msg.c

   1 /*
   2  * Copyright (C) 2008 Intel Corporation
   3  *
   4  *      mdmon socket / message handling
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms and conditions of the GNU General Public License,
   8  * version 2, as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  13  * more details.
  14  *
  15  * You should have received a copy of the GNU General Public License along with
  16  * this program; if not, write to the Free Software Foundation, Inc.,
  17  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  18  */
  19 #ifndef _GNU_SOURCE
  20 #define _GNU_SOURCE
  21 #endif
  22 #include <unistd.h>
  23 #include <stdio.h>
  24 #include <string.h>
  25 #include <stdlib.h>
  26 #include <errno.h>
  27 #include <fcntl.h>
  28 #include <sys/types.h>
  29 #include <sys/socket.h>
  30 #include <sys/un.h>
  31 #include "mdadm.h"
  32 #include "mdmon.h"
  33
  34 static const __u32 start_magic = 0x5a5aa5a5;
  35 static const __u32 end_magic = 0xa5a55a5a;
  36
  37 static int send_buf(int fd, const void* buf, int len, int tmo)
  38 {
  39         fd_set set;
  40         int rv;
  41         struct timeval timeout = {tmo, 0};
  42         struct timeval *ptmo = tmo ? &timeout : NULL;
  43
  44         while (len) {
  45                 FD_ZERO(&set);
  46                 FD_SET(fd, &set);
  47                 rv = select(fd+1, NULL, &set, NULL, ptmo);
  48                 if (rv <= 0)
  49                         return -1;
  50                 rv = write(fd, buf, len);
  51                 if (rv <= 0)
  52                         return -1;
  53                 len -= rv;
  54                 buf += rv;
  55         }
  56         return 0;
  57 }
  58
  59 static int recv_buf(int fd, void* buf, int len, int tmo)
  60 {
  61         fd_set set;
  62         int rv;
  63         struct timeval timeout = {tmo, 0};
  64         struct timeval *ptmo = tmo ? &timeout : NULL;
  65
  66         while (len) {
  67                 FD_ZERO(&set);
  68                 FD_SET(fd, &set);
  69                 rv = select(fd+1, &set, NULL, NULL, ptmo);
  70                 if (rv <= 0)
  71                         return -1;
  72                 rv = read(fd, buf, len);
  73                 if (rv <= 0)
  74                         return -1;
  75                 len -= rv;
  76                 buf += rv;
  77         }
  78         return 0;
  79 }
  80
  81
  82 int send_message(int fd, struct metadata_update *msg, int tmo)
  83 {
  84         __s32 len = msg->len;
  85         int rv;
  86
  87         rv = send_buf(fd, &start_magic, 4, tmo);
  88         rv = rv ?: send_buf(fd, &len, 4, tmo);
  89         if (len > 0)
  90                 rv = rv ?: send_buf(fd, msg->buf, msg->len, tmo);
  91         rv = send_buf(fd, &end_magic, 4, tmo);
  92
  93         return rv;
  94 }
  95
  96 int receive_message(int fd, struct metadata_update *msg, int tmo)
  97 {
  98         __u32 magic;
  99         __s32 len;
 100         int rv;
 101
 102         rv = recv_buf(fd, &magic, 4, tmo);
 103         if (rv < 0 || magic != start_magic)
 104                 return -1;
 105         rv = recv_buf(fd, &len, 4, tmo);
 106         if (rv < 0 || len > MSG_MAX_LEN)
 107                 return -1;
 108         if (len > 0) {
 109                 msg->buf = xmalloc(len);
 110                 rv = recv_buf(fd, msg->buf, len, tmo);
 111                 if (rv < 0) {
 112                         free(msg->buf);
 113                         return -1;
 114                 }
 115         } else
 116                 msg->buf = NULL;
 117         rv = recv_buf(fd, &magic, 4, tmo);
 118         if (rv < 0 || magic != end_magic) {
 119                 free(msg->buf);
 120                 return -1;
 121         }
 122         msg->len = len;
 123         return 0;
 124 }
 125
 126 int ack(int fd, int tmo)
 127 {
 128         struct metadata_update msg = { .len = 0 };
 129
 130         return send_message(fd, &msg, tmo);
 131 }
 132
 133 int wait_reply(int fd, int tmo)
 134 {
 135         struct metadata_update msg;
 136         int err = receive_message(fd, &msg, tmo);
 137
 138         /* mdmon sent extra data, but caller only cares that we got a
 139          * successful reply
 140          */
 141         if (err == 0 && msg.len > 0)
 142                 free(msg.buf);
 143
 144         return err;
 145 }
 146
 147 int connect_monitor(char *devname)
 148 {
 149         char path[100];
 150         int sfd;
 151         long fl;
 152         struct sockaddr_un addr;
 153         int pos;
 154         char *c;
 155
 156         pos = sprintf(path, "%s/", MDMON_DIR);
 157         if (is_subarray(devname)) {
 158                 devname++;
 159                 c = strchr(devname, '/');
 160                 if (!c)
 161                         return -1;
 162                 snprintf(&path[pos], c - devname + 1, "%s", devname);
 163                 pos += c - devname;
 164         } else
 165                 pos += sprintf(&path[pos], "%s", devname);
 166         sprintf(&path[pos], ".sock");
 167
 168         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 169         if (sfd < 0)
 170                 return -1;
 171
 172         addr.sun_family = PF_LOCAL;
 173         strcpy(addr.sun_path, path);
 174         if (connect(sfd, &addr, sizeof(addr)) < 0) {
 175                 close(sfd);
 176                 return -1;
 177         }
 178
 179         fl = fcntl(sfd, F_GETFL, 0);
 180         fl |= O_NONBLOCK;
 181         fcntl(sfd, F_SETFL, fl);
 182
 183         return sfd;
 184 }
 185
 186 int fping_monitor(int sfd)
 187 {
 188         int err = 0;
 189
 190         if (sfd < 0)
 191                 return sfd;
 192
 193         /* try to ping existing socket */
 194         if (ack(sfd, 20) != 0)
 195                 err = -1;
 196
 197         /* check the reply */
 198         if (!err && wait_reply(sfd, 20) != 0)
 199                 err = -1;
 200
 201         return err;
 202 }
 203
 204 /* give the monitor a chance to update the metadata */
 205 int ping_monitor(char *devname)
 206 {
 207         int sfd = connect_monitor(devname);
 208         int err;
 209
 210         if (sfd >= 0) {
 211                 err = fping_monitor(sfd);
 212                 close(sfd);
 213         } else
 214                 err = -1;
 215
 216         return err;
 217 }
 218
 219 /* ping monitor using device number */
 220 int ping_monitor_by_id(int devnum)
 221 {
 222         int err = -1;
 223         char *container = devnum2devname(devnum);
 224
 225         if (container) {
 226                 err = ping_monitor(container);
 227                 free(container);
 228         }
 229
 230         return err;
 231 }
 232
 233 static char *ping_monitor_version(char *devname)
 234 {
 235         int sfd = connect_monitor(devname);
 236         struct metadata_update msg;
 237         int err = 0;
 238
 239         if (sfd < 0)
 240                 return NULL;
 241
 242         if (ack(sfd, 20) != 0)
 243                 err = -1;
 244
 245         if (!err && receive_message(sfd, &msg, 20) != 0)
 246                 err = -1;
 247
 248         close(sfd);
 249
 250         if (err || !msg.len || !msg.buf)
 251                 return NULL;
 252         return msg.buf;
 253 }
 254
 255 int unblock_subarray(struct mdinfo *sra, const int unfreeze)
 256 {
 257         char buf[64];
 258         int rc = 0;
 259
 260         if (sra) {
 261                 sprintf(buf, "external:%s\n", sra->text_version);
 262                 buf[9] = '/';
 263         } else
 264                 buf[9] = '-';
 265
 266         if (buf[9] == '-' ||
 267             sysfs_set_str(sra, NULL, "metadata_version", buf) ||
 268             (unfreeze &&
 269              sysfs_attribute_available(sra, NULL, "sync_action") &&
 270              sysfs_set_str(sra, NULL, "sync_action", "idle")))
 271                 rc = -1;
 272         return rc;
 273 }
 274
 275 int block_subarray(struct mdinfo *sra)
 276 {
 277         char buf[64];
 278         int rc = 0;
 279
 280         sprintf(buf, "external:%s\n", sra->text_version);
 281         buf[9] = '-';
 282         if (sysfs_set_str(sra, NULL, "metadata_version", buf))
 283                 rc = -1;
 284
 285         return rc;
 286 }
 287
 288 /* check mdmon version if it supports
 289  * array blocking mechanism
 290  */
 291 int check_mdmon_version(char *container)
 292 {
 293         char *version = NULL;
 294         int devnum = devname2devnum(container);
 295
 296         if (!mdmon_running(devnum)) {
 297                 /* if mdmon is not active we assume that any instance that is
 298                  * later started will match the current mdadm version, if this
 299                  * assumption is violated we may inadvertantly rebuild an array
 300                  * that was meant for reshape, or start rebuild on a spare that
 301                  * was to be moved to another container
 302                  */
 303                 /* pass */;
 304         } else {
 305                 int ver;
 306
 307                 version = ping_monitor_version(container);
 308                 ver = version ? mdadm_version(version) : -1;
 309                 free(version);
 310                 if (ver < 3002000) {
 311                         pr_err("mdmon instance for %s cannot be disabled\n",
 312                                container);
 313                         return -1;
 314                 }
 315         }
 316
 317         return 0;
 318 }
 319
 320 /**
 321  * block_monitor - prevent mdmon spare assignment
 322  * @container - container to block
 323  * @freeze - flag to additionally freeze sync_action
 324  *
 325  * This is used by the reshape code to freeze the container, and the
 326  * auto-rebuild implementation to atomically move spares.
 327  * In both cases we need to stop mdmon from assigning spares to replace
 328  * failed devices as we might have other plans for the spare.
 329  * For the reshape case we also need to 'freeze' sync_action so that
 330  * no recovery happens until we have fully prepared for the reshape.
 331  *
 332  * We tell mdmon that the array is frozen by marking the 'metadata' name
 333  * with a leading '-'.  The previously told mdmon "Don't make this array
 334  * read/write, leave it readonly".  Now it means a more general "Don't
 335  * reconfigure this array at all".
 336  * As older versions of mdmon (which might run from initrd) don't understand
 337  * this, we first check that the running mdmon is new enough.
 338  */
 339 int block_monitor(char *container, const int freeze)
 340 {
 341         struct mdstat_ent *ent, *e, *e2;
 342         struct mdinfo *sra = NULL;
 343         char buf[64];
 344         int rv = 0;
 345
 346         if (check_mdmon_version(container))
 347                 return -1;
 348
 349         ent = mdstat_read(0, 0);
 350         if (!ent) {
 351                 pr_err("failed to read /proc/mdstat while disabling mdmon\n");
 352                 return -1;
 353         }
 354
 355         /* freeze container contents */
 356         for (e = ent; e; e = e->next) {
 357                 if (!is_container_member(e, container))
 358                         continue;
 359                 sysfs_free(sra);
 360                 sra = sysfs_read(-1, e->devnum, GET_VERSION);
 361                 if (!sra) {
 362                         pr_err("failed to read sysfs for subarray%s\n",
 363                                to_subarray(e, container));
 364                         break;
 365                 }
 366                 /* can't reshape an array that we can't monitor */
 367                 if (sra->text_version[0] == '-')
 368                         break;
 369
 370                 if (freeze && sysfs_freeze_array(sra) < 1)
 371                         break;
 372                 /* flag this array to not be modified by mdmon (close race with
 373                  * takeover in reshape case and spare reassignment in the
 374                  * auto-rebuild case)
 375                  */
 376                 if (block_subarray(sra))
 377                         break;
 378                 ping_monitor(container);
 379
 380                 /* check that we did not race with recovery */
 381                 if ((freeze &&
 382                      !sysfs_attribute_available(sra, NULL, "sync_action")) ||
 383                     (freeze &&
 384                      sysfs_attribute_available(sra, NULL, "sync_action") &&
 385                      sysfs_get_str(sra, NULL, "sync_action", buf, 20) > 0 &&
 386                      strcmp(buf, "frozen\n") == 0))
 387                         /* pass */;
 388                 else {
 389                         unblock_subarray(sra, 0);
 390                         break;
 391                 }
 392                 /* Double check against races - there should be no spares
 393                  * or part-spares
 394                  */
 395                 sysfs_free(sra);
 396                 sra = sysfs_read(-1, e->devnum, GET_DEVS | GET_STATE);
 397                 if (sra && sra->array.spare_disks > 0) {
 398                         unblock_subarray(sra, freeze);
 399                         break;
 400                 }
 401         }
 402
 403         if (e) {
 404                 pr_err("failed to freeze subarray%s\n",
 405                         to_subarray(e, container));
 406
 407                 /* thaw the partially frozen container */
 408                 for (e2 = ent; e2 && e2 != e; e2 = e2->next) {
 409                         if (!is_container_member(e2, container))
 410                                 continue;
 411                         sysfs_free(sra);
 412                         sra = sysfs_read(-1, e2->devnum, GET_VERSION);
 413                         if (unblock_subarray(sra, freeze))
 414                                 pr_err("Failed to unfreeze %s\n", e2->dev);
 415                 }
 416
 417                 ping_monitor(container); /* cleared frozen */
 418                 rv = -1;
 419         }
 420
 421         sysfs_free(sra);
 422         free_mdstat(ent);
 423
 424         return rv;
 425 }
 426
 427 void unblock_monitor(char *container, const int unfreeze)
 428 {
 429         struct mdstat_ent *ent, *e;
 430         struct mdinfo *sra = NULL;
 431         int to_ping = 0;
 432
 433         ent = mdstat_read(0, 0);
 434         if (!ent) {
 435                 pr_err("failed to read /proc/mdstat while unblocking container\n");
 436                 return;
 437         }
 438
 439         /* unfreeze container contents */
 440         for (e = ent; e; e = e->next) {
 441                 if (!is_container_member(e, container))
 442                         continue;
 443                 sysfs_free(sra);
 444                 sra = sysfs_read(-1, e->devnum, GET_VERSION|GET_LEVEL);
 445                 if (!sra)
 446                         continue;
 447                 if (sra->array.level > 0)
 448                         to_ping++;
 449                 if (unblock_subarray(sra, unfreeze))
 450                         pr_err("Failed to unfreeze %s\n", e->dev);
 451         }
 452         if (to_ping)
 453                 ping_monitor(container);
 454
 455         sysfs_free(sra);
 456         free_mdstat(ent);
 457 }
 458
 459
 460
 461 /* give the manager a chance to view the updated container state.  This
 462  * would naturally happen due to the manager noticing a change in
 463  * /proc/mdstat; however, pinging encourages this detection to happen
 464  * while an exclusive open() on the container is active
 465  */
 466 int ping_manager(char *devname)
 467 {
 468         int sfd = connect_monitor(devname);
 469         struct metadata_update msg = { .len = -1 };
 470         int err = 0;
 471
 472         if (sfd < 0)
 473                 return sfd;
 474
 475         err = send_message(sfd, &msg, 20);
 476
 477         /* check the reply */
 478         if (!err && wait_reply(sfd, 20) != 0)
 479                 err = -1;
 480
 481         close(sfd);
 482         return err;
 483 }
 484
 485 /* using takeover operation for grow purposes, mdadm has to be sure
 486  * that mdmon processes all updates, and if necessary it will be closed
 487  * at takeover to raid0 operation
 488   */
 489 void flush_mdmon(char *container)
 490 {
 491         ping_manager(container);
 492         ping_monitor(container);
 493 }