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 #include <linux/xattr.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12
  13 /*
  14  * Directory operations: readdir, lookup, create, link, unlink,
  15  * rename, etc.
  16  */
  17
  18 /*
  19  * Ceph MDS operations are specified in terms of a base ino and
  20  * relative path.  Thus, the client can specify an operation on a
  21  * specific inode (e.g., a getattr due to fstat(2)), or as a path
  22  * relative to, say, the root directory.
  23  *
  24  * Normally, we limit ourselves to strict inode ops (no path component)
  25  * or dentry operations (a single path component relative to an ino).  The
  26  * exception to this is open_root_dentry(), which will open the mount
  27  * point by name.
  28  */
  29
  30 const struct dentry_operations ceph_dentry_ops;
  31
  32 /*
  33  * Initialize ceph dentry state.
  34  */
  35 static int ceph_d_init(struct dentry *dentry)
  36 {
  37         struct ceph_dentry_info *di;
  38
  39         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  40         if (!di)
  41                 return -ENOMEM;          /* oh well */
  42
  43         di->dentry = dentry;
  44         di->lease_session = NULL;
  45         di->time = jiffies;
  46         dentry->d_fsdata = di;
  47         ceph_dentry_lru_add(dentry);
  48         return 0;
  49 }
  50
  51 /*
  52  * for f_pos for readdir:
  53  * - hash order:
  54  *      (0xff << 52) | ((24 bits hash) << 28) |
  55  *      (the nth entry has hash collision);
  56  * - frag+name order;
  57  *      ((frag value) << 28) | (the nth entry in frag);
  58  */
  59 #define OFFSET_BITS     28
  60 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  61 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  62 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  63 {
  64         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  65         if (hash_order)
  66                 fpos |= HASH_ORDER;
  67         return fpos;
  68 }
  69
  70 static bool is_hash_order(loff_t p)
  71 {
  72         return (p & HASH_ORDER) == HASH_ORDER;
  73 }
  74
  75 static unsigned fpos_frag(loff_t p)
  76 {
  77         return p >> OFFSET_BITS;
  78 }
  79
  80 static unsigned fpos_hash(loff_t p)
  81 {
  82         return ceph_frag_value(fpos_frag(p));
  83 }
  84
  85 static unsigned fpos_off(loff_t p)
  86 {
  87         return p & OFFSET_MASK;
  88 }
  89
  90 static int fpos_cmp(loff_t l, loff_t r)
  91 {
  92         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
  93         if (v)
  94                 return v;
  95         return (int)(fpos_off(l) - fpos_off(r));
  96 }
  97
  98 /*
  99  * make note of the last dentry we read, so we can
 100  * continue at the same lexicographical point,
 101  * regardless of what dir changes take place on the
 102  * server.
 103  */
 104 static int note_last_dentry(struct ceph_file_info *fi, const char *name,
 105                             int len, unsigned next_offset)
 106 {
 107         char *buf = kmalloc(len+1, GFP_KERNEL);
 108         if (!buf)
 109                 return -ENOMEM;
 110         kfree(fi->last_name);
 111         fi->last_name = buf;
 112         memcpy(fi->last_name, name, len);
 113         fi->last_name[len] = 0;
 114         fi->next_offset = next_offset;
 115         dout("note_last_dentry '%s'\n", fi->last_name);
 116         return 0;
 117 }
 118
 119
 120 static struct dentry *
 121 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 122                         struct ceph_readdir_cache_control *cache_ctl)
 123 {
 124         struct inode *dir = d_inode(parent);
 125         struct dentry *dentry;
 126         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 127         loff_t ptr_pos = idx * sizeof(struct dentry *);
 128         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 129
 130         if (ptr_pos >= i_size_read(dir))
 131                 return NULL;
 132
 133         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 134                 ceph_readdir_cache_release(cache_ctl);
 135                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 136                 if (!cache_ctl->page) {
 137                         dout(" page %lu not found\n", ptr_pgoff);
 138                         return ERR_PTR(-EAGAIN);
 139                 }
 140                 /* reading/filling the cache are serialized by
 141                    i_mutex, no need to use page lock */
 142                 unlock_page(cache_ctl->page);
 143                 cache_ctl->dentries = kmap(cache_ctl->page);
 144         }
 145
 146         cache_ctl->index = idx & idx_mask;
 147
 148         rcu_read_lock();
 149         spin_lock(&parent->d_lock);
 150         /* check i_size again here, because empty directory can be
 151          * marked as complete while not holding the i_mutex. */
 152         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 153                 dentry = cache_ctl->dentries[cache_ctl->index];
 154         else
 155                 dentry = NULL;
 156         spin_unlock(&parent->d_lock);
 157         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 158                 dentry = NULL;
 159         rcu_read_unlock();
 160         return dentry ? : ERR_PTR(-EAGAIN);
 161 }
 162
 163 /*
 164  * When possible, we try to satisfy a readdir by peeking at the
 165  * dcache.  We make this work by carefully ordering dentries on
 166  * d_child when we initially get results back from the MDS, and
 167  * falling back to a "normal" sync readdir if any dentries in the dir
 168  * are dropped.
 169  *
 170  * Complete dir indicates that we have all dentries in the dir.  It is
 171  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 172  * the MDS if/when the directory is modified).
 173  */
 174 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 175                             u32 shared_gen)
 176 {
 177         struct ceph_file_info *fi = file->private_data;
 178         struct dentry *parent = file->f_path.dentry;
 179         struct inode *dir = d_inode(parent);
 180         struct dentry *dentry, *last = NULL;
 181         struct ceph_dentry_info *di;
 182         struct ceph_readdir_cache_control cache_ctl = {};
 183         u64 idx = 0;
 184         int err = 0;
 185
 186         dout("__dcache_readdir %p v%u at %llx\n", dir, shared_gen, ctx->pos);
 187
 188         /* search start position */
 189         if (ctx->pos > 2) {
 190                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 191                 while (count > 0) {
 192                         u64 step = count >> 1;
 193                         dentry = __dcache_find_get_entry(parent, idx + step,
 194                                                          &cache_ctl);
 195                         if (!dentry) {
 196                                 /* use linar search */
 197                                 idx = 0;
 198                                 break;
 199                         }
 200                         if (IS_ERR(dentry)) {
 201                                 err = PTR_ERR(dentry);
 202                                 goto out;
 203                         }
 204                         di = ceph_dentry(dentry);
 205                         spin_lock(&dentry->d_lock);
 206                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 207                                 idx += step + 1;
 208                                 count -= step + 1;
 209                         } else {
 210                                 count = step;
 211                         }
 212                         spin_unlock(&dentry->d_lock);
 213                         dput(dentry);
 214                 }
 215
 216                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 217         }
 218
 219
 220         for (;;) {
 221                 bool emit_dentry = false;
 222                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 223                 if (!dentry) {
 224                         fi->flags |= CEPH_F_ATEND;
 225                         err = 0;
 226                         break;
 227                 }
 228                 if (IS_ERR(dentry)) {
 229                         err = PTR_ERR(dentry);
 230                         goto out;
 231                 }
 232
 233                 di = ceph_dentry(dentry);
 234                 spin_lock(&dentry->d_lock);
 235                 if (di->lease_shared_gen == shared_gen &&
 236                     d_really_is_positive(dentry) &&
 237                     fpos_cmp(ctx->pos, di->offset) <= 0) {
 238                         emit_dentry = true;
 239                 }
 240                 spin_unlock(&dentry->d_lock);
 241
 242                 if (emit_dentry) {
 243                         dout(" %llx dentry %p %pd %p\n", di->offset,
 244                              dentry, dentry, d_inode(dentry));
 245                         ctx->pos = di->offset;
 246                         if (!dir_emit(ctx, dentry->d_name.name,
 247                                       dentry->d_name.len,
 248                                       ceph_translate_ino(dentry->d_sb,
 249                                                          d_inode(dentry)->i_ino),
 250                                       d_inode(dentry)->i_mode >> 12)) {
 251                                 dput(dentry);
 252                                 err = 0;
 253                                 break;
 254                         }
 255                         ctx->pos++;
 256
 257                         if (last)
 258                                 dput(last);
 259                         last = dentry;
 260                 } else {
 261                         dput(dentry);
 262                 }
 263         }
 264 out:
 265         ceph_readdir_cache_release(&cache_ctl);
 266         if (last) {
 267                 int ret;
 268                 di = ceph_dentry(last);
 269                 ret = note_last_dentry(fi, last->d_name.name, last->d_name.len,
 270                                        fpos_off(di->offset) + 1);
 271                 if (ret < 0)
 272                         err = ret;
 273                 dput(last);
 274         }
 275         return err;
 276 }
 277
 278 static bool need_send_readdir(struct ceph_file_info *fi, loff_t pos)
 279 {
 280         if (!fi->last_readdir)
 281                 return true;
 282         if (is_hash_order(pos))
 283                 return !ceph_frag_contains_value(fi->frag, fpos_hash(pos));
 284         else
 285                 return fi->frag != fpos_frag(pos);
 286 }
 287
 288 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 289 {
 290         struct ceph_file_info *fi = file->private_data;
 291         struct inode *inode = file_inode(file);
 292         struct ceph_inode_info *ci = ceph_inode(inode);
 293         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 294         struct ceph_mds_client *mdsc = fsc->mdsc;
 295         int i;
 296         int err;
 297         u32 ftype;
 298         struct ceph_mds_reply_info_parsed *rinfo;
 299
 300         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 301         if (fi->flags & CEPH_F_ATEND)
 302                 return 0;
 303
 304         /* always start with . and .. */
 305         if (ctx->pos == 0) {
 306                 dout("readdir off 0 -> '.'\n");
 307                 if (!dir_emit(ctx, ".", 1,
 308                             ceph_translate_ino(inode->i_sb, inode->i_ino),
 309                             inode->i_mode >> 12))
 310                         return 0;
 311                 ctx->pos = 1;
 312         }
 313         if (ctx->pos == 1) {
 314                 ino_t ino = parent_ino(file->f_path.dentry);
 315                 dout("readdir off 1 -> '..'\n");
 316                 if (!dir_emit(ctx, "..", 2,
 317                             ceph_translate_ino(inode->i_sb, ino),
 318                             inode->i_mode >> 12))
 319                         return 0;
 320                 ctx->pos = 2;
 321         }
 322
 323         /* can we use the dcache? */
 324         spin_lock(&ci->i_ceph_lock);
 325         if (ceph_test_mount_opt(fsc, DCACHE) &&
 326             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 327             ceph_snap(inode) != CEPH_SNAPDIR &&
 328             __ceph_dir_is_complete_ordered(ci) &&
 329             __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
 330                 u32 shared_gen = ci->i_shared_gen;
 331                 spin_unlock(&ci->i_ceph_lock);
 332                 err = __dcache_readdir(file, ctx, shared_gen);
 333                 if (err != -EAGAIN)
 334                         return err;
 335         } else {
 336                 spin_unlock(&ci->i_ceph_lock);
 337         }
 338
 339         /* proceed with a normal readdir */
 340 more:
 341         /* do we have the correct frag content buffered? */
 342         if (need_send_readdir(fi, ctx->pos)) {
 343                 struct ceph_mds_request *req;
 344                 unsigned frag;
 345                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 346                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 347
 348                 /* discard old result, if any */
 349                 if (fi->last_readdir) {
 350                         ceph_mdsc_put_request(fi->last_readdir);
 351                         fi->last_readdir = NULL;
 352                 }
 353
 354                 if (is_hash_order(ctx->pos)) {
 355                         frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 356                                                 NULL, NULL);
 357                 } else {
 358                         frag = fpos_frag(ctx->pos);
 359                 }
 360
 361                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 362                      ceph_vinop(inode), frag, fi->last_name);
 363                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 364                 if (IS_ERR(req))
 365                         return PTR_ERR(req);
 366                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 367                 if (err) {
 368                         ceph_mdsc_put_request(req);
 369                         return err;
 370                 }
 371                 /* hints to request -> mds selection code */
 372                 req->r_direct_mode = USE_AUTH_MDS;
 373                 req->r_direct_hash = ceph_frag_value(frag);
 374                 __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 375                 if (fi->last_name) {
 376                         req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL);
 377                         if (!req->r_path2) {
 378                                 ceph_mdsc_put_request(req);
 379                                 return -ENOMEM;
 380                         }
 381                 } else if (is_hash_order(ctx->pos)) {
 382                         req->r_args.readdir.offset_hash =
 383                                 cpu_to_le32(fpos_hash(ctx->pos));
 384                 }
 385
 386                 req->r_dir_release_cnt = fi->dir_release_count;
 387                 req->r_dir_ordered_cnt = fi->dir_ordered_count;
 388                 req->r_readdir_cache_idx = fi->readdir_cache_idx;
 389                 req->r_readdir_offset = fi->next_offset;
 390                 req->r_args.readdir.frag = cpu_to_le32(frag);
 391                 req->r_args.readdir.flags =
 392                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 393
 394                 req->r_inode = inode;
 395                 ihold(inode);
 396                 req->r_dentry = dget(file->f_path.dentry);
 397                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 398                 if (err < 0) {
 399                         ceph_mdsc_put_request(req);
 400                         return err;
 401                 }
 402                 dout("readdir got and parsed readdir result=%d on "
 403                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 404                      err, frag,
 405                      (int)req->r_reply_info.dir_end,
 406                      (int)req->r_reply_info.dir_complete,
 407                      (int)req->r_reply_info.hash_order);
 408
 409                 rinfo = &req->r_reply_info;
 410                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 411                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 412                         if (!rinfo->hash_order) {
 413                                 fi->next_offset = req->r_readdir_offset;
 414                                 /* adjust ctx->pos to beginning of frag */
 415                                 ctx->pos = ceph_make_fpos(frag,
 416                                                           fi->next_offset,
 417                                                           false);
 418                         }
 419                 }
 420
 421                 fi->frag = frag;
 422                 fi->last_readdir = req;
 423
 424                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 425                         fi->readdir_cache_idx = req->r_readdir_cache_idx;
 426                         if (fi->readdir_cache_idx < 0) {
 427                                 /* preclude from marking dir ordered */
 428                                 fi->dir_ordered_count = 0;
 429                         } else if (ceph_frag_is_leftmost(frag) &&
 430                                    fi->next_offset == 2) {
 431                                 /* note dir version at start of readdir so
 432                                  * we can tell if any dentries get dropped */
 433                                 fi->dir_release_count = req->r_dir_release_cnt;
 434                                 fi->dir_ordered_count = req->r_dir_ordered_cnt;
 435                         }
 436                 } else {
 437                         dout("readdir !did_prepopulate");
 438                         /* disable readdir cache */
 439                         fi->readdir_cache_idx = -1;
 440                         /* preclude from marking dir complete */
 441                         fi->dir_release_count = 0;
 442                 }
 443
 444                 /* note next offset and last dentry name */
 445                 if (rinfo->dir_nr > 0) {
 446                         struct ceph_mds_reply_dir_entry *rde =
 447                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 448                         unsigned next_offset = req->r_reply_info.dir_end ?
 449                                         2 : (fpos_off(rde->offset) + 1);
 450                         err = note_last_dentry(fi, rde->name, rde->name_len,
 451                                                next_offset);
 452                         if (err)
 453                                 return err;
 454                 } else if (req->r_reply_info.dir_end) {
 455                         fi->next_offset = 2;
 456                         /* keep last name */
 457                 }
 458         }
 459
 460         rinfo = &fi->last_readdir->r_reply_info;
 461         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 462              fi->frag, rinfo->dir_nr, ctx->pos,
 463              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 464
 465         i = 0;
 466         /* search start position */
 467         if (rinfo->dir_nr > 0) {
 468                 int step, nr = rinfo->dir_nr;
 469                 while (nr > 0) {
 470                         step = nr >> 1;
 471                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 472                                 i +=  step + 1;
 473                                 nr -= step + 1;
 474                         } else {
 475                                 nr = step;
 476                         }
 477                 }
 478         }
 479         for (; i < rinfo->dir_nr; i++) {
 480                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 481                 struct ceph_vino vino;
 482                 ino_t ino;
 483
 484                 BUG_ON(rde->offset < ctx->pos);
 485
 486                 ctx->pos = rde->offset;
 487                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 488                      i, rinfo->dir_nr, ctx->pos,
 489                      rde->name_len, rde->name, &rde->inode.in);
 490
 491                 BUG_ON(!rde->inode.in);
 492                 ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
 493                 vino.ino = le64_to_cpu(rde->inode.in->ino);
 494                 vino.snap = le64_to_cpu(rde->inode.in->snapid);
 495                 ino = ceph_vino_to_ino(vino);
 496
 497                 if (!dir_emit(ctx, rde->name, rde->name_len,
 498                               ceph_translate_ino(inode->i_sb, ino), ftype)) {
 499                         dout("filldir stopping us...\n");
 500                         return 0;
 501                 }
 502                 ctx->pos++;
 503         }
 504
 505         if (fi->next_offset > 2) {
 506                 ceph_mdsc_put_request(fi->last_readdir);
 507                 fi->last_readdir = NULL;
 508                 goto more;
 509         }
 510
 511         /* more frags? */
 512         if (!ceph_frag_is_rightmost(fi->frag)) {
 513                 unsigned frag = ceph_frag_next(fi->frag);
 514                 if (is_hash_order(ctx->pos)) {
 515                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 516                                                         fi->next_offset, true);
 517                         if (new_pos > ctx->pos)
 518                                 ctx->pos = new_pos;
 519                         /* keep last_name */
 520                 } else {
 521                         ctx->pos = ceph_make_fpos(frag, fi->next_offset, false);
 522                         kfree(fi->last_name);
 523                         fi->last_name = NULL;
 524                 }
 525                 dout("readdir next frag is %x\n", frag);
 526                 goto more;
 527         }
 528         fi->flags |= CEPH_F_ATEND;
 529
 530         /*
 531          * if dir_release_count still matches the dir, no dentries
 532          * were released during the whole readdir, and we should have
 533          * the complete dir contents in our cache.
 534          */
 535         if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) {
 536                 spin_lock(&ci->i_ceph_lock);
 537                 if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) {
 538                         dout(" marking %p complete and ordered\n", inode);
 539                         /* use i_size to track number of entries in
 540                          * readdir cache */
 541                         BUG_ON(fi->readdir_cache_idx < 0);
 542                         i_size_write(inode, fi->readdir_cache_idx *
 543                                      sizeof(struct dentry*));
 544                 } else {
 545                         dout(" marking %p complete\n", inode);
 546                 }
 547                 __ceph_dir_set_complete(ci, fi->dir_release_count,
 548                                         fi->dir_ordered_count);
 549                 spin_unlock(&ci->i_ceph_lock);
 550         }
 551
 552         dout("readdir %p file %p done.\n", inode, file);
 553         return 0;
 554 }
 555
 556 static void reset_readdir(struct ceph_file_info *fi)
 557 {
 558         if (fi->last_readdir) {
 559                 ceph_mdsc_put_request(fi->last_readdir);
 560                 fi->last_readdir = NULL;
 561         }
 562         kfree(fi->last_name);
 563         fi->last_name = NULL;
 564         fi->dir_release_count = 0;
 565         fi->readdir_cache_idx = -1;
 566         fi->next_offset = 2;  /* compensate for . and .. */
 567         fi->flags &= ~CEPH_F_ATEND;
 568 }
 569
 570 /*
 571  * discard buffered readdir content on seekdir(0), or seek to new frag,
 572  * or seek prior to current chunk
 573  */
 574 static bool need_reset_readdir(struct ceph_file_info *fi, loff_t new_pos)
 575 {
 576         struct ceph_mds_reply_info_parsed *rinfo;
 577         loff_t chunk_offset;
 578         if (new_pos == 0)
 579                 return true;
 580         if (is_hash_order(new_pos)) {
 581                 /* no need to reset last_name for a forward seek when
 582                  * dentries are sotred in hash order */
 583         } else if (fi->frag != fpos_frag(new_pos)) {
 584                 return true;
 585         }
 586         rinfo = fi->last_readdir ? &fi->last_readdir->r_reply_info : NULL;
 587         if (!rinfo || !rinfo->dir_nr)
 588                 return true;
 589         chunk_offset = rinfo->dir_entries[0].offset;
 590         return new_pos < chunk_offset ||
 591                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 592 }
 593
 594 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 595 {
 596         struct ceph_file_info *fi = file->private_data;
 597         struct inode *inode = file->f_mapping->host;
 598         loff_t retval;
 599
 600         inode_lock(inode);
 601         retval = -EINVAL;
 602         switch (whence) {
 603         case SEEK_CUR:
 604                 offset += file->f_pos;
 605         case SEEK_SET:
 606                 break;
 607         case SEEK_END:
 608                 retval = -EOPNOTSUPP;
 609         default:
 610                 goto out;
 611         }
 612
 613         if (offset >= 0) {
 614                 if (need_reset_readdir(fi, offset)) {
 615                         dout("dir_llseek dropping %p content\n", file);
 616                         reset_readdir(fi);
 617                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 618                         /* for hash offset, we don't know if a forward seek
 619                          * is within same frag */
 620                         fi->dir_release_count = 0;
 621                         fi->readdir_cache_idx = -1;
 622                 }
 623
 624                 if (offset != file->f_pos) {
 625                         file->f_pos = offset;
 626                         file->f_version = 0;
 627                         fi->flags &= ~CEPH_F_ATEND;
 628                 }
 629                 retval = offset;
 630         }
 631 out:
 632         inode_unlock(inode);
 633         return retval;
 634 }
 635
 636 /*
 637  * Handle lookups for the hidden .snap directory.
 638  */
 639 int ceph_handle_snapdir(struct ceph_mds_request *req,
 640                         struct dentry *dentry, int err)
 641 {
 642         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 643         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 644
 645         /* .snap dir? */
 646         if (err == -ENOENT &&
 647             ceph_snap(parent) == CEPH_NOSNAP &&
 648             strcmp(dentry->d_name.name,
 649                    fsc->mount_options->snapdir_name) == 0) {
 650                 struct inode *inode = ceph_get_snapdir(parent);
 651                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 652                      dentry, dentry, inode);
 653                 BUG_ON(!d_unhashed(dentry));
 654                 d_add(dentry, inode);
 655                 err = 0;
 656         }
 657         return err;
 658 }
 659
 660 /*
 661  * Figure out final result of a lookup/open request.
 662  *
 663  * Mainly, make sure we return the final req->r_dentry (if it already
 664  * existed) in place of the original VFS-provided dentry when they
 665  * differ.
 666  *
 667  * Gracefully handle the case where the MDS replies with -ENOENT and
 668  * no trace (which it may do, at its discretion, e.g., if it doesn't
 669  * care to issue a lease on the negative dentry).
 670  */
 671 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 672                                   struct dentry *dentry, int err)
 673 {
 674         if (err == -ENOENT) {
 675                 /* no trace? */
 676                 err = 0;
 677                 if (!req->r_reply_info.head->is_dentry) {
 678                         dout("ENOENT and no trace, dentry %p inode %p\n",
 679                              dentry, d_inode(dentry));
 680                         if (d_really_is_positive(dentry)) {
 681                                 d_drop(dentry);
 682                                 err = -ENOENT;
 683                         } else {
 684                                 d_add(dentry, NULL);
 685                         }
 686                 }
 687         }
 688         if (err)
 689                 dentry = ERR_PTR(err);
 690         else if (dentry != req->r_dentry)
 691                 dentry = dget(req->r_dentry);   /* we got spliced */
 692         else
 693                 dentry = NULL;
 694         return dentry;
 695 }
 696
 697 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 698 {
 699         return ceph_ino(inode) == CEPH_INO_ROOT &&
 700                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 701 }
 702
 703 /*
 704  * Look up a single dir entry.  If there is a lookup intent, inform
 705  * the MDS so that it gets our 'caps wanted' value in a single op.
 706  */
 707 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 708                                   unsigned int flags)
 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         int op;
 714         int mask;
 715         int err;
 716
 717         dout("lookup %p dentry %p '%pd'\n",
 718              dir, dentry, dentry);
 719
 720         if (dentry->d_name.len > NAME_MAX)
 721                 return ERR_PTR(-ENAMETOOLONG);
 722
 723         /* can we conclude ENOENT locally? */
 724         if (d_really_is_negative(dentry)) {
 725                 struct ceph_inode_info *ci = ceph_inode(dir);
 726                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 727
 728                 spin_lock(&ci->i_ceph_lock);
 729                 dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
 730                 if (strncmp(dentry->d_name.name,
 731                             fsc->mount_options->snapdir_name,
 732                             dentry->d_name.len) &&
 733                     !is_root_ceph_dentry(dir, dentry) &&
 734                     ceph_test_mount_opt(fsc, DCACHE) &&
 735                     __ceph_dir_is_complete(ci) &&
 736                     (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
 737                         spin_unlock(&ci->i_ceph_lock);
 738                         dout(" dir %p complete, -ENOENT\n", dir);
 739                         d_add(dentry, NULL);
 740                         di->lease_shared_gen = ci->i_shared_gen;
 741                         return NULL;
 742                 }
 743                 spin_unlock(&ci->i_ceph_lock);
 744         }
 745
 746         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 747                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 748         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 749         if (IS_ERR(req))
 750                 return ERR_CAST(req);
 751         req->r_dentry = dget(dentry);
 752         req->r_num_caps = 2;
 753
 754         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 755         if (ceph_security_xattr_wanted(dir))
 756                 mask |= CEPH_CAP_XATTR_SHARED;
 757         req->r_args.getattr.mask = cpu_to_le32(mask);
 758
 759         req->r_parent = dir;
 760         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 761         err = ceph_mdsc_do_request(mdsc, NULL, req);
 762         err = ceph_handle_snapdir(req, dentry, err);
 763         dentry = ceph_finish_lookup(req, dentry, err);
 764         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 765         dout("lookup result=%p\n", dentry);
 766         return dentry;
 767 }
 768
 769 /*
 770  * If we do a create but get no trace back from the MDS, follow up with
 771  * a lookup (the VFS expects us to link up the provided dentry).
 772  */
 773 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 774 {
 775         struct dentry *result = ceph_lookup(dir, dentry, 0);
 776
 777         if (result && !IS_ERR(result)) {
 778                 /*
 779                  * We created the item, then did a lookup, and found
 780                  * it was already linked to another inode we already
 781                  * had in our cache (and thus got spliced). To not
 782                  * confuse VFS (especially when inode is a directory),
 783                  * we don't link our dentry to that inode, return an
 784                  * error instead.
 785                  *
 786                  * This event should be rare and it happens only when
 787                  * we talk to old MDS. Recent MDS does not send traceless
 788                  * reply for request that creates new inode.
 789                  */
 790                 d_drop(result);
 791                 return -ESTALE;
 792         }
 793         return PTR_ERR(result);
 794 }
 795
 796 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 797                       umode_t mode, dev_t rdev)
 798 {
 799         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 800         struct ceph_mds_client *mdsc = fsc->mdsc;
 801         struct ceph_mds_request *req;
 802         struct ceph_acls_info acls = {};
 803         int err;
 804
 805         if (ceph_snap(dir) != CEPH_NOSNAP)
 806                 return -EROFS;
 807
 808         err = ceph_pre_init_acls(dir, &mode, &acls);
 809         if (err < 0)
 810                 return err;
 811
 812         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 813              dir, dentry, mode, rdev);
 814         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 815         if (IS_ERR(req)) {
 816                 err = PTR_ERR(req);
 817                 goto out;
 818         }
 819         req->r_dentry = dget(dentry);
 820         req->r_num_caps = 2;
 821         req->r_parent = dir;
 822         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 823         req->r_args.mknod.mode = cpu_to_le32(mode);
 824         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 825         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 826         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 827         if (acls.pagelist) {
 828                 req->r_pagelist = acls.pagelist;
 829                 acls.pagelist = NULL;
 830         }
 831         err = ceph_mdsc_do_request(mdsc, dir, req);
 832         if (!err && !req->r_reply_info.head->is_dentry)
 833                 err = ceph_handle_notrace_create(dir, dentry);
 834         ceph_mdsc_put_request(req);
 835 out:
 836         if (!err)
 837                 ceph_init_inode_acls(d_inode(dentry), &acls);
 838         else
 839                 d_drop(dentry);
 840         ceph_release_acls_info(&acls);
 841         return err;
 842 }
 843
 844 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 845                        bool excl)
 846 {
 847         return ceph_mknod(dir, dentry, mode, 0);
 848 }
 849
 850 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 851                             const char *dest)
 852 {
 853         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 854         struct ceph_mds_client *mdsc = fsc->mdsc;
 855         struct ceph_mds_request *req;
 856         int err;
 857
 858         if (ceph_snap(dir) != CEPH_NOSNAP)
 859                 return -EROFS;
 860
 861         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 862         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 863         if (IS_ERR(req)) {
 864                 err = PTR_ERR(req);
 865                 goto out;
 866         }
 867         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 868         if (!req->r_path2) {
 869                 err = -ENOMEM;
 870                 ceph_mdsc_put_request(req);
 871                 goto out;
 872         }
 873         req->r_parent = dir;
 874         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 875         req->r_dentry = dget(dentry);
 876         req->r_num_caps = 2;
 877         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 878         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 879         err = ceph_mdsc_do_request(mdsc, dir, req);
 880         if (!err && !req->r_reply_info.head->is_dentry)
 881                 err = ceph_handle_notrace_create(dir, dentry);
 882         ceph_mdsc_put_request(req);
 883 out:
 884         if (err)
 885                 d_drop(dentry);
 886         return err;
 887 }
 888
 889 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 890 {
 891         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 892         struct ceph_mds_client *mdsc = fsc->mdsc;
 893         struct ceph_mds_request *req;
 894         struct ceph_acls_info acls = {};
 895         int err = -EROFS;
 896         int op;
 897
 898         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 899                 /* mkdir .snap/foo is a MKSNAP */
 900                 op = CEPH_MDS_OP_MKSNAP;
 901                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 902                      dentry, dentry);
 903         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 904                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 905                 op = CEPH_MDS_OP_MKDIR;
 906         } else {
 907                 goto out;
 908         }
 909
 910         mode |= S_IFDIR;
 911         err = ceph_pre_init_acls(dir, &mode, &acls);
 912         if (err < 0)
 913                 goto out;
 914
 915         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 916         if (IS_ERR(req)) {
 917                 err = PTR_ERR(req);
 918                 goto out;
 919         }
 920
 921         req->r_dentry = dget(dentry);
 922         req->r_num_caps = 2;
 923         req->r_parent = dir;
 924         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 925         req->r_args.mkdir.mode = cpu_to_le32(mode);
 926         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 927         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 928         if (acls.pagelist) {
 929                 req->r_pagelist = acls.pagelist;
 930                 acls.pagelist = NULL;
 931         }
 932         err = ceph_mdsc_do_request(mdsc, dir, req);
 933         if (!err &&
 934             !req->r_reply_info.head->is_target &&
 935             !req->r_reply_info.head->is_dentry)
 936                 err = ceph_handle_notrace_create(dir, dentry);
 937         ceph_mdsc_put_request(req);
 938 out:
 939         if (!err)
 940                 ceph_init_inode_acls(d_inode(dentry), &acls);
 941         else
 942                 d_drop(dentry);
 943         ceph_release_acls_info(&acls);
 944         return err;
 945 }
 946
 947 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
 948                      struct dentry *dentry)
 949 {
 950         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 951         struct ceph_mds_client *mdsc = fsc->mdsc;
 952         struct ceph_mds_request *req;
 953         int err;
 954
 955         if (ceph_snap(dir) != CEPH_NOSNAP)
 956                 return -EROFS;
 957
 958         dout("link in dir %p old_dentry %p dentry %p\n", dir,
 959              old_dentry, dentry);
 960         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
 961         if (IS_ERR(req)) {
 962                 d_drop(dentry);
 963                 return PTR_ERR(req);
 964         }
 965         req->r_dentry = dget(dentry);
 966         req->r_num_caps = 2;
 967         req->r_old_dentry = dget(old_dentry);
 968         req->r_parent = dir;
 969         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 970         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 971         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 972         /* release LINK_SHARED on source inode (mds will lock it) */
 973         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
 974         err = ceph_mdsc_do_request(mdsc, dir, req);
 975         if (err) {
 976                 d_drop(dentry);
 977         } else if (!req->r_reply_info.head->is_dentry) {
 978                 ihold(d_inode(old_dentry));
 979                 d_instantiate(dentry, d_inode(old_dentry));
 980         }
 981         ceph_mdsc_put_request(req);
 982         return err;
 983 }
 984
 985 /*
 986  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
 987  * looks like the link count will hit 0, drop any other caps (other
 988  * than PIN) we don't specifically want (due to the file still being
 989  * open).
 990  */
 991 static int drop_caps_for_unlink(struct inode *inode)
 992 {
 993         struct ceph_inode_info *ci = ceph_inode(inode);
 994         int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
 995
 996         spin_lock(&ci->i_ceph_lock);
 997         if (inode->i_nlink == 1) {
 998                 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
 999                 ci->i_ceph_flags |= CEPH_I_NODELAY;
1000         }
1001         spin_unlock(&ci->i_ceph_lock);
1002         return drop;
1003 }
1004
1005 /*
1006  * rmdir and unlink are differ only by the metadata op code
1007  */
1008 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1009 {
1010         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1011         struct ceph_mds_client *mdsc = fsc->mdsc;
1012         struct inode *inode = d_inode(dentry);
1013         struct ceph_mds_request *req;
1014         int err = -EROFS;
1015         int op;
1016
1017         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1018                 /* rmdir .snap/foo is RMSNAP */
1019                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1020                 op = CEPH_MDS_OP_RMSNAP;
1021         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1022                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1023                      dir, dentry, inode);
1024                 op = d_is_dir(dentry) ?
1025                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1026         } else
1027                 goto out;
1028         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1029         if (IS_ERR(req)) {
1030                 err = PTR_ERR(req);
1031                 goto out;
1032         }
1033         req->r_dentry = dget(dentry);
1034         req->r_num_caps = 2;
1035         req->r_parent = dir;
1036         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1037         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1038         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1039         req->r_inode_drop = drop_caps_for_unlink(inode);
1040         err = ceph_mdsc_do_request(mdsc, dir, req);
1041         if (!err && !req->r_reply_info.head->is_dentry)
1042                 d_delete(dentry);
1043         ceph_mdsc_put_request(req);
1044 out:
1045         return err;
1046 }
1047
1048 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1049                        struct inode *new_dir, struct dentry *new_dentry,
1050                        unsigned int flags)
1051 {
1052         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1053         struct ceph_mds_client *mdsc = fsc->mdsc;
1054         struct ceph_mds_request *req;
1055         int op = CEPH_MDS_OP_RENAME;
1056         int err;
1057
1058         if (flags)
1059                 return -EINVAL;
1060
1061         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1062                 return -EXDEV;
1063         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1064                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1065                         op = CEPH_MDS_OP_RENAMESNAP;
1066                 else
1067                         return -EROFS;
1068         }
1069         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1070              old_dir, old_dentry, new_dir, new_dentry);
1071         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1072         if (IS_ERR(req))
1073                 return PTR_ERR(req);
1074         ihold(old_dir);
1075         req->r_dentry = dget(new_dentry);
1076         req->r_num_caps = 2;
1077         req->r_old_dentry = dget(old_dentry);
1078         req->r_old_dentry_dir = old_dir;
1079         req->r_parent = new_dir;
1080         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1081         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1082         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1083         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1084         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1085         /* release LINK_RDCACHE on source inode (mds will lock it) */
1086         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
1087         if (d_really_is_positive(new_dentry))
1088                 req->r_inode_drop = drop_caps_for_unlink(d_inode(new_dentry));
1089         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1090         if (!err && !req->r_reply_info.head->is_dentry) {
1091                 /*
1092                  * Normally d_move() is done by fill_trace (called by
1093                  * do_request, above).  If there is no trace, we need
1094                  * to do it here.
1095                  */
1096
1097                 /* d_move screws up sibling dentries' offsets */
1098                 ceph_dir_clear_complete(old_dir);
1099                 ceph_dir_clear_complete(new_dir);
1100
1101                 d_move(old_dentry, new_dentry);
1102
1103                 /* ensure target dentry is invalidated, despite
1104                    rehashing bug in vfs_rename_dir */
1105                 ceph_invalidate_dentry_lease(new_dentry);
1106         }
1107         ceph_mdsc_put_request(req);
1108         return err;
1109 }
1110
1111 /*
1112  * Ensure a dentry lease will no longer revalidate.
1113  */
1114 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1115 {
1116         spin_lock(&dentry->d_lock);
1117         ceph_dentry(dentry)->time = jiffies;
1118         ceph_dentry(dentry)->lease_shared_gen = 0;
1119         spin_unlock(&dentry->d_lock);
1120 }
1121
1122 /*
1123  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1124  * renew if the least is more than half up.
1125  */
1126 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags,
1127                                  struct inode *dir)
1128 {
1129         struct ceph_dentry_info *di;
1130         struct ceph_mds_session *s;
1131         int valid = 0;
1132         u32 gen;
1133         unsigned long ttl;
1134         struct ceph_mds_session *session = NULL;
1135         u32 seq = 0;
1136
1137         spin_lock(&dentry->d_lock);
1138         di = ceph_dentry(dentry);
1139         if (di && di->lease_session) {
1140                 s = di->lease_session;
1141                 spin_lock(&s->s_gen_ttl_lock);
1142                 gen = s->s_cap_gen;
1143                 ttl = s->s_cap_ttl;
1144                 spin_unlock(&s->s_gen_ttl_lock);
1145
1146                 if (di->lease_gen == gen &&
1147                     time_before(jiffies, di->time) &&
1148                     time_before(jiffies, ttl)) {
1149                         valid = 1;
1150                         if (di->lease_renew_after &&
1151                             time_after(jiffies, di->lease_renew_after)) {
1152                                 /*
1153                                  * We should renew. If we're in RCU walk mode
1154                                  * though, we can't do that so just return
1155                                  * -ECHILD.
1156                                  */
1157                                 if (flags & LOOKUP_RCU) {
1158                                         valid = -ECHILD;
1159                                 } else {
1160                                         session = ceph_get_mds_session(s);
1161                                         seq = di->lease_seq;
1162                                         di->lease_renew_after = 0;
1163                                         di->lease_renew_from = jiffies;
1164                                 }
1165                         }
1166                 }
1167         }
1168         spin_unlock(&dentry->d_lock);
1169
1170         if (session) {
1171                 ceph_mdsc_lease_send_msg(session, dir, dentry,
1172                                          CEPH_MDS_LEASE_RENEW, seq);
1173                 ceph_put_mds_session(session);
1174         }
1175         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1176         return valid;
1177 }
1178
1179 /*
1180  * Check if directory-wide content lease/cap is valid.
1181  */
1182 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1183 {
1184         struct ceph_inode_info *ci = ceph_inode(dir);
1185         struct ceph_dentry_info *di = ceph_dentry(dentry);
1186         int valid = 0;
1187
1188         spin_lock(&ci->i_ceph_lock);
1189         if (ci->i_shared_gen == di->lease_shared_gen)
1190                 valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1191         spin_unlock(&ci->i_ceph_lock);
1192         dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1193              dir, (unsigned)ci->i_shared_gen, dentry,
1194              (unsigned)di->lease_shared_gen, valid);
1195         return valid;
1196 }
1197
1198 /*
1199  * Check if cached dentry can be trusted.
1200  */
1201 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1202 {
1203         int valid = 0;
1204         struct dentry *parent;
1205         struct inode *dir;
1206
1207         if (flags & LOOKUP_RCU) {
1208                 parent = READ_ONCE(dentry->d_parent);
1209                 dir = d_inode_rcu(parent);
1210                 if (!dir)
1211                         return -ECHILD;
1212         } else {
1213                 parent = dget_parent(dentry);
1214                 dir = d_inode(parent);
1215         }
1216
1217         dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1218              dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1219
1220         /* always trust cached snapped dentries, snapdir dentry */
1221         if (ceph_snap(dir) != CEPH_NOSNAP) {
1222                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1223                      dentry, d_inode(dentry));
1224                 valid = 1;
1225         } else if (d_really_is_positive(dentry) &&
1226                    ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1227                 valid = 1;
1228         } else {
1229                 valid = dentry_lease_is_valid(dentry, flags, dir);
1230                 if (valid == -ECHILD)
1231                         return valid;
1232                 if (valid || dir_lease_is_valid(dir, dentry)) {
1233                         if (d_really_is_positive(dentry))
1234                                 valid = ceph_is_any_caps(d_inode(dentry));
1235                         else
1236                                 valid = 1;
1237                 }
1238         }
1239
1240         if (!valid) {
1241                 struct ceph_mds_client *mdsc =
1242                         ceph_sb_to_client(dir->i_sb)->mdsc;
1243                 struct ceph_mds_request *req;
1244                 int op, err;
1245                 u32 mask;
1246
1247                 if (flags & LOOKUP_RCU)
1248                         return -ECHILD;
1249
1250                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1251                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1252                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1253                 if (!IS_ERR(req)) {
1254                         req->r_dentry = dget(dentry);
1255                         req->r_num_caps = 2;
1256                         req->r_parent = dir;
1257
1258                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1259                         if (ceph_security_xattr_wanted(dir))
1260                                 mask |= CEPH_CAP_XATTR_SHARED;
1261                         req->r_args.getattr.mask = cpu_to_le32(mask);
1262
1263                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1264                         switch (err) {
1265                         case 0:
1266                                 if (d_really_is_positive(dentry) &&
1267                                     d_inode(dentry) == req->r_target_inode)
1268                                         valid = 1;
1269                                 break;
1270                         case -ENOENT:
1271                                 if (d_really_is_negative(dentry))
1272                                         valid = 1;
1273                                 /* Fallthrough */
1274                         default:
1275                                 break;
1276                         }
1277                         ceph_mdsc_put_request(req);
1278                         dout("d_revalidate %p lookup result=%d\n",
1279                              dentry, err);
1280                 }
1281         }
1282
1283         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1284         if (valid) {
1285                 ceph_dentry_lru_touch(dentry);
1286         } else {
1287                 ceph_dir_clear_complete(dir);
1288         }
1289
1290         if (!(flags & LOOKUP_RCU))
1291                 dput(parent);
1292         return valid;
1293 }
1294
1295 /*
1296  * Release our ceph_dentry_info.
1297  */
1298 static void ceph_d_release(struct dentry *dentry)
1299 {
1300         struct ceph_dentry_info *di = ceph_dentry(dentry);
1301
1302         dout("d_release %p\n", dentry);
1303         ceph_dentry_lru_del(dentry);
1304
1305         spin_lock(&dentry->d_lock);
1306         dentry->d_fsdata = NULL;
1307         spin_unlock(&dentry->d_lock);
1308
1309         if (di->lease_session)
1310                 ceph_put_mds_session(di->lease_session);
1311         kmem_cache_free(ceph_dentry_cachep, di);
1312 }
1313
1314 /*
1315  * When the VFS prunes a dentry from the cache, we need to clear the
1316  * complete flag on the parent directory.
1317  *
1318  * Called under dentry->d_lock.
1319  */
1320 static void ceph_d_prune(struct dentry *dentry)
1321 {
1322         dout("ceph_d_prune %p\n", dentry);
1323
1324         /* do we have a valid parent? */
1325         if (IS_ROOT(dentry))
1326                 return;
1327
1328         /* if we are not hashed, we don't affect dir's completeness */
1329         if (d_unhashed(dentry))
1330                 return;
1331
1332         if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_SNAPDIR)
1333                 return;
1334
1335         /*
1336          * we hold d_lock, so d_parent is stable, and d_fsdata is never
1337          * cleared until d_release
1338          */
1339         ceph_dir_clear_complete(d_inode(dentry->d_parent));
1340 }
1341
1342 /*
1343  * read() on a dir.  This weird interface hack only works if mounted
1344  * with '-o dirstat'.
1345  */
1346 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1347                              loff_t *ppos)
1348 {
1349         struct ceph_file_info *cf = file->private_data;
1350         struct inode *inode = file_inode(file);
1351         struct ceph_inode_info *ci = ceph_inode(inode);
1352         int left;
1353         const int bufsize = 1024;
1354
1355         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1356                 return -EISDIR;
1357
1358         if (!cf->dir_info) {
1359                 cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1360                 if (!cf->dir_info)
1361                         return -ENOMEM;
1362                 cf->dir_info_len =
1363                         snprintf(cf->dir_info, bufsize,
1364                                 "entries:   %20lld\n"
1365                                 " files:    %20lld\n"
1366                                 " subdirs:  %20lld\n"
1367                                 "rentries:  %20lld\n"
1368                                 " rfiles:   %20lld\n"
1369                                 " rsubdirs: %20lld\n"
1370                                 "rbytes:    %20lld\n"
1371                                 "rctime:    %10ld.%09ld\n",
1372                                 ci->i_files + ci->i_subdirs,
1373                                 ci->i_files,
1374                                 ci->i_subdirs,
1375                                 ci->i_rfiles + ci->i_rsubdirs,
1376                                 ci->i_rfiles,
1377                                 ci->i_rsubdirs,
1378                                 ci->i_rbytes,
1379                                 (long)ci->i_rctime.tv_sec,
1380                                 (long)ci->i_rctime.tv_nsec);
1381         }
1382
1383         if (*ppos >= cf->dir_info_len)
1384                 return 0;
1385         size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1386         left = copy_to_user(buf, cf->dir_info + *ppos, size);
1387         if (left == size)
1388                 return -EFAULT;
1389         *ppos += (size - left);
1390         return size - left;
1391 }
1392
1393 /*
1394  * We maintain a private dentry LRU.
1395  *
1396  * FIXME: this needs to be changed to a per-mds lru to be useful.
1397  */
1398 void ceph_dentry_lru_add(struct dentry *dn)
1399 {
1400         struct ceph_dentry_info *di = ceph_dentry(dn);
1401         struct ceph_mds_client *mdsc;
1402
1403         dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1404         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1405         spin_lock(&mdsc->dentry_lru_lock);
1406         list_add_tail(&di->lru, &mdsc->dentry_lru);
1407         mdsc->num_dentry++;
1408         spin_unlock(&mdsc->dentry_lru_lock);
1409 }
1410
1411 void ceph_dentry_lru_touch(struct dentry *dn)
1412 {
1413         struct ceph_dentry_info *di = ceph_dentry(dn);
1414         struct ceph_mds_client *mdsc;
1415
1416         dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1417              di->offset);
1418         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1419         spin_lock(&mdsc->dentry_lru_lock);
1420         list_move_tail(&di->lru, &mdsc->dentry_lru);
1421         spin_unlock(&mdsc->dentry_lru_lock);
1422 }
1423
1424 void ceph_dentry_lru_del(struct dentry *dn)
1425 {
1426         struct ceph_dentry_info *di = ceph_dentry(dn);
1427         struct ceph_mds_client *mdsc;
1428
1429         dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1430         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1431         spin_lock(&mdsc->dentry_lru_lock);
1432         list_del_init(&di->lru);
1433         mdsc->num_dentry--;
1434         spin_unlock(&mdsc->dentry_lru_lock);
1435 }
1436
1437 /*
1438  * Return name hash for a given dentry.  This is dependent on
1439  * the parent directory's hash function.
1440  */
1441 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1442 {
1443         struct ceph_inode_info *dci = ceph_inode(dir);
1444
1445         switch (dci->i_dir_layout.dl_dir_hash) {
1446         case 0: /* for backward compat */
1447         case CEPH_STR_HASH_LINUX:
1448                 return dn->d_name.hash;
1449
1450         default:
1451                 return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1452                                      dn->d_name.name, dn->d_name.len);
1453         }
1454 }
1455
1456 const struct file_operations ceph_dir_fops = {
1457         .read = ceph_read_dir,
1458         .iterate = ceph_readdir,
1459         .llseek = ceph_dir_llseek,
1460         .open = ceph_open,
1461         .release = ceph_release,
1462         .unlocked_ioctl = ceph_ioctl,
1463         .fsync = ceph_fsync,
1464 };
1465
1466 const struct file_operations ceph_snapdir_fops = {
1467         .iterate = ceph_readdir,
1468         .llseek = ceph_dir_llseek,
1469         .open = ceph_open,
1470         .release = ceph_release,
1471 };
1472
1473 const struct inode_operations ceph_dir_iops = {
1474         .lookup = ceph_lookup,
1475         .permission = ceph_permission,
1476         .getattr = ceph_getattr,
1477         .setattr = ceph_setattr,
1478         .listxattr = ceph_listxattr,
1479         .get_acl = ceph_get_acl,
1480         .set_acl = ceph_set_acl,
1481         .mknod = ceph_mknod,
1482         .symlink = ceph_symlink,
1483         .mkdir = ceph_mkdir,
1484         .link = ceph_link,
1485         .unlink = ceph_unlink,
1486         .rmdir = ceph_unlink,
1487         .rename = ceph_rename,
1488         .create = ceph_create,
1489         .atomic_open = ceph_atomic_open,
1490 };
1491
1492 const struct inode_operations ceph_snapdir_iops = {
1493         .lookup = ceph_lookup,
1494         .permission = ceph_permission,
1495         .getattr = ceph_getattr,
1496         .mkdir = ceph_mkdir,
1497         .rmdir = ceph_unlink,
1498         .rename = ceph_rename,
1499 };
1500
1501 const struct dentry_operations ceph_dentry_ops = {
1502         .d_revalidate = ceph_d_revalidate,
1503         .d_release = ceph_d_release,
1504         .d_prune = ceph_d_prune,
1505         .d_init = ceph_d_init,
1506 };