fs/orangefs/dir.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 /*
  12  * decode routine used by kmod to deal with the blob sent from
  13  * userspace for readdirs. The blob contains zero or more of these
  14  * sub-blobs:
  15  *   __u32 - represents length of the character string that follows.
  16  *   string - between 1 and ORANGEFS_NAME_MAX bytes long.
  17  *   padding - (if needed) to cause the __u32 plus the string to be
  18  *             eight byte aligned.
  19  *   khandle - sizeof(khandle) bytes.
  20  */
  21 static long decode_dirents(char *ptr, size_t size,
  22                            struct orangefs_readdir_response_s *readdir)
  23 {
  24         int i;
  25         struct orangefs_readdir_response_s *rd =
  26                 (struct orangefs_readdir_response_s *) ptr;
  27         char *buf = ptr;
  28         int khandle_size = sizeof(struct orangefs_khandle);
  29         size_t offset = offsetof(struct orangefs_readdir_response_s,
  30                                 dirent_array);
  31         /* 8 reflects eight byte alignment */
  32         int smallest_blob = khandle_size + 8;
  33         __u32 len;
  34         int aligned_len;
  35         int sizeof_u32 = sizeof(__u32);
  36         long ret;
  37
  38         gossip_debug(GOSSIP_DIR_DEBUG, "%s: size:%zu:\n", __func__, size);
  39
  40         /* size is = offset on empty dirs, > offset on non-empty dirs... */
  41         if (size < offset) {
  42                 gossip_err("%s: size:%zu: offset:%zu:\n",
  43                            __func__,
  44                            size,
  45                            offset);
  46                 ret = -EINVAL;
  47                 goto out;
  48         }
  49
  50         if ((size == offset) && (readdir->orangefs_dirent_outcount != 0)) {
  51                 gossip_err("%s: size:%zu: dirent_outcount:%d:\n",
  52                            __func__,
  53                            size,
  54                            readdir->orangefs_dirent_outcount);
  55                 ret = -EINVAL;
  56                 goto out;
  57         }
  58
  59         readdir->token = rd->token;
  60         readdir->orangefs_dirent_outcount = rd->orangefs_dirent_outcount;
  61         readdir->dirent_array = kcalloc(readdir->orangefs_dirent_outcount,
  62                                         sizeof(*readdir->dirent_array),
  63                                         GFP_KERNEL);
  64         if (readdir->dirent_array == NULL) {
  65                 gossip_err("%s: kcalloc failed.\n", __func__);
  66                 ret = -ENOMEM;
  67                 goto out;
  68         }
  69
  70         buf += offset;
  71         size -= offset;
  72
  73         for (i = 0; i < readdir->orangefs_dirent_outcount; i++) {
  74                 if (size < smallest_blob) {
  75                         gossip_err("%s: size:%zu: smallest_blob:%d:\n",
  76                                    __func__,
  77                                    size,
  78                                    smallest_blob);
  79                         ret = -EINVAL;
  80                         goto free;
  81                 }
  82
  83                 len = *(__u32 *)buf;
  84                 if ((len < 1) || (len > ORANGEFS_NAME_MAX)) {
  85                         gossip_err("%s: len:%d:\n", __func__, len);
  86                         ret = -EINVAL;
  87                         goto free;
  88                 }
  89
  90                 gossip_debug(GOSSIP_DIR_DEBUG,
  91                              "%s: size:%zu: len:%d:\n",
  92                              __func__,
  93                              size,
  94                              len);
  95
  96                 readdir->dirent_array[i].d_name = buf + sizeof_u32;
  97                 readdir->dirent_array[i].d_length = len;
  98
  99                 /*
 100                  * Calculate "aligned" length of this string and its
 101                  * associated __u32 descriptor.
 102                  */
 103                 aligned_len = ((sizeof_u32 + len + 1) + 7) & ~7;
 104                 gossip_debug(GOSSIP_DIR_DEBUG,
 105                              "%s: aligned_len:%d:\n",
 106                              __func__,
 107                              aligned_len);
 108
 109                 /*
 110                  * The end of the blob should coincide with the end
 111                  * of the last sub-blob.
 112                  */
 113                 if (size < aligned_len + khandle_size) {
 114                         gossip_err("%s: ran off the end of the blob.\n",
 115                                    __func__);
 116                         ret = -EINVAL;
 117                         goto free;
 118                 }
 119                 size -= aligned_len + khandle_size;
 120
 121                 buf += aligned_len;
 122
 123                 readdir->dirent_array[i].khandle =
 124                         *(struct orangefs_khandle *) buf;
 125                 buf += khandle_size;
 126         }
 127         ret = buf - ptr;
 128         gossip_debug(GOSSIP_DIR_DEBUG, "%s: returning:%ld:\n", __func__, ret);
 129         goto out;
 130
 131 free:
 132         kfree(readdir->dirent_array);
 133         readdir->dirent_array = NULL;
 134
 135 out:
 136         return ret;
 137 }
 138
 139 /*
 140  * Read directory entries from an instance of an open directory.
 141  */
 142 static int orangefs_readdir(struct file *file, struct dir_context *ctx)
 143 {
 144         int ret = 0;
 145         int buffer_index;
 146         /*
 147          * ptoken supports Orangefs' distributed directory logic, added
 148          * in 2.9.2.
 149          */
 150         __u64 *ptoken = file->private_data;
 151         __u64 pos = 0;
 152         ino_t ino = 0;
 153         struct dentry *dentry = file->f_path.dentry;
 154         struct orangefs_kernel_op_s *new_op = NULL;
 155         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(dentry->d_inode);
 156         struct orangefs_readdir_response_s readdir_response;
 157         void *dents_buf;
 158         int i = 0;
 159         int len = 0;
 160         ino_t current_ino = 0;
 161         char *current_entry = NULL;
 162         long bytes_decoded;
 163
 164         gossip_debug(GOSSIP_DIR_DEBUG,
 165                      "%s: ctx->pos:%lld, ptoken = %llu\n",
 166                      __func__,
 167                      lld(ctx->pos),
 168                      llu(*ptoken));
 169
 170         pos = (__u64) ctx->pos;
 171
 172         /* are we done? */
 173         if (pos == ORANGEFS_READDIR_END) {
 174                 gossip_debug(GOSSIP_DIR_DEBUG,
 175                              "Skipping to termination path\n");
 176                 return 0;
 177         }
 178
 179         gossip_debug(GOSSIP_DIR_DEBUG,
 180                      "orangefs_readdir called on %s (pos=%llu)\n",
 181                      dentry->d_name.name, llu(pos));
 182
 183         memset(&readdir_response, 0, sizeof(readdir_response));
 184
 185         new_op = op_alloc(ORANGEFS_VFS_OP_READDIR);
 186         if (!new_op)
 187                 return -ENOMEM;
 188
 189         /*
 190          * Only the indices are shared. No memory is actually shared, but the
 191          * mechanism is used.
 192          */
 193         new_op->uses_shared_memory = 1;
 194         new_op->upcall.req.readdir.refn = orangefs_inode->refn;
 195         new_op->upcall.req.readdir.max_dirent_count =
 196             ORANGEFS_MAX_DIRENT_COUNT_READDIR;
 197
 198         gossip_debug(GOSSIP_DIR_DEBUG,
 199                      "%s: upcall.req.readdir.refn.khandle: %pU\n",
 200                      __func__,
 201                      &new_op->upcall.req.readdir.refn.khandle);
 202
 203         new_op->upcall.req.readdir.token = *ptoken;
 204
 205 get_new_buffer_index:
 206         buffer_index = orangefs_readdir_index_get();
 207         if (buffer_index < 0) {
 208                 ret = buffer_index;
 209                 gossip_lerr("orangefs_readdir: orangefs_readdir_index_get() failure (%d)\n",
 210                             ret);
 211                 goto out_free_op;
 212         }
 213         new_op->upcall.req.readdir.buf_index = buffer_index;
 214
 215         ret = service_operation(new_op,
 216                                 "orangefs_readdir",
 217                                 get_interruptible_flag(dentry->d_inode));
 218
 219         gossip_debug(GOSSIP_DIR_DEBUG,
 220                      "Readdir downcall status is %d.  ret:%d\n",
 221                      new_op->downcall.status,
 222                      ret);
 223
 224         orangefs_readdir_index_put(buffer_index);
 225
 226         if (ret == -EAGAIN && op_state_purged(new_op)) {
 227                 /* Client-core indices are invalid after it restarted. */
 228                 gossip_debug(GOSSIP_DIR_DEBUG,
 229                         "%s: Getting new buffer_index for retry of readdir..\n",
 230                          __func__);
 231                 goto get_new_buffer_index;
 232         }
 233
 234         if (ret == -EIO && op_state_purged(new_op)) {
 235                 gossip_err("%s: Client is down. Aborting readdir call.\n",
 236                         __func__);
 237                 goto out_free_op;
 238         }
 239
 240         if (ret < 0 || new_op->downcall.status != 0) {
 241                 gossip_debug(GOSSIP_DIR_DEBUG,
 242                              "Readdir request failed.  Status:%d\n",
 243                              new_op->downcall.status);
 244                 if (ret >= 0)
 245                         ret = new_op->downcall.status;
 246                 goto out_free_op;
 247         }
 248
 249         dents_buf = new_op->downcall.trailer_buf;
 250         if (dents_buf == NULL) {
 251                 gossip_err("Invalid NULL buffer in readdir response\n");
 252                 ret = -ENOMEM;
 253                 goto out_free_op;
 254         }
 255
 256         bytes_decoded = decode_dirents(dents_buf, new_op->downcall.trailer_size,
 257                                         &readdir_response);
 258         if (bytes_decoded < 0) {
 259                 ret = bytes_decoded;
 260                 gossip_err("Could not decode readdir from buffer %d\n", ret);
 261                 goto out_vfree;
 262         }
 263
 264         if (bytes_decoded != new_op->downcall.trailer_size) {
 265                 gossip_err("orangefs_readdir: # bytes decoded (%ld) "
 266                            "!= trailer size (%ld)\n",
 267                            bytes_decoded,
 268                            (long)new_op->downcall.trailer_size);
 269                 ret = -EINVAL;
 270                 goto out_destroy_handle;
 271         }
 272
 273         /*
 274          *  orangefs doesn't actually store dot and dot-dot, but
 275          *  we need to have them represented.
 276          */
 277         if (pos == 0) {
 278                 ino = get_ino_from_khandle(dentry->d_inode);
 279                 gossip_debug(GOSSIP_DIR_DEBUG,
 280                              "%s: calling dir_emit of \".\" with pos = %llu\n",
 281                              __func__,
 282                              llu(pos));
 283                 ret = dir_emit(ctx, ".", 1, ino, DT_DIR);
 284                 pos += 1;
 285         }
 286
 287         if (pos == 1) {
 288                 ino = get_parent_ino_from_dentry(dentry);
 289                 gossip_debug(GOSSIP_DIR_DEBUG,
 290                              "%s: calling dir_emit of \"..\" with pos = %llu\n",
 291                              __func__,
 292                              llu(pos));
 293                 ret = dir_emit(ctx, "..", 2, ino, DT_DIR);
 294                 pos += 1;
 295         }
 296
 297         /*
 298          * we stored ORANGEFS_ITERATE_NEXT in ctx->pos last time around
 299          * to prevent "finding" dot and dot-dot on any iteration
 300          * other than the first.
 301          */
 302         if (ctx->pos == ORANGEFS_ITERATE_NEXT)
 303                 ctx->pos = 0;
 304
 305         gossip_debug(GOSSIP_DIR_DEBUG,
 306                      "%s: dirent_outcount:%d:\n",
 307                      __func__,
 308                      readdir_response.orangefs_dirent_outcount);
 309         for (i = ctx->pos;
 310              i < readdir_response.orangefs_dirent_outcount;
 311              i++) {
 312                 len = readdir_response.dirent_array[i].d_length;
 313                 current_entry = readdir_response.dirent_array[i].d_name;
 314                 current_ino = orangefs_khandle_to_ino(
 315                         &readdir_response.dirent_array[i].khandle);
 316
 317                 gossip_debug(GOSSIP_DIR_DEBUG,
 318                              "calling dir_emit for %s with len %d"
 319                              ", ctx->pos %ld\n",
 320                              current_entry,
 321                              len,
 322                              (unsigned long)ctx->pos);
 323                 /*
 324                  * type is unknown. We don't return object type
 325                  * in the dirent_array. This leaves getdents
 326                  * clueless about type.
 327                  */
 328                 ret =
 329                     dir_emit(ctx, current_entry, len, current_ino, DT_UNKNOWN);
 330                 if (!ret)
 331                         break;
 332                 ctx->pos++;
 333                 gossip_debug(GOSSIP_DIR_DEBUG,
 334                               "%s: ctx->pos:%lld\n",
 335                               __func__,
 336                               lld(ctx->pos));
 337
 338         }
 339
 340         /*
 341          * we ran all the way through the last batch, set up for
 342          * getting another batch...
 343          */
 344         if (ret) {
 345                 *ptoken = readdir_response.token;
 346                 ctx->pos = ORANGEFS_ITERATE_NEXT;
 347         }
 348
 349         /*
 350          * Did we hit the end of the directory?
 351          */
 352         if (readdir_response.token == ORANGEFS_READDIR_END) {
 353                 gossip_debug(GOSSIP_DIR_DEBUG,
 354                 "End of dir detected; setting ctx->pos to ORANGEFS_READDIR_END.\n");
 355                 ctx->pos = ORANGEFS_READDIR_END;
 356         }
 357
 358 out_destroy_handle:
 359         /* kfree(NULL) is safe */
 360         kfree(readdir_response.dirent_array);
 361 out_vfree:
 362         gossip_debug(GOSSIP_DIR_DEBUG, "vfree %p\n", dents_buf);
 363         vfree(dents_buf);
 364 out_free_op:
 365         op_release(new_op);
 366         gossip_debug(GOSSIP_DIR_DEBUG, "orangefs_readdir returning %d\n", ret);
 367         return ret;
 368 }
 369
 370 static int orangefs_dir_open(struct inode *inode, struct file *file)
 371 {
 372         __u64 *ptoken;
 373
 374         file->private_data = kmalloc(sizeof(__u64), GFP_KERNEL);
 375         if (!file->private_data)
 376                 return -ENOMEM;
 377
 378         ptoken = file->private_data;
 379         *ptoken = ORANGEFS_READDIR_START;
 380         return 0;
 381 }
 382
 383 static int orangefs_dir_release(struct inode *inode, struct file *file)
 384 {
 385         orangefs_flush_inode(inode);
 386         kfree(file->private_data);
 387         return 0;
 388 }
 389
 390 /** ORANGEFS implementation of VFS directory operations */
 391 const struct file_operations orangefs_dir_operations = {
 392         .read = generic_read_dir,
 393         .iterate = orangefs_readdir,
 394         .open = orangefs_dir_open,
 395         .release = orangefs_dir_release,
 396 };