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                                 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
 511                 BUG_ON(rde->offset < ctx->pos);
 512
 513                 ctx->pos = rde->offset;
 514                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 515                      i, rinfo->dir_nr, ctx->pos,
 516                      rde->name_len, rde->name, &rde->inode.in);
 517
 518                 BUG_ON(!rde->inode.in);
 519
 520                 if (!dir_emit(ctx, rde->name, rde->name_len,
 521                               ceph_present_ino(inode->i_sb, le64_to_cpu(rde->inode.in->ino)),
 522                               le32_to_cpu(rde->inode.in->mode) >> 12)) {
 523                         dout("filldir stopping us...\n");
 524                         return 0;
 525                 }
 526                 ctx->pos++;
 527         }
 528
 529         ceph_mdsc_put_request(dfi->last_readdir);
 530         dfi->last_readdir = NULL;
 531
 532         if (dfi->next_offset > 2) {
 533                 frag = dfi->frag;
 534                 goto more;
 535         }
 536
 537         /* more frags? */
 538         if (!ceph_frag_is_rightmost(dfi->frag)) {
 539                 frag = ceph_frag_next(dfi->frag);
 540                 if (is_hash_order(ctx->pos)) {
 541                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 542                                                         dfi->next_offset, true);
 543                         if (new_pos > ctx->pos)
 544                                 ctx->pos = new_pos;
 545                         /* keep last_name */
 546                 } else {
 547                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 548                                                         false);
 549                         kfree(dfi->last_name);
 550                         dfi->last_name = NULL;
 551                 }
 552                 dout("readdir next frag is %x\n", frag);
 553                 goto more;
 554         }
 555         dfi->file_info.flags |= CEPH_F_ATEND;
 556
 557         /*
 558          * if dir_release_count still matches the dir, no dentries
 559          * were released during the whole readdir, and we should have
 560          * the complete dir contents in our cache.
 561          */
 562         if (atomic64_read(&ci->i_release_count) ==
 563                                         dfi->dir_release_count) {
 564                 spin_lock(&ci->i_ceph_lock);
 565                 if (dfi->dir_ordered_count ==
 566                                 atomic64_read(&ci->i_ordered_count)) {
 567                         dout(" marking %p complete and ordered\n", inode);
 568                         /* use i_size to track number of entries in
 569                          * readdir cache */
 570                         BUG_ON(dfi->readdir_cache_idx < 0);
 571                         i_size_write(inode, dfi->readdir_cache_idx *
 572                                      sizeof(struct dentry*));
 573                 } else {
 574                         dout(" marking %p complete\n", inode);
 575                 }
 576                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 577                                         dfi->dir_ordered_count);
 578                 spin_unlock(&ci->i_ceph_lock);
 579         }
 580
 581         dout("readdir %p file %p done.\n", inode, file);
 582         return 0;
 583 }
 584
 585 static void reset_readdir(struct ceph_dir_file_info *dfi)
 586 {
 587         if (dfi->last_readdir) {
 588                 ceph_mdsc_put_request(dfi->last_readdir);
 589                 dfi->last_readdir = NULL;
 590         }
 591         kfree(dfi->last_name);
 592         dfi->last_name = NULL;
 593         dfi->dir_release_count = 0;
 594         dfi->readdir_cache_idx = -1;
 595         dfi->next_offset = 2;  /* compensate for . and .. */
 596         dfi->file_info.flags &= ~CEPH_F_ATEND;
 597 }
 598
 599 /*
 600  * discard buffered readdir content on seekdir(0), or seek to new frag,
 601  * or seek prior to current chunk
 602  */
 603 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 604 {
 605         struct ceph_mds_reply_info_parsed *rinfo;
 606         loff_t chunk_offset;
 607         if (new_pos == 0)
 608                 return true;
 609         if (is_hash_order(new_pos)) {
 610                 /* no need to reset last_name for a forward seek when
 611                  * dentries are sotred in hash order */
 612         } else if (dfi->frag != fpos_frag(new_pos)) {
 613                 return true;
 614         }
 615         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 616         if (!rinfo || !rinfo->dir_nr)
 617                 return true;
 618         chunk_offset = rinfo->dir_entries[0].offset;
 619         return new_pos < chunk_offset ||
 620                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 621 }
 622
 623 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 624 {
 625         struct ceph_dir_file_info *dfi = file->private_data;
 626         struct inode *inode = file->f_mapping->host;
 627         loff_t retval;
 628
 629         inode_lock(inode);
 630         retval = -EINVAL;
 631         switch (whence) {
 632         case SEEK_CUR:
 633                 offset += file->f_pos;
 634         case SEEK_SET:
 635                 break;
 636         case SEEK_END:
 637                 retval = -EOPNOTSUPP;
 638         default:
 639                 goto out;
 640         }
 641
 642         if (offset >= 0) {
 643                 if (need_reset_readdir(dfi, offset)) {
 644                         dout("dir_llseek dropping %p content\n", file);
 645                         reset_readdir(dfi);
 646                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 647                         /* for hash offset, we don't know if a forward seek
 648                          * is within same frag */
 649                         dfi->dir_release_count = 0;
 650                         dfi->readdir_cache_idx = -1;
 651                 }
 652
 653                 if (offset != file->f_pos) {
 654                         file->f_pos = offset;
 655                         file->f_version = 0;
 656                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 657                 }
 658                 retval = offset;
 659         }
 660 out:
 661         inode_unlock(inode);
 662         return retval;
 663 }
 664
 665 /*
 666  * Handle lookups for the hidden .snap directory.
 667  */
 668 int ceph_handle_snapdir(struct ceph_mds_request *req,
 669                         struct dentry *dentry, int err)
 670 {
 671         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 672         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 673
 674         /* .snap dir? */
 675         if (err == -ENOENT &&
 676             ceph_snap(parent) == CEPH_NOSNAP &&
 677             strcmp(dentry->d_name.name,
 678                    fsc->mount_options->snapdir_name) == 0) {
 679                 struct inode *inode = ceph_get_snapdir(parent);
 680                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 681                      dentry, dentry, inode);
 682                 BUG_ON(!d_unhashed(dentry));
 683                 d_add(dentry, inode);
 684                 err = 0;
 685         }
 686         return err;
 687 }
 688
 689 /*
 690  * Figure out final result of a lookup/open request.
 691  *
 692  * Mainly, make sure we return the final req->r_dentry (if it already
 693  * existed) in place of the original VFS-provided dentry when they
 694  * differ.
 695  *
 696  * Gracefully handle the case where the MDS replies with -ENOENT and
 697  * no trace (which it may do, at its discretion, e.g., if it doesn't
 698  * care to issue a lease on the negative dentry).
 699  */
 700 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 701                                   struct dentry *dentry, int err)
 702 {
 703         if (err == -ENOENT) {
 704                 /* no trace? */
 705                 err = 0;
 706                 if (!req->r_reply_info.head->is_dentry) {
 707                         dout("ENOENT and no trace, dentry %p inode %p\n",
 708                              dentry, d_inode(dentry));
 709                         if (d_really_is_positive(dentry)) {
 710                                 d_drop(dentry);
 711                                 err = -ENOENT;
 712                         } else {
 713                                 d_add(dentry, NULL);
 714                         }
 715                 }
 716         }
 717         if (err)
 718                 dentry = ERR_PTR(err);
 719         else if (dentry != req->r_dentry)
 720                 dentry = dget(req->r_dentry);   /* we got spliced */
 721         else
 722                 dentry = NULL;
 723         return dentry;
 724 }
 725
 726 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 727 {
 728         return ceph_ino(inode) == CEPH_INO_ROOT &&
 729                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 730 }
 731
 732 /*
 733  * Look up a single dir entry.  If there is a lookup intent, inform
 734  * the MDS so that it gets our 'caps wanted' value in a single op.
 735  */
 736 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 737                                   unsigned int flags)
 738 {
 739         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 740         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 741         struct ceph_mds_request *req;
 742         int op;
 743         int mask;
 744         int err;
 745
 746         dout("lookup %p dentry %p '%pd'\n",
 747              dir, dentry, dentry);
 748
 749         if (dentry->d_name.len > NAME_MAX)
 750                 return ERR_PTR(-ENAMETOOLONG);
 751
 752         /* can we conclude ENOENT locally? */
 753         if (d_really_is_negative(dentry)) {
 754                 struct ceph_inode_info *ci = ceph_inode(dir);
 755                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 756
 757                 spin_lock(&ci->i_ceph_lock);
 758                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 759                 if (strncmp(dentry->d_name.name,
 760                             fsc->mount_options->snapdir_name,
 761                             dentry->d_name.len) &&
 762                     !is_root_ceph_dentry(dir, dentry) &&
 763                     ceph_test_mount_opt(fsc, DCACHE) &&
 764                     __ceph_dir_is_complete(ci) &&
 765                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 766                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 767                         spin_unlock(&ci->i_ceph_lock);
 768                         dout(" dir %p complete, -ENOENT\n", dir);
 769                         d_add(dentry, NULL);
 770                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 771                         return NULL;
 772                 }
 773                 spin_unlock(&ci->i_ceph_lock);
 774         }
 775
 776         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 777                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 778         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 779         if (IS_ERR(req))
 780                 return ERR_CAST(req);
 781         req->r_dentry = dget(dentry);
 782         req->r_num_caps = 2;
 783
 784         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 785         if (ceph_security_xattr_wanted(dir))
 786                 mask |= CEPH_CAP_XATTR_SHARED;
 787         req->r_args.getattr.mask = cpu_to_le32(mask);
 788
 789         req->r_parent = dir;
 790         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 791         err = ceph_mdsc_do_request(mdsc, NULL, req);
 792         err = ceph_handle_snapdir(req, dentry, err);
 793         dentry = ceph_finish_lookup(req, dentry, err);
 794         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 795         dout("lookup result=%p\n", dentry);
 796         return dentry;
 797 }
 798
 799 /*
 800  * If we do a create but get no trace back from the MDS, follow up with
 801  * a lookup (the VFS expects us to link up the provided dentry).
 802  */
 803 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 804 {
 805         struct dentry *result = ceph_lookup(dir, dentry, 0);
 806
 807         if (result && !IS_ERR(result)) {
 808                 /*
 809                  * We created the item, then did a lookup, and found
 810                  * it was already linked to another inode we already
 811                  * had in our cache (and thus got spliced). To not
 812                  * confuse VFS (especially when inode is a directory),
 813                  * we don't link our dentry to that inode, return an
 814                  * error instead.
 815                  *
 816                  * This event should be rare and it happens only when
 817                  * we talk to old MDS. Recent MDS does not send traceless
 818                  * reply for request that creates new inode.
 819                  */
 820                 d_drop(result);
 821                 return -ESTALE;
 822         }
 823         return PTR_ERR(result);
 824 }
 825
 826 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 827                       umode_t mode, dev_t rdev)
 828 {
 829         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 830         struct ceph_mds_request *req;
 831         struct ceph_acl_sec_ctx as_ctx = {};
 832         int err;
 833
 834         if (ceph_snap(dir) != CEPH_NOSNAP)
 835                 return -EROFS;
 836
 837         if (ceph_quota_is_max_files_exceeded(dir)) {
 838                 err = -EDQUOT;
 839                 goto out;
 840         }
 841
 842         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 843         if (err < 0)
 844                 goto out;
 845         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 846         if (err < 0)
 847                 goto out;
 848
 849         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 850              dir, dentry, mode, rdev);
 851         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 852         if (IS_ERR(req)) {
 853                 err = PTR_ERR(req);
 854                 goto out;
 855         }
 856         req->r_dentry = dget(dentry);
 857         req->r_num_caps = 2;
 858         req->r_parent = dir;
 859         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 860         req->r_args.mknod.mode = cpu_to_le32(mode);
 861         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 862         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 863         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 864         if (as_ctx.pagelist) {
 865                 req->r_pagelist = as_ctx.pagelist;
 866                 as_ctx.pagelist = NULL;
 867         }
 868         err = ceph_mdsc_do_request(mdsc, dir, req);
 869         if (!err && !req->r_reply_info.head->is_dentry)
 870                 err = ceph_handle_notrace_create(dir, dentry);
 871         ceph_mdsc_put_request(req);
 872 out:
 873         if (!err)
 874                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 875         else
 876                 d_drop(dentry);
 877         ceph_release_acl_sec_ctx(&as_ctx);
 878         return err;
 879 }
 880
 881 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 882                        bool excl)
 883 {
 884         return ceph_mknod(dir, dentry, mode, 0);
 885 }
 886
 887 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 888                             const char *dest)
 889 {
 890         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 891         struct ceph_mds_request *req;
 892         struct ceph_acl_sec_ctx as_ctx = {};
 893         int err;
 894
 895         if (ceph_snap(dir) != CEPH_NOSNAP)
 896                 return -EROFS;
 897
 898         if (ceph_quota_is_max_files_exceeded(dir)) {
 899                 err = -EDQUOT;
 900                 goto out;
 901         }
 902
 903         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 904         if (err < 0)
 905                 goto out;
 906
 907         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 908         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 909         if (IS_ERR(req)) {
 910                 err = PTR_ERR(req);
 911                 goto out;
 912         }
 913         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 914         if (!req->r_path2) {
 915                 err = -ENOMEM;
 916                 ceph_mdsc_put_request(req);
 917                 goto out;
 918         }
 919         req->r_parent = dir;
 920         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 921         req->r_dentry = dget(dentry);
 922         req->r_num_caps = 2;
 923         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 924         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 925         if (as_ctx.pagelist) {
 926                 req->r_pagelist = as_ctx.pagelist;
 927                 as_ctx.pagelist = NULL;
 928         }
 929         err = ceph_mdsc_do_request(mdsc, dir, req);
 930         if (!err && !req->r_reply_info.head->is_dentry)
 931                 err = ceph_handle_notrace_create(dir, dentry);
 932         ceph_mdsc_put_request(req);
 933 out:
 934         if (err)
 935                 d_drop(dentry);
 936         ceph_release_acl_sec_ctx(&as_ctx);
 937         return err;
 938 }
 939
 940 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 941 {
 942         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 943         struct ceph_mds_request *req;
 944         struct ceph_acl_sec_ctx as_ctx = {};
 945         int err = -EROFS;
 946         int op;
 947
 948         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 949                 /* mkdir .snap/foo is a MKSNAP */
 950                 op = CEPH_MDS_OP_MKSNAP;
 951                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 952                      dentry, dentry);
 953         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 954                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 955                 op = CEPH_MDS_OP_MKDIR;
 956         } else {
 957                 goto out;
 958         }
 959
 960         if (op == CEPH_MDS_OP_MKDIR &&
 961             ceph_quota_is_max_files_exceeded(dir)) {
 962                 err = -EDQUOT;
 963                 goto out;
 964         }
 965
 966         mode |= S_IFDIR;
 967         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 968         if (err < 0)
 969                 goto out;
 970         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 971         if (err < 0)
 972                 goto out;
 973
 974         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 975         if (IS_ERR(req)) {
 976                 err = PTR_ERR(req);
 977                 goto out;
 978         }
 979
 980         req->r_dentry = dget(dentry);
 981         req->r_num_caps = 2;
 982         req->r_parent = dir;
 983         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 984         req->r_args.mkdir.mode = cpu_to_le32(mode);
 985         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 986         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 987         if (as_ctx.pagelist) {
 988                 req->r_pagelist = as_ctx.pagelist;
 989                 as_ctx.pagelist = NULL;
 990         }
 991         err = ceph_mdsc_do_request(mdsc, dir, req);
 992         if (!err &&
 993             !req->r_reply_info.head->is_target &&
 994             !req->r_reply_info.head->is_dentry)
 995                 err = ceph_handle_notrace_create(dir, dentry);
 996         ceph_mdsc_put_request(req);
 997 out:
 998         if (!err)
 999                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1000         else
1001                 d_drop(dentry);
1002         ceph_release_acl_sec_ctx(&as_ctx);
1003         return err;
1004 }
1005
1006 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1007                      struct dentry *dentry)
1008 {
1009         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1010         struct ceph_mds_request *req;
1011         int err;
1012
1013         if (ceph_snap(dir) != CEPH_NOSNAP)
1014                 return -EROFS;
1015
1016         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1017              old_dentry, dentry);
1018         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1019         if (IS_ERR(req)) {
1020                 d_drop(dentry);
1021                 return PTR_ERR(req);
1022         }
1023         req->r_dentry = dget(dentry);
1024         req->r_num_caps = 2;
1025         req->r_old_dentry = dget(old_dentry);
1026         req->r_parent = dir;
1027         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1028         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1029         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1030         /* release LINK_SHARED on source inode (mds will lock it) */
1031         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1032         err = ceph_mdsc_do_request(mdsc, dir, req);
1033         if (err) {
1034                 d_drop(dentry);
1035         } else if (!req->r_reply_info.head->is_dentry) {
1036                 ihold(d_inode(old_dentry));
1037                 d_instantiate(dentry, d_inode(old_dentry));
1038         }
1039         ceph_mdsc_put_request(req);
1040         return err;
1041 }
1042
1043 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1044                                  struct ceph_mds_request *req)
1045 {
1046         int result = req->r_err ? req->r_err :
1047                         le32_to_cpu(req->r_reply_info.head->result);
1048
1049         if (result == -EJUKEBOX)
1050                 goto out;
1051
1052         /* If op failed, mark everyone involved for errors */
1053         if (result) {
1054                 int pathlen = 0;
1055                 u64 base = 0;
1056                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1057                                                   &base, 0);
1058
1059                 /* mark error on parent + clear complete */
1060                 mapping_set_error(req->r_parent->i_mapping, result);
1061                 ceph_dir_clear_complete(req->r_parent);
1062
1063                 /* drop the dentry -- we don't know its status */
1064                 if (!d_unhashed(req->r_dentry))
1065                         d_drop(req->r_dentry);
1066
1067                 /* mark inode itself for an error (since metadata is bogus) */
1068                 mapping_set_error(req->r_old_inode->i_mapping, result);
1069
1070                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1071                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1072                 ceph_mdsc_free_path(path, pathlen);
1073         }
1074 out:
1075         iput(req->r_old_inode);
1076         ceph_mdsc_release_dir_caps(req);
1077 }
1078
1079 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1080 {
1081         struct ceph_inode_info *ci = ceph_inode(dir);
1082         struct ceph_dentry_info *di;
1083         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1084
1085         spin_lock(&ci->i_ceph_lock);
1086         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1087                 ceph_take_cap_refs(ci, want, false);
1088                 got = want;
1089         }
1090         spin_unlock(&ci->i_ceph_lock);
1091
1092         /* If we didn't get anything, return 0 */
1093         if (!got)
1094                 return 0;
1095
1096         spin_lock(&dentry->d_lock);
1097         di = ceph_dentry(dentry);
1098         /*
1099          * - We are holding Fx, which implies Fs caps.
1100          * - Only support async unlink for primary linkage
1101          */
1102         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1103             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1104                 want = 0;
1105         spin_unlock(&dentry->d_lock);
1106
1107         /* Do we still want what we've got? */
1108         if (want == got)
1109                 return got;
1110
1111         ceph_put_cap_refs(ci, got);
1112         return 0;
1113 }
1114
1115 /*
1116  * rmdir and unlink are differ only by the metadata op code
1117  */
1118 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1119 {
1120         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1121         struct ceph_mds_client *mdsc = fsc->mdsc;
1122         struct inode *inode = d_inode(dentry);
1123         struct ceph_mds_request *req;
1124         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1125         int err = -EROFS;
1126         int op;
1127
1128         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1129                 /* rmdir .snap/foo is RMSNAP */
1130                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1131                 op = CEPH_MDS_OP_RMSNAP;
1132         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1133                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1134                      dir, dentry, inode);
1135                 op = d_is_dir(dentry) ?
1136                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1137         } else
1138                 goto out;
1139 retry:
1140         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1141         if (IS_ERR(req)) {
1142                 err = PTR_ERR(req);
1143                 goto out;
1144         }
1145         req->r_dentry = dget(dentry);
1146         req->r_num_caps = 2;
1147         req->r_parent = dir;
1148         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1149         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1150         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1151
1152         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1153             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1154                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1155                      dentry->d_name.len, dentry->d_name.name,
1156                      ceph_cap_string(req->r_dir_caps));
1157                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1158                 req->r_callback = ceph_async_unlink_cb;
1159                 req->r_old_inode = d_inode(dentry);
1160                 ihold(req->r_old_inode);
1161                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1162                 if (!err) {
1163                         /*
1164                          * We have enough caps, so we assume that the unlink
1165                          * will succeed. Fix up the target inode and dcache.
1166                          */
1167                         drop_nlink(inode);
1168                         d_delete(dentry);
1169                 } else if (err == -EJUKEBOX) {
1170                         try_async = false;
1171                         ceph_mdsc_put_request(req);
1172                         goto retry;
1173                 }
1174         } else {
1175                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1176                 err = ceph_mdsc_do_request(mdsc, dir, req);
1177                 if (!err && !req->r_reply_info.head->is_dentry)
1178                         d_delete(dentry);
1179         }
1180
1181         ceph_mdsc_put_request(req);
1182 out:
1183         return err;
1184 }
1185
1186 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1187                        struct inode *new_dir, struct dentry *new_dentry,
1188                        unsigned int flags)
1189 {
1190         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1191         struct ceph_mds_request *req;
1192         int op = CEPH_MDS_OP_RENAME;
1193         int err;
1194
1195         if (flags)
1196                 return -EINVAL;
1197
1198         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1199                 return -EXDEV;
1200         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1201                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1202                         op = CEPH_MDS_OP_RENAMESNAP;
1203                 else
1204                         return -EROFS;
1205         }
1206         /* don't allow cross-quota renames */
1207         if ((old_dir != new_dir) &&
1208             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1209                 return -EXDEV;
1210
1211         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1212              old_dir, old_dentry, new_dir, new_dentry);
1213         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1214         if (IS_ERR(req))
1215                 return PTR_ERR(req);
1216         ihold(old_dir);
1217         req->r_dentry = dget(new_dentry);
1218         req->r_num_caps = 2;
1219         req->r_old_dentry = dget(old_dentry);
1220         req->r_old_dentry_dir = old_dir;
1221         req->r_parent = new_dir;
1222         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1223         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1224         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1225         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1226         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1227         /* release LINK_RDCACHE on source inode (mds will lock it) */
1228         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1229         if (d_really_is_positive(new_dentry)) {
1230                 req->r_inode_drop =
1231                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1232         }
1233         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1234         if (!err && !req->r_reply_info.head->is_dentry) {
1235                 /*
1236                  * Normally d_move() is done by fill_trace (called by
1237                  * do_request, above).  If there is no trace, we need
1238                  * to do it here.
1239                  */
1240                 d_move(old_dentry, new_dentry);
1241         }
1242         ceph_mdsc_put_request(req);
1243         return err;
1244 }
1245
1246 /*
1247  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1248  * Leases at front of the list will expire first. (Assume all leases have
1249  * similar duration)
1250  *
1251  * Called under dentry->d_lock.
1252  */
1253 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1254 {
1255         struct dentry *dn = di->dentry;
1256         struct ceph_mds_client *mdsc;
1257
1258         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1259
1260         di->flags |= CEPH_DENTRY_LEASE_LIST;
1261         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1262                 di->flags |= CEPH_DENTRY_REFERENCED;
1263                 return;
1264         }
1265
1266         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1267         spin_lock(&mdsc->dentry_list_lock);
1268         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1269         spin_unlock(&mdsc->dentry_list_lock);
1270 }
1271
1272 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1273                                      struct ceph_dentry_info *di)
1274 {
1275         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1276         di->lease_gen = 0;
1277         di->time = jiffies;
1278         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1279 }
1280
1281 /*
1282  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1283  * list if it's not in the list, otherwise set 'referenced' flag.
1284  *
1285  * Called under dentry->d_lock.
1286  */
1287 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1288 {
1289         struct dentry *dn = di->dentry;
1290         struct ceph_mds_client *mdsc;
1291
1292         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1293              di, dn, dn, di->offset);
1294
1295         if (!list_empty(&di->lease_list)) {
1296                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1297                         /* don't remove dentry from dentry lease list
1298                          * if its lease is valid */
1299                         if (__dentry_lease_is_valid(di))
1300                                 return;
1301                 } else {
1302                         di->flags |= CEPH_DENTRY_REFERENCED;
1303                         return;
1304                 }
1305         }
1306
1307         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1308                 di->flags |= CEPH_DENTRY_REFERENCED;
1309                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1310                 return;
1311         }
1312
1313         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1314         spin_lock(&mdsc->dentry_list_lock);
1315         __dentry_dir_lease_touch(mdsc, di),
1316         spin_unlock(&mdsc->dentry_list_lock);
1317 }
1318
1319 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1320 {
1321         struct ceph_mds_client *mdsc;
1322         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1323                 return;
1324         if (list_empty(&di->lease_list))
1325                 return;
1326
1327         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1328         spin_lock(&mdsc->dentry_list_lock);
1329         list_del_init(&di->lease_list);
1330         spin_unlock(&mdsc->dentry_list_lock);
1331 }
1332
1333 enum {
1334         KEEP    = 0,
1335         DELETE  = 1,
1336         TOUCH   = 2,
1337         STOP    = 4,
1338 };
1339
1340 struct ceph_lease_walk_control {
1341         bool dir_lease;
1342         bool expire_dir_lease;
1343         unsigned long nr_to_scan;
1344         unsigned long dir_lease_ttl;
1345 };
1346
1347 static unsigned long
1348 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1349                      struct ceph_lease_walk_control *lwc,
1350                      int (*check)(struct dentry*, void*))
1351 {
1352         struct ceph_dentry_info *di, *tmp;
1353         struct dentry *dentry, *last = NULL;
1354         struct list_head* list;
1355         LIST_HEAD(dispose);
1356         unsigned long freed = 0;
1357         int ret = 0;
1358
1359         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1360         spin_lock(&mdsc->dentry_list_lock);
1361         list_for_each_entry_safe(di, tmp, list, lease_list) {
1362                 if (!lwc->nr_to_scan)
1363                         break;
1364                 --lwc->nr_to_scan;
1365
1366                 dentry = di->dentry;
1367                 if (last == dentry)
1368                         break;
1369
1370                 if (!spin_trylock(&dentry->d_lock))
1371                         continue;
1372
1373                 if (__lockref_is_dead(&dentry->d_lockref)) {
1374                         list_del_init(&di->lease_list);
1375                         goto next;
1376                 }
1377
1378                 ret = check(dentry, lwc);
1379                 if (ret & TOUCH) {
1380                         /* move it into tail of dir lease list */
1381                         __dentry_dir_lease_touch(mdsc, di);
1382                         if (!last)
1383                                 last = dentry;
1384                 }
1385                 if (ret & DELETE) {
1386                         /* stale lease */
1387                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1388                         if (dentry->d_lockref.count > 0) {
1389                                 /* update_dentry_lease() will re-add
1390                                  * it to lease list, or
1391                                  * ceph_d_delete() will return 1 when
1392                                  * last reference is dropped */
1393                                 list_del_init(&di->lease_list);
1394                         } else {
1395                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1396                                 list_move_tail(&di->lease_list, &dispose);
1397                                 dget_dlock(dentry);
1398                         }
1399                 }
1400 next:
1401                 spin_unlock(&dentry->d_lock);
1402                 if (ret & STOP)
1403                         break;
1404         }
1405         spin_unlock(&mdsc->dentry_list_lock);
1406
1407         while (!list_empty(&dispose)) {
1408                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1409                                       lease_list);
1410                 dentry = di->dentry;
1411                 spin_lock(&dentry->d_lock);
1412
1413                 list_del_init(&di->lease_list);
1414                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1415                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1416                         spin_lock(&mdsc->dentry_list_lock);
1417                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1418                                 list_add_tail(&di->lease_list,
1419                                               &mdsc->dentry_leases);
1420                         } else {
1421                                 __dentry_dir_lease_touch(mdsc, di);
1422                         }
1423                         spin_unlock(&mdsc->dentry_list_lock);
1424                 } else {
1425                         freed++;
1426                 }
1427
1428                 spin_unlock(&dentry->d_lock);
1429                 /* ceph_d_delete() does the trick */
1430                 dput(dentry);
1431         }
1432         return freed;
1433 }
1434
1435 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1436 {
1437         struct ceph_dentry_info *di = ceph_dentry(dentry);
1438         int ret;
1439
1440         if (__dentry_lease_is_valid(di))
1441                 return STOP;
1442         ret = __dir_lease_try_check(dentry);
1443         if (ret == -EBUSY)
1444                 return KEEP;
1445         if (ret > 0)
1446                 return TOUCH;
1447         return DELETE;
1448 }
1449
1450 static int __dir_lease_check(struct dentry *dentry, void *arg)
1451 {
1452         struct ceph_lease_walk_control *lwc = arg;
1453         struct ceph_dentry_info *di = ceph_dentry(dentry);
1454
1455         int ret = __dir_lease_try_check(dentry);
1456         if (ret == -EBUSY)
1457                 return KEEP;
1458         if (ret > 0) {
1459                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1460                         return STOP;
1461                 /* Move dentry to tail of dir lease list if we don't want
1462                  * to delete it. So dentries in the list are checked in a
1463                  * round robin manner */
1464                 if (!lwc->expire_dir_lease)
1465                         return TOUCH;
1466                 if (dentry->d_lockref.count > 0 ||
1467                     (di->flags & CEPH_DENTRY_REFERENCED))
1468                         return TOUCH;
1469                 /* invalidate dir lease */
1470                 di->lease_shared_gen = 0;
1471         }
1472         return DELETE;
1473 }
1474
1475 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1476 {
1477         struct ceph_lease_walk_control lwc;
1478         unsigned long count;
1479         unsigned long freed;
1480
1481         spin_lock(&mdsc->caps_list_lock);
1482         if (mdsc->caps_use_max > 0 &&
1483             mdsc->caps_use_count > mdsc->caps_use_max)
1484                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1485         else
1486                 count = 0;
1487         spin_unlock(&mdsc->caps_list_lock);
1488
1489         lwc.dir_lease = false;
1490         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1491         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1492         if (!lwc.nr_to_scan) /* more invalid leases */
1493                 return -EAGAIN;
1494
1495         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1496                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1497
1498         lwc.dir_lease = true;
1499         lwc.expire_dir_lease = freed < count;
1500         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1501         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1502         if (!lwc.nr_to_scan) /* more to check */
1503                 return -EAGAIN;
1504
1505         return freed > 0 ? 1 : 0;
1506 }
1507
1508 /*
1509  * Ensure a dentry lease will no longer revalidate.
1510  */
1511 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1512 {
1513         struct ceph_dentry_info *di = ceph_dentry(dentry);
1514         spin_lock(&dentry->d_lock);
1515         di->time = jiffies;
1516         di->lease_shared_gen = 0;
1517         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1518         __dentry_lease_unlist(di);
1519         spin_unlock(&dentry->d_lock);
1520 }
1521
1522 /*
1523  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1524  * renew if the least is more than half up.
1525  */
1526 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1527 {
1528         struct ceph_mds_session *session;
1529
1530         if (!di->lease_gen)
1531                 return false;
1532
1533         session = di->lease_session;
1534         if (session) {
1535                 u32 gen;
1536                 unsigned long ttl;
1537
1538                 spin_lock(&session->s_gen_ttl_lock);
1539                 gen = session->s_cap_gen;
1540                 ttl = session->s_cap_ttl;
1541                 spin_unlock(&session->s_gen_ttl_lock);
1542
1543                 if (di->lease_gen == gen &&
1544                     time_before(jiffies, ttl) &&
1545                     time_before(jiffies, di->time))
1546                         return true;
1547         }
1548         di->lease_gen = 0;
1549         return false;
1550 }
1551
1552 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1553 {
1554         struct ceph_dentry_info *di;
1555         struct ceph_mds_session *session = NULL;
1556         u32 seq = 0;
1557         int valid = 0;
1558
1559         spin_lock(&dentry->d_lock);
1560         di = ceph_dentry(dentry);
1561         if (di && __dentry_lease_is_valid(di)) {
1562                 valid = 1;
1563
1564                 if (di->lease_renew_after &&
1565                     time_after(jiffies, di->lease_renew_after)) {
1566                         /*
1567                          * We should renew. If we're in RCU walk mode
1568                          * though, we can't do that so just return
1569                          * -ECHILD.
1570                          */
1571                         if (flags & LOOKUP_RCU) {
1572                                 valid = -ECHILD;
1573                         } else {
1574                                 session = ceph_get_mds_session(di->lease_session);
1575                                 seq = di->lease_seq;
1576                                 di->lease_renew_after = 0;
1577                                 di->lease_renew_from = jiffies;
1578                         }
1579                 }
1580         }
1581         spin_unlock(&dentry->d_lock);
1582
1583         if (session) {
1584                 ceph_mdsc_lease_send_msg(session, dentry,
1585                                          CEPH_MDS_LEASE_RENEW, seq);
1586                 ceph_put_mds_session(session);
1587         }
1588         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1589         return valid;
1590 }
1591
1592 /*
1593  * Called under dentry->d_lock.
1594  */
1595 static int __dir_lease_try_check(const struct dentry *dentry)
1596 {
1597         struct ceph_dentry_info *di = ceph_dentry(dentry);
1598         struct inode *dir;
1599         struct ceph_inode_info *ci;
1600         int valid = 0;
1601
1602         if (!di->lease_shared_gen)
1603                 return 0;
1604         if (IS_ROOT(dentry))
1605                 return 0;
1606
1607         dir = d_inode(dentry->d_parent);
1608         ci = ceph_inode(dir);
1609
1610         if (spin_trylock(&ci->i_ceph_lock)) {
1611                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1612                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1613                         valid = 1;
1614                 spin_unlock(&ci->i_ceph_lock);
1615         } else {
1616                 valid = -EBUSY;
1617         }
1618
1619         if (!valid)
1620                 di->lease_shared_gen = 0;
1621         return valid;
1622 }
1623
1624 /*
1625  * Check if directory-wide content lease/cap is valid.
1626  */
1627 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1628                               struct ceph_mds_client *mdsc)
1629 {
1630         struct ceph_inode_info *ci = ceph_inode(dir);
1631         int valid;
1632         int shared_gen;
1633
1634         spin_lock(&ci->i_ceph_lock);
1635         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1636         if (valid) {
1637                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1638                 shared_gen = atomic_read(&ci->i_shared_gen);
1639         }
1640         spin_unlock(&ci->i_ceph_lock);
1641         if (valid) {
1642                 struct ceph_dentry_info *di;
1643                 spin_lock(&dentry->d_lock);
1644                 di = ceph_dentry(dentry);
1645                 if (dir == d_inode(dentry->d_parent) &&
1646                     di && di->lease_shared_gen == shared_gen)
1647                         __ceph_dentry_dir_lease_touch(di);
1648                 else
1649                         valid = 0;
1650                 spin_unlock(&dentry->d_lock);
1651         }
1652         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1653              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1654         return valid;
1655 }
1656
1657 /*
1658  * Check if cached dentry can be trusted.
1659  */
1660 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1661 {
1662         int valid = 0;
1663         struct dentry *parent;
1664         struct inode *dir, *inode;
1665         struct ceph_mds_client *mdsc;
1666
1667         if (flags & LOOKUP_RCU) {
1668                 parent = READ_ONCE(dentry->d_parent);
1669                 dir = d_inode_rcu(parent);
1670                 if (!dir)
1671                         return -ECHILD;
1672                 inode = d_inode_rcu(dentry);
1673         } else {
1674                 parent = dget_parent(dentry);
1675                 dir = d_inode(parent);
1676                 inode = d_inode(dentry);
1677         }
1678
1679         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1680              dentry, inode, ceph_dentry(dentry)->offset);
1681
1682         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1683
1684         /* always trust cached snapped dentries, snapdir dentry */
1685         if (ceph_snap(dir) != CEPH_NOSNAP) {
1686                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1687                      dentry, inode);
1688                 valid = 1;
1689         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1690                 valid = 1;
1691         } else {
1692                 valid = dentry_lease_is_valid(dentry, flags);
1693                 if (valid == -ECHILD)
1694                         return valid;
1695                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1696                         if (inode)
1697                                 valid = ceph_is_any_caps(inode);
1698                         else
1699                                 valid = 1;
1700                 }
1701         }
1702
1703         if (!valid) {
1704                 struct ceph_mds_request *req;
1705                 int op, err;
1706                 u32 mask;
1707
1708                 if (flags & LOOKUP_RCU)
1709                         return -ECHILD;
1710
1711                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1712
1713                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1714                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1715                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1716                 if (!IS_ERR(req)) {
1717                         req->r_dentry = dget(dentry);
1718                         req->r_num_caps = 2;
1719                         req->r_parent = dir;
1720
1721                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1722                         if (ceph_security_xattr_wanted(dir))
1723                                 mask |= CEPH_CAP_XATTR_SHARED;
1724                         req->r_args.getattr.mask = cpu_to_le32(mask);
1725
1726                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1727                         switch (err) {
1728                         case 0:
1729                                 if (d_really_is_positive(dentry) &&
1730                                     d_inode(dentry) == req->r_target_inode)
1731                                         valid = 1;
1732                                 break;
1733                         case -ENOENT:
1734                                 if (d_really_is_negative(dentry))
1735                                         valid = 1;
1736                                 fallthrough;
1737                         default:
1738                                 break;
1739                         }
1740                         ceph_mdsc_put_request(req);
1741                         dout("d_revalidate %p lookup result=%d\n",
1742                              dentry, err);
1743                 }
1744         } else {
1745                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1746         }
1747
1748         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1749         if (!valid)
1750                 ceph_dir_clear_complete(dir);
1751
1752         if (!(flags & LOOKUP_RCU))
1753                 dput(parent);
1754         return valid;
1755 }
1756
1757 /*
1758  * Delete unused dentry that doesn't have valid lease
1759  *
1760  * Called under dentry->d_lock.
1761  */
1762 static int ceph_d_delete(const struct dentry *dentry)
1763 {
1764         struct ceph_dentry_info *di;
1765
1766         /* won't release caps */
1767         if (d_really_is_negative(dentry))
1768                 return 0;
1769         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1770                 return 0;
1771         /* vaild lease? */
1772         di = ceph_dentry(dentry);
1773         if (di) {
1774                 if (__dentry_lease_is_valid(di))
1775                         return 0;
1776                 if (__dir_lease_try_check(dentry))
1777                         return 0;
1778         }
1779         return 1;
1780 }
1781
1782 /*
1783  * Release our ceph_dentry_info.
1784  */
1785 static void ceph_d_release(struct dentry *dentry)
1786 {
1787         struct ceph_dentry_info *di = ceph_dentry(dentry);
1788         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1789
1790         dout("d_release %p\n", dentry);
1791
1792         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1793
1794         spin_lock(&dentry->d_lock);
1795         __dentry_lease_unlist(di);
1796         dentry->d_fsdata = NULL;
1797         spin_unlock(&dentry->d_lock);
1798
1799         if (di->lease_session)
1800                 ceph_put_mds_session(di->lease_session);
1801         kmem_cache_free(ceph_dentry_cachep, di);
1802 }
1803
1804 /*
1805  * When the VFS prunes a dentry from the cache, we need to clear the
1806  * complete flag on the parent directory.
1807  *
1808  * Called under dentry->d_lock.
1809  */
1810 static void ceph_d_prune(struct dentry *dentry)
1811 {
1812         struct ceph_inode_info *dir_ci;
1813         struct ceph_dentry_info *di;
1814
1815         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1816
1817         /* do we have a valid parent? */
1818         if (IS_ROOT(dentry))
1819                 return;
1820
1821         /* we hold d_lock, so d_parent is stable */
1822         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1823         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1824                 return;
1825
1826         /* who calls d_delete() should also disable dcache readdir */
1827         if (d_really_is_negative(dentry))
1828                 return;
1829
1830         /* d_fsdata does not get cleared until d_release */
1831         if (!d_unhashed(dentry)) {
1832                 __ceph_dir_clear_complete(dir_ci);
1833                 return;
1834         }
1835
1836         /* Disable dcache readdir just in case that someone called d_drop()
1837          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1838          * properly (dcache readdir is still enabled) */
1839         di = ceph_dentry(dentry);
1840         if (di->offset > 0 &&
1841             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1842                 __ceph_dir_clear_ordered(dir_ci);
1843 }
1844
1845 /*
1846  * read() on a dir.  This weird interface hack only works if mounted
1847  * with '-o dirstat'.
1848  */
1849 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1850                              loff_t *ppos)
1851 {
1852         struct ceph_dir_file_info *dfi = file->private_data;
1853         struct inode *inode = file_inode(file);
1854         struct ceph_inode_info *ci = ceph_inode(inode);
1855         int left;
1856         const int bufsize = 1024;
1857
1858         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1859                 return -EISDIR;
1860
1861         if (!dfi->dir_info) {
1862                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1863                 if (!dfi->dir_info)
1864                         return -ENOMEM;
1865                 dfi->dir_info_len =
1866                         snprintf(dfi->dir_info, bufsize,
1867                                 "entries:   %20lld\n"
1868                                 " files:    %20lld\n"
1869                                 " subdirs:  %20lld\n"
1870                                 "rentries:  %20lld\n"
1871                                 " rfiles:   %20lld\n"
1872                                 " rsubdirs: %20lld\n"
1873                                 "rbytes:    %20lld\n"
1874                                 "rctime:    %10lld.%09ld\n",
1875                                 ci->i_files + ci->i_subdirs,
1876                                 ci->i_files,
1877                                 ci->i_subdirs,
1878                                 ci->i_rfiles + ci->i_rsubdirs,
1879                                 ci->i_rfiles,
1880                                 ci->i_rsubdirs,
1881                                 ci->i_rbytes,
1882                                 ci->i_rctime.tv_sec,
1883                                 ci->i_rctime.tv_nsec);
1884         }
1885
1886         if (*ppos >= dfi->dir_info_len)
1887                 return 0;
1888         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1889         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1890         if (left == size)
1891                 return -EFAULT;
1892         *ppos += (size - left);
1893         return size - left;
1894 }
1895
1896
1897
1898 /*
1899  * Return name hash for a given dentry.  This is dependent on
1900  * the parent directory's hash function.
1901  */
1902 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1903 {
1904         struct ceph_inode_info *dci = ceph_inode(dir);
1905         unsigned hash;
1906
1907         switch (dci->i_dir_layout.dl_dir_hash) {
1908         case 0: /* for backward compat */
1909         case CEPH_STR_HASH_LINUX:
1910                 return dn->d_name.hash;
1911
1912         default:
1913                 spin_lock(&dn->d_lock);
1914                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1915                                      dn->d_name.name, dn->d_name.len);
1916                 spin_unlock(&dn->d_lock);
1917                 return hash;
1918         }
1919 }
1920
1921 const struct file_operations ceph_dir_fops = {
1922         .read = ceph_read_dir,
1923         .iterate = ceph_readdir,
1924         .llseek = ceph_dir_llseek,
1925         .open = ceph_open,
1926         .release = ceph_release,
1927         .unlocked_ioctl = ceph_ioctl,
1928         .compat_ioctl = compat_ptr_ioctl,
1929         .fsync = ceph_fsync,
1930         .lock = ceph_lock,
1931         .flock = ceph_flock,
1932 };
1933
1934 const struct file_operations ceph_snapdir_fops = {
1935         .iterate = ceph_readdir,
1936         .llseek = ceph_dir_llseek,
1937         .open = ceph_open,
1938         .release = ceph_release,
1939 };
1940
1941 const struct inode_operations ceph_dir_iops = {
1942         .lookup = ceph_lookup,
1943         .permission = ceph_permission,
1944         .getattr = ceph_getattr,
1945         .setattr = ceph_setattr,
1946         .listxattr = ceph_listxattr,
1947         .get_acl = ceph_get_acl,
1948         .set_acl = ceph_set_acl,
1949         .mknod = ceph_mknod,
1950         .symlink = ceph_symlink,
1951         .mkdir = ceph_mkdir,
1952         .link = ceph_link,
1953         .unlink = ceph_unlink,
1954         .rmdir = ceph_unlink,
1955         .rename = ceph_rename,
1956         .create = ceph_create,
1957         .atomic_open = ceph_atomic_open,
1958 };
1959
1960 const struct inode_operations ceph_snapdir_iops = {
1961         .lookup = ceph_lookup,
1962         .permission = ceph_permission,
1963         .getattr = ceph_getattr,
1964         .mkdir = ceph_mkdir,
1965         .rmdir = ceph_unlink,
1966         .rename = ceph_rename,
1967 };
1968
1969 const struct dentry_operations ceph_dentry_ops = {
1970         .d_revalidate = ceph_d_revalidate,
1971         .d_delete = ceph_d_delete,
1972         .d_release = ceph_d_release,
1973         .d_prune = ceph_d_prune,
1974         .d_init = ceph_d_init,
1975 };