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 = sb->s_blocksize;
 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", 0);
 267                 orangefs_features = new_op->downcall.resp.features.features;
 268                 op_release(new_op);
 269         } else {
 270                 orangefs_features = 0;
 271         }
 272
 273         return ret;
 274 }
 275
 276 int fsid_key_table_initialize(void)
 277 {
 278         return 0;
 279 }
 280
 281 void fsid_key_table_finalize(void)
 282 {
 283 }
 284
 285 /* Called whenever the VFS dirties the inode in response to atime updates */
 286 static void orangefs_dirty_inode(struct inode *inode, int flags)
 287 {
 288         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 289
 290         gossip_debug(GOSSIP_SUPER_DEBUG,
 291                      "orangefs_dirty_inode: %pU\n",
 292                      get_khandle_from_ino(inode));
 293         SetAtimeFlag(orangefs_inode);
 294 }
 295
 296 static const struct super_operations orangefs_s_ops = {
 297         .alloc_inode = orangefs_alloc_inode,
 298         .destroy_inode = orangefs_destroy_inode,
 299         .dirty_inode = orangefs_dirty_inode,
 300         .drop_inode = generic_delete_inode,
 301         .statfs = orangefs_statfs,
 302         .remount_fs = orangefs_remount_fs,
 303         .show_options = generic_show_options,
 304 };
 305
 306 static struct dentry *orangefs_fh_to_dentry(struct super_block *sb,
 307                                   struct fid *fid,
 308                                   int fh_len,
 309                                   int fh_type)
 310 {
 311         struct orangefs_object_kref refn;
 312
 313         if (fh_len < 5 || fh_type > 2)
 314                 return NULL;
 315
 316         ORANGEFS_khandle_from(&(refn.khandle), fid->raw, 16);
 317         refn.fs_id = (u32) fid->raw[4];
 318         gossip_debug(GOSSIP_SUPER_DEBUG,
 319                      "fh_to_dentry: handle %pU, fs_id %d\n",
 320                      &refn.khandle,
 321                      refn.fs_id);
 322
 323         return d_obtain_alias(orangefs_iget(sb, &refn));
 324 }
 325
 326 static int orangefs_encode_fh(struct inode *inode,
 327                     __u32 *fh,
 328                     int *max_len,
 329                     struct inode *parent)
 330 {
 331         int len = parent ? 10 : 5;
 332         int type = 1;
 333         struct orangefs_object_kref refn;
 334
 335         if (*max_len < len) {
 336                 gossip_lerr("fh buffer is too small for encoding\n");
 337                 *max_len = len;
 338                 type = 255;
 339                 goto out;
 340         }
 341
 342         refn = ORANGEFS_I(inode)->refn;
 343         ORANGEFS_khandle_to(&refn.khandle, fh, 16);
 344         fh[4] = refn.fs_id;
 345
 346         gossip_debug(GOSSIP_SUPER_DEBUG,
 347                      "Encoding fh: handle %pU, fsid %u\n",
 348                      &refn.khandle,
 349                      refn.fs_id);
 350
 351
 352         if (parent) {
 353                 refn = ORANGEFS_I(parent)->refn;
 354                 ORANGEFS_khandle_to(&refn.khandle, (char *) fh + 20, 16);
 355                 fh[9] = refn.fs_id;
 356
 357                 type = 2;
 358                 gossip_debug(GOSSIP_SUPER_DEBUG,
 359                              "Encoding parent: handle %pU, fsid %u\n",
 360                              &refn.khandle,
 361                              refn.fs_id);
 362         }
 363         *max_len = len;
 364
 365 out:
 366         return type;
 367 }
 368
 369 static const struct export_operations orangefs_export_ops = {
 370         .encode_fh = orangefs_encode_fh,
 371         .fh_to_dentry = orangefs_fh_to_dentry,
 372 };
 373
 374 static int orangefs_fill_sb(struct super_block *sb,
 375                 struct orangefs_fs_mount_response *fs_mount,
 376                 void *data, int silent)
 377 {
 378         int ret = -EINVAL;
 379         struct inode *root = NULL;
 380         struct dentry *root_dentry = NULL;
 381         struct orangefs_object_kref root_object;
 382
 383         /* alloc and init our private orangefs sb info */
 384         sb->s_fs_info = kzalloc(sizeof(struct orangefs_sb_info_s), GFP_KERNEL);
 385         if (!ORANGEFS_SB(sb))
 386                 return -ENOMEM;
 387         ORANGEFS_SB(sb)->sb = sb;
 388
 389         ORANGEFS_SB(sb)->root_khandle = fs_mount->root_khandle;
 390         ORANGEFS_SB(sb)->fs_id = fs_mount->fs_id;
 391         ORANGEFS_SB(sb)->id = fs_mount->id;
 392
 393         if (data) {
 394                 ret = parse_mount_options(sb, data, silent);
 395                 if (ret)
 396                         return ret;
 397         }
 398
 399         /* Hang the xattr handlers off the superblock */
 400         sb->s_xattr = orangefs_xattr_handlers;
 401         sb->s_magic = ORANGEFS_SUPER_MAGIC;
 402         sb->s_op = &orangefs_s_ops;
 403         sb->s_d_op = &orangefs_dentry_operations;
 404
 405         sb->s_blocksize = orangefs_bufmap_size_query();
 406         sb->s_blocksize_bits = orangefs_bufmap_shift_query();
 407         sb->s_maxbytes = MAX_LFS_FILESIZE;
 408
 409         root_object.khandle = ORANGEFS_SB(sb)->root_khandle;
 410         root_object.fs_id = ORANGEFS_SB(sb)->fs_id;
 411         gossip_debug(GOSSIP_SUPER_DEBUG,
 412                      "get inode %pU, fsid %d\n",
 413                      &root_object.khandle,
 414                      root_object.fs_id);
 415
 416         root = orangefs_iget(sb, &root_object);
 417         if (IS_ERR(root))
 418                 return PTR_ERR(root);
 419
 420         gossip_debug(GOSSIP_SUPER_DEBUG,
 421                      "Allocated root inode [%p] with mode %x\n",
 422                      root,
 423                      root->i_mode);
 424
 425         /* allocates and places root dentry in dcache */
 426         root_dentry = d_make_root(root);
 427         if (!root_dentry)
 428                 return -ENOMEM;
 429
 430         sb->s_export_op = &orangefs_export_ops;
 431         sb->s_root = root_dentry;
 432         return 0;
 433 }
 434
 435 struct dentry *orangefs_mount(struct file_system_type *fst,
 436                            int flags,
 437                            const char *devname,
 438                            void *data)
 439 {
 440         int ret = -EINVAL;
 441         struct super_block *sb = ERR_PTR(-EINVAL);
 442         struct orangefs_kernel_op_s *new_op;
 443         struct dentry *d = ERR_PTR(-EINVAL);
 444
 445         gossip_debug(GOSSIP_SUPER_DEBUG,
 446                      "orangefs_mount: called with devname %s\n",
 447                      devname);
 448
 449         if (!devname) {
 450                 gossip_err("ERROR: device name not specified.\n");
 451                 return ERR_PTR(-EINVAL);
 452         }
 453
 454         new_op = op_alloc(ORANGEFS_VFS_OP_FS_MOUNT);
 455         if (!new_op)
 456                 return ERR_PTR(-ENOMEM);
 457
 458         strncpy(new_op->upcall.req.fs_mount.orangefs_config_server,
 459                 devname,
 460                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 461
 462         gossip_debug(GOSSIP_SUPER_DEBUG,
 463                      "Attempting ORANGEFS Mount via host %s\n",
 464                      new_op->upcall.req.fs_mount.orangefs_config_server);
 465
 466         ret = service_operation(new_op, "orangefs_mount", 0);
 467         gossip_debug(GOSSIP_SUPER_DEBUG,
 468                      "orangefs_mount: mount got return value of %d\n", ret);
 469         if (ret)
 470                 goto free_op;
 471
 472         if (new_op->downcall.resp.fs_mount.fs_id == ORANGEFS_FS_ID_NULL) {
 473                 gossip_err("ERROR: Retrieved null fs_id\n");
 474                 ret = -EINVAL;
 475                 goto free_op;
 476         }
 477
 478         sb = sget(fst, NULL, set_anon_super, flags, NULL);
 479
 480         if (IS_ERR(sb)) {
 481                 d = ERR_CAST(sb);
 482                 goto free_op;
 483         }
 484
 485         ret = orangefs_fill_sb(sb,
 486               &new_op->downcall.resp.fs_mount, data,
 487               flags & MS_SILENT ? 1 : 0);
 488
 489         if (ret) {
 490                 d = ERR_PTR(ret);
 491                 goto free_op;
 492         }
 493
 494         /*
 495          * on successful mount, store the devname and data
 496          * used
 497          */
 498         strncpy(ORANGEFS_SB(sb)->devname,
 499                 devname,
 500                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 501
 502         /* mount_pending must be cleared */
 503         ORANGEFS_SB(sb)->mount_pending = 0;
 504
 505         /*
 506          * finally, add this sb to our list of known orangefs
 507          * sb's
 508          */
 509         gossip_debug(GOSSIP_SUPER_DEBUG,
 510                      "Adding SB %p to orangefs superblocks\n",
 511                      ORANGEFS_SB(sb));
 512         spin_lock(&orangefs_superblocks_lock);
 513         list_add_tail(&ORANGEFS_SB(sb)->list, &orangefs_superblocks);
 514         spin_unlock(&orangefs_superblocks_lock);
 515         op_release(new_op);
 516
 517         if (orangefs_userspace_version >= 20906) {
 518                 new_op = op_alloc(ORANGEFS_VFS_OP_FEATURES);
 519                 if (!new_op)
 520                         return ERR_PTR(-ENOMEM);
 521                 new_op->upcall.req.features.features = 0;
 522                 ret = service_operation(new_op, "orangefs_features", 0);
 523                 orangefs_features = new_op->downcall.resp.features.features;
 524                 op_release(new_op);
 525         } else {
 526                 orangefs_features = 0;
 527         }
 528
 529         return dget(sb->s_root);
 530
 531 free_op:
 532         gossip_err("orangefs_mount: mount request failed with %d\n", ret);
 533         if (ret == -EINVAL) {
 534                 gossip_err("Ensure that all orangefs-servers have the same FS configuration files\n");
 535                 gossip_err("Look at pvfs2-client-core log file (typically /tmp/pvfs2-client.log) for more details\n");
 536         }
 537
 538         op_release(new_op);
 539
 540         return d;
 541 }
 542
 543 void orangefs_kill_sb(struct super_block *sb)
 544 {
 545         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_kill_sb: called\n");
 546
 547         /* provided sb cleanup */
 548         kill_anon_super(sb);
 549
 550         /*
 551          * issue the unmount to userspace to tell it to remove the
 552          * dynamic mount info it has for this superblock
 553          */
 554          orangefs_unmount_sb(sb);
 555
 556         /* remove the sb from our list of orangefs specific sb's */
 557
 558         spin_lock(&orangefs_superblocks_lock);
 559         __list_del_entry(&ORANGEFS_SB(sb)->list);       /* not list_del_init */
 560         ORANGEFS_SB(sb)->list.prev = NULL;
 561         spin_unlock(&orangefs_superblocks_lock);
 562
 563         /*
 564          * make sure that ORANGEFS_DEV_REMOUNT_ALL loop that might've seen us
 565          * gets completed before we free the dang thing.
 566          */
 567         mutex_lock(&orangefs_request_mutex);
 568         mutex_unlock(&orangefs_request_mutex);
 569
 570         /* free the orangefs superblock private data */
 571         kfree(ORANGEFS_SB(sb));
 572 }
 573
 574 int orangefs_inode_cache_initialize(void)
 575 {
 576         orangefs_inode_cache = kmem_cache_create("orangefs_inode_cache",
 577                                               sizeof(struct orangefs_inode_s),
 578                                               0,
 579                                               ORANGEFS_CACHE_CREATE_FLAGS,
 580                                               orangefs_inode_cache_ctor);
 581
 582         if (!orangefs_inode_cache) {
 583                 gossip_err("Cannot create orangefs_inode_cache\n");
 584                 return -ENOMEM;
 585         }
 586         return 0;
 587 }
 588
 589 int orangefs_inode_cache_finalize(void)
 590 {
 591         kmem_cache_destroy(orangefs_inode_cache);
 592         return 0;
 593 }