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 int send_message(int fd, struct metadata_update *msg, int tmo)
  82 {
  83         __s32 len = msg->len;
  84         int rv;
  85
  86         rv = send_buf(fd, &start_magic, 4, tmo);
  87         rv = rv ?: send_buf(fd, &len, 4, tmo);
  88         if (len > 0)
  89                 rv = rv ?: send_buf(fd, msg->buf, msg->len, tmo);
  90         rv = send_buf(fd, &end_magic, 4, tmo);
  91
  92         return rv;
  93 }
  94
  95 int receive_message(int fd, struct metadata_update *msg, int tmo)
  96 {
  97         __u32 magic;
  98         __s32 len;
  99         int rv;
 100
 101         rv = recv_buf(fd, &magic, 4, tmo);
 102         if (rv < 0 || magic != start_magic)
 103                 return -1;
 104         rv = recv_buf(fd, &len, 4, tmo);
 105         if (rv < 0 || len > MSG_MAX_LEN)
 106                 return -1;
 107         if (len > 0) {
 108                 msg->buf = xmalloc(len);
 109                 rv = recv_buf(fd, msg->buf, len, tmo);
 110                 if (rv < 0) {
 111                         free(msg->buf);
 112                         return -1;
 113                 }
 114         } else
 115                 msg->buf = NULL;
 116         rv = recv_buf(fd, &magic, 4, tmo);
 117         if (rv < 0 || magic != end_magic) {
 118                 free(msg->buf);
 119                 return -1;
 120         }
 121         msg->len = len;
 122         return 0;
 123 }
 124
 125 int ack(int fd, int tmo)
 126 {
 127         struct metadata_update msg = { .len = 0 };
 128
 129         return send_message(fd, &msg, tmo);
 130 }
 131
 132 int wait_reply(int fd, int tmo)
 133 {
 134         struct metadata_update msg;
 135         int err = receive_message(fd, &msg, tmo);
 136
 137         /* mdmon sent extra data, but caller only cares that we got a
 138          * successful reply
 139          */
 140         if (err == 0 && msg.len > 0)
 141                 free(msg.buf);
 142
 143         return err;
 144 }
 145
 146 int connect_monitor(char *devname)
 147 {
 148         char path[100];
 149         int sfd;
 150         long fl;
 151         struct sockaddr_un addr;
 152         int pos;
 153         char *c;
 154
 155         pos = sprintf(path, "%s/", MDMON_DIR);
 156         if (is_subarray(devname)) {
 157                 devname++;
 158                 c = strchr(devname, '/');
 159                 if (!c)
 160                         return -1;
 161                 snprintf(&path[pos], c - devname + 1, "%s", devname);
 162                 pos += c - devname;
 163         } else
 164                 pos += sprintf(&path[pos], "%s", devname);
 165         sprintf(&path[pos], ".sock");
 166
 167         sfd = socket(PF_LOCAL, SOCK_STREAM, 0);
 168         if (sfd < 0)
 169                 return -1;
 170
 171         addr.sun_family = PF_LOCAL;
 172         strcpy(addr.sun_path, path);
 173         if (connect(sfd, (struct sockaddr*)&addr, sizeof(addr)) < 0) {
 174                 close(sfd);
 175                 return -1;
 176         }
 177
 178         fl = fcntl(sfd, F_GETFL, 0);
 179         fl |= O_NONBLOCK;
 180         fcntl(sfd, F_SETFL, fl);
 181
 182         return sfd;
 183 }
 184
 185 int fping_monitor(int sfd)
 186 {
 187         int err = 0;
 188
 189         if (sfd < 0)
 190                 return sfd;
 191
 192         /* try to ping existing socket */
 193         if (ack(sfd, 20) != 0)
 194                 err = -1;
 195
 196         /* check the reply */
 197         if (!err && wait_reply(sfd, 20) != 0)
 198                 err = -1;
 199
 200         return err;
 201 }
 202
 203 /* give the monitor a chance to update the metadata */
 204 int ping_monitor(char *devname)
 205 {
 206         int sfd = connect_monitor(devname);
 207         int err;
 208
 209         if (sfd >= 0) {
 210                 err = fping_monitor(sfd);
 211                 close(sfd);
 212         } else
 213                 err = -1;
 214
 215         return err;
 216 }
 217
 218 static char *ping_monitor_version(char *devname)
 219 {
 220         int sfd = connect_monitor(devname);
 221         struct metadata_update msg;
 222         int err = 0;
 223
 224         if (sfd < 0)
 225                 return NULL;
 226
 227         if (ack(sfd, 20) != 0)
 228                 err = -1;
 229
 230         if (!err && receive_message(sfd, &msg, 20) != 0)
 231                 err = -1;
 232
 233         close(sfd);
 234
 235         if (err || !msg.len || !msg.buf)
 236                 return NULL;
 237         return msg.buf;
 238 }
 239
 240 int unblock_subarray(struct mdinfo *sra, const int unfreeze)
 241 {
 242         char buf[64];
 243         int rc = 0;
 244
 245         if (sra) {
 246                 sprintf(buf, "external:%s\n", sra->text_version);
 247                 buf[9] = '/';
 248         } else
 249                 buf[9] = '-';
 250
 251         if (buf[9] == '-' ||
 252             sysfs_set_str(sra, NULL, "metadata_version", buf) ||
 253             (unfreeze &&
 254              sysfs_attribute_available(sra, NULL, "sync_action") &&
 255              sysfs_set_str(sra, NULL, "sync_action", "idle")))
 256                 rc = -1;
 257         return rc;
 258 }
 259
 260 int block_subarray(struct mdinfo *sra)
 261 {
 262         char buf[64];
 263         int rc = 0;
 264
 265         sprintf(buf, "external:%s\n", sra->text_version);
 266         buf[9] = '-';
 267         if (sysfs_set_str(sra, NULL, "metadata_version", buf))
 268                 rc = -1;
 269
 270         return rc;
 271 }
 272
 273 /* check mdmon version if it supports
 274  * array blocking mechanism
 275  */
 276 int check_mdmon_version(char *container)
 277 {
 278         char *version = NULL;
 279
 280         if (!mdmon_running(container)) {
 281                 /* if mdmon is not active we assume that any instance that is
 282                  * later started will match the current mdadm version, if this
 283                  * assumption is violated we may inadvertantly rebuild an array
 284                  * that was meant for reshape, or start rebuild on a spare that
 285                  * was to be moved to another container
 286                  */
 287                 /* pass */;
 288         } else {
 289                 int ver;
 290
 291                 version = ping_monitor_version(container);
 292                 ver = version ? mdadm_version(version) : -1;
 293                 free(version);
 294                 if (ver < 3002000) {
 295                         pr_err("mdmon instance for %s cannot be disabled\n",
 296                                container);
 297                         return -1;
 298                 }
 299         }
 300
 301         return 0;
 302 }
 303
 304 /**
 305  * block_monitor - prevent mdmon spare assignment
 306  * @container - container to block
 307  * @freeze - flag to additionally freeze sync_action
 308  *
 309  * This is used by the reshape code to freeze the container, and the
 310  * auto-rebuild implementation to atomically move spares.
 311  * In both cases we need to stop mdmon from assigning spares to replace
 312  * failed devices as we might have other plans for the spare.
 313  * For the reshape case we also need to 'freeze' sync_action so that
 314  * no recovery happens until we have fully prepared for the reshape.
 315  *
 316  * We tell mdmon that the array is frozen by marking the 'metadata' name
 317  * with a leading '-'.  The previously told mdmon "Don't make this array
 318  * read/write, leave it readonly".  Now it means a more general "Don't
 319  * reconfigure this array at all".
 320  * As older versions of mdmon (which might run from initrd) don't understand
 321  * this, we first check that the running mdmon is new enough.
 322  */
 323 int block_monitor(char *container, const int freeze)
 324 {
 325         struct mdstat_ent *ent, *e, *e2;
 326         struct mdinfo *sra = NULL;
 327         char buf[SYSFS_MAX_BUF_SIZE];
 328         int rv = 0;
 329
 330         if (check_mdmon_version(container))
 331                 return -1;
 332
 333         ent = mdstat_read(0, 0);
 334         if (!ent) {
 335                 pr_err("failed to read /proc/mdstat while disabling mdmon\n");
 336                 return -1;
 337         }
 338
 339         /* freeze container contents */
 340         for (e = ent; e; e = e->next) {
 341                 if (!is_container_member(e, container))
 342                         continue;
 343                 sysfs_free(sra);
 344                 sra = sysfs_read(-1, e->devnm, GET_VERSION);
 345                 if (!sra) {
 346                         pr_err("failed to read sysfs for subarray%s\n",
 347                                to_subarray(e, container));
 348                         break;
 349                 }
 350                 /* can't reshape an array that we can't monitor */
 351                 if (sra->text_version[0] == '-')
 352                         break;
 353
 354                 if (freeze && sysfs_freeze_array(sra) < 1)
 355                         break;
 356                 /* flag this array to not be modified by mdmon (close race with
 357                  * takeover in reshape case and spare reassignment in the
 358                  * auto-rebuild case)
 359                  */
 360                 if (block_subarray(sra))
 361                         break;
 362                 ping_monitor(container);
 363
 364                 /* check that we did not race with recovery */
 365                 if ((freeze &&
 366                      !sysfs_attribute_available(sra, NULL, "sync_action")) ||
 367                     (freeze &&
 368                      sysfs_attribute_available(sra, NULL, "sync_action") &&
 369                      sysfs_get_str(sra, NULL, "sync_action", buf, sizeof(buf)) > 0 &&
 370                      strcmp(buf, "frozen\n") == 0))
 371                         /* pass */;
 372                 else {
 373                         unblock_subarray(sra, 0);
 374                         break;
 375                 }
 376                 /* Double check against races - there should be no spares
 377                  * or part-spares
 378                  */
 379                 sysfs_free(sra);
 380                 sra = sysfs_read(-1, e->devnm, GET_DEVS | GET_STATE);
 381                 if (sra && sra->array.spare_disks > 0) {
 382                         unblock_subarray(sra, freeze);
 383                         break;
 384                 }
 385         }
 386
 387         if (e) {
 388                 pr_err("failed to freeze subarray%s\n",
 389                         to_subarray(e, container));
 390
 391                 /* thaw the partially frozen container */
 392                 for (e2 = ent; e2 && e2 != e; e2 = e2->next) {
 393                         if (!is_container_member(e2, container))
 394                                 continue;
 395                         sysfs_free(sra);
 396                         sra = sysfs_read(-1, e2->devnm, GET_VERSION);
 397                         if (unblock_subarray(sra, freeze))
 398                                 pr_err("Failed to unfreeze %s\n", e2->devnm);
 399                 }
 400
 401                 ping_monitor(container); /* cleared frozen */
 402                 rv = -1;
 403         }
 404
 405         sysfs_free(sra);
 406         free_mdstat(ent);
 407
 408         return rv;
 409 }
 410
 411 void unblock_monitor(char *container, const int unfreeze)
 412 {
 413         struct mdstat_ent *ent, *e;
 414         struct mdinfo *sra = NULL;
 415         int to_ping = 0;
 416
 417         ent = mdstat_read(0, 0);
 418         if (!ent) {
 419                 pr_err("failed to read /proc/mdstat while unblocking container\n");
 420                 return;
 421         }
 422
 423         /* unfreeze container contents */
 424         for (e = ent; e; e = e->next) {
 425                 if (!is_container_member(e, container))
 426                         continue;
 427                 sysfs_free(sra);
 428                 sra = sysfs_read(-1, e->devnm, GET_VERSION|GET_LEVEL);
 429                 if (!sra)
 430                         continue;
 431                 if (sra->array.level > 0)
 432                         to_ping++;
 433                 if (unblock_subarray(sra, unfreeze))
 434                         pr_err("Failed to unfreeze %s\n", e->devnm);
 435         }
 436         if (to_ping)
 437                 ping_monitor(container);
 438
 439         sysfs_free(sra);
 440         free_mdstat(ent);
 441 }
 442
 443 /* give the manager a chance to view the updated container state.  This
 444  * would naturally happen due to the manager noticing a change in
 445  * /proc/mdstat; however, pinging encourages this detection to happen
 446  * while an exclusive open() on the container is active
 447  */
 448 int ping_manager(char *devname)
 449 {
 450         int sfd = connect_monitor(devname);
 451         struct metadata_update msg = { .len = -1 };
 452         int err = 0;
 453
 454         if (sfd < 0)
 455                 return sfd;
 456
 457         err = send_message(sfd, &msg, 20);
 458
 459         /* check the reply */
 460         if (!err && wait_reply(sfd, 20) != 0)
 461                 err = -1;
 462
 463         close(sfd);
 464         return err;
 465 }
 466
 467 /* using takeover operation for grow purposes, mdadm has to be sure
 468  * that mdmon processes all updates, and if necessary it will be closed
 469  * at takeover to raid0 operation
 470   */
 471 void flush_mdmon(char *container)
 472 {
 473         ping_manager(container);
 474         ping_monitor(container);
 475 }