679 pjdlog_error("Unable to allocate memory for remote activemap (mapsize=%ju).",
685 * Remote node have some dirty extents on its own, lets
686 * download its activemap.
688 if (hast_proto_recv_data(res, out, nvin, map,
690 pjdlog_errno(LOG_ERR,
691 "Unable to receive remote activemap");
697 * Merge local and remote bitmaps.
699 activemap_merge(res->hr_amp, map, mapsize);
702 * Now that we merged bitmaps from both nodes, flush it to the
703 * disk before we start to synchronize.
705 (void)hast_activemap_flush(res);
708 pjdlog_info("Connected to %s.", res->hr_remoteaddr);
709 if (inp != NULL && outp != NULL) {
713 res->hr_remotein = in;
714 res->hr_remoteout = out;
716 event_send(res, EVENT_CONNECT);
719 if (errmsg != NULL && strcmp(errmsg, "Split-brain condition!") == 0)
720 event_send(res, EVENT_SPLITBRAIN);
731 mtx_lock(&sync_lock);
732 sync_inprogress = true;
733 mtx_unlock(&sync_lock);
734 cv_signal(&sync_cond);
741 mtx_lock(&sync_lock);
743 sync_inprogress = false;
744 mtx_unlock(&sync_lock);
748 init_ggate(struct hast_resource *res)
750 struct g_gate_ctl_create ggiocreate;
751 struct g_gate_ctl_cancel ggiocancel;
754 * We communicate with ggate via /dev/ggctl. Open it.
756 res->hr_ggatefd = open("/dev/" G_GATE_CTL_NAME, O_RDWR);
757 if (res->hr_ggatefd < 0)
758 primary_exit(EX_OSFILE, "Unable to open /dev/" G_GATE_CTL_NAME);
760 * Create provider before trying to connect, as connection failure
761 * is not critical, but may take some time.
763 bzero(&ggiocreate, sizeof(ggiocreate));
764 ggiocreate.gctl_version = G_GATE_VERSION;
765 ggiocreate.gctl_mediasize = res->hr_datasize;
766 ggiocreate.gctl_sectorsize = res->hr_local_sectorsize;
767 ggiocreate.gctl_flags = 0;
768 ggiocreate.gctl_maxcount = G_GATE_MAX_QUEUE_SIZE;
769 ggiocreate.gctl_timeout = 0;
770 ggiocreate.gctl_unit = G_GATE_NAME_GIVEN;
771 snprintf(ggiocreate.gctl_name, sizeof(ggiocreate.gctl_name), "hast/%s",
773 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CREATE, &ggiocreate) == 0) {
774 pjdlog_info("Device hast/%s created.", res->hr_provname);
775 res->hr_ggateunit = ggiocreate.gctl_unit;
778 if (errno != EEXIST) {
779 primary_exit(EX_OSERR, "Unable to create hast/%s device",
783 "Device hast/%s already exists, we will try to take it over.",
786 * If we received EEXIST, we assume that the process who created the
787 * provider died and didn't clean up. In that case we will start from
790 bzero(&ggiocancel, sizeof(ggiocancel));
791 ggiocancel.gctl_version = G_GATE_VERSION;
792 ggiocancel.gctl_unit = G_GATE_NAME_GIVEN;
793 snprintf(ggiocancel.gctl_name, sizeof(ggiocancel.gctl_name), "hast/%s",
795 if (ioctl(res->hr_ggatefd, G_GATE_CMD_CANCEL, &ggiocancel) == 0) {
796 pjdlog_info("Device hast/%s recovered.", res->hr_provname);
797 res->hr_ggateunit = ggiocancel.gctl_unit;
800 primary_exit(EX_OSERR, "Unable to take over hast/%s device",
805 hastd_primary(struct hast_resource *res)
812 * Create communication channel for sending control commands from
815 if (proto_client("socketpair://", &res->hr_ctrl) < 0) {
816 /* TODO: There's no need for this to be fatal error. */
817 KEEP_ERRNO((void)pidfile_remove(pfh));
818 pjdlog_exit(EX_OSERR,
819 "Unable to create control sockets between parent and child");
822 * Create communication channel for sending events from child to parent.
824 if (proto_client("socketpair://", &res->hr_event) < 0) {
825 /* TODO: There's no need for this to be fatal error. */
826 KEEP_ERRNO((void)pidfile_remove(pfh));
827 pjdlog_exit(EX_OSERR,
828 "Unable to create event sockets between child and parent");
831 * Create communication channel for sending connection requests from
834 if (proto_client("socketpair://", &res->hr_conn) < 0) {
835 /* TODO: There's no need for this to be fatal error. */
836 KEEP_ERRNO((void)pidfile_remove(pfh));
837 pjdlog_exit(EX_OSERR,
838 "Unable to create connection sockets between child and parent");
843 /* TODO: There's no need for this to be fatal error. */
844 KEEP_ERRNO((void)pidfile_remove(pfh));
845 pjdlog_exit(EX_TEMPFAIL, "Unable to fork");
849 /* This is parent. */
850 /* Declare that we are receiver. */
851 proto_recv(res->hr_event, NULL, 0);
852 proto_recv(res->hr_conn, NULL, 0);
853 /* Declare that we are sender. */
854 proto_send(res->hr_ctrl, NULL, 0);
855 res->hr_workerpid = pid;
860 mode = pjdlog_mode_get();
862 /* Declare that we are sender. */
863 proto_send(res->hr_event, NULL, 0);
864 proto_send(res->hr_conn, NULL, 0);
865 /* Declare that we are receiver. */
866 proto_recv(res->hr_ctrl, NULL, 0);
867 descriptors_cleanup(res);
869 descriptors_assert(res, mode);
872 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
873 setproctitle("%s (primary)", res->hr_name);
877 init_environment(res);
879 if (drop_privs() != 0) {
883 pjdlog_info("Privileges successfully dropped.");
886 * Create the guard thread first, so we can handle signals from the
889 error = pthread_create(&td, NULL, guard_thread, res);
890 PJDLOG_ASSERT(error == 0);
892 * Create the control thread before sending any event to the parent,
893 * as we can deadlock when parent sends control request to worker,
894 * but worker has no control thread started yet, so parent waits.
895 * In the meantime worker sends an event to the parent, but parent
896 * is unable to handle the event, because it waits for control
899 error = pthread_create(&td, NULL, ctrl_thread, res);
900 PJDLOG_ASSERT(error == 0);
901 if (real_remote(res) && init_remote(res, NULL, NULL))
903 error = pthread_create(&td, NULL, ggate_recv_thread, res);
904 PJDLOG_ASSERT(error == 0);
905 error = pthread_create(&td, NULL, local_send_thread, res);
906 PJDLOG_ASSERT(error == 0);
907 error = pthread_create(&td, NULL, remote_send_thread, res);
908 PJDLOG_ASSERT(error == 0);
909 error = pthread_create(&td, NULL, remote_recv_thread, res);
910 PJDLOG_ASSERT(error == 0);
911 error = pthread_create(&td, NULL, ggate_send_thread, res);
912 PJDLOG_ASSERT(error == 0);
913 (void)sync_thread(res);
917 reqlog(int loglevel, int debuglevel, struct g_gate_ctl_io *ggio, const char *fmt, ...)
924 len = vsnprintf(msg, sizeof(msg), fmt, ap);
926 if ((size_t)len < sizeof(msg)) {
927 switch (ggio->gctl_cmd) {
929 (void)snprintf(msg + len, sizeof(msg) - len,
930 "READ(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
931 (uintmax_t)ggio->gctl_length);
934 (void)snprintf(msg + len, sizeof(msg) - len,
935 "DELETE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
936 (uintmax_t)ggio->gctl_length);
939 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
942 (void)snprintf(msg + len, sizeof(msg) - len,
943 "WRITE(%ju, %ju).", (uintmax_t)ggio->gctl_offset,
944 (uintmax_t)ggio->gctl_length);
947 (void)snprintf(msg + len, sizeof(msg) - len,
948 "UNKNOWN(%u).", (unsigned int)ggio->gctl_cmd);
952 pjdlog_common(loglevel, debuglevel, -1, "%s", msg);
956 remote_close(struct hast_resource *res, int ncomp)
959 rw_wlock(&hio_remote_lock[ncomp]);
961 * A race is possible between dropping rlock and acquiring wlock -
962 * another thread can close connection in-between.
964 if (!ISCONNECTED(res, ncomp)) {
965 PJDLOG_ASSERT(res->hr_remotein == NULL);
966 PJDLOG_ASSERT(res->hr_remoteout == NULL);
967 rw_unlock(&hio_remote_lock[ncomp]);
971 PJDLOG_ASSERT(res->hr_remotein != NULL);
972 PJDLOG_ASSERT(res->hr_remoteout != NULL);
974 pjdlog_debug(2, "Closing incoming connection to %s.",
976 proto_close(res->hr_remotein);
977 res->hr_remotein = NULL;
978 pjdlog_debug(2, "Closing outgoing connection to %s.",
980 proto_close(res->hr_remoteout);
981 res->hr_remoteout = NULL;
983 rw_unlock(&hio_remote_lock[ncomp]);
985 pjdlog_warning("Disconnected from %s.", res->hr_remoteaddr);
988 * Stop synchronization if in-progress.
992 event_send(res, EVENT_DISCONNECT);
996 * Thread receives ggate I/O requests from the kernel and passes them to
997 * appropriate threads:
998 * WRITE - always goes to both local_send and remote_send threads
999 * READ (when the block is up-to-date on local component) -
1000 * only local_send thread
1001 * READ (when the block isn't up-to-date on local component) -
1002 * only remote_send thread
1003 * DELETE - always goes to both local_send and remote_send threads
1004 * FLUSH - always goes to both local_send and remote_send threads
1007 ggate_recv_thread(void *arg)
1009 struct hast_resource *res = arg;
1010 struct g_gate_ctl_io *ggio;
1012 unsigned int ii, ncomp, ncomps;
1015 ncomps = HAST_NCOMPONENTS;
1018 pjdlog_debug(2, "ggate_recv: Taking free request.");
1019 QUEUE_TAKE2(hio, free);
1020 pjdlog_debug(2, "ggate_recv: (%p) Got free request.", hio);
1021 ggio = &hio->hio_ggio;
1022 ggio->gctl_unit = res->hr_ggateunit;
1023 ggio->gctl_length = MAXPHYS;
1024 ggio->gctl_error = 0;
1026 "ggate_recv: (%p) Waiting for request from the kernel.",
1028 if (ioctl(res->hr_ggatefd, G_GATE_CMD_START, ggio) < 0) {
1029 if (sigexit_received)
1031 primary_exit(EX_OSERR, "G_GATE_CMD_START failed");
1033 error = ggio->gctl_error;
1038 /* Exit gracefully. */
1039 if (!sigexit_received) {
1041 "ggate_recv: (%p) Received cancel from the kernel.",
1043 pjdlog_info("Received cancel from the kernel, exiting.");
1048 * Buffer too small? Impossible, we allocate MAXPHYS
1049 * bytes - request can't be bigger than that.
1054 primary_exitx(EX_OSERR, "G_GATE_CMD_START failed: %s.",
1057 for (ii = 0; ii < ncomps; ii++)
1058 hio->hio_errors[ii] = EINVAL;
1059 reqlog(LOG_DEBUG, 2, ggio,
1060 "ggate_recv: (%p) Request received from the kernel: ",
1063 * Inform all components about new write request.
1064 * For read request prefer local component unless the given
1065 * range is out-of-date, then use remote component.
1067 switch (ggio->gctl_cmd) {
1070 "ggate_recv: (%p) Moving request to the send queue.",
1072 refcount_init(&hio->hio_countdown, 1);
1073 mtx_lock(&metadata_lock);
1074 if (res->hr_syncsrc == HAST_SYNCSRC_UNDEF ||
1075 res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1077 * This range is up-to-date on local component,
1078 * so handle request locally.
1080 /* Local component is 0 for now. */
1082 } else /* if (res->hr_syncsrc ==
1083 HAST_SYNCSRC_SECONDARY) */ {
1084 PJDLOG_ASSERT(res->hr_syncsrc ==
1085 HAST_SYNCSRC_SECONDARY);
1087 * This range is out-of-date on local component,
1088 * so send request to the remote node.
1090 /* Remote component is 1 for now. */
1093 mtx_unlock(&metadata_lock);
1094 QUEUE_INSERT1(hio, send, ncomp);
1097 if (res->hr_resuid == 0) {
1098 /* This is first write, initialize resuid. */
1099 (void)init_resuid(res);
1102 mtx_lock(&range_lock);
1103 if (rangelock_islocked(range_sync,
1104 ggio->gctl_offset, ggio->gctl_length)) {
1106 "regular: Range offset=%jd length=%zu locked.",
1107 (intmax_t)ggio->gctl_offset,
1108 (size_t)ggio->gctl_length);
1109 range_regular_wait = true;
1110 cv_wait(&range_regular_cond, &range_lock);
1111 range_regular_wait = false;
1112 mtx_unlock(&range_lock);
1115 if (rangelock_add(range_regular,
1116 ggio->gctl_offset, ggio->gctl_length) < 0) {
1117 mtx_unlock(&range_lock);
1119 "regular: Range offset=%jd length=%zu is already locked, waiting.",
1120 (intmax_t)ggio->gctl_offset,
1121 (size_t)ggio->gctl_length);
1125 mtx_unlock(&range_lock);
1128 mtx_lock(&res->hr_amp_lock);
1129 if (activemap_write_start(res->hr_amp,
1130 ggio->gctl_offset, ggio->gctl_length)) {
1131 (void)hast_activemap_flush(res);
1133 mtx_unlock(&res->hr_amp_lock);
1138 "ggate_recv: (%p) Moving request to the send queues.",
1140 refcount_init(&hio->hio_countdown, ncomps);
1141 for (ii = 0; ii < ncomps; ii++)
1142 QUEUE_INSERT1(hio, send, ii);
1151 * Thread reads from or writes to local component.
1152 * If local read fails, it redirects it to remote_send thread.
1155 local_send_thread(void *arg)
1157 struct hast_resource *res = arg;
1158 struct g_gate_ctl_io *ggio;
1160 unsigned int ncomp, rncomp;
1163 /* Local component is 0 for now. */
1165 /* Remote component is 1 for now. */
1169 pjdlog_debug(2, "local_send: Taking request.");
1170 QUEUE_TAKE1(hio, send, ncomp, 0);
1171 pjdlog_debug(2, "local_send: (%p) Got request.", hio);
1172 ggio = &hio->hio_ggio;
1173 switch (ggio->gctl_cmd) {
1175 ret = pread(res->hr_localfd, ggio->gctl_data,
1177 ggio->gctl_offset + res->hr_localoff);
1178 if (ret == ggio->gctl_length)
1179 hio->hio_errors[ncomp] = 0;
1182 * If READ failed, try to read from remote node.
1185 reqlog(LOG_WARNING, 0, ggio,
1186 "Local request failed (%s), trying remote node. ",
1188 } else if (ret != ggio->gctl_length) {
1189 reqlog(LOG_WARNING, 0, ggio,
1190 "Local request failed (%zd != %jd), trying remote node. ",
1191 ret, (intmax_t)ggio->gctl_length);
1193 QUEUE_INSERT1(hio, send, rncomp);
1198 ret = pwrite(res->hr_localfd, ggio->gctl_data,
1200 ggio->gctl_offset + res->hr_localoff);
1202 hio->hio_errors[ncomp] = errno;
1203 reqlog(LOG_WARNING, 0, ggio,
1204 "Local request failed (%s): ",
1206 } else if (ret != ggio->gctl_length) {
1207 hio->hio_errors[ncomp] = EIO;
1208 reqlog(LOG_WARNING, 0, ggio,
1209 "Local request failed (%zd != %jd): ",
1210 ret, (intmax_t)ggio->gctl_length);
1212 hio->hio_errors[ncomp] = 0;
1216 ret = g_delete(res->hr_localfd,
1217 ggio->gctl_offset + res->hr_localoff,
1220 hio->hio_errors[ncomp] = errno;
1221 reqlog(LOG_WARNING, 0, ggio,
1222 "Local request failed (%s): ",
1225 hio->hio_errors[ncomp] = 0;
1229 ret = g_flush(res->hr_localfd);
1231 hio->hio_errors[ncomp] = errno;
1232 reqlog(LOG_WARNING, 0, ggio,
1233 "Local request failed (%s): ",
1236 hio->hio_errors[ncomp] = 0;
1240 if (refcount_release(&hio->hio_countdown)) {
1241 if (ISSYNCREQ(hio)) {
1242 mtx_lock(&sync_lock);
1244 mtx_unlock(&sync_lock);
1245 cv_signal(&sync_cond);
1248 "local_send: (%p) Moving request to the done queue.",
1250 QUEUE_INSERT2(hio, done);
1259 keepalive_send(struct hast_resource *res, unsigned int ncomp)
1263 rw_rlock(&hio_remote_lock[ncomp]);
1265 if (!ISCONNECTED(res, ncomp)) {
1266 rw_unlock(&hio_remote_lock[ncomp]);
1270 PJDLOG_ASSERT(res->hr_remotein != NULL);
1271 PJDLOG_ASSERT(res->hr_remoteout != NULL);
1274 nv_add_uint8(nv, HIO_KEEPALIVE, "cmd");
1275 if (nv_error(nv) != 0) {
1276 rw_unlock(&hio_remote_lock[ncomp]);
1279 "keepalive_send: Unable to prepare header to send.");
1282 if (hast_proto_send(res, res->hr_remoteout, nv, NULL, 0) < 0) {
1283 rw_unlock(&hio_remote_lock[ncomp]);
1284 pjdlog_common(LOG_DEBUG, 1, errno,
1285 "keepalive_send: Unable to send request");
1287 remote_close(res, ncomp);
1291 rw_unlock(&hio_remote_lock[ncomp]);
1293 pjdlog_debug(2, "keepalive_send: Request sent.");
1297 * Thread sends request to secondary node.
1300 remote_send_thread(void *arg)
1302 struct hast_resource *res = arg;
1303 struct g_gate_ctl_io *ggio;
1304 time_t lastcheck, now;
1309 uint64_t offset, length;
1313 /* Remote component is 1 for now. */
1315 lastcheck = time(NULL);
1318 pjdlog_debug(2, "remote_send: Taking request.");
1319 QUEUE_TAKE1(hio, send, ncomp, RETRY_SLEEP);
1322 if (lastcheck + RETRY_SLEEP <= now) {
1323 keepalive_send(res, ncomp);
1328 pjdlog_debug(2, "remote_send: (%p) Got request.", hio);
1329 ggio = &hio->hio_ggio;
1330 switch (ggio->gctl_cmd) {
1334 offset = ggio->gctl_offset;
1335 length = ggio->gctl_length;
1339 data = ggio->gctl_data;
1340 offset = ggio->gctl_offset;
1341 length = ggio->gctl_length;
1346 offset = ggio->gctl_offset;
1347 length = ggio->gctl_length;
1356 PJDLOG_ASSERT(!"invalid condition");
1360 nv_add_uint8(nv, cmd, "cmd");
1361 nv_add_uint64(nv, (uint64_t)ggio->gctl_seq, "seq");
1362 nv_add_uint64(nv, offset, "offset");
1363 nv_add_uint64(nv, length, "length");
1364 if (nv_error(nv) != 0) {
1365 hio->hio_errors[ncomp] = nv_error(nv);
1367 "remote_send: (%p) Unable to prepare header to send.",
1369 reqlog(LOG_ERR, 0, ggio,
1370 "Unable to prepare header to send (%s): ",
1371 strerror(nv_error(nv)));
1372 /* Move failed request immediately to the done queue. */
1376 "remote_send: (%p) Moving request to the recv queue.",
1379 * Protect connection from disappearing.
1381 rw_rlock(&hio_remote_lock[ncomp]);
1382 if (!ISCONNECTED(res, ncomp)) {
1383 rw_unlock(&hio_remote_lock[ncomp]);
1384 hio->hio_errors[ncomp] = ENOTCONN;
1388 * Move the request to recv queue before sending it, because
1389 * in different order we can get reply before we move request
1392 mtx_lock(&hio_recv_list_lock[ncomp]);
1393 wakeup = TAILQ_EMPTY(&hio_recv_list[ncomp]);
1394 TAILQ_INSERT_TAIL(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1395 mtx_unlock(&hio_recv_list_lock[ncomp]);
1396 if (hast_proto_send(res, res->hr_remoteout, nv, data,
1397 data != NULL ? length : 0) < 0) {
1398 hio->hio_errors[ncomp] = errno;
1399 rw_unlock(&hio_remote_lock[ncomp]);
1401 "remote_send: (%p) Unable to send request.", hio);
1402 reqlog(LOG_ERR, 0, ggio,
1403 "Unable to send request (%s): ",
1404 strerror(hio->hio_errors[ncomp]));
1405 remote_close(res, ncomp);
1407 * Take request back from the receive queue and move
1408 * it immediately to the done queue.
1410 mtx_lock(&hio_recv_list_lock[ncomp]);
1411 TAILQ_REMOVE(&hio_recv_list[ncomp], hio, hio_next[ncomp]);
1412 mtx_unlock(&hio_recv_list_lock[ncomp]);
1415 rw_unlock(&hio_remote_lock[ncomp]);
1418 cv_signal(&hio_recv_list_cond[ncomp]);
1422 if (ISSYNCREQ(hio)) {
1423 if (!refcount_release(&hio->hio_countdown))
1425 mtx_lock(&sync_lock);
1427 mtx_unlock(&sync_lock);
1428 cv_signal(&sync_cond);
1431 if (ggio->gctl_cmd == BIO_WRITE) {
1432 mtx_lock(&res->hr_amp_lock);
1433 if (activemap_need_sync(res->hr_amp, ggio->gctl_offset,
1434 ggio->gctl_length)) {
1435 (void)hast_activemap_flush(res);
1437 mtx_unlock(&res->hr_amp_lock);
1439 if (!refcount_release(&hio->hio_countdown))
1442 "remote_send: (%p) Moving request to the done queue.",
1444 QUEUE_INSERT2(hio, done);
1451 * Thread receives answer from secondary node and passes it to ggate_send
1455 remote_recv_thread(void *arg)
1457 struct hast_resource *res = arg;
1458 struct g_gate_ctl_io *ggio;
1465 /* Remote component is 1 for now. */
1469 /* Wait until there is anything to receive. */
1470 mtx_lock(&hio_recv_list_lock[ncomp]);
1471 while (TAILQ_EMPTY(&hio_recv_list[ncomp])) {
1472 pjdlog_debug(2, "remote_recv: No requests, waiting.");
1473 cv_wait(&hio_recv_list_cond[ncomp],
1474 &hio_recv_list_lock[ncomp]);
1476 mtx_unlock(&hio_recv_list_lock[ncomp]);
1477 rw_rlock(&hio_remote_lock[ncomp]);
1478 if (!ISCONNECTED(res, ncomp)) {
1479 rw_unlock(&hio_remote_lock[ncomp]);
1481 * Connection is dead, so move all pending requests to
1482 * the done queue (one-by-one).
1484 mtx_lock(&hio_recv_list_lock[ncomp]);
1485 hio = TAILQ_FIRST(&hio_recv_list[ncomp]);
1486 PJDLOG_ASSERT(hio != NULL);
1487 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1489 mtx_unlock(&hio_recv_list_lock[ncomp]);
1492 if (hast_proto_recv_hdr(res->hr_remotein, &nv) < 0) {
1493 pjdlog_errno(LOG_ERR,
1494 "Unable to receive reply header");
1495 rw_unlock(&hio_remote_lock[ncomp]);
1496 remote_close(res, ncomp);
1499 rw_unlock(&hio_remote_lock[ncomp]);
1500 seq = nv_get_uint64(nv, "seq");
1502 pjdlog_error("Header contains no 'seq' field.");
1506 mtx_lock(&hio_recv_list_lock[ncomp]);
1507 TAILQ_FOREACH(hio, &hio_recv_list[ncomp], hio_next[ncomp]) {
1508 if (hio->hio_ggio.gctl_seq == seq) {
1509 TAILQ_REMOVE(&hio_recv_list[ncomp], hio,
1514 mtx_unlock(&hio_recv_list_lock[ncomp]);
1516 pjdlog_error("Found no request matching received 'seq' field (%ju).",
1521 error = nv_get_int16(nv, "error");
1523 /* Request failed on remote side. */
1524 hio->hio_errors[ncomp] = error;
1525 reqlog(LOG_WARNING, 0, &hio->hio_ggio,
1526 "Remote request failed (%s): ", strerror(error));
1530 ggio = &hio->hio_ggio;
1531 switch (ggio->gctl_cmd) {
1533 rw_rlock(&hio_remote_lock[ncomp]);
1534 if (!ISCONNECTED(res, ncomp)) {
1535 rw_unlock(&hio_remote_lock[ncomp]);
1539 if (hast_proto_recv_data(res, res->hr_remotein, nv,
1540 ggio->gctl_data, ggio->gctl_length) < 0) {
1541 hio->hio_errors[ncomp] = errno;
1542 pjdlog_errno(LOG_ERR,
1543 "Unable to receive reply data");
1544 rw_unlock(&hio_remote_lock[ncomp]);
1546 remote_close(res, ncomp);
1549 rw_unlock(&hio_remote_lock[ncomp]);
1556 PJDLOG_ASSERT(!"invalid condition");
1559 hio->hio_errors[ncomp] = 0;
1562 if (refcount_release(&hio->hio_countdown)) {
1563 if (ISSYNCREQ(hio)) {
1564 mtx_lock(&sync_lock);
1566 mtx_unlock(&sync_lock);
1567 cv_signal(&sync_cond);
1570 "remote_recv: (%p) Moving request to the done queue.",
1572 QUEUE_INSERT2(hio, done);
1581 * Thread sends answer to the kernel.
1584 ggate_send_thread(void *arg)
1586 struct hast_resource *res = arg;
1587 struct g_gate_ctl_io *ggio;
1589 unsigned int ii, ncomp, ncomps;
1591 ncomps = HAST_NCOMPONENTS;
1594 pjdlog_debug(2, "ggate_send: Taking request.");
1595 QUEUE_TAKE2(hio, done);
1596 pjdlog_debug(2, "ggate_send: (%p) Got request.", hio);
1597 ggio = &hio->hio_ggio;
1598 for (ii = 0; ii < ncomps; ii++) {
1599 if (hio->hio_errors[ii] == 0) {
1601 * One successful request is enough to declare
1604 ggio->gctl_error = 0;
1610 * None of the requests were successful.
1613 ggio->gctl_error = hio->hio_errors[0];
1615 if (ggio->gctl_error == 0 && ggio->gctl_cmd == BIO_WRITE) {
1616 mtx_lock(&res->hr_amp_lock);
1617 activemap_write_complete(res->hr_amp,
1618 ggio->gctl_offset, ggio->gctl_length);
1619 mtx_unlock(&res->hr_amp_lock);
1621 if (ggio->gctl_cmd == BIO_WRITE) {
1623 * Unlock range we locked.
1625 mtx_lock(&range_lock);
1626 rangelock_del(range_regular, ggio->gctl_offset,
1628 if (range_sync_wait)
1629 cv_signal(&range_sync_cond);
1630 mtx_unlock(&range_lock);
1632 * Bump local count if this is first write after
1633 * connection failure with remote node.
1636 rw_rlock(&hio_remote_lock[ncomp]);
1637 if (!ISCONNECTED(res, ncomp)) {
1638 mtx_lock(&metadata_lock);
1639 if (res->hr_primary_localcnt ==
1640 res->hr_secondary_remotecnt) {
1641 res->hr_primary_localcnt++;
1643 "Increasing localcnt to %ju.",
1644 (uintmax_t)res->hr_primary_localcnt);
1645 (void)metadata_write(res);
1647 mtx_unlock(&metadata_lock);
1649 rw_unlock(&hio_remote_lock[ncomp]);
1651 if (ioctl(res->hr_ggatefd, G_GATE_CMD_DONE, ggio) < 0)
1652 primary_exit(EX_OSERR, "G_GATE_CMD_DONE failed");
1654 "ggate_send: (%p) Moving request to the free queue.", hio);
1655 QUEUE_INSERT2(hio, free);
1662 * Thread synchronize local and remote components.
1665 sync_thread(void *arg __unused)
1667 struct hast_resource *res = arg;
1669 struct g_gate_ctl_io *ggio;
1670 unsigned int ii, ncomp, ncomps;
1671 off_t offset, length, synced;
1675 ncomps = HAST_NCOMPONENTS;
1681 mtx_lock(&sync_lock);
1682 if (offset >= 0 && !sync_inprogress) {
1683 pjdlog_info("Synchronization interrupted. "
1684 "%jd bytes synchronized so far.",
1686 event_send(res, EVENT_SYNCINTR);
1688 while (!sync_inprogress) {
1691 cv_wait(&sync_cond, &sync_lock);
1693 mtx_unlock(&sync_lock);
1695 * Obtain offset at which we should synchronize.
1696 * Rewind synchronization if needed.
1698 mtx_lock(&res->hr_amp_lock);
1700 activemap_sync_rewind(res->hr_amp);
1701 offset = activemap_sync_offset(res->hr_amp, &length, &syncext);
1702 if (syncext != -1) {
1704 * We synchronized entire syncext extent, we can mark
1707 if (activemap_extent_complete(res->hr_amp, syncext))
1708 (void)hast_activemap_flush(res);
1710 mtx_unlock(&res->hr_amp_lock);
1714 pjdlog_info("Nodes are in sync.");
1716 pjdlog_info("Synchronization started. %ju bytes to go.",
1717 (uintmax_t)(res->hr_extentsize *
1718 activemap_ndirty(res->hr_amp)));
1719 event_send(res, EVENT_SYNCSTART);
1724 pjdlog_debug(1, "Nothing to synchronize.");
1726 * Synchronization complete, make both localcnt and
1730 rw_rlock(&hio_remote_lock[ncomp]);
1731 if (ISCONNECTED(res, ncomp)) {
1733 pjdlog_info("Synchronization complete. "
1734 "%jd bytes synchronized.",
1736 event_send(res, EVENT_SYNCDONE);
1738 mtx_lock(&metadata_lock);
1739 res->hr_syncsrc = HAST_SYNCSRC_UNDEF;
1740 res->hr_primary_localcnt =
1741 res->hr_secondary_localcnt;
1742 res->hr_primary_remotecnt =
1743 res->hr_secondary_remotecnt;
1745 "Setting localcnt to %ju and remotecnt to %ju.",
1746 (uintmax_t)res->hr_primary_localcnt,
1747 (uintmax_t)res->hr_secondary_localcnt);
1748 (void)metadata_write(res);
1749 mtx_unlock(&metadata_lock);
1751 rw_unlock(&hio_remote_lock[ncomp]);
1754 pjdlog_debug(2, "sync: Taking free request.");
1755 QUEUE_TAKE2(hio, free);
1756 pjdlog_debug(2, "sync: (%p) Got free request.", hio);
1758 * Lock the range we are going to synchronize. We don't want
1759 * race where someone writes between our read and write.
1762 mtx_lock(&range_lock);
1763 if (rangelock_islocked(range_regular, offset, length)) {
1765 "sync: Range offset=%jd length=%jd locked.",
1766 (intmax_t)offset, (intmax_t)length);
1767 range_sync_wait = true;
1768 cv_wait(&range_sync_cond, &range_lock);
1769 range_sync_wait = false;
1770 mtx_unlock(&range_lock);
1773 if (rangelock_add(range_sync, offset, length) < 0) {
1774 mtx_unlock(&range_lock);
1776 "sync: Range offset=%jd length=%jd is already locked, waiting.",
1777 (intmax_t)offset, (intmax_t)length);
1781 mtx_unlock(&range_lock);
1785 * First read the data from synchronization source.
1788 ggio = &hio->hio_ggio;
1789 ggio->gctl_cmd = BIO_READ;
1790 ggio->gctl_offset = offset;
1791 ggio->gctl_length = length;
1792 ggio->gctl_error = 0;
1793 for (ii = 0; ii < ncomps; ii++)
1794 hio->hio_errors[ii] = EINVAL;
1795 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1797 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1799 mtx_lock(&metadata_lock);
1800 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1802 * This range is up-to-date on local component,
1803 * so handle request locally.
1805 /* Local component is 0 for now. */
1807 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1808 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1810 * This range is out-of-date on local component,
1811 * so send request to the remote node.
1813 /* Remote component is 1 for now. */
1816 mtx_unlock(&metadata_lock);
1817 refcount_init(&hio->hio_countdown, 1);
1818 QUEUE_INSERT1(hio, send, ncomp);
1821 * Let's wait for READ to finish.
1823 mtx_lock(&sync_lock);
1824 while (!ISSYNCREQDONE(hio))
1825 cv_wait(&sync_cond, &sync_lock);
1826 mtx_unlock(&sync_lock);
1828 if (hio->hio_errors[ncomp] != 0) {
1829 pjdlog_error("Unable to read synchronization data: %s.",
1830 strerror(hio->hio_errors[ncomp]));
1835 * We read the data from synchronization source, now write it
1836 * to synchronization target.
1839 ggio->gctl_cmd = BIO_WRITE;
1840 for (ii = 0; ii < ncomps; ii++)
1841 hio->hio_errors[ii] = EINVAL;
1842 reqlog(LOG_DEBUG, 2, ggio, "sync: (%p) Sending sync request: ",
1844 pjdlog_debug(2, "sync: (%p) Moving request to the send queue.",
1846 mtx_lock(&metadata_lock);
1847 if (res->hr_syncsrc == HAST_SYNCSRC_PRIMARY) {
1849 * This range is up-to-date on local component,
1850 * so we update remote component.
1852 /* Remote component is 1 for now. */
1854 } else /* if (res->hr_syncsrc == HAST_SYNCSRC_SECONDARY) */ {
1855 PJDLOG_ASSERT(res->hr_syncsrc == HAST_SYNCSRC_SECONDARY);
1857 * This range is out-of-date on local component,
1860 /* Local component is 0 for now. */
1863 mtx_unlock(&metadata_lock);
1865 pjdlog_debug(2, "sync: (%p) Moving request to the send queues.",
1867 refcount_init(&hio->hio_countdown, 1);
1868 QUEUE_INSERT1(hio, send, ncomp);
1871 * Let's wait for WRITE to finish.
1873 mtx_lock(&sync_lock);
1874 while (!ISSYNCREQDONE(hio))
1875 cv_wait(&sync_cond, &sync_lock);
1876 mtx_unlock(&sync_lock);
1878 if (hio->hio_errors[ncomp] != 0) {
1879 pjdlog_error("Unable to write synchronization data: %s.",
1880 strerror(hio->hio_errors[ncomp]));
1886 mtx_lock(&range_lock);
1887 rangelock_del(range_sync, offset, length);
1888 if (range_regular_wait)
1889 cv_signal(&range_regular_cond);
1890 mtx_unlock(&range_lock);
1891 pjdlog_debug(2, "sync: (%p) Moving request to the free queue.",
1893 QUEUE_INSERT2(hio, free);
1900 primary_config_reload(struct hast_resource *res, struct nv *nv)
1902 unsigned int ii, ncomps;
1906 pjdlog_info("Reloading configuration...");
1908 PJDLOG_ASSERT(res->hr_role == HAST_ROLE_PRIMARY);
1909 PJDLOG_ASSERT(gres == res);
1910 nv_assert(nv, "remoteaddr");
1911 nv_assert(nv, "replication");
1912 nv_assert(nv, "timeout");
1913 nv_assert(nv, "exec");
1915 ncomps = HAST_NCOMPONENTS;
1917 #define MODIFIED_REMOTEADDR 0x1
1918 #define MODIFIED_REPLICATION 0x2
1919 #define MODIFIED_TIMEOUT 0x4
1920 #define MODIFIED_EXEC 0x8
1923 vstr = nv_get_string(nv, "remoteaddr");
1924 if (strcmp(gres->hr_remoteaddr, vstr) != 0) {
1926 * Don't copy res->hr_remoteaddr to gres just yet.
1927 * We want remote_close() to log disconnect from the old
1928 * addresses, not from the new ones.
1930 modified |= MODIFIED_REMOTEADDR;
1932 vint = nv_get_int32(nv, "replication");
1933 if (gres->hr_replication != vint) {
1934 gres->hr_replication = vint;
1935 modified |= MODIFIED_REPLICATION;
1937 vint = nv_get_int32(nv, "timeout");
1938 if (gres->hr_timeout != vint) {
1939 gres->hr_timeout = vint;
1940 modified |= MODIFIED_TIMEOUT;
1942 vstr = nv_get_string(nv, "exec");
1943 if (strcmp(gres->hr_exec, vstr) != 0) {
1944 strlcpy(gres->hr_exec, vstr, sizeof(gres->hr_exec));
1945 modified |= MODIFIED_EXEC;
1949 * If only timeout was modified we only need to change it without
1952 if (modified == MODIFIED_TIMEOUT) {
1953 for (ii = 0; ii < ncomps; ii++) {
1956 rw_rlock(&hio_remote_lock[ii]);
1957 if (!ISCONNECTED(gres, ii)) {
1958 rw_unlock(&hio_remote_lock[ii]);
1961 rw_unlock(&hio_remote_lock[ii]);
1962 if (proto_timeout(gres->hr_remotein,
1963 gres->hr_timeout) < 0) {
1964 pjdlog_errno(LOG_WARNING,
1965 "Unable to set connection timeout");
1967 if (proto_timeout(gres->hr_remoteout,
1968 gres->hr_timeout) < 0) {
1969 pjdlog_errno(LOG_WARNING,
1970 "Unable to set connection timeout");
1973 } else if ((modified &
1974 (MODIFIED_REMOTEADDR | MODIFIED_REPLICATION)) != 0) {
1975 for (ii = 0; ii < ncomps; ii++) {
1978 remote_close(gres, ii);
1980 if (modified & MODIFIED_REMOTEADDR) {
1981 vstr = nv_get_string(nv, "remoteaddr");
1982 strlcpy(gres->hr_remoteaddr, vstr,
1983 sizeof(gres->hr_remoteaddr));
1986 #undef MODIFIED_REMOTEADDR
1987 #undef MODIFIED_REPLICATION
1988 #undef MODIFIED_TIMEOUT
1989 #undef MODIFIED_EXEC
1991 pjdlog_info("Configuration reloaded successfully.");
1995 guard_one(struct hast_resource *res, unsigned int ncomp)
1997 struct proto_conn *in, *out;
1999 if (!ISREMOTE(ncomp))
2002 rw_rlock(&hio_remote_lock[ncomp]);
2004 if (!real_remote(res)) {
2005 rw_unlock(&hio_remote_lock[ncomp]);
2009 if (ISCONNECTED(res, ncomp)) {
2010 PJDLOG_ASSERT(res->hr_remotein != NULL);
2011 PJDLOG_ASSERT(res->hr_remoteout != NULL);
2012 rw_unlock(&hio_remote_lock[ncomp]);
2013 pjdlog_debug(2, "remote_guard: Connection to %s is ok.",
2014 res->hr_remoteaddr);
2018 PJDLOG_ASSERT(res->hr_remotein == NULL);
2019 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2021 * Upgrade the lock. It doesn't have to be atomic as no other thread
2022 * can change connection status from disconnected to connected.
2024 rw_unlock(&hio_remote_lock[ncomp]);
2025 pjdlog_debug(2, "remote_guard: Reconnecting to %s.",
2026 res->hr_remoteaddr);
2028 if (init_remote(res, &in, &out)) {
2029 rw_wlock(&hio_remote_lock[ncomp]);
2030 PJDLOG_ASSERT(res->hr_remotein == NULL);
2031 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2032 PJDLOG_ASSERT(in != NULL && out != NULL);
2033 res->hr_remotein = in;
2034 res->hr_remoteout = out;
2035 rw_unlock(&hio_remote_lock[ncomp]);
2036 pjdlog_info("Successfully reconnected to %s.",
2037 res->hr_remoteaddr);
2040 /* Both connections should be NULL. */
2041 PJDLOG_ASSERT(res->hr_remotein == NULL);
2042 PJDLOG_ASSERT(res->hr_remoteout == NULL);
2043 PJDLOG_ASSERT(in == NULL && out == NULL);
2044 pjdlog_debug(2, "remote_guard: Reconnect to %s failed.",
2045 res->hr_remoteaddr);
2050 * Thread guards remote connections and reconnects when needed, handles
2054 guard_thread(void *arg)
2056 struct hast_resource *res = arg;
2057 unsigned int ii, ncomps;
2058 struct timespec timeout;
2059 time_t lastcheck, now;
2063 ncomps = HAST_NCOMPONENTS;
2064 lastcheck = time(NULL);
2066 PJDLOG_VERIFY(sigemptyset(&mask) == 0);
2067 PJDLOG_VERIFY(sigaddset(&mask, SIGINT) == 0);
2068 PJDLOG_VERIFY(sigaddset(&mask, SIGTERM) == 0);
2070 timeout.tv_sec = RETRY_SLEEP;
2071 timeout.tv_nsec = 0;
2078 sigexit_received = true;
2079 primary_exitx(EX_OK,
2080 "Termination signal received, exiting.");
2086 pjdlog_debug(2, "remote_guard: Checking connections.");
2088 if (lastcheck + RETRY_SLEEP <= now) {
2089 for (ii = 0; ii < ncomps; ii++)
2093 signo = sigtimedwait(&mask, NULL, &timeout);