fs/ceph/dir.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/spinlock.h>
   4 #include <linux/fs_struct.h>
   5 #include <linux/namei.h>
   6 #include <linux/slab.h>
   7 #include <linux/sched.h>
   8
   9 #include "super.h"
  10 #include "mds_client.h"
  11
  12 /*
  13  * Directory operations: readdir, lookup, create, link, unlink,
  14  * rename, etc.
  15  */
  16
  17 /*
  18  * Ceph MDS operations are specified in terms of a base ino and
  19  * relative path.  Thus, the client can specify an operation on a
  20  * specific inode (e.g., a getattr due to fstat(2)), or as a path
  21  * relative to, say, the root directory.
  22  *
  23  * Normally, we limit ourselves to strict inode ops (no path component)
  24  * or dentry operations (a single path component relative to an ino).  The
  25  * exception to this is open_root_dentry(), which will open the mount
  26  * point by name.
  27  */
  28
  29 const struct dentry_operations ceph_dentry_ops;
  30
  31 /*
  32  * Initialize ceph dentry state.
  33  */
  34 int ceph_init_dentry(struct dentry *dentry)
  35 {
  36         struct ceph_dentry_info *di;
  37
  38         if (dentry->d_fsdata)
  39                 return 0;
  40
  41         di = kmem_cache_alloc(ceph_dentry_cachep, GFP_KERNEL | __GFP_ZERO);
  42         if (!di)
  43                 return -ENOMEM;          /* oh well */
  44
  45         spin_lock(&dentry->d_lock);
  46         if (dentry->d_fsdata) {
  47                 /* lost a race */
  48                 kmem_cache_free(ceph_dentry_cachep, di);
  49                 goto out_unlock;
  50         }
  51
  52         if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_NOSNAP)
  53                 d_set_d_op(dentry, &ceph_dentry_ops);
  54         else if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_SNAPDIR)
  55                 d_set_d_op(dentry, &ceph_snapdir_dentry_ops);
  56         else
  57                 d_set_d_op(dentry, &ceph_snap_dentry_ops);
  58
  59         di->dentry = dentry;
  60         di->lease_session = NULL;
  61         dentry->d_time = jiffies;
  62         /* avoid reordering d_fsdata setup so that the check above is safe */
  63         smp_mb();
  64         dentry->d_fsdata = di;
  65         ceph_dentry_lru_add(dentry);
  66 out_unlock:
  67         spin_unlock(&dentry->d_lock);
  68         return 0;
  69 }
  70
  71 struct inode *ceph_get_dentry_parent_inode(struct dentry *dentry)
  72 {
  73         struct inode *inode = NULL;
  74
  75         if (!dentry)
  76                 return NULL;
  77
  78         spin_lock(&dentry->d_lock);
  79         if (!IS_ROOT(dentry)) {
  80                 inode = d_inode(dentry->d_parent);
  81                 ihold(inode);
  82         }
  83         spin_unlock(&dentry->d_lock);
  84         return inode;
  85 }
  86
  87
  88 /*
  89  * for readdir, we encode the directory frag and offset within that
  90  * frag into f_pos.
  91  */
  92 static unsigned fpos_frag(loff_t p)
  93 {
  94         return p >> 32;
  95 }
  96 static unsigned fpos_off(loff_t p)
  97 {
  98         return p & 0xffffffff;
  99 }
 100
 101 static int fpos_cmp(loff_t l, loff_t r)
 102 {
 103         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
 104         if (v)
 105                 return v;
 106         return (int)(fpos_off(l) - fpos_off(r));
 107 }
 108
 109 /*
 110  * make note of the last dentry we read, so we can
 111  * continue at the same lexicographical point,
 112  * regardless of what dir changes take place on the
 113  * server.
 114  */
 115 static int note_last_dentry(struct ceph_file_info *fi, const char *name,
 116                             int len, unsigned next_offset)
 117 {
 118         char *buf = kmalloc(len+1, GFP_KERNEL);
 119         if (!buf)
 120                 return -ENOMEM;
 121         kfree(fi->last_name);
 122         fi->last_name = buf;
 123         memcpy(fi->last_name, name, len);
 124         fi->last_name[len] = 0;
 125         fi->next_offset = next_offset;
 126         dout("note_last_dentry '%s'\n", fi->last_name);
 127         return 0;
 128 }
 129
 130 /*
 131  * When possible, we try to satisfy a readdir by peeking at the
 132  * dcache.  We make this work by carefully ordering dentries on
 133  * d_child when we initially get results back from the MDS, and
 134  * falling back to a "normal" sync readdir if any dentries in the dir
 135  * are dropped.
 136  *
 137  * Complete dir indicates that we have all dentries in the dir.  It is
 138  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 139  * the MDS if/when the directory is modified).
 140  */
 141 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 142                             u32 shared_gen)
 143 {
 144         struct ceph_file_info *fi = file->private_data;
 145         struct dentry *parent = file->f_path.dentry;
 146         struct inode *dir = d_inode(parent);
 147         struct dentry *dentry, *last = NULL;
 148         struct ceph_dentry_info *di;
 149         unsigned nsize = PAGE_CACHE_SIZE / sizeof(struct dentry *);
 150         int err = 0;
 151         loff_t ptr_pos = 0;
 152         struct ceph_readdir_cache_control cache_ctl = {};
 153
 154         dout("__dcache_readdir %p v%u at %llu\n", dir, shared_gen, ctx->pos);
 155
 156         /* we can calculate cache index for the first dirfrag */
 157         if (ceph_frag_is_leftmost(fpos_frag(ctx->pos))) {
 158                 cache_ctl.index = fpos_off(ctx->pos) - 2;
 159                 BUG_ON(cache_ctl.index < 0);
 160                 ptr_pos = cache_ctl.index * sizeof(struct dentry *);
 161         }
 162
 163         while (true) {
 164                 pgoff_t pgoff;
 165                 bool emit_dentry;
 166
 167                 if (ptr_pos >= i_size_read(dir)) {
 168                         fi->flags |= CEPH_F_ATEND;
 169                         err = 0;
 170                         break;
 171                 }
 172
 173                 err = -EAGAIN;
 174                 pgoff = ptr_pos >> PAGE_CACHE_SHIFT;
 175                 if (!cache_ctl.page || pgoff != page_index(cache_ctl.page)) {
 176                         ceph_readdir_cache_release(&cache_ctl);
 177                         cache_ctl.page = find_lock_page(&dir->i_data, pgoff);
 178                         if (!cache_ctl.page) {
 179                                 dout(" page %lu not found\n", pgoff);
 180                                 break;
 181                         }
 182                         /* reading/filling the cache are serialized by
 183                          * i_mutex, no need to use page lock */
 184                         unlock_page(cache_ctl.page);
 185                         cache_ctl.dentries = kmap(cache_ctl.page);
 186                 }
 187
 188                 rcu_read_lock();
 189                 spin_lock(&parent->d_lock);
 190                 /* check i_size again here, because empty directory can be
 191                  * marked as complete while not holding the i_mutex. */
 192                 if (ceph_dir_is_complete_ordered(dir) &&
 193                     ptr_pos < i_size_read(dir))
 194                         dentry = cache_ctl.dentries[cache_ctl.index % nsize];
 195                 else
 196                         dentry = NULL;
 197                 spin_unlock(&parent->d_lock);
 198                 if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 199                         dentry = NULL;
 200                 rcu_read_unlock();
 201                 if (!dentry)
 202                         break;
 203
 204                 emit_dentry = false;
 205                 di = ceph_dentry(dentry);
 206                 spin_lock(&dentry->d_lock);
 207                 if (di->lease_shared_gen == shared_gen &&
 208                     d_really_is_positive(dentry) &&
 209                     ceph_snap(d_inode(dentry)) != CEPH_SNAPDIR &&
 210                     ceph_ino(d_inode(dentry)) != CEPH_INO_CEPH &&
 211                     fpos_cmp(ctx->pos, di->offset) <= 0) {
 212                         emit_dentry = true;
 213                 }
 214                 spin_unlock(&dentry->d_lock);
 215
 216                 if (emit_dentry) {
 217                         dout(" %llu (%llu) dentry %p %pd %p\n", di->offset, ctx->pos,
 218                              dentry, dentry, d_inode(dentry));
 219                         ctx->pos = di->offset;
 220                         if (!dir_emit(ctx, dentry->d_name.name,
 221                                       dentry->d_name.len,
 222                                       ceph_translate_ino(dentry->d_sb,
 223                                                          d_inode(dentry)->i_ino),
 224                                       d_inode(dentry)->i_mode >> 12)) {
 225                                 dput(dentry);
 226                                 err = 0;
 227                                 break;
 228                         }
 229                         ctx->pos++;
 230
 231                         if (last)
 232                                 dput(last);
 233                         last = dentry;
 234                 } else {
 235                         dput(dentry);
 236                 }
 237
 238                 cache_ctl.index++;
 239                 ptr_pos += sizeof(struct dentry *);
 240         }
 241         ceph_readdir_cache_release(&cache_ctl);
 242         if (last) {
 243                 int ret;
 244                 di = ceph_dentry(last);
 245                 ret = note_last_dentry(fi, last->d_name.name, last->d_name.len,
 246                                        fpos_off(di->offset) + 1);
 247                 if (ret < 0)
 248                         err = ret;
 249                 dput(last);
 250         }
 251         return err;
 252 }
 253
 254 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 255 {
 256         struct ceph_file_info *fi = file->private_data;
 257         struct inode *inode = file_inode(file);
 258         struct ceph_inode_info *ci = ceph_inode(inode);
 259         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 260         struct ceph_mds_client *mdsc = fsc->mdsc;
 261         unsigned frag = fpos_frag(ctx->pos);
 262         int off = fpos_off(ctx->pos);
 263         int err;
 264         u32 ftype;
 265         struct ceph_mds_reply_info_parsed *rinfo;
 266
 267         dout("readdir %p file %p frag %u off %u\n", inode, file, frag, off);
 268         if (fi->flags & CEPH_F_ATEND)
 269                 return 0;
 270
 271         /* always start with . and .. */
 272         if (ctx->pos == 0) {
 273                 dout("readdir off 0 -> '.'\n");
 274                 if (!dir_emit(ctx, ".", 1,
 275                             ceph_translate_ino(inode->i_sb, inode->i_ino),
 276                             inode->i_mode >> 12))
 277                         return 0;
 278                 ctx->pos = 1;
 279                 off = 1;
 280         }
 281         if (ctx->pos == 1) {
 282                 ino_t ino = parent_ino(file->f_path.dentry);
 283                 dout("readdir off 1 -> '..'\n");
 284                 if (!dir_emit(ctx, "..", 2,
 285                             ceph_translate_ino(inode->i_sb, ino),
 286                             inode->i_mode >> 12))
 287                         return 0;
 288                 ctx->pos = 2;
 289                 off = 2;
 290         }
 291
 292         /* can we use the dcache? */
 293         spin_lock(&ci->i_ceph_lock);
 294         if (ceph_test_mount_opt(fsc, DCACHE) &&
 295             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 296             ceph_snap(inode) != CEPH_SNAPDIR &&
 297             __ceph_dir_is_complete_ordered(ci) &&
 298             __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
 299                 u32 shared_gen = ci->i_shared_gen;
 300                 spin_unlock(&ci->i_ceph_lock);
 301                 err = __dcache_readdir(file, ctx, shared_gen);
 302                 if (err != -EAGAIN)
 303                         return err;
 304                 frag = fpos_frag(ctx->pos);
 305                 off = fpos_off(ctx->pos);
 306         } else {
 307                 spin_unlock(&ci->i_ceph_lock);
 308         }
 309
 310         /* proceed with a normal readdir */
 311 more:
 312         /* do we have the correct frag content buffered? */
 313         if (fi->frag != frag || fi->last_readdir == NULL) {
 314                 struct ceph_mds_request *req;
 315                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 316                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 317
 318                 /* discard old result, if any */
 319                 if (fi->last_readdir) {
 320                         ceph_mdsc_put_request(fi->last_readdir);
 321                         fi->last_readdir = NULL;
 322                 }
 323
 324                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 325                      ceph_vinop(inode), frag, fi->last_name);
 326                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 327                 if (IS_ERR(req))
 328                         return PTR_ERR(req);
 329                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 330                 if (err) {
 331                         ceph_mdsc_put_request(req);
 332                         return err;
 333                 }
 334                 /* hints to request -> mds selection code */
 335                 req->r_direct_mode = USE_AUTH_MDS;
 336                 req->r_direct_hash = ceph_frag_value(frag);
 337                 req->r_direct_is_hash = true;
 338                 if (fi->last_name) {
 339                         req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL);
 340                         if (!req->r_path2) {
 341                                 ceph_mdsc_put_request(req);
 342                                 return -ENOMEM;
 343                         }
 344                 }
 345                 req->r_dir_release_cnt = fi->dir_release_count;
 346                 req->r_dir_ordered_cnt = fi->dir_ordered_count;
 347                 req->r_readdir_cache_idx = fi->readdir_cache_idx;
 348                 req->r_readdir_offset = fi->next_offset;
 349                 req->r_args.readdir.frag = cpu_to_le32(frag);
 350
 351                 req->r_inode = inode;
 352                 ihold(inode);
 353                 req->r_dentry = dget(file->f_path.dentry);
 354                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 355                 if (err < 0) {
 356                         ceph_mdsc_put_request(req);
 357                         return err;
 358                 }
 359                 dout("readdir got and parsed readdir result=%d"
 360                      " on frag %x, end=%d, complete=%d\n", err, frag,
 361                      (int)req->r_reply_info.dir_end,
 362                      (int)req->r_reply_info.dir_complete);
 363
 364
 365                 /* note next offset and last dentry name */
 366                 rinfo = &req->r_reply_info;
 367                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 368                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 369                         off = req->r_readdir_offset;
 370                         fi->next_offset = off;
 371                 }
 372
 373                 fi->frag = frag;
 374                 fi->offset = fi->next_offset;
 375                 fi->last_readdir = req;
 376
 377                 if (req->r_did_prepopulate) {
 378                         fi->readdir_cache_idx = req->r_readdir_cache_idx;
 379                         if (fi->readdir_cache_idx < 0) {
 380                                 /* preclude from marking dir ordered */
 381                                 fi->dir_ordered_count = 0;
 382                         } else if (ceph_frag_is_leftmost(frag) && off == 2) {
 383                                 /* note dir version at start of readdir so
 384                                  * we can tell if any dentries get dropped */
 385                                 fi->dir_release_count = req->r_dir_release_cnt;
 386                                 fi->dir_ordered_count = req->r_dir_ordered_cnt;
 387                         }
 388                 } else {
 389                         dout("readdir !did_prepopulate");
 390                         /* disable readdir cache */
 391                         fi->readdir_cache_idx = -1;
 392                         /* preclude from marking dir complete */
 393                         fi->dir_release_count = 0;
 394                 }
 395
 396                 if (req->r_reply_info.dir_end) {
 397                         kfree(fi->last_name);
 398                         fi->last_name = NULL;
 399                         if (ceph_frag_is_rightmost(frag))
 400                                 fi->next_offset = 2;
 401                         else
 402                                 fi->next_offset = 0;
 403                 } else {
 404                         err = note_last_dentry(fi,
 405                                        rinfo->dir_dname[rinfo->dir_nr-1],
 406                                        rinfo->dir_dname_len[rinfo->dir_nr-1],
 407                                        fi->next_offset + rinfo->dir_nr);
 408                         if (err)
 409                                 return err;
 410                 }
 411         }
 412
 413         rinfo = &fi->last_readdir->r_reply_info;
 414         dout("readdir frag %x num %d off %d chunkoff %d\n", frag,
 415              rinfo->dir_nr, off, fi->offset);
 416
 417         ctx->pos = ceph_make_fpos(frag, off);
 418         while (off >= fi->offset && off - fi->offset < rinfo->dir_nr) {
 419                 struct ceph_mds_reply_inode *in =
 420                         rinfo->dir_in[off - fi->offset].in;
 421                 struct ceph_vino vino;
 422                 ino_t ino;
 423
 424                 dout("readdir off %d (%d/%d) -> %lld '%.*s' %p\n",
 425                      off, off - fi->offset, rinfo->dir_nr, ctx->pos,
 426                      rinfo->dir_dname_len[off - fi->offset],
 427                      rinfo->dir_dname[off - fi->offset], in);
 428                 BUG_ON(!in);
 429                 ftype = le32_to_cpu(in->mode) >> 12;
 430                 vino.ino = le64_to_cpu(in->ino);
 431                 vino.snap = le64_to_cpu(in->snapid);
 432                 ino = ceph_vino_to_ino(vino);
 433                 if (!dir_emit(ctx,
 434                             rinfo->dir_dname[off - fi->offset],
 435                             rinfo->dir_dname_len[off - fi->offset],
 436                             ceph_translate_ino(inode->i_sb, ino), ftype)) {
 437                         dout("filldir stopping us...\n");
 438                         return 0;
 439                 }
 440                 off++;
 441                 ctx->pos++;
 442         }
 443
 444         if (fi->last_name) {
 445                 ceph_mdsc_put_request(fi->last_readdir);
 446                 fi->last_readdir = NULL;
 447                 goto more;
 448         }
 449
 450         /* more frags? */
 451         if (!ceph_frag_is_rightmost(frag)) {
 452                 frag = ceph_frag_next(frag);
 453                 off = 0;
 454                 ctx->pos = ceph_make_fpos(frag, off);
 455                 dout("readdir next frag is %x\n", frag);
 456                 goto more;
 457         }
 458         fi->flags |= CEPH_F_ATEND;
 459
 460         /*
 461          * if dir_release_count still matches the dir, no dentries
 462          * were released during the whole readdir, and we should have
 463          * the complete dir contents in our cache.
 464          */
 465         if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) {
 466                 spin_lock(&ci->i_ceph_lock);
 467                 if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) {
 468                         dout(" marking %p complete and ordered\n", inode);
 469                         /* use i_size to track number of entries in
 470                          * readdir cache */
 471                         BUG_ON(fi->readdir_cache_idx < 0);
 472                         i_size_write(inode, fi->readdir_cache_idx *
 473                                      sizeof(struct dentry*));
 474                 } else {
 475                         dout(" marking %p complete\n", inode);
 476                 }
 477                 __ceph_dir_set_complete(ci, fi->dir_release_count,
 478                                         fi->dir_ordered_count);
 479                 spin_unlock(&ci->i_ceph_lock);
 480         }
 481
 482         dout("readdir %p file %p done.\n", inode, file);
 483         return 0;
 484 }
 485
 486 static void reset_readdir(struct ceph_file_info *fi, unsigned frag)
 487 {
 488         if (fi->last_readdir) {
 489                 ceph_mdsc_put_request(fi->last_readdir);
 490                 fi->last_readdir = NULL;
 491         }
 492         kfree(fi->last_name);
 493         fi->last_name = NULL;
 494         fi->dir_release_count = 0;
 495         fi->readdir_cache_idx = -1;
 496         if (ceph_frag_is_leftmost(frag))
 497                 fi->next_offset = 2;  /* compensate for . and .. */
 498         else
 499                 fi->next_offset = 0;
 500         fi->flags &= ~CEPH_F_ATEND;
 501 }
 502
 503 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 504 {
 505         struct ceph_file_info *fi = file->private_data;
 506         struct inode *inode = file->f_mapping->host;
 507         loff_t old_offset = ceph_make_fpos(fi->frag, fi->next_offset);
 508         loff_t retval;
 509
 510         inode_lock(inode);
 511         retval = -EINVAL;
 512         switch (whence) {
 513         case SEEK_CUR:
 514                 offset += file->f_pos;
 515         case SEEK_SET:
 516                 break;
 517         case SEEK_END:
 518                 retval = -EOPNOTSUPP;
 519         default:
 520                 goto out;
 521         }
 522
 523         if (offset >= 0) {
 524                 if (offset != file->f_pos) {
 525                         file->f_pos = offset;
 526                         file->f_version = 0;
 527                         fi->flags &= ~CEPH_F_ATEND;
 528                 }
 529                 retval = offset;
 530
 531                 if (offset == 0 ||
 532                     fpos_frag(offset) != fi->frag ||
 533                     fpos_off(offset) < fi->offset) {
 534                         /* discard buffered readdir content on seekdir(0), or
 535                          * seek to new frag, or seek prior to current chunk */
 536                         dout("dir_llseek dropping %p content\n", file);
 537                         reset_readdir(fi, fpos_frag(offset));
 538                 } else if (fpos_cmp(offset, old_offset) > 0) {
 539                         /* reset dir_release_count if we did a forward seek */
 540                         fi->dir_release_count = 0;
 541                         fi->readdir_cache_idx = -1;
 542                 }
 543         }
 544 out:
 545         inode_unlock(inode);
 546         return retval;
 547 }
 548
 549 /*
 550  * Handle lookups for the hidden .snap directory.
 551  */
 552 int ceph_handle_snapdir(struct ceph_mds_request *req,
 553                         struct dentry *dentry, int err)
 554 {
 555         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 556         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 557
 558         /* .snap dir? */
 559         if (err == -ENOENT &&
 560             ceph_snap(parent) == CEPH_NOSNAP &&
 561             strcmp(dentry->d_name.name,
 562                    fsc->mount_options->snapdir_name) == 0) {
 563                 struct inode *inode = ceph_get_snapdir(parent);
 564                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 565                      dentry, dentry, inode);
 566                 BUG_ON(!d_unhashed(dentry));
 567                 d_add(dentry, inode);
 568                 err = 0;
 569         }
 570         return err;
 571 }
 572
 573 /*
 574  * Figure out final result of a lookup/open request.
 575  *
 576  * Mainly, make sure we return the final req->r_dentry (if it already
 577  * existed) in place of the original VFS-provided dentry when they
 578  * differ.
 579  *
 580  * Gracefully handle the case where the MDS replies with -ENOENT and
 581  * no trace (which it may do, at its discretion, e.g., if it doesn't
 582  * care to issue a lease on the negative dentry).
 583  */
 584 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 585                                   struct dentry *dentry, int err)
 586 {
 587         if (err == -ENOENT) {
 588                 /* no trace? */
 589                 err = 0;
 590                 if (!req->r_reply_info.head->is_dentry) {
 591                         dout("ENOENT and no trace, dentry %p inode %p\n",
 592                              dentry, d_inode(dentry));
 593                         if (d_really_is_positive(dentry)) {
 594                                 d_drop(dentry);
 595                                 err = -ENOENT;
 596                         } else {
 597                                 d_add(dentry, NULL);
 598                         }
 599                 }
 600         }
 601         if (err)
 602                 dentry = ERR_PTR(err);
 603         else if (dentry != req->r_dentry)
 604                 dentry = dget(req->r_dentry);   /* we got spliced */
 605         else
 606                 dentry = NULL;
 607         return dentry;
 608 }
 609
 610 static int is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 611 {
 612         return ceph_ino(inode) == CEPH_INO_ROOT &&
 613                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 614 }
 615
 616 /*
 617  * Look up a single dir entry.  If there is a lookup intent, inform
 618  * the MDS so that it gets our 'caps wanted' value in a single op.
 619  */
 620 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 621                                   unsigned int flags)
 622 {
 623         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 624         struct ceph_mds_client *mdsc = fsc->mdsc;
 625         struct ceph_mds_request *req;
 626         int op;
 627         int mask;
 628         int err;
 629
 630         dout("lookup %p dentry %p '%pd'\n",
 631              dir, dentry, dentry);
 632
 633         if (dentry->d_name.len > NAME_MAX)
 634                 return ERR_PTR(-ENAMETOOLONG);
 635
 636         err = ceph_init_dentry(dentry);
 637         if (err < 0)
 638                 return ERR_PTR(err);
 639
 640         /* can we conclude ENOENT locally? */
 641         if (d_really_is_negative(dentry)) {
 642                 struct ceph_inode_info *ci = ceph_inode(dir);
 643                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 644
 645                 spin_lock(&ci->i_ceph_lock);
 646                 dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
 647                 if (strncmp(dentry->d_name.name,
 648                             fsc->mount_options->snapdir_name,
 649                             dentry->d_name.len) &&
 650                     !is_root_ceph_dentry(dir, dentry) &&
 651                     ceph_test_mount_opt(fsc, DCACHE) &&
 652                     __ceph_dir_is_complete(ci) &&
 653                     (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
 654                         spin_unlock(&ci->i_ceph_lock);
 655                         dout(" dir %p complete, -ENOENT\n", dir);
 656                         d_add(dentry, NULL);
 657                         di->lease_shared_gen = ci->i_shared_gen;
 658                         return NULL;
 659                 }
 660                 spin_unlock(&ci->i_ceph_lock);
 661         }
 662
 663         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 664                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 665         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 666         if (IS_ERR(req))
 667                 return ERR_CAST(req);
 668         req->r_dentry = dget(dentry);
 669         req->r_num_caps = 2;
 670
 671         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 672         if (ceph_security_xattr_wanted(dir))
 673                 mask |= CEPH_CAP_XATTR_SHARED;
 674         req->r_args.getattr.mask = cpu_to_le32(mask);
 675
 676         req->r_locked_dir = dir;
 677         err = ceph_mdsc_do_request(mdsc, NULL, req);
 678         err = ceph_handle_snapdir(req, dentry, err);
 679         dentry = ceph_finish_lookup(req, dentry, err);
 680         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 681         dout("lookup result=%p\n", dentry);
 682         return dentry;
 683 }
 684
 685 /*
 686  * If we do a create but get no trace back from the MDS, follow up with
 687  * a lookup (the VFS expects us to link up the provided dentry).
 688  */
 689 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 690 {
 691         struct dentry *result = ceph_lookup(dir, dentry, 0);
 692
 693         if (result && !IS_ERR(result)) {
 694                 /*
 695                  * We created the item, then did a lookup, and found
 696                  * it was already linked to another inode we already
 697                  * had in our cache (and thus got spliced). To not
 698                  * confuse VFS (especially when inode is a directory),
 699                  * we don't link our dentry to that inode, return an
 700                  * error instead.
 701                  *
 702                  * This event should be rare and it happens only when
 703                  * we talk to old MDS. Recent MDS does not send traceless
 704                  * reply for request that creates new inode.
 705                  */
 706                 d_drop(result);
 707                 return -ESTALE;
 708         }
 709         return PTR_ERR(result);
 710 }
 711
 712 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 713                       umode_t mode, dev_t rdev)
 714 {
 715         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 716         struct ceph_mds_client *mdsc = fsc->mdsc;
 717         struct ceph_mds_request *req;
 718         struct ceph_acls_info acls = {};
 719         int err;
 720
 721         if (ceph_snap(dir) != CEPH_NOSNAP)
 722                 return -EROFS;
 723
 724         err = ceph_pre_init_acls(dir, &mode, &acls);
 725         if (err < 0)
 726                 return err;
 727
 728         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 729              dir, dentry, mode, rdev);
 730         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 731         if (IS_ERR(req)) {
 732                 err = PTR_ERR(req);
 733                 goto out;
 734         }
 735         req->r_dentry = dget(dentry);
 736         req->r_num_caps = 2;
 737         req->r_locked_dir = dir;
 738         req->r_args.mknod.mode = cpu_to_le32(mode);
 739         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 740         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 741         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 742         if (acls.pagelist) {
 743                 req->r_pagelist = acls.pagelist;
 744                 acls.pagelist = NULL;
 745         }
 746         err = ceph_mdsc_do_request(mdsc, dir, req);
 747         if (!err && !req->r_reply_info.head->is_dentry)
 748                 err = ceph_handle_notrace_create(dir, dentry);
 749         ceph_mdsc_put_request(req);
 750 out:
 751         if (!err)
 752                 ceph_init_inode_acls(d_inode(dentry), &acls);
 753         else
 754                 d_drop(dentry);
 755         ceph_release_acls_info(&acls);
 756         return err;
 757 }
 758
 759 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 760                        bool excl)
 761 {
 762         return ceph_mknod(dir, dentry, mode, 0);
 763 }
 764
 765 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 766                             const char *dest)
 767 {
 768         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 769         struct ceph_mds_client *mdsc = fsc->mdsc;
 770         struct ceph_mds_request *req;
 771         int err;
 772
 773         if (ceph_snap(dir) != CEPH_NOSNAP)
 774                 return -EROFS;
 775
 776         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 777         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 778         if (IS_ERR(req)) {
 779                 err = PTR_ERR(req);
 780                 goto out;
 781         }
 782         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 783         if (!req->r_path2) {
 784                 err = -ENOMEM;
 785                 ceph_mdsc_put_request(req);
 786                 goto out;
 787         }
 788         req->r_locked_dir = dir;
 789         req->r_dentry = dget(dentry);
 790         req->r_num_caps = 2;
 791         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 792         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 793         err = ceph_mdsc_do_request(mdsc, dir, req);
 794         if (!err && !req->r_reply_info.head->is_dentry)
 795                 err = ceph_handle_notrace_create(dir, dentry);
 796         ceph_mdsc_put_request(req);
 797 out:
 798         if (err)
 799                 d_drop(dentry);
 800         return err;
 801 }
 802
 803 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 804 {
 805         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 806         struct ceph_mds_client *mdsc = fsc->mdsc;
 807         struct ceph_mds_request *req;
 808         struct ceph_acls_info acls = {};
 809         int err = -EROFS;
 810         int op;
 811
 812         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 813                 /* mkdir .snap/foo is a MKSNAP */
 814                 op = CEPH_MDS_OP_MKSNAP;
 815                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 816                      dentry, dentry);
 817         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 818                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 819                 op = CEPH_MDS_OP_MKDIR;
 820         } else {
 821                 goto out;
 822         }
 823
 824         mode |= S_IFDIR;
 825         err = ceph_pre_init_acls(dir, &mode, &acls);
 826         if (err < 0)
 827                 goto out;
 828
 829         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 830         if (IS_ERR(req)) {
 831                 err = PTR_ERR(req);
 832                 goto out;
 833         }
 834
 835         req->r_dentry = dget(dentry);
 836         req->r_num_caps = 2;
 837         req->r_locked_dir = dir;
 838         req->r_args.mkdir.mode = cpu_to_le32(mode);
 839         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 840         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 841         if (acls.pagelist) {
 842                 req->r_pagelist = acls.pagelist;
 843                 acls.pagelist = NULL;
 844         }
 845         err = ceph_mdsc_do_request(mdsc, dir, req);
 846         if (!err &&
 847             !req->r_reply_info.head->is_target &&
 848             !req->r_reply_info.head->is_dentry)
 849                 err = ceph_handle_notrace_create(dir, dentry);
 850         ceph_mdsc_put_request(req);
 851 out:
 852         if (!err)
 853                 ceph_init_inode_acls(d_inode(dentry), &acls);
 854         else
 855                 d_drop(dentry);
 856         ceph_release_acls_info(&acls);
 857         return err;
 858 }
 859
 860 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
 861                      struct dentry *dentry)
 862 {
 863         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 864         struct ceph_mds_client *mdsc = fsc->mdsc;
 865         struct ceph_mds_request *req;
 866         int err;
 867
 868         if (ceph_snap(dir) != CEPH_NOSNAP)
 869                 return -EROFS;
 870
 871         dout("link in dir %p old_dentry %p dentry %p\n", dir,
 872              old_dentry, dentry);
 873         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
 874         if (IS_ERR(req)) {
 875                 d_drop(dentry);
 876                 return PTR_ERR(req);
 877         }
 878         req->r_dentry = dget(dentry);
 879         req->r_num_caps = 2;
 880         req->r_old_dentry = dget(old_dentry);
 881         req->r_locked_dir = dir;
 882         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 883         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 884         /* release LINK_SHARED on source inode (mds will lock it) */
 885         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
 886         err = ceph_mdsc_do_request(mdsc, dir, req);
 887         if (err) {
 888                 d_drop(dentry);
 889         } else if (!req->r_reply_info.head->is_dentry) {
 890                 ihold(d_inode(old_dentry));
 891                 d_instantiate(dentry, d_inode(old_dentry));
 892         }
 893         ceph_mdsc_put_request(req);
 894         return err;
 895 }
 896
 897 /*
 898  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
 899  * looks like the link count will hit 0, drop any other caps (other
 900  * than PIN) we don't specifically want (due to the file still being
 901  * open).
 902  */
 903 static int drop_caps_for_unlink(struct inode *inode)
 904 {
 905         struct ceph_inode_info *ci = ceph_inode(inode);
 906         int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
 907
 908         spin_lock(&ci->i_ceph_lock);
 909         if (inode->i_nlink == 1) {
 910                 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
 911                 ci->i_ceph_flags |= CEPH_I_NODELAY;
 912         }
 913         spin_unlock(&ci->i_ceph_lock);
 914         return drop;
 915 }
 916
 917 /*
 918  * rmdir and unlink are differ only by the metadata op code
 919  */
 920 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
 921 {
 922         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 923         struct ceph_mds_client *mdsc = fsc->mdsc;
 924         struct inode *inode = d_inode(dentry);
 925         struct ceph_mds_request *req;
 926         int err = -EROFS;
 927         int op;
 928
 929         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 930                 /* rmdir .snap/foo is RMSNAP */
 931                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
 932                 op = CEPH_MDS_OP_RMSNAP;
 933         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 934                 dout("unlink/rmdir dir %p dn %p inode %p\n",
 935                      dir, dentry, inode);
 936                 op = d_is_dir(dentry) ?
 937                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
 938         } else
 939                 goto out;
 940         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 941         if (IS_ERR(req)) {
 942                 err = PTR_ERR(req);
 943                 goto out;
 944         }
 945         req->r_dentry = dget(dentry);
 946         req->r_num_caps = 2;
 947         req->r_locked_dir = dir;
 948         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 949         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 950         req->r_inode_drop = drop_caps_for_unlink(inode);
 951         err = ceph_mdsc_do_request(mdsc, dir, req);
 952         if (!err && !req->r_reply_info.head->is_dentry)
 953                 d_delete(dentry);
 954         ceph_mdsc_put_request(req);
 955 out:
 956         return err;
 957 }
 958
 959 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
 960                        struct inode *new_dir, struct dentry *new_dentry)
 961 {
 962         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
 963         struct ceph_mds_client *mdsc = fsc->mdsc;
 964         struct ceph_mds_request *req;
 965         int op = CEPH_MDS_OP_RENAME;
 966         int err;
 967
 968         if (ceph_snap(old_dir) != ceph_snap(new_dir))
 969                 return -EXDEV;
 970         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
 971                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
 972                         op = CEPH_MDS_OP_RENAMESNAP;
 973                 else
 974                         return -EROFS;
 975         }
 976         dout("rename dir %p dentry %p to dir %p dentry %p\n",
 977              old_dir, old_dentry, new_dir, new_dentry);
 978         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 979         if (IS_ERR(req))
 980                 return PTR_ERR(req);
 981         ihold(old_dir);
 982         req->r_dentry = dget(new_dentry);
 983         req->r_num_caps = 2;
 984         req->r_old_dentry = dget(old_dentry);
 985         req->r_old_dentry_dir = old_dir;
 986         req->r_locked_dir = new_dir;
 987         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
 988         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
 989         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 990         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 991         /* release LINK_RDCACHE on source inode (mds will lock it) */
 992         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
 993         if (d_really_is_positive(new_dentry))
 994                 req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry));
 995         err = ceph_mdsc_do_request(mdsc, old_dir, req);
 996         if (!err && !req->r_reply_info.head->is_dentry) {
 997                 /*
 998                  * Normally d_move() is done by fill_trace (called by
 999                  * do_request, above).  If there is no trace, we need
1000                  * to do it here.
1001                  */
1002
1003                 /* d_move screws up sibling dentries' offsets */
1004                 ceph_dir_clear_complete(old_dir);
1005                 ceph_dir_clear_complete(new_dir);
1006
1007                 d_move(old_dentry, new_dentry);
1008
1009                 /* ensure target dentry is invalidated, despite
1010                    rehashing bug in vfs_rename_dir */
1011                 ceph_invalidate_dentry_lease(new_dentry);
1012         }
1013         ceph_mdsc_put_request(req);
1014         return err;
1015 }
1016
1017 /*
1018  * Ensure a dentry lease will no longer revalidate.
1019  */
1020 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1021 {
1022         spin_lock(&dentry->d_lock);
1023         dentry->d_time = jiffies;
1024         ceph_dentry(dentry)->lease_shared_gen = 0;
1025         spin_unlock(&dentry->d_lock);
1026 }
1027
1028 /*
1029  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1030  * renew if the least is more than half up.
1031  */
1032 static int dentry_lease_is_valid(struct dentry *dentry)
1033 {
1034         struct ceph_dentry_info *di;
1035         struct ceph_mds_session *s;
1036         int valid = 0;
1037         u32 gen;
1038         unsigned long ttl;
1039         struct ceph_mds_session *session = NULL;
1040         struct inode *dir = NULL;
1041         u32 seq = 0;
1042
1043         spin_lock(&dentry->d_lock);
1044         di = ceph_dentry(dentry);
1045         if (di->lease_session) {
1046                 s = di->lease_session;
1047                 spin_lock(&s->s_gen_ttl_lock);
1048                 gen = s->s_cap_gen;
1049                 ttl = s->s_cap_ttl;
1050                 spin_unlock(&s->s_gen_ttl_lock);
1051
1052                 if (di->lease_gen == gen &&
1053                     time_before(jiffies, dentry->d_time) &&
1054                     time_before(jiffies, ttl)) {
1055                         valid = 1;
1056                         if (di->lease_renew_after &&
1057                             time_after(jiffies, di->lease_renew_after)) {
1058                                 /* we should renew */
1059                                 dir = d_inode(dentry->d_parent);
1060                                 session = ceph_get_mds_session(s);
1061                                 seq = di->lease_seq;
1062                                 di->lease_renew_after = 0;
1063                                 di->lease_renew_from = jiffies;
1064                         }
1065                 }
1066         }
1067         spin_unlock(&dentry->d_lock);
1068
1069         if (session) {
1070                 ceph_mdsc_lease_send_msg(session, dir, dentry,
1071                                          CEPH_MDS_LEASE_RENEW, seq);
1072                 ceph_put_mds_session(session);
1073         }
1074         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1075         return valid;
1076 }
1077
1078 /*
1079  * Check if directory-wide content lease/cap is valid.
1080  */
1081 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1082 {
1083         struct ceph_inode_info *ci = ceph_inode(dir);
1084         struct ceph_dentry_info *di = ceph_dentry(dentry);
1085         int valid = 0;
1086
1087         spin_lock(&ci->i_ceph_lock);
1088         if (ci->i_shared_gen == di->lease_shared_gen)
1089                 valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1090         spin_unlock(&ci->i_ceph_lock);
1091         dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1092              dir, (unsigned)ci->i_shared_gen, dentry,
1093              (unsigned)di->lease_shared_gen, valid);
1094         return valid;
1095 }
1096
1097 /*
1098  * Check if cached dentry can be trusted.
1099  */
1100 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1101 {
1102         int valid = 0;
1103         struct inode *dir;
1104
1105         if (flags & LOOKUP_RCU)
1106                 return -ECHILD;
1107
1108         dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1109              dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1110
1111         dir = ceph_get_dentry_parent_inode(dentry);
1112
1113         /* always trust cached snapped dentries, snapdir dentry */
1114         if (ceph_snap(dir) != CEPH_NOSNAP) {
1115                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1116                      dentry, d_inode(dentry));
1117                 valid = 1;
1118         } else if (d_really_is_positive(dentry) &&
1119                    ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1120                 valid = 1;
1121         } else if (dentry_lease_is_valid(dentry) ||
1122                    dir_lease_is_valid(dir, dentry)) {
1123                 if (d_really_is_positive(dentry))
1124                         valid = ceph_is_any_caps(d_inode(dentry));
1125                 else
1126                         valid = 1;
1127         }
1128
1129         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1130         if (valid) {
1131                 ceph_dentry_lru_touch(dentry);
1132         } else {
1133                 ceph_dir_clear_complete(dir);
1134         }
1135         iput(dir);
1136         return valid;
1137 }
1138
1139 /*
1140  * Release our ceph_dentry_info.
1141  */
1142 static void ceph_d_release(struct dentry *dentry)
1143 {
1144         struct ceph_dentry_info *di = ceph_dentry(dentry);
1145
1146         dout("d_release %p\n", dentry);
1147         ceph_dentry_lru_del(dentry);
1148         if (di->lease_session)
1149                 ceph_put_mds_session(di->lease_session);
1150         kmem_cache_free(ceph_dentry_cachep, di);
1151         dentry->d_fsdata = NULL;
1152 }
1153
1154 static int ceph_snapdir_d_revalidate(struct dentry *dentry,
1155                                           unsigned int flags)
1156 {
1157         /*
1158          * Eventually, we'll want to revalidate snapped metadata
1159          * too... probably...
1160          */
1161         return 1;
1162 }
1163
1164 /*
1165  * When the VFS prunes a dentry from the cache, we need to clear the
1166  * complete flag on the parent directory.
1167  *
1168  * Called under dentry->d_lock.
1169  */
1170 static void ceph_d_prune(struct dentry *dentry)
1171 {
1172         dout("ceph_d_prune %p\n", dentry);
1173
1174         /* do we have a valid parent? */
1175         if (IS_ROOT(dentry))
1176                 return;
1177
1178         /* if we are not hashed, we don't affect dir's completeness */
1179         if (d_unhashed(dentry))
1180                 return;
1181
1182         /*
1183          * we hold d_lock, so d_parent is stable, and d_fsdata is never
1184          * cleared until d_release
1185          */
1186         ceph_dir_clear_complete(d_inode(dentry->d_parent));
1187 }
1188
1189 /*
1190  * read() on a dir.  This weird interface hack only works if mounted
1191  * with '-o dirstat'.
1192  */
1193 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1194                              loff_t *ppos)
1195 {
1196         struct ceph_file_info *cf = file->private_data;
1197         struct inode *inode = file_inode(file);
1198         struct ceph_inode_info *ci = ceph_inode(inode);
1199         int left;
1200         const int bufsize = 1024;
1201
1202         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1203                 return -EISDIR;
1204
1205         if (!cf->dir_info) {
1206                 cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1207                 if (!cf->dir_info)
1208                         return -ENOMEM;
1209                 cf->dir_info_len =
1210                         snprintf(cf->dir_info, bufsize,
1211                                 "entries:   %20lld\n"
1212                                 " files:    %20lld\n"
1213                                 " subdirs:  %20lld\n"
1214                                 "rentries:  %20lld\n"
1215                                 " rfiles:   %20lld\n"
1216                                 " rsubdirs: %20lld\n"
1217                                 "rbytes:    %20lld\n"
1218                                 "rctime:    %10ld.%09ld\n",
1219                                 ci->i_files + ci->i_subdirs,
1220                                 ci->i_files,
1221                                 ci->i_subdirs,
1222                                 ci->i_rfiles + ci->i_rsubdirs,
1223                                 ci->i_rfiles,
1224                                 ci->i_rsubdirs,
1225                                 ci->i_rbytes,
1226                                 (long)ci->i_rctime.tv_sec,
1227                                 (long)ci->i_rctime.tv_nsec);
1228         }
1229
1230         if (*ppos >= cf->dir_info_len)
1231                 return 0;
1232         size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1233         left = copy_to_user(buf, cf->dir_info + *ppos, size);
1234         if (left == size)
1235                 return -EFAULT;
1236         *ppos += (size - left);
1237         return size - left;
1238 }
1239
1240 /*
1241  * We maintain a private dentry LRU.
1242  *
1243  * FIXME: this needs to be changed to a per-mds lru to be useful.
1244  */
1245 void ceph_dentry_lru_add(struct dentry *dn)
1246 {
1247         struct ceph_dentry_info *di = ceph_dentry(dn);
1248         struct ceph_mds_client *mdsc;
1249
1250         dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1251         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1252         spin_lock(&mdsc->dentry_lru_lock);
1253         list_add_tail(&di->lru, &mdsc->dentry_lru);
1254         mdsc->num_dentry++;
1255         spin_unlock(&mdsc->dentry_lru_lock);
1256 }
1257
1258 void ceph_dentry_lru_touch(struct dentry *dn)
1259 {
1260         struct ceph_dentry_info *di = ceph_dentry(dn);
1261         struct ceph_mds_client *mdsc;
1262
1263         dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1264              di->offset);
1265         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1266         spin_lock(&mdsc->dentry_lru_lock);
1267         list_move_tail(&di->lru, &mdsc->dentry_lru);
1268         spin_unlock(&mdsc->dentry_lru_lock);
1269 }
1270
1271 void ceph_dentry_lru_del(struct dentry *dn)
1272 {
1273         struct ceph_dentry_info *di = ceph_dentry(dn);
1274         struct ceph_mds_client *mdsc;
1275
1276         dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1277         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1278         spin_lock(&mdsc->dentry_lru_lock);
1279         list_del_init(&di->lru);
1280         mdsc->num_dentry--;
1281         spin_unlock(&mdsc->dentry_lru_lock);
1282 }
1283
1284 /*
1285  * Return name hash for a given dentry.  This is dependent on
1286  * the parent directory's hash function.
1287  */
1288 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1289 {
1290         struct ceph_inode_info *dci = ceph_inode(dir);
1291
1292         switch (dci->i_dir_layout.dl_dir_hash) {
1293         case 0: /* for backward compat */
1294         case CEPH_STR_HASH_LINUX:
1295                 return dn->d_name.hash;
1296
1297         default:
1298                 return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1299                                      dn->d_name.name, dn->d_name.len);
1300         }
1301 }
1302
1303 const struct file_operations ceph_dir_fops = {
1304         .read = ceph_read_dir,
1305         .iterate = ceph_readdir,
1306         .llseek = ceph_dir_llseek,
1307         .open = ceph_open,
1308         .release = ceph_release,
1309         .unlocked_ioctl = ceph_ioctl,
1310         .fsync = ceph_fsync,
1311 };
1312
1313 const struct file_operations ceph_snapdir_fops = {
1314         .iterate = ceph_readdir,
1315         .llseek = ceph_dir_llseek,
1316         .open = ceph_open,
1317         .release = ceph_release,
1318 };
1319
1320 const struct inode_operations ceph_dir_iops = {
1321         .lookup = ceph_lookup,
1322         .permission = ceph_permission,
1323         .getattr = ceph_getattr,
1324         .setattr = ceph_setattr,
1325         .setxattr = ceph_setxattr,
1326         .getxattr = ceph_getxattr,
1327         .listxattr = ceph_listxattr,
1328         .removexattr = ceph_removexattr,
1329         .get_acl = ceph_get_acl,
1330         .set_acl = ceph_set_acl,
1331         .mknod = ceph_mknod,
1332         .symlink = ceph_symlink,
1333         .mkdir = ceph_mkdir,
1334         .link = ceph_link,
1335         .unlink = ceph_unlink,
1336         .rmdir = ceph_unlink,
1337         .rename = ceph_rename,
1338         .create = ceph_create,
1339         .atomic_open = ceph_atomic_open,
1340 };
1341
1342 const struct inode_operations ceph_snapdir_iops = {
1343         .lookup = ceph_lookup,
1344         .permission = ceph_permission,
1345         .getattr = ceph_getattr,
1346         .mkdir = ceph_mkdir,
1347         .rmdir = ceph_unlink,
1348         .rename = ceph_rename,
1349 };
1350
1351 const struct dentry_operations ceph_dentry_ops = {
1352         .d_revalidate = ceph_d_revalidate,
1353         .d_release = ceph_d_release,
1354         .d_prune = ceph_d_prune,
1355 };
1356
1357 const struct dentry_operations ceph_snapdir_dentry_ops = {
1358         .d_revalidate = ceph_snapdir_d_revalidate,
1359         .d_release = ceph_d_release,
1360 };
1361
1362 const struct dentry_operations ceph_snap_dentry_ops = {
1363         .d_release = ceph_d_release,
1364         .d_prune = ceph_d_prune,
1365 };