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         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dentry->d_sb);
  42
  43         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  44         if (!di)
  45                 return -ENOMEM;          /* oh well */
  46
  47         di->dentry = dentry;
  48         di->lease_session = NULL;
  49         di->time = jiffies;
  50         dentry->d_fsdata = di;
  51         INIT_LIST_HEAD(&di->lease_list);
  52
  53         atomic64_inc(&mdsc->metric.total_dentries);
  54
  55         return 0;
  56 }
  57
  58 /*
  59  * for f_pos for readdir:
  60  * - hash order:
  61  *      (0xff << 52) | ((24 bits hash) << 28) |
  62  *      (the nth entry has hash collision);
  63  * - frag+name order;
  64  *      ((frag value) << 28) | (the nth entry in frag);
  65  */
  66 #define OFFSET_BITS     28
  67 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  68 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  69 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  70 {
  71         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  72         if (hash_order)
  73                 fpos |= HASH_ORDER;
  74         return fpos;
  75 }
  76
  77 static bool is_hash_order(loff_t p)
  78 {
  79         return (p & HASH_ORDER) == HASH_ORDER;
  80 }
  81
  82 static unsigned fpos_frag(loff_t p)
  83 {
  84         return p >> OFFSET_BITS;
  85 }
  86
  87 static unsigned fpos_hash(loff_t p)
  88 {
  89         return ceph_frag_value(fpos_frag(p));
  90 }
  91
  92 static unsigned fpos_off(loff_t p)
  93 {
  94         return p & OFFSET_MASK;
  95 }
  96
  97 static int fpos_cmp(loff_t l, loff_t r)
  98 {
  99         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
 100         if (v)
 101                 return v;
 102         return (int)(fpos_off(l) - fpos_off(r));
 103 }
 104
 105 /*
 106  * make note of the last dentry we read, so we can
 107  * continue at the same lexicographical point,
 108  * regardless of what dir changes take place on the
 109  * server.
 110  */
 111 static int note_last_dentry(struct ceph_dir_file_info *dfi, const char *name,
 112                             int len, unsigned next_offset)
 113 {
 114         char *buf = kmalloc(len+1, GFP_KERNEL);
 115         if (!buf)
 116                 return -ENOMEM;
 117         kfree(dfi->last_name);
 118         dfi->last_name = buf;
 119         memcpy(dfi->last_name, name, len);
 120         dfi->last_name[len] = 0;
 121         dfi->next_offset = next_offset;
 122         dout("note_last_dentry '%s'\n", dfi->last_name);
 123         return 0;
 124 }
 125
 126
 127 static struct dentry *
 128 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 129                         struct ceph_readdir_cache_control *cache_ctl)
 130 {
 131         struct inode *dir = d_inode(parent);
 132         struct dentry *dentry;
 133         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 134         loff_t ptr_pos = idx * sizeof(struct dentry *);
 135         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 136
 137         if (ptr_pos >= i_size_read(dir))
 138                 return NULL;
 139
 140         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 141                 ceph_readdir_cache_release(cache_ctl);
 142                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 143                 if (!cache_ctl->page) {
 144                         dout(" page %lu not found\n", ptr_pgoff);
 145                         return ERR_PTR(-EAGAIN);
 146                 }
 147                 /* reading/filling the cache are serialized by
 148                    i_mutex, no need to use page lock */
 149                 unlock_page(cache_ctl->page);
 150                 cache_ctl->dentries = kmap(cache_ctl->page);
 151         }
 152
 153         cache_ctl->index = idx & idx_mask;
 154
 155         rcu_read_lock();
 156         spin_lock(&parent->d_lock);
 157         /* check i_size again here, because empty directory can be
 158          * marked as complete while not holding the i_mutex. */
 159         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 160                 dentry = cache_ctl->dentries[cache_ctl->index];
 161         else
 162                 dentry = NULL;
 163         spin_unlock(&parent->d_lock);
 164         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 165                 dentry = NULL;
 166         rcu_read_unlock();
 167         return dentry ? : ERR_PTR(-EAGAIN);
 168 }
 169
 170 /*
 171  * When possible, we try to satisfy a readdir by peeking at the
 172  * dcache.  We make this work by carefully ordering dentries on
 173  * d_child when we initially get results back from the MDS, and
 174  * falling back to a "normal" sync readdir if any dentries in the dir
 175  * are dropped.
 176  *
 177  * Complete dir indicates that we have all dentries in the dir.  It is
 178  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 179  * the MDS if/when the directory is modified).
 180  */
 181 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 182                             int shared_gen)
 183 {
 184         struct ceph_dir_file_info *dfi = file->private_data;
 185         struct dentry *parent = file->f_path.dentry;
 186         struct inode *dir = d_inode(parent);
 187         struct dentry *dentry, *last = NULL;
 188         struct ceph_dentry_info *di;
 189         struct ceph_readdir_cache_control cache_ctl = {};
 190         u64 idx = 0;
 191         int err = 0;
 192
 193         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 194
 195         /* search start position */
 196         if (ctx->pos > 2) {
 197                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 198                 while (count > 0) {
 199                         u64 step = count >> 1;
 200                         dentry = __dcache_find_get_entry(parent, idx + step,
 201                                                          &cache_ctl);
 202                         if (!dentry) {
 203                                 /* use linar search */
 204                                 idx = 0;
 205                                 break;
 206                         }
 207                         if (IS_ERR(dentry)) {
 208                                 err = PTR_ERR(dentry);
 209                                 goto out;
 210                         }
 211                         di = ceph_dentry(dentry);
 212                         spin_lock(&dentry->d_lock);
 213                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 214                                 idx += step + 1;
 215                                 count -= step + 1;
 216                         } else {
 217                                 count = step;
 218                         }
 219                         spin_unlock(&dentry->d_lock);
 220                         dput(dentry);
 221                 }
 222
 223                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 224         }
 225
 226
 227         for (;;) {
 228                 bool emit_dentry = false;
 229                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 230                 if (!dentry) {
 231                         dfi->file_info.flags |= CEPH_F_ATEND;
 232                         err = 0;
 233                         break;
 234                 }
 235                 if (IS_ERR(dentry)) {
 236                         err = PTR_ERR(dentry);
 237                         goto out;
 238                 }
 239
 240                 spin_lock(&dentry->d_lock);
 241                 di = ceph_dentry(dentry);
 242                 if (d_unhashed(dentry) ||
 243                     d_really_is_negative(dentry) ||
 244                     di->lease_shared_gen != shared_gen) {
 245                         spin_unlock(&dentry->d_lock);
 246                         dput(dentry);
 247                         err = -EAGAIN;
 248                         goto out;
 249                 }
 250                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 251                         __ceph_dentry_dir_lease_touch(di);
 252                         emit_dentry = true;
 253                 }
 254                 spin_unlock(&dentry->d_lock);
 255
 256                 if (emit_dentry) {
 257                         dout(" %llx dentry %p %pd %p\n", di->offset,
 258                              dentry, dentry, d_inode(dentry));
 259                         ctx->pos = di->offset;
 260                         if (!dir_emit(ctx, dentry->d_name.name,
 261                                       dentry->d_name.len, ceph_present_inode(d_inode(dentry)),
 262                                       d_inode(dentry)->i_mode >> 12)) {
 263                                 dput(dentry);
 264                                 err = 0;
 265                                 break;
 266                         }
 267                         ctx->pos++;
 268
 269                         if (last)
 270                                 dput(last);
 271                         last = dentry;
 272                 } else {
 273                         dput(dentry);
 274                 }
 275         }
 276 out:
 277         ceph_readdir_cache_release(&cache_ctl);
 278         if (last) {
 279                 int ret;
 280                 di = ceph_dentry(last);
 281                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 282                                        fpos_off(di->offset) + 1);
 283                 if (ret < 0)
 284                         err = ret;
 285                 dput(last);
 286                 /* last_name no longer match cache index */
 287                 if (dfi->readdir_cache_idx >= 0) {
 288                         dfi->readdir_cache_idx = -1;
 289                         dfi->dir_release_count = 0;
 290                 }
 291         }
 292         return err;
 293 }
 294
 295 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 296 {
 297         if (!dfi->last_readdir)
 298                 return true;
 299         if (is_hash_order(pos))
 300                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 301         else
 302                 return dfi->frag != fpos_frag(pos);
 303 }
 304
 305 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 306 {
 307         struct ceph_dir_file_info *dfi = file->private_data;
 308         struct inode *inode = file_inode(file);
 309         struct ceph_inode_info *ci = ceph_inode(inode);
 310         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 311         struct ceph_mds_client *mdsc = fsc->mdsc;
 312         int i;
 313         int err;
 314         unsigned frag = -1;
 315         struct ceph_mds_reply_info_parsed *rinfo;
 316
 317         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 318         if (dfi->file_info.flags & CEPH_F_ATEND)
 319                 return 0;
 320
 321         /* always start with . and .. */
 322         if (ctx->pos == 0) {
 323                 dout("readdir off 0 -> '.'\n");
 324                 if (!dir_emit(ctx, ".", 1, ceph_present_inode(inode),
 325                             inode->i_mode >> 12))
 326                         return 0;
 327                 ctx->pos = 1;
 328         }
 329         if (ctx->pos == 1) {
 330                 u64 ino;
 331                 struct dentry *dentry = file->f_path.dentry;
 332
 333                 spin_lock(&dentry->d_lock);
 334                 ino = ceph_present_inode(dentry->d_parent->d_inode);
 335                 spin_unlock(&dentry->d_lock);
 336
 337                 dout("readdir off 1 -> '..'\n");
 338                 if (!dir_emit(ctx, "..", 2, ino, inode->i_mode >> 12))
 339                         return 0;
 340                 ctx->pos = 2;
 341         }
 342
 343         spin_lock(&ci->i_ceph_lock);
 344         /* request Fx cap. if have Fx, we don't need to release Fs cap
 345          * for later create/unlink. */
 346         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_WR);
 347         /* can we use the dcache? */
 348         if (ceph_test_mount_opt(fsc, DCACHE) &&
 349             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 350             ceph_snap(inode) != CEPH_SNAPDIR &&
 351             __ceph_dir_is_complete_ordered(ci) &&
 352             __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 353                 int shared_gen = atomic_read(&ci->i_shared_gen);
 354
 355                 spin_unlock(&ci->i_ceph_lock);
 356                 err = __dcache_readdir(file, ctx, shared_gen);
 357                 if (err != -EAGAIN)
 358                         return err;
 359         } else {
 360                 spin_unlock(&ci->i_ceph_lock);
 361         }
 362
 363         /* proceed with a normal readdir */
 364 more:
 365         /* do we have the correct frag content buffered? */
 366         if (need_send_readdir(dfi, ctx->pos)) {
 367                 struct ceph_mds_request *req;
 368                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 369                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 370
 371                 /* discard old result, if any */
 372                 if (dfi->last_readdir) {
 373                         ceph_mdsc_put_request(dfi->last_readdir);
 374                         dfi->last_readdir = NULL;
 375                 }
 376
 377                 if (is_hash_order(ctx->pos)) {
 378                         /* fragtree isn't always accurate. choose frag
 379                          * based on previous reply when possible. */
 380                         if (frag == (unsigned)-1)
 381                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 382                                                         NULL, NULL);
 383                 } else {
 384                         frag = fpos_frag(ctx->pos);
 385                 }
 386
 387                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 388                      ceph_vinop(inode), frag, dfi->last_name);
 389                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 390                 if (IS_ERR(req))
 391                         return PTR_ERR(req);
 392                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 393                 if (err) {
 394                         ceph_mdsc_put_request(req);
 395                         return err;
 396                 }
 397                 /* hints to request -> mds selection code */
 398                 req->r_direct_mode = USE_AUTH_MDS;
 399                 if (op == CEPH_MDS_OP_READDIR) {
 400                         req->r_direct_hash = ceph_frag_value(frag);
 401                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 402                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 403                 }
 404                 if (dfi->last_name) {
 405                         req->r_path2 = kstrdup(dfi->last_name, GFP_KERNEL);
 406                         if (!req->r_path2) {
 407                                 ceph_mdsc_put_request(req);
 408                                 return -ENOMEM;
 409                         }
 410                 } else if (is_hash_order(ctx->pos)) {
 411                         req->r_args.readdir.offset_hash =
 412                                 cpu_to_le32(fpos_hash(ctx->pos));
 413                 }
 414
 415                 req->r_dir_release_cnt = dfi->dir_release_count;
 416                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 417                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 418                 req->r_readdir_offset = dfi->next_offset;
 419                 req->r_args.readdir.frag = cpu_to_le32(frag);
 420                 req->r_args.readdir.flags =
 421                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 422
 423                 req->r_inode = inode;
 424                 ihold(inode);
 425                 req->r_dentry = dget(file->f_path.dentry);
 426                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 427                 if (err < 0) {
 428                         ceph_mdsc_put_request(req);
 429                         return err;
 430                 }
 431                 dout("readdir got and parsed readdir result=%d on "
 432                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 433                      err, frag,
 434                      (int)req->r_reply_info.dir_end,
 435                      (int)req->r_reply_info.dir_complete,
 436                      (int)req->r_reply_info.hash_order);
 437
 438                 rinfo = &req->r_reply_info;
 439                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 440                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 441                         if (!rinfo->hash_order) {
 442                                 dfi->next_offset = req->r_readdir_offset;
 443                                 /* adjust ctx->pos to beginning of frag */
 444                                 ctx->pos = ceph_make_fpos(frag,
 445                                                           dfi->next_offset,
 446                                                           false);
 447                         }
 448                 }
 449
 450                 dfi->frag = frag;
 451                 dfi->last_readdir = req;
 452
 453                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 454                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 455                         if (dfi->readdir_cache_idx < 0) {
 456                                 /* preclude from marking dir ordered */
 457                                 dfi->dir_ordered_count = 0;
 458                         } else if (ceph_frag_is_leftmost(frag) &&
 459                                    dfi->next_offset == 2) {
 460                                 /* note dir version at start of readdir so
 461                                  * we can tell if any dentries get dropped */
 462                                 dfi->dir_release_count = req->r_dir_release_cnt;
 463                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 464                         }
 465                 } else {
 466                         dout("readdir !did_prepopulate\n");
 467                         /* disable readdir cache */
 468                         dfi->readdir_cache_idx = -1;
 469                         /* preclude from marking dir complete */
 470                         dfi->dir_release_count = 0;
 471                 }
 472
 473                 /* note next offset and last dentry name */
 474                 if (rinfo->dir_nr > 0) {
 475                         struct ceph_mds_reply_dir_entry *rde =
 476                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 477                         unsigned next_offset = req->r_reply_info.dir_end ?
 478                                         2 : (fpos_off(rde->offset) + 1);
 479                         err = note_last_dentry(dfi, rde->name, rde->name_len,
 480                                                next_offset);
 481                         if (err) {
 482                                 ceph_mdsc_put_request(dfi->last_readdir);
 483                                 dfi->last_readdir = NULL;
 484                                 return err;
 485                         }
 486                 } else if (req->r_reply_info.dir_end) {
 487                         dfi->next_offset = 2;
 488                         /* keep last name */
 489                 }
 490         }
 491
 492         rinfo = &dfi->last_readdir->r_reply_info;
 493         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 494              dfi->frag, rinfo->dir_nr, ctx->pos,
 495              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 496
 497         i = 0;
 498         /* search start position */
 499         if (rinfo->dir_nr > 0) {
 500                 int step, nr = rinfo->dir_nr;
 501                 while (nr > 0) {
 502                         step = nr >> 1;
 503                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 504                                 i +=  step + 1;
 505                                 nr -= step + 1;
 506                         } else {
 507                                 nr = step;
 508                         }
 509                 }
 510         }
 511         for (; i < rinfo->dir_nr; i++) {
 512                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 513
 514                 BUG_ON(rde->offset < ctx->pos);
 515
 516                 ctx->pos = rde->offset;
 517                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 518                      i, rinfo->dir_nr, ctx->pos,
 519                      rde->name_len, rde->name, &rde->inode.in);
 520
 521                 BUG_ON(!rde->inode.in);
 522
 523                 if (!dir_emit(ctx, rde->name, rde->name_len,
 524                               ceph_present_ino(inode->i_sb, le64_to_cpu(rde->inode.in->ino)),
 525                               le32_to_cpu(rde->inode.in->mode) >> 12)) {
 526                         /*
 527                          * NOTE: Here no need to put the 'dfi->last_readdir',
 528                          * because when dir_emit stops us it's most likely
 529                          * doesn't have enough memory, etc. So for next readdir
 530                          * it will continue.
 531                          */
 532                         dout("filldir stopping us...\n");
 533                         return 0;
 534                 }
 535                 ctx->pos++;
 536         }
 537
 538         ceph_mdsc_put_request(dfi->last_readdir);
 539         dfi->last_readdir = NULL;
 540
 541         if (dfi->next_offset > 2) {
 542                 frag = dfi->frag;
 543                 goto more;
 544         }
 545
 546         /* more frags? */
 547         if (!ceph_frag_is_rightmost(dfi->frag)) {
 548                 frag = ceph_frag_next(dfi->frag);
 549                 if (is_hash_order(ctx->pos)) {
 550                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 551                                                         dfi->next_offset, true);
 552                         if (new_pos > ctx->pos)
 553                                 ctx->pos = new_pos;
 554                         /* keep last_name */
 555                 } else {
 556                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 557                                                         false);
 558                         kfree(dfi->last_name);
 559                         dfi->last_name = NULL;
 560                 }
 561                 dout("readdir next frag is %x\n", frag);
 562                 goto more;
 563         }
 564         dfi->file_info.flags |= CEPH_F_ATEND;
 565
 566         /*
 567          * if dir_release_count still matches the dir, no dentries
 568          * were released during the whole readdir, and we should have
 569          * the complete dir contents in our cache.
 570          */
 571         if (atomic64_read(&ci->i_release_count) ==
 572                                         dfi->dir_release_count) {
 573                 spin_lock(&ci->i_ceph_lock);
 574                 if (dfi->dir_ordered_count ==
 575                                 atomic64_read(&ci->i_ordered_count)) {
 576                         dout(" marking %p complete and ordered\n", inode);
 577                         /* use i_size to track number of entries in
 578                          * readdir cache */
 579                         BUG_ON(dfi->readdir_cache_idx < 0);
 580                         i_size_write(inode, dfi->readdir_cache_idx *
 581                                      sizeof(struct dentry*));
 582                 } else {
 583                         dout(" marking %p complete\n", inode);
 584                 }
 585                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 586                                         dfi->dir_ordered_count);
 587                 spin_unlock(&ci->i_ceph_lock);
 588         }
 589
 590         dout("readdir %p file %p done.\n", inode, file);
 591         return 0;
 592 }
 593
 594 static void reset_readdir(struct ceph_dir_file_info *dfi)
 595 {
 596         if (dfi->last_readdir) {
 597                 ceph_mdsc_put_request(dfi->last_readdir);
 598                 dfi->last_readdir = NULL;
 599         }
 600         kfree(dfi->last_name);
 601         dfi->last_name = NULL;
 602         dfi->dir_release_count = 0;
 603         dfi->readdir_cache_idx = -1;
 604         dfi->next_offset = 2;  /* compensate for . and .. */
 605         dfi->file_info.flags &= ~CEPH_F_ATEND;
 606 }
 607
 608 /*
 609  * discard buffered readdir content on seekdir(0), or seek to new frag,
 610  * or seek prior to current chunk
 611  */
 612 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 613 {
 614         struct ceph_mds_reply_info_parsed *rinfo;
 615         loff_t chunk_offset;
 616         if (new_pos == 0)
 617                 return true;
 618         if (is_hash_order(new_pos)) {
 619                 /* no need to reset last_name for a forward seek when
 620                  * dentries are sotred in hash order */
 621         } else if (dfi->frag != fpos_frag(new_pos)) {
 622                 return true;
 623         }
 624         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 625         if (!rinfo || !rinfo->dir_nr)
 626                 return true;
 627         chunk_offset = rinfo->dir_entries[0].offset;
 628         return new_pos < chunk_offset ||
 629                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 630 }
 631
 632 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 633 {
 634         struct ceph_dir_file_info *dfi = file->private_data;
 635         struct inode *inode = file->f_mapping->host;
 636         loff_t retval;
 637
 638         inode_lock(inode);
 639         retval = -EINVAL;
 640         switch (whence) {
 641         case SEEK_CUR:
 642                 offset += file->f_pos;
 643         case SEEK_SET:
 644                 break;
 645         case SEEK_END:
 646                 retval = -EOPNOTSUPP;
 647         default:
 648                 goto out;
 649         }
 650
 651         if (offset >= 0) {
 652                 if (need_reset_readdir(dfi, offset)) {
 653                         dout("dir_llseek dropping %p content\n", file);
 654                         reset_readdir(dfi);
 655                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 656                         /* for hash offset, we don't know if a forward seek
 657                          * is within same frag */
 658                         dfi->dir_release_count = 0;
 659                         dfi->readdir_cache_idx = -1;
 660                 }
 661
 662                 if (offset != file->f_pos) {
 663                         file->f_pos = offset;
 664                         file->f_version = 0;
 665                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 666                 }
 667                 retval = offset;
 668         }
 669 out:
 670         inode_unlock(inode);
 671         return retval;
 672 }
 673
 674 /*
 675  * Handle lookups for the hidden .snap directory.
 676  */
 677 int ceph_handle_snapdir(struct ceph_mds_request *req,
 678                         struct dentry *dentry, int err)
 679 {
 680         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 681         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 682
 683         /* .snap dir? */
 684         if (err == -ENOENT &&
 685             ceph_snap(parent) == CEPH_NOSNAP &&
 686             strcmp(dentry->d_name.name,
 687                    fsc->mount_options->snapdir_name) == 0) {
 688                 struct inode *inode = ceph_get_snapdir(parent);
 689                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 690                      dentry, dentry, inode);
 691                 BUG_ON(!d_unhashed(dentry));
 692                 d_add(dentry, inode);
 693                 err = 0;
 694         }
 695         return err;
 696 }
 697
 698 /*
 699  * Figure out final result of a lookup/open request.
 700  *
 701  * Mainly, make sure we return the final req->r_dentry (if it already
 702  * existed) in place of the original VFS-provided dentry when they
 703  * differ.
 704  *
 705  * Gracefully handle the case where the MDS replies with -ENOENT and
 706  * no trace (which it may do, at its discretion, e.g., if it doesn't
 707  * care to issue a lease on the negative dentry).
 708  */
 709 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 710                                   struct dentry *dentry, int err)
 711 {
 712         if (err == -ENOENT) {
 713                 /* no trace? */
 714                 err = 0;
 715                 if (!req->r_reply_info.head->is_dentry) {
 716                         dout("ENOENT and no trace, dentry %p inode %p\n",
 717                              dentry, d_inode(dentry));
 718                         if (d_really_is_positive(dentry)) {
 719                                 d_drop(dentry);
 720                                 err = -ENOENT;
 721                         } else {
 722                                 d_add(dentry, NULL);
 723                         }
 724                 }
 725         }
 726         if (err)
 727                 dentry = ERR_PTR(err);
 728         else if (dentry != req->r_dentry)
 729                 dentry = dget(req->r_dentry);   /* we got spliced */
 730         else
 731                 dentry = NULL;
 732         return dentry;
 733 }
 734
 735 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 736 {
 737         return ceph_ino(inode) == CEPH_INO_ROOT &&
 738                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 739 }
 740
 741 /*
 742  * Look up a single dir entry.  If there is a lookup intent, inform
 743  * the MDS so that it gets our 'caps wanted' value in a single op.
 744  */
 745 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 746                                   unsigned int flags)
 747 {
 748         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 749         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 750         struct ceph_mds_request *req;
 751         int op;
 752         int mask;
 753         int err;
 754
 755         dout("lookup %p dentry %p '%pd'\n",
 756              dir, dentry, dentry);
 757
 758         if (dentry->d_name.len > NAME_MAX)
 759                 return ERR_PTR(-ENAMETOOLONG);
 760
 761         /* can we conclude ENOENT locally? */
 762         if (d_really_is_negative(dentry)) {
 763                 struct ceph_inode_info *ci = ceph_inode(dir);
 764                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 765
 766                 spin_lock(&ci->i_ceph_lock);
 767                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 768                 if (strncmp(dentry->d_name.name,
 769                             fsc->mount_options->snapdir_name,
 770                             dentry->d_name.len) &&
 771                     !is_root_ceph_dentry(dir, dentry) &&
 772                     ceph_test_mount_opt(fsc, DCACHE) &&
 773                     __ceph_dir_is_complete(ci) &&
 774                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 775                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 776                         spin_unlock(&ci->i_ceph_lock);
 777                         dout(" dir %p complete, -ENOENT\n", dir);
 778                         d_add(dentry, NULL);
 779                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 780                         return NULL;
 781                 }
 782                 spin_unlock(&ci->i_ceph_lock);
 783         }
 784
 785         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 786                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 787         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 788         if (IS_ERR(req))
 789                 return ERR_CAST(req);
 790         req->r_dentry = dget(dentry);
 791         req->r_num_caps = 2;
 792
 793         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 794         if (ceph_security_xattr_wanted(dir))
 795                 mask |= CEPH_CAP_XATTR_SHARED;
 796         req->r_args.getattr.mask = cpu_to_le32(mask);
 797
 798         req->r_parent = dir;
 799         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 800         err = ceph_mdsc_do_request(mdsc, NULL, req);
 801         err = ceph_handle_snapdir(req, dentry, err);
 802         dentry = ceph_finish_lookup(req, dentry, err);
 803         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 804         dout("lookup result=%p\n", dentry);
 805         return dentry;
 806 }
 807
 808 /*
 809  * If we do a create but get no trace back from the MDS, follow up with
 810  * a lookup (the VFS expects us to link up the provided dentry).
 811  */
 812 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 813 {
 814         struct dentry *result = ceph_lookup(dir, dentry, 0);
 815
 816         if (result && !IS_ERR(result)) {
 817                 /*
 818                  * We created the item, then did a lookup, and found
 819                  * it was already linked to another inode we already
 820                  * had in our cache (and thus got spliced). To not
 821                  * confuse VFS (especially when inode is a directory),
 822                  * we don't link our dentry to that inode, return an
 823                  * error instead.
 824                  *
 825                  * This event should be rare and it happens only when
 826                  * we talk to old MDS. Recent MDS does not send traceless
 827                  * reply for request that creates new inode.
 828                  */
 829                 d_drop(result);
 830                 return -ESTALE;
 831         }
 832         return PTR_ERR(result);
 833 }
 834
 835 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 836                       umode_t mode, dev_t rdev)
 837 {
 838         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 839         struct ceph_mds_request *req;
 840         struct ceph_acl_sec_ctx as_ctx = {};
 841         int err;
 842
 843         if (ceph_snap(dir) != CEPH_NOSNAP)
 844                 return -EROFS;
 845
 846         if (ceph_quota_is_max_files_exceeded(dir)) {
 847                 err = -EDQUOT;
 848                 goto out;
 849         }
 850
 851         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 852         if (err < 0)
 853                 goto out;
 854         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 855         if (err < 0)
 856                 goto out;
 857
 858         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 859              dir, dentry, mode, rdev);
 860         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 861         if (IS_ERR(req)) {
 862                 err = PTR_ERR(req);
 863                 goto out;
 864         }
 865         req->r_dentry = dget(dentry);
 866         req->r_num_caps = 2;
 867         req->r_parent = dir;
 868         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 869         req->r_args.mknod.mode = cpu_to_le32(mode);
 870         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 871         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 872         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 873         if (as_ctx.pagelist) {
 874                 req->r_pagelist = as_ctx.pagelist;
 875                 as_ctx.pagelist = NULL;
 876         }
 877         err = ceph_mdsc_do_request(mdsc, dir, req);
 878         if (!err && !req->r_reply_info.head->is_dentry)
 879                 err = ceph_handle_notrace_create(dir, dentry);
 880         ceph_mdsc_put_request(req);
 881 out:
 882         if (!err)
 883                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 884         else
 885                 d_drop(dentry);
 886         ceph_release_acl_sec_ctx(&as_ctx);
 887         return err;
 888 }
 889
 890 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 891                        bool excl)
 892 {
 893         return ceph_mknod(dir, dentry, mode, 0);
 894 }
 895
 896 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 897                             const char *dest)
 898 {
 899         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 900         struct ceph_mds_request *req;
 901         struct ceph_acl_sec_ctx as_ctx = {};
 902         int err;
 903
 904         if (ceph_snap(dir) != CEPH_NOSNAP)
 905                 return -EROFS;
 906
 907         if (ceph_quota_is_max_files_exceeded(dir)) {
 908                 err = -EDQUOT;
 909                 goto out;
 910         }
 911
 912         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 913         if (err < 0)
 914                 goto out;
 915
 916         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 917         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 918         if (IS_ERR(req)) {
 919                 err = PTR_ERR(req);
 920                 goto out;
 921         }
 922         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 923         if (!req->r_path2) {
 924                 err = -ENOMEM;
 925                 ceph_mdsc_put_request(req);
 926                 goto out;
 927         }
 928         req->r_parent = dir;
 929         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 930         req->r_dentry = dget(dentry);
 931         req->r_num_caps = 2;
 932         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 933         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 934         if (as_ctx.pagelist) {
 935                 req->r_pagelist = as_ctx.pagelist;
 936                 as_ctx.pagelist = NULL;
 937         }
 938         err = ceph_mdsc_do_request(mdsc, dir, req);
 939         if (!err && !req->r_reply_info.head->is_dentry)
 940                 err = ceph_handle_notrace_create(dir, dentry);
 941         ceph_mdsc_put_request(req);
 942 out:
 943         if (err)
 944                 d_drop(dentry);
 945         ceph_release_acl_sec_ctx(&as_ctx);
 946         return err;
 947 }
 948
 949 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 950 {
 951         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 952         struct ceph_mds_request *req;
 953         struct ceph_acl_sec_ctx as_ctx = {};
 954         int err = -EROFS;
 955         int op;
 956
 957         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 958                 /* mkdir .snap/foo is a MKSNAP */
 959                 op = CEPH_MDS_OP_MKSNAP;
 960                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 961                      dentry, dentry);
 962         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 963                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 964                 op = CEPH_MDS_OP_MKDIR;
 965         } else {
 966                 goto out;
 967         }
 968
 969         if (op == CEPH_MDS_OP_MKDIR &&
 970             ceph_quota_is_max_files_exceeded(dir)) {
 971                 err = -EDQUOT;
 972                 goto out;
 973         }
 974
 975         mode |= S_IFDIR;
 976         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 977         if (err < 0)
 978                 goto out;
 979         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 980         if (err < 0)
 981                 goto out;
 982
 983         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 984         if (IS_ERR(req)) {
 985                 err = PTR_ERR(req);
 986                 goto out;
 987         }
 988
 989         req->r_dentry = dget(dentry);
 990         req->r_num_caps = 2;
 991         req->r_parent = dir;
 992         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 993         req->r_args.mkdir.mode = cpu_to_le32(mode);
 994         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 995         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 996         if (as_ctx.pagelist) {
 997                 req->r_pagelist = as_ctx.pagelist;
 998                 as_ctx.pagelist = NULL;
 999         }
1000         err = ceph_mdsc_do_request(mdsc, dir, req);
1001         if (!err &&
1002             !req->r_reply_info.head->is_target &&
1003             !req->r_reply_info.head->is_dentry)
1004                 err = ceph_handle_notrace_create(dir, dentry);
1005         ceph_mdsc_put_request(req);
1006 out:
1007         if (!err)
1008                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1009         else
1010                 d_drop(dentry);
1011         ceph_release_acl_sec_ctx(&as_ctx);
1012         return err;
1013 }
1014
1015 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1016                      struct dentry *dentry)
1017 {
1018         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1019         struct ceph_mds_request *req;
1020         int err;
1021
1022         if (ceph_snap(dir) != CEPH_NOSNAP)
1023                 return -EROFS;
1024
1025         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1026              old_dentry, dentry);
1027         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1028         if (IS_ERR(req)) {
1029                 d_drop(dentry);
1030                 return PTR_ERR(req);
1031         }
1032         req->r_dentry = dget(dentry);
1033         req->r_num_caps = 2;
1034         req->r_old_dentry = dget(old_dentry);
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         /* release LINK_SHARED on source inode (mds will lock it) */
1040         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1041         err = ceph_mdsc_do_request(mdsc, dir, req);
1042         if (err) {
1043                 d_drop(dentry);
1044         } else if (!req->r_reply_info.head->is_dentry) {
1045                 ihold(d_inode(old_dentry));
1046                 d_instantiate(dentry, d_inode(old_dentry));
1047         }
1048         ceph_mdsc_put_request(req);
1049         return err;
1050 }
1051
1052 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1053                                  struct ceph_mds_request *req)
1054 {
1055         int result = req->r_err ? req->r_err :
1056                         le32_to_cpu(req->r_reply_info.head->result);
1057
1058         if (result == -EJUKEBOX)
1059                 goto out;
1060
1061         /* If op failed, mark everyone involved for errors */
1062         if (result) {
1063                 int pathlen = 0;
1064                 u64 base = 0;
1065                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1066                                                   &base, 0);
1067
1068                 /* mark error on parent + clear complete */
1069                 mapping_set_error(req->r_parent->i_mapping, result);
1070                 ceph_dir_clear_complete(req->r_parent);
1071
1072                 /* drop the dentry -- we don't know its status */
1073                 if (!d_unhashed(req->r_dentry))
1074                         d_drop(req->r_dentry);
1075
1076                 /* mark inode itself for an error (since metadata is bogus) */
1077                 mapping_set_error(req->r_old_inode->i_mapping, result);
1078
1079                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1080                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1081                 ceph_mdsc_free_path(path, pathlen);
1082         }
1083 out:
1084         iput(req->r_old_inode);
1085         ceph_mdsc_release_dir_caps(req);
1086 }
1087
1088 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1089 {
1090         struct ceph_inode_info *ci = ceph_inode(dir);
1091         struct ceph_dentry_info *di;
1092         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1093
1094         spin_lock(&ci->i_ceph_lock);
1095         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1096                 ceph_take_cap_refs(ci, want, false);
1097                 got = want;
1098         }
1099         spin_unlock(&ci->i_ceph_lock);
1100
1101         /* If we didn't get anything, return 0 */
1102         if (!got)
1103                 return 0;
1104
1105         spin_lock(&dentry->d_lock);
1106         di = ceph_dentry(dentry);
1107         /*
1108          * - We are holding Fx, which implies Fs caps.
1109          * - Only support async unlink for primary linkage
1110          */
1111         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1112             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1113                 want = 0;
1114         spin_unlock(&dentry->d_lock);
1115
1116         /* Do we still want what we've got? */
1117         if (want == got)
1118                 return got;
1119
1120         ceph_put_cap_refs(ci, got);
1121         return 0;
1122 }
1123
1124 /*
1125  * rmdir and unlink are differ only by the metadata op code
1126  */
1127 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1128 {
1129         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1130         struct ceph_mds_client *mdsc = fsc->mdsc;
1131         struct inode *inode = d_inode(dentry);
1132         struct ceph_mds_request *req;
1133         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1134         int err = -EROFS;
1135         int op;
1136
1137         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1138                 /* rmdir .snap/foo is RMSNAP */
1139                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1140                 op = CEPH_MDS_OP_RMSNAP;
1141         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1142                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1143                      dir, dentry, inode);
1144                 op = d_is_dir(dentry) ?
1145                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1146         } else
1147                 goto out;
1148 retry:
1149         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1150         if (IS_ERR(req)) {
1151                 err = PTR_ERR(req);
1152                 goto out;
1153         }
1154         req->r_dentry = dget(dentry);
1155         req->r_num_caps = 2;
1156         req->r_parent = dir;
1157         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1158         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1159         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1160
1161         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1162             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1163                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1164                      dentry->d_name.len, dentry->d_name.name,
1165                      ceph_cap_string(req->r_dir_caps));
1166                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1167                 req->r_callback = ceph_async_unlink_cb;
1168                 req->r_old_inode = d_inode(dentry);
1169                 ihold(req->r_old_inode);
1170                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1171                 if (!err) {
1172                         /*
1173                          * We have enough caps, so we assume that the unlink
1174                          * will succeed. Fix up the target inode and dcache.
1175                          */
1176                         drop_nlink(inode);
1177                         d_delete(dentry);
1178                 } else if (err == -EJUKEBOX) {
1179                         try_async = false;
1180                         ceph_mdsc_put_request(req);
1181                         goto retry;
1182                 }
1183         } else {
1184                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1185                 err = ceph_mdsc_do_request(mdsc, dir, req);
1186                 if (!err && !req->r_reply_info.head->is_dentry)
1187                         d_delete(dentry);
1188         }
1189
1190         ceph_mdsc_put_request(req);
1191 out:
1192         return err;
1193 }
1194
1195 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1196                        struct inode *new_dir, struct dentry *new_dentry,
1197                        unsigned int flags)
1198 {
1199         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1200         struct ceph_mds_request *req;
1201         int op = CEPH_MDS_OP_RENAME;
1202         int err;
1203
1204         if (flags)
1205                 return -EINVAL;
1206
1207         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1208                 return -EXDEV;
1209         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1210                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1211                         op = CEPH_MDS_OP_RENAMESNAP;
1212                 else
1213                         return -EROFS;
1214         } else if (old_dir != new_dir) {
1215                 err = ceph_quota_check_rename(mdsc, d_inode(old_dentry),
1216                                               new_dir);
1217                 if (err)
1218                         return err;
1219         }
1220
1221         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1222              old_dir, old_dentry, new_dir, new_dentry);
1223         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1224         if (IS_ERR(req))
1225                 return PTR_ERR(req);
1226         ihold(old_dir);
1227         req->r_dentry = dget(new_dentry);
1228         req->r_num_caps = 2;
1229         req->r_old_dentry = dget(old_dentry);
1230         req->r_old_dentry_dir = old_dir;
1231         req->r_parent = new_dir;
1232         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1233         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1234         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1235         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1236         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1237         /* release LINK_RDCACHE on source inode (mds will lock it) */
1238         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1239         if (d_really_is_positive(new_dentry)) {
1240                 req->r_inode_drop =
1241                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1242         }
1243         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1244         if (!err && !req->r_reply_info.head->is_dentry) {
1245                 /*
1246                  * Normally d_move() is done by fill_trace (called by
1247                  * do_request, above).  If there is no trace, we need
1248                  * to do it here.
1249                  */
1250                 d_move(old_dentry, new_dentry);
1251         }
1252         ceph_mdsc_put_request(req);
1253         return err;
1254 }
1255
1256 /*
1257  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1258  * Leases at front of the list will expire first. (Assume all leases have
1259  * similar duration)
1260  *
1261  * Called under dentry->d_lock.
1262  */
1263 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1264 {
1265         struct dentry *dn = di->dentry;
1266         struct ceph_mds_client *mdsc;
1267
1268         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1269
1270         di->flags |= CEPH_DENTRY_LEASE_LIST;
1271         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1272                 di->flags |= CEPH_DENTRY_REFERENCED;
1273                 return;
1274         }
1275
1276         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1277         spin_lock(&mdsc->dentry_list_lock);
1278         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1279         spin_unlock(&mdsc->dentry_list_lock);
1280 }
1281
1282 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1283                                      struct ceph_dentry_info *di)
1284 {
1285         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1286         di->lease_gen = 0;
1287         di->time = jiffies;
1288         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1289 }
1290
1291 /*
1292  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1293  * list if it's not in the list, otherwise set 'referenced' flag.
1294  *
1295  * Called under dentry->d_lock.
1296  */
1297 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1298 {
1299         struct dentry *dn = di->dentry;
1300         struct ceph_mds_client *mdsc;
1301
1302         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1303              di, dn, dn, di->offset);
1304
1305         if (!list_empty(&di->lease_list)) {
1306                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1307                         /* don't remove dentry from dentry lease list
1308                          * if its lease is valid */
1309                         if (__dentry_lease_is_valid(di))
1310                                 return;
1311                 } else {
1312                         di->flags |= CEPH_DENTRY_REFERENCED;
1313                         return;
1314                 }
1315         }
1316
1317         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1318                 di->flags |= CEPH_DENTRY_REFERENCED;
1319                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1320                 return;
1321         }
1322
1323         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1324         spin_lock(&mdsc->dentry_list_lock);
1325         __dentry_dir_lease_touch(mdsc, di),
1326         spin_unlock(&mdsc->dentry_list_lock);
1327 }
1328
1329 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1330 {
1331         struct ceph_mds_client *mdsc;
1332         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1333                 return;
1334         if (list_empty(&di->lease_list))
1335                 return;
1336
1337         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1338         spin_lock(&mdsc->dentry_list_lock);
1339         list_del_init(&di->lease_list);
1340         spin_unlock(&mdsc->dentry_list_lock);
1341 }
1342
1343 enum {
1344         KEEP    = 0,
1345         DELETE  = 1,
1346         TOUCH   = 2,
1347         STOP    = 4,
1348 };
1349
1350 struct ceph_lease_walk_control {
1351         bool dir_lease;
1352         bool expire_dir_lease;
1353         unsigned long nr_to_scan;
1354         unsigned long dir_lease_ttl;
1355 };
1356
1357 static unsigned long
1358 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1359                      struct ceph_lease_walk_control *lwc,
1360                      int (*check)(struct dentry*, void*))
1361 {
1362         struct ceph_dentry_info *di, *tmp;
1363         struct dentry *dentry, *last = NULL;
1364         struct list_head* list;
1365         LIST_HEAD(dispose);
1366         unsigned long freed = 0;
1367         int ret = 0;
1368
1369         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1370         spin_lock(&mdsc->dentry_list_lock);
1371         list_for_each_entry_safe(di, tmp, list, lease_list) {
1372                 if (!lwc->nr_to_scan)
1373                         break;
1374                 --lwc->nr_to_scan;
1375
1376                 dentry = di->dentry;
1377                 if (last == dentry)
1378                         break;
1379
1380                 if (!spin_trylock(&dentry->d_lock))
1381                         continue;
1382
1383                 if (__lockref_is_dead(&dentry->d_lockref)) {
1384                         list_del_init(&di->lease_list);
1385                         goto next;
1386                 }
1387
1388                 ret = check(dentry, lwc);
1389                 if (ret & TOUCH) {
1390                         /* move it into tail of dir lease list */
1391                         __dentry_dir_lease_touch(mdsc, di);
1392                         if (!last)
1393                                 last = dentry;
1394                 }
1395                 if (ret & DELETE) {
1396                         /* stale lease */
1397                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1398                         if (dentry->d_lockref.count > 0) {
1399                                 /* update_dentry_lease() will re-add
1400                                  * it to lease list, or
1401                                  * ceph_d_delete() will return 1 when
1402                                  * last reference is dropped */
1403                                 list_del_init(&di->lease_list);
1404                         } else {
1405                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1406                                 list_move_tail(&di->lease_list, &dispose);
1407                                 dget_dlock(dentry);
1408                         }
1409                 }
1410 next:
1411                 spin_unlock(&dentry->d_lock);
1412                 if (ret & STOP)
1413                         break;
1414         }
1415         spin_unlock(&mdsc->dentry_list_lock);
1416
1417         while (!list_empty(&dispose)) {
1418                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1419                                       lease_list);
1420                 dentry = di->dentry;
1421                 spin_lock(&dentry->d_lock);
1422
1423                 list_del_init(&di->lease_list);
1424                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1425                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1426                         spin_lock(&mdsc->dentry_list_lock);
1427                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1428                                 list_add_tail(&di->lease_list,
1429                                               &mdsc->dentry_leases);
1430                         } else {
1431                                 __dentry_dir_lease_touch(mdsc, di);
1432                         }
1433                         spin_unlock(&mdsc->dentry_list_lock);
1434                 } else {
1435                         freed++;
1436                 }
1437
1438                 spin_unlock(&dentry->d_lock);
1439                 /* ceph_d_delete() does the trick */
1440                 dput(dentry);
1441         }
1442         return freed;
1443 }
1444
1445 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1446 {
1447         struct ceph_dentry_info *di = ceph_dentry(dentry);
1448         int ret;
1449
1450         if (__dentry_lease_is_valid(di))
1451                 return STOP;
1452         ret = __dir_lease_try_check(dentry);
1453         if (ret == -EBUSY)
1454                 return KEEP;
1455         if (ret > 0)
1456                 return TOUCH;
1457         return DELETE;
1458 }
1459
1460 static int __dir_lease_check(struct dentry *dentry, void *arg)
1461 {
1462         struct ceph_lease_walk_control *lwc = arg;
1463         struct ceph_dentry_info *di = ceph_dentry(dentry);
1464
1465         int ret = __dir_lease_try_check(dentry);
1466         if (ret == -EBUSY)
1467                 return KEEP;
1468         if (ret > 0) {
1469                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1470                         return STOP;
1471                 /* Move dentry to tail of dir lease list if we don't want
1472                  * to delete it. So dentries in the list are checked in a
1473                  * round robin manner */
1474                 if (!lwc->expire_dir_lease)
1475                         return TOUCH;
1476                 if (dentry->d_lockref.count > 0 ||
1477                     (di->flags & CEPH_DENTRY_REFERENCED))
1478                         return TOUCH;
1479                 /* invalidate dir lease */
1480                 di->lease_shared_gen = 0;
1481         }
1482         return DELETE;
1483 }
1484
1485 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1486 {
1487         struct ceph_lease_walk_control lwc;
1488         unsigned long count;
1489         unsigned long freed;
1490
1491         spin_lock(&mdsc->caps_list_lock);
1492         if (mdsc->caps_use_max > 0 &&
1493             mdsc->caps_use_count > mdsc->caps_use_max)
1494                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1495         else
1496                 count = 0;
1497         spin_unlock(&mdsc->caps_list_lock);
1498
1499         lwc.dir_lease = false;
1500         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1501         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1502         if (!lwc.nr_to_scan) /* more invalid leases */
1503                 return -EAGAIN;
1504
1505         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1506                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1507
1508         lwc.dir_lease = true;
1509         lwc.expire_dir_lease = freed < count;
1510         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1511         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1512         if (!lwc.nr_to_scan) /* more to check */
1513                 return -EAGAIN;
1514
1515         return freed > 0 ? 1 : 0;
1516 }
1517
1518 /*
1519  * Ensure a dentry lease will no longer revalidate.
1520  */
1521 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1522 {
1523         struct ceph_dentry_info *di = ceph_dentry(dentry);
1524         spin_lock(&dentry->d_lock);
1525         di->time = jiffies;
1526         di->lease_shared_gen = 0;
1527         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1528         __dentry_lease_unlist(di);
1529         spin_unlock(&dentry->d_lock);
1530 }
1531
1532 /*
1533  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1534  * renew if the least is more than half up.
1535  */
1536 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1537 {
1538         struct ceph_mds_session *session;
1539
1540         if (!di->lease_gen)
1541                 return false;
1542
1543         session = di->lease_session;
1544         if (session) {
1545                 u32 gen;
1546                 unsigned long ttl;
1547
1548                 spin_lock(&session->s_gen_ttl_lock);
1549                 gen = session->s_cap_gen;
1550                 ttl = session->s_cap_ttl;
1551                 spin_unlock(&session->s_gen_ttl_lock);
1552
1553                 if (di->lease_gen == gen &&
1554                     time_before(jiffies, ttl) &&
1555                     time_before(jiffies, di->time))
1556                         return true;
1557         }
1558         di->lease_gen = 0;
1559         return false;
1560 }
1561
1562 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1563 {
1564         struct ceph_dentry_info *di;
1565         struct ceph_mds_session *session = NULL;
1566         u32 seq = 0;
1567         int valid = 0;
1568
1569         spin_lock(&dentry->d_lock);
1570         di = ceph_dentry(dentry);
1571         if (di && __dentry_lease_is_valid(di)) {
1572                 valid = 1;
1573
1574                 if (di->lease_renew_after &&
1575                     time_after(jiffies, di->lease_renew_after)) {
1576                         /*
1577                          * We should renew. If we're in RCU walk mode
1578                          * though, we can't do that so just return
1579                          * -ECHILD.
1580                          */
1581                         if (flags & LOOKUP_RCU) {
1582                                 valid = -ECHILD;
1583                         } else {
1584                                 session = ceph_get_mds_session(di->lease_session);
1585                                 seq = di->lease_seq;
1586                                 di->lease_renew_after = 0;
1587                                 di->lease_renew_from = jiffies;
1588                         }
1589                 }
1590         }
1591         spin_unlock(&dentry->d_lock);
1592
1593         if (session) {
1594                 ceph_mdsc_lease_send_msg(session, dentry,
1595                                          CEPH_MDS_LEASE_RENEW, seq);
1596                 ceph_put_mds_session(session);
1597         }
1598         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1599         return valid;
1600 }
1601
1602 /*
1603  * Called under dentry->d_lock.
1604  */
1605 static int __dir_lease_try_check(const struct dentry *dentry)
1606 {
1607         struct ceph_dentry_info *di = ceph_dentry(dentry);
1608         struct inode *dir;
1609         struct ceph_inode_info *ci;
1610         int valid = 0;
1611
1612         if (!di->lease_shared_gen)
1613                 return 0;
1614         if (IS_ROOT(dentry))
1615                 return 0;
1616
1617         dir = d_inode(dentry->d_parent);
1618         ci = ceph_inode(dir);
1619
1620         if (spin_trylock(&ci->i_ceph_lock)) {
1621                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1622                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1623                         valid = 1;
1624                 spin_unlock(&ci->i_ceph_lock);
1625         } else {
1626                 valid = -EBUSY;
1627         }
1628
1629         if (!valid)
1630                 di->lease_shared_gen = 0;
1631         return valid;
1632 }
1633
1634 /*
1635  * Check if directory-wide content lease/cap is valid.
1636  */
1637 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1638                               struct ceph_mds_client *mdsc)
1639 {
1640         struct ceph_inode_info *ci = ceph_inode(dir);
1641         int valid;
1642         int shared_gen;
1643
1644         spin_lock(&ci->i_ceph_lock);
1645         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1646         if (valid) {
1647                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1648                 shared_gen = atomic_read(&ci->i_shared_gen);
1649         }
1650         spin_unlock(&ci->i_ceph_lock);
1651         if (valid) {
1652                 struct ceph_dentry_info *di;
1653                 spin_lock(&dentry->d_lock);
1654                 di = ceph_dentry(dentry);
1655                 if (dir == d_inode(dentry->d_parent) &&
1656                     di && di->lease_shared_gen == shared_gen)
1657                         __ceph_dentry_dir_lease_touch(di);
1658                 else
1659                         valid = 0;
1660                 spin_unlock(&dentry->d_lock);
1661         }
1662         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1663              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1664         return valid;
1665 }
1666
1667 /*
1668  * Check if cached dentry can be trusted.
1669  */
1670 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1671 {
1672         int valid = 0;
1673         struct dentry *parent;
1674         struct inode *dir, *inode;
1675         struct ceph_mds_client *mdsc;
1676
1677         if (flags & LOOKUP_RCU) {
1678                 parent = READ_ONCE(dentry->d_parent);
1679                 dir = d_inode_rcu(parent);
1680                 if (!dir)
1681                         return -ECHILD;
1682                 inode = d_inode_rcu(dentry);
1683         } else {
1684                 parent = dget_parent(dentry);
1685                 dir = d_inode(parent);
1686                 inode = d_inode(dentry);
1687         }
1688
1689         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1690              dentry, inode, ceph_dentry(dentry)->offset);
1691
1692         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1693
1694         /* always trust cached snapped dentries, snapdir dentry */
1695         if (ceph_snap(dir) != CEPH_NOSNAP) {
1696                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1697                      dentry, inode);
1698                 valid = 1;
1699         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1700                 valid = 1;
1701         } else {
1702                 valid = dentry_lease_is_valid(dentry, flags);
1703                 if (valid == -ECHILD)
1704                         return valid;
1705                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1706                         if (inode)
1707                                 valid = ceph_is_any_caps(inode);
1708                         else
1709                                 valid = 1;
1710                 }
1711         }
1712
1713         if (!valid) {
1714                 struct ceph_mds_request *req;
1715                 int op, err;
1716                 u32 mask;
1717
1718                 if (flags & LOOKUP_RCU)
1719                         return -ECHILD;
1720
1721                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1722
1723                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1724                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1725                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1726                 if (!IS_ERR(req)) {
1727                         req->r_dentry = dget(dentry);
1728                         req->r_num_caps = 2;
1729                         req->r_parent = dir;
1730
1731                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1732                         if (ceph_security_xattr_wanted(dir))
1733                                 mask |= CEPH_CAP_XATTR_SHARED;
1734                         req->r_args.getattr.mask = cpu_to_le32(mask);
1735
1736                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1737                         switch (err) {
1738                         case 0:
1739                                 if (d_really_is_positive(dentry) &&
1740                                     d_inode(dentry) == req->r_target_inode)
1741                                         valid = 1;
1742                                 break;
1743                         case -ENOENT:
1744                                 if (d_really_is_negative(dentry))
1745                                         valid = 1;
1746                                 fallthrough;
1747                         default:
1748                                 break;
1749                         }
1750                         ceph_mdsc_put_request(req);
1751                         dout("d_revalidate %p lookup result=%d\n",
1752                              dentry, err);
1753                 }
1754         } else {
1755                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1756         }
1757
1758         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1759         if (!valid)
1760                 ceph_dir_clear_complete(dir);
1761
1762         if (!(flags & LOOKUP_RCU))
1763                 dput(parent);
1764         return valid;
1765 }
1766
1767 /*
1768  * Delete unused dentry that doesn't have valid lease
1769  *
1770  * Called under dentry->d_lock.
1771  */
1772 static int ceph_d_delete(const struct dentry *dentry)
1773 {
1774         struct ceph_dentry_info *di;
1775
1776         /* won't release caps */
1777         if (d_really_is_negative(dentry))
1778                 return 0;
1779         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1780                 return 0;
1781         /* vaild lease? */
1782         di = ceph_dentry(dentry);
1783         if (di) {
1784                 if (__dentry_lease_is_valid(di))
1785                         return 0;
1786                 if (__dir_lease_try_check(dentry))
1787                         return 0;
1788         }
1789         return 1;
1790 }
1791
1792 /*
1793  * Release our ceph_dentry_info.
1794  */
1795 static void ceph_d_release(struct dentry *dentry)
1796 {
1797         struct ceph_dentry_info *di = ceph_dentry(dentry);
1798         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1799
1800         dout("d_release %p\n", dentry);
1801
1802         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1803
1804         spin_lock(&dentry->d_lock);
1805         __dentry_lease_unlist(di);
1806         dentry->d_fsdata = NULL;
1807         spin_unlock(&dentry->d_lock);
1808
1809         ceph_put_mds_session(di->lease_session);
1810         kmem_cache_free(ceph_dentry_cachep, di);
1811 }
1812
1813 /*
1814  * When the VFS prunes a dentry from the cache, we need to clear the
1815  * complete flag on the parent directory.
1816  *
1817  * Called under dentry->d_lock.
1818  */
1819 static void ceph_d_prune(struct dentry *dentry)
1820 {
1821         struct ceph_inode_info *dir_ci;
1822         struct ceph_dentry_info *di;
1823
1824         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1825
1826         /* do we have a valid parent? */
1827         if (IS_ROOT(dentry))
1828                 return;
1829
1830         /* we hold d_lock, so d_parent is stable */
1831         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1832         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1833                 return;
1834
1835         /* who calls d_delete() should also disable dcache readdir */
1836         if (d_really_is_negative(dentry))
1837                 return;
1838
1839         /* d_fsdata does not get cleared until d_release */
1840         if (!d_unhashed(dentry)) {
1841                 __ceph_dir_clear_complete(dir_ci);
1842                 return;
1843         }
1844
1845         /* Disable dcache readdir just in case that someone called d_drop()
1846          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1847          * properly (dcache readdir is still enabled) */
1848         di = ceph_dentry(dentry);
1849         if (di->offset > 0 &&
1850             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1851                 __ceph_dir_clear_ordered(dir_ci);
1852 }
1853
1854 /*
1855  * read() on a dir.  This weird interface hack only works if mounted
1856  * with '-o dirstat'.
1857  */
1858 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1859                              loff_t *ppos)
1860 {
1861         struct ceph_dir_file_info *dfi = file->private_data;
1862         struct inode *inode = file_inode(file);
1863         struct ceph_inode_info *ci = ceph_inode(inode);
1864         int left;
1865         const int bufsize = 1024;
1866
1867         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1868                 return -EISDIR;
1869
1870         if (!dfi->dir_info) {
1871                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1872                 if (!dfi->dir_info)
1873                         return -ENOMEM;
1874                 dfi->dir_info_len =
1875                         snprintf(dfi->dir_info, bufsize,
1876                                 "entries:   %20lld\n"
1877                                 " files:    %20lld\n"
1878                                 " subdirs:  %20lld\n"
1879                                 "rentries:  %20lld\n"
1880                                 " rfiles:   %20lld\n"
1881                                 " rsubdirs: %20lld\n"
1882                                 "rbytes:    %20lld\n"
1883                                 "rctime:    %10lld.%09ld\n",
1884                                 ci->i_files + ci->i_subdirs,
1885                                 ci->i_files,
1886                                 ci->i_subdirs,
1887                                 ci->i_rfiles + ci->i_rsubdirs,
1888                                 ci->i_rfiles,
1889                                 ci->i_rsubdirs,
1890                                 ci->i_rbytes,
1891                                 ci->i_rctime.tv_sec,
1892                                 ci->i_rctime.tv_nsec);
1893         }
1894
1895         if (*ppos >= dfi->dir_info_len)
1896                 return 0;
1897         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1898         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1899         if (left == size)
1900                 return -EFAULT;
1901         *ppos += (size - left);
1902         return size - left;
1903 }
1904
1905
1906
1907 /*
1908  * Return name hash for a given dentry.  This is dependent on
1909  * the parent directory's hash function.
1910  */
1911 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1912 {
1913         struct ceph_inode_info *dci = ceph_inode(dir);
1914         unsigned hash;
1915
1916         switch (dci->i_dir_layout.dl_dir_hash) {
1917         case 0: /* for backward compat */
1918         case CEPH_STR_HASH_LINUX:
1919                 return dn->d_name.hash;
1920
1921         default:
1922                 spin_lock(&dn->d_lock);
1923                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1924                                      dn->d_name.name, dn->d_name.len);
1925                 spin_unlock(&dn->d_lock);
1926                 return hash;
1927         }
1928 }
1929
1930 const struct file_operations ceph_dir_fops = {
1931         .read = ceph_read_dir,
1932         .iterate = ceph_readdir,
1933         .llseek = ceph_dir_llseek,
1934         .open = ceph_open,
1935         .release = ceph_release,
1936         .unlocked_ioctl = ceph_ioctl,
1937         .compat_ioctl = compat_ptr_ioctl,
1938         .fsync = ceph_fsync,
1939         .lock = ceph_lock,
1940         .flock = ceph_flock,
1941 };
1942
1943 const struct file_operations ceph_snapdir_fops = {
1944         .iterate = ceph_readdir,
1945         .llseek = ceph_dir_llseek,
1946         .open = ceph_open,
1947         .release = ceph_release,
1948 };
1949
1950 const struct inode_operations ceph_dir_iops = {
1951         .lookup = ceph_lookup,
1952         .permission = ceph_permission,
1953         .getattr = ceph_getattr,
1954         .setattr = ceph_setattr,
1955         .listxattr = ceph_listxattr,
1956         .get_acl = ceph_get_acl,
1957         .set_acl = ceph_set_acl,
1958         .mknod = ceph_mknod,
1959         .symlink = ceph_symlink,
1960         .mkdir = ceph_mkdir,
1961         .link = ceph_link,
1962         .unlink = ceph_unlink,
1963         .rmdir = ceph_unlink,
1964         .rename = ceph_rename,
1965         .create = ceph_create,
1966         .atomic_open = ceph_atomic_open,
1967 };
1968
1969 const struct inode_operations ceph_snapdir_iops = {
1970         .lookup = ceph_lookup,
1971         .permission = ceph_permission,
1972         .getattr = ceph_getattr,
1973         .mkdir = ceph_mkdir,
1974         .rmdir = ceph_unlink,
1975         .rename = ceph_rename,
1976 };
1977
1978 const struct dentry_operations ceph_dentry_ops = {
1979         .d_revalidate = ceph_d_revalidate,
1980         .d_delete = ceph_d_delete,
1981         .d_release = ceph_d_release,
1982         .d_prune = ceph_d_prune,
1983         .d_init = ceph_d_init,
1984 };