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