secondary.c revision 219482
1117397Skan/*- 2117397Skan * Copyright (c) 2009-2010 The FreeBSD Foundation 3169691Skan * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org> 4169691Skan * All rights reserved. 5117397Skan * 6117397Skan * This software was developed by Pawel Jakub Dawidek under sponsorship from 7117397Skan * the FreeBSD Foundation. 8117397Skan * 9117397Skan * Redistribution and use in source and binary forms, with or without 10117397Skan * modification, are permitted provided that the following conditions 11117397Skan * are met: 12117397Skan * 1. Redistributions of source code must retain the above copyright 13117397Skan * notice, this list of conditions and the following disclaimer. 14117397Skan * 2. Redistributions in binary form must reproduce the above copyright 15117397Skan * notice, this list of conditions and the following disclaimer in the 16117397Skan * documentation and/or other materials provided with the distribution. 17117397Skan * 18117397Skan * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 19169691Skan * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20117397Skan * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21117397Skan * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 22117397Skan * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23117397Skan * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24117397Skan * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25117397Skan * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26117397Skan * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27117397Skan * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28117397Skan * SUCH DAMAGE. 29117397Skan */ 30117397Skan 31117397Skan#include <sys/cdefs.h> 32117397Skan__FBSDID("$FreeBSD: head/sbin/hastd/secondary.c 219482 2011-03-11 12:12:35Z trociny $"); 33117397Skan 34117397Skan#include <sys/param.h> 35117397Skan#include <sys/time.h> 36117397Skan#include <sys/bio.h> 37117397Skan#include <sys/disk.h> 38117397Skan#include <sys/stat.h> 39117397Skan 40117397Skan#include <err.h> 41117397Skan#include <errno.h> 42117397Skan#include <fcntl.h> 43117397Skan#include <libgeom.h> 44117397Skan#include <pthread.h> 45117397Skan#include <signal.h> 46117397Skan#include <stdint.h> 47117397Skan#include <stdio.h> 48117397Skan#include <string.h> 49117397Skan#include <sysexits.h> 50117397Skan#include <unistd.h> 51117397Skan 52117397Skan#include <activemap.h> 53117397Skan#include <nv.h> 54117397Skan#include <pjdlog.h> 55117397Skan 56117397Skan#include "control.h" 57117397Skan#include "event.h" 58117397Skan#include "hast.h" 59117397Skan#include "hast_proto.h" 60117397Skan#include "hastd.h" 61117397Skan#include "hooks.h" 62132720Skan#include "metadata.h" 63132720Skan#include "proto.h" 64117397Skan#include "subr.h" 65169691Skan#include "synch.h" 66169691Skan 67117397Skanstruct hio { 68117397Skan uint64_t hio_seq; 69169691Skan int hio_error; 70132720Skan struct nv *hio_nv; 71117397Skan void *hio_data; 72117397Skan uint8_t hio_cmd; 73132720Skan uint64_t hio_offset; 74132720Skan uint64_t hio_length; 75169691Skan TAILQ_ENTRY(hio) hio_next; 76169691Skan}; 77169691Skan 78169691Skanstatic struct hast_resource *gres; 79169691Skan 80169691Skan/* 81117397Skan * Free list holds unused structures. When free list is empty, we have to wait 82132720Skan * until some in-progress requests are freed. 83117397Skan */ 84169691Skanstatic TAILQ_HEAD(, hio) hio_free_list; 85169691Skanstatic pthread_mutex_t hio_free_list_lock; 86117397Skanstatic pthread_cond_t hio_free_list_cond; 87117397Skan/* 88117397Skan * Disk thread (the one that do I/O requests) takes requests from this list. 89132720Skan */ 90169691Skanstatic TAILQ_HEAD(, hio) hio_disk_list; 91117397Skanstatic pthread_mutex_t hio_disk_list_lock; 92132720Skanstatic pthread_cond_t hio_disk_list_cond; 93117397Skan/* 94169691Skan * There is one recv list for every component, although local components don't 95169691Skan * use recv lists as local requests are done synchronously. 96117397Skan */ 97117397Skanstatic TAILQ_HEAD(, hio) hio_send_list; 98169691Skanstatic pthread_mutex_t hio_send_list_lock; 99132720Skanstatic pthread_cond_t hio_send_list_cond; 100132720Skan 101117397Skan/* 102132720Skan * Maximum number of outstanding I/O requests. 103117397Skan */ 104117397Skan#define HAST_HIO_MAX 256 105169691Skan 106169691Skanstatic void *recv_thread(void *arg); 107117397Skanstatic void *disk_thread(void *arg); 108117397Skanstatic void *send_thread(void *arg); 109117397Skan 110169691Skan#define QUEUE_INSERT(name, hio) do { \ 111117397Skan bool _wakeup; \ 112117397Skan \ 113117397Skan mtx_lock(&hio_##name##_list_lock); \ 114117397Skan _wakeup = TAILQ_EMPTY(&hio_##name##_list); \ 115117397Skan TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_next); \ 116117397Skan mtx_unlock(&hio_##name##_list_lock); \ 117132720Skan if (_wakeup) \ 118117397Skan cv_signal(&hio_##name##_list_cond); \ 119169691Skan} while (0) 120169691Skan#define QUEUE_TAKE(name, hio) do { \ 121117397Skan mtx_lock(&hio_##name##_list_lock); \ 122117397Skan while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 123117397Skan cv_wait(&hio_##name##_list_cond, \ 124132720Skan &hio_##name##_list_lock); \ 125132720Skan } \ 126132720Skan TAILQ_REMOVE(&hio_##name##_list, (hio), hio_next); \ 127132720Skan mtx_unlock(&hio_##name##_list_lock); \ 128132720Skan} while (0) 129169691Skan 130169691Skanstatic void 131169691Skaninit_environment(void) 132132720Skan{ 133132720Skan struct hio *hio; 134132720Skan unsigned int ii; 135132720Skan 136132720Skan /* 137117397Skan * Initialize lists, their locks and theirs condition variables. 138117397Skan */ 139132720Skan TAILQ_INIT(&hio_free_list); 140117397Skan mtx_init(&hio_free_list_lock); 141117397Skan cv_init(&hio_free_list_cond); 142169691Skan TAILQ_INIT(&hio_disk_list); 143117397Skan mtx_init(&hio_disk_list_lock); 144117397Skan cv_init(&hio_disk_list_cond); 145117397Skan TAILQ_INIT(&hio_send_list); 146169691Skan mtx_init(&hio_send_list_lock); 147117397Skan cv_init(&hio_send_list_cond); 148117397Skan 149117397Skan /* 150117397Skan * Allocate requests pool and initialize requests. 151117397Skan */ 152117397Skan for (ii = 0; ii < HAST_HIO_MAX; ii++) { 153132720Skan hio = malloc(sizeof(*hio)); 154117397Skan if (hio == NULL) { 155132720Skan pjdlog_exitx(EX_TEMPFAIL, 156117397Skan "Unable to allocate memory (%zu bytes) for hio request.", 157169691Skan sizeof(*hio)); 158132720Skan } 159132720Skan hio->hio_error = 0; 160117397Skan hio->hio_data = malloc(MAXPHYS); 161117397Skan if (hio->hio_data == NULL) { 162117397Skan pjdlog_exitx(EX_TEMPFAIL, 163132720Skan "Unable to allocate memory (%zu bytes) for gctl_data.", 164132720Skan (size_t)MAXPHYS); 165117397Skan } 166117397Skan TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_next); 167117397Skan } 168117397Skan} 169117397Skan 170117397Skanstatic void 171132720Skaninit_local(struct hast_resource *res) 172117397Skan{ 173117397Skan 174169691Skan if (metadata_read(res, true) < 0) 175117397Skan exit(EX_NOINPUT); 176117397Skan} 177117397Skan 178117397Skanstatic void 179117397Skaninit_remote(struct hast_resource *res, struct nv *nvin) 180169691Skan{ 181169691Skan uint64_t resuid; 182169691Skan struct nv *nvout; 183169691Skan unsigned char *map; 184169691Skan size_t mapsize; 185169691Skan 186169691Skan map = NULL; 187117397Skan mapsize = 0; 188132720Skan nvout = nv_alloc(); 189117397Skan nv_add_int64(nvout, (int64_t)res->hr_datasize, "datasize"); 190117397Skan nv_add_int32(nvout, (int32_t)res->hr_extentsize, "extentsize"); 191169691Skan resuid = nv_get_uint64(nvin, "resuid"); 192117397Skan res->hr_primary_localcnt = nv_get_uint64(nvin, "localcnt"); 193117397Skan res->hr_primary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 194117397Skan nv_add_uint64(nvout, res->hr_secondary_localcnt, "localcnt"); 195117397Skan nv_add_uint64(nvout, res->hr_secondary_remotecnt, "remotecnt"); 196132720Skan mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize - 197132720Skan METADATA_SIZE, res->hr_extentsize, res->hr_local_sectorsize); 198117397Skan map = malloc(mapsize); 199117397Skan if (map == NULL) { 200169691Skan pjdlog_exitx(EX_TEMPFAIL, 201169691Skan "Unable to allocate memory (%zu bytes) for activemap.", 202169691Skan mapsize); 203169691Skan } 204169691Skan nv_add_uint32(nvout, (uint32_t)mapsize, "mapsize"); 205169691Skan /* 206169691Skan * When we work as primary and secondary is missing we will increase 207117397Skan * localcnt in our metadata. When secondary is connected and synced 208132720Skan * we make localcnt be equal to remotecnt, which means nodes are more 209117397Skan * or less in sync. 210117397Skan * Split-brain condition is when both nodes are not able to communicate 211169691Skan * and are both configured as primary nodes. In turn, they can both 212117397Skan * make incompatible changes to the data and we have to detect that. 213117397Skan * Under split-brain condition we will increase our localcnt on first 214132720Skan * write and remote node will increase its localcnt on first write. 215132720Skan * When we connect we can see that primary's localcnt is greater than 216132720Skan * our remotecnt (primary was modified while we weren't watching) and 217132720Skan * our localcnt is greater than primary's remotecnt (we were modified 218169691Skan * while primary wasn't watching). 219169691Skan * There are many possible combinations which are all gathered below. 220132720Skan * Don't pay too much attention to exact numbers, the more important 221132720Skan * is to compare them. We compare secondary's local with primary's 222132720Skan * remote and secondary's remote with primary's local. 223132720Skan * Note that every case where primary's localcnt is smaller than 224132720Skan * secondary's remotecnt and where secondary's localcnt is smaller than 225132720Skan * primary's remotecnt should be impossible in practise. We will perform 226132720Skan * full synchronization then. Those cases are marked with an asterisk. 227132720Skan * Regular synchronization means that only extents marked as dirty are 228132720Skan * synchronized (regular synchronization). 229132720Skan * 230132720Skan * SECONDARY METADATA PRIMARY METADATA 231132720Skan * local=3 remote=3 local=2 remote=2* ?! Full sync from secondary. 232132720Skan * local=3 remote=3 local=2 remote=3* ?! Full sync from primary. 233132720Skan * local=3 remote=3 local=2 remote=4* ?! Full sync from primary. 234132720Skan * local=3 remote=3 local=3 remote=2 Primary is out-of-date, 235132720Skan * regular sync from secondary. 236117397Skan * local=3 remote=3 local=3 remote=3 Regular sync just in case. 237132720Skan * local=3 remote=3 local=3 remote=4* ?! Full sync from primary. 238117397Skan * local=3 remote=3 local=4 remote=2 Split-brain condition. 239169691Skan * local=3 remote=3 local=4 remote=3 Secondary out-of-date, 240169691Skan * regular sync from primary. 241169691Skan * local=3 remote=3 local=4 remote=4* ?! Full sync from primary. 242169691Skan */ 243169691Skan if (res->hr_resuid == 0) { 244169691Skan /* 245169691Skan * Provider is used for the first time. If primary node done no 246169691Skan * writes yet as well (we will find "virgin" argument) then 247169691Skan * there is no need to synchronize anything. If primary node 248169691Skan * done any writes already we have to synchronize everything. 249169691Skan */ 250169691Skan PJDLOG_ASSERT(res->hr_secondary_localcnt == 0); 251169691Skan res->hr_resuid = resuid; 252169691Skan if (metadata_write(res) < 0) 253169691Skan exit(EX_NOINPUT); 254169691Skan if (nv_exists(nvin, "virgin")) { 255169691Skan free(map); 256169691Skan map = NULL; 257169691Skan mapsize = 0; 258169691Skan } else { 259169691Skan memset(map, 0xff, mapsize); 260169691Skan } 261169691Skan nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 262169691Skan } else if ( 263169691Skan /* Is primary is out-of-date? */ 264169691Skan (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 265169691Skan res->hr_secondary_remotecnt == res->hr_primary_localcnt) || 266169691Skan /* Node are more or less in sync? */ 267169691Skan (res->hr_secondary_localcnt == res->hr_primary_remotecnt && 268169691Skan res->hr_secondary_remotecnt == res->hr_primary_localcnt) || 269117397Skan /* Is secondary is out-of-date? */ 270169691Skan (res->hr_secondary_localcnt == res->hr_primary_remotecnt && 271117397Skan res->hr_secondary_remotecnt < res->hr_primary_localcnt)) { 272117397Skan /* 273117397Skan * Nodes are more or less in sync or one of the nodes is 274132720Skan * out-of-date. 275132720Skan * It doesn't matter at this point which one, we just have to 276117397Skan * send out local bitmap to the remote node. 277117397Skan */ 278132720Skan if (pread(res->hr_localfd, map, mapsize, METADATA_SIZE) != 279132720Skan (ssize_t)mapsize) { 280132720Skan pjdlog_exit(LOG_ERR, "Unable to read activemap"); 281132720Skan } 282132720Skan if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 283132720Skan res->hr_secondary_remotecnt == res->hr_primary_localcnt) { 284132720Skan /* Primary is out-of-date, sync from secondary. */ 285132720Skan nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc"); 286132720Skan } else { 287169691Skan /* 288132720Skan * Secondary is out-of-date or counts match. 289132720Skan * Sync from primary. 290132720Skan */ 291132720Skan nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 292132720Skan } 293132720Skan } else if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 294132720Skan res->hr_primary_localcnt > res->hr_secondary_remotecnt) { 295132720Skan /* 296132720Skan * Not good, we have split-brain condition. 297132720Skan */ 298132720Skan pjdlog_error("Split-brain detected, exiting."); 299169691Skan nv_add_string(nvout, "Split-brain condition!", "errmsg"); 300169691Skan free(map); 301169691Skan map = NULL; 302117397Skan mapsize = 0; 303117397Skan } else /* if (res->hr_secondary_localcnt < res->hr_primary_remotecnt || 304132720Skan res->hr_primary_localcnt < res->hr_secondary_remotecnt) */ { 305117397Skan /* 306117397Skan * This should never happen in practise, but we will perform 307117397Skan * full synchronization. 308169691Skan */ 309117397Skan PJDLOG_ASSERT(res->hr_secondary_localcnt < res->hr_primary_remotecnt || 310117397Skan res->hr_primary_localcnt < res->hr_secondary_remotecnt); 311117397Skan mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize - 312117397Skan METADATA_SIZE, res->hr_extentsize, 313117397Skan res->hr_local_sectorsize); 314169691Skan memset(map, 0xff, mapsize); 315169691Skan if (res->hr_secondary_localcnt > res->hr_primary_remotecnt) { 316169691Skan /* In this one of five cases sync from secondary. */ 317169691Skan nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc"); 318169691Skan } else { 319169691Skan /* For the rest four cases sync from primary. */ 320169691Skan nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 321117397Skan } 322132720Skan pjdlog_warning("This should never happen, asking for full synchronization (primary(local=%ju, remote=%ju), secondary(local=%ju, remote=%ju)).", 323117397Skan (uintmax_t)res->hr_primary_localcnt, 324117397Skan (uintmax_t)res->hr_primary_remotecnt, 325117397Skan (uintmax_t)res->hr_secondary_localcnt, 326169691Skan (uintmax_t)res->hr_secondary_remotecnt); 327117397Skan } 328117397Skan if (hast_proto_send(res, res->hr_remotein, nvout, map, mapsize) < 0) { 329117397Skan pjdlog_exit(EX_TEMPFAIL, "Unable to send activemap to %s", 330117397Skan res->hr_remoteaddr); 331169691Skan } 332169691Skan if (map != NULL) 333117397Skan free(map); 334117397Skan nv_free(nvout); 335132720Skan if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 336169691Skan res->hr_primary_localcnt > res->hr_secondary_remotecnt) { 337169691Skan /* Exit on split-brain. */ 338169691Skan event_send(res, EVENT_SPLITBRAIN); 339169691Skan exit(EX_CONFIG); 340169691Skan } 341132720Skan} 342117397Skan 343132720Skanvoid 344117397Skanhastd_secondary(struct hast_resource *res, struct nv *nvin) 345117397Skan{ 346132720Skan sigset_t mask; 347169691Skan pthread_t td; 348132720Skan pid_t pid; 349117397Skan int error, mode, debuglevel; 350132720Skan 351132720Skan /* 352132720Skan * Create communication channel between parent and child. 353132720Skan */ 354132720Skan if (proto_client("socketpair://", &res->hr_ctrl) < 0) { 355132720Skan KEEP_ERRNO((void)pidfile_remove(pfh)); 356132720Skan pjdlog_exit(EX_OSERR, 357132720Skan "Unable to create control sockets between parent and child"); 358132720Skan } 359132720Skan /* 360132720Skan * Create communication channel between child and parent. 361132720Skan */ 362132720Skan if (proto_client("socketpair://", &res->hr_event) < 0) { 363132720Skan KEEP_ERRNO((void)pidfile_remove(pfh)); 364169691Skan pjdlog_exit(EX_OSERR, 365132720Skan "Unable to create event sockets between child and parent"); 366132720Skan } 367132720Skan /* 368132720Skan * Create communication channel for sending connection requests from 369132720Skan * parent to child. 370132720Skan */ 371132720Skan if (proto_client("socketpair://", &res->hr_conn) < 0) { 372132720Skan /* TODO: There's no need for this to be fatal error. */ 373132720Skan KEEP_ERRNO((void)pidfile_remove(pfh)); 374132720Skan pjdlog_exit(EX_OSERR, 375132720Skan "Unable to create connection sockets between parent and child"); 376169691Skan } 377169691Skan 378169691Skan pid = fork(); 379132720Skan if (pid < 0) { 380117397Skan KEEP_ERRNO((void)pidfile_remove(pfh)); 381117397Skan pjdlog_exit(EX_OSERR, "Unable to fork"); 382169691Skan } 383169691Skan 384132720Skan if (pid > 0) { 385132720Skan /* This is parent. */ 386 proto_close(res->hr_remotein); 387 res->hr_remotein = NULL; 388 proto_close(res->hr_remoteout); 389 res->hr_remoteout = NULL; 390 /* Declare that we are receiver. */ 391 proto_recv(res->hr_event, NULL, 0); 392 /* Declare that we are sender. */ 393 proto_send(res->hr_ctrl, NULL, 0); 394 proto_send(res->hr_conn, NULL, 0); 395 res->hr_workerpid = pid; 396 return; 397 } 398 399 gres = res; 400 mode = pjdlog_mode_get(); 401 debuglevel = pjdlog_debug_get(); 402 403 /* Declare that we are sender. */ 404 proto_send(res->hr_event, NULL, 0); 405 /* Declare that we are receiver. */ 406 proto_recv(res->hr_ctrl, NULL, 0); 407 proto_recv(res->hr_conn, NULL, 0); 408 descriptors_cleanup(res); 409 410 descriptors_assert(res, mode); 411 412 pjdlog_init(mode); 413 pjdlog_debug_set(debuglevel); 414 pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 415 setproctitle("%s (secondary)", res->hr_name); 416 417 PJDLOG_VERIFY(sigemptyset(&mask) == 0); 418 PJDLOG_VERIFY(sigprocmask(SIG_SETMASK, &mask, NULL) == 0); 419 420 /* Error in setting timeout is not critical, but why should it fail? */ 421 if (proto_timeout(res->hr_remotein, 0) < 0) 422 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 423 if (proto_timeout(res->hr_remoteout, res->hr_timeout) < 0) 424 pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 425 426 init_local(res); 427 init_environment(); 428 429 if (drop_privs() != 0) 430 exit(EX_CONFIG); 431 pjdlog_info("Privileges successfully dropped."); 432 433 /* 434 * Create the control thread before sending any event to the parent, 435 * as we can deadlock when parent sends control request to worker, 436 * but worker has no control thread started yet, so parent waits. 437 * In the meantime worker sends an event to the parent, but parent 438 * is unable to handle the event, because it waits for control 439 * request response. 440 */ 441 error = pthread_create(&td, NULL, ctrl_thread, res); 442 PJDLOG_ASSERT(error == 0); 443 444 init_remote(res, nvin); 445 event_send(res, EVENT_CONNECT); 446 447 error = pthread_create(&td, NULL, recv_thread, res); 448 PJDLOG_ASSERT(error == 0); 449 error = pthread_create(&td, NULL, disk_thread, res); 450 PJDLOG_ASSERT(error == 0); 451 (void)send_thread(res); 452} 453 454static void 455reqlog(int loglevel, int debuglevel, int error, struct hio *hio, const char *fmt, ...) 456{ 457 char msg[1024]; 458 va_list ap; 459 int len; 460 461 va_start(ap, fmt); 462 len = vsnprintf(msg, sizeof(msg), fmt, ap); 463 va_end(ap); 464 if ((size_t)len < sizeof(msg)) { 465 switch (hio->hio_cmd) { 466 case HIO_READ: 467 (void)snprintf(msg + len, sizeof(msg) - len, 468 "READ(%ju, %ju).", (uintmax_t)hio->hio_offset, 469 (uintmax_t)hio->hio_length); 470 break; 471 case HIO_DELETE: 472 (void)snprintf(msg + len, sizeof(msg) - len, 473 "DELETE(%ju, %ju).", (uintmax_t)hio->hio_offset, 474 (uintmax_t)hio->hio_length); 475 break; 476 case HIO_FLUSH: 477 (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 478 break; 479 case HIO_WRITE: 480 (void)snprintf(msg + len, sizeof(msg) - len, 481 "WRITE(%ju, %ju).", (uintmax_t)hio->hio_offset, 482 (uintmax_t)hio->hio_length); 483 break; 484 case HIO_KEEPALIVE: 485 (void)snprintf(msg + len, sizeof(msg) - len, "KEEPALIVE."); 486 break; 487 default: 488 (void)snprintf(msg + len, sizeof(msg) - len, 489 "UNKNOWN(%u).", (unsigned int)hio->hio_cmd); 490 break; 491 } 492 } 493 pjdlog_common(loglevel, debuglevel, error, "%s", msg); 494} 495 496static int 497requnpack(struct hast_resource *res, struct hio *hio) 498{ 499 500 hio->hio_cmd = nv_get_uint8(hio->hio_nv, "cmd"); 501 if (hio->hio_cmd == 0) { 502 pjdlog_error("Header contains no 'cmd' field."); 503 hio->hio_error = EINVAL; 504 goto end; 505 } 506 switch (hio->hio_cmd) { 507 case HIO_KEEPALIVE: 508 break; 509 case HIO_READ: 510 case HIO_WRITE: 511 case HIO_DELETE: 512 hio->hio_offset = nv_get_uint64(hio->hio_nv, "offset"); 513 if (nv_error(hio->hio_nv) != 0) { 514 pjdlog_error("Header is missing 'offset' field."); 515 hio->hio_error = EINVAL; 516 goto end; 517 } 518 hio->hio_length = nv_get_uint64(hio->hio_nv, "length"); 519 if (nv_error(hio->hio_nv) != 0) { 520 pjdlog_error("Header is missing 'length' field."); 521 hio->hio_error = EINVAL; 522 goto end; 523 } 524 if (hio->hio_length == 0) { 525 pjdlog_error("Data length is zero."); 526 hio->hio_error = EINVAL; 527 goto end; 528 } 529 if (hio->hio_length > MAXPHYS) { 530 pjdlog_error("Data length is too large (%ju > %ju).", 531 (uintmax_t)hio->hio_length, (uintmax_t)MAXPHYS); 532 hio->hio_error = EINVAL; 533 goto end; 534 } 535 if ((hio->hio_offset % res->hr_local_sectorsize) != 0) { 536 pjdlog_error("Offset %ju is not multiple of sector size.", 537 (uintmax_t)hio->hio_offset); 538 hio->hio_error = EINVAL; 539 goto end; 540 } 541 if ((hio->hio_length % res->hr_local_sectorsize) != 0) { 542 pjdlog_error("Length %ju is not multiple of sector size.", 543 (uintmax_t)hio->hio_length); 544 hio->hio_error = EINVAL; 545 goto end; 546 } 547 if (hio->hio_offset + hio->hio_length > 548 (uint64_t)res->hr_datasize) { 549 pjdlog_error("Data offset is too large (%ju > %ju).", 550 (uintmax_t)(hio->hio_offset + hio->hio_length), 551 (uintmax_t)res->hr_datasize); 552 hio->hio_error = EINVAL; 553 goto end; 554 } 555 break; 556 default: 557 pjdlog_error("Header contains invalid 'cmd' (%hhu).", 558 hio->hio_cmd); 559 hio->hio_error = EINVAL; 560 goto end; 561 } 562 hio->hio_error = 0; 563end: 564 return (hio->hio_error); 565} 566 567static __dead2 void 568secondary_exit(int exitcode, const char *fmt, ...) 569{ 570 va_list ap; 571 572 PJDLOG_ASSERT(exitcode != EX_OK); 573 va_start(ap, fmt); 574 pjdlogv_errno(LOG_ERR, fmt, ap); 575 va_end(ap); 576 event_send(gres, EVENT_DISCONNECT); 577 exit(exitcode); 578} 579 580/* 581 * Thread receives requests from the primary node. 582 */ 583static void * 584recv_thread(void *arg) 585{ 586 struct hast_resource *res = arg; 587 struct hio *hio; 588 589 for (;;) { 590 pjdlog_debug(2, "recv: Taking free request."); 591 QUEUE_TAKE(free, hio); 592 pjdlog_debug(2, "recv: (%p) Got request.", hio); 593 if (hast_proto_recv_hdr(res->hr_remotein, &hio->hio_nv) < 0) { 594 secondary_exit(EX_TEMPFAIL, 595 "Unable to receive request header"); 596 } 597 if (requnpack(res, hio) != 0) { 598 pjdlog_debug(2, 599 "recv: (%p) Moving request to the send queue.", 600 hio); 601 QUEUE_INSERT(send, hio); 602 continue; 603 } 604 reqlog(LOG_DEBUG, 2, -1, hio, 605 "recv: (%p) Got request header: ", hio); 606 if (hio->hio_cmd == HIO_KEEPALIVE) { 607 pjdlog_debug(2, 608 "recv: (%p) Moving request to the free queue.", 609 hio); 610 nv_free(hio->hio_nv); 611 QUEUE_INSERT(free, hio); 612 continue; 613 } else if (hio->hio_cmd == HIO_WRITE) { 614 if (hast_proto_recv_data(res, res->hr_remotein, 615 hio->hio_nv, hio->hio_data, MAXPHYS) < 0) { 616 secondary_exit(EX_TEMPFAIL, 617 "Unable to receive request data"); 618 } 619 } 620 pjdlog_debug(2, "recv: (%p) Moving request to the disk queue.", 621 hio); 622 QUEUE_INSERT(disk, hio); 623 } 624 /* NOTREACHED */ 625 return (NULL); 626} 627 628/* 629 * Thread reads from or writes to local component and also handles DELETE and 630 * FLUSH requests. 631 */ 632static void * 633disk_thread(void *arg) 634{ 635 struct hast_resource *res = arg; 636 struct hio *hio; 637 ssize_t ret; 638 bool clear_activemap; 639 640 clear_activemap = true; 641 642 for (;;) { 643 pjdlog_debug(2, "disk: Taking request."); 644 QUEUE_TAKE(disk, hio); 645 while (clear_activemap) { 646 unsigned char *map; 647 size_t mapsize; 648 649 /* 650 * When first request is received, it means that primary 651 * already received our activemap, merged it and stored 652 * locally. We can now safely clear our activemap. 653 */ 654 mapsize = 655 activemap_calc_ondisk_size(res->hr_local_mediasize - 656 METADATA_SIZE, res->hr_extentsize, 657 res->hr_local_sectorsize); 658 map = calloc(1, mapsize); 659 if (map == NULL) { 660 pjdlog_warning("Unable to allocate memory to clear local activemap."); 661 break; 662 } 663 if (pwrite(res->hr_localfd, map, mapsize, 664 METADATA_SIZE) != (ssize_t)mapsize) { 665 pjdlog_errno(LOG_WARNING, 666 "Unable to store cleared activemap"); 667 free(map); 668 break; 669 } 670 free(map); 671 clear_activemap = false; 672 pjdlog_debug(1, "Local activemap cleared."); 673 } 674 reqlog(LOG_DEBUG, 2, -1, hio, "disk: (%p) Got request: ", hio); 675 /* Handle the actual request. */ 676 switch (hio->hio_cmd) { 677 case HIO_READ: 678 ret = pread(res->hr_localfd, hio->hio_data, 679 hio->hio_length, 680 hio->hio_offset + res->hr_localoff); 681 if (ret < 0) 682 hio->hio_error = errno; 683 else if (ret != (int64_t)hio->hio_length) 684 hio->hio_error = EIO; 685 else 686 hio->hio_error = 0; 687 break; 688 case HIO_WRITE: 689 ret = pwrite(res->hr_localfd, hio->hio_data, 690 hio->hio_length, 691 hio->hio_offset + res->hr_localoff); 692 if (ret < 0) 693 hio->hio_error = errno; 694 else if (ret != (int64_t)hio->hio_length) 695 hio->hio_error = EIO; 696 else 697 hio->hio_error = 0; 698 break; 699 case HIO_DELETE: 700 ret = g_delete(res->hr_localfd, 701 hio->hio_offset + res->hr_localoff, 702 hio->hio_length); 703 if (ret < 0) 704 hio->hio_error = errno; 705 else 706 hio->hio_error = 0; 707 break; 708 case HIO_FLUSH: 709 ret = g_flush(res->hr_localfd); 710 if (ret < 0) 711 hio->hio_error = errno; 712 else 713 hio->hio_error = 0; 714 break; 715 } 716 if (hio->hio_error != 0) { 717 reqlog(LOG_ERR, 0, hio->hio_error, hio, 718 "Request failed: "); 719 } 720 pjdlog_debug(2, "disk: (%p) Moving request to the send queue.", 721 hio); 722 QUEUE_INSERT(send, hio); 723 } 724 /* NOTREACHED */ 725 return (NULL); 726} 727 728/* 729 * Thread sends requests back to primary node. 730 */ 731static void * 732send_thread(void *arg) 733{ 734 struct hast_resource *res = arg; 735 struct nv *nvout; 736 struct hio *hio; 737 void *data; 738 size_t length; 739 740 for (;;) { 741 pjdlog_debug(2, "send: Taking request."); 742 QUEUE_TAKE(send, hio); 743 reqlog(LOG_DEBUG, 2, -1, hio, "send: (%p) Got request: ", hio); 744 nvout = nv_alloc(); 745 /* Copy sequence number. */ 746 nv_add_uint64(nvout, nv_get_uint64(hio->hio_nv, "seq"), "seq"); 747 switch (hio->hio_cmd) { 748 case HIO_READ: 749 if (hio->hio_error == 0) { 750 data = hio->hio_data; 751 length = hio->hio_length; 752 break; 753 } 754 /* 755 * We send no data in case of an error. 756 */ 757 /* FALLTHROUGH */ 758 case HIO_DELETE: 759 case HIO_FLUSH: 760 case HIO_WRITE: 761 data = NULL; 762 length = 0; 763 break; 764 default: 765 abort(); 766 break; 767 } 768 if (hio->hio_error != 0) 769 nv_add_int16(nvout, hio->hio_error, "error"); 770 if (hast_proto_send(res, res->hr_remoteout, nvout, data, 771 length) < 0) { 772 secondary_exit(EX_TEMPFAIL, "Unable to send reply."); 773 } 774 nv_free(nvout); 775 pjdlog_debug(2, "send: (%p) Moving request to the free queue.", 776 hio); 777 nv_free(hio->hio_nv); 778 hio->hio_error = 0; 779 QUEUE_INSERT(free, hio); 780 } 781 /* NOTREACHED */ 782 return (NULL); 783} 784