kernel/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         mutex_lock(&inode->i_mutex);
 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         mutex_unlock(&inode->i_mutex);
 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 err;
 628
 629         dout("lookup %p dentry %p '%pd'\n",
 630              dir, dentry, dentry);
 631
 632         if (dentry->d_name.len > NAME_MAX)
 633                 return ERR_PTR(-ENAMETOOLONG);
 634
 635         err = ceph_init_dentry(dentry);
 636         if (err < 0)
 637                 return ERR_PTR(err);
 638
 639         /* can we conclude ENOENT locally? */
 640         if (d_really_is_negative(dentry)) {
 641                 struct ceph_inode_info *ci = ceph_inode(dir);
 642                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 643
 644                 spin_lock(&ci->i_ceph_lock);
 645                 dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
 646                 if (strncmp(dentry->d_name.name,
 647                             fsc->mount_options->snapdir_name,
 648                             dentry->d_name.len) &&
 649                     !is_root_ceph_dentry(dir, dentry) &&
 650                     ceph_test_mount_opt(fsc, DCACHE) &&
 651                     __ceph_dir_is_complete(ci) &&
 652                     (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
 653                         spin_unlock(&ci->i_ceph_lock);
 654                         dout(" dir %p complete, -ENOENT\n", dir);
 655                         d_add(dentry, NULL);
 656                         di->lease_shared_gen = ci->i_shared_gen;
 657                         return NULL;
 658                 }
 659                 spin_unlock(&ci->i_ceph_lock);
 660         }
 661
 662         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 663                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 664         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 665         if (IS_ERR(req))
 666                 return ERR_CAST(req);
 667         req->r_dentry = dget(dentry);
 668         req->r_num_caps = 2;
 669         /* we only need inode linkage */
 670         req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INODE);
 671         req->r_locked_dir = dir;
 672         err = ceph_mdsc_do_request(mdsc, NULL, req);
 673         err = ceph_handle_snapdir(req, dentry, err);
 674         dentry = ceph_finish_lookup(req, dentry, err);
 675         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 676         dout("lookup result=%p\n", dentry);
 677         return dentry;
 678 }
 679
 680 /*
 681  * If we do a create but get no trace back from the MDS, follow up with
 682  * a lookup (the VFS expects us to link up the provided dentry).
 683  */
 684 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 685 {
 686         struct dentry *result = ceph_lookup(dir, dentry, 0);
 687
 688         if (result && !IS_ERR(result)) {
 689                 /*
 690                  * We created the item, then did a lookup, and found
 691                  * it was already linked to another inode we already
 692                  * had in our cache (and thus got spliced). To not
 693                  * confuse VFS (especially when inode is a directory),
 694                  * we don't link our dentry to that inode, return an
 695                  * error instead.
 696                  *
 697                  * This event should be rare and it happens only when
 698                  * we talk to old MDS. Recent MDS does not send traceless
 699                  * reply for request that creates new inode.
 700                  */
 701                 d_drop(result);
 702                 return -ESTALE;
 703         }
 704         return PTR_ERR(result);
 705 }
 706
 707 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 708                       umode_t mode, dev_t rdev)
 709 {
 710         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 711         struct ceph_mds_client *mdsc = fsc->mdsc;
 712         struct ceph_mds_request *req;
 713         struct ceph_acls_info acls = {};
 714         int err;
 715
 716         if (ceph_snap(dir) != CEPH_NOSNAP)
 717                 return -EROFS;
 718
 719         err = ceph_pre_init_acls(dir, &mode, &acls);
 720         if (err < 0)
 721                 return err;
 722
 723         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 724              dir, dentry, mode, rdev);
 725         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 726         if (IS_ERR(req)) {
 727                 err = PTR_ERR(req);
 728                 goto out;
 729         }
 730         req->r_dentry = dget(dentry);
 731         req->r_num_caps = 2;
 732         req->r_locked_dir = dir;
 733         req->r_args.mknod.mode = cpu_to_le32(mode);
 734         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 735         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 736         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 737         if (acls.pagelist) {
 738                 req->r_pagelist = acls.pagelist;
 739                 acls.pagelist = NULL;
 740         }
 741         err = ceph_mdsc_do_request(mdsc, dir, req);
 742         if (!err && !req->r_reply_info.head->is_dentry)
 743                 err = ceph_handle_notrace_create(dir, dentry);
 744         ceph_mdsc_put_request(req);
 745 out:
 746         if (!err)
 747                 ceph_init_inode_acls(d_inode(dentry), &acls);
 748         else
 749                 d_drop(dentry);
 750         ceph_release_acls_info(&acls);
 751         return err;
 752 }
 753
 754 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 755                        bool excl)
 756 {
 757         return ceph_mknod(dir, dentry, mode, 0);
 758 }
 759
 760 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 761                             const char *dest)
 762 {
 763         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 764         struct ceph_mds_client *mdsc = fsc->mdsc;
 765         struct ceph_mds_request *req;
 766         int err;
 767
 768         if (ceph_snap(dir) != CEPH_NOSNAP)
 769                 return -EROFS;
 770
 771         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 772         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 773         if (IS_ERR(req)) {
 774                 err = PTR_ERR(req);
 775                 goto out;
 776         }
 777         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 778         if (!req->r_path2) {
 779                 err = -ENOMEM;
 780                 ceph_mdsc_put_request(req);
 781                 goto out;
 782         }
 783         req->r_locked_dir = dir;
 784         req->r_dentry = dget(dentry);
 785         req->r_num_caps = 2;
 786         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 787         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 788         err = ceph_mdsc_do_request(mdsc, dir, req);
 789         if (!err && !req->r_reply_info.head->is_dentry)
 790                 err = ceph_handle_notrace_create(dir, dentry);
 791         ceph_mdsc_put_request(req);
 792 out:
 793         if (err)
 794                 d_drop(dentry);
 795         return err;
 796 }
 797
 798 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 799 {
 800         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 801         struct ceph_mds_client *mdsc = fsc->mdsc;
 802         struct ceph_mds_request *req;
 803         struct ceph_acls_info acls = {};
 804         int err = -EROFS;
 805         int op;
 806
 807         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 808                 /* mkdir .snap/foo is a MKSNAP */
 809                 op = CEPH_MDS_OP_MKSNAP;
 810                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 811                      dentry, dentry);
 812         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 813                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 814                 op = CEPH_MDS_OP_MKDIR;
 815         } else {
 816                 goto out;
 817         }
 818
 819         mode |= S_IFDIR;
 820         err = ceph_pre_init_acls(dir, &mode, &acls);
 821         if (err < 0)
 822                 goto out;
 823
 824         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 825         if (IS_ERR(req)) {
 826                 err = PTR_ERR(req);
 827                 goto out;
 828         }
 829
 830         req->r_dentry = dget(dentry);
 831         req->r_num_caps = 2;
 832         req->r_locked_dir = dir;
 833         req->r_args.mkdir.mode = cpu_to_le32(mode);
 834         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 835         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 836         if (acls.pagelist) {
 837                 req->r_pagelist = acls.pagelist;
 838                 acls.pagelist = NULL;
 839         }
 840         err = ceph_mdsc_do_request(mdsc, dir, req);
 841         if (!err &&
 842             !req->r_reply_info.head->is_target &&
 843             !req->r_reply_info.head->is_dentry)
 844                 err = ceph_handle_notrace_create(dir, dentry);
 845         ceph_mdsc_put_request(req);
 846 out:
 847         if (!err)
 848                 ceph_init_inode_acls(d_inode(dentry), &acls);
 849         else
 850                 d_drop(dentry);
 851         ceph_release_acls_info(&acls);
 852         return err;
 853 }
 854
 855 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
 856                      struct dentry *dentry)
 857 {
 858         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 859         struct ceph_mds_client *mdsc = fsc->mdsc;
 860         struct ceph_mds_request *req;
 861         int err;
 862
 863         if (ceph_snap(dir) != CEPH_NOSNAP)
 864                 return -EROFS;
 865
 866         dout("link in dir %p old_dentry %p dentry %p\n", dir,
 867              old_dentry, dentry);
 868         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
 869         if (IS_ERR(req)) {
 870                 d_drop(dentry);
 871                 return PTR_ERR(req);
 872         }
 873         req->r_dentry = dget(dentry);
 874         req->r_num_caps = 2;
 875         req->r_old_dentry = dget(old_dentry);
 876         req->r_locked_dir = dir;
 877         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 878         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 879         /* release LINK_SHARED on source inode (mds will lock it) */
 880         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
 881         err = ceph_mdsc_do_request(mdsc, dir, req);
 882         if (err) {
 883                 d_drop(dentry);
 884         } else if (!req->r_reply_info.head->is_dentry) {
 885                 ihold(d_inode(old_dentry));
 886                 d_instantiate(dentry, d_inode(old_dentry));
 887         }
 888         ceph_mdsc_put_request(req);
 889         return err;
 890 }
 891
 892 /*
 893  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
 894  * looks like the link count will hit 0, drop any other caps (other
 895  * than PIN) we don't specifically want (due to the file still being
 896  * open).
 897  */
 898 static int drop_caps_for_unlink(struct inode *inode)
 899 {
 900         struct ceph_inode_info *ci = ceph_inode(inode);
 901         int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
 902
 903         spin_lock(&ci->i_ceph_lock);
 904         if (inode->i_nlink == 1) {
 905                 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
 906                 ci->i_ceph_flags |= CEPH_I_NODELAY;
 907         }
 908         spin_unlock(&ci->i_ceph_lock);
 909         return drop;
 910 }
 911
 912 /*
 913  * rmdir and unlink are differ only by the metadata op code
 914  */
 915 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
 916 {
 917         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 918         struct ceph_mds_client *mdsc = fsc->mdsc;
 919         struct inode *inode = d_inode(dentry);
 920         struct ceph_mds_request *req;
 921         int err = -EROFS;
 922         int op;
 923
 924         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 925                 /* rmdir .snap/foo is RMSNAP */
 926                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
 927                 op = CEPH_MDS_OP_RMSNAP;
 928         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 929                 dout("unlink/rmdir dir %p dn %p inode %p\n",
 930                      dir, dentry, inode);
 931                 op = d_is_dir(dentry) ?
 932                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
 933         } else
 934                 goto out;
 935         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 936         if (IS_ERR(req)) {
 937                 err = PTR_ERR(req);
 938                 goto out;
 939         }
 940         req->r_dentry = dget(dentry);
 941         req->r_num_caps = 2;
 942         req->r_locked_dir = dir;
 943         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 944         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 945         req->r_inode_drop = drop_caps_for_unlink(inode);
 946         err = ceph_mdsc_do_request(mdsc, dir, req);
 947         if (!err && !req->r_reply_info.head->is_dentry)
 948                 d_delete(dentry);
 949         ceph_mdsc_put_request(req);
 950 out:
 951         return err;
 952 }
 953
 954 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
 955                        struct inode *new_dir, struct dentry *new_dentry)
 956 {
 957         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
 958         struct ceph_mds_client *mdsc = fsc->mdsc;
 959         struct ceph_mds_request *req;
 960         int op = CEPH_MDS_OP_RENAME;
 961         int err;
 962
 963         if (ceph_snap(old_dir) != ceph_snap(new_dir))
 964                 return -EXDEV;
 965         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
 966                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
 967                         op = CEPH_MDS_OP_RENAMESNAP;
 968                 else
 969                         return -EROFS;
 970         }
 971         dout("rename dir %p dentry %p to dir %p dentry %p\n",
 972              old_dir, old_dentry, new_dir, new_dentry);
 973         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 974         if (IS_ERR(req))
 975                 return PTR_ERR(req);
 976         ihold(old_dir);
 977         req->r_dentry = dget(new_dentry);
 978         req->r_num_caps = 2;
 979         req->r_old_dentry = dget(old_dentry);
 980         req->r_old_dentry_dir = old_dir;
 981         req->r_locked_dir = new_dir;
 982         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
 983         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
 984         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 985         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 986         /* release LINK_RDCACHE on source inode (mds will lock it) */
 987         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
 988         if (d_really_is_positive(new_dentry))
 989                 req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry));
 990         err = ceph_mdsc_do_request(mdsc, old_dir, req);
 991         if (!err && !req->r_reply_info.head->is_dentry) {
 992                 /*
 993                  * Normally d_move() is done by fill_trace (called by
 994                  * do_request, above).  If there is no trace, we need
 995                  * to do it here.
 996                  */
 997
 998                 /* d_move screws up sibling dentries' offsets */
 999                 ceph_dir_clear_complete(old_dir);
1000                 ceph_dir_clear_complete(new_dir);
1001
1002                 d_move(old_dentry, new_dentry);
1003
1004                 /* ensure target dentry is invalidated, despite
1005                    rehashing bug in vfs_rename_dir */
1006                 ceph_invalidate_dentry_lease(new_dentry);
1007         }
1008         ceph_mdsc_put_request(req);
1009         return err;
1010 }
1011
1012 /*
1013  * Ensure a dentry lease will no longer revalidate.
1014  */
1015 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1016 {
1017         spin_lock(&dentry->d_lock);
1018         dentry->d_time = jiffies;
1019         ceph_dentry(dentry)->lease_shared_gen = 0;
1020         spin_unlock(&dentry->d_lock);
1021 }
1022
1023 /*
1024  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1025  * renew if the least is more than half up.
1026  */
1027 static int dentry_lease_is_valid(struct dentry *dentry)
1028 {
1029         struct ceph_dentry_info *di;
1030         struct ceph_mds_session *s;
1031         int valid = 0;
1032         u32 gen;
1033         unsigned long ttl;
1034         struct ceph_mds_session *session = NULL;
1035         struct inode *dir = NULL;
1036         u32 seq = 0;
1037
1038         spin_lock(&dentry->d_lock);
1039         di = ceph_dentry(dentry);
1040         if (di->lease_session) {
1041                 s = di->lease_session;
1042                 spin_lock(&s->s_gen_ttl_lock);
1043                 gen = s->s_cap_gen;
1044                 ttl = s->s_cap_ttl;
1045                 spin_unlock(&s->s_gen_ttl_lock);
1046
1047                 if (di->lease_gen == gen &&
1048                     time_before(jiffies, dentry->d_time) &&
1049                     time_before(jiffies, ttl)) {
1050                         valid = 1;
1051                         if (di->lease_renew_after &&
1052                             time_after(jiffies, di->lease_renew_after)) {
1053                                 /* we should renew */
1054                                 dir = d_inode(dentry->d_parent);
1055                                 session = ceph_get_mds_session(s);
1056                                 seq = di->lease_seq;
1057                                 di->lease_renew_after = 0;
1058                                 di->lease_renew_from = jiffies;
1059                         }
1060                 }
1061         }
1062         spin_unlock(&dentry->d_lock);
1063
1064         if (session) {
1065                 ceph_mdsc_lease_send_msg(session, dir, dentry,
1066                                          CEPH_MDS_LEASE_RENEW, seq);
1067                 ceph_put_mds_session(session);
1068         }
1069         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1070         return valid;
1071 }
1072
1073 /*
1074  * Check if directory-wide content lease/cap is valid.
1075  */
1076 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1077 {
1078         struct ceph_inode_info *ci = ceph_inode(dir);
1079         struct ceph_dentry_info *di = ceph_dentry(dentry);
1080         int valid = 0;
1081
1082         spin_lock(&ci->i_ceph_lock);
1083         if (ci->i_shared_gen == di->lease_shared_gen)
1084                 valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1085         spin_unlock(&ci->i_ceph_lock);
1086         dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1087              dir, (unsigned)ci->i_shared_gen, dentry,
1088              (unsigned)di->lease_shared_gen, valid);
1089         return valid;
1090 }
1091
1092 /*
1093  * Check if cached dentry can be trusted.
1094  */
1095 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1096 {
1097         int valid = 0;
1098         struct inode *dir;
1099
1100         if (flags & LOOKUP_RCU)
1101                 return -ECHILD;
1102
1103         dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1104              dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1105
1106         dir = ceph_get_dentry_parent_inode(dentry);
1107
1108         /* always trust cached snapped dentries, snapdir dentry */
1109         if (ceph_snap(dir) != CEPH_NOSNAP) {
1110                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1111                      dentry, d_inode(dentry));
1112                 valid = 1;
1113         } else if (d_really_is_positive(dentry) &&
1114                    ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1115                 valid = 1;
1116         } else if (dentry_lease_is_valid(dentry) ||
1117                    dir_lease_is_valid(dir, dentry)) {
1118                 if (d_really_is_positive(dentry))
1119                         valid = ceph_is_any_caps(d_inode(dentry));
1120                 else
1121                         valid = 1;
1122         }
1123
1124         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1125         if (valid) {
1126                 ceph_dentry_lru_touch(dentry);
1127         } else {
1128                 ceph_dir_clear_complete(dir);
1129         }
1130         iput(dir);
1131         return valid;
1132 }
1133
1134 /*
1135  * Release our ceph_dentry_info.
1136  */
1137 static void ceph_d_release(struct dentry *dentry)
1138 {
1139         struct ceph_dentry_info *di = ceph_dentry(dentry);
1140
1141         dout("d_release %p\n", dentry);
1142         ceph_dentry_lru_del(dentry);
1143         if (di->lease_session)
1144                 ceph_put_mds_session(di->lease_session);
1145         kmem_cache_free(ceph_dentry_cachep, di);
1146         dentry->d_fsdata = NULL;
1147 }
1148
1149 static int ceph_snapdir_d_revalidate(struct dentry *dentry,
1150                                           unsigned int flags)
1151 {
1152         /*
1153          * Eventually, we'll want to revalidate snapped metadata
1154          * too... probably...
1155          */
1156         return 1;
1157 }
1158
1159 /*
1160  * When the VFS prunes a dentry from the cache, we need to clear the
1161  * complete flag on the parent directory.
1162  *
1163  * Called under dentry->d_lock.
1164  */
1165 static void ceph_d_prune(struct dentry *dentry)
1166 {
1167         dout("ceph_d_prune %p\n", dentry);
1168
1169         /* do we have a valid parent? */
1170         if (IS_ROOT(dentry))
1171                 return;
1172
1173         /* if we are not hashed, we don't affect dir's completeness */
1174         if (d_unhashed(dentry))
1175                 return;
1176
1177         /*
1178          * we hold d_lock, so d_parent is stable, and d_fsdata is never
1179          * cleared until d_release
1180          */
1181         ceph_dir_clear_complete(d_inode(dentry->d_parent));
1182 }
1183
1184 /*
1185  * read() on a dir.  This weird interface hack only works if mounted
1186  * with '-o dirstat'.
1187  */
1188 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1189                              loff_t *ppos)
1190 {
1191         struct ceph_file_info *cf = file->private_data;
1192         struct inode *inode = file_inode(file);
1193         struct ceph_inode_info *ci = ceph_inode(inode);
1194         int left;
1195         const int bufsize = 1024;
1196
1197         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1198                 return -EISDIR;
1199
1200         if (!cf->dir_info) {
1201                 cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1202                 if (!cf->dir_info)
1203                         return -ENOMEM;
1204                 cf->dir_info_len =
1205                         snprintf(cf->dir_info, bufsize,
1206                                 "entries:   %20lld\n"
1207                                 " files:    %20lld\n"
1208                                 " subdirs:  %20lld\n"
1209                                 "rentries:  %20lld\n"
1210                                 " rfiles:   %20lld\n"
1211                                 " rsubdirs: %20lld\n"
1212                                 "rbytes:    %20lld\n"
1213                                 "rctime:    %10ld.%09ld\n",
1214                                 ci->i_files + ci->i_subdirs,
1215                                 ci->i_files,
1216                                 ci->i_subdirs,
1217                                 ci->i_rfiles + ci->i_rsubdirs,
1218                                 ci->i_rfiles,
1219                                 ci->i_rsubdirs,
1220                                 ci->i_rbytes,
1221                                 (long)ci->i_rctime.tv_sec,
1222                                 (long)ci->i_rctime.tv_nsec);
1223         }
1224
1225         if (*ppos >= cf->dir_info_len)
1226                 return 0;
1227         size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1228         left = copy_to_user(buf, cf->dir_info + *ppos, size);
1229         if (left == size)
1230                 return -EFAULT;
1231         *ppos += (size - left);
1232         return size - left;
1233 }
1234
1235 /*
1236  * We maintain a private dentry LRU.
1237  *
1238  * FIXME: this needs to be changed to a per-mds lru to be useful.
1239  */
1240 void ceph_dentry_lru_add(struct dentry *dn)
1241 {
1242         struct ceph_dentry_info *di = ceph_dentry(dn);
1243         struct ceph_mds_client *mdsc;
1244
1245         dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1246         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1247         spin_lock(&mdsc->dentry_lru_lock);
1248         list_add_tail(&di->lru, &mdsc->dentry_lru);
1249         mdsc->num_dentry++;
1250         spin_unlock(&mdsc->dentry_lru_lock);
1251 }
1252
1253 void ceph_dentry_lru_touch(struct dentry *dn)
1254 {
1255         struct ceph_dentry_info *di = ceph_dentry(dn);
1256         struct ceph_mds_client *mdsc;
1257
1258         dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1259              di->offset);
1260         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1261         spin_lock(&mdsc->dentry_lru_lock);
1262         list_move_tail(&di->lru, &mdsc->dentry_lru);
1263         spin_unlock(&mdsc->dentry_lru_lock);
1264 }
1265
1266 void ceph_dentry_lru_del(struct dentry *dn)
1267 {
1268         struct ceph_dentry_info *di = ceph_dentry(dn);
1269         struct ceph_mds_client *mdsc;
1270
1271         dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1272         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1273         spin_lock(&mdsc->dentry_lru_lock);
1274         list_del_init(&di->lru);
1275         mdsc->num_dentry--;
1276         spin_unlock(&mdsc->dentry_lru_lock);
1277 }
1278
1279 /*
1280  * Return name hash for a given dentry.  This is dependent on
1281  * the parent directory's hash function.
1282  */
1283 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1284 {
1285         struct ceph_inode_info *dci = ceph_inode(dir);
1286
1287         switch (dci->i_dir_layout.dl_dir_hash) {
1288         case 0: /* for backward compat */
1289         case CEPH_STR_HASH_LINUX:
1290                 return dn->d_name.hash;
1291
1292         default:
1293                 return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1294                                      dn->d_name.name, dn->d_name.len);
1295         }
1296 }
1297
1298 const struct file_operations ceph_dir_fops = {
1299         .read = ceph_read_dir,
1300         .iterate = ceph_readdir,
1301         .llseek = ceph_dir_llseek,
1302         .open = ceph_open,
1303         .release = ceph_release,
1304         .unlocked_ioctl = ceph_ioctl,
1305         .fsync = ceph_fsync,
1306 };
1307
1308 const struct file_operations ceph_snapdir_fops = {
1309         .iterate = ceph_readdir,
1310         .llseek = ceph_dir_llseek,
1311         .open = ceph_open,
1312         .release = ceph_release,
1313 };
1314
1315 const struct inode_operations ceph_dir_iops = {
1316         .lookup = ceph_lookup,
1317         .permission = ceph_permission,
1318         .getattr = ceph_getattr,
1319         .setattr = ceph_setattr,
1320         .setxattr = ceph_setxattr,
1321         .getxattr = ceph_getxattr,
1322         .listxattr = ceph_listxattr,
1323         .removexattr = ceph_removexattr,
1324         .get_acl = ceph_get_acl,
1325         .set_acl = ceph_set_acl,
1326         .mknod = ceph_mknod,
1327         .symlink = ceph_symlink,
1328         .mkdir = ceph_mkdir,
1329         .link = ceph_link,
1330         .unlink = ceph_unlink,
1331         .rmdir = ceph_unlink,
1332         .rename = ceph_rename,
1333         .create = ceph_create,
1334         .atomic_open = ceph_atomic_open,
1335 };
1336
1337 const struct inode_operations ceph_snapdir_iops = {
1338         .lookup = ceph_lookup,
1339         .permission = ceph_permission,
1340         .getattr = ceph_getattr,
1341         .mkdir = ceph_mkdir,
1342         .rmdir = ceph_unlink,
1343         .rename = ceph_rename,
1344 };
1345
1346 const struct dentry_operations ceph_dentry_ops = {
1347         .d_revalidate = ceph_d_revalidate,
1348         .d_release = ceph_d_release,
1349         .d_prune = ceph_d_prune,
1350 };
1351
1352 const struct dentry_operations ceph_snapdir_dentry_ops = {
1353         .d_revalidate = ceph_snapdir_d_revalidate,
1354         .d_release = ceph_d_release,
1355 };
1356
1357 const struct dentry_operations ceph_snap_dentry_ops = {
1358         .d_release = ceph_d_release,
1359         .d_prune = ceph_d_prune,
1360 };