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