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_destroy_inode(struct inode *inode)
 119 {
 120         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 121
 122         gossip_debug(GOSSIP_SUPER_DEBUG,
 123                         "%s: deallocated %p destroying inode %pU\n",
 124                         __func__, orangefs_inode, get_khandle_from_ino(inode));
 125
 126         kmem_cache_free(orangefs_inode_cache, orangefs_inode);
 127 }
 128
 129 /*
 130  * NOTE: information filled in here is typically reflected in the
 131  * output of the system command 'df'
 132 */
 133 static int orangefs_statfs(struct dentry *dentry, struct kstatfs *buf)
 134 {
 135         int ret = -ENOMEM;
 136         struct orangefs_kernel_op_s *new_op = NULL;
 137         int flags = 0;
 138         struct super_block *sb = NULL;
 139
 140         sb = dentry->d_sb;
 141
 142         gossip_debug(GOSSIP_SUPER_DEBUG,
 143                      "orangefs_statfs: called on sb %p (fs_id is %d)\n",
 144                      sb,
 145                      (int)(ORANGEFS_SB(sb)->fs_id));
 146
 147         new_op = op_alloc(ORANGEFS_VFS_OP_STATFS);
 148         if (!new_op)
 149                 return ret;
 150         new_op->upcall.req.statfs.fs_id = ORANGEFS_SB(sb)->fs_id;
 151
 152         if (ORANGEFS_SB(sb)->flags & ORANGEFS_OPT_INTR)
 153                 flags = ORANGEFS_OP_INTERRUPTIBLE;
 154
 155         ret = service_operation(new_op, "orangefs_statfs", flags);
 156
 157         if (new_op->downcall.status < 0)
 158                 goto out_op_release;
 159
 160         gossip_debug(GOSSIP_SUPER_DEBUG,
 161                      "%s: got %ld blocks available | "
 162                      "%ld blocks total | %ld block size | "
 163                      "%ld files total | %ld files avail\n",
 164                      __func__,
 165                      (long)new_op->downcall.resp.statfs.blocks_avail,
 166                      (long)new_op->downcall.resp.statfs.blocks_total,
 167                      (long)new_op->downcall.resp.statfs.block_size,
 168                      (long)new_op->downcall.resp.statfs.files_total,
 169                      (long)new_op->downcall.resp.statfs.files_avail);
 170
 171         buf->f_type = sb->s_magic;
 172         memcpy(&buf->f_fsid, &ORANGEFS_SB(sb)->fs_id, sizeof(buf->f_fsid));
 173         buf->f_bsize = new_op->downcall.resp.statfs.block_size;
 174         buf->f_namelen = ORANGEFS_NAME_MAX;
 175
 176         buf->f_blocks = (sector_t) new_op->downcall.resp.statfs.blocks_total;
 177         buf->f_bfree = (sector_t) new_op->downcall.resp.statfs.blocks_avail;
 178         buf->f_bavail = (sector_t) new_op->downcall.resp.statfs.blocks_avail;
 179         buf->f_files = (sector_t) new_op->downcall.resp.statfs.files_total;
 180         buf->f_ffree = (sector_t) new_op->downcall.resp.statfs.files_avail;
 181         buf->f_frsize = sb->s_blocksize;
 182
 183 out_op_release:
 184         op_release(new_op);
 185         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_statfs: returning %d\n", ret);
 186         return ret;
 187 }
 188
 189 /*
 190  * Remount as initiated by VFS layer.  We just need to reparse the mount
 191  * options, no need to signal pvfs2-client-core about it.
 192  */
 193 static int orangefs_remount_fs(struct super_block *sb, int *flags, char *data)
 194 {
 195         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_remount_fs: called\n");
 196         return parse_mount_options(sb, data, 1);
 197 }
 198
 199 /*
 200  * Remount as initiated by pvfs2-client-core on restart.  This is used to
 201  * repopulate mount information left from previous pvfs2-client-core.
 202  *
 203  * the idea here is that given a valid superblock, we're
 204  * re-initializing the user space client with the initial mount
 205  * information specified when the super block was first initialized.
 206  * this is very different than the first initialization/creation of a
 207  * superblock.  we use the special service_priority_operation to make
 208  * sure that the mount gets ahead of any other pending operation that
 209  * is waiting for servicing.  this means that the pvfs2-client won't
 210  * fail to start several times for all other pending operations before
 211  * the client regains all of the mount information from us.
 212  * NOTE: this function assumes that the request_mutex is already acquired!
 213  */
 214 int orangefs_remount(struct orangefs_sb_info_s *orangefs_sb)
 215 {
 216         struct orangefs_kernel_op_s *new_op;
 217         int ret = -EINVAL;
 218
 219         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_remount: called\n");
 220
 221         new_op = op_alloc(ORANGEFS_VFS_OP_FS_MOUNT);
 222         if (!new_op)
 223                 return -ENOMEM;
 224         strncpy(new_op->upcall.req.fs_mount.orangefs_config_server,
 225                 orangefs_sb->devname,
 226                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 227
 228         gossip_debug(GOSSIP_SUPER_DEBUG,
 229                      "Attempting ORANGEFS Remount via host %s\n",
 230                      new_op->upcall.req.fs_mount.orangefs_config_server);
 231
 232         /*
 233          * we assume that the calling function has already acquired the
 234          * request_mutex to prevent other operations from bypassing
 235          * this one
 236          */
 237         ret = service_operation(new_op, "orangefs_remount",
 238                 ORANGEFS_OP_PRIORITY | ORANGEFS_OP_NO_MUTEX);
 239         gossip_debug(GOSSIP_SUPER_DEBUG,
 240                      "orangefs_remount: mount got return value of %d\n",
 241                      ret);
 242         if (ret == 0) {
 243                 /*
 244                  * store the id assigned to this sb -- it's just a
 245                  * short-lived mapping that the system interface uses
 246                  * to map this superblock to a particular mount entry
 247                  */
 248                 orangefs_sb->id = new_op->downcall.resp.fs_mount.id;
 249                 orangefs_sb->mount_pending = 0;
 250         }
 251
 252         op_release(new_op);
 253
 254         if (orangefs_userspace_version >= 20906) {
 255                 new_op = op_alloc(ORANGEFS_VFS_OP_FEATURES);
 256                 if (!new_op)
 257                         return -ENOMEM;
 258                 new_op->upcall.req.features.features = 0;
 259                 ret = service_operation(new_op, "orangefs_features", 0);
 260                 orangefs_features = new_op->downcall.resp.features.features;
 261                 op_release(new_op);
 262         } else {
 263                 orangefs_features = 0;
 264         }
 265
 266         return ret;
 267 }
 268
 269 int fsid_key_table_initialize(void)
 270 {
 271         return 0;
 272 }
 273
 274 void fsid_key_table_finalize(void)
 275 {
 276 }
 277
 278 /* Called whenever the VFS dirties the inode in response to atime updates */
 279 static void orangefs_dirty_inode(struct inode *inode, int flags)
 280 {
 281         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 282
 283         gossip_debug(GOSSIP_SUPER_DEBUG,
 284                      "orangefs_dirty_inode: %pU\n",
 285                      get_khandle_from_ino(inode));
 286         SetAtimeFlag(orangefs_inode);
 287 }
 288
 289 static const struct super_operations orangefs_s_ops = {
 290         .alloc_inode = orangefs_alloc_inode,
 291         .destroy_inode = orangefs_destroy_inode,
 292         .dirty_inode = orangefs_dirty_inode,
 293         .drop_inode = generic_delete_inode,
 294         .statfs = orangefs_statfs,
 295         .remount_fs = orangefs_remount_fs,
 296         .show_options = generic_show_options,
 297 };
 298
 299 static struct dentry *orangefs_fh_to_dentry(struct super_block *sb,
 300                                   struct fid *fid,
 301                                   int fh_len,
 302                                   int fh_type)
 303 {
 304         struct orangefs_object_kref refn;
 305
 306         if (fh_len < 5 || fh_type > 2)
 307                 return NULL;
 308
 309         ORANGEFS_khandle_from(&(refn.khandle), fid->raw, 16);
 310         refn.fs_id = (u32) fid->raw[4];
 311         gossip_debug(GOSSIP_SUPER_DEBUG,
 312                      "fh_to_dentry: handle %pU, fs_id %d\n",
 313                      &refn.khandle,
 314                      refn.fs_id);
 315
 316         return d_obtain_alias(orangefs_iget(sb, &refn));
 317 }
 318
 319 static int orangefs_encode_fh(struct inode *inode,
 320                     __u32 *fh,
 321                     int *max_len,
 322                     struct inode *parent)
 323 {
 324         int len = parent ? 10 : 5;
 325         int type = 1;
 326         struct orangefs_object_kref refn;
 327
 328         if (*max_len < len) {
 329                 gossip_lerr("fh buffer is too small for encoding\n");
 330                 *max_len = len;
 331                 type = 255;
 332                 goto out;
 333         }
 334
 335         refn = ORANGEFS_I(inode)->refn;
 336         ORANGEFS_khandle_to(&refn.khandle, fh, 16);
 337         fh[4] = refn.fs_id;
 338
 339         gossip_debug(GOSSIP_SUPER_DEBUG,
 340                      "Encoding fh: handle %pU, fsid %u\n",
 341                      &refn.khandle,
 342                      refn.fs_id);
 343
 344
 345         if (parent) {
 346                 refn = ORANGEFS_I(parent)->refn;
 347                 ORANGEFS_khandle_to(&refn.khandle, (char *) fh + 20, 16);
 348                 fh[9] = refn.fs_id;
 349
 350                 type = 2;
 351                 gossip_debug(GOSSIP_SUPER_DEBUG,
 352                              "Encoding parent: handle %pU, fsid %u\n",
 353                              &refn.khandle,
 354                              refn.fs_id);
 355         }
 356         *max_len = len;
 357
 358 out:
 359         return type;
 360 }
 361
 362 static const struct export_operations orangefs_export_ops = {
 363         .encode_fh = orangefs_encode_fh,
 364         .fh_to_dentry = orangefs_fh_to_dentry,
 365 };
 366
 367 static int orangefs_fill_sb(struct super_block *sb,
 368                 struct orangefs_fs_mount_response *fs_mount,
 369                 void *data, int silent)
 370 {
 371         int ret = -EINVAL;
 372         struct inode *root = NULL;
 373         struct dentry *root_dentry = NULL;
 374         struct orangefs_object_kref root_object;
 375
 376         /* alloc and init our private orangefs sb info */
 377         sb->s_fs_info = kzalloc(sizeof(struct orangefs_sb_info_s), GFP_KERNEL);
 378         if (!ORANGEFS_SB(sb))
 379                 return -ENOMEM;
 380         ORANGEFS_SB(sb)->sb = sb;
 381
 382         ORANGEFS_SB(sb)->root_khandle = fs_mount->root_khandle;
 383         ORANGEFS_SB(sb)->fs_id = fs_mount->fs_id;
 384         ORANGEFS_SB(sb)->id = fs_mount->id;
 385
 386         if (data) {
 387                 ret = parse_mount_options(sb, data, silent);
 388                 if (ret)
 389                         return ret;
 390         }
 391
 392         /* Hang the xattr handlers off the superblock */
 393         sb->s_xattr = orangefs_xattr_handlers;
 394         sb->s_magic = ORANGEFS_SUPER_MAGIC;
 395         sb->s_op = &orangefs_s_ops;
 396         sb->s_d_op = &orangefs_dentry_operations;
 397
 398         sb->s_blocksize = orangefs_bufmap_size_query();
 399         sb->s_blocksize_bits = orangefs_bufmap_shift_query();
 400         sb->s_maxbytes = MAX_LFS_FILESIZE;
 401
 402         root_object.khandle = ORANGEFS_SB(sb)->root_khandle;
 403         root_object.fs_id = ORANGEFS_SB(sb)->fs_id;
 404         gossip_debug(GOSSIP_SUPER_DEBUG,
 405                      "get inode %pU, fsid %d\n",
 406                      &root_object.khandle,
 407                      root_object.fs_id);
 408
 409         root = orangefs_iget(sb, &root_object);
 410         if (IS_ERR(root))
 411                 return PTR_ERR(root);
 412
 413         gossip_debug(GOSSIP_SUPER_DEBUG,
 414                      "Allocated root inode [%p] with mode %x\n",
 415                      root,
 416                      root->i_mode);
 417
 418         /* allocates and places root dentry in dcache */
 419         root_dentry = d_make_root(root);
 420         if (!root_dentry)
 421                 return -ENOMEM;
 422
 423         sb->s_export_op = &orangefs_export_ops;
 424         sb->s_root = root_dentry;
 425         return 0;
 426 }
 427
 428 struct dentry *orangefs_mount(struct file_system_type *fst,
 429                            int flags,
 430                            const char *devname,
 431                            void *data)
 432 {
 433         int ret = -EINVAL;
 434         struct super_block *sb = ERR_PTR(-EINVAL);
 435         struct orangefs_kernel_op_s *new_op;
 436         struct dentry *d = ERR_PTR(-EINVAL);
 437
 438         gossip_debug(GOSSIP_SUPER_DEBUG,
 439                      "orangefs_mount: called with devname %s\n",
 440                      devname);
 441
 442         if (!devname) {
 443                 gossip_err("ERROR: device name not specified.\n");
 444                 return ERR_PTR(-EINVAL);
 445         }
 446
 447         new_op = op_alloc(ORANGEFS_VFS_OP_FS_MOUNT);
 448         if (!new_op)
 449                 return ERR_PTR(-ENOMEM);
 450
 451         strncpy(new_op->upcall.req.fs_mount.orangefs_config_server,
 452                 devname,
 453                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 454
 455         gossip_debug(GOSSIP_SUPER_DEBUG,
 456                      "Attempting ORANGEFS Mount via host %s\n",
 457                      new_op->upcall.req.fs_mount.orangefs_config_server);
 458
 459         ret = service_operation(new_op, "orangefs_mount", 0);
 460         gossip_debug(GOSSIP_SUPER_DEBUG,
 461                      "orangefs_mount: mount got return value of %d\n", ret);
 462         if (ret)
 463                 goto free_op;
 464
 465         if (new_op->downcall.resp.fs_mount.fs_id == ORANGEFS_FS_ID_NULL) {
 466                 gossip_err("ERROR: Retrieved null fs_id\n");
 467                 ret = -EINVAL;
 468                 goto free_op;
 469         }
 470
 471         sb = sget(fst, NULL, set_anon_super, flags, NULL);
 472
 473         if (IS_ERR(sb)) {
 474                 d = ERR_CAST(sb);
 475                 goto free_op;
 476         }
 477
 478         ret = orangefs_fill_sb(sb,
 479               &new_op->downcall.resp.fs_mount, data,
 480               flags & MS_SILENT ? 1 : 0);
 481
 482         if (ret) {
 483                 d = ERR_PTR(ret);
 484                 goto free_op;
 485         }
 486
 487         /*
 488          * on successful mount, store the devname and data
 489          * used
 490          */
 491         strncpy(ORANGEFS_SB(sb)->devname,
 492                 devname,
 493                 ORANGEFS_MAX_SERVER_ADDR_LEN);
 494
 495         /* mount_pending must be cleared */
 496         ORANGEFS_SB(sb)->mount_pending = 0;
 497
 498         /*
 499          * finally, add this sb to our list of known orangefs
 500          * sb's
 501          */
 502         gossip_debug(GOSSIP_SUPER_DEBUG,
 503                      "Adding SB %p to orangefs superblocks\n",
 504                      ORANGEFS_SB(sb));
 505         spin_lock(&orangefs_superblocks_lock);
 506         list_add_tail(&ORANGEFS_SB(sb)->list, &orangefs_superblocks);
 507         spin_unlock(&orangefs_superblocks_lock);
 508         op_release(new_op);
 509
 510         if (orangefs_userspace_version >= 20906) {
 511                 new_op = op_alloc(ORANGEFS_VFS_OP_FEATURES);
 512                 if (!new_op)
 513                         return ERR_PTR(-ENOMEM);
 514                 new_op->upcall.req.features.features = 0;
 515                 ret = service_operation(new_op, "orangefs_features", 0);
 516                 orangefs_features = new_op->downcall.resp.features.features;
 517                 op_release(new_op);
 518         } else {
 519                 orangefs_features = 0;
 520         }
 521
 522         return dget(sb->s_root);
 523
 524 free_op:
 525         gossip_err("orangefs_mount: mount request failed with %d\n", ret);
 526         if (ret == -EINVAL) {
 527                 gossip_err("Ensure that all orangefs-servers have the same FS configuration files\n");
 528                 gossip_err("Look at pvfs2-client-core log file (typically /tmp/pvfs2-client.log) for more details\n");
 529         }
 530
 531         op_release(new_op);
 532
 533         return d;
 534 }
 535
 536 void orangefs_kill_sb(struct super_block *sb)
 537 {
 538         gossip_debug(GOSSIP_SUPER_DEBUG, "orangefs_kill_sb: called\n");
 539
 540         /* provided sb cleanup */
 541         kill_anon_super(sb);
 542
 543         /*
 544          * issue the unmount to userspace to tell it to remove the
 545          * dynamic mount info it has for this superblock
 546          */
 547          orangefs_unmount_sb(sb);
 548
 549         /* remove the sb from our list of orangefs specific sb's */
 550
 551         spin_lock(&orangefs_superblocks_lock);
 552         __list_del_entry(&ORANGEFS_SB(sb)->list);       /* not list_del_init */
 553         ORANGEFS_SB(sb)->list.prev = NULL;
 554         spin_unlock(&orangefs_superblocks_lock);
 555
 556         /*
 557          * make sure that ORANGEFS_DEV_REMOUNT_ALL loop that might've seen us
 558          * gets completed before we free the dang thing.
 559          */
 560         mutex_lock(&orangefs_request_mutex);
 561         mutex_unlock(&orangefs_request_mutex);
 562
 563         /* free the orangefs superblock private data */
 564         kfree(ORANGEFS_SB(sb));
 565 }
 566
 567 int orangefs_inode_cache_initialize(void)
 568 {
 569         orangefs_inode_cache = kmem_cache_create("orangefs_inode_cache",
 570                                               sizeof(struct orangefs_inode_s),
 571                                               0,
 572                                               ORANGEFS_CACHE_CREATE_FLAGS,
 573                                               orangefs_inode_cache_ctor);
 574
 575         if (!orangefs_inode_cache) {
 576                 gossip_err("Cannot create orangefs_inode_cache\n");
 577                 return -ENOMEM;
 578         }
 579         return 0;
 580 }
 581
 582 int orangefs_inode_cache_finalize(void)
 583 {
 584         kmem_cache_destroy(orangefs_inode_cache);
 585         return 0;
 586 }