]> CyberLeo.Net >> Repos - FreeBSD/stable/10.git/blob - sbin/hastd/secondary.c
MFC r368207,368607:
[FreeBSD/stable/10.git] / sbin / hastd / secondary.c
1 /*-
2  * Copyright (c) 2009-2010 The FreeBSD Foundation
3  * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org>
4  * All rights reserved.
5  *
6  * This software was developed by Pawel Jakub Dawidek under sponsorship from
7  * the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33
34 #include <sys/param.h>
35 #include <sys/time.h>
36 #include <sys/bio.h>
37 #include <sys/disk.h>
38 #include <sys/stat.h>
39
40 #include <err.h>
41 #include <errno.h>
42 #include <fcntl.h>
43 #include <libgeom.h>
44 #include <pthread.h>
45 #include <signal.h>
46 #include <stdint.h>
47 #include <stdio.h>
48 #include <string.h>
49 #include <sysexits.h>
50 #include <unistd.h>
51
52 #include <activemap.h>
53 #include <nv.h>
54 #include <pjdlog.h>
55
56 #include "control.h"
57 #include "event.h"
58 #include "hast.h"
59 #include "hast_proto.h"
60 #include "hastd.h"
61 #include "hooks.h"
62 #include "metadata.h"
63 #include "proto.h"
64 #include "subr.h"
65 #include "synch.h"
66
67 struct hio {
68         uint64_t         hio_seq;
69         int              hio_error;
70         void            *hio_data;
71         uint8_t          hio_cmd;
72         uint64_t         hio_offset;
73         uint64_t         hio_length;
74         bool             hio_memsync;
75         TAILQ_ENTRY(hio) hio_next;
76 };
77
78 static struct hast_resource *gres;
79
80 /*
81  * Free list holds unused structures. When free list is empty, we have to wait
82  * until some in-progress requests are freed.
83  */
84 static TAILQ_HEAD(, hio) hio_free_list;
85 static size_t hio_free_list_size;
86 static pthread_mutex_t hio_free_list_lock;
87 static pthread_cond_t hio_free_list_cond;
88 /*
89  * Disk thread (the one that does I/O requests) takes requests from this list.
90  */
91 static TAILQ_HEAD(, hio) hio_disk_list;
92 static size_t hio_disk_list_size;
93 static pthread_mutex_t hio_disk_list_lock;
94 static pthread_cond_t hio_disk_list_cond;
95 /*
96  * Thread that sends requests back to primary takes requests from this list.
97  */
98 static TAILQ_HEAD(, hio) hio_send_list;
99 static size_t hio_send_list_size;
100 static pthread_mutex_t hio_send_list_lock;
101 static pthread_cond_t hio_send_list_cond;
102
103 /*
104  * Maximum number of outstanding I/O requests.
105  */
106 #define HAST_HIO_MAX    256
107
108 static void *recv_thread(void *arg);
109 static void *disk_thread(void *arg);
110 static void *send_thread(void *arg);
111
112 #define QUEUE_INSERT(name, hio) do {                                    \
113         mtx_lock(&hio_##name##_list_lock);                              \
114         if (TAILQ_EMPTY(&hio_##name##_list))                            \
115                 cv_broadcast(&hio_##name##_list_cond);                  \
116         TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_next);         \
117         hio_##name##_list_size++;                                       \
118         mtx_unlock(&hio_##name##_list_lock);                            \
119 } while (0)
120 #define QUEUE_TAKE(name, hio)   do {                                    \
121         mtx_lock(&hio_##name##_list_lock);                              \
122         while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) {     \
123                 cv_wait(&hio_##name##_list_cond,                        \
124                     &hio_##name##_list_lock);                           \
125         }                                                               \
126         PJDLOG_ASSERT(hio_##name##_list_size != 0);                     \
127         hio_##name##_list_size--;                                       \
128         TAILQ_REMOVE(&hio_##name##_list, (hio), hio_next);              \
129         mtx_unlock(&hio_##name##_list_lock);                            \
130 } while (0)
131
132 static void
133 output_status_aux(struct nv *nvout)
134 {
135
136         nv_add_uint64(nvout, (uint64_t)hio_free_list_size, "idle_queue_size");
137         nv_add_uint64(nvout, (uint64_t)hio_disk_list_size, "local_queue_size");
138         nv_add_uint64(nvout, (uint64_t)hio_send_list_size, "send_queue_size");
139 }
140
141 static void
142 hio_clear(struct hio *hio)
143 {
144
145         hio->hio_seq = 0;
146         hio->hio_error = 0;
147         hio->hio_cmd = HIO_UNDEF;
148         hio->hio_offset = 0;
149         hio->hio_length = 0;
150         hio->hio_memsync = false;
151 }
152
153 static void
154 hio_copy(const struct hio *srchio, struct hio *dsthio)
155 {
156
157         /*
158          * We don't copy hio_error, hio_data and hio_next fields.
159          */
160
161         dsthio->hio_seq = srchio->hio_seq;
162         dsthio->hio_cmd = srchio->hio_cmd;
163         dsthio->hio_offset = srchio->hio_offset;
164         dsthio->hio_length = srchio->hio_length;
165         dsthio->hio_memsync = srchio->hio_memsync;
166 }
167
168 static void
169 init_environment(void)
170 {
171         struct hio *hio;
172         unsigned int ii;
173
174         /*
175          * Initialize lists, their locks and theirs condition variables.
176          */
177         TAILQ_INIT(&hio_free_list);
178         mtx_init(&hio_free_list_lock);
179         cv_init(&hio_free_list_cond);
180         TAILQ_INIT(&hio_disk_list);
181         mtx_init(&hio_disk_list_lock);
182         cv_init(&hio_disk_list_cond);
183         TAILQ_INIT(&hio_send_list);
184         mtx_init(&hio_send_list_lock);
185         cv_init(&hio_send_list_cond);
186
187         /*
188          * Allocate requests pool and initialize requests.
189          */
190         for (ii = 0; ii < HAST_HIO_MAX; ii++) {
191                 hio = malloc(sizeof(*hio));
192                 if (hio == NULL) {
193                         pjdlog_exitx(EX_TEMPFAIL,
194                             "Unable to allocate memory (%zu bytes) for hio request.",
195                             sizeof(*hio));
196                 }
197                 hio->hio_data = malloc(MAXPHYS);
198                 if (hio->hio_data == NULL) {
199                         pjdlog_exitx(EX_TEMPFAIL,
200                             "Unable to allocate memory (%zu bytes) for gctl_data.",
201                             (size_t)MAXPHYS);
202                 }
203                 hio_clear(hio);
204                 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_next);
205                 hio_free_list_size++;
206         }
207 }
208
209 static void
210 init_local(struct hast_resource *res)
211 {
212
213         if (metadata_read(res, true) == -1)
214                 exit(EX_NOINPUT);
215 }
216
217 static void
218 init_remote(struct hast_resource *res, struct nv *nvin)
219 {
220         uint64_t resuid;
221         struct nv *nvout;
222         unsigned char *map;
223         size_t mapsize;
224
225 #ifdef notyet
226         /* Setup direction. */
227         if (proto_send(res->hr_remoteout, NULL, 0) == -1)
228                 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
229 #endif
230
231         nvout = nv_alloc();
232         nv_add_int64(nvout, (int64_t)res->hr_datasize, "datasize");
233         nv_add_int32(nvout, (int32_t)res->hr_extentsize, "extentsize");
234         resuid = nv_get_uint64(nvin, "resuid");
235         res->hr_primary_localcnt = nv_get_uint64(nvin, "localcnt");
236         res->hr_primary_remotecnt = nv_get_uint64(nvin, "remotecnt");
237         nv_add_uint64(nvout, res->hr_secondary_localcnt, "localcnt");
238         nv_add_uint64(nvout, res->hr_secondary_remotecnt, "remotecnt");
239         mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize -
240             METADATA_SIZE, res->hr_extentsize, res->hr_local_sectorsize);
241         map = malloc(mapsize);
242         if (map == NULL) {
243                 pjdlog_exitx(EX_TEMPFAIL,
244                     "Unable to allocate memory (%zu bytes) for activemap.",
245                     mapsize);
246         }
247         /*
248          * When we work as primary and secondary is missing we will increase
249          * localcnt in our metadata. When secondary is connected and synced
250          * we make localcnt be equal to remotecnt, which means nodes are more
251          * or less in sync.
252          * Split-brain condition is when both nodes are not able to communicate
253          * and are both configured as primary nodes. In turn, they can both
254          * make incompatible changes to the data and we have to detect that.
255          * Under split-brain condition we will increase our localcnt on first
256          * write and remote node will increase its localcnt on first write.
257          * When we connect we can see that primary's localcnt is greater than
258          * our remotecnt (primary was modified while we weren't watching) and
259          * our localcnt is greater than primary's remotecnt (we were modified
260          * while primary wasn't watching).
261          * There are many possible combinations which are all gathered below.
262          * Don't pay too much attention to exact numbers, the more important
263          * is to compare them. We compare secondary's local with primary's
264          * remote and secondary's remote with primary's local.
265          * Note that every case where primary's localcnt is smaller than
266          * secondary's remotecnt and where secondary's localcnt is smaller than
267          * primary's remotecnt should be impossible in practise. We will perform
268          * full synchronization then. Those cases are marked with an asterisk.
269          * Regular synchronization means that only extents marked as dirty are
270          * synchronized (regular synchronization).
271          *
272          * SECONDARY METADATA PRIMARY METADATA
273          * local=3 remote=3   local=2 remote=2*  ?! Full sync from secondary.
274          * local=3 remote=3   local=2 remote=3*  ?! Full sync from primary.
275          * local=3 remote=3   local=2 remote=4*  ?! Full sync from primary.
276          * local=3 remote=3   local=3 remote=2   Primary is out-of-date,
277          *                                       regular sync from secondary.
278          * local=3 remote=3   local=3 remote=3   Regular sync just in case.
279          * local=3 remote=3   local=3 remote=4*  ?! Full sync from primary.
280          * local=3 remote=3   local=4 remote=2   Split-brain condition.
281          * local=3 remote=3   local=4 remote=3   Secondary out-of-date,
282          *                                       regular sync from primary.
283          * local=3 remote=3   local=4 remote=4*  ?! Full sync from primary.
284          */
285         if (res->hr_resuid == 0) {
286                 /*
287                  * Provider is used for the first time. If primary node done no
288                  * writes yet as well (we will find "virgin" argument) then
289                  * there is no need to synchronize anything. If primary node
290                  * done any writes already we have to synchronize everything.
291                  */
292                 PJDLOG_ASSERT(res->hr_secondary_localcnt == 0);
293                 res->hr_resuid = resuid;
294                 if (metadata_write(res) == -1)
295                         exit(EX_NOINPUT);
296                 if (nv_exists(nvin, "virgin")) {
297                         free(map);
298                         map = NULL;
299                         mapsize = 0;
300                 } else {
301                         memset(map, 0xff, mapsize);
302                 }
303                 nv_add_int8(nvout, 1, "virgin");
304                 nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc");
305         } else if (res->hr_resuid != resuid) {
306                 char errmsg[256];
307
308                 free(map);
309                 (void)snprintf(errmsg, sizeof(errmsg),
310                     "Resource unique ID mismatch (primary=%ju, secondary=%ju).",
311                     (uintmax_t)resuid, (uintmax_t)res->hr_resuid);
312                 pjdlog_error("%s", errmsg);
313                 nv_add_string(nvout, errmsg, "errmsg");
314                 if (hast_proto_send(res, res->hr_remotein, nvout,
315                     NULL, 0) == -1) {
316                         pjdlog_exit(EX_TEMPFAIL,
317                             "Unable to send response to %s",
318                             res->hr_remoteaddr);
319                 }
320                 nv_free(nvout);
321                 exit(EX_CONFIG);
322         } else if (
323             /* Is primary out-of-date? */
324             (res->hr_secondary_localcnt > res->hr_primary_remotecnt &&
325              res->hr_secondary_remotecnt == res->hr_primary_localcnt) ||
326             /* Are the nodes more or less in sync? */
327             (res->hr_secondary_localcnt == res->hr_primary_remotecnt &&
328              res->hr_secondary_remotecnt == res->hr_primary_localcnt) ||
329             /* Is secondary out-of-date? */
330             (res->hr_secondary_localcnt == res->hr_primary_remotecnt &&
331              res->hr_secondary_remotecnt < res->hr_primary_localcnt)) {
332                 /*
333                  * Nodes are more or less in sync or one of the nodes is
334                  * out-of-date.
335                  * It doesn't matter at this point which one, we just have to
336                  * send out local bitmap to the remote node.
337                  */
338                 if (pread(res->hr_localfd, map, mapsize, METADATA_SIZE) !=
339                     (ssize_t)mapsize) {
340                         pjdlog_exit(LOG_ERR, "Unable to read activemap");
341                 }
342                 if (res->hr_secondary_localcnt > res->hr_primary_remotecnt &&
343                      res->hr_secondary_remotecnt == res->hr_primary_localcnt) {
344                         /* Primary is out-of-date, sync from secondary. */
345                         nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc");
346                 } else {
347                         /*
348                          * Secondary is out-of-date or counts match.
349                          * Sync from primary.
350                          */
351                         nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc");
352                 }
353         } else if (res->hr_secondary_localcnt > res->hr_primary_remotecnt &&
354              res->hr_primary_localcnt > res->hr_secondary_remotecnt) {
355                 /*
356                  * Not good, we have split-brain condition.
357                  */
358                 free(map);
359                 pjdlog_error("Split-brain detected, exiting.");
360                 nv_add_string(nvout, "Split-brain condition!", "errmsg");
361                 if (hast_proto_send(res, res->hr_remotein, nvout,
362                     NULL, 0) == -1) {
363                         pjdlog_exit(EX_TEMPFAIL,
364                             "Unable to send response to %s",
365                             res->hr_remoteaddr);
366                 }
367                 nv_free(nvout);
368                 /* Exit on split-brain. */
369                 event_send(res, EVENT_SPLITBRAIN);
370                 exit(EX_CONFIG);
371         } else /* if (res->hr_secondary_localcnt < res->hr_primary_remotecnt ||
372             res->hr_primary_localcnt < res->hr_secondary_remotecnt) */ {
373                 /*
374                  * This should never happen in practise, but we will perform
375                  * full synchronization.
376                  */
377                 PJDLOG_ASSERT(res->hr_secondary_localcnt < res->hr_primary_remotecnt ||
378                     res->hr_primary_localcnt < res->hr_secondary_remotecnt);
379                 mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize -
380                     METADATA_SIZE, res->hr_extentsize,
381                     res->hr_local_sectorsize);
382                 memset(map, 0xff, mapsize);
383                 if (res->hr_secondary_localcnt > res->hr_primary_remotecnt) {
384                         /* In this one of five cases sync from secondary. */
385                         nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc");
386                 } else {
387                         /* For the rest four cases sync from primary. */
388                         nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc");
389                 }
390                 pjdlog_warning("This should never happen, asking for full synchronization (primary(local=%ju, remote=%ju), secondary(local=%ju, remote=%ju)).",
391                     (uintmax_t)res->hr_primary_localcnt,
392                     (uintmax_t)res->hr_primary_remotecnt,
393                     (uintmax_t)res->hr_secondary_localcnt,
394                     (uintmax_t)res->hr_secondary_remotecnt);
395         }
396         nv_add_uint32(nvout, (uint32_t)mapsize, "mapsize");
397         if (hast_proto_send(res, res->hr_remotein, nvout, map, mapsize) == -1) {
398                 pjdlog_exit(EX_TEMPFAIL, "Unable to send activemap to %s",
399                     res->hr_remoteaddr);
400         }
401         if (map != NULL)
402                 free(map);
403         nv_free(nvout);
404 #ifdef notyet
405         /* Setup direction. */
406         if (proto_recv(res->hr_remotein, NULL, 0) == -1)
407                 pjdlog_errno(LOG_WARNING, "Unable to set connection direction");
408 #endif
409 }
410
411 void
412 hastd_secondary(struct hast_resource *res, struct nv *nvin)
413 {
414         sigset_t mask;
415         pthread_t td;
416         pid_t pid;
417         int error, mode, debuglevel;
418
419         /*
420          * Create communication channel between parent and child.
421          */
422         if (proto_client(NULL, "socketpair://", &res->hr_ctrl) == -1) {
423                 KEEP_ERRNO((void)pidfile_remove(pfh));
424                 pjdlog_exit(EX_OSERR,
425                     "Unable to create control sockets between parent and child");
426         }
427         /*
428          * Create communication channel between child and parent.
429          */
430         if (proto_client(NULL, "socketpair://", &res->hr_event) == -1) {
431                 KEEP_ERRNO((void)pidfile_remove(pfh));
432                 pjdlog_exit(EX_OSERR,
433                     "Unable to create event sockets between child and parent");
434         }
435
436         pid = fork();
437         if (pid == -1) {
438                 KEEP_ERRNO((void)pidfile_remove(pfh));
439                 pjdlog_exit(EX_OSERR, "Unable to fork");
440         }
441
442         if (pid > 0) {
443                 /* This is parent. */
444                 proto_close(res->hr_remotein);
445                 res->hr_remotein = NULL;
446                 proto_close(res->hr_remoteout);
447                 res->hr_remoteout = NULL;
448                 /* Declare that we are receiver. */
449                 proto_recv(res->hr_event, NULL, 0);
450                 /* Declare that we are sender. */
451                 proto_send(res->hr_ctrl, NULL, 0);
452                 res->hr_workerpid = pid;
453                 return;
454         }
455
456         gres = res;
457         res->output_status_aux = output_status_aux;
458         mode = pjdlog_mode_get();
459         debuglevel = pjdlog_debug_get();
460
461         /* Declare that we are sender. */
462         proto_send(res->hr_event, NULL, 0);
463         /* Declare that we are receiver. */
464         proto_recv(res->hr_ctrl, NULL, 0);
465         descriptors_cleanup(res);
466
467         descriptors_assert(res, mode);
468
469         pjdlog_init(mode);
470         pjdlog_debug_set(debuglevel);
471         pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role));
472         setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role));
473
474         PJDLOG_VERIFY(sigemptyset(&mask) == 0);
475         PJDLOG_VERIFY(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
476
477         /* Error in setting timeout is not critical, but why should it fail? */
478         if (proto_timeout(res->hr_remotein, 2 * HAST_KEEPALIVE) == -1)
479                 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
480         if (proto_timeout(res->hr_remoteout, res->hr_timeout) == -1)
481                 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout");
482
483         init_local(res);
484         init_environment();
485
486         if (drop_privs(res) != 0)
487                 exit(EX_CONFIG);
488         pjdlog_info("Privileges successfully dropped.");
489
490         /*
491          * Create the control thread before sending any event to the parent,
492          * as we can deadlock when parent sends control request to worker,
493          * but worker has no control thread started yet, so parent waits.
494          * In the meantime worker sends an event to the parent, but parent
495          * is unable to handle the event, because it waits for control
496          * request response.
497          */
498         error = pthread_create(&td, NULL, ctrl_thread, res);
499         PJDLOG_ASSERT(error == 0);
500
501         init_remote(res, nvin);
502         event_send(res, EVENT_CONNECT);
503
504         error = pthread_create(&td, NULL, recv_thread, res);
505         PJDLOG_ASSERT(error == 0);
506         error = pthread_create(&td, NULL, disk_thread, res);
507         PJDLOG_ASSERT(error == 0);
508         (void)send_thread(res);
509 }
510
511 static void
512 reqlog(int loglevel, int debuglevel, int error, struct hio *hio,
513     const char *fmt, ...)
514 {
515         char msg[1024];
516         va_list ap;
517         int len;
518
519         va_start(ap, fmt);
520         len = vsnprintf(msg, sizeof(msg), fmt, ap);
521         va_end(ap);
522         if ((size_t)len < sizeof(msg)) {
523                 switch (hio->hio_cmd) {
524                 case HIO_READ:
525                         (void)snprintf(msg + len, sizeof(msg) - len,
526                             "READ(%ju, %ju).", (uintmax_t)hio->hio_offset,
527                             (uintmax_t)hio->hio_length);
528                         break;
529                 case HIO_DELETE:
530                         (void)snprintf(msg + len, sizeof(msg) - len,
531                             "DELETE(%ju, %ju).", (uintmax_t)hio->hio_offset,
532                             (uintmax_t)hio->hio_length);
533                         break;
534                 case HIO_FLUSH:
535                         (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH.");
536                         break;
537                 case HIO_WRITE:
538                         (void)snprintf(msg + len, sizeof(msg) - len,
539                             "WRITE(%ju, %ju).", (uintmax_t)hio->hio_offset,
540                             (uintmax_t)hio->hio_length);
541                         break;
542                 case HIO_KEEPALIVE:
543                         (void)snprintf(msg + len, sizeof(msg) - len, "KEEPALIVE.");
544                         break;
545                 default:
546                         (void)snprintf(msg + len, sizeof(msg) - len,
547                             "UNKNOWN(%u).", (unsigned int)hio->hio_cmd);
548                         break;
549                 }
550         }
551         pjdlog_common(loglevel, debuglevel, error, "%s", msg);
552 }
553
554 static int
555 requnpack(struct hast_resource *res, struct hio *hio, struct nv *nv)
556 {
557
558         hio->hio_cmd = nv_get_uint8(nv, "cmd");
559         if (hio->hio_cmd == 0) {
560                 pjdlog_error("Header contains no 'cmd' field.");
561                 hio->hio_error = EINVAL;
562                 goto end;
563         }
564         if (hio->hio_cmd != HIO_KEEPALIVE) {
565                 hio->hio_seq = nv_get_uint64(nv, "seq");
566                 if (hio->hio_seq == 0) {
567                         pjdlog_error("Header contains no 'seq' field.");
568                         hio->hio_error = EINVAL;
569                         goto end;
570                 }
571         }
572         switch (hio->hio_cmd) {
573         case HIO_FLUSH:
574         case HIO_KEEPALIVE:
575                 break;
576         case HIO_WRITE:
577                 hio->hio_memsync = nv_exists(nv, "memsync");
578                 /* FALLTHROUGH */
579         case HIO_READ:
580         case HIO_DELETE:
581                 hio->hio_offset = nv_get_uint64(nv, "offset");
582                 if (nv_error(nv) != 0) {
583                         pjdlog_error("Header is missing 'offset' field.");
584                         hio->hio_error = EINVAL;
585                         goto end;
586                 }
587                 hio->hio_length = nv_get_uint64(nv, "length");
588                 if (nv_error(nv) != 0) {
589                         pjdlog_error("Header is missing 'length' field.");
590                         hio->hio_error = EINVAL;
591                         goto end;
592                 }
593                 if (hio->hio_length == 0) {
594                         pjdlog_error("Data length is zero.");
595                         hio->hio_error = EINVAL;
596                         goto end;
597                 }
598                 if (hio->hio_cmd != HIO_DELETE && hio->hio_length > MAXPHYS) {
599                         pjdlog_error("Data length is too large (%ju > %ju).",
600                             (uintmax_t)hio->hio_length, (uintmax_t)MAXPHYS);
601                         hio->hio_error = EINVAL;
602                         goto end;
603                 }
604                 if ((hio->hio_offset % res->hr_local_sectorsize) != 0) {
605                         pjdlog_error("Offset %ju is not multiple of sector size.",
606                             (uintmax_t)hio->hio_offset);
607                         hio->hio_error = EINVAL;
608                         goto end;
609                 }
610                 if ((hio->hio_length % res->hr_local_sectorsize) != 0) {
611                         pjdlog_error("Length %ju is not multiple of sector size.",
612                             (uintmax_t)hio->hio_length);
613                         hio->hio_error = EINVAL;
614                         goto end;
615                 }
616                 if (hio->hio_offset + hio->hio_length >
617                     (uint64_t)res->hr_datasize) {
618                         pjdlog_error("Data offset is too large (%ju > %ju).",
619                             (uintmax_t)(hio->hio_offset + hio->hio_length),
620                             (uintmax_t)res->hr_datasize);
621                         hio->hio_error = EINVAL;
622                         goto end;
623                 }
624                 break;
625         default:
626                 pjdlog_error("Header contains invalid 'cmd' (%hhu).",
627                     hio->hio_cmd);
628                 hio->hio_error = EINVAL;
629                 goto end;
630         }
631         hio->hio_error = 0;
632 end:
633         return (hio->hio_error);
634 }
635
636 static __dead2 void
637 secondary_exit(int exitcode, const char *fmt, ...)
638 {
639         va_list ap;
640
641         PJDLOG_ASSERT(exitcode != EX_OK);
642         va_start(ap, fmt);
643         pjdlogv_errno(LOG_ERR, fmt, ap);
644         va_end(ap);
645         event_send(gres, EVENT_DISCONNECT);
646         exit(exitcode);
647 }
648
649 /*
650  * Thread receives requests from the primary node.
651  */
652 static void *
653 recv_thread(void *arg)
654 {
655         struct hast_resource *res = arg;
656         struct hio *hio, *mshio;
657         struct nv *nv;
658
659         for (;;) {
660                 pjdlog_debug(2, "recv: Taking free request.");
661                 QUEUE_TAKE(free, hio);
662                 pjdlog_debug(2, "recv: (%p) Got request.", hio);
663                 if (hast_proto_recv_hdr(res->hr_remotein, &nv) == -1) {
664                         secondary_exit(EX_TEMPFAIL,
665                             "Unable to receive request header");
666                 }
667                 if (requnpack(res, hio, nv) != 0) {
668                         nv_free(nv);
669                         pjdlog_debug(2,
670                             "recv: (%p) Moving request to the send queue.",
671                             hio);
672                         QUEUE_INSERT(send, hio);
673                         continue;
674                 }
675                 switch (hio->hio_cmd) {
676                 case HIO_READ:
677                         res->hr_stat_read++;
678                         break;
679                 case HIO_WRITE:
680                         res->hr_stat_write++;
681                         break;
682                 case HIO_DELETE:
683                         res->hr_stat_delete++;
684                         break;
685                 case HIO_FLUSH:
686                         res->hr_stat_flush++;
687                         break;
688                 case HIO_KEEPALIVE:
689                         break;
690                 default:
691                         PJDLOG_ABORT("Unexpected command (cmd=%hhu).",
692                             hio->hio_cmd);
693                 }
694                 reqlog(LOG_DEBUG, 2, -1, hio,
695                     "recv: (%p) Got request header: ", hio);
696                 if (hio->hio_cmd == HIO_KEEPALIVE) {
697                         nv_free(nv);
698                         pjdlog_debug(2,
699                             "recv: (%p) Moving request to the free queue.",
700                             hio);
701                         hio_clear(hio);
702                         QUEUE_INSERT(free, hio);
703                         continue;
704                 } else if (hio->hio_cmd == HIO_WRITE) {
705                         if (hast_proto_recv_data(res, res->hr_remotein, nv,
706                             hio->hio_data, MAXPHYS) == -1) {
707                                 secondary_exit(EX_TEMPFAIL,
708                                     "Unable to receive request data");
709                         }
710                         if (hio->hio_memsync) {
711                                 /*
712                                  * For memsync requests we expect two replies.
713                                  * Clone the hio so we can handle both of them.
714                                  */
715                                 pjdlog_debug(2, "recv: Taking free request.");
716                                 QUEUE_TAKE(free, mshio);
717                                 pjdlog_debug(2, "recv: (%p) Got request.",
718                                     mshio);
719                                 hio_copy(hio, mshio);
720                                 mshio->hio_error = 0;
721                                 /*
722                                  * We want to keep 'memsync' tag only on the
723                                  * request going onto send queue (mshio).
724                                  */
725                                 hio->hio_memsync = false;
726                                 pjdlog_debug(2,
727                                     "recv: (%p) Moving memsync request to the send queue.",
728                                     mshio);
729                                 QUEUE_INSERT(send, mshio);
730                         }
731                 }
732                 nv_free(nv);
733                 pjdlog_debug(2, "recv: (%p) Moving request to the disk queue.",
734                     hio);
735                 QUEUE_INSERT(disk, hio);
736         }
737         /* NOTREACHED */
738         return (NULL);
739 }
740
741 /*
742  * Thread reads from or writes to local component and also handles DELETE and
743  * FLUSH requests.
744  */
745 static void *
746 disk_thread(void *arg)
747 {
748         struct hast_resource *res = arg;
749         struct hio *hio;
750         ssize_t ret;
751         bool clear_activemap, logerror;
752
753         clear_activemap = true;
754
755         for (;;) {
756                 pjdlog_debug(2, "disk: Taking request.");
757                 QUEUE_TAKE(disk, hio);
758                 while (clear_activemap) {
759                         unsigned char *map;
760                         size_t mapsize;
761
762                         /*
763                          * When first request is received, it means that primary
764                          * already received our activemap, merged it and stored
765                          * locally. We can now safely clear our activemap.
766                          */
767                         mapsize =
768                             activemap_calc_ondisk_size(res->hr_local_mediasize -
769                             METADATA_SIZE, res->hr_extentsize,
770                             res->hr_local_sectorsize);
771                         map = calloc(1, mapsize);
772                         if (map == NULL) {
773                                 pjdlog_warning("Unable to allocate memory to clear local activemap.");
774                                 break;
775                         }
776                         if (pwrite(res->hr_localfd, map, mapsize,
777                             METADATA_SIZE) != (ssize_t)mapsize) {
778                                 pjdlog_errno(LOG_WARNING,
779                                     "Unable to store cleared activemap");
780                                 free(map);
781                                 res->hr_stat_activemap_write_error++;
782                                 break;
783                         }
784                         free(map);
785                         clear_activemap = false;
786                         pjdlog_debug(1, "Local activemap cleared.");
787                         break;
788                 }
789                 reqlog(LOG_DEBUG, 2, -1, hio, "disk: (%p) Got request: ", hio);
790                 logerror = true;
791                 /* Handle the actual request. */
792                 switch (hio->hio_cmd) {
793                 case HIO_READ:
794                         ret = pread(res->hr_localfd, hio->hio_data,
795                             hio->hio_length,
796                             hio->hio_offset + res->hr_localoff);
797                         if (ret == -1)
798                                 hio->hio_error = errno;
799                         else if (ret != (int64_t)hio->hio_length)
800                                 hio->hio_error = EIO;
801                         else
802                                 hio->hio_error = 0;
803                         break;
804                 case HIO_WRITE:
805                         ret = pwrite(res->hr_localfd, hio->hio_data,
806                             hio->hio_length,
807                             hio->hio_offset + res->hr_localoff);
808                         if (ret == -1)
809                                 hio->hio_error = errno;
810                         else if (ret != (int64_t)hio->hio_length)
811                                 hio->hio_error = EIO;
812                         else
813                                 hio->hio_error = 0;
814                         break;
815                 case HIO_DELETE:
816                         ret = g_delete(res->hr_localfd,
817                             hio->hio_offset + res->hr_localoff,
818                             hio->hio_length);
819                         if (ret == -1)
820                                 hio->hio_error = errno;
821                         else
822                                 hio->hio_error = 0;
823                         break;
824                 case HIO_FLUSH:
825                         if (!res->hr_localflush) {
826                                 ret = -1;
827                                 hio->hio_error = EOPNOTSUPP;
828                                 logerror = false;
829                                 break;
830                         }
831                         ret = g_flush(res->hr_localfd);
832                         if (ret == -1) {
833                                 if (errno == EOPNOTSUPP)
834                                         res->hr_localflush = false;
835                                 hio->hio_error = errno;
836                         } else {
837                                 hio->hio_error = 0;
838                         }
839                         break;
840                 default:
841                         PJDLOG_ABORT("Unexpected command (cmd=%hhu).",
842                             hio->hio_cmd);
843                 }
844                 if (logerror && hio->hio_error != 0) {
845                         reqlog(LOG_ERR, 0, hio->hio_error, hio,
846                             "Request failed: ");
847                 }
848                 pjdlog_debug(2, "disk: (%p) Moving request to the send queue.",
849                     hio);
850                 QUEUE_INSERT(send, hio);
851         }
852         /* NOTREACHED */
853         return (NULL);
854 }
855
856 /*
857  * Thread sends requests back to primary node.
858  */
859 static void *
860 send_thread(void *arg)
861 {
862         struct hast_resource *res = arg;
863         struct nv *nvout;
864         struct hio *hio;
865         void *data;
866         size_t length;
867
868         for (;;) {
869                 pjdlog_debug(2, "send: Taking request.");
870                 QUEUE_TAKE(send, hio);
871                 reqlog(LOG_DEBUG, 2, -1, hio, "send: (%p) Got request: ", hio);
872                 nvout = nv_alloc();
873                 /* Copy sequence number. */
874                 nv_add_uint64(nvout, hio->hio_seq, "seq");
875                 if (hio->hio_memsync) {
876                         PJDLOG_ASSERT(hio->hio_cmd == HIO_WRITE);
877                         nv_add_int8(nvout, 1, "received");
878                 }
879                 switch (hio->hio_cmd) {
880                 case HIO_READ:
881                         if (hio->hio_error == 0) {
882                                 data = hio->hio_data;
883                                 length = hio->hio_length;
884                                 break;
885                         }
886                         /*
887                          * We send no data in case of an error.
888                          */
889                         /* FALLTHROUGH */
890                 case HIO_DELETE:
891                 case HIO_FLUSH:
892                 case HIO_WRITE:
893                         data = NULL;
894                         length = 0;
895                         break;
896                 default:
897                         PJDLOG_ABORT("Unexpected command (cmd=%hhu).",
898                             hio->hio_cmd);
899                 }
900                 if (hio->hio_error != 0) {
901                         switch (hio->hio_cmd) {
902                         case HIO_READ:
903                                 res->hr_stat_read_error++;
904                                 break;
905                         case HIO_WRITE:
906                                 res->hr_stat_write_error++;
907                                 break;
908                         case HIO_DELETE:
909                                 res->hr_stat_delete_error++;
910                                 break;
911                         case HIO_FLUSH:
912                                 res->hr_stat_flush_error++;
913                                 break;
914                         }
915                         nv_add_int16(nvout, hio->hio_error, "error");
916                 }
917                 if (hast_proto_send(res, res->hr_remoteout, nvout, data,
918                     length) == -1) {
919                         secondary_exit(EX_TEMPFAIL, "Unable to send reply");
920                 }
921                 nv_free(nvout);
922                 pjdlog_debug(2, "send: (%p) Moving request to the free queue.",
923                     hio);
924                 hio_clear(hio);
925                 QUEUE_INSERT(free, hio);
926         }
927         /* NOTREACHED */
928         return (NULL);
929 }