fs/ceph/dir.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/spinlock.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 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di);
  33 static int __dir_lease_try_check(const struct dentry *dentry);
  34
  35 /*
  36  * Initialize ceph dentry state.
  37  */
  38 static int ceph_d_init(struct dentry *dentry)
  39 {
  40         struct ceph_dentry_info *di;
  41
  42         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  43         if (!di)
  44                 return -ENOMEM;          /* oh well */
  45
  46         di->dentry = dentry;
  47         di->lease_session = NULL;
  48         di->time = jiffies;
  49         dentry->d_fsdata = di;
  50         INIT_LIST_HEAD(&di->lease_list);
  51         return 0;
  52 }
  53
  54 /*
  55  * for f_pos for readdir:
  56  * - hash order:
  57  *      (0xff << 52) | ((24 bits hash) << 28) |
  58  *      (the nth entry has hash collision);
  59  * - frag+name order;
  60  *      ((frag value) << 28) | (the nth entry in frag);
  61  */
  62 #define OFFSET_BITS     28
  63 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  64 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  65 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  66 {
  67         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  68         if (hash_order)
  69                 fpos |= HASH_ORDER;
  70         return fpos;
  71 }
  72
  73 static bool is_hash_order(loff_t p)
  74 {
  75         return (p & HASH_ORDER) == HASH_ORDER;
  76 }
  77
  78 static unsigned fpos_frag(loff_t p)
  79 {
  80         return p >> OFFSET_BITS;
  81 }
  82
  83 static unsigned fpos_hash(loff_t p)
  84 {
  85         return ceph_frag_value(fpos_frag(p));
  86 }
  87
  88 static unsigned fpos_off(loff_t p)
  89 {
  90         return p & OFFSET_MASK;
  91 }
  92
  93 static int fpos_cmp(loff_t l, loff_t r)
  94 {
  95         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
  96         if (v)
  97                 return v;
  98         return (int)(fpos_off(l) - fpos_off(r));
  99 }
 100
 101 /*
 102  * make note of the last dentry we read, so we can
 103  * continue at the same lexicographical point,
 104  * regardless of what dir changes take place on the
 105  * server.
 106  */
 107 static int note_last_dentry(struct ceph_dir_file_info *dfi, const char *name,
 108                             int len, unsigned next_offset)
 109 {
 110         char *buf = kmalloc(len+1, GFP_KERNEL);
 111         if (!buf)
 112                 return -ENOMEM;
 113         kfree(dfi->last_name);
 114         dfi->last_name = buf;
 115         memcpy(dfi->last_name, name, len);
 116         dfi->last_name[len] = 0;
 117         dfi->next_offset = next_offset;
 118         dout("note_last_dentry '%s'\n", dfi->last_name);
 119         return 0;
 120 }
 121
 122
 123 static struct dentry *
 124 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 125                         struct ceph_readdir_cache_control *cache_ctl)
 126 {
 127         struct inode *dir = d_inode(parent);
 128         struct dentry *dentry;
 129         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 130         loff_t ptr_pos = idx * sizeof(struct dentry *);
 131         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 132
 133         if (ptr_pos >= i_size_read(dir))
 134                 return NULL;
 135
 136         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 137                 ceph_readdir_cache_release(cache_ctl);
 138                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 139                 if (!cache_ctl->page) {
 140                         dout(" page %lu not found\n", ptr_pgoff);
 141                         return ERR_PTR(-EAGAIN);
 142                 }
 143                 /* reading/filling the cache are serialized by
 144                    i_mutex, no need to use page lock */
 145                 unlock_page(cache_ctl->page);
 146                 cache_ctl->dentries = kmap(cache_ctl->page);
 147         }
 148
 149         cache_ctl->index = idx & idx_mask;
 150
 151         rcu_read_lock();
 152         spin_lock(&parent->d_lock);
 153         /* check i_size again here, because empty directory can be
 154          * marked as complete while not holding the i_mutex. */
 155         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 156                 dentry = cache_ctl->dentries[cache_ctl->index];
 157         else
 158                 dentry = NULL;
 159         spin_unlock(&parent->d_lock);
 160         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 161                 dentry = NULL;
 162         rcu_read_unlock();
 163         return dentry ? : ERR_PTR(-EAGAIN);
 164 }
 165
 166 /*
 167  * When possible, we try to satisfy a readdir by peeking at the
 168  * dcache.  We make this work by carefully ordering dentries on
 169  * d_child when we initially get results back from the MDS, and
 170  * falling back to a "normal" sync readdir if any dentries in the dir
 171  * are dropped.
 172  *
 173  * Complete dir indicates that we have all dentries in the dir.  It is
 174  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 175  * the MDS if/when the directory is modified).
 176  */
 177 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 178                             int shared_gen)
 179 {
 180         struct ceph_dir_file_info *dfi = file->private_data;
 181         struct dentry *parent = file->f_path.dentry;
 182         struct inode *dir = d_inode(parent);
 183         struct dentry *dentry, *last = NULL;
 184         struct ceph_dentry_info *di;
 185         struct ceph_readdir_cache_control cache_ctl = {};
 186         u64 idx = 0;
 187         int err = 0;
 188
 189         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 190
 191         /* search start position */
 192         if (ctx->pos > 2) {
 193                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 194                 while (count > 0) {
 195                         u64 step = count >> 1;
 196                         dentry = __dcache_find_get_entry(parent, idx + step,
 197                                                          &cache_ctl);
 198                         if (!dentry) {
 199                                 /* use linar search */
 200                                 idx = 0;
 201                                 break;
 202                         }
 203                         if (IS_ERR(dentry)) {
 204                                 err = PTR_ERR(dentry);
 205                                 goto out;
 206                         }
 207                         di = ceph_dentry(dentry);
 208                         spin_lock(&dentry->d_lock);
 209                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 210                                 idx += step + 1;
 211                                 count -= step + 1;
 212                         } else {
 213                                 count = step;
 214                         }
 215                         spin_unlock(&dentry->d_lock);
 216                         dput(dentry);
 217                 }
 218
 219                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 220         }
 221
 222
 223         for (;;) {
 224                 bool emit_dentry = false;
 225                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 226                 if (!dentry) {
 227                         dfi->file_info.flags |= CEPH_F_ATEND;
 228                         err = 0;
 229                         break;
 230                 }
 231                 if (IS_ERR(dentry)) {
 232                         err = PTR_ERR(dentry);
 233                         goto out;
 234                 }
 235
 236                 spin_lock(&dentry->d_lock);
 237                 di = ceph_dentry(dentry);
 238                 if (d_unhashed(dentry) ||
 239                     d_really_is_negative(dentry) ||
 240                     di->lease_shared_gen != shared_gen) {
 241                         spin_unlock(&dentry->d_lock);
 242                         dput(dentry);
 243                         err = -EAGAIN;
 244                         goto out;
 245                 }
 246                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 247                         __ceph_dentry_dir_lease_touch(di);
 248                         emit_dentry = true;
 249                 }
 250                 spin_unlock(&dentry->d_lock);
 251
 252                 if (emit_dentry) {
 253                         dout(" %llx dentry %p %pd %p\n", di->offset,
 254                              dentry, dentry, d_inode(dentry));
 255                         ctx->pos = di->offset;
 256                         if (!dir_emit(ctx, dentry->d_name.name,
 257                                       dentry->d_name.len,
 258                                       ceph_translate_ino(dentry->d_sb,
 259                                                          d_inode(dentry)->i_ino),
 260                                       d_inode(dentry)->i_mode >> 12)) {
 261                                 dput(dentry);
 262                                 err = 0;
 263                                 break;
 264                         }
 265                         ctx->pos++;
 266
 267                         if (last)
 268                                 dput(last);
 269                         last = dentry;
 270                 } else {
 271                         dput(dentry);
 272                 }
 273         }
 274 out:
 275         ceph_readdir_cache_release(&cache_ctl);
 276         if (last) {
 277                 int ret;
 278                 di = ceph_dentry(last);
 279                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 280                                        fpos_off(di->offset) + 1);
 281                 if (ret < 0)
 282                         err = ret;
 283                 dput(last);
 284                 /* last_name no longer match cache index */
 285                 if (dfi->readdir_cache_idx >= 0) {
 286                         dfi->readdir_cache_idx = -1;
 287                         dfi->dir_release_count = 0;
 288                 }
 289         }
 290         return err;
 291 }
 292
 293 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 294 {
 295         if (!dfi->last_readdir)
 296                 return true;
 297         if (is_hash_order(pos))
 298                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 299         else
 300                 return dfi->frag != fpos_frag(pos);
 301 }
 302
 303 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 304 {
 305         struct ceph_dir_file_info *dfi = file->private_data;
 306         struct inode *inode = file_inode(file);
 307         struct ceph_inode_info *ci = ceph_inode(inode);
 308         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 309         struct ceph_mds_client *mdsc = fsc->mdsc;
 310         int i;
 311         int err;
 312         unsigned frag = -1;
 313         struct ceph_mds_reply_info_parsed *rinfo;
 314
 315         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 316         if (dfi->file_info.flags & CEPH_F_ATEND)
 317                 return 0;
 318
 319         /* always start with . and .. */
 320         if (ctx->pos == 0) {
 321                 dout("readdir off 0 -> '.'\n");
 322                 if (!dir_emit(ctx, ".", 1,
 323                             ceph_translate_ino(inode->i_sb, inode->i_ino),
 324                             inode->i_mode >> 12))
 325                         return 0;
 326                 ctx->pos = 1;
 327         }
 328         if (ctx->pos == 1) {
 329                 ino_t ino = parent_ino(file->f_path.dentry);
 330                 dout("readdir off 1 -> '..'\n");
 331                 if (!dir_emit(ctx, "..", 2,
 332                             ceph_translate_ino(inode->i_sb, ino),
 333                             inode->i_mode >> 12))
 334                         return 0;
 335                 ctx->pos = 2;
 336         }
 337
 338         /* can we use the dcache? */
 339         spin_lock(&ci->i_ceph_lock);
 340         if (ceph_test_mount_opt(fsc, DCACHE) &&
 341             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 342             ceph_snap(inode) != CEPH_SNAPDIR &&
 343             __ceph_dir_is_complete_ordered(ci) &&
 344             __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
 345                 int shared_gen = atomic_read(&ci->i_shared_gen);
 346                 spin_unlock(&ci->i_ceph_lock);
 347                 err = __dcache_readdir(file, ctx, shared_gen);
 348                 if (err != -EAGAIN)
 349                         return err;
 350         } else {
 351                 spin_unlock(&ci->i_ceph_lock);
 352         }
 353
 354         /* proceed with a normal readdir */
 355 more:
 356         /* do we have the correct frag content buffered? */
 357         if (need_send_readdir(dfi, ctx->pos)) {
 358                 struct ceph_mds_request *req;
 359                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 360                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 361
 362                 /* discard old result, if any */
 363                 if (dfi->last_readdir) {
 364                         ceph_mdsc_put_request(dfi->last_readdir);
 365                         dfi->last_readdir = NULL;
 366                 }
 367
 368                 if (is_hash_order(ctx->pos)) {
 369                         /* fragtree isn't always accurate. choose frag
 370                          * based on previous reply when possible. */
 371                         if (frag == (unsigned)-1)
 372                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 373                                                         NULL, NULL);
 374                 } else {
 375                         frag = fpos_frag(ctx->pos);
 376                 }
 377
 378                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 379                      ceph_vinop(inode), frag, dfi->last_name);
 380                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 381                 if (IS_ERR(req))
 382                         return PTR_ERR(req);
 383                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 384                 if (err) {
 385                         ceph_mdsc_put_request(req);
 386                         return err;
 387                 }
 388                 /* hints to request -> mds selection code */
 389                 req->r_direct_mode = USE_AUTH_MDS;
 390                 if (op == CEPH_MDS_OP_READDIR) {
 391                         req->r_direct_hash = ceph_frag_value(frag);
 392                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 393                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 394                 }
 395                 if (dfi->last_name) {
 396                         req->r_path2 = kstrdup(dfi->last_name, GFP_KERNEL);
 397                         if (!req->r_path2) {
 398                                 ceph_mdsc_put_request(req);
 399                                 return -ENOMEM;
 400                         }
 401                 } else if (is_hash_order(ctx->pos)) {
 402                         req->r_args.readdir.offset_hash =
 403                                 cpu_to_le32(fpos_hash(ctx->pos));
 404                 }
 405
 406                 req->r_dir_release_cnt = dfi->dir_release_count;
 407                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 408                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 409                 req->r_readdir_offset = dfi->next_offset;
 410                 req->r_args.readdir.frag = cpu_to_le32(frag);
 411                 req->r_args.readdir.flags =
 412                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 413
 414                 req->r_inode = inode;
 415                 ihold(inode);
 416                 req->r_dentry = dget(file->f_path.dentry);
 417                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 418                 if (err < 0) {
 419                         ceph_mdsc_put_request(req);
 420                         return err;
 421                 }
 422                 dout("readdir got and parsed readdir result=%d on "
 423                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 424                      err, frag,
 425                      (int)req->r_reply_info.dir_end,
 426                      (int)req->r_reply_info.dir_complete,
 427                      (int)req->r_reply_info.hash_order);
 428
 429                 rinfo = &req->r_reply_info;
 430                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 431                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 432                         if (!rinfo->hash_order) {
 433                                 dfi->next_offset = req->r_readdir_offset;
 434                                 /* adjust ctx->pos to beginning of frag */
 435                                 ctx->pos = ceph_make_fpos(frag,
 436                                                           dfi->next_offset,
 437                                                           false);
 438                         }
 439                 }
 440
 441                 dfi->frag = frag;
 442                 dfi->last_readdir = req;
 443
 444                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 445                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 446                         if (dfi->readdir_cache_idx < 0) {
 447                                 /* preclude from marking dir ordered */
 448                                 dfi->dir_ordered_count = 0;
 449                         } else if (ceph_frag_is_leftmost(frag) &&
 450                                    dfi->next_offset == 2) {
 451                                 /* note dir version at start of readdir so
 452                                  * we can tell if any dentries get dropped */
 453                                 dfi->dir_release_count = req->r_dir_release_cnt;
 454                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 455                         }
 456                 } else {
 457                         dout("readdir !did_prepopulate\n");
 458                         /* disable readdir cache */
 459                         dfi->readdir_cache_idx = -1;
 460                         /* preclude from marking dir complete */
 461                         dfi->dir_release_count = 0;
 462                 }
 463
 464                 /* note next offset and last dentry name */
 465                 if (rinfo->dir_nr > 0) {
 466                         struct ceph_mds_reply_dir_entry *rde =
 467                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 468                         unsigned next_offset = req->r_reply_info.dir_end ?
 469                                         2 : (fpos_off(rde->offset) + 1);
 470                         err = note_last_dentry(dfi, rde->name, rde->name_len,
 471                                                next_offset);
 472                         if (err)
 473                                 return err;
 474                 } else if (req->r_reply_info.dir_end) {
 475                         dfi->next_offset = 2;
 476                         /* keep last name */
 477                 }
 478         }
 479
 480         rinfo = &dfi->last_readdir->r_reply_info;
 481         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 482              dfi->frag, rinfo->dir_nr, ctx->pos,
 483              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 484
 485         i = 0;
 486         /* search start position */
 487         if (rinfo->dir_nr > 0) {
 488                 int step, nr = rinfo->dir_nr;
 489                 while (nr > 0) {
 490                         step = nr >> 1;
 491                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 492                                 i +=  step + 1;
 493                                 nr -= step + 1;
 494                         } else {
 495                                 nr = step;
 496                         }
 497                 }
 498         }
 499         for (; i < rinfo->dir_nr; i++) {
 500                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 501                 struct ceph_vino vino;
 502                 ino_t ino;
 503                 u32 ftype;
 504
 505                 BUG_ON(rde->offset < ctx->pos);
 506
 507                 ctx->pos = rde->offset;
 508                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 509                      i, rinfo->dir_nr, ctx->pos,
 510                      rde->name_len, rde->name, &rde->inode.in);
 511
 512                 BUG_ON(!rde->inode.in);
 513                 ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
 514                 vino.ino = le64_to_cpu(rde->inode.in->ino);
 515                 vino.snap = le64_to_cpu(rde->inode.in->snapid);
 516                 ino = ceph_vino_to_ino(vino);
 517
 518                 if (!dir_emit(ctx, rde->name, rde->name_len,
 519                               ceph_translate_ino(inode->i_sb, ino), ftype)) {
 520                         dout("filldir stopping us...\n");
 521                         return 0;
 522                 }
 523                 ctx->pos++;
 524         }
 525
 526         ceph_mdsc_put_request(dfi->last_readdir);
 527         dfi->last_readdir = NULL;
 528
 529         if (dfi->next_offset > 2) {
 530                 frag = dfi->frag;
 531                 goto more;
 532         }
 533
 534         /* more frags? */
 535         if (!ceph_frag_is_rightmost(dfi->frag)) {
 536                 frag = ceph_frag_next(dfi->frag);
 537                 if (is_hash_order(ctx->pos)) {
 538                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 539                                                         dfi->next_offset, true);
 540                         if (new_pos > ctx->pos)
 541                                 ctx->pos = new_pos;
 542                         /* keep last_name */
 543                 } else {
 544                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 545                                                         false);
 546                         kfree(dfi->last_name);
 547                         dfi->last_name = NULL;
 548                 }
 549                 dout("readdir next frag is %x\n", frag);
 550                 goto more;
 551         }
 552         dfi->file_info.flags |= CEPH_F_ATEND;
 553
 554         /*
 555          * if dir_release_count still matches the dir, no dentries
 556          * were released during the whole readdir, and we should have
 557          * the complete dir contents in our cache.
 558          */
 559         if (atomic64_read(&ci->i_release_count) ==
 560                                         dfi->dir_release_count) {
 561                 spin_lock(&ci->i_ceph_lock);
 562                 if (dfi->dir_ordered_count ==
 563                                 atomic64_read(&ci->i_ordered_count)) {
 564                         dout(" marking %p complete and ordered\n", inode);
 565                         /* use i_size to track number of entries in
 566                          * readdir cache */
 567                         BUG_ON(dfi->readdir_cache_idx < 0);
 568                         i_size_write(inode, dfi->readdir_cache_idx *
 569                                      sizeof(struct dentry*));
 570                 } else {
 571                         dout(" marking %p complete\n", inode);
 572                 }
 573                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 574                                         dfi->dir_ordered_count);
 575                 spin_unlock(&ci->i_ceph_lock);
 576         }
 577
 578         dout("readdir %p file %p done.\n", inode, file);
 579         return 0;
 580 }
 581
 582 static void reset_readdir(struct ceph_dir_file_info *dfi)
 583 {
 584         if (dfi->last_readdir) {
 585                 ceph_mdsc_put_request(dfi->last_readdir);
 586                 dfi->last_readdir = NULL;
 587         }
 588         kfree(dfi->last_name);
 589         dfi->last_name = NULL;
 590         dfi->dir_release_count = 0;
 591         dfi->readdir_cache_idx = -1;
 592         dfi->next_offset = 2;  /* compensate for . and .. */
 593         dfi->file_info.flags &= ~CEPH_F_ATEND;
 594 }
 595
 596 /*
 597  * discard buffered readdir content on seekdir(0), or seek to new frag,
 598  * or seek prior to current chunk
 599  */
 600 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 601 {
 602         struct ceph_mds_reply_info_parsed *rinfo;
 603         loff_t chunk_offset;
 604         if (new_pos == 0)
 605                 return true;
 606         if (is_hash_order(new_pos)) {
 607                 /* no need to reset last_name for a forward seek when
 608                  * dentries are sotred in hash order */
 609         } else if (dfi->frag != fpos_frag(new_pos)) {
 610                 return true;
 611         }
 612         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 613         if (!rinfo || !rinfo->dir_nr)
 614                 return true;
 615         chunk_offset = rinfo->dir_entries[0].offset;
 616         return new_pos < chunk_offset ||
 617                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 618 }
 619
 620 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 621 {
 622         struct ceph_dir_file_info *dfi = file->private_data;
 623         struct inode *inode = file->f_mapping->host;
 624         loff_t retval;
 625
 626         inode_lock(inode);
 627         retval = -EINVAL;
 628         switch (whence) {
 629         case SEEK_CUR:
 630                 offset += file->f_pos;
 631         case SEEK_SET:
 632                 break;
 633         case SEEK_END:
 634                 retval = -EOPNOTSUPP;
 635         default:
 636                 goto out;
 637         }
 638
 639         if (offset >= 0) {
 640                 if (need_reset_readdir(dfi, offset)) {
 641                         dout("dir_llseek dropping %p content\n", file);
 642                         reset_readdir(dfi);
 643                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 644                         /* for hash offset, we don't know if a forward seek
 645                          * is within same frag */
 646                         dfi->dir_release_count = 0;
 647                         dfi->readdir_cache_idx = -1;
 648                 }
 649
 650                 if (offset != file->f_pos) {
 651                         file->f_pos = offset;
 652                         file->f_version = 0;
 653                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 654                 }
 655                 retval = offset;
 656         }
 657 out:
 658         inode_unlock(inode);
 659         return retval;
 660 }
 661
 662 /*
 663  * Handle lookups for the hidden .snap directory.
 664  */
 665 int ceph_handle_snapdir(struct ceph_mds_request *req,
 666                         struct dentry *dentry, int err)
 667 {
 668         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 669         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 670
 671         /* .snap dir? */
 672         if (err == -ENOENT &&
 673             ceph_snap(parent) == CEPH_NOSNAP &&
 674             strcmp(dentry->d_name.name,
 675                    fsc->mount_options->snapdir_name) == 0) {
 676                 struct inode *inode = ceph_get_snapdir(parent);
 677                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 678                      dentry, dentry, inode);
 679                 BUG_ON(!d_unhashed(dentry));
 680                 d_add(dentry, inode);
 681                 err = 0;
 682         }
 683         return err;
 684 }
 685
 686 /*
 687  * Figure out final result of a lookup/open request.
 688  *
 689  * Mainly, make sure we return the final req->r_dentry (if it already
 690  * existed) in place of the original VFS-provided dentry when they
 691  * differ.
 692  *
 693  * Gracefully handle the case where the MDS replies with -ENOENT and
 694  * no trace (which it may do, at its discretion, e.g., if it doesn't
 695  * care to issue a lease on the negative dentry).
 696  */
 697 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 698                                   struct dentry *dentry, int err)
 699 {
 700         if (err == -ENOENT) {
 701                 /* no trace? */
 702                 err = 0;
 703                 if (!req->r_reply_info.head->is_dentry) {
 704                         dout("ENOENT and no trace, dentry %p inode %p\n",
 705                              dentry, d_inode(dentry));
 706                         if (d_really_is_positive(dentry)) {
 707                                 d_drop(dentry);
 708                                 err = -ENOENT;
 709                         } else {
 710                                 d_add(dentry, NULL);
 711                         }
 712                 }
 713         }
 714         if (err)
 715                 dentry = ERR_PTR(err);
 716         else if (dentry != req->r_dentry)
 717                 dentry = dget(req->r_dentry);   /* we got spliced */
 718         else
 719                 dentry = NULL;
 720         return dentry;
 721 }
 722
 723 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 724 {
 725         return ceph_ino(inode) == CEPH_INO_ROOT &&
 726                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 727 }
 728
 729 /*
 730  * Look up a single dir entry.  If there is a lookup intent, inform
 731  * the MDS so that it gets our 'caps wanted' value in a single op.
 732  */
 733 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 734                                   unsigned int flags)
 735 {
 736         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 737         struct ceph_mds_client *mdsc = fsc->mdsc;
 738         struct ceph_mds_request *req;
 739         int op;
 740         int mask;
 741         int err;
 742
 743         dout("lookup %p dentry %p '%pd'\n",
 744              dir, dentry, dentry);
 745
 746         if (dentry->d_name.len > NAME_MAX)
 747                 return ERR_PTR(-ENAMETOOLONG);
 748
 749         /* can we conclude ENOENT locally? */
 750         if (d_really_is_negative(dentry)) {
 751                 struct ceph_inode_info *ci = ceph_inode(dir);
 752                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 753
 754                 spin_lock(&ci->i_ceph_lock);
 755                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 756                 if (strncmp(dentry->d_name.name,
 757                             fsc->mount_options->snapdir_name,
 758                             dentry->d_name.len) &&
 759                     !is_root_ceph_dentry(dir, dentry) &&
 760                     ceph_test_mount_opt(fsc, DCACHE) &&
 761                     __ceph_dir_is_complete(ci) &&
 762                     (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
 763                         spin_unlock(&ci->i_ceph_lock);
 764                         dout(" dir %p complete, -ENOENT\n", dir);
 765                         d_add(dentry, NULL);
 766                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 767                         return NULL;
 768                 }
 769                 spin_unlock(&ci->i_ceph_lock);
 770         }
 771
 772         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 773                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 774         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 775         if (IS_ERR(req))
 776                 return ERR_CAST(req);
 777         req->r_dentry = dget(dentry);
 778         req->r_num_caps = 2;
 779
 780         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 781         if (ceph_security_xattr_wanted(dir))
 782                 mask |= CEPH_CAP_XATTR_SHARED;
 783         req->r_args.getattr.mask = cpu_to_le32(mask);
 784
 785         req->r_parent = dir;
 786         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 787         err = ceph_mdsc_do_request(mdsc, NULL, req);
 788         err = ceph_handle_snapdir(req, dentry, err);
 789         dentry = ceph_finish_lookup(req, dentry, err);
 790         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 791         dout("lookup result=%p\n", dentry);
 792         return dentry;
 793 }
 794
 795 /*
 796  * If we do a create but get no trace back from the MDS, follow up with
 797  * a lookup (the VFS expects us to link up the provided dentry).
 798  */
 799 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 800 {
 801         struct dentry *result = ceph_lookup(dir, dentry, 0);
 802
 803         if (result && !IS_ERR(result)) {
 804                 /*
 805                  * We created the item, then did a lookup, and found
 806                  * it was already linked to another inode we already
 807                  * had in our cache (and thus got spliced). To not
 808                  * confuse VFS (especially when inode is a directory),
 809                  * we don't link our dentry to that inode, return an
 810                  * error instead.
 811                  *
 812                  * This event should be rare and it happens only when
 813                  * we talk to old MDS. Recent MDS does not send traceless
 814                  * reply for request that creates new inode.
 815                  */
 816                 d_drop(result);
 817                 return -ESTALE;
 818         }
 819         return PTR_ERR(result);
 820 }
 821
 822 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 823                       umode_t mode, dev_t rdev)
 824 {
 825         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 826         struct ceph_mds_client *mdsc = fsc->mdsc;
 827         struct ceph_mds_request *req;
 828         struct ceph_acl_sec_ctx as_ctx = {};
 829         int err;
 830
 831         if (ceph_snap(dir) != CEPH_NOSNAP)
 832                 return -EROFS;
 833
 834         if (ceph_quota_is_max_files_exceeded(dir)) {
 835                 err = -EDQUOT;
 836                 goto out;
 837         }
 838
 839         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 840         if (err < 0)
 841                 goto out;
 842         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 843         if (err < 0)
 844                 goto out;
 845
 846         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 847              dir, dentry, mode, rdev);
 848         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 849         if (IS_ERR(req)) {
 850                 err = PTR_ERR(req);
 851                 goto out;
 852         }
 853         req->r_dentry = dget(dentry);
 854         req->r_num_caps = 2;
 855         req->r_parent = dir;
 856         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 857         req->r_args.mknod.mode = cpu_to_le32(mode);
 858         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 859         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 860         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 861         if (as_ctx.pagelist) {
 862                 req->r_pagelist = as_ctx.pagelist;
 863                 as_ctx.pagelist = NULL;
 864         }
 865         err = ceph_mdsc_do_request(mdsc, dir, req);
 866         if (!err && !req->r_reply_info.head->is_dentry)
 867                 err = ceph_handle_notrace_create(dir, dentry);
 868         ceph_mdsc_put_request(req);
 869 out:
 870         if (!err)
 871                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 872         else
 873                 d_drop(dentry);
 874         ceph_release_acl_sec_ctx(&as_ctx);
 875         return err;
 876 }
 877
 878 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 879                        bool excl)
 880 {
 881         return ceph_mknod(dir, dentry, mode, 0);
 882 }
 883
 884 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 885                             const char *dest)
 886 {
 887         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 888         struct ceph_mds_client *mdsc = fsc->mdsc;
 889         struct ceph_mds_request *req;
 890         struct ceph_acl_sec_ctx as_ctx = {};
 891         int err;
 892
 893         if (ceph_snap(dir) != CEPH_NOSNAP)
 894                 return -EROFS;
 895
 896         if (ceph_quota_is_max_files_exceeded(dir)) {
 897                 err = -EDQUOT;
 898                 goto out;
 899         }
 900
 901         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 902         if (err < 0)
 903                 goto out;
 904
 905         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 906         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 907         if (IS_ERR(req)) {
 908                 err = PTR_ERR(req);
 909                 goto out;
 910         }
 911         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 912         if (!req->r_path2) {
 913                 err = -ENOMEM;
 914                 ceph_mdsc_put_request(req);
 915                 goto out;
 916         }
 917         req->r_parent = dir;
 918         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 919         req->r_dentry = dget(dentry);
 920         req->r_num_caps = 2;
 921         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 922         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 923         err = ceph_mdsc_do_request(mdsc, dir, req);
 924         if (!err && !req->r_reply_info.head->is_dentry)
 925                 err = ceph_handle_notrace_create(dir, dentry);
 926         ceph_mdsc_put_request(req);
 927 out:
 928         if (err)
 929                 d_drop(dentry);
 930         ceph_release_acl_sec_ctx(&as_ctx);
 931         return err;
 932 }
 933
 934 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 935 {
 936         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 937         struct ceph_mds_client *mdsc = fsc->mdsc;
 938         struct ceph_mds_request *req;
 939         struct ceph_acl_sec_ctx as_ctx = {};
 940         int err = -EROFS;
 941         int op;
 942
 943         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 944                 /* mkdir .snap/foo is a MKSNAP */
 945                 op = CEPH_MDS_OP_MKSNAP;
 946                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 947                      dentry, dentry);
 948         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 949                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 950                 op = CEPH_MDS_OP_MKDIR;
 951         } else {
 952                 goto out;
 953         }
 954
 955         if (op == CEPH_MDS_OP_MKDIR &&
 956             ceph_quota_is_max_files_exceeded(dir)) {
 957                 err = -EDQUOT;
 958                 goto out;
 959         }
 960
 961         mode |= S_IFDIR;
 962         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 963         if (err < 0)
 964                 goto out;
 965         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 966         if (err < 0)
 967                 goto out;
 968
 969         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 970         if (IS_ERR(req)) {
 971                 err = PTR_ERR(req);
 972                 goto out;
 973         }
 974
 975         req->r_dentry = dget(dentry);
 976         req->r_num_caps = 2;
 977         req->r_parent = dir;
 978         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 979         req->r_args.mkdir.mode = cpu_to_le32(mode);
 980         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 981         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 982         if (as_ctx.pagelist) {
 983                 req->r_pagelist = as_ctx.pagelist;
 984                 as_ctx.pagelist = NULL;
 985         }
 986         err = ceph_mdsc_do_request(mdsc, dir, req);
 987         if (!err &&
 988             !req->r_reply_info.head->is_target &&
 989             !req->r_reply_info.head->is_dentry)
 990                 err = ceph_handle_notrace_create(dir, dentry);
 991         ceph_mdsc_put_request(req);
 992 out:
 993         if (!err)
 994                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 995         else
 996                 d_drop(dentry);
 997         ceph_release_acl_sec_ctx(&as_ctx);
 998         return err;
 999 }
1000
1001 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1002                      struct dentry *dentry)
1003 {
1004         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1005         struct ceph_mds_client *mdsc = fsc->mdsc;
1006         struct ceph_mds_request *req;
1007         int err;
1008
1009         if (ceph_snap(dir) != CEPH_NOSNAP)
1010                 return -EROFS;
1011
1012         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1013              old_dentry, dentry);
1014         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1015         if (IS_ERR(req)) {
1016                 d_drop(dentry);
1017                 return PTR_ERR(req);
1018         }
1019         req->r_dentry = dget(dentry);
1020         req->r_num_caps = 2;
1021         req->r_old_dentry = dget(old_dentry);
1022         req->r_parent = dir;
1023         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1024         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1025         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1026         /* release LINK_SHARED on source inode (mds will lock it) */
1027         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1028         err = ceph_mdsc_do_request(mdsc, dir, req);
1029         if (err) {
1030                 d_drop(dentry);
1031         } else if (!req->r_reply_info.head->is_dentry) {
1032                 ihold(d_inode(old_dentry));
1033                 d_instantiate(dentry, d_inode(old_dentry));
1034         }
1035         ceph_mdsc_put_request(req);
1036         return err;
1037 }
1038
1039 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1040                                  struct ceph_mds_request *req)
1041 {
1042         int result = req->r_err ? req->r_err :
1043                         le32_to_cpu(req->r_reply_info.head->result);
1044
1045         if (result == -EJUKEBOX)
1046                 goto out;
1047
1048         /* If op failed, mark everyone involved for errors */
1049         if (result) {
1050                 int pathlen;
1051                 u64 base;
1052                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1053                                                   &base, 0);
1054
1055                 /* mark error on parent + clear complete */
1056                 mapping_set_error(req->r_parent->i_mapping, result);
1057                 ceph_dir_clear_complete(req->r_parent);
1058
1059                 /* drop the dentry -- we don't know its status */
1060                 if (!d_unhashed(req->r_dentry))
1061                         d_drop(req->r_dentry);
1062
1063                 /* mark inode itself for an error (since metadata is bogus) */
1064                 mapping_set_error(req->r_old_inode->i_mapping, result);
1065
1066                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1067                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1068                 ceph_mdsc_free_path(path, pathlen);
1069         }
1070 out:
1071         iput(req->r_old_inode);
1072         ceph_mdsc_release_dir_caps(req);
1073 }
1074
1075 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1076 {
1077         struct ceph_inode_info *ci = ceph_inode(dir);
1078         struct ceph_dentry_info *di;
1079         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1080
1081         spin_lock(&ci->i_ceph_lock);
1082         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1083                 ceph_take_cap_refs(ci, want, false);
1084                 got = want;
1085         }
1086         spin_unlock(&ci->i_ceph_lock);
1087
1088         /* If we didn't get anything, return 0 */
1089         if (!got)
1090                 return 0;
1091
1092         spin_lock(&dentry->d_lock);
1093         di = ceph_dentry(dentry);
1094         /*
1095          * - We are holding Fx, which implies Fs caps.
1096          * - Only support async unlink for primary linkage
1097          */
1098         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1099             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1100                 want = 0;
1101         spin_unlock(&dentry->d_lock);
1102
1103         /* Do we still want what we've got? */
1104         if (want == got)
1105                 return got;
1106
1107         ceph_put_cap_refs(ci, got);
1108         return 0;
1109 }
1110
1111 /*
1112  * rmdir and unlink are differ only by the metadata op code
1113  */
1114 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1115 {
1116         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1117         struct ceph_mds_client *mdsc = fsc->mdsc;
1118         struct inode *inode = d_inode(dentry);
1119         struct ceph_mds_request *req;
1120         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1121         int err = -EROFS;
1122         int op;
1123
1124         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1125                 /* rmdir .snap/foo is RMSNAP */
1126                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1127                 op = CEPH_MDS_OP_RMSNAP;
1128         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1129                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1130                      dir, dentry, inode);
1131                 op = d_is_dir(dentry) ?
1132                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1133         } else
1134                 goto out;
1135 retry:
1136         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1137         if (IS_ERR(req)) {
1138                 err = PTR_ERR(req);
1139                 goto out;
1140         }
1141         req->r_dentry = dget(dentry);
1142         req->r_num_caps = 2;
1143         req->r_parent = dir;
1144         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1145         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1146         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1147
1148         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1149             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1150                 dout("async unlink on %lu/%.*s caps=%s", dir->i_ino,
1151                      dentry->d_name.len, dentry->d_name.name,
1152                      ceph_cap_string(req->r_dir_caps));
1153                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1154                 req->r_callback = ceph_async_unlink_cb;
1155                 req->r_old_inode = d_inode(dentry);
1156                 ihold(req->r_old_inode);
1157                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1158                 if (!err) {
1159                         /*
1160                          * We have enough caps, so we assume that the unlink
1161                          * will succeed. Fix up the target inode and dcache.
1162                          */
1163                         drop_nlink(inode);
1164                         d_delete(dentry);
1165                 } else if (err == -EJUKEBOX) {
1166                         try_async = false;
1167                         ceph_mdsc_put_request(req);
1168                         goto retry;
1169                 }
1170         } else {
1171                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1172                 err = ceph_mdsc_do_request(mdsc, dir, req);
1173                 if (!err && !req->r_reply_info.head->is_dentry)
1174                         d_delete(dentry);
1175         }
1176
1177         ceph_mdsc_put_request(req);
1178 out:
1179         return err;
1180 }
1181
1182 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1183                        struct inode *new_dir, struct dentry *new_dentry,
1184                        unsigned int flags)
1185 {
1186         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1187         struct ceph_mds_client *mdsc = fsc->mdsc;
1188         struct ceph_mds_request *req;
1189         int op = CEPH_MDS_OP_RENAME;
1190         int err;
1191
1192         if (flags)
1193                 return -EINVAL;
1194
1195         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1196                 return -EXDEV;
1197         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1198                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1199                         op = CEPH_MDS_OP_RENAMESNAP;
1200                 else
1201                         return -EROFS;
1202         }
1203         /* don't allow cross-quota renames */
1204         if ((old_dir != new_dir) &&
1205             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1206                 return -EXDEV;
1207
1208         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1209              old_dir, old_dentry, new_dir, new_dentry);
1210         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1211         if (IS_ERR(req))
1212                 return PTR_ERR(req);
1213         ihold(old_dir);
1214         req->r_dentry = dget(new_dentry);
1215         req->r_num_caps = 2;
1216         req->r_old_dentry = dget(old_dentry);
1217         req->r_old_dentry_dir = old_dir;
1218         req->r_parent = new_dir;
1219         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1220         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1221         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1222         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1223         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1224         /* release LINK_RDCACHE on source inode (mds will lock it) */
1225         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1226         if (d_really_is_positive(new_dentry)) {
1227                 req->r_inode_drop =
1228                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1229         }
1230         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1231         if (!err && !req->r_reply_info.head->is_dentry) {
1232                 /*
1233                  * Normally d_move() is done by fill_trace (called by
1234                  * do_request, above).  If there is no trace, we need
1235                  * to do it here.
1236                  */
1237                 d_move(old_dentry, new_dentry);
1238         }
1239         ceph_mdsc_put_request(req);
1240         return err;
1241 }
1242
1243 /*
1244  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1245  * Leases at front of the list will expire first. (Assume all leases have
1246  * similar duration)
1247  *
1248  * Called under dentry->d_lock.
1249  */
1250 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1251 {
1252         struct dentry *dn = di->dentry;
1253         struct ceph_mds_client *mdsc;
1254
1255         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1256
1257         di->flags |= CEPH_DENTRY_LEASE_LIST;
1258         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1259                 di->flags |= CEPH_DENTRY_REFERENCED;
1260                 return;
1261         }
1262
1263         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1264         spin_lock(&mdsc->dentry_list_lock);
1265         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1266         spin_unlock(&mdsc->dentry_list_lock);
1267 }
1268
1269 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1270                                      struct ceph_dentry_info *di)
1271 {
1272         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1273         di->lease_gen = 0;
1274         di->time = jiffies;
1275         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1276 }
1277
1278 /*
1279  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1280  * list if it's not in the list, otherwise set 'referenced' flag.
1281  *
1282  * Called under dentry->d_lock.
1283  */
1284 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1285 {
1286         struct dentry *dn = di->dentry;
1287         struct ceph_mds_client *mdsc;
1288
1289         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1290              di, dn, dn, di->offset);
1291
1292         if (!list_empty(&di->lease_list)) {
1293                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1294                         /* don't remove dentry from dentry lease list
1295                          * if its lease is valid */
1296                         if (__dentry_lease_is_valid(di))
1297                                 return;
1298                 } else {
1299                         di->flags |= CEPH_DENTRY_REFERENCED;
1300                         return;
1301                 }
1302         }
1303
1304         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1305                 di->flags |= CEPH_DENTRY_REFERENCED;
1306                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1307                 return;
1308         }
1309
1310         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1311         spin_lock(&mdsc->dentry_list_lock);
1312         __dentry_dir_lease_touch(mdsc, di),
1313         spin_unlock(&mdsc->dentry_list_lock);
1314 }
1315
1316 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1317 {
1318         struct ceph_mds_client *mdsc;
1319         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1320                 return;
1321         if (list_empty(&di->lease_list))
1322                 return;
1323
1324         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1325         spin_lock(&mdsc->dentry_list_lock);
1326         list_del_init(&di->lease_list);
1327         spin_unlock(&mdsc->dentry_list_lock);
1328 }
1329
1330 enum {
1331         KEEP    = 0,
1332         DELETE  = 1,
1333         TOUCH   = 2,
1334         STOP    = 4,
1335 };
1336
1337 struct ceph_lease_walk_control {
1338         bool dir_lease;
1339         bool expire_dir_lease;
1340         unsigned long nr_to_scan;
1341         unsigned long dir_lease_ttl;
1342 };
1343
1344 static unsigned long
1345 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1346                      struct ceph_lease_walk_control *lwc,
1347                      int (*check)(struct dentry*, void*))
1348 {
1349         struct ceph_dentry_info *di, *tmp;
1350         struct dentry *dentry, *last = NULL;
1351         struct list_head* list;
1352         LIST_HEAD(dispose);
1353         unsigned long freed = 0;
1354         int ret = 0;
1355
1356         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1357         spin_lock(&mdsc->dentry_list_lock);
1358         list_for_each_entry_safe(di, tmp, list, lease_list) {
1359                 if (!lwc->nr_to_scan)
1360                         break;
1361                 --lwc->nr_to_scan;
1362
1363                 dentry = di->dentry;
1364                 if (last == dentry)
1365                         break;
1366
1367                 if (!spin_trylock(&dentry->d_lock))
1368                         continue;
1369
1370                 if (__lockref_is_dead(&dentry->d_lockref)) {
1371                         list_del_init(&di->lease_list);
1372                         goto next;
1373                 }
1374
1375                 ret = check(dentry, lwc);
1376                 if (ret & TOUCH) {
1377                         /* move it into tail of dir lease list */
1378                         __dentry_dir_lease_touch(mdsc, di);
1379                         if (!last)
1380                                 last = dentry;
1381                 }
1382                 if (ret & DELETE) {
1383                         /* stale lease */
1384                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1385                         if (dentry->d_lockref.count > 0) {
1386                                 /* update_dentry_lease() will re-add
1387                                  * it to lease list, or
1388                                  * ceph_d_delete() will return 1 when
1389                                  * last reference is dropped */
1390                                 list_del_init(&di->lease_list);
1391                         } else {
1392                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1393                                 list_move_tail(&di->lease_list, &dispose);
1394                                 dget_dlock(dentry);
1395                         }
1396                 }
1397 next:
1398                 spin_unlock(&dentry->d_lock);
1399                 if (ret & STOP)
1400                         break;
1401         }
1402         spin_unlock(&mdsc->dentry_list_lock);
1403
1404         while (!list_empty(&dispose)) {
1405                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1406                                       lease_list);
1407                 dentry = di->dentry;
1408                 spin_lock(&dentry->d_lock);
1409
1410                 list_del_init(&di->lease_list);
1411                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1412                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1413                         spin_lock(&mdsc->dentry_list_lock);
1414                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1415                                 list_add_tail(&di->lease_list,
1416                                               &mdsc->dentry_leases);
1417                         } else {
1418                                 __dentry_dir_lease_touch(mdsc, di);
1419                         }
1420                         spin_unlock(&mdsc->dentry_list_lock);
1421                 } else {
1422                         freed++;
1423                 }
1424
1425                 spin_unlock(&dentry->d_lock);
1426                 /* ceph_d_delete() does the trick */
1427                 dput(dentry);
1428         }
1429         return freed;
1430 }
1431
1432 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1433 {
1434         struct ceph_dentry_info *di = ceph_dentry(dentry);
1435         int ret;
1436
1437         if (__dentry_lease_is_valid(di))
1438                 return STOP;
1439         ret = __dir_lease_try_check(dentry);
1440         if (ret == -EBUSY)
1441                 return KEEP;
1442         if (ret > 0)
1443                 return TOUCH;
1444         return DELETE;
1445 }
1446
1447 static int __dir_lease_check(struct dentry *dentry, void *arg)
1448 {
1449         struct ceph_lease_walk_control *lwc = arg;
1450         struct ceph_dentry_info *di = ceph_dentry(dentry);
1451
1452         int ret = __dir_lease_try_check(dentry);
1453         if (ret == -EBUSY)
1454                 return KEEP;
1455         if (ret > 0) {
1456                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1457                         return STOP;
1458                 /* Move dentry to tail of dir lease list if we don't want
1459                  * to delete it. So dentries in the list are checked in a
1460                  * round robin manner */
1461                 if (!lwc->expire_dir_lease)
1462                         return TOUCH;
1463                 if (dentry->d_lockref.count > 0 ||
1464                     (di->flags & CEPH_DENTRY_REFERENCED))
1465                         return TOUCH;
1466                 /* invalidate dir lease */
1467                 di->lease_shared_gen = 0;
1468         }
1469         return DELETE;
1470 }
1471
1472 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1473 {
1474         struct ceph_lease_walk_control lwc;
1475         unsigned long count;
1476         unsigned long freed;
1477
1478         spin_lock(&mdsc->caps_list_lock);
1479         if (mdsc->caps_use_max > 0 &&
1480             mdsc->caps_use_count > mdsc->caps_use_max)
1481                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1482         else
1483                 count = 0;
1484         spin_unlock(&mdsc->caps_list_lock);
1485
1486         lwc.dir_lease = false;
1487         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1488         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1489         if (!lwc.nr_to_scan) /* more invalid leases */
1490                 return -EAGAIN;
1491
1492         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1493                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1494
1495         lwc.dir_lease = true;
1496         lwc.expire_dir_lease = freed < count;
1497         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1498         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1499         if (!lwc.nr_to_scan) /* more to check */
1500                 return -EAGAIN;
1501
1502         return freed > 0 ? 1 : 0;
1503 }
1504
1505 /*
1506  * Ensure a dentry lease will no longer revalidate.
1507  */
1508 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1509 {
1510         struct ceph_dentry_info *di = ceph_dentry(dentry);
1511         spin_lock(&dentry->d_lock);
1512         di->time = jiffies;
1513         di->lease_shared_gen = 0;
1514         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1515         __dentry_lease_unlist(di);
1516         spin_unlock(&dentry->d_lock);
1517 }
1518
1519 /*
1520  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1521  * renew if the least is more than half up.
1522  */
1523 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1524 {
1525         struct ceph_mds_session *session;
1526
1527         if (!di->lease_gen)
1528                 return false;
1529
1530         session = di->lease_session;
1531         if (session) {
1532                 u32 gen;
1533                 unsigned long ttl;
1534
1535                 spin_lock(&session->s_gen_ttl_lock);
1536                 gen = session->s_cap_gen;
1537                 ttl = session->s_cap_ttl;
1538                 spin_unlock(&session->s_gen_ttl_lock);
1539
1540                 if (di->lease_gen == gen &&
1541                     time_before(jiffies, ttl) &&
1542                     time_before(jiffies, di->time))
1543                         return true;
1544         }
1545         di->lease_gen = 0;
1546         return false;
1547 }
1548
1549 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1550 {
1551         struct ceph_dentry_info *di;
1552         struct ceph_mds_session *session = NULL;
1553         u32 seq = 0;
1554         int valid = 0;
1555
1556         spin_lock(&dentry->d_lock);
1557         di = ceph_dentry(dentry);
1558         if (di && __dentry_lease_is_valid(di)) {
1559                 valid = 1;
1560
1561                 if (di->lease_renew_after &&
1562                     time_after(jiffies, di->lease_renew_after)) {
1563                         /*
1564                          * We should renew. If we're in RCU walk mode
1565                          * though, we can't do that so just return
1566                          * -ECHILD.
1567                          */
1568                         if (flags & LOOKUP_RCU) {
1569                                 valid = -ECHILD;
1570                         } else {
1571                                 session = ceph_get_mds_session(di->lease_session);
1572                                 seq = di->lease_seq;
1573                                 di->lease_renew_after = 0;
1574                                 di->lease_renew_from = jiffies;
1575                         }
1576                 }
1577         }
1578         spin_unlock(&dentry->d_lock);
1579
1580         if (session) {
1581                 ceph_mdsc_lease_send_msg(session, dentry,
1582                                          CEPH_MDS_LEASE_RENEW, seq);
1583                 ceph_put_mds_session(session);
1584         }
1585         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1586         return valid;
1587 }
1588
1589 /*
1590  * Called under dentry->d_lock.
1591  */
1592 static int __dir_lease_try_check(const struct dentry *dentry)
1593 {
1594         struct ceph_dentry_info *di = ceph_dentry(dentry);
1595         struct inode *dir;
1596         struct ceph_inode_info *ci;
1597         int valid = 0;
1598
1599         if (!di->lease_shared_gen)
1600                 return 0;
1601         if (IS_ROOT(dentry))
1602                 return 0;
1603
1604         dir = d_inode(dentry->d_parent);
1605         ci = ceph_inode(dir);
1606
1607         if (spin_trylock(&ci->i_ceph_lock)) {
1608                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1609                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1610                         valid = 1;
1611                 spin_unlock(&ci->i_ceph_lock);
1612         } else {
1613                 valid = -EBUSY;
1614         }
1615
1616         if (!valid)
1617                 di->lease_shared_gen = 0;
1618         return valid;
1619 }
1620
1621 /*
1622  * Check if directory-wide content lease/cap is valid.
1623  */
1624 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1625 {
1626         struct ceph_inode_info *ci = ceph_inode(dir);
1627         int valid;
1628         int shared_gen;
1629
1630         spin_lock(&ci->i_ceph_lock);
1631         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1632         shared_gen = atomic_read(&ci->i_shared_gen);
1633         spin_unlock(&ci->i_ceph_lock);
1634         if (valid) {
1635                 struct ceph_dentry_info *di;
1636                 spin_lock(&dentry->d_lock);
1637                 di = ceph_dentry(dentry);
1638                 if (dir == d_inode(dentry->d_parent) &&
1639                     di && di->lease_shared_gen == shared_gen)
1640                         __ceph_dentry_dir_lease_touch(di);
1641                 else
1642                         valid = 0;
1643                 spin_unlock(&dentry->d_lock);
1644         }
1645         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1646              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1647         return valid;
1648 }
1649
1650 /*
1651  * Check if cached dentry can be trusted.
1652  */
1653 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1654 {
1655         int valid = 0;
1656         struct dentry *parent;
1657         struct inode *dir, *inode;
1658
1659         if (flags & LOOKUP_RCU) {
1660                 parent = READ_ONCE(dentry->d_parent);
1661                 dir = d_inode_rcu(parent);
1662                 if (!dir)
1663                         return -ECHILD;
1664                 inode = d_inode_rcu(dentry);
1665         } else {
1666                 parent = dget_parent(dentry);
1667                 dir = d_inode(parent);
1668                 inode = d_inode(dentry);
1669         }
1670
1671         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1672              dentry, inode, ceph_dentry(dentry)->offset);
1673
1674         /* always trust cached snapped dentries, snapdir dentry */
1675         if (ceph_snap(dir) != CEPH_NOSNAP) {
1676                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1677                      dentry, inode);
1678                 valid = 1;
1679         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1680                 valid = 1;
1681         } else {
1682                 valid = dentry_lease_is_valid(dentry, flags);
1683                 if (valid == -ECHILD)
1684                         return valid;
1685                 if (valid || dir_lease_is_valid(dir, dentry)) {
1686                         if (inode)
1687                                 valid = ceph_is_any_caps(inode);
1688                         else
1689                                 valid = 1;
1690                 }
1691         }
1692
1693         if (!valid) {
1694                 struct ceph_mds_client *mdsc =
1695                         ceph_sb_to_client(dir->i_sb)->mdsc;
1696                 struct ceph_mds_request *req;
1697                 int op, err;
1698                 u32 mask;
1699
1700                 if (flags & LOOKUP_RCU)
1701                         return -ECHILD;
1702
1703                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1704                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1705                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1706                 if (!IS_ERR(req)) {
1707                         req->r_dentry = dget(dentry);
1708                         req->r_num_caps = 2;
1709                         req->r_parent = dir;
1710
1711                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1712                         if (ceph_security_xattr_wanted(dir))
1713                                 mask |= CEPH_CAP_XATTR_SHARED;
1714                         req->r_args.getattr.mask = cpu_to_le32(mask);
1715
1716                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1717                         switch (err) {
1718                         case 0:
1719                                 if (d_really_is_positive(dentry) &&
1720                                     d_inode(dentry) == req->r_target_inode)
1721                                         valid = 1;
1722                                 break;
1723                         case -ENOENT:
1724                                 if (d_really_is_negative(dentry))
1725                                         valid = 1;
1726                                 /* Fallthrough */
1727                         default:
1728                                 break;
1729                         }
1730                         ceph_mdsc_put_request(req);
1731                         dout("d_revalidate %p lookup result=%d\n",
1732                              dentry, err);
1733                 }
1734         }
1735
1736         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1737         if (!valid)
1738                 ceph_dir_clear_complete(dir);
1739
1740         if (!(flags & LOOKUP_RCU))
1741                 dput(parent);
1742         return valid;
1743 }
1744
1745 /*
1746  * Delete unused dentry that doesn't have valid lease
1747  *
1748  * Called under dentry->d_lock.
1749  */
1750 static int ceph_d_delete(const struct dentry *dentry)
1751 {
1752         struct ceph_dentry_info *di;
1753
1754         /* won't release caps */
1755         if (d_really_is_negative(dentry))
1756                 return 0;
1757         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1758                 return 0;
1759         /* vaild lease? */
1760         di = ceph_dentry(dentry);
1761         if (di) {
1762                 if (__dentry_lease_is_valid(di))
1763                         return 0;
1764                 if (__dir_lease_try_check(dentry))
1765                         return 0;
1766         }
1767         return 1;
1768 }
1769
1770 /*
1771  * Release our ceph_dentry_info.
1772  */
1773 static void ceph_d_release(struct dentry *dentry)
1774 {
1775         struct ceph_dentry_info *di = ceph_dentry(dentry);
1776
1777         dout("d_release %p\n", dentry);
1778
1779         spin_lock(&dentry->d_lock);
1780         __dentry_lease_unlist(di);
1781         dentry->d_fsdata = NULL;
1782         spin_unlock(&dentry->d_lock);
1783
1784         if (di->lease_session)
1785                 ceph_put_mds_session(di->lease_session);
1786         kmem_cache_free(ceph_dentry_cachep, di);
1787 }
1788
1789 /*
1790  * When the VFS prunes a dentry from the cache, we need to clear the
1791  * complete flag on the parent directory.
1792  *
1793  * Called under dentry->d_lock.
1794  */
1795 static void ceph_d_prune(struct dentry *dentry)
1796 {
1797         struct ceph_inode_info *dir_ci;
1798         struct ceph_dentry_info *di;
1799
1800         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1801
1802         /* do we have a valid parent? */
1803         if (IS_ROOT(dentry))
1804                 return;
1805
1806         /* we hold d_lock, so d_parent is stable */
1807         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1808         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1809                 return;
1810
1811         /* who calls d_delete() should also disable dcache readdir */
1812         if (d_really_is_negative(dentry))
1813                 return;
1814
1815         /* d_fsdata does not get cleared until d_release */
1816         if (!d_unhashed(dentry)) {
1817                 __ceph_dir_clear_complete(dir_ci);
1818                 return;
1819         }
1820
1821         /* Disable dcache readdir just in case that someone called d_drop()
1822          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1823          * properly (dcache readdir is still enabled) */
1824         di = ceph_dentry(dentry);
1825         if (di->offset > 0 &&
1826             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1827                 __ceph_dir_clear_ordered(dir_ci);
1828 }
1829
1830 /*
1831  * read() on a dir.  This weird interface hack only works if mounted
1832  * with '-o dirstat'.
1833  */
1834 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1835                              loff_t *ppos)
1836 {
1837         struct ceph_dir_file_info *dfi = file->private_data;
1838         struct inode *inode = file_inode(file);
1839         struct ceph_inode_info *ci = ceph_inode(inode);
1840         int left;
1841         const int bufsize = 1024;
1842
1843         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1844                 return -EISDIR;
1845
1846         if (!dfi->dir_info) {
1847                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1848                 if (!dfi->dir_info)
1849                         return -ENOMEM;
1850                 dfi->dir_info_len =
1851                         snprintf(dfi->dir_info, bufsize,
1852                                 "entries:   %20lld\n"
1853                                 " files:    %20lld\n"
1854                                 " subdirs:  %20lld\n"
1855                                 "rentries:  %20lld\n"
1856                                 " rfiles:   %20lld\n"
1857                                 " rsubdirs: %20lld\n"
1858                                 "rbytes:    %20lld\n"
1859                                 "rctime:    %10lld.%09ld\n",
1860                                 ci->i_files + ci->i_subdirs,
1861                                 ci->i_files,
1862                                 ci->i_subdirs,
1863                                 ci->i_rfiles + ci->i_rsubdirs,
1864                                 ci->i_rfiles,
1865                                 ci->i_rsubdirs,
1866                                 ci->i_rbytes,
1867                                 ci->i_rctime.tv_sec,
1868                                 ci->i_rctime.tv_nsec);
1869         }
1870
1871         if (*ppos >= dfi->dir_info_len)
1872                 return 0;
1873         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1874         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1875         if (left == size)
1876                 return -EFAULT;
1877         *ppos += (size - left);
1878         return size - left;
1879 }
1880
1881
1882
1883 /*
1884  * Return name hash for a given dentry.  This is dependent on
1885  * the parent directory's hash function.
1886  */
1887 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1888 {
1889         struct ceph_inode_info *dci = ceph_inode(dir);
1890         unsigned hash;
1891
1892         switch (dci->i_dir_layout.dl_dir_hash) {
1893         case 0: /* for backward compat */
1894         case CEPH_STR_HASH_LINUX:
1895                 return dn->d_name.hash;
1896
1897         default:
1898                 spin_lock(&dn->d_lock);
1899                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1900                                      dn->d_name.name, dn->d_name.len);
1901                 spin_unlock(&dn->d_lock);
1902                 return hash;
1903         }
1904 }
1905
1906 const struct file_operations ceph_dir_fops = {
1907         .read = ceph_read_dir,
1908         .iterate = ceph_readdir,
1909         .llseek = ceph_dir_llseek,
1910         .open = ceph_open,
1911         .release = ceph_release,
1912         .unlocked_ioctl = ceph_ioctl,
1913         .compat_ioctl = compat_ptr_ioctl,
1914         .fsync = ceph_fsync,
1915         .lock = ceph_lock,
1916         .flock = ceph_flock,
1917 };
1918
1919 const struct file_operations ceph_snapdir_fops = {
1920         .iterate = ceph_readdir,
1921         .llseek = ceph_dir_llseek,
1922         .open = ceph_open,
1923         .release = ceph_release,
1924 };
1925
1926 const struct inode_operations ceph_dir_iops = {
1927         .lookup = ceph_lookup,
1928         .permission = ceph_permission,
1929         .getattr = ceph_getattr,
1930         .setattr = ceph_setattr,
1931         .listxattr = ceph_listxattr,
1932         .get_acl = ceph_get_acl,
1933         .set_acl = ceph_set_acl,
1934         .mknod = ceph_mknod,
1935         .symlink = ceph_symlink,
1936         .mkdir = ceph_mkdir,
1937         .link = ceph_link,
1938         .unlink = ceph_unlink,
1939         .rmdir = ceph_unlink,
1940         .rename = ceph_rename,
1941         .create = ceph_create,
1942         .atomic_open = ceph_atomic_open,
1943 };
1944
1945 const struct inode_operations ceph_snapdir_iops = {
1946         .lookup = ceph_lookup,
1947         .permission = ceph_permission,
1948         .getattr = ceph_getattr,
1949         .mkdir = ceph_mkdir,
1950         .rmdir = ceph_unlink,
1951         .rename = ceph_rename,
1952 };
1953
1954 const struct dentry_operations ceph_dentry_ops = {
1955         .d_revalidate = ceph_d_revalidate,
1956         .d_delete = ceph_d_delete,
1957         .d_release = ceph_d_release,
1958         .d_prune = ceph_d_prune,
1959         .d_init = ceph_d_init,
1960 };