secondary.c revision 222164
1283625Sdim/*- 2283625Sdim * Copyright (c) 2009-2010 The FreeBSD Foundation 3283625Sdim * Copyright (c) 2010 Pawel Jakub Dawidek <pjd@FreeBSD.org> 4283625Sdim * All rights reserved. 5283625Sdim * 6283625Sdim * This software was developed by Pawel Jakub Dawidek under sponsorship from 7283625Sdim * the FreeBSD Foundation. 8283625Sdim * 9283625Sdim * Redistribution and use in source and binary forms, with or without 10283625Sdim * modification, are permitted provided that the following conditions 11283625Sdim * are met: 12283625Sdim * 1. Redistributions of source code must retain the above copyright 13283625Sdim * notice, this list of conditions and the following disclaimer. 14283625Sdim * 2. Redistributions in binary form must reproduce the above copyright 15283625Sdim * notice, this list of conditions and the following disclaimer in the 16283625Sdim * documentation and/or other materials provided with the distribution. 17283625Sdim * 18283625Sdim * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND 19283625Sdim * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20283625Sdim * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21283625Sdim * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE 22283625Sdim * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23283625Sdim * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24283625Sdim * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25283625Sdim * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26283625Sdim * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27283625Sdim * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28283625Sdim * SUCH DAMAGE. 29283625Sdim */ 30283625Sdim 31283625Sdim#include <sys/cdefs.h> 32283625Sdim__FBSDID("$FreeBSD: head/sbin/hastd/secondary.c 222164 2011-05-21 20:21:20Z pjd $"); 33283625Sdim 34283625Sdim#include <sys/param.h> 35283625Sdim#include <sys/time.h> 36283625Sdim#include <sys/bio.h> 37283625Sdim#include <sys/disk.h> 38283625Sdim#include <sys/stat.h> 39283625Sdim 40283625Sdim#include <err.h> 41283625Sdim#include <errno.h> 42283625Sdim#include <fcntl.h> 43283625Sdim#include <libgeom.h> 44283625Sdim#include <pthread.h> 45283625Sdim#include <signal.h> 46296417Sdim#include <stdint.h> 47296417Sdim#include <stdio.h> 48283625Sdim#include <string.h> 49296417Sdim#include <sysexits.h> 50296417Sdim#include <unistd.h> 51296417Sdim 52283625Sdim#include <activemap.h> 53283625Sdim#include <nv.h> 54283625Sdim#include <pjdlog.h> 55283625Sdim 56283625Sdim#include "control.h" 57283625Sdim#include "event.h" 58283625Sdim#include "hast.h" 59283625Sdim#include "hast_proto.h" 60283625Sdim#include "hastd.h" 61283625Sdim#include "hooks.h" 62283625Sdim#include "metadata.h" 63283625Sdim#include "proto.h" 64283625Sdim#include "subr.h" 65283625Sdim#include "synch.h" 66283625Sdim 67296417Sdimstruct hio { 68296417Sdim uint64_t hio_seq; 69283625Sdim int hio_error; 70283625Sdim struct nv *hio_nv; 71283625Sdim void *hio_data; 72283625Sdim uint8_t hio_cmd; 73283625Sdim uint64_t hio_offset; 74296417Sdim uint64_t hio_length; 75283625Sdim TAILQ_ENTRY(hio) hio_next; 76283625Sdim}; 77283625Sdim 78283625Sdimstatic struct hast_resource *gres; 79283625Sdim 80283625Sdim/* 81283625Sdim * Free list holds unused structures. When free list is empty, we have to wait 82283625Sdim * until some in-progress requests are freed. 83283625Sdim */ 84283625Sdimstatic TAILQ_HEAD(, hio) hio_free_list; 85283625Sdimstatic pthread_mutex_t hio_free_list_lock; 86283625Sdimstatic pthread_cond_t hio_free_list_cond; 87296417Sdim/* 88296417Sdim * Disk thread (the one that do I/O requests) takes requests from this list. 89296417Sdim */ 90296417Sdimstatic TAILQ_HEAD(, hio) hio_disk_list; 91283625Sdimstatic pthread_mutex_t hio_disk_list_lock; 92283625Sdimstatic pthread_cond_t hio_disk_list_cond; 93283625Sdim/* 94283625Sdim * There is one recv list for every component, although local components don't 95283625Sdim * use recv lists as local requests are done synchronously. 96283625Sdim */ 97283625Sdimstatic TAILQ_HEAD(, hio) hio_send_list; 98283625Sdimstatic pthread_mutex_t hio_send_list_lock; 99283625Sdimstatic pthread_cond_t hio_send_list_cond; 100283625Sdim 101283625Sdim/* 102283625Sdim * Maximum number of outstanding I/O requests. 103283625Sdim */ 104283625Sdim#define HAST_HIO_MAX 256 105283625Sdim 106283625Sdimstatic void *recv_thread(void *arg); 107283625Sdimstatic void *disk_thread(void *arg); 108283625Sdimstatic void *send_thread(void *arg); 109283625Sdim 110283625Sdim#define QUEUE_INSERT(name, hio) do { \ 111283625Sdim bool _wakeup; \ 112283625Sdim \ 113283625Sdim mtx_lock(&hio_##name##_list_lock); \ 114283625Sdim _wakeup = TAILQ_EMPTY(&hio_##name##_list); \ 115283625Sdim TAILQ_INSERT_TAIL(&hio_##name##_list, (hio), hio_next); \ 116283625Sdim mtx_unlock(&hio_##name##_list_lock); \ 117283625Sdim if (_wakeup) \ 118283625Sdim cv_signal(&hio_##name##_list_cond); \ 119283625Sdim} while (0) 120283625Sdim#define QUEUE_TAKE(name, hio) do { \ 121283625Sdim mtx_lock(&hio_##name##_list_lock); \ 122283625Sdim while (((hio) = TAILQ_FIRST(&hio_##name##_list)) == NULL) { \ 123283625Sdim cv_wait(&hio_##name##_list_cond, \ 124283625Sdim &hio_##name##_list_lock); \ 125283625Sdim } \ 126296417Sdim TAILQ_REMOVE(&hio_##name##_list, (hio), hio_next); \ 127296417Sdim mtx_unlock(&hio_##name##_list_lock); \ 128296417Sdim} while (0) 129296417Sdim 130296417Sdimstatic void 131283625Sdiminit_environment(void) 132283625Sdim{ 133283625Sdim struct hio *hio; 134283625Sdim unsigned int ii; 135296417Sdim 136283625Sdim /* 137296417Sdim * Initialize lists, their locks and theirs condition variables. 138296417Sdim */ 139296417Sdim TAILQ_INIT(&hio_free_list); 140296417Sdim mtx_init(&hio_free_list_lock); 141296417Sdim cv_init(&hio_free_list_cond); 142296417Sdim TAILQ_INIT(&hio_disk_list); 143296417Sdim mtx_init(&hio_disk_list_lock); 144296417Sdim cv_init(&hio_disk_list_cond); 145296417Sdim TAILQ_INIT(&hio_send_list); 146296417Sdim mtx_init(&hio_send_list_lock); 147296417Sdim cv_init(&hio_send_list_cond); 148296417Sdim 149296417Sdim /* 150296417Sdim * Allocate requests pool and initialize requests. 151296417Sdim */ 152283625Sdim for (ii = 0; ii < HAST_HIO_MAX; ii++) { 153283625Sdim hio = malloc(sizeof(*hio)); 154283625Sdim if (hio == NULL) { 155283625Sdim pjdlog_exitx(EX_TEMPFAIL, 156283625Sdim "Unable to allocate memory (%zu bytes) for hio request.", 157296417Sdim sizeof(*hio)); 158283625Sdim } 159283625Sdim hio->hio_error = 0; 160283625Sdim hio->hio_data = malloc(MAXPHYS); 161283625Sdim if (hio->hio_data == NULL) { 162283625Sdim pjdlog_exitx(EX_TEMPFAIL, 163283625Sdim "Unable to allocate memory (%zu bytes) for gctl_data.", 164283625Sdim (size_t)MAXPHYS); 165283625Sdim } 166283625Sdim TAILQ_INSERT_HEAD(&hio_free_list, hio, hio_next); 167283625Sdim } 168283625Sdim} 169283625Sdim 170283625Sdimstatic void 171283625Sdiminit_local(struct hast_resource *res) 172283625Sdim{ 173296417Sdim 174296417Sdim if (metadata_read(res, true) < 0) 175283625Sdim exit(EX_NOINPUT); 176283625Sdim} 177283625Sdim 178283625Sdimstatic void 179283625Sdiminit_remote(struct hast_resource *res, struct nv *nvin) 180283625Sdim{ 181283625Sdim uint64_t resuid; 182283625Sdim struct nv *nvout; 183296417Sdim unsigned char *map; 184296417Sdim size_t mapsize; 185296417Sdim 186283625Sdim /* Setup direction. */ 187283625Sdim if (proto_send(res->hr_remoteout, NULL, 0) == -1) 188283625Sdim pjdlog_errno(LOG_WARNING, "Unable to set connection direction"); 189283625Sdim 190283625Sdim map = NULL; 191283625Sdim mapsize = 0; 192283625Sdim nvout = nv_alloc(); 193283625Sdim nv_add_int64(nvout, (int64_t)res->hr_datasize, "datasize"); 194283625Sdim nv_add_int32(nvout, (int32_t)res->hr_extentsize, "extentsize"); 195283625Sdim resuid = nv_get_uint64(nvin, "resuid"); 196283625Sdim res->hr_primary_localcnt = nv_get_uint64(nvin, "localcnt"); 197283625Sdim res->hr_primary_remotecnt = nv_get_uint64(nvin, "remotecnt"); 198283625Sdim nv_add_uint64(nvout, res->hr_secondary_localcnt, "localcnt"); 199283625Sdim nv_add_uint64(nvout, res->hr_secondary_remotecnt, "remotecnt"); 200283625Sdim mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize - 201283625Sdim METADATA_SIZE, res->hr_extentsize, res->hr_local_sectorsize); 202283625Sdim map = malloc(mapsize); 203283625Sdim if (map == NULL) { 204283625Sdim pjdlog_exitx(EX_TEMPFAIL, 205283625Sdim "Unable to allocate memory (%zu bytes) for activemap.", 206283625Sdim mapsize); 207283625Sdim } 208283625Sdim /* 209283625Sdim * When we work as primary and secondary is missing we will increase 210283625Sdim * localcnt in our metadata. When secondary is connected and synced 211283625Sdim * we make localcnt be equal to remotecnt, which means nodes are more 212283625Sdim * or less in sync. 213283625Sdim * Split-brain condition is when both nodes are not able to communicate 214283625Sdim * and are both configured as primary nodes. In turn, they can both 215283625Sdim * make incompatible changes to the data and we have to detect that. 216283625Sdim * Under split-brain condition we will increase our localcnt on first 217283625Sdim * write and remote node will increase its localcnt on first write. 218283625Sdim * When we connect we can see that primary's localcnt is greater than 219283625Sdim * our remotecnt (primary was modified while we weren't watching) and 220283625Sdim * our localcnt is greater than primary's remotecnt (we were modified 221283625Sdim * while primary wasn't watching). 222283625Sdim * There are many possible combinations which are all gathered below. 223296417Sdim * Don't pay too much attention to exact numbers, the more important 224296417Sdim * is to compare them. We compare secondary's local with primary's 225283625Sdim * remote and secondary's remote with primary's local. 226283625Sdim * Note that every case where primary's localcnt is smaller than 227283625Sdim * secondary's remotecnt and where secondary's localcnt is smaller than 228283625Sdim * primary's remotecnt should be impossible in practise. We will perform 229283625Sdim * full synchronization then. Those cases are marked with an asterisk. 230283625Sdim * Regular synchronization means that only extents marked as dirty are 231296417Sdim * synchronized (regular synchronization). 232283625Sdim * 233283625Sdim * SECONDARY METADATA PRIMARY METADATA 234283625Sdim * local=3 remote=3 local=2 remote=2* ?! Full sync from secondary. 235283625Sdim * local=3 remote=3 local=2 remote=3* ?! Full sync from primary. 236283625Sdim * local=3 remote=3 local=2 remote=4* ?! Full sync from primary. 237283625Sdim * local=3 remote=3 local=3 remote=2 Primary is out-of-date, 238296417Sdim * regular sync from secondary. 239296417Sdim * local=3 remote=3 local=3 remote=3 Regular sync just in case. 240296417Sdim * local=3 remote=3 local=3 remote=4* ?! Full sync from primary. 241296417Sdim * local=3 remote=3 local=4 remote=2 Split-brain condition. 242296417Sdim * local=3 remote=3 local=4 remote=3 Secondary out-of-date, 243296417Sdim * regular sync from primary. 244296417Sdim * local=3 remote=3 local=4 remote=4* ?! Full sync from primary. 245296417Sdim */ 246296417Sdim if (res->hr_resuid == 0) { 247283625Sdim /* 248296417Sdim * Provider is used for the first time. If primary node done no 249296417Sdim * writes yet as well (we will find "virgin" argument) then 250296417Sdim * there is no need to synchronize anything. If primary node 251283625Sdim * done any writes already we have to synchronize everything. 252283625Sdim */ 253283625Sdim PJDLOG_ASSERT(res->hr_secondary_localcnt == 0); 254283625Sdim res->hr_resuid = resuid; 255283625Sdim if (metadata_write(res) < 0) 256283625Sdim exit(EX_NOINPUT); 257283625Sdim if (nv_exists(nvin, "virgin")) { 258283625Sdim free(map); 259283625Sdim map = NULL; 260283625Sdim mapsize = 0; 261283625Sdim } else { 262283625Sdim memset(map, 0xff, mapsize); 263283625Sdim } 264283625Sdim nv_add_int8(nvout, 1, "virgin"); 265283625Sdim nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 266283625Sdim } else if (res->hr_resuid != resuid) { 267296417Sdim char errmsg[256]; 268296417Sdim 269283625Sdim (void)snprintf(errmsg, sizeof(errmsg), 270283625Sdim "Resource unique ID mismatch (primary=%ju, secondary=%ju).", 271283625Sdim (uintmax_t)resuid, (uintmax_t)res->hr_resuid); 272296417Sdim pjdlog_error("%s", errmsg); 273296417Sdim nv_add_string(nvout, errmsg, "errmsg"); 274283625Sdim if (hast_proto_send(res, res->hr_remotein, nvout, NULL, 0) < 0) { 275283625Sdim pjdlog_exit(EX_TEMPFAIL, "Unable to send response to %s", 276283625Sdim res->hr_remoteaddr); 277283625Sdim } 278283625Sdim nv_free(nvout); 279283625Sdim exit(EX_CONFIG); 280283625Sdim } else if ( 281283625Sdim /* Is primary is out-of-date? */ 282283625Sdim (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 283283625Sdim res->hr_secondary_remotecnt == res->hr_primary_localcnt) || 284283625Sdim /* Nodes are more or less in sync? */ 285283625Sdim (res->hr_secondary_localcnt == res->hr_primary_remotecnt && 286283625Sdim res->hr_secondary_remotecnt == res->hr_primary_localcnt) || 287283625Sdim /* Is secondary is out-of-date? */ 288283625Sdim (res->hr_secondary_localcnt == res->hr_primary_remotecnt && 289283625Sdim res->hr_secondary_remotecnt < res->hr_primary_localcnt)) { 290283625Sdim /* 291283625Sdim * Nodes are more or less in sync or one of the nodes is 292283625Sdim * out-of-date. 293283625Sdim * It doesn't matter at this point which one, we just have to 294296417Sdim * send out local bitmap to the remote node. 295283625Sdim */ 296283625Sdim if (pread(res->hr_localfd, map, mapsize, METADATA_SIZE) != 297283625Sdim (ssize_t)mapsize) { 298283625Sdim pjdlog_exit(LOG_ERR, "Unable to read activemap"); 299283625Sdim } 300283625Sdim if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 301283625Sdim res->hr_secondary_remotecnt == res->hr_primary_localcnt) { 302283625Sdim /* Primary is out-of-date, sync from secondary. */ 303283625Sdim nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc"); 304283625Sdim } else { 305283625Sdim /* 306283625Sdim * Secondary is out-of-date or counts match. 307283625Sdim * Sync from primary. 308283625Sdim */ 309283625Sdim nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 310283625Sdim } 311283625Sdim } else if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 312283625Sdim res->hr_primary_localcnt > res->hr_secondary_remotecnt) { 313283625Sdim /* 314283625Sdim * Not good, we have split-brain condition. 315283625Sdim */ 316283625Sdim pjdlog_error("Split-brain detected, exiting."); 317283625Sdim nv_add_string(nvout, "Split-brain condition!", "errmsg"); 318283625Sdim free(map); 319283625Sdim map = NULL; 320283625Sdim mapsize = 0; 321283625Sdim } else /* if (res->hr_secondary_localcnt < res->hr_primary_remotecnt || 322283625Sdim res->hr_primary_localcnt < res->hr_secondary_remotecnt) */ { 323283625Sdim /* 324283625Sdim * This should never happen in practise, but we will perform 325296417Sdim * full synchronization. 326296417Sdim */ 327296417Sdim PJDLOG_ASSERT(res->hr_secondary_localcnt < res->hr_primary_remotecnt || 328296417Sdim res->hr_primary_localcnt < res->hr_secondary_remotecnt); 329296417Sdim mapsize = activemap_calc_ondisk_size(res->hr_local_mediasize - 330296417Sdim METADATA_SIZE, res->hr_extentsize, 331296417Sdim res->hr_local_sectorsize); 332296417Sdim memset(map, 0xff, mapsize); 333296417Sdim if (res->hr_secondary_localcnt > res->hr_primary_remotecnt) { 334296417Sdim /* In this one of five cases sync from secondary. */ 335296417Sdim nv_add_uint8(nvout, HAST_SYNCSRC_SECONDARY, "syncsrc"); 336296417Sdim } else { 337296417Sdim /* For the rest four cases sync from primary. */ 338296417Sdim nv_add_uint8(nvout, HAST_SYNCSRC_PRIMARY, "syncsrc"); 339296417Sdim } 340296417Sdim pjdlog_warning("This should never happen, asking for full synchronization (primary(local=%ju, remote=%ju), secondary(local=%ju, remote=%ju)).", 341296417Sdim (uintmax_t)res->hr_primary_localcnt, 342296417Sdim (uintmax_t)res->hr_primary_remotecnt, 343283625Sdim (uintmax_t)res->hr_secondary_localcnt, 344283625Sdim (uintmax_t)res->hr_secondary_remotecnt); 345283625Sdim } 346283625Sdim nv_add_uint32(nvout, (uint32_t)mapsize, "mapsize"); 347283625Sdim if (hast_proto_send(res, res->hr_remotein, nvout, map, mapsize) < 0) { 348283625Sdim pjdlog_exit(EX_TEMPFAIL, "Unable to send activemap to %s", 349283625Sdim res->hr_remoteaddr); 350283625Sdim } 351283625Sdim if (map != NULL) 352283625Sdim free(map); 353296417Sdim nv_free(nvout); 354296417Sdim /* Setup direction. */ 355296417Sdim if (proto_recv(res->hr_remotein, NULL, 0) == -1) 356296417Sdim pjdlog_errno(LOG_WARNING, "Unable to set connection direction"); 357296417Sdim if (res->hr_secondary_localcnt > res->hr_primary_remotecnt && 358283625Sdim res->hr_primary_localcnt > res->hr_secondary_remotecnt) { 359296417Sdim /* Exit on split-brain. */ 360296417Sdim event_send(res, EVENT_SPLITBRAIN); 361296417Sdim exit(EX_CONFIG); 362296417Sdim } 363296417Sdim} 364296417Sdim 365296417Sdimvoid 366283625Sdimhastd_secondary(struct hast_resource *res, struct nv *nvin) 367296417Sdim{ 368296417Sdim sigset_t mask; 369296417Sdim pthread_t td; 370296417Sdim pid_t pid; 371296417Sdim int error, mode, debuglevel; 372296417Sdim 373296417Sdim /* 374296417Sdim * Create communication channel between parent and child. 375296417Sdim */ 376296417Sdim if (proto_client(NULL, "socketpair://", &res->hr_ctrl) < 0) { 377296417Sdim KEEP_ERRNO((void)pidfile_remove(pfh)); 378296417Sdim pjdlog_exit(EX_OSERR, 379296417Sdim "Unable to create control sockets between parent and child"); 380296417Sdim } 381296417Sdim /* 382296417Sdim * Create communication channel between child and parent. 383296417Sdim */ 384283625Sdim if (proto_client(NULL, "socketpair://", &res->hr_event) < 0) { 385283625Sdim KEEP_ERRNO((void)pidfile_remove(pfh)); 386283625Sdim pjdlog_exit(EX_OSERR, 387283625Sdim "Unable to create event sockets between child and parent"); 388296417Sdim } 389296417Sdim 390296417Sdim pid = fork(); 391296417Sdim if (pid < 0) { 392296417Sdim KEEP_ERRNO((void)pidfile_remove(pfh)); 393296417Sdim pjdlog_exit(EX_OSERR, "Unable to fork"); 394296417Sdim } 395296417Sdim 396296417Sdim if (pid > 0) { 397296417Sdim /* This is parent. */ 398296417Sdim proto_close(res->hr_remotein); 399296417Sdim res->hr_remotein = NULL; 400296417Sdim proto_close(res->hr_remoteout); 401296417Sdim res->hr_remoteout = NULL; 402296417Sdim /* Declare that we are receiver. */ 403296417Sdim proto_recv(res->hr_event, NULL, 0); 404296417Sdim /* Declare that we are sender. */ 405296417Sdim proto_send(res->hr_ctrl, NULL, 0); 406296417Sdim res->hr_workerpid = pid; 407296417Sdim return; 408296417Sdim } 409296417Sdim 410296417Sdim gres = res; 411296417Sdim mode = pjdlog_mode_get(); 412296417Sdim debuglevel = pjdlog_debug_get(); 413296417Sdim 414296417Sdim /* Declare that we are sender. */ 415296417Sdim proto_send(res->hr_event, NULL, 0); 416296417Sdim /* Declare that we are receiver. */ 417296417Sdim proto_recv(res->hr_ctrl, NULL, 0); 418296417Sdim descriptors_cleanup(res); 419296417Sdim 420296417Sdim descriptors_assert(res, mode); 421296417Sdim 422296417Sdim pjdlog_init(mode); 423283625Sdim pjdlog_debug_set(debuglevel); 424296417Sdim pjdlog_prefix_set("[%s] (%s) ", res->hr_name, role2str(res->hr_role)); 425283625Sdim setproctitle("%s (%s)", res->hr_name, role2str(res->hr_role)); 426296417Sdim 427283625Sdim PJDLOG_VERIFY(sigemptyset(&mask) == 0); 428283625Sdim PJDLOG_VERIFY(sigprocmask(SIG_SETMASK, &mask, NULL) == 0); 429283625Sdim 430283625Sdim /* Error in setting timeout is not critical, but why should it fail? */ 431296417Sdim if (proto_timeout(res->hr_remotein, 2 * HAST_KEEPALIVE) < 0) 432296417Sdim pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 433296417Sdim if (proto_timeout(res->hr_remoteout, res->hr_timeout) < 0) 434296417Sdim pjdlog_errno(LOG_WARNING, "Unable to set connection timeout"); 435296417Sdim 436296417Sdim init_local(res); 437296417Sdim init_environment(); 438296417Sdim 439296417Sdim if (drop_privs(res) != 0) 440296417Sdim exit(EX_CONFIG); 441296417Sdim pjdlog_info("Privileges successfully dropped."); 442296417Sdim 443296417Sdim /* 444296417Sdim * Create the control thread before sending any event to the parent, 445296417Sdim * as we can deadlock when parent sends control request to worker, 446283625Sdim * but worker has no control thread started yet, so parent waits. 447283625Sdim * In the meantime worker sends an event to the parent, but parent 448283625Sdim * is unable to handle the event, because it waits for control 449283625Sdim * request response. 450296417Sdim */ 451296417Sdim error = pthread_create(&td, NULL, ctrl_thread, res); 452296417Sdim PJDLOG_ASSERT(error == 0); 453296417Sdim 454296417Sdim init_remote(res, nvin); 455283625Sdim event_send(res, EVENT_CONNECT); 456296417Sdim 457283625Sdim error = pthread_create(&td, NULL, recv_thread, res); 458283625Sdim PJDLOG_ASSERT(error == 0); 459283625Sdim error = pthread_create(&td, NULL, disk_thread, res); 460296417Sdim PJDLOG_ASSERT(error == 0); 461283625Sdim (void)send_thread(res); 462283625Sdim} 463283625Sdim 464283625Sdimstatic void 465283625Sdimreqlog(int loglevel, int debuglevel, int error, struct hio *hio, const char *fmt, ...) 466283625Sdim{ 467283625Sdim char msg[1024]; 468283625Sdim va_list ap; 469283625Sdim int len; 470283625Sdim 471283625Sdim va_start(ap, fmt); 472283625Sdim len = vsnprintf(msg, sizeof(msg), fmt, ap); 473283625Sdim va_end(ap); 474283625Sdim if ((size_t)len < sizeof(msg)) { 475283625Sdim switch (hio->hio_cmd) { 476283625Sdim case HIO_READ: 477283625Sdim (void)snprintf(msg + len, sizeof(msg) - len, 478283625Sdim "READ(%ju, %ju).", (uintmax_t)hio->hio_offset, 479283625Sdim (uintmax_t)hio->hio_length); 480283625Sdim break; 481283625Sdim case HIO_DELETE: 482283625Sdim (void)snprintf(msg + len, sizeof(msg) - len, 483283625Sdim "DELETE(%ju, %ju).", (uintmax_t)hio->hio_offset, 484283625Sdim (uintmax_t)hio->hio_length); 485283625Sdim break; 486283625Sdim case HIO_FLUSH: 487283625Sdim (void)snprintf(msg + len, sizeof(msg) - len, "FLUSH."); 488283625Sdim break; 489283625Sdim case HIO_WRITE: 490283625Sdim (void)snprintf(msg + len, sizeof(msg) - len, 491283625Sdim "WRITE(%ju, %ju).", (uintmax_t)hio->hio_offset, 492283625Sdim (uintmax_t)hio->hio_length); 493283625Sdim break; 494283625Sdim case HIO_KEEPALIVE: 495283625Sdim (void)snprintf(msg + len, sizeof(msg) - len, "KEEPALIVE."); 496283625Sdim break; 497283625Sdim default: 498283625Sdim (void)snprintf(msg + len, sizeof(msg) - len, 499283625Sdim "UNKNOWN(%u).", (unsigned int)hio->hio_cmd); 500283625Sdim break; 501283625Sdim } 502283625Sdim } 503283625Sdim pjdlog_common(loglevel, debuglevel, error, "%s", msg); 504283625Sdim} 505283625Sdim 506283625Sdimstatic int 507283625Sdimrequnpack(struct hast_resource *res, struct hio *hio) 508283625Sdim{ 509283625Sdim 510283625Sdim hio->hio_cmd = nv_get_uint8(hio->hio_nv, "cmd"); 511283625Sdim if (hio->hio_cmd == 0) { 512296417Sdim pjdlog_error("Header contains no 'cmd' field."); 513283625Sdim hio->hio_error = EINVAL; 514283625Sdim goto end; 515283625Sdim } 516283625Sdim switch (hio->hio_cmd) { 517283625Sdim case HIO_FLUSH: 518283625Sdim case HIO_KEEPALIVE: 519283625Sdim break; 520283625Sdim case HIO_READ: 521283625Sdim case HIO_WRITE: 522283625Sdim case HIO_DELETE: 523283625Sdim hio->hio_offset = nv_get_uint64(hio->hio_nv, "offset"); 524283625Sdim if (nv_error(hio->hio_nv) != 0) { 525283625Sdim pjdlog_error("Header is missing 'offset' field."); 526283625Sdim hio->hio_error = EINVAL; 527283625Sdim goto end; 528283625Sdim } 529283625Sdim hio->hio_length = nv_get_uint64(hio->hio_nv, "length"); 530296417Sdim if (nv_error(hio->hio_nv) != 0) { 531296417Sdim pjdlog_error("Header is missing 'length' field."); 532296417Sdim hio->hio_error = EINVAL; 533296417Sdim goto end; 534296417Sdim } 535296417Sdim if (hio->hio_length == 0) { 536296417Sdim pjdlog_error("Data length is zero."); 537296417Sdim hio->hio_error = EINVAL; 538296417Sdim goto end; 539296417Sdim } 540296417Sdim if (hio->hio_length > MAXPHYS) { 541296417Sdim pjdlog_error("Data length is too large (%ju > %ju).", 542296417Sdim (uintmax_t)hio->hio_length, (uintmax_t)MAXPHYS); 543296417Sdim hio->hio_error = EINVAL; 544296417Sdim goto end; 545296417Sdim } 546296417Sdim if ((hio->hio_offset % res->hr_local_sectorsize) != 0) { 547296417Sdim pjdlog_error("Offset %ju is not multiple of sector size.", 548296417Sdim (uintmax_t)hio->hio_offset); 549296417Sdim hio->hio_error = EINVAL; 550296417Sdim goto end; 551296417Sdim } 552296417Sdim if ((hio->hio_length % res->hr_local_sectorsize) != 0) { 553296417Sdim pjdlog_error("Length %ju is not multiple of sector size.", 554296417Sdim (uintmax_t)hio->hio_length); 555296417Sdim hio->hio_error = EINVAL; 556296417Sdim goto end; 557 } 558 if (hio->hio_offset + hio->hio_length > 559 (uint64_t)res->hr_datasize) { 560 pjdlog_error("Data offset is too large (%ju > %ju).", 561 (uintmax_t)(hio->hio_offset + hio->hio_length), 562 (uintmax_t)res->hr_datasize); 563 hio->hio_error = EINVAL; 564 goto end; 565 } 566 break; 567 default: 568 pjdlog_error("Header contains invalid 'cmd' (%hhu).", 569 hio->hio_cmd); 570 hio->hio_error = EINVAL; 571 goto end; 572 } 573 hio->hio_error = 0; 574end: 575 return (hio->hio_error); 576} 577 578static __dead2 void 579secondary_exit(int exitcode, const char *fmt, ...) 580{ 581 va_list ap; 582 583 PJDLOG_ASSERT(exitcode != EX_OK); 584 va_start(ap, fmt); 585 pjdlogv_errno(LOG_ERR, fmt, ap); 586 va_end(ap); 587 event_send(gres, EVENT_DISCONNECT); 588 exit(exitcode); 589} 590 591/* 592 * Thread receives requests from the primary node. 593 */ 594static void * 595recv_thread(void *arg) 596{ 597 struct hast_resource *res = arg; 598 struct hio *hio; 599 600 for (;;) { 601 pjdlog_debug(2, "recv: Taking free request."); 602 QUEUE_TAKE(free, hio); 603 pjdlog_debug(2, "recv: (%p) Got request.", hio); 604 if (hast_proto_recv_hdr(res->hr_remotein, &hio->hio_nv) < 0) { 605 secondary_exit(EX_TEMPFAIL, 606 "Unable to receive request header"); 607 } 608 if (requnpack(res, hio) != 0) { 609 pjdlog_debug(2, 610 "recv: (%p) Moving request to the send queue.", 611 hio); 612 QUEUE_INSERT(send, hio); 613 continue; 614 } 615 reqlog(LOG_DEBUG, 2, -1, hio, 616 "recv: (%p) Got request header: ", hio); 617 if (hio->hio_cmd == HIO_KEEPALIVE) { 618 pjdlog_debug(2, 619 "recv: (%p) Moving request to the free queue.", 620 hio); 621 nv_free(hio->hio_nv); 622 QUEUE_INSERT(free, hio); 623 continue; 624 } else if (hio->hio_cmd == HIO_WRITE) { 625 if (hast_proto_recv_data(res, res->hr_remotein, 626 hio->hio_nv, hio->hio_data, MAXPHYS) < 0) { 627 secondary_exit(EX_TEMPFAIL, 628 "Unable to receive request data"); 629 } 630 } 631 pjdlog_debug(2, "recv: (%p) Moving request to the disk queue.", 632 hio); 633 QUEUE_INSERT(disk, hio); 634 } 635 /* NOTREACHED */ 636 return (NULL); 637} 638 639/* 640 * Thread reads from or writes to local component and also handles DELETE and 641 * FLUSH requests. 642 */ 643static void * 644disk_thread(void *arg) 645{ 646 struct hast_resource *res = arg; 647 struct hio *hio; 648 ssize_t ret; 649 bool clear_activemap; 650 651 clear_activemap = true; 652 653 for (;;) { 654 pjdlog_debug(2, "disk: Taking request."); 655 QUEUE_TAKE(disk, hio); 656 while (clear_activemap) { 657 unsigned char *map; 658 size_t mapsize; 659 660 /* 661 * When first request is received, it means that primary 662 * already received our activemap, merged it and stored 663 * locally. We can now safely clear our activemap. 664 */ 665 mapsize = 666 activemap_calc_ondisk_size(res->hr_local_mediasize - 667 METADATA_SIZE, res->hr_extentsize, 668 res->hr_local_sectorsize); 669 map = calloc(1, mapsize); 670 if (map == NULL) { 671 pjdlog_warning("Unable to allocate memory to clear local activemap."); 672 break; 673 } 674 if (pwrite(res->hr_localfd, map, mapsize, 675 METADATA_SIZE) != (ssize_t)mapsize) { 676 pjdlog_errno(LOG_WARNING, 677 "Unable to store cleared activemap"); 678 free(map); 679 break; 680 } 681 free(map); 682 clear_activemap = false; 683 pjdlog_debug(1, "Local activemap cleared."); 684 } 685 reqlog(LOG_DEBUG, 2, -1, hio, "disk: (%p) Got request: ", hio); 686 /* Handle the actual request. */ 687 switch (hio->hio_cmd) { 688 case HIO_READ: 689 ret = pread(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_WRITE: 700 ret = pwrite(res->hr_localfd, hio->hio_data, 701 hio->hio_length, 702 hio->hio_offset + res->hr_localoff); 703 if (ret < 0) 704 hio->hio_error = errno; 705 else if (ret != (int64_t)hio->hio_length) 706 hio->hio_error = EIO; 707 else 708 hio->hio_error = 0; 709 break; 710 case HIO_DELETE: 711 ret = g_delete(res->hr_localfd, 712 hio->hio_offset + res->hr_localoff, 713 hio->hio_length); 714 if (ret < 0) 715 hio->hio_error = errno; 716 else 717 hio->hio_error = 0; 718 break; 719 case HIO_FLUSH: 720 ret = g_flush(res->hr_localfd); 721 if (ret < 0) 722 hio->hio_error = errno; 723 else 724 hio->hio_error = 0; 725 break; 726 } 727 if (hio->hio_error != 0) { 728 reqlog(LOG_ERR, 0, hio->hio_error, hio, 729 "Request failed: "); 730 } 731 pjdlog_debug(2, "disk: (%p) Moving request to the send queue.", 732 hio); 733 QUEUE_INSERT(send, hio); 734 } 735 /* NOTREACHED */ 736 return (NULL); 737} 738 739/* 740 * Thread sends requests back to primary node. 741 */ 742static void * 743send_thread(void *arg) 744{ 745 struct hast_resource *res = arg; 746 struct nv *nvout; 747 struct hio *hio; 748 void *data; 749 size_t length; 750 751 for (;;) { 752 pjdlog_debug(2, "send: Taking request."); 753 QUEUE_TAKE(send, hio); 754 reqlog(LOG_DEBUG, 2, -1, hio, "send: (%p) Got request: ", hio); 755 nvout = nv_alloc(); 756 /* Copy sequence number. */ 757 nv_add_uint64(nvout, nv_get_uint64(hio->hio_nv, "seq"), "seq"); 758 switch (hio->hio_cmd) { 759 case HIO_READ: 760 if (hio->hio_error == 0) { 761 data = hio->hio_data; 762 length = hio->hio_length; 763 break; 764 } 765 /* 766 * We send no data in case of an error. 767 */ 768 /* FALLTHROUGH */ 769 case HIO_DELETE: 770 case HIO_FLUSH: 771 case HIO_WRITE: 772 data = NULL; 773 length = 0; 774 break; 775 default: 776 abort(); 777 break; 778 } 779 if (hio->hio_error != 0) 780 nv_add_int16(nvout, hio->hio_error, "error"); 781 if (hast_proto_send(res, res->hr_remoteout, nvout, data, 782 length) < 0) { 783 secondary_exit(EX_TEMPFAIL, "Unable to send reply."); 784 } 785 nv_free(nvout); 786 pjdlog_debug(2, "send: (%p) Moving request to the free queue.", 787 hio); 788 nv_free(hio->hio_nv); 789 hio->hio_error = 0; 790 QUEUE_INSERT(free, hio); 791 } 792 /* NOTREACHED */ 793 return (NULL); 794} 795