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                 pr_err("Error connecting monitor with %s: %s\n",
 175                         addr.sun_path, strerror(errno));
 176                 close(sfd);
 177                 return -1;
 178         }
 179
 180         fl = fcntl(sfd, F_GETFL, 0);
 181         fl |= O_NONBLOCK;
 182         fcntl(sfd, F_SETFL, fl);
 183
 184         return sfd;
 185 }
 186
 187 int fping_monitor(int sfd)
 188 {
 189         int err = 0;
 190
 191         if (sfd < 0)
 192                 return sfd;
 193
 194         /* try to ping existing socket */
 195         if (ack(sfd, 20) != 0)
 196                 err = -1;
 197
 198         /* check the reply */
 199         if (!err && wait_reply(sfd, 20) != 0)
 200                 err = -1;
 201
 202         return err;
 203 }
 204
 205 /* give the monitor a chance to update the metadata */
 206 int ping_monitor(char *devname)
 207 {
 208         int sfd = connect_monitor(devname);
 209         int err;
 210
 211         if (sfd >= 0) {
 212                 err = fping_monitor(sfd);
 213                 close(sfd);
 214         } else
 215                 err = -1;
 216
 217         return err;
 218 }
 219
 220 static char *ping_monitor_version(char *devname)
 221 {
 222         int sfd = connect_monitor(devname);
 223         struct metadata_update msg;
 224         int err = 0;
 225
 226         if (sfd < 0)
 227                 return NULL;
 228
 229         if (ack(sfd, 20) != 0)
 230                 err = -1;
 231
 232         if (!err && receive_message(sfd, &msg, 20) != 0)
 233                 err = -1;
 234
 235         close(sfd);
 236
 237         if (err || !msg.len || !msg.buf)
 238                 return NULL;
 239         return msg.buf;
 240 }
 241
 242 int unblock_subarray(struct mdinfo *sra, const int unfreeze)
 243 {
 244         char buf[64];
 245         int rc = 0;
 246
 247         if (sra) {
 248                 sprintf(buf, "external:%s\n", sra->text_version);
 249                 buf[9] = '/';
 250         } else
 251                 buf[9] = '-';
 252
 253         if (buf[9] == '-' ||
 254             sysfs_set_str(sra, NULL, "metadata_version", buf) ||
 255             (unfreeze &&
 256              sysfs_attribute_available(sra, NULL, "sync_action") &&
 257              sysfs_set_str(sra, NULL, "sync_action", "idle")))
 258                 rc = -1;
 259         return rc;
 260 }
 261
 262 int block_subarray(struct mdinfo *sra)
 263 {
 264         char buf[64];
 265         int rc = 0;
 266
 267         sprintf(buf, "external:%s\n", sra->text_version);
 268         buf[9] = '-';
 269         if (sysfs_set_str(sra, NULL, "metadata_version", buf))
 270                 rc = -1;
 271
 272         return rc;
 273 }
 274
 275 /* check mdmon version if it supports
 276  * array blocking mechanism
 277  */
 278 int check_mdmon_version(char *container)
 279 {
 280         char *version = NULL;
 281
 282         if (!mdmon_running(container)) {
 283                 /* if mdmon is not active we assume that any instance that is
 284                  * later started will match the current mdadm version, if this
 285                  * assumption is violated we may inadvertantly rebuild an array
 286                  * that was meant for reshape, or start rebuild on a spare that
 287                  * was to be moved to another container
 288                  */
 289                 /* pass */;
 290         } else {
 291                 int ver;
 292
 293                 version = ping_monitor_version(container);
 294                 ver = version ? mdadm_version(version) : -1;
 295                 free(version);
 296                 if (ver < 3002000) {
 297                         pr_err("mdmon instance for %s cannot be disabled\n",
 298                                container);
 299                         return -1;
 300                 }
 301         }
 302
 303         return 0;
 304 }
 305
 306 /**
 307  * block_monitor - prevent mdmon spare assignment
 308  * @container - container to block
 309  * @freeze - flag to additionally freeze sync_action
 310  *
 311  * This is used by the reshape code to freeze the container, and the
 312  * auto-rebuild implementation to atomically move spares.
 313  * In both cases we need to stop mdmon from assigning spares to replace
 314  * failed devices as we might have other plans for the spare.
 315  * For the reshape case we also need to 'freeze' sync_action so that
 316  * no recovery happens until we have fully prepared for the reshape.
 317  *
 318  * We tell mdmon that the array is frozen by marking the 'metadata' name
 319  * with a leading '-'.  The previously told mdmon "Don't make this array
 320  * read/write, leave it readonly".  Now it means a more general "Don't
 321  * reconfigure this array at all".
 322  * As older versions of mdmon (which might run from initrd) don't understand
 323  * this, we first check that the running mdmon is new enough.
 324  */
 325 int block_monitor(char *container, const int freeze)
 326 {
 327         struct mdstat_ent *ent, *e, *e2;
 328         struct mdinfo *sra = NULL;
 329         char buf[64];
 330         int rv = 0;
 331
 332         if (check_mdmon_version(container))
 333                 return -1;
 334
 335         ent = mdstat_read(0, 0);
 336         if (!ent) {
 337                 pr_err("failed to read /proc/mdstat while disabling mdmon\n");
 338                 return -1;
 339         }
 340
 341         /* freeze container contents */
 342         for (e = ent; e; e = e->next) {
 343                 if (!is_container_member(e, container))
 344                         continue;
 345                 sysfs_free(sra);
 346                 sra = sysfs_read(-1, e->devnm, GET_VERSION);
 347                 if (!sra) {
 348                         pr_err("failed to read sysfs for subarray%s\n",
 349                                to_subarray(e, container));
 350                         break;
 351                 }
 352                 /* can't reshape an array that we can't monitor */
 353                 if (sra->text_version[0] == '-')
 354                         break;
 355
 356                 if (freeze && sysfs_freeze_array(sra) < 1)
 357                         break;
 358                 /* flag this array to not be modified by mdmon (close race with
 359                  * takeover in reshape case and spare reassignment in the
 360                  * auto-rebuild case)
 361                  */
 362                 if (block_subarray(sra))
 363                         break;
 364                 ping_monitor(container);
 365
 366                 /* check that we did not race with recovery */
 367                 if ((freeze &&
 368                      !sysfs_attribute_available(sra, NULL, "sync_action")) ||
 369                     (freeze &&
 370                      sysfs_attribute_available(sra, NULL, "sync_action") &&
 371                      sysfs_get_str(sra, NULL, "sync_action", buf, 20) > 0 &&
 372                      strcmp(buf, "frozen\n") == 0))
 373                         /* pass */;
 374                 else {
 375                         unblock_subarray(sra, 0);
 376                         break;
 377                 }
 378                 /* Double check against races - there should be no spares
 379                  * or part-spares
 380                  */
 381                 sysfs_free(sra);
 382                 sra = sysfs_read(-1, e->devnm, GET_DEVS | GET_STATE);
 383                 if (sra && sra->array.spare_disks > 0) {
 384                         unblock_subarray(sra, freeze);
 385                         break;
 386                 }
 387         }
 388
 389         if (e) {
 390                 pr_err("failed to freeze subarray%s\n",
 391                         to_subarray(e, container));
 392
 393                 /* thaw the partially frozen container */
 394                 for (e2 = ent; e2 && e2 != e; e2 = e2->next) {
 395                         if (!is_container_member(e2, container))
 396                                 continue;
 397                         sysfs_free(sra);
 398                         sra = sysfs_read(-1, e2->devnm, GET_VERSION);
 399                         if (unblock_subarray(sra, freeze))
 400                                 pr_err("Failed to unfreeze %s\n", e2->devnm);
 401                 }
 402
 403                 ping_monitor(container); /* cleared frozen */
 404                 rv = -1;
 405         }
 406
 407         sysfs_free(sra);
 408         free_mdstat(ent);
 409
 410         return rv;
 411 }
 412
 413 void unblock_monitor(char *container, const int unfreeze)
 414 {
 415         struct mdstat_ent *ent, *e;
 416         struct mdinfo *sra = NULL;
 417         int to_ping = 0;
 418
 419         ent = mdstat_read(0, 0);
 420         if (!ent) {
 421                 pr_err("failed to read /proc/mdstat while unblocking container\n");
 422                 return;
 423         }
 424
 425         /* unfreeze container contents */
 426         for (e = ent; e; e = e->next) {
 427                 if (!is_container_member(e, container))
 428                         continue;
 429                 sysfs_free(sra);
 430                 sra = sysfs_read(-1, e->devnm, GET_VERSION|GET_LEVEL);
 431                 if (!sra)
 432                         continue;
 433                 if (sra->array.level > 0)
 434                         to_ping++;
 435                 if (unblock_subarray(sra, unfreeze))
 436                         pr_err("Failed to unfreeze %s\n", e->devnm);
 437         }
 438         if (to_ping)
 439                 ping_monitor(container);
 440
 441         sysfs_free(sra);
 442         free_mdstat(ent);
 443 }
 444
 445 /* give the manager a chance to view the updated container state.  This
 446  * would naturally happen due to the manager noticing a change in
 447  * /proc/mdstat; however, pinging encourages this detection to happen
 448  * while an exclusive open() on the container is active
 449  */
 450 int ping_manager(char *devname)
 451 {
 452         int sfd = connect_monitor(devname);
 453         struct metadata_update msg = { .len = -1 };
 454         int err = 0;
 455
 456         if (sfd < 0)
 457                 return sfd;
 458
 459         err = send_message(sfd, &msg, 20);
 460
 461         /* check the reply */
 462         if (!err && wait_reply(sfd, 20) != 0)
 463                 err = -1;
 464
 465         close(sfd);
 466         return err;
 467 }
 468
 469 /* using takeover operation for grow purposes, mdadm has to be sure
 470  * that mdmon processes all updates, and if necessary it will be closed
 471  * at takeover to raid0 operation
 472   */
 473 void flush_mdmon(char *container)
 474 {
 475         ping_manager(container);
 476         ping_monitor(container);
 477 }