fs/orangefs/super.c

   1 /*
   2  * (C) 2001 Clemson University and The University of Chicago
   3  *
   4  * See COPYING in top-level directory.
   5  */
   6
   7 #include "protocol.h"
   8 #include "orangefs-kernel.h"
   9 #include "orangefs-bufmap.h"
  10
  11 #include <linux/parser.h>
  12
  13 /* a cache for orangefs-inode objects (i.e. orangefs inode private data) */
  14 static struct kmem_cache *orangefs_inode_cache;
  15
  16 /* list for storing orangefs specific superblocks in use */
  17 LIST_HEAD(orangefs_superblocks);
  18
  19 DEFINE_SPINLOCK(orangefs_superblocks_lock);
  20
  21 enum {
  22         Opt_intr,
  23         Opt_acl,
  24         Opt_local_lock,
  25
  26         Opt_err
  27 };
  28
  29 static const match_table_t tokens = {
  30         { Opt_acl,              "acl" },
  31         { Opt_intr,             "intr" },
  32         { Opt_local_lock,       "local_lock" },
  33         { Opt_err,      NULL }
  34 };
  35
  36 uint64_t orangefs_features;
  37
  38 static int parse_mount_options(struct super_block *sb, char *options,
  39                 int silent)
  40 {
  41         struct orangefs_sb_info_s *orangefs_sb = ORANGEFS_SB(sb);
  42         substring_t args[MAX_OPT_ARGS];
  43         char *p;
  44
  45         /*
  46          * Force any potential flags that might be set from the mount
  47          * to zero, ie, initialize to unset.
  48          */
  49         sb->s_flags &= ~MS_POSIXACL;
  50         orangefs_sb->flags &= ~ORANGEFS_OPT_INTR;
  51         orangefs_sb->flags &= ~ORANGEFS_OPT_LOCAL_LOCK;
  52
  53         while ((p = strsep(&options, ",")) != NULL) {
  54                 int token;
  55
  56                 if (!*p)
  57                         continue;
  58
  59                 token = match_token(p, tokens, args);
  60                 switch (token) {
  61                 case Opt_acl:
  62                         sb->s_flags |= MS_POSIXACL;
  63                         break;
  64                 case Opt_intr:
  65                         orangefs_sb->flags |= ORANGEFS_OPT_INTR;
  66                         break;
  67                 case Opt_local_lock:
  68                         orangefs_sb->flags |= ORANGEFS_OPT_LOCAL_LOCK;
  69                         break;
  70                 default:
  71                         goto fail;
  72                 }
  73         }
  74
  75         return 0;
  76 fail:
  77         if (!silent)
  78                 gossip_err("Error: mount option [%s] is not supported.\n", p);
  79         return -EINVAL;
  80 }
  81
  82 static void orangefs_inode_cache_ctor(void *req)
  83 {
  84         struct orangefs_inode_s *orangefs_inode = req;
  85
  86         inode_init_once(&orangefs_inode->vfs_inode);
  87         init_rwsem(&orangefs_inode->xattr_sem);
  88
  89         orangefs_inode->vfs_inode.i_version = 1;
  90 }
  91
  92 static struct inode *orangefs_alloc_inode(struct super_block *sb)
  93 {
  94         struct orangefs_inode_s *orangefs_inode;
  95
  96         orangefs_inode = kmem_cache_alloc(orangefs_inode_cache, GFP_KERNEL);
  97         if (orangefs_inode == NULL) {
  98                 gossip_err("Failed to allocate orangefs_inode\n");
  99                 return NULL;
 100         }
 101
 102         /*
 103          * We want to clear everything except for rw_semaphore and the
 104          * vfs_inode.
 105          */
 106         memset(&orangefs_inode->refn.khandle, 0, 16);
 107         orangefs_inode->refn.fs_id = ORANGEFS_FS_ID_NULL;
 108         orangefs_inode->last_failed_block_index_read = 0;
 109         memset(orangefs_inode->link_target, 0, sizeof(orangefs_inode->link_target));
 110         orangefs_inode->pinode_flags = 0;
 111
 112         gossip_debug(GOSSIP_SUPER_DEBUG,
 113                      "orangefs_alloc_inode: allocated %p\n",
 114                      &orangefs_inode->vfs_inode);
 115         return &orangefs_inode->vfs_inode;
 116 }
 117
 118 static void orangefs_i_callback(struct rcu_head *head)
 119 {
 120         struct inode *inode = container_of(head, struct inode, i_rcu);
 121         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 122         kmem_cache_free(orangefs_inode_cache, orangefs_inode);
 123 }
 124
 125 static void orangefs_destroy_inode(struct inode *inode)
 126 {
 127         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 128
 129         gossip_debug(GOSSIP_SUPER_DEBUG,
 130                         "%s: deallocated %p destroying inode %pU\n",
 131                         __func__, orangefs_inode, get_khandle_from_ino(inode));
 132
 133         call_rcu(&inode->i_rcu, orangefs_i_callback);
 134 }
 135
 136 /*
 137  * NOTE: information filled in here is typically reflected in the
 138  * output of the system command 'df'
 139 */
 140 static int orangefs_statfs(struct dentry *dentry, struct kstatfs *buf)
 141 {
 142         int ret = -ENOMEM;
 143         struct orangefs_kernel_op_s *new_op = NULL;
 144         int flags = 0;
 145         struct super_block *sb = NULL;
 146
 147         sb = dentry->d_sb;
 148
 149         gossip_debug(GOSSIP_SUPER_DEBUG,
 150                      "orangefs_statfs: called on sb %p (fs_id is %d)\n",
 151                      sb,
 152                      (int)(ORANGEFS_SB(sb)->fs_id));
 153
 154         new_op = op_alloc(ORANGEFS_VFS_OP_STATFS);
 155         if (!new_op)
 156                 return ret;
 157         new_op->upcall.req.statfs.fs_id = ORANGEFS_SB(sb)->fs_id;
 158
 159         if (ORANGEFS_SB(sb)->flags & ORANGEFS_OPT_INTR)
 160                 flags = ORANGEFS_OP_INTERRUPTIBLE;
 161
 162         ret = service_operation(new_op, "orangefs_statfs", flags);
 163
 164         if (new_op->downcall.status < 0)
 165                 goto out_op_release;
 166
 167         gossip_debug(GOSSIP_SUPER_DEBUG,
 168                      "%s: got %ld blocks available | "
 169                      "%ld blocks total | %ld block size | "
 170                      "%ld files total | %ld files avail\n",
 171                      __func__,
 172                      (long)new_op->downcall.resp.statfs.blocks_avail,
 173                      (long)new_op->downcall.resp.statfs.blocks_total,
 174                      (long)new_op->downcall.resp.statfs.block_size,
 175                      (long)new_op->downcall.resp.statfs.files_total,
 176                      (long)new_op->downcall.resp.statfs.files_avail);
 177
 178         buf->f_type = sb->s_magic;
 179         memcpy(&buf->f_fsid, &ORANGEFS_SB(sb)->fs_id, sizeof(buf->f_fsid));
 180         buf->f_bsize = new_op->downcall.resp.statfs.block_size;
 181         buf->f_namelen = ORANGEFS_NAME_MAX;
 182
 183         buf->f_blocks = (sector_t) new_op->downcall.resp.statfs.blocks_total;
 184         buf->f_bfree = (sector_t) new_op->downcall.resp.statfs.blocks_avail;
 185         buf->f_bavail = (sector_t) new_op->downcall.resp.statfs.blocks_avail;
 186         buf->f_files = (sector_t) new_op->downcall.resp.statfs.files_total;
 187         buf->f_ffree = (sector_t) new_op->downcall.resp.statfs.files_avail;
 188         buf->f_frsize = 0;
 189
 190 out_op_release:
 191         op_release(new_op);
 192         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_statfs: returning %d\n", ret);
 193         return ret;
 194 }
 195
 196 /*
 197  * Remount as initiated by VFS layer.  We just need to reparse the mount
 198  * options, no need to signal pvfs2-client-core about it.
 199  */
 200 static int orangefs_remount_fs(struct super_block *sb, int *flags, char *data)
 201 {
 202         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_remount_fs: called\n");
 203         return parse_mount_options(sb, data, 1);
 204 }
 205
 206 /*
 207  * Remount as initiated by pvfs2-client-core on restart.  This is used to
 208  * repopulate mount information left from previous pvfs2-client-core.
 209  *
 210  * the idea here is that given a valid superblock, we're
 211  * re-initializing the user space client with the initial mount
 212  * information specified when the super block was first initialized.
 213  * this is very different than the first initialization/creation of a
 214  * superblock.  we use the special service_priority_operation to make
 215  * sure that the mount gets ahead of any other pending operation that
 216  * is waiting for servicing.  this means that the pvfs2-client won't
 217  * fail to start several times for all other pending operations before
 218  * the client regains all of the mount information from us.
 219  * NOTE: this function assumes that the request_mutex is already acquired!
 220  */
 221 int orangefs_remount(struct orangefs_sb_info_s *orangefs_sb)
 222 {
 223         struct orangefs_kernel_op_s *new_op;
 224         int ret = -EINVAL;
 225
 226         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_remount: called\n");
 227
 228         new_op = op_alloc(ORANGEFS_VFS_OP_FS_MOUNT);
 229         if (!new_op)
 230                 return -ENOMEM;
 231         strncpy(new_op->upcall.req.fs_mount.orangefs_config_server,
 232                 orangefs_sb->devname,
 233                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 234
 235         gossip_debug(GOSSIP_SUPER_DEBUG,
 236                      "Attempting ORANGEFS Remount via host %s\n",
 237                      new_op->upcall.req.fs_mount.orangefs_config_server);
 238
 239         /*
 240          * we assume that the calling function has already acquired the
 241          * request_mutex to prevent other operations from bypassing
 242          * this one
 243          */
 244         ret = service_operation(new_op, "orangefs_remount",
 245                 ORANGEFS_OP_PRIORITY | ORANGEFS_OP_NO_MUTEX);
 246         gossip_debug(GOSSIP_SUPER_DEBUG,
 247                      "orangefs_remount: mount got return value of %d\n",
 248                      ret);
 249         if (ret == 0) {
 250                 /*
 251                  * store the id assigned to this sb -- it's just a
 252                  * short-lived mapping that the system interface uses
 253                  * to map this superblock to a particular mount entry
 254                  */
 255                 orangefs_sb->id = new_op->downcall.resp.fs_mount.id;
 256                 orangefs_sb->mount_pending = 0;
 257         }
 258
 259         op_release(new_op);
 260
 261         if (orangefs_userspace_version >= 20906) {
 262                 new_op = op_alloc(ORANGEFS_VFS_OP_FEATURES);
 263                 if (!new_op)
 264                         return -ENOMEM;
 265                 new_op->upcall.req.features.features = 0;
 266                 ret = service_operation(new_op, "orangefs_features",
 267                     ORANGEFS_OP_PRIORITY | ORANGEFS_OP_NO_MUTEX);
 268                 if (!ret)
 269                         orangefs_features =
 270                             new_op->downcall.resp.features.features;
 271                 else
 272                         orangefs_features = 0;
 273                 op_release(new_op);
 274         } else {
 275                 orangefs_features = 0;
 276         }
 277
 278         return ret;
 279 }
 280
 281 int fsid_key_table_initialize(void)
 282 {
 283         return 0;
 284 }
 285
 286 void fsid_key_table_finalize(void)
 287 {
 288 }
 289
 290 /* Called whenever the VFS dirties the inode in response to atime updates */
 291 static void orangefs_dirty_inode(struct inode *inode, int flags)
 292 {
 293         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 294
 295         gossip_debug(GOSSIP_SUPER_DEBUG,
 296                      "orangefs_dirty_inode: %pU\n",
 297                      get_khandle_from_ino(inode));
 298         SetAtimeFlag(orangefs_inode);
 299 }
 300
 301 static const struct super_operations orangefs_s_ops = {
 302         .alloc_inode = orangefs_alloc_inode,
 303         .destroy_inode = orangefs_destroy_inode,
 304         .dirty_inode = orangefs_dirty_inode,
 305         .drop_inode = generic_delete_inode,
 306         .statfs = orangefs_statfs,
 307         .remount_fs = orangefs_remount_fs,
 308         .show_options = generic_show_options,
 309 };
 310
 311 static struct dentry *orangefs_fh_to_dentry(struct super_block *sb,
 312                                   struct fid *fid,
 313                                   int fh_len,
 314                                   int fh_type)
 315 {
 316         struct orangefs_object_kref refn;
 317
 318         if (fh_len < 5 || fh_type > 2)
 319                 return NULL;
 320
 321         ORANGEFS_khandle_from(&(refn.khandle), fid->raw, 16);
 322         refn.fs_id = (u32) fid->raw[4];
 323         gossip_debug(GOSSIP_SUPER_DEBUG,
 324                      "fh_to_dentry: handle %pU, fs_id %d\n",
 325                      &refn.khandle,
 326                      refn.fs_id);
 327
 328         return d_obtain_alias(orangefs_iget(sb, &refn));
 329 }
 330
 331 static int orangefs_encode_fh(struct inode *inode,
 332                     __u32 *fh,
 333                     int *max_len,
 334                     struct inode *parent)
 335 {
 336         int len = parent ? 10 : 5;
 337         int type = 1;
 338         struct orangefs_object_kref refn;
 339
 340         if (*max_len < len) {
 341                 gossip_lerr("fh buffer is too small for encoding\n");
 342                 *max_len = len;
 343                 type = 255;
 344                 goto out;
 345         }
 346
 347         refn = ORANGEFS_I(inode)->refn;
 348         ORANGEFS_khandle_to(&refn.khandle, fh, 16);
 349         fh[4] = refn.fs_id;
 350
 351         gossip_debug(GOSSIP_SUPER_DEBUG,
 352                      "Encoding fh: handle %pU, fsid %u\n",
 353                      &refn.khandle,
 354                      refn.fs_id);
 355
 356
 357         if (parent) {
 358                 refn = ORANGEFS_I(parent)->refn;
 359                 ORANGEFS_khandle_to(&refn.khandle, (char *) fh + 20, 16);
 360                 fh[9] = refn.fs_id;
 361
 362                 type = 2;
 363                 gossip_debug(GOSSIP_SUPER_DEBUG,
 364                              "Encoding parent: handle %pU, fsid %u\n",
 365                              &refn.khandle,
 366                              refn.fs_id);
 367         }
 368         *max_len = len;
 369
 370 out:
 371         return type;
 372 }
 373
 374 static const struct export_operations orangefs_export_ops = {
 375         .encode_fh = orangefs_encode_fh,
 376         .fh_to_dentry = orangefs_fh_to_dentry,
 377 };
 378
 379 static int orangefs_fill_sb(struct super_block *sb,
 380                 struct orangefs_fs_mount_response *fs_mount,
 381                 void *data, int silent)
 382 {
 383         int ret = -EINVAL;
 384         struct inode *root = NULL;
 385         struct dentry *root_dentry = NULL;
 386         struct orangefs_object_kref root_object;
 387
 388         /* alloc and init our private orangefs sb info */
 389         sb->s_fs_info = kzalloc(sizeof(struct orangefs_sb_info_s), GFP_KERNEL);
 390         if (!ORANGEFS_SB(sb))
 391                 return -ENOMEM;
 392         ORANGEFS_SB(sb)->sb = sb;
 393
 394         ORANGEFS_SB(sb)->root_khandle = fs_mount->root_khandle;
 395         ORANGEFS_SB(sb)->fs_id = fs_mount->fs_id;
 396         ORANGEFS_SB(sb)->id = fs_mount->id;
 397
 398         if (data) {
 399                 ret = parse_mount_options(sb, data, silent);
 400                 if (ret)
 401                         return ret;
 402         }
 403
 404         /* Hang the xattr handlers off the superblock */
 405         sb->s_xattr = orangefs_xattr_handlers;
 406         sb->s_magic = ORANGEFS_SUPER_MAGIC;
 407         sb->s_op = &orangefs_s_ops;
 408         sb->s_d_op = &orangefs_dentry_operations;
 409
 410         sb->s_blocksize = orangefs_bufmap_size_query();
 411         sb->s_blocksize_bits = orangefs_bufmap_shift_query();
 412         sb->s_maxbytes = MAX_LFS_FILESIZE;
 413
 414         root_object.khandle = ORANGEFS_SB(sb)->root_khandle;
 415         root_object.fs_id = ORANGEFS_SB(sb)->fs_id;
 416         gossip_debug(GOSSIP_SUPER_DEBUG,
 417                      "get inode %pU, fsid %d\n",
 418                      &root_object.khandle,
 419                      root_object.fs_id);
 420
 421         root = orangefs_iget(sb, &root_object);
 422         if (IS_ERR(root))
 423                 return PTR_ERR(root);
 424
 425         gossip_debug(GOSSIP_SUPER_DEBUG,
 426                      "Allocated root inode [%p] with mode %x\n",
 427                      root,
 428                      root->i_mode);
 429
 430         /* allocates and places root dentry in dcache */
 431         root_dentry = d_make_root(root);
 432         if (!root_dentry)
 433                 return -ENOMEM;
 434
 435         sb->s_export_op = &orangefs_export_ops;
 436         sb->s_root = root_dentry;
 437         return 0;
 438 }
 439
 440 struct dentry *orangefs_mount(struct file_system_type *fst,
 441                            int flags,
 442                            const char *devname,
 443                            void *data)
 444 {
 445         int ret = -EINVAL;
 446         struct super_block *sb = ERR_PTR(-EINVAL);
 447         struct orangefs_kernel_op_s *new_op;
 448         struct dentry *d = ERR_PTR(-EINVAL);
 449
 450         gossip_debug(GOSSIP_SUPER_DEBUG,
 451                      "orangefs_mount: called with devname %s\n",
 452                      devname);
 453
 454         if (!devname) {
 455                 gossip_err("ERROR: device name not specified.\n");
 456                 return ERR_PTR(-EINVAL);
 457         }
 458
 459         new_op = op_alloc(ORANGEFS_VFS_OP_FS_MOUNT);
 460         if (!new_op)
 461                 return ERR_PTR(-ENOMEM);
 462
 463         strncpy(new_op->upcall.req.fs_mount.orangefs_config_server,
 464                 devname,
 465                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 466
 467         gossip_debug(GOSSIP_SUPER_DEBUG,
 468                      "Attempting ORANGEFS Mount via host %s\n",
 469                      new_op->upcall.req.fs_mount.orangefs_config_server);
 470
 471         ret = service_operation(new_op, "orangefs_mount", 0);
 472         gossip_debug(GOSSIP_SUPER_DEBUG,
 473                      "orangefs_mount: mount got return value of %d\n", ret);
 474         if (ret)
 475                 goto free_op;
 476
 477         if (new_op->downcall.resp.fs_mount.fs_id == ORANGEFS_FS_ID_NULL) {
 478                 gossip_err("ERROR: Retrieved null fs_id\n");
 479                 ret = -EINVAL;
 480                 goto free_op;
 481         }
 482
 483         sb = sget(fst, NULL, set_anon_super, flags, NULL);
 484
 485         if (IS_ERR(sb)) {
 486                 d = ERR_CAST(sb);
 487                 goto free_op;
 488         }
 489
 490         ret = orangefs_fill_sb(sb,
 491               &new_op->downcall.resp.fs_mount, data,
 492               flags & MS_SILENT ? 1 : 0);
 493
 494         if (ret) {
 495                 d = ERR_PTR(ret);
 496                 goto free_sb_and_op;
 497         }
 498
 499         /*
 500          * on successful mount, store the devname and data
 501          * used
 502          */
 503         strncpy(ORANGEFS_SB(sb)->devname,
 504                 devname,
 505                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 506
 507         /* mount_pending must be cleared */
 508         ORANGEFS_SB(sb)->mount_pending = 0;
 509
 510         /*
 511          * finally, add this sb to our list of known orangefs
 512          * sb's
 513          */
 514         gossip_debug(GOSSIP_SUPER_DEBUG,
 515                      "Adding SB %p to orangefs superblocks\n",
 516                      ORANGEFS_SB(sb));
 517         spin_lock(&orangefs_superblocks_lock);
 518         list_add_tail(&ORANGEFS_SB(sb)->list, &orangefs_superblocks);
 519         spin_unlock(&orangefs_superblocks_lock);
 520         op_release(new_op);
 521
 522         /* Must be removed from the list now. */
 523         ORANGEFS_SB(sb)->no_list = 0;
 524
 525         if (orangefs_userspace_version >= 20906) {
 526                 new_op = op_alloc(ORANGEFS_VFS_OP_FEATURES);
 527                 if (!new_op)
 528                         return ERR_PTR(-ENOMEM);
 529                 new_op->upcall.req.features.features = 0;
 530                 ret = service_operation(new_op, "orangefs_features", 0);
 531                 orangefs_features = new_op->downcall.resp.features.features;
 532                 op_release(new_op);
 533         } else {
 534                 orangefs_features = 0;
 535         }
 536
 537         return dget(sb->s_root);
 538
 539 free_sb_and_op:
 540         /* Will call orangefs_kill_sb with sb not in list. */
 541         ORANGEFS_SB(sb)->no_list = 1;
 542         deactivate_locked_super(sb);
 543 free_op:
 544         gossip_err("orangefs_mount: mount request failed with %d\n", ret);
 545         if (ret == -EINVAL) {
 546                 gossip_err("Ensure that all orangefs-servers have the same FS configuration files\n");
 547                 gossip_err("Look at pvfs2-client-core log file (typically /tmp/pvfs2-client.log) for more details\n");
 548         }
 549
 550         op_release(new_op);
 551
 552         return d;
 553 }
 554
 555 void orangefs_kill_sb(struct super_block *sb)
 556 {
 557         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_kill_sb: called\n");
 558
 559         /* provided sb cleanup */
 560         kill_anon_super(sb);
 561
 562         if (!ORANGEFS_SB(sb)) {
 563                 mutex_lock(&orangefs_request_mutex);
 564                 mutex_unlock(&orangefs_request_mutex);
 565                 return;
 566         }
 567         /*
 568          * issue the unmount to userspace to tell it to remove the
 569          * dynamic mount info it has for this superblock
 570          */
 571          orangefs_unmount_sb(sb);
 572
 573         if (!ORANGEFS_SB(sb)->no_list) {
 574                 /* remove the sb from our list of orangefs specific sb's */
 575                 spin_lock(&orangefs_superblocks_lock);
 576                 /* not list_del_init */
 577                 __list_del_entry(&ORANGEFS_SB(sb)->list);
 578                 ORANGEFS_SB(sb)->list.prev = NULL;
 579                 spin_unlock(&orangefs_superblocks_lock);
 580         }
 581
 582         /*
 583          * make sure that ORANGEFS_DEV_REMOUNT_ALL loop that might've seen us
 584          * gets completed before we free the dang thing.
 585          */
 586         mutex_lock(&orangefs_request_mutex);
 587         mutex_unlock(&orangefs_request_mutex);
 588
 589         /* free the orangefs superblock private data */
 590         kfree(ORANGEFS_SB(sb));
 591 }
 592
 593 int orangefs_inode_cache_initialize(void)
 594 {
 595         orangefs_inode_cache = kmem_cache_create("orangefs_inode_cache",
 596                                               sizeof(struct orangefs_inode_s),
 597                                               0,
 598                                               ORANGEFS_CACHE_CREATE_FLAGS,
 599                                               orangefs_inode_cache_ctor);
 600
 601         if (!orangefs_inode_cache) {
 602                 gossip_err("Cannot create orangefs_inode_cache\n");
 603                 return -ENOMEM;
 604         }
 605         return 0;
 606 }
 607
 608 int orangefs_inode_cache_finalize(void)
 609 {
 610         kmem_cache_destroy(orangefs_inode_cache);
 611         return 0;
 612 }