secondary.c revision 219830
1127128Snectar/*- 2127128Snectar * Copyright (c) 2009-2010 The FreeBSD Foundation 3127128Snectar * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org> 4127128Snectar * All rights reserved. 5127128Snectar * 6127128Snectar * This software was developed by Pawel Jakub Dawidek under sponsorship from 7127128Snectar * the FreeBSD Foundation. 8127128Snectar * 9127128Snectar * Redistribution and use in source and binary forms, with or without 10127128Snectar * modification, are permitted provided that the following conditions 11127128Snectar * are met: 12127128Snectar * 1. Redistributions of source code must retain the above copyright 13127128Snectar * notice, this list of conditions and the following disclaimer. 14127128Snectar * 2. Redistributions in binary form must reproduce the above copyright 15127128Snectar * notice, this list of conditions and the following disclaimer in the 16127128Snectar * documentation and/or other materials provided with the distribution. 17127128Snectar * 18296341Sdelphij * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 19296341Sdelphij * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20296341Sdelphij * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21296341Sdelphij * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 22127128Snectar * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23127128Snectar * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24127128Snectar * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25127128Snectar * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26127128Snectar * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27127128Snectar * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28296341Sdelphij * SUCH DAMAGE. 29296341Sdelphij */ 30296341Sdelphij 31296341Sdelphij#include <sys/cdefs.h> 32296341Sdelphij__FBSDID("$FreeBSD: head/sbin/hastd/secondary.c 219830 2011-03-21 14:50:12Z pjd $"); 33296341Sdelphij 34296341Sdelphij#include <sys/param.h> 35296341Sdelphij#include <sys/time.h> 36127128Snectar#include <sys/bio.h> 37127128Snectar#include <sys/disk.h> 38127128Snectar#include <sys/stat.h> 39127128Snectar 40127128Snectar#include <err.h> 41127128Snectar#include <errno.h> 42127128Snectar#include <fcntl.h> 43127128Snectar#include <libgeom.h> 44127128Snectar#include <pthread.h> 45127128Snectar#include <signal.h> 46127128Snectar#include <stdint.h> 47127128Snectar#include <stdio.h> 48127128Snectar#include <string.h> 49127128Snectar#include <sysexits.h> 50127128Snectar#include <unistd.h> 51127128Snectar 52127128Snectar#include <activemap.h> 53127128Snectar#include <nv.h> 54127128Snectar#include <pjdlog.h> 55127128Snectar 56127128Snectar#include "control.h" 57127128Snectar#include "event.h" 58127128Snectar#include "hast.h" 59127128Snectar#include "hast_proto.h" 60127128Snectar#include "hastd.h" 61127128Snectar#include "hooks.h" 62127128Snectar#include "metadata.h" 63127128Snectar#include "proto.h" 64127128Snectar#include "subr.h" 65127128Snectar#include "synch.h" 66127128Snectar 67127128Snectarstruct hio { 68127128Snectar uint64_t hio_seq; 69267258Sjkim int hio_error; 70127128Snectar struct nv *hio_nv; 71127128Snectar void *hio_data; 72127128Snectar uint8_t hio_cmd; 73127128Snectar uint64_t hio_offset; 74127128Snectar uint64_t hio_length; 75127128Snectar TAILQ_ENTRY(hio) hio_next; 76}; 77 78static 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 */ 84static TAILQ_HEAD(, hio) hio_free_list; 85static pthread_mutex_t hio_free_list_lock; 86static pthread_cond_t hio_free_list_cond; 87/* 88 * Disk thread (the one that do I/O requests) takes requests from this list. 89 */ 90static TAILQ_HEAD(, hio) hio_disk_list; 91static pthread_mutex_t hio_disk_list_lock; 92static pthread_cond_t hio_disk_list_cond; 93/* 94 * There is one recv list for every component, although local components don't 95 * use recv lists as local requests are done synchronously. 96 */ 97static TAILQ_HEAD(, hio) hio_send_list; 98static pthread_mutex_t hio_send_list_lock; 99static pthread_cond_t hio_send_list_cond; 100 101/* 102 * Maximum number of outstanding I/O requests. 103 */ 104#define HAST_HIO_MAX 256 105 106static void *recv_thread(void *arg); 107static void *disk_thread(void *arg); 108static void *send_thread(void *arg); 109 110#define QUEUE_INSERT(name, hio) do { \ 111 bool _wakeup; \ 112 \ 113 mtx_lock(&hio_##name##_list_lock); \ 114 _wakeup = TAILQ_EMPTY(&hio_##name##_list); \ 115 TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_next); \ 116 mtx_unlock(&hio_##name##_list_lock); \ 117 if (_wakeup) \ 118 cv_signal(&hio_##name##_list_cond); \ 119} while (0) 120#define QUEUE_TAKE(name, hio) do { \ 121 mtx_lock(&hio_##name##_list_lock); \ 122 while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 123 cv_wait(&hio_##name##_list_cond, \ 124 &hio_##name##_list_lock); \ 125 } \ 126 TAILQ_REMOVE(&hio_##name##_list, (hio), hio_next); \ 127 mtx_unlock(&hio_##name##_list_lock); \ 128} while (0) 129 130static void 131init_environment(void) 132{ 133 struct hio *hio; 134 unsigned int ii; 135 136 /* 137 * Initialize lists, their locks and theirs condition variables. 138 */ 139 TAILQ_INIT(&hio_free_list); 140 mtx_init(&hio_free_list_lock); 141 cv_init(&hio_free_list_cond); 142 TAILQ_INIT(&hio_disk_list); 143 mtx_init(&hio_disk_list_lock); 144 cv_init(&hio_disk_list_cond); 145 TAILQ_INIT(&hio_send_list); 146 mtx_init(&hio_send_list_lock); 147 cv_init(&hio_send_list_cond); 148 149 /* 150 * Allocate requests pool and initialize requests. 151 */ 152 for (ii = 0; ii < HAST_HIO_MAX; ii++) { 153 hio = malloc(sizeof(*hio)); 154 if (hio == NULL) { 155 pjdlog_exitx(EX_TEMPFAIL, 156 "Unable to allocate memory (%zu bytes) for hio request.", 157 sizeof(*hio)); 158 } 159 hio->hio_error = 0; 160 hio->hio_data = malloc(MAXPHYS); 161 if (hio->hio_data == NULL) { 162 pjdlog_exitx(EX_TEMPFAIL, 163 "Unable to allocate memory (%zu bytes) for gctl_data.", 164 (size_t)MAXPHYS); 165 } 166 TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_next); 167 } 168} 169 170static void 171init_local(struct hast_resource *res) 172{ 173 174 if (metadata_read(res, true) < 0) 175 exit(EX_NOINPUT); 176} 177 178static void 179init_remote(struct hast_resource *res, struct nv *nvin) 180{ 181 uint64_t resuid; 182 struct nv *nvout; 183 unsigned char *map; 184 size_t mapsize; 185 186 map = NULL; 187 mapsize = 0; 188 nvout = nv_alloc(); 189 nv_add_int64(nvout, (int64_t)res->hr_datasize, "datasize"); 190 nv_add_int32(nvout, (int32_t)res->hr_extentsize, "extentsize"); 191 resuid = nv_get_uint64(nvin, "resuid"); 192 res->hr_primary_localcnt = nv_get_uint64(nvin, "localcnt"); 193 res->hr_primary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 194 nv_add_uint64(nvout, res->hr_secondary_localcnt, "localcnt"); 195 nv_add_uint64(nvout, res->hr_secondary_remotecnt, "remotecnt"); 196 mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize - 197 METADATA_SIZE, res->hr_extentsize, res->hr_local_sectorsize); 198 map = malloc(mapsize); 199 if (map == NULL) { 200 pjdlog_exitx(EX_TEMPFAIL, 201 "Unable to allocate memory (%zu bytes) for activemap.", 202 mapsize); 203 } 204 nv_add_uint32(nvout, (uint32_t)mapsize, "mapsize"); 205 /* 206 * When we work as primary and secondary is missing we will increase 207 * localcnt in our metadata. When secondary is connected and synced 208 * we make localcnt be equal to remotecnt, which means nodes are more 209 * or less in sync. 210 * Split-brain condition is when both nodes are not able to communicate 211 * and are both configured as primary nodes. In turn, they can both 212 * make incompatible changes to the data and we have to detect that. 213 * Under split-brain condition we will increase our localcnt on first 214 * write and remote node will increase its localcnt on first write. 215 * When we connect we can see that primary's localcnt is greater than 216 * our remotecnt (primary was modified while we weren't watching) and 217 * our localcnt is greater than primary's remotecnt (we were modified 218 * while primary wasn't watching). 219 * There are many possible combinations which are all gathered below. 220 * Don't pay too much attention to exact numbers, the more important 221 * is to compare them. We compare secondary's local with primary's 222 * remote and secondary's remote with primary's local. 223 * Note that every case where primary's localcnt is smaller than 224 * secondary's remotecnt and where secondary's localcnt is smaller than 225 * primary's remotecnt should be impossible in practise. We will perform 226 * full synchronization then. Those cases are marked with an asterisk. 227 * Regular synchronization means that only extents marked as dirty are 228 * synchronized (regular synchronization). 229 * 230 * SECONDARY METADATA PRIMARY METADATA 231 * local=3 remote=3 local=2 remote=2* ?! Full sync from secondary. 232 * local=3 remote=3 local=2 remote=3* ?! Full sync from primary. 233 * local=3 remote=3 local=2 remote=4* ?! Full sync from primary. 234 * local=3 remote=3 local=3 remote=2 Primary is out-of-date, 235 * regular sync from secondary. 236 * local=3 remote=3 local=3 remote=3 Regular sync just in case. 237 * local=3 remote=3 local=3 remote=4* ?! Full sync from primary. 238 * local=3 remote=3 local=4 remote=2 Split-brain condition. 239 * local=3 remote=3 local=4 remote=3 Secondary out-of-date, 240 * regular sync from primary. 241 * local=3 remote=3 local=4 remote=4* ?! Full sync from primary. 242 */ 243 if (res->hr_resuid == 0) { 244 /* 245 * Provider is used for the first time. If primary node done no 246 * writes yet as well (we will find "virgin" argument) then 247 * there is no need to synchronize anything. If primary node 248 * done any writes already we have to synchronize everything. 249 */ 250 PJDLOG_ASSERT(res->hr_secondary_localcnt == 0); 251 res->hr_resuid = resuid; 252 if (metadata_write(res) < 0) 253 exit(EX_NOINPUT); 254 if (nv_exists(nvin, "virgin")) { 255 free(map); 256 map = NULL; 257 mapsize = 0; 258 } else { 259 memset(map, 0xff, mapsize); 260 } 261 nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 262 } else if (res->hr_resuid != resuid) { 263 char errmsg[256]; 264 265 (void)snprintf(errmsg, sizeof(errmsg), 266 "Resource unique ID mismatch (primary=%ju, secondary=%ju).", 267 (uintmax_t)resuid, (uintmax_t)res->hr_resuid); 268 pjdlog_error("%s", errmsg); 269 nv_add_string(nvout, errmsg, "errmsg"); 270 if (hast_proto_send(res, res->hr_remotein, nvout, NULL, 0) < 0) { 271 pjdlog_exit(EX_TEMPFAIL, "Unable to send response to %s", 272 res->hr_remoteaddr); 273 } 274 exit(EX_CONFIG); 275 } else if ( 276 /* Is primary is out-of-date? */ 277 (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 278 res->hr_secondary_remotecnt == res->hr_primary_localcnt) || 279 /* Node are more or less in sync? */ 280 (res->hr_secondary_localcnt == res->hr_primary_remotecnt && 281 res->hr_secondary_remotecnt == res->hr_primary_localcnt) || 282 /* Is secondary is out-of-date? */ 283 (res->hr_secondary_localcnt == res->hr_primary_remotecnt && 284 res->hr_secondary_remotecnt < res->hr_primary_localcnt)) { 285 /* 286 * Nodes are more or less in sync or one of the nodes is 287 * out-of-date. 288 * It doesn't matter at this point which one, we just have to 289 * send out local bitmap to the remote node. 290 */ 291 if (pread(res->hr_localfd, map, mapsize, METADATA_SIZE) != 292 (ssize_t)mapsize) { 293 pjdlog_exit(LOG_ERR, "Unable to read activemap"); 294 } 295 if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 296 res->hr_secondary_remotecnt == res->hr_primary_localcnt) { 297 /* Primary is out-of-date, sync from secondary. */ 298 nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc"); 299 } else { 300 /* 301 * Secondary is out-of-date or counts match. 302 * Sync from primary. 303 */ 304 nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 305 } 306 } else if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 307 res->hr_primary_localcnt > res->hr_secondary_remotecnt) { 308 /* 309 * Not good, we have split-brain condition. 310 */ 311 pjdlog_error("Split-brain detected, exiting."); 312 nv_add_string(nvout, "Split-brain condition!", "errmsg"); 313 free(map); 314 map = NULL; 315 mapsize = 0; 316 } else /* if (res->hr_secondary_localcnt < res->hr_primary_remotecnt || 317 res->hr_primary_localcnt < res->hr_secondary_remotecnt) */ { 318 /* 319 * This should never happen in practise, but we will perform 320 * full synchronization. 321 */ 322 PJDLOG_ASSERT(res->hr_secondary_localcnt < res->hr_primary_remotecnt || 323 res->hr_primary_localcnt < res->hr_secondary_remotecnt); 324 mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize - 325 METADATA_SIZE, res->hr_extentsize, 326 res->hr_local_sectorsize); 327 memset(map, 0xff, mapsize); 328 if (res->hr_secondary_localcnt > res->hr_primary_remotecnt) { 329 /* In this one of five cases sync from secondary. */ 330 nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc"); 331 } else { 332 /* For the rest four cases sync from primary. */ 333 nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 334 } 335 pjdlog_warning("This should never happen, asking for full synchronization (primary(local=%ju, remote=%ju), secondary(local=%ju, remote=%ju)).", 336 (uintmax_t)res->hr_primary_localcnt, 337 (uintmax_t)res->hr_primary_remotecnt, 338 (uintmax_t)res->hr_secondary_localcnt, 339 (uintmax_t)res->hr_secondary_remotecnt); 340 } 341 if (hast_proto_send(res, res->hr_remotein, nvout, map, mapsize) < 0) { 342 pjdlog_exit(EX_TEMPFAIL, "Unable to send activemap to %s", 343 res->hr_remoteaddr); 344 } 345 if (map != NULL) 346 free(map); 347 nv_free(nvout); 348 if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 349 res->hr_primary_localcnt > res->hr_secondary_remotecnt) { 350 /* Exit on split-brain. */ 351 event_send(res, EVENT_SPLITBRAIN); 352 exit(EX_CONFIG); 353 } 354} 355 356void 357hastd_secondary(struct hast_resource *res, struct nv *nvin) 358{ 359 sigset_t mask; 360 pthread_t td; 361 pid_t pid; 362 int error, mode, debuglevel; 363 364 /* 365 * Create communication channel between parent and child. 366 */ 367 if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) { 368 KEEP_ERRNO((void)pidfile_remove(pfh)); 369 pjdlog_exit(EX_OSERR, 370 "Unable to create control sockets between parent and child"); 371 } 372 /* 373 * Create communication channel between child and parent. 374 */ 375 if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) { 376 KEEP_ERRNO((void)pidfile_remove(pfh)); 377 pjdlog_exit(EX_OSERR, 378 "Unable to create event sockets between child and parent"); 379 } 380 /* 381 * Create communication channel for sending connection requests from 382 * parent to child. 383 */ 384 if (proto_client(NULL, "socketpair://", &res->hr_conn) < 0) { 385 /* TODO: There's no need for this to be fatal error. */ 386 KEEP_ERRNO((void)pidfile_remove(pfh)); 387 pjdlog_exit(EX_OSERR, 388 "Unable to create connection sockets between parent and child"); 389 } 390 391 pid = fork(); 392 if (pid < 0) { 393 KEEP_ERRNO((void)pidfile_remove(pfh)); 394 pjdlog_exit(EX_OSERR, "Unable to fork"); 395 } 396 397 if (pid > 0) { 398 /* This is parent. */ 399 proto_close(res->hr_remotein); 400 res->hr_remotein = NULL; 401 proto_close(res->hr_remoteout); 402 res->hr_remoteout = NULL; 403 /* Declare that we are receiver. */ 404 proto_recv(res->hr_event, NULL, 0); 405 /* Declare that we are sender. */ 406 proto_send(res->hr_ctrl, NULL, 0); 407 proto_send(res->hr_conn, NULL, 0); 408 res->hr_workerpid = pid; 409 return; 410 } 411 412 gres = res; 413 mode = pjdlog_mode_get(); 414 debuglevel = pjdlog_debug_get(); 415 416 /* Declare that we are sender. */ 417 proto_send(res->hr_event, NULL, 0); 418 /* Declare that we are receiver. */ 419 proto_recv(res->hr_ctrl, NULL, 0); 420 proto_recv(res->hr_conn, NULL, 0); 421 descriptors_cleanup(res); 422 423 descriptors_assert(res, mode); 424 425 pjdlog_init(mode); 426 pjdlog_debug_set(debuglevel); 427 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 428 setproctitle("%s (secondary)", res->hr_name); 429 430 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 431 PJDLOG_VERIFY(sigprocmask(SIG_SETMASK, &mask, NULL) == 0); 432 433 /* Error in setting timeout is not critical, but why should it fail? */ 434 if (proto_timeout(res->hr_remotein, 2 * HAST_KEEPALIVE) < 0) 435 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 436 if (proto_timeout(res->hr_remoteout, res->hr_timeout) < 0) 437 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 438 439 init_local(res); 440 init_environment(); 441 442 if (drop_privs() != 0) 443 exit(EX_CONFIG); 444 pjdlog_info("Privileges successfully dropped."); 445 446 /* 447 * Create the control thread before sending any event to the parent, 448 * as we can deadlock when parent sends control request to worker, 449 * but worker has no control thread started yet, so parent waits. 450 * In the meantime worker sends an event to the parent, but parent 451 * is unable to handle the event, because it waits for control 452 * request response. 453 */ 454 error = pthread_create(&td, NULL, ctrl_thread, res); 455 PJDLOG_ASSERT(error == 0); 456 457 init_remote(res, nvin); 458 event_send(res, EVENT_CONNECT); 459 460 error = pthread_create(&td, NULL, recv_thread, res); 461 PJDLOG_ASSERT(error == 0); 462 error = pthread_create(&td, NULL, disk_thread, res); 463 PJDLOG_ASSERT(error == 0); 464 (void)send_thread(res); 465} 466 467static void 468reqlog(int loglevel, int debuglevel, int error, struct hio *hio, const char *fmt, ...) 469{ 470 char msg[1024]; 471 va_list ap; 472 int len; 473 474 va_start(ap, fmt); 475 len = vsnprintf(msg, sizeof(msg), fmt, ap); 476 va_end(ap); 477 if ((size_t)len < sizeof(msg)) { 478 switch (hio->hio_cmd) { 479 case HIO_READ: 480 (void)snprintf(msg + len, sizeof(msg) - len, 481 "READ(%ju, %ju).", (uintmax_t)hio->hio_offset, 482 (uintmax_t)hio->hio_length); 483 break; 484 case HIO_DELETE: 485 (void)snprintf(msg + len, sizeof(msg) - len, 486 "DELETE(%ju, %ju).", (uintmax_t)hio->hio_offset, 487 (uintmax_t)hio->hio_length); 488 break; 489 case HIO_FLUSH: 490 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 491 break; 492 case HIO_WRITE: 493 (void)snprintf(msg + len, sizeof(msg) - len, 494 "WRITE(%ju, %ju).", (uintmax_t)hio->hio_offset, 495 (uintmax_t)hio->hio_length); 496 break; 497 case HIO_KEEPALIVE: 498 (void)snprintf(msg + len, sizeof(msg) - len, "KEEPALIVE."); 499 break; 500 default: 501 (void)snprintf(msg + len, sizeof(msg) - len, 502 "UNKNOWN(%u).", (unsigned int)hio->hio_cmd); 503 break; 504 } 505 } 506 pjdlog_common(loglevel, debuglevel, error, "%s", msg); 507} 508 509static int 510requnpack(struct hast_resource *res, struct hio *hio) 511{ 512 513 hio->hio_cmd = nv_get_uint8(hio->hio_nv, "cmd"); 514 if (hio->hio_cmd == 0) { 515 pjdlog_error("Header contains no 'cmd' field."); 516 hio->hio_error = EINVAL; 517 goto end; 518 } 519 switch (hio->hio_cmd) { 520 case HIO_KEEPALIVE: 521 break; 522 case HIO_READ: 523 case HIO_WRITE: 524 case HIO_DELETE: 525 hio->hio_offset = nv_get_uint64(hio->hio_nv, "offset"); 526 if (nv_error(hio->hio_nv) != 0) { 527 pjdlog_error("Header is missing 'offset' field."); 528 hio->hio_error = EINVAL; 529 goto end; 530 } 531 hio->hio_length = nv_get_uint64(hio->hio_nv, "length"); 532 if (nv_error(hio->hio_nv) != 0) { 533 pjdlog_error("Header is missing 'length' field."); 534 hio->hio_error = EINVAL; 535 goto end; 536 } 537 if (hio->hio_length == 0) { 538 pjdlog_error("Data length is zero."); 539 hio->hio_error = EINVAL; 540 goto end; 541 } 542 if (hio->hio_length > MAXPHYS) { 543 pjdlog_error("Data length is too large (%ju > %ju).", 544 (uintmax_t)hio->hio_length, (uintmax_t)MAXPHYS); 545 hio->hio_error = EINVAL; 546 goto end; 547 } 548 if ((hio->hio_offset % res->hr_local_sectorsize) != 0) { 549 pjdlog_error("Offset %ju is not multiple of sector size.", 550 (uintmax_t)hio->hio_offset); 551 hio->hio_error = EINVAL; 552 goto end; 553 } 554 if ((hio->hio_length % res->hr_local_sectorsize) != 0) { 555 pjdlog_error("Length %ju is not multiple of sector size.", 556 (uintmax_t)hio->hio_length); 557 hio->hio_error = EINVAL; 558 goto end; 559 } 560 if (hio->hio_offset + hio->hio_length > 561 (uint64_t)res->hr_datasize) { 562 pjdlog_error("Data offset is too large (%ju > %ju).", 563 (uintmax_t)(hio->hio_offset + hio->hio_length), 564 (uintmax_t)res->hr_datasize); 565 hio->hio_error = EINVAL; 566 goto end; 567 } 568 break; 569 default: 570 pjdlog_error("Header contains invalid 'cmd' (%hhu).", 571 hio->hio_cmd); 572 hio->hio_error = EINVAL; 573 goto end; 574 } 575 hio->hio_error = 0; 576end: 577 return (hio->hio_error); 578} 579 580static __dead2 void 581secondary_exit(int exitcode, const char *fmt, ...) 582{ 583 va_list ap; 584 585 PJDLOG_ASSERT(exitcode != EX_OK); 586 va_start(ap, fmt); 587 pjdlogv_errno(LOG_ERR, fmt, ap); 588 va_end(ap); 589 event_send(gres, EVENT_DISCONNECT); 590 exit(exitcode); 591} 592 593/* 594 * Thread receives requests from the primary node. 595 */ 596static void * 597recv_thread(void *arg) 598{ 599 struct hast_resource *res = arg; 600 struct hio *hio; 601 602 for (;;) { 603 pjdlog_debug(2, "recv: Taking free request."); 604 QUEUE_TAKE(free, hio); 605 pjdlog_debug(2, "recv: (%p) Got request.", hio); 606 if (hast_proto_recv_hdr(res->hr_remotein, &hio->hio_nv) < 0) { 607 secondary_exit(EX_TEMPFAIL, 608 "Unable to receive request header"); 609 } 610 if (requnpack(res, hio) != 0) { 611 pjdlog_debug(2, 612 "recv: (%p) Moving request to the send queue.", 613 hio); 614 QUEUE_INSERT(send, hio); 615 continue; 616 } 617 reqlog(LOG_DEBUG, 2, -1, hio, 618 "recv: (%p) Got request header: ", hio); 619 if (hio->hio_cmd == HIO_KEEPALIVE) { 620 pjdlog_debug(2, 621 "recv: (%p) Moving request to the free queue.", 622 hio); 623 nv_free(hio->hio_nv); 624 QUEUE_INSERT(free, hio); 625 continue; 626 } else if (hio->hio_cmd == HIO_WRITE) { 627 if (hast_proto_recv_data(res, res->hr_remotein, 628 hio->hio_nv, hio->hio_data, MAXPHYS) < 0) { 629 secondary_exit(EX_TEMPFAIL, 630 "Unable to receive request data"); 631 } 632 } 633 pjdlog_debug(2, "recv: (%p) Moving request to the disk queue.", 634 hio); 635 QUEUE_INSERT(disk, hio); 636 } 637 /* NOTREACHED */ 638 return (NULL); 639} 640 641/* 642 * Thread reads from or writes to local component and also handles DELETE and 643 * FLUSH requests. 644 */ 645static void * 646disk_thread(void *arg) 647{ 648 struct hast_resource *res = arg; 649 struct hio *hio; 650 ssize_t ret; 651 bool clear_activemap; 652 653 clear_activemap = true; 654 655 for (;;) { 656 pjdlog_debug(2, "disk: Taking request."); 657 QUEUE_TAKE(disk, hio); 658 while (clear_activemap) { 659 unsigned char *map; 660 size_t mapsize; 661 662 /* 663 * When first request is received, it means that primary 664 * already received our activemap, merged it and stored 665 * locally. We can now safely clear our activemap. 666 */ 667 mapsize = 668 activemap_calc_ondisk_size(res->hr_local_mediasize - 669 METADATA_SIZE, res->hr_extentsize, 670 res->hr_local_sectorsize); 671 map = calloc(1, mapsize); 672 if (map == NULL) { 673 pjdlog_warning("Unable to allocate memory to clear local activemap."); 674 break; 675 } 676 if (pwrite(res->hr_localfd, map, mapsize, 677 METADATA_SIZE) != (ssize_t)mapsize) { 678 pjdlog_errno(LOG_WARNING, 679 "Unable to store cleared activemap"); 680 free(map); 681 break; 682 } 683 free(map); 684 clear_activemap = false; 685 pjdlog_debug(1, "Local activemap cleared."); 686 } 687 reqlog(LOG_DEBUG, 2, -1, hio, "disk: (%p) Got request: ", hio); 688 /* Handle the actual request. */ 689 switch (hio->hio_cmd) { 690 case HIO_READ: 691 ret = pread(res->hr_localfd, hio->hio_data, 692 hio->hio_length, 693 hio->hio_offset + res->hr_localoff); 694 if (ret < 0) 695 hio->hio_error = errno; 696 else if (ret != (int64_t)hio->hio_length) 697 hio->hio_error = EIO; 698 else 699 hio->hio_error = 0; 700 break; 701 case HIO_WRITE: 702 ret = pwrite(res->hr_localfd, hio->hio_data, 703 hio->hio_length, 704 hio->hio_offset + res->hr_localoff); 705 if (ret < 0) 706 hio->hio_error = errno; 707 else if (ret != (int64_t)hio->hio_length) 708 hio->hio_error = EIO; 709 else 710 hio->hio_error = 0; 711 break; 712 case HIO_DELETE: 713 ret = g_delete(res->hr_localfd, 714 hio->hio_offset + res->hr_localoff, 715 hio->hio_length); 716 if (ret < 0) 717 hio->hio_error = errno; 718 else 719 hio->hio_error = 0; 720 break; 721 case HIO_FLUSH: 722 ret = g_flush(res->hr_localfd); 723 if (ret < 0) 724 hio->hio_error = errno; 725 else 726 hio->hio_error = 0; 727 break; 728 } 729 if (hio->hio_error != 0) { 730 reqlog(LOG_ERR, 0, hio->hio_error, hio, 731 "Request failed: "); 732 } 733 pjdlog_debug(2, "disk: (%p) Moving request to the send queue.", 734 hio); 735 QUEUE_INSERT(send, hio); 736 } 737 /* NOTREACHED */ 738 return (NULL); 739} 740 741/* 742 * Thread sends requests back to primary node. 743 */ 744static void * 745send_thread(void *arg) 746{ 747 struct hast_resource *res = arg; 748 struct nv *nvout; 749 struct hio *hio; 750 void *data; 751 size_t length; 752 753 for (;;) { 754 pjdlog_debug(2, "send: Taking request."); 755 QUEUE_TAKE(send, hio); 756 reqlog(LOG_DEBUG, 2, -1, hio, "send: (%p) Got request: ", hio); 757 nvout = nv_alloc(); 758 /* Copy sequence number. */ 759 nv_add_uint64(nvout, nv_get_uint64(hio->hio_nv, "seq"), "seq"); 760 switch (hio->hio_cmd) { 761 case HIO_READ: 762 if (hio->hio_error == 0) { 763 data = hio->hio_data; 764 length = hio->hio_length; 765 break; 766 } 767 /* 768 * We send no data in case of an error. 769 */ 770 /* FALLTHROUGH */ 771 case HIO_DELETE: 772 case HIO_FLUSH: 773 case HIO_WRITE: 774 data = NULL; 775 length = 0; 776 break; 777 default: 778 abort(); 779 break; 780 } 781 if (hio->hio_error != 0) 782 nv_add_int16(nvout, hio->hio_error, "error"); 783 if (hast_proto_send(res, res->hr_remoteout, nvout, data, 784 length) < 0) { 785 secondary_exit(EX_TEMPFAIL, "Unable to send reply."); 786 } 787 nv_free(nvout); 788 pjdlog_debug(2, "send: (%p) Moving request to the free queue.", 789 hio); 790 nv_free(hio->hio_nv); 791 hio->hio_error = 0; 792 QUEUE_INSERT(free, hio); 793 } 794 /* NOTREACHED */ 795 return (NULL); 796} 797