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_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
  42         struct ceph_mds_client *mdsc = fsc->mdsc;
  43
  44         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  45         if (!di)
  46                 return -ENOMEM;          /* oh well */
  47
  48         di->dentry = dentry;
  49         di->lease_session = NULL;
  50         di->time = jiffies;
  51         dentry->d_fsdata = di;
  52         INIT_LIST_HEAD(&di->lease_list);
  53
  54         atomic64_inc(&mdsc->metric.total_dentries);
  55
  56         return 0;
  57 }
  58
  59 /*
  60  * for f_pos for readdir:
  61  * - hash order:
  62  *      (0xff << 52) | ((24 bits hash) << 28) |
  63  *      (the nth entry has hash collision);
  64  * - frag+name order;
  65  *      ((frag value) << 28) | (the nth entry in frag);
  66  */
  67 #define OFFSET_BITS     28
  68 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  69 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  70 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  71 {
  72         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  73         if (hash_order)
  74                 fpos |= HASH_ORDER;
  75         return fpos;
  76 }
  77
  78 static bool is_hash_order(loff_t p)
  79 {
  80         return (p & HASH_ORDER) == HASH_ORDER;
  81 }
  82
  83 static unsigned fpos_frag(loff_t p)
  84 {
  85         return p >> OFFSET_BITS;
  86 }
  87
  88 static unsigned fpos_hash(loff_t p)
  89 {
  90         return ceph_frag_value(fpos_frag(p));
  91 }
  92
  93 static unsigned fpos_off(loff_t p)
  94 {
  95         return p & OFFSET_MASK;
  96 }
  97
  98 static int fpos_cmp(loff_t l, loff_t r)
  99 {
 100         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
 101         if (v)
 102                 return v;
 103         return (int)(fpos_off(l) - fpos_off(r));
 104 }
 105
 106 /*
 107  * make note of the last dentry we read, so we can
 108  * continue at the same lexicographical point,
 109  * regardless of what dir changes take place on the
 110  * server.
 111  */
 112 static int note_last_dentry(struct ceph_dir_file_info *dfi, const char *name,
 113                             int len, unsigned next_offset)
 114 {
 115         char *buf = kmalloc(len+1, GFP_KERNEL);
 116         if (!buf)
 117                 return -ENOMEM;
 118         kfree(dfi->last_name);
 119         dfi->last_name = buf;
 120         memcpy(dfi->last_name, name, len);
 121         dfi->last_name[len] = 0;
 122         dfi->next_offset = next_offset;
 123         dout("note_last_dentry '%s'\n", dfi->last_name);
 124         return 0;
 125 }
 126
 127
 128 static struct dentry *
 129 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 130                         struct ceph_readdir_cache_control *cache_ctl)
 131 {
 132         struct inode *dir = d_inode(parent);
 133         struct dentry *dentry;
 134         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 135         loff_t ptr_pos = idx * sizeof(struct dentry *);
 136         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 137
 138         if (ptr_pos >= i_size_read(dir))
 139                 return NULL;
 140
 141         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 142                 ceph_readdir_cache_release(cache_ctl);
 143                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 144                 if (!cache_ctl->page) {
 145                         dout(" page %lu not found\n", ptr_pgoff);
 146                         return ERR_PTR(-EAGAIN);
 147                 }
 148                 /* reading/filling the cache are serialized by
 149                    i_mutex, no need to use page lock */
 150                 unlock_page(cache_ctl->page);
 151                 cache_ctl->dentries = kmap(cache_ctl->page);
 152         }
 153
 154         cache_ctl->index = idx & idx_mask;
 155
 156         rcu_read_lock();
 157         spin_lock(&parent->d_lock);
 158         /* check i_size again here, because empty directory can be
 159          * marked as complete while not holding the i_mutex. */
 160         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 161                 dentry = cache_ctl->dentries[cache_ctl->index];
 162         else
 163                 dentry = NULL;
 164         spin_unlock(&parent->d_lock);
 165         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 166                 dentry = NULL;
 167         rcu_read_unlock();
 168         return dentry ? : ERR_PTR(-EAGAIN);
 169 }
 170
 171 /*
 172  * When possible, we try to satisfy a readdir by peeking at the
 173  * dcache.  We make this work by carefully ordering dentries on
 174  * d_child when we initially get results back from the MDS, and
 175  * falling back to a "normal" sync readdir if any dentries in the dir
 176  * are dropped.
 177  *
 178  * Complete dir indicates that we have all dentries in the dir.  It is
 179  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 180  * the MDS if/when the directory is modified).
 181  */
 182 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 183                             int shared_gen)
 184 {
 185         struct ceph_dir_file_info *dfi = file->private_data;
 186         struct dentry *parent = file->f_path.dentry;
 187         struct inode *dir = d_inode(parent);
 188         struct dentry *dentry, *last = NULL;
 189         struct ceph_dentry_info *di;
 190         struct ceph_readdir_cache_control cache_ctl = {};
 191         u64 idx = 0;
 192         int err = 0;
 193
 194         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 195
 196         /* search start position */
 197         if (ctx->pos > 2) {
 198                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 199                 while (count > 0) {
 200                         u64 step = count >> 1;
 201                         dentry = __dcache_find_get_entry(parent, idx + step,
 202                                                          &cache_ctl);
 203                         if (!dentry) {
 204                                 /* use linar search */
 205                                 idx = 0;
 206                                 break;
 207                         }
 208                         if (IS_ERR(dentry)) {
 209                                 err = PTR_ERR(dentry);
 210                                 goto out;
 211                         }
 212                         di = ceph_dentry(dentry);
 213                         spin_lock(&dentry->d_lock);
 214                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 215                                 idx += step + 1;
 216                                 count -= step + 1;
 217                         } else {
 218                                 count = step;
 219                         }
 220                         spin_unlock(&dentry->d_lock);
 221                         dput(dentry);
 222                 }
 223
 224                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 225         }
 226
 227
 228         for (;;) {
 229                 bool emit_dentry = false;
 230                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 231                 if (!dentry) {
 232                         dfi->file_info.flags |= CEPH_F_ATEND;
 233                         err = 0;
 234                         break;
 235                 }
 236                 if (IS_ERR(dentry)) {
 237                         err = PTR_ERR(dentry);
 238                         goto out;
 239                 }
 240
 241                 spin_lock(&dentry->d_lock);
 242                 di = ceph_dentry(dentry);
 243                 if (d_unhashed(dentry) ||
 244                     d_really_is_negative(dentry) ||
 245                     di->lease_shared_gen != shared_gen) {
 246                         spin_unlock(&dentry->d_lock);
 247                         dput(dentry);
 248                         err = -EAGAIN;
 249                         goto out;
 250                 }
 251                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 252                         __ceph_dentry_dir_lease_touch(di);
 253                         emit_dentry = true;
 254                 }
 255                 spin_unlock(&dentry->d_lock);
 256
 257                 if (emit_dentry) {
 258                         dout(" %llx dentry %p %pd %p\n", di->offset,
 259                              dentry, dentry, d_inode(dentry));
 260                         ctx->pos = di->offset;
 261                         if (!dir_emit(ctx, dentry->d_name.name,
 262                                       dentry->d_name.len,
 263                                       ceph_translate_ino(dentry->d_sb,
 264                                                          d_inode(dentry)->i_ino),
 265                                       d_inode(dentry)->i_mode >> 12)) {
 266                                 dput(dentry);
 267                                 err = 0;
 268                                 break;
 269                         }
 270                         ctx->pos++;
 271
 272                         if (last)
 273                                 dput(last);
 274                         last = dentry;
 275                 } else {
 276                         dput(dentry);
 277                 }
 278         }
 279 out:
 280         ceph_readdir_cache_release(&cache_ctl);
 281         if (last) {
 282                 int ret;
 283                 di = ceph_dentry(last);
 284                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 285                                        fpos_off(di->offset) + 1);
 286                 if (ret < 0)
 287                         err = ret;
 288                 dput(last);
 289                 /* last_name no longer match cache index */
 290                 if (dfi->readdir_cache_idx >= 0) {
 291                         dfi->readdir_cache_idx = -1;
 292                         dfi->dir_release_count = 0;
 293                 }
 294         }
 295         return err;
 296 }
 297
 298 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 299 {
 300         if (!dfi->last_readdir)
 301                 return true;
 302         if (is_hash_order(pos))
 303                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 304         else
 305                 return dfi->frag != fpos_frag(pos);
 306 }
 307
 308 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 309 {
 310         struct ceph_dir_file_info *dfi = file->private_data;
 311         struct inode *inode = file_inode(file);
 312         struct ceph_inode_info *ci = ceph_inode(inode);
 313         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 314         struct ceph_mds_client *mdsc = fsc->mdsc;
 315         int i;
 316         int err;
 317         unsigned frag = -1;
 318         struct ceph_mds_reply_info_parsed *rinfo;
 319
 320         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 321         if (dfi->file_info.flags & CEPH_F_ATEND)
 322                 return 0;
 323
 324         /* always start with . and .. */
 325         if (ctx->pos == 0) {
 326                 dout("readdir off 0 -> '.'\n");
 327                 if (!dir_emit(ctx, ".", 1,
 328                             ceph_translate_ino(inode->i_sb, inode->i_ino),
 329                             inode->i_mode >> 12))
 330                         return 0;
 331                 ctx->pos = 1;
 332         }
 333         if (ctx->pos == 1) {
 334                 ino_t ino = parent_ino(file->f_path.dentry);
 335                 dout("readdir off 1 -> '..'\n");
 336                 if (!dir_emit(ctx, "..", 2,
 337                             ceph_translate_ino(inode->i_sb, ino),
 338                             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                                 return err;
 483                 } else if (req->r_reply_info.dir_end) {
 484                         dfi->next_offset = 2;
 485                         /* keep last name */
 486                 }
 487         }
 488
 489         rinfo = &dfi->last_readdir->r_reply_info;
 490         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 491              dfi->frag, rinfo->dir_nr, ctx->pos,
 492              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 493
 494         i = 0;
 495         /* search start position */
 496         if (rinfo->dir_nr > 0) {
 497                 int step, nr = rinfo->dir_nr;
 498                 while (nr > 0) {
 499                         step = nr >> 1;
 500                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 501                                 i +=  step + 1;
 502                                 nr -= step + 1;
 503                         } else {
 504                                 nr = step;
 505                         }
 506                 }
 507         }
 508         for (; i < rinfo->dir_nr; i++) {
 509                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 510                 struct ceph_vino vino;
 511                 ino_t ino;
 512                 u32 ftype;
 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                 ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
 523                 vino.ino = le64_to_cpu(rde->inode.in->ino);
 524                 vino.snap = le64_to_cpu(rde->inode.in->snapid);
 525                 ino = ceph_vino_to_ino(vino);
 526
 527                 if (!dir_emit(ctx, rde->name, rde->name_len,
 528                               ceph_translate_ino(inode->i_sb, ino), ftype)) {
 529                         dout("filldir stopping us...\n");
 530                         return 0;
 531                 }
 532                 ctx->pos++;
 533         }
 534
 535         ceph_mdsc_put_request(dfi->last_readdir);
 536         dfi->last_readdir = NULL;
 537
 538         if (dfi->next_offset > 2) {
 539                 frag = dfi->frag;
 540                 goto more;
 541         }
 542
 543         /* more frags? */
 544         if (!ceph_frag_is_rightmost(dfi->frag)) {
 545                 frag = ceph_frag_next(dfi->frag);
 546                 if (is_hash_order(ctx->pos)) {
 547                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 548                                                         dfi->next_offset, true);
 549                         if (new_pos > ctx->pos)
 550                                 ctx->pos = new_pos;
 551                         /* keep last_name */
 552                 } else {
 553                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 554                                                         false);
 555                         kfree(dfi->last_name);
 556                         dfi->last_name = NULL;
 557                 }
 558                 dout("readdir next frag is %x\n", frag);
 559                 goto more;
 560         }
 561         dfi->file_info.flags |= CEPH_F_ATEND;
 562
 563         /*
 564          * if dir_release_count still matches the dir, no dentries
 565          * were released during the whole readdir, and we should have
 566          * the complete dir contents in our cache.
 567          */
 568         if (atomic64_read(&ci->i_release_count) ==
 569                                         dfi->dir_release_count) {
 570                 spin_lock(&ci->i_ceph_lock);
 571                 if (dfi->dir_ordered_count ==
 572                                 atomic64_read(&ci->i_ordered_count)) {
 573                         dout(" marking %p complete and ordered\n", inode);
 574                         /* use i_size to track number of entries in
 575                          * readdir cache */
 576                         BUG_ON(dfi->readdir_cache_idx < 0);
 577                         i_size_write(inode, dfi->readdir_cache_idx *
 578                                      sizeof(struct dentry*));
 579                 } else {
 580                         dout(" marking %p complete\n", inode);
 581                 }
 582                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 583                                         dfi->dir_ordered_count);
 584                 spin_unlock(&ci->i_ceph_lock);
 585         }
 586
 587         dout("readdir %p file %p done.\n", inode, file);
 588         return 0;
 589 }
 590
 591 static void reset_readdir(struct ceph_dir_file_info *dfi)
 592 {
 593         if (dfi->last_readdir) {
 594                 ceph_mdsc_put_request(dfi->last_readdir);
 595                 dfi->last_readdir = NULL;
 596         }
 597         kfree(dfi->last_name);
 598         dfi->last_name = NULL;
 599         dfi->dir_release_count = 0;
 600         dfi->readdir_cache_idx = -1;
 601         dfi->next_offset = 2;  /* compensate for . and .. */
 602         dfi->file_info.flags &= ~CEPH_F_ATEND;
 603 }
 604
 605 /*
 606  * discard buffered readdir content on seekdir(0), or seek to new frag,
 607  * or seek prior to current chunk
 608  */
 609 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 610 {
 611         struct ceph_mds_reply_info_parsed *rinfo;
 612         loff_t chunk_offset;
 613         if (new_pos == 0)
 614                 return true;
 615         if (is_hash_order(new_pos)) {
 616                 /* no need to reset last_name for a forward seek when
 617                  * dentries are sotred in hash order */
 618         } else if (dfi->frag != fpos_frag(new_pos)) {
 619                 return true;
 620         }
 621         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 622         if (!rinfo || !rinfo->dir_nr)
 623                 return true;
 624         chunk_offset = rinfo->dir_entries[0].offset;
 625         return new_pos < chunk_offset ||
 626                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 627 }
 628
 629 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 630 {
 631         struct ceph_dir_file_info *dfi = file->private_data;
 632         struct inode *inode = file->f_mapping->host;
 633         loff_t retval;
 634
 635         inode_lock(inode);
 636         retval = -EINVAL;
 637         switch (whence) {
 638         case SEEK_CUR:
 639                 offset += file->f_pos;
 640         case SEEK_SET:
 641                 break;
 642         case SEEK_END:
 643                 retval = -EOPNOTSUPP;
 644         default:
 645                 goto out;
 646         }
 647
 648         if (offset >= 0) {
 649                 if (need_reset_readdir(dfi, offset)) {
 650                         dout("dir_llseek dropping %p content\n", file);
 651                         reset_readdir(dfi);
 652                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 653                         /* for hash offset, we don't know if a forward seek
 654                          * is within same frag */
 655                         dfi->dir_release_count = 0;
 656                         dfi->readdir_cache_idx = -1;
 657                 }
 658
 659                 if (offset != file->f_pos) {
 660                         file->f_pos = offset;
 661                         file->f_version = 0;
 662                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 663                 }
 664                 retval = offset;
 665         }
 666 out:
 667         inode_unlock(inode);
 668         return retval;
 669 }
 670
 671 /*
 672  * Handle lookups for the hidden .snap directory.
 673  */
 674 int ceph_handle_snapdir(struct ceph_mds_request *req,
 675                         struct dentry *dentry, int err)
 676 {
 677         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 678         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 679
 680         /* .snap dir? */
 681         if (err == -ENOENT &&
 682             ceph_snap(parent) == CEPH_NOSNAP &&
 683             strcmp(dentry->d_name.name,
 684                    fsc->mount_options->snapdir_name) == 0) {
 685                 struct inode *inode = ceph_get_snapdir(parent);
 686                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 687                      dentry, dentry, inode);
 688                 BUG_ON(!d_unhashed(dentry));
 689                 d_add(dentry, inode);
 690                 err = 0;
 691         }
 692         return err;
 693 }
 694
 695 /*
 696  * Figure out final result of a lookup/open request.
 697  *
 698  * Mainly, make sure we return the final req->r_dentry (if it already
 699  * existed) in place of the original VFS-provided dentry when they
 700  * differ.
 701  *
 702  * Gracefully handle the case where the MDS replies with -ENOENT and
 703  * no trace (which it may do, at its discretion, e.g., if it doesn't
 704  * care to issue a lease on the negative dentry).
 705  */
 706 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 707                                   struct dentry *dentry, int err)
 708 {
 709         if (err == -ENOENT) {
 710                 /* no trace? */
 711                 err = 0;
 712                 if (!req->r_reply_info.head->is_dentry) {
 713                         dout("ENOENT and no trace, dentry %p inode %p\n",
 714                              dentry, d_inode(dentry));
 715                         if (d_really_is_positive(dentry)) {
 716                                 d_drop(dentry);
 717                                 err = -ENOENT;
 718                         } else {
 719                                 d_add(dentry, NULL);
 720                         }
 721                 }
 722         }
 723         if (err)
 724                 dentry = ERR_PTR(err);
 725         else if (dentry != req->r_dentry)
 726                 dentry = dget(req->r_dentry);   /* we got spliced */
 727         else
 728                 dentry = NULL;
 729         return dentry;
 730 }
 731
 732 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 733 {
 734         return ceph_ino(inode) == CEPH_INO_ROOT &&
 735                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 736 }
 737
 738 /*
 739  * Look up a single dir entry.  If there is a lookup intent, inform
 740  * the MDS so that it gets our 'caps wanted' value in a single op.
 741  */
 742 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 743                                   unsigned int flags)
 744 {
 745         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 746         struct ceph_mds_client *mdsc = fsc->mdsc;
 747         struct ceph_mds_request *req;
 748         int op;
 749         int mask;
 750         int err;
 751
 752         dout("lookup %p dentry %p '%pd'\n",
 753              dir, dentry, dentry);
 754
 755         if (dentry->d_name.len > NAME_MAX)
 756                 return ERR_PTR(-ENAMETOOLONG);
 757
 758         /* can we conclude ENOENT locally? */
 759         if (d_really_is_negative(dentry)) {
 760                 struct ceph_inode_info *ci = ceph_inode(dir);
 761                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 762
 763                 spin_lock(&ci->i_ceph_lock);
 764                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 765                 if (strncmp(dentry->d_name.name,
 766                             fsc->mount_options->snapdir_name,
 767                             dentry->d_name.len) &&
 768                     !is_root_ceph_dentry(dir, dentry) &&
 769                     ceph_test_mount_opt(fsc, DCACHE) &&
 770                     __ceph_dir_is_complete(ci) &&
 771                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 772                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 773                         spin_unlock(&ci->i_ceph_lock);
 774                         dout(" dir %p complete, -ENOENT\n", dir);
 775                         d_add(dentry, NULL);
 776                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 777                         return NULL;
 778                 }
 779                 spin_unlock(&ci->i_ceph_lock);
 780         }
 781
 782         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 783                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 784         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 785         if (IS_ERR(req))
 786                 return ERR_CAST(req);
 787         req->r_dentry = dget(dentry);
 788         req->r_num_caps = 2;
 789
 790         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 791         if (ceph_security_xattr_wanted(dir))
 792                 mask |= CEPH_CAP_XATTR_SHARED;
 793         req->r_args.getattr.mask = cpu_to_le32(mask);
 794
 795         req->r_parent = dir;
 796         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 797         err = ceph_mdsc_do_request(mdsc, NULL, req);
 798         err = ceph_handle_snapdir(req, dentry, err);
 799         dentry = ceph_finish_lookup(req, dentry, err);
 800         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 801         dout("lookup result=%p\n", dentry);
 802         return dentry;
 803 }
 804
 805 /*
 806  * If we do a create but get no trace back from the MDS, follow up with
 807  * a lookup (the VFS expects us to link up the provided dentry).
 808  */
 809 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 810 {
 811         struct dentry *result = ceph_lookup(dir, dentry, 0);
 812
 813         if (result && !IS_ERR(result)) {
 814                 /*
 815                  * We created the item, then did a lookup, and found
 816                  * it was already linked to another inode we already
 817                  * had in our cache (and thus got spliced). To not
 818                  * confuse VFS (especially when inode is a directory),
 819                  * we don't link our dentry to that inode, return an
 820                  * error instead.
 821                  *
 822                  * This event should be rare and it happens only when
 823                  * we talk to old MDS. Recent MDS does not send traceless
 824                  * reply for request that creates new inode.
 825                  */
 826                 d_drop(result);
 827                 return -ESTALE;
 828         }
 829         return PTR_ERR(result);
 830 }
 831
 832 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 833                       umode_t mode, dev_t rdev)
 834 {
 835         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 836         struct ceph_mds_client *mdsc = fsc->mdsc;
 837         struct ceph_mds_request *req;
 838         struct ceph_acl_sec_ctx as_ctx = {};
 839         int err;
 840
 841         if (ceph_snap(dir) != CEPH_NOSNAP)
 842                 return -EROFS;
 843
 844         if (ceph_quota_is_max_files_exceeded(dir)) {
 845                 err = -EDQUOT;
 846                 goto out;
 847         }
 848
 849         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 850         if (err < 0)
 851                 goto out;
 852         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 853         if (err < 0)
 854                 goto out;
 855
 856         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 857              dir, dentry, mode, rdev);
 858         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 859         if (IS_ERR(req)) {
 860                 err = PTR_ERR(req);
 861                 goto out;
 862         }
 863         req->r_dentry = dget(dentry);
 864         req->r_num_caps = 2;
 865         req->r_parent = dir;
 866         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 867         req->r_args.mknod.mode = cpu_to_le32(mode);
 868         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 869         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 870         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 871         if (as_ctx.pagelist) {
 872                 req->r_pagelist = as_ctx.pagelist;
 873                 as_ctx.pagelist = NULL;
 874         }
 875         err = ceph_mdsc_do_request(mdsc, dir, req);
 876         if (!err && !req->r_reply_info.head->is_dentry)
 877                 err = ceph_handle_notrace_create(dir, dentry);
 878         ceph_mdsc_put_request(req);
 879 out:
 880         if (!err)
 881                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 882         else
 883                 d_drop(dentry);
 884         ceph_release_acl_sec_ctx(&as_ctx);
 885         return err;
 886 }
 887
 888 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 889                        bool excl)
 890 {
 891         return ceph_mknod(dir, dentry, mode, 0);
 892 }
 893
 894 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 895                             const char *dest)
 896 {
 897         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 898         struct ceph_mds_client *mdsc = fsc->mdsc;
 899         struct ceph_mds_request *req;
 900         struct ceph_acl_sec_ctx as_ctx = {};
 901         int err;
 902
 903         if (ceph_snap(dir) != CEPH_NOSNAP)
 904                 return -EROFS;
 905
 906         if (ceph_quota_is_max_files_exceeded(dir)) {
 907                 err = -EDQUOT;
 908                 goto out;
 909         }
 910
 911         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 912         if (err < 0)
 913                 goto out;
 914
 915         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 916         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 917         if (IS_ERR(req)) {
 918                 err = PTR_ERR(req);
 919                 goto out;
 920         }
 921         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 922         if (!req->r_path2) {
 923                 err = -ENOMEM;
 924                 ceph_mdsc_put_request(req);
 925                 goto out;
 926         }
 927         req->r_parent = dir;
 928         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 929         req->r_dentry = dget(dentry);
 930         req->r_num_caps = 2;
 931         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 932         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 933         err = ceph_mdsc_do_request(mdsc, dir, req);
 934         if (!err && !req->r_reply_info.head->is_dentry)
 935                 err = ceph_handle_notrace_create(dir, dentry);
 936         ceph_mdsc_put_request(req);
 937 out:
 938         if (err)
 939                 d_drop(dentry);
 940         ceph_release_acl_sec_ctx(&as_ctx);
 941         return err;
 942 }
 943
 944 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 945 {
 946         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 947         struct ceph_mds_client *mdsc = fsc->mdsc;
 948         struct ceph_mds_request *req;
 949         struct ceph_acl_sec_ctx as_ctx = {};
 950         int err = -EROFS;
 951         int op;
 952
 953         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 954                 /* mkdir .snap/foo is a MKSNAP */
 955                 op = CEPH_MDS_OP_MKSNAP;
 956                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 957                      dentry, dentry);
 958         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 959                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 960                 op = CEPH_MDS_OP_MKDIR;
 961         } else {
 962                 goto out;
 963         }
 964
 965         if (op == CEPH_MDS_OP_MKDIR &&
 966             ceph_quota_is_max_files_exceeded(dir)) {
 967                 err = -EDQUOT;
 968                 goto out;
 969         }
 970
 971         mode |= S_IFDIR;
 972         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 973         if (err < 0)
 974                 goto out;
 975         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 976         if (err < 0)
 977                 goto out;
 978
 979         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 980         if (IS_ERR(req)) {
 981                 err = PTR_ERR(req);
 982                 goto out;
 983         }
 984
 985         req->r_dentry = dget(dentry);
 986         req->r_num_caps = 2;
 987         req->r_parent = dir;
 988         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 989         req->r_args.mkdir.mode = cpu_to_le32(mode);
 990         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 991         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 992         if (as_ctx.pagelist) {
 993                 req->r_pagelist = as_ctx.pagelist;
 994                 as_ctx.pagelist = NULL;
 995         }
 996         err = ceph_mdsc_do_request(mdsc, dir, req);
 997         if (!err &&
 998             !req->r_reply_info.head->is_target &&
 999             !req->r_reply_info.head->is_dentry)
1000                 err = ceph_handle_notrace_create(dir, dentry);
1001         ceph_mdsc_put_request(req);
1002 out:
1003         if (!err)
1004                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1005         else
1006                 d_drop(dentry);
1007         ceph_release_acl_sec_ctx(&as_ctx);
1008         return err;
1009 }
1010
1011 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1012                      struct dentry *dentry)
1013 {
1014         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1015         struct ceph_mds_client *mdsc = fsc->mdsc;
1016         struct ceph_mds_request *req;
1017         int err;
1018
1019         if (ceph_snap(dir) != CEPH_NOSNAP)
1020                 return -EROFS;
1021
1022         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1023              old_dentry, dentry);
1024         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1025         if (IS_ERR(req)) {
1026                 d_drop(dentry);
1027                 return PTR_ERR(req);
1028         }
1029         req->r_dentry = dget(dentry);
1030         req->r_num_caps = 2;
1031         req->r_old_dentry = dget(old_dentry);
1032         req->r_parent = dir;
1033         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1034         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1035         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1036         /* release LINK_SHARED on source inode (mds will lock it) */
1037         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1038         err = ceph_mdsc_do_request(mdsc, dir, req);
1039         if (err) {
1040                 d_drop(dentry);
1041         } else if (!req->r_reply_info.head->is_dentry) {
1042                 ihold(d_inode(old_dentry));
1043                 d_instantiate(dentry, d_inode(old_dentry));
1044         }
1045         ceph_mdsc_put_request(req);
1046         return err;
1047 }
1048
1049 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1050                                  struct ceph_mds_request *req)
1051 {
1052         int result = req->r_err ? req->r_err :
1053                         le32_to_cpu(req->r_reply_info.head->result);
1054
1055         if (result == -EJUKEBOX)
1056                 goto out;
1057
1058         /* If op failed, mark everyone involved for errors */
1059         if (result) {
1060                 int pathlen = 0;
1061                 u64 base = 0;
1062                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1063                                                   &base, 0);
1064
1065                 /* mark error on parent + clear complete */
1066                 mapping_set_error(req->r_parent->i_mapping, result);
1067                 ceph_dir_clear_complete(req->r_parent);
1068
1069                 /* drop the dentry -- we don't know its status */
1070                 if (!d_unhashed(req->r_dentry))
1071                         d_drop(req->r_dentry);
1072
1073                 /* mark inode itself for an error (since metadata is bogus) */
1074                 mapping_set_error(req->r_old_inode->i_mapping, result);
1075
1076                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1077                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1078                 ceph_mdsc_free_path(path, pathlen);
1079         }
1080 out:
1081         iput(req->r_old_inode);
1082         ceph_mdsc_release_dir_caps(req);
1083 }
1084
1085 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1086 {
1087         struct ceph_inode_info *ci = ceph_inode(dir);
1088         struct ceph_dentry_info *di;
1089         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1090
1091         spin_lock(&ci->i_ceph_lock);
1092         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1093                 ceph_take_cap_refs(ci, want, false);
1094                 got = want;
1095         }
1096         spin_unlock(&ci->i_ceph_lock);
1097
1098         /* If we didn't get anything, return 0 */
1099         if (!got)
1100                 return 0;
1101
1102         spin_lock(&dentry->d_lock);
1103         di = ceph_dentry(dentry);
1104         /*
1105          * - We are holding Fx, which implies Fs caps.
1106          * - Only support async unlink for primary linkage
1107          */
1108         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1109             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1110                 want = 0;
1111         spin_unlock(&dentry->d_lock);
1112
1113         /* Do we still want what we've got? */
1114         if (want == got)
1115                 return got;
1116
1117         ceph_put_cap_refs(ci, got);
1118         return 0;
1119 }
1120
1121 /*
1122  * rmdir and unlink are differ only by the metadata op code
1123  */
1124 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1125 {
1126         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1127         struct ceph_mds_client *mdsc = fsc->mdsc;
1128         struct inode *inode = d_inode(dentry);
1129         struct ceph_mds_request *req;
1130         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1131         int err = -EROFS;
1132         int op;
1133
1134         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1135                 /* rmdir .snap/foo is RMSNAP */
1136                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1137                 op = CEPH_MDS_OP_RMSNAP;
1138         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1139                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1140                      dir, dentry, inode);
1141                 op = d_is_dir(dentry) ?
1142                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1143         } else
1144                 goto out;
1145 retry:
1146         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1147         if (IS_ERR(req)) {
1148                 err = PTR_ERR(req);
1149                 goto out;
1150         }
1151         req->r_dentry = dget(dentry);
1152         req->r_num_caps = 2;
1153         req->r_parent = dir;
1154         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1155         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1156         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1157
1158         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1159             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1160                 dout("async unlink on %lu/%.*s caps=%s", dir->i_ino,
1161                      dentry->d_name.len, dentry->d_name.name,
1162                      ceph_cap_string(req->r_dir_caps));
1163                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1164                 req->r_callback = ceph_async_unlink_cb;
1165                 req->r_old_inode = d_inode(dentry);
1166                 ihold(req->r_old_inode);
1167                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1168                 if (!err) {
1169                         /*
1170                          * We have enough caps, so we assume that the unlink
1171                          * will succeed. Fix up the target inode and dcache.
1172                          */
1173                         drop_nlink(inode);
1174                         d_delete(dentry);
1175                 } else if (err == -EJUKEBOX) {
1176                         try_async = false;
1177                         ceph_mdsc_put_request(req);
1178                         goto retry;
1179                 }
1180         } else {
1181                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1182                 err = ceph_mdsc_do_request(mdsc, dir, req);
1183                 if (!err && !req->r_reply_info.head->is_dentry)
1184                         d_delete(dentry);
1185         }
1186
1187         ceph_mdsc_put_request(req);
1188 out:
1189         return err;
1190 }
1191
1192 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1193                        struct inode *new_dir, struct dentry *new_dentry,
1194                        unsigned int flags)
1195 {
1196         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1197         struct ceph_mds_client *mdsc = fsc->mdsc;
1198         struct ceph_mds_request *req;
1199         int op = CEPH_MDS_OP_RENAME;
1200         int err;
1201
1202         if (flags)
1203                 return -EINVAL;
1204
1205         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1206                 return -EXDEV;
1207         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1208                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1209                         op = CEPH_MDS_OP_RENAMESNAP;
1210                 else
1211                         return -EROFS;
1212         } else if (old_dir != new_dir) {
1213                 err = ceph_quota_check_rename(mdsc, d_inode(old_dentry),
1214                                               new_dir);
1215                 if (err)
1216                         return err;
1217         }
1218
1219         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1220              old_dir, old_dentry, new_dir, new_dentry);
1221         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1222         if (IS_ERR(req))
1223                 return PTR_ERR(req);
1224         ihold(old_dir);
1225         req->r_dentry = dget(new_dentry);
1226         req->r_num_caps = 2;
1227         req->r_old_dentry = dget(old_dentry);
1228         req->r_old_dentry_dir = old_dir;
1229         req->r_parent = new_dir;
1230         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1231         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1232         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1233         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1234         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1235         /* release LINK_RDCACHE on source inode (mds will lock it) */
1236         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1237         if (d_really_is_positive(new_dentry)) {
1238                 req->r_inode_drop =
1239                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1240         }
1241         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1242         if (!err && !req->r_reply_info.head->is_dentry) {
1243                 /*
1244                  * Normally d_move() is done by fill_trace (called by
1245                  * do_request, above).  If there is no trace, we need
1246                  * to do it here.
1247                  */
1248                 d_move(old_dentry, new_dentry);
1249         }
1250         ceph_mdsc_put_request(req);
1251         return err;
1252 }
1253
1254 /*
1255  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1256  * Leases at front of the list will expire first. (Assume all leases have
1257  * similar duration)
1258  *
1259  * Called under dentry->d_lock.
1260  */
1261 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1262 {
1263         struct dentry *dn = di->dentry;
1264         struct ceph_mds_client *mdsc;
1265
1266         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1267
1268         di->flags |= CEPH_DENTRY_LEASE_LIST;
1269         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1270                 di->flags |= CEPH_DENTRY_REFERENCED;
1271                 return;
1272         }
1273
1274         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1275         spin_lock(&mdsc->dentry_list_lock);
1276         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1277         spin_unlock(&mdsc->dentry_list_lock);
1278 }
1279
1280 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1281                                      struct ceph_dentry_info *di)
1282 {
1283         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1284         di->lease_gen = 0;
1285         di->time = jiffies;
1286         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1287 }
1288
1289 /*
1290  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1291  * list if it's not in the list, otherwise set 'referenced' flag.
1292  *
1293  * Called under dentry->d_lock.
1294  */
1295 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1296 {
1297         struct dentry *dn = di->dentry;
1298         struct ceph_mds_client *mdsc;
1299
1300         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1301              di, dn, dn, di->offset);
1302
1303         if (!list_empty(&di->lease_list)) {
1304                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1305                         /* don't remove dentry from dentry lease list
1306                          * if its lease is valid */
1307                         if (__dentry_lease_is_valid(di))
1308                                 return;
1309                 } else {
1310                         di->flags |= CEPH_DENTRY_REFERENCED;
1311                         return;
1312                 }
1313         }
1314
1315         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1316                 di->flags |= CEPH_DENTRY_REFERENCED;
1317                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1318                 return;
1319         }
1320
1321         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1322         spin_lock(&mdsc->dentry_list_lock);
1323         __dentry_dir_lease_touch(mdsc, di),
1324         spin_unlock(&mdsc->dentry_list_lock);
1325 }
1326
1327 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1328 {
1329         struct ceph_mds_client *mdsc;
1330         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1331                 return;
1332         if (list_empty(&di->lease_list))
1333                 return;
1334
1335         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1336         spin_lock(&mdsc->dentry_list_lock);
1337         list_del_init(&di->lease_list);
1338         spin_unlock(&mdsc->dentry_list_lock);
1339 }
1340
1341 enum {
1342         KEEP    = 0,
1343         DELETE  = 1,
1344         TOUCH   = 2,
1345         STOP    = 4,
1346 };
1347
1348 struct ceph_lease_walk_control {
1349         bool dir_lease;
1350         bool expire_dir_lease;
1351         unsigned long nr_to_scan;
1352         unsigned long dir_lease_ttl;
1353 };
1354
1355 static unsigned long
1356 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1357                      struct ceph_lease_walk_control *lwc,
1358                      int (*check)(struct dentry*, void*))
1359 {
1360         struct ceph_dentry_info *di, *tmp;
1361         struct dentry *dentry, *last = NULL;
1362         struct list_head* list;
1363         LIST_HEAD(dispose);
1364         unsigned long freed = 0;
1365         int ret = 0;
1366
1367         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1368         spin_lock(&mdsc->dentry_list_lock);
1369         list_for_each_entry_safe(di, tmp, list, lease_list) {
1370                 if (!lwc->nr_to_scan)
1371                         break;
1372                 --lwc->nr_to_scan;
1373
1374                 dentry = di->dentry;
1375                 if (last == dentry)
1376                         break;
1377
1378                 if (!spin_trylock(&dentry->d_lock))
1379                         continue;
1380
1381                 if (__lockref_is_dead(&dentry->d_lockref)) {
1382                         list_del_init(&di->lease_list);
1383                         goto next;
1384                 }
1385
1386                 ret = check(dentry, lwc);
1387                 if (ret & TOUCH) {
1388                         /* move it into tail of dir lease list */
1389                         __dentry_dir_lease_touch(mdsc, di);
1390                         if (!last)
1391                                 last = dentry;
1392                 }
1393                 if (ret & DELETE) {
1394                         /* stale lease */
1395                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1396                         if (dentry->d_lockref.count > 0) {
1397                                 /* update_dentry_lease() will re-add
1398                                  * it to lease list, or
1399                                  * ceph_d_delete() will return 1 when
1400                                  * last reference is dropped */
1401                                 list_del_init(&di->lease_list);
1402                         } else {
1403                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1404                                 list_move_tail(&di->lease_list, &dispose);
1405                                 dget_dlock(dentry);
1406                         }
1407                 }
1408 next:
1409                 spin_unlock(&dentry->d_lock);
1410                 if (ret & STOP)
1411                         break;
1412         }
1413         spin_unlock(&mdsc->dentry_list_lock);
1414
1415         while (!list_empty(&dispose)) {
1416                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1417                                       lease_list);
1418                 dentry = di->dentry;
1419                 spin_lock(&dentry->d_lock);
1420
1421                 list_del_init(&di->lease_list);
1422                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1423                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1424                         spin_lock(&mdsc->dentry_list_lock);
1425                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1426                                 list_add_tail(&di->lease_list,
1427                                               &mdsc->dentry_leases);
1428                         } else {
1429                                 __dentry_dir_lease_touch(mdsc, di);
1430                         }
1431                         spin_unlock(&mdsc->dentry_list_lock);
1432                 } else {
1433                         freed++;
1434                 }
1435
1436                 spin_unlock(&dentry->d_lock);
1437                 /* ceph_d_delete() does the trick */
1438                 dput(dentry);
1439         }
1440         return freed;
1441 }
1442
1443 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1444 {
1445         struct ceph_dentry_info *di = ceph_dentry(dentry);
1446         int ret;
1447
1448         if (__dentry_lease_is_valid(di))
1449                 return STOP;
1450         ret = __dir_lease_try_check(dentry);
1451         if (ret == -EBUSY)
1452                 return KEEP;
1453         if (ret > 0)
1454                 return TOUCH;
1455         return DELETE;
1456 }
1457
1458 static int __dir_lease_check(struct dentry *dentry, void *arg)
1459 {
1460         struct ceph_lease_walk_control *lwc = arg;
1461         struct ceph_dentry_info *di = ceph_dentry(dentry);
1462
1463         int ret = __dir_lease_try_check(dentry);
1464         if (ret == -EBUSY)
1465                 return KEEP;
1466         if (ret > 0) {
1467                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1468                         return STOP;
1469                 /* Move dentry to tail of dir lease list if we don't want
1470                  * to delete it. So dentries in the list are checked in a
1471                  * round robin manner */
1472                 if (!lwc->expire_dir_lease)
1473                         return TOUCH;
1474                 if (dentry->d_lockref.count > 0 ||
1475                     (di->flags & CEPH_DENTRY_REFERENCED))
1476                         return TOUCH;
1477                 /* invalidate dir lease */
1478                 di->lease_shared_gen = 0;
1479         }
1480         return DELETE;
1481 }
1482
1483 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1484 {
1485         struct ceph_lease_walk_control lwc;
1486         unsigned long count;
1487         unsigned long freed;
1488
1489         spin_lock(&mdsc->caps_list_lock);
1490         if (mdsc->caps_use_max > 0 &&
1491             mdsc->caps_use_count > mdsc->caps_use_max)
1492                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1493         else
1494                 count = 0;
1495         spin_unlock(&mdsc->caps_list_lock);
1496
1497         lwc.dir_lease = false;
1498         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1499         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1500         if (!lwc.nr_to_scan) /* more invalid leases */
1501                 return -EAGAIN;
1502
1503         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1504                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1505
1506         lwc.dir_lease = true;
1507         lwc.expire_dir_lease = freed < count;
1508         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1509         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1510         if (!lwc.nr_to_scan) /* more to check */
1511                 return -EAGAIN;
1512
1513         return freed > 0 ? 1 : 0;
1514 }
1515
1516 /*
1517  * Ensure a dentry lease will no longer revalidate.
1518  */
1519 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1520 {
1521         struct ceph_dentry_info *di = ceph_dentry(dentry);
1522         spin_lock(&dentry->d_lock);
1523         di->time = jiffies;
1524         di->lease_shared_gen = 0;
1525         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1526         __dentry_lease_unlist(di);
1527         spin_unlock(&dentry->d_lock);
1528 }
1529
1530 /*
1531  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1532  * renew if the least is more than half up.
1533  */
1534 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1535 {
1536         struct ceph_mds_session *session;
1537
1538         if (!di->lease_gen)
1539                 return false;
1540
1541         session = di->lease_session;
1542         if (session) {
1543                 u32 gen;
1544                 unsigned long ttl;
1545
1546                 spin_lock(&session->s_gen_ttl_lock);
1547                 gen = session->s_cap_gen;
1548                 ttl = session->s_cap_ttl;
1549                 spin_unlock(&session->s_gen_ttl_lock);
1550
1551                 if (di->lease_gen == gen &&
1552                     time_before(jiffies, ttl) &&
1553                     time_before(jiffies, di->time))
1554                         return true;
1555         }
1556         di->lease_gen = 0;
1557         return false;
1558 }
1559
1560 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1561 {
1562         struct ceph_dentry_info *di;
1563         struct ceph_mds_session *session = NULL;
1564         u32 seq = 0;
1565         int valid = 0;
1566
1567         spin_lock(&dentry->d_lock);
1568         di = ceph_dentry(dentry);
1569         if (di && __dentry_lease_is_valid(di)) {
1570                 valid = 1;
1571
1572                 if (di->lease_renew_after &&
1573                     time_after(jiffies, di->lease_renew_after)) {
1574                         /*
1575                          * We should renew. If we're in RCU walk mode
1576                          * though, we can't do that so just return
1577                          * -ECHILD.
1578                          */
1579                         if (flags & LOOKUP_RCU) {
1580                                 valid = -ECHILD;
1581                         } else {
1582                                 session = ceph_get_mds_session(di->lease_session);
1583                                 seq = di->lease_seq;
1584                                 di->lease_renew_after = 0;
1585                                 di->lease_renew_from = jiffies;
1586                         }
1587                 }
1588         }
1589         spin_unlock(&dentry->d_lock);
1590
1591         if (session) {
1592                 ceph_mdsc_lease_send_msg(session, dentry,
1593                                          CEPH_MDS_LEASE_RENEW, seq);
1594                 ceph_put_mds_session(session);
1595         }
1596         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1597         return valid;
1598 }
1599
1600 /*
1601  * Called under dentry->d_lock.
1602  */
1603 static int __dir_lease_try_check(const struct dentry *dentry)
1604 {
1605         struct ceph_dentry_info *di = ceph_dentry(dentry);
1606         struct inode *dir;
1607         struct ceph_inode_info *ci;
1608         int valid = 0;
1609
1610         if (!di->lease_shared_gen)
1611                 return 0;
1612         if (IS_ROOT(dentry))
1613                 return 0;
1614
1615         dir = d_inode(dentry->d_parent);
1616         ci = ceph_inode(dir);
1617
1618         if (spin_trylock(&ci->i_ceph_lock)) {
1619                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1620                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1621                         valid = 1;
1622                 spin_unlock(&ci->i_ceph_lock);
1623         } else {
1624                 valid = -EBUSY;
1625         }
1626
1627         if (!valid)
1628                 di->lease_shared_gen = 0;
1629         return valid;
1630 }
1631
1632 /*
1633  * Check if directory-wide content lease/cap is valid.
1634  */
1635 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1636                               struct ceph_mds_client *mdsc)
1637 {
1638         struct ceph_inode_info *ci = ceph_inode(dir);
1639         int valid;
1640         int shared_gen;
1641
1642         spin_lock(&ci->i_ceph_lock);
1643         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1644         if (valid) {
1645                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1646                 shared_gen = atomic_read(&ci->i_shared_gen);
1647         }
1648         spin_unlock(&ci->i_ceph_lock);
1649         if (valid) {
1650                 struct ceph_dentry_info *di;
1651                 spin_lock(&dentry->d_lock);
1652                 di = ceph_dentry(dentry);
1653                 if (dir == d_inode(dentry->d_parent) &&
1654                     di && di->lease_shared_gen == shared_gen)
1655                         __ceph_dentry_dir_lease_touch(di);
1656                 else
1657                         valid = 0;
1658                 spin_unlock(&dentry->d_lock);
1659         }
1660         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1661              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1662         return valid;
1663 }
1664
1665 /*
1666  * Check if cached dentry can be trusted.
1667  */
1668 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1669 {
1670         int valid = 0;
1671         struct dentry *parent;
1672         struct inode *dir, *inode;
1673         struct ceph_mds_client *mdsc;
1674
1675         if (flags & LOOKUP_RCU) {
1676                 parent = READ_ONCE(dentry->d_parent);
1677                 dir = d_inode_rcu(parent);
1678                 if (!dir)
1679                         return -ECHILD;
1680                 inode = d_inode_rcu(dentry);
1681         } else {
1682                 parent = dget_parent(dentry);
1683                 dir = d_inode(parent);
1684                 inode = d_inode(dentry);
1685         }
1686
1687         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1688              dentry, inode, ceph_dentry(dentry)->offset);
1689
1690         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1691
1692         /* always trust cached snapped dentries, snapdir dentry */
1693         if (ceph_snap(dir) != CEPH_NOSNAP) {
1694                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1695                      dentry, inode);
1696                 valid = 1;
1697         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1698                 valid = 1;
1699         } else {
1700                 valid = dentry_lease_is_valid(dentry, flags);
1701                 if (valid == -ECHILD)
1702                         return valid;
1703                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1704                         if (inode)
1705                                 valid = ceph_is_any_caps(inode);
1706                         else
1707                                 valid = 1;
1708                 }
1709         }
1710
1711         if (!valid) {
1712                 struct ceph_mds_request *req;
1713                 int op, err;
1714                 u32 mask;
1715
1716                 if (flags & LOOKUP_RCU)
1717                         return -ECHILD;
1718
1719                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1720
1721                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1722                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1723                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1724                 if (!IS_ERR(req)) {
1725                         req->r_dentry = dget(dentry);
1726                         req->r_num_caps = 2;
1727                         req->r_parent = dir;
1728
1729                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1730                         if (ceph_security_xattr_wanted(dir))
1731                                 mask |= CEPH_CAP_XATTR_SHARED;
1732                         req->r_args.getattr.mask = cpu_to_le32(mask);
1733
1734                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1735                         switch (err) {
1736                         case 0:
1737                                 if (d_really_is_positive(dentry) &&
1738                                     d_inode(dentry) == req->r_target_inode)
1739                                         valid = 1;
1740                                 break;
1741                         case -ENOENT:
1742                                 if (d_really_is_negative(dentry))
1743                                         valid = 1;
1744                                 /* Fallthrough */
1745                         default:
1746                                 break;
1747                         }
1748                         ceph_mdsc_put_request(req);
1749                         dout("d_revalidate %p lookup result=%d\n",
1750                              dentry, err);
1751                 }
1752         } else {
1753                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1754         }
1755
1756         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1757         if (!valid)
1758                 ceph_dir_clear_complete(dir);
1759
1760         if (!(flags & LOOKUP_RCU))
1761                 dput(parent);
1762         return valid;
1763 }
1764
1765 /*
1766  * Delete unused dentry that doesn't have valid lease
1767  *
1768  * Called under dentry->d_lock.
1769  */
1770 static int ceph_d_delete(const struct dentry *dentry)
1771 {
1772         struct ceph_dentry_info *di;
1773
1774         /* won't release caps */
1775         if (d_really_is_negative(dentry))
1776                 return 0;
1777         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1778                 return 0;
1779         /* vaild lease? */
1780         di = ceph_dentry(dentry);
1781         if (di) {
1782                 if (__dentry_lease_is_valid(di))
1783                         return 0;
1784                 if (__dir_lease_try_check(dentry))
1785                         return 0;
1786         }
1787         return 1;
1788 }
1789
1790 /*
1791  * Release our ceph_dentry_info.
1792  */
1793 static void ceph_d_release(struct dentry *dentry)
1794 {
1795         struct ceph_dentry_info *di = ceph_dentry(dentry);
1796         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1797
1798         dout("d_release %p\n", dentry);
1799
1800         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1801
1802         spin_lock(&dentry->d_lock);
1803         __dentry_lease_unlist(di);
1804         dentry->d_fsdata = NULL;
1805         spin_unlock(&dentry->d_lock);
1806
1807         if (di->lease_session)
1808                 ceph_put_mds_session(di->lease_session);
1809         kmem_cache_free(ceph_dentry_cachep, di);
1810 }
1811
1812 /*
1813  * When the VFS prunes a dentry from the cache, we need to clear the
1814  * complete flag on the parent directory.
1815  *
1816  * Called under dentry->d_lock.
1817  */
1818 static void ceph_d_prune(struct dentry *dentry)
1819 {
1820         struct ceph_inode_info *dir_ci;
1821         struct ceph_dentry_info *di;
1822
1823         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1824
1825         /* do we have a valid parent? */
1826         if (IS_ROOT(dentry))
1827                 return;
1828
1829         /* we hold d_lock, so d_parent is stable */
1830         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1831         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1832                 return;
1833
1834         /* who calls d_delete() should also disable dcache readdir */
1835         if (d_really_is_negative(dentry))
1836                 return;
1837
1838         /* d_fsdata does not get cleared until d_release */
1839         if (!d_unhashed(dentry)) {
1840                 __ceph_dir_clear_complete(dir_ci);
1841                 return;
1842         }
1843
1844         /* Disable dcache readdir just in case that someone called d_drop()
1845          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1846          * properly (dcache readdir is still enabled) */
1847         di = ceph_dentry(dentry);
1848         if (di->offset > 0 &&
1849             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1850                 __ceph_dir_clear_ordered(dir_ci);
1851 }
1852
1853 /*
1854  * read() on a dir.  This weird interface hack only works if mounted
1855  * with '-o dirstat'.
1856  */
1857 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1858                              loff_t *ppos)
1859 {
1860         struct ceph_dir_file_info *dfi = file->private_data;
1861         struct inode *inode = file_inode(file);
1862         struct ceph_inode_info *ci = ceph_inode(inode);
1863         int left;
1864         const int bufsize = 1024;
1865
1866         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1867                 return -EISDIR;
1868
1869         if (!dfi->dir_info) {
1870                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1871                 if (!dfi->dir_info)
1872                         return -ENOMEM;
1873                 dfi->dir_info_len =
1874                         snprintf(dfi->dir_info, bufsize,
1875                                 "entries:   %20lld\n"
1876                                 " files:    %20lld\n"
1877                                 " subdirs:  %20lld\n"
1878                                 "rentries:  %20lld\n"
1879                                 " rfiles:   %20lld\n"
1880                                 " rsubdirs: %20lld\n"
1881                                 "rbytes:    %20lld\n"
1882                                 "rctime:    %10lld.%09ld\n",
1883                                 ci->i_files + ci->i_subdirs,
1884                                 ci->i_files,
1885                                 ci->i_subdirs,
1886                                 ci->i_rfiles + ci->i_rsubdirs,
1887                                 ci->i_rfiles,
1888                                 ci->i_rsubdirs,
1889                                 ci->i_rbytes,
1890                                 ci->i_rctime.tv_sec,
1891                                 ci->i_rctime.tv_nsec);
1892         }
1893
1894         if (*ppos >= dfi->dir_info_len)
1895                 return 0;
1896         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1897         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1898         if (left == size)
1899                 return -EFAULT;
1900         *ppos += (size - left);
1901         return size - left;
1902 }
1903
1904
1905
1906 /*
1907  * Return name hash for a given dentry.  This is dependent on
1908  * the parent directory's hash function.
1909  */
1910 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1911 {
1912         struct ceph_inode_info *dci = ceph_inode(dir);
1913         unsigned hash;
1914
1915         switch (dci->i_dir_layout.dl_dir_hash) {
1916         case 0: /* for backward compat */
1917         case CEPH_STR_HASH_LINUX:
1918                 return dn->d_name.hash;
1919
1920         default:
1921                 spin_lock(&dn->d_lock);
1922                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1923                                      dn->d_name.name, dn->d_name.len);
1924                 spin_unlock(&dn->d_lock);
1925                 return hash;
1926         }
1927 }
1928
1929 const struct file_operations ceph_dir_fops = {
1930         .read = ceph_read_dir,
1931         .iterate = ceph_readdir,
1932         .llseek = ceph_dir_llseek,
1933         .open = ceph_open,
1934         .release = ceph_release,
1935         .unlocked_ioctl = ceph_ioctl,
1936         .compat_ioctl = compat_ptr_ioctl,
1937         .fsync = ceph_fsync,
1938         .lock = ceph_lock,
1939         .flock = ceph_flock,
1940 };
1941
1942 const struct file_operations ceph_snapdir_fops = {
1943         .iterate = ceph_readdir,
1944         .llseek = ceph_dir_llseek,
1945         .open = ceph_open,
1946         .release = ceph_release,
1947 };
1948
1949 const struct inode_operations ceph_dir_iops = {
1950         .lookup = ceph_lookup,
1951         .permission = ceph_permission,
1952         .getattr = ceph_getattr,
1953         .setattr = ceph_setattr,
1954         .listxattr = ceph_listxattr,
1955         .get_acl = ceph_get_acl,
1956         .set_acl = ceph_set_acl,
1957         .mknod = ceph_mknod,
1958         .symlink = ceph_symlink,
1959         .mkdir = ceph_mkdir,
1960         .link = ceph_link,
1961         .unlink = ceph_unlink,
1962         .rmdir = ceph_unlink,
1963         .rename = ceph_rename,
1964         .create = ceph_create,
1965         .atomic_open = ceph_atomic_open,
1966 };
1967
1968 const struct inode_operations ceph_snapdir_iops = {
1969         .lookup = ceph_lookup,
1970         .permission = ceph_permission,
1971         .getattr = ceph_getattr,
1972         .mkdir = ceph_mkdir,
1973         .rmdir = ceph_unlink,
1974         .rename = ceph_rename,
1975 };
1976
1977 const struct dentry_operations ceph_dentry_ops = {
1978         .d_revalidate = ceph_d_revalidate,
1979         .d_delete = ceph_d_delete,
1980         .d_release = ceph_d_release,
1981         .d_prune = ceph_d_prune,
1982         .d_init = ceph_d_init,
1983 };