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         } else if (old_dir != new_dir) {
1206                 err = ceph_quota_check_rename(mdsc, d_inode(old_dentry),
1207                                               new_dir);
1208                 if (err)
1209                         return err;
1210         }
1211
1212         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1213              old_dir, old_dentry, new_dir, new_dentry);
1214         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1215         if (IS_ERR(req))
1216                 return PTR_ERR(req);
1217         ihold(old_dir);
1218         req->r_dentry = dget(new_dentry);
1219         req->r_num_caps = 2;
1220         req->r_old_dentry = dget(old_dentry);
1221         req->r_old_dentry_dir = old_dir;
1222         req->r_parent = new_dir;
1223         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1224         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1225         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1226         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1227         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1228         /* release LINK_RDCACHE on source inode (mds will lock it) */
1229         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1230         if (d_really_is_positive(new_dentry)) {
1231                 req->r_inode_drop =
1232                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1233         }
1234         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1235         if (!err && !req->r_reply_info.head->is_dentry) {
1236                 /*
1237                  * Normally d_move() is done by fill_trace (called by
1238                  * do_request, above).  If there is no trace, we need
1239                  * to do it here.
1240                  */
1241                 d_move(old_dentry, new_dentry);
1242         }
1243         ceph_mdsc_put_request(req);
1244         return err;
1245 }
1246
1247 /*
1248  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1249  * Leases at front of the list will expire first. (Assume all leases have
1250  * similar duration)
1251  *
1252  * Called under dentry->d_lock.
1253  */
1254 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1255 {
1256         struct dentry *dn = di->dentry;
1257         struct ceph_mds_client *mdsc;
1258
1259         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1260
1261         di->flags |= CEPH_DENTRY_LEASE_LIST;
1262         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1263                 di->flags |= CEPH_DENTRY_REFERENCED;
1264                 return;
1265         }
1266
1267         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1268         spin_lock(&mdsc->dentry_list_lock);
1269         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1270         spin_unlock(&mdsc->dentry_list_lock);
1271 }
1272
1273 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1274                                      struct ceph_dentry_info *di)
1275 {
1276         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1277         di->lease_gen = 0;
1278         di->time = jiffies;
1279         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1280 }
1281
1282 /*
1283  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1284  * list if it's not in the list, otherwise set 'referenced' flag.
1285  *
1286  * Called under dentry->d_lock.
1287  */
1288 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1289 {
1290         struct dentry *dn = di->dentry;
1291         struct ceph_mds_client *mdsc;
1292
1293         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1294              di, dn, dn, di->offset);
1295
1296         if (!list_empty(&di->lease_list)) {
1297                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1298                         /* don't remove dentry from dentry lease list
1299                          * if its lease is valid */
1300                         if (__dentry_lease_is_valid(di))
1301                                 return;
1302                 } else {
1303                         di->flags |= CEPH_DENTRY_REFERENCED;
1304                         return;
1305                 }
1306         }
1307
1308         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1309                 di->flags |= CEPH_DENTRY_REFERENCED;
1310                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1311                 return;
1312         }
1313
1314         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1315         spin_lock(&mdsc->dentry_list_lock);
1316         __dentry_dir_lease_touch(mdsc, di),
1317         spin_unlock(&mdsc->dentry_list_lock);
1318 }
1319
1320 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1321 {
1322         struct ceph_mds_client *mdsc;
1323         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1324                 return;
1325         if (list_empty(&di->lease_list))
1326                 return;
1327
1328         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1329         spin_lock(&mdsc->dentry_list_lock);
1330         list_del_init(&di->lease_list);
1331         spin_unlock(&mdsc->dentry_list_lock);
1332 }
1333
1334 enum {
1335         KEEP    = 0,
1336         DELETE  = 1,
1337         TOUCH   = 2,
1338         STOP    = 4,
1339 };
1340
1341 struct ceph_lease_walk_control {
1342         bool dir_lease;
1343         bool expire_dir_lease;
1344         unsigned long nr_to_scan;
1345         unsigned long dir_lease_ttl;
1346 };
1347
1348 static unsigned long
1349 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1350                      struct ceph_lease_walk_control *lwc,
1351                      int (*check)(struct dentry*, void*))
1352 {
1353         struct ceph_dentry_info *di, *tmp;
1354         struct dentry *dentry, *last = NULL;
1355         struct list_head* list;
1356         LIST_HEAD(dispose);
1357         unsigned long freed = 0;
1358         int ret = 0;
1359
1360         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1361         spin_lock(&mdsc->dentry_list_lock);
1362         list_for_each_entry_safe(di, tmp, list, lease_list) {
1363                 if (!lwc->nr_to_scan)
1364                         break;
1365                 --lwc->nr_to_scan;
1366
1367                 dentry = di->dentry;
1368                 if (last == dentry)
1369                         break;
1370
1371                 if (!spin_trylock(&dentry->d_lock))
1372                         continue;
1373
1374                 if (__lockref_is_dead(&dentry->d_lockref)) {
1375                         list_del_init(&di->lease_list);
1376                         goto next;
1377                 }
1378
1379                 ret = check(dentry, lwc);
1380                 if (ret & TOUCH) {
1381                         /* move it into tail of dir lease list */
1382                         __dentry_dir_lease_touch(mdsc, di);
1383                         if (!last)
1384                                 last = dentry;
1385                 }
1386                 if (ret & DELETE) {
1387                         /* stale lease */
1388                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1389                         if (dentry->d_lockref.count > 0) {
1390                                 /* update_dentry_lease() will re-add
1391                                  * it to lease list, or
1392                                  * ceph_d_delete() will return 1 when
1393                                  * last reference is dropped */
1394                                 list_del_init(&di->lease_list);
1395                         } else {
1396                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1397                                 list_move_tail(&di->lease_list, &dispose);
1398                                 dget_dlock(dentry);
1399                         }
1400                 }
1401 next:
1402                 spin_unlock(&dentry->d_lock);
1403                 if (ret & STOP)
1404                         break;
1405         }
1406         spin_unlock(&mdsc->dentry_list_lock);
1407
1408         while (!list_empty(&dispose)) {
1409                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1410                                       lease_list);
1411                 dentry = di->dentry;
1412                 spin_lock(&dentry->d_lock);
1413
1414                 list_del_init(&di->lease_list);
1415                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1416                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1417                         spin_lock(&mdsc->dentry_list_lock);
1418                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1419                                 list_add_tail(&di->lease_list,
1420                                               &mdsc->dentry_leases);
1421                         } else {
1422                                 __dentry_dir_lease_touch(mdsc, di);
1423                         }
1424                         spin_unlock(&mdsc->dentry_list_lock);
1425                 } else {
1426                         freed++;
1427                 }
1428
1429                 spin_unlock(&dentry->d_lock);
1430                 /* ceph_d_delete() does the trick */
1431                 dput(dentry);
1432         }
1433         return freed;
1434 }
1435
1436 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1437 {
1438         struct ceph_dentry_info *di = ceph_dentry(dentry);
1439         int ret;
1440
1441         if (__dentry_lease_is_valid(di))
1442                 return STOP;
1443         ret = __dir_lease_try_check(dentry);
1444         if (ret == -EBUSY)
1445                 return KEEP;
1446         if (ret > 0)
1447                 return TOUCH;
1448         return DELETE;
1449 }
1450
1451 static int __dir_lease_check(struct dentry *dentry, void *arg)
1452 {
1453         struct ceph_lease_walk_control *lwc = arg;
1454         struct ceph_dentry_info *di = ceph_dentry(dentry);
1455
1456         int ret = __dir_lease_try_check(dentry);
1457         if (ret == -EBUSY)
1458                 return KEEP;
1459         if (ret > 0) {
1460                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1461                         return STOP;
1462                 /* Move dentry to tail of dir lease list if we don't want
1463                  * to delete it. So dentries in the list are checked in a
1464                  * round robin manner */
1465                 if (!lwc->expire_dir_lease)
1466                         return TOUCH;
1467                 if (dentry->d_lockref.count > 0 ||
1468                     (di->flags & CEPH_DENTRY_REFERENCED))
1469                         return TOUCH;
1470                 /* invalidate dir lease */
1471                 di->lease_shared_gen = 0;
1472         }
1473         return DELETE;
1474 }
1475
1476 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1477 {
1478         struct ceph_lease_walk_control lwc;
1479         unsigned long count;
1480         unsigned long freed;
1481
1482         spin_lock(&mdsc->caps_list_lock);
1483         if (mdsc->caps_use_max > 0 &&
1484             mdsc->caps_use_count > mdsc->caps_use_max)
1485                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1486         else
1487                 count = 0;
1488         spin_unlock(&mdsc->caps_list_lock);
1489
1490         lwc.dir_lease = false;
1491         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1492         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1493         if (!lwc.nr_to_scan) /* more invalid leases */
1494                 return -EAGAIN;
1495
1496         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1497                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1498
1499         lwc.dir_lease = true;
1500         lwc.expire_dir_lease = freed < count;
1501         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1502         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1503         if (!lwc.nr_to_scan) /* more to check */
1504                 return -EAGAIN;
1505
1506         return freed > 0 ? 1 : 0;
1507 }
1508
1509 /*
1510  * Ensure a dentry lease will no longer revalidate.
1511  */
1512 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1513 {
1514         struct ceph_dentry_info *di = ceph_dentry(dentry);
1515         spin_lock(&dentry->d_lock);
1516         di->time = jiffies;
1517         di->lease_shared_gen = 0;
1518         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1519         __dentry_lease_unlist(di);
1520         spin_unlock(&dentry->d_lock);
1521 }
1522
1523 /*
1524  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1525  * renew if the least is more than half up.
1526  */
1527 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1528 {
1529         struct ceph_mds_session *session;
1530
1531         if (!di->lease_gen)
1532                 return false;
1533
1534         session = di->lease_session;
1535         if (session) {
1536                 u32 gen;
1537                 unsigned long ttl;
1538
1539                 spin_lock(&session->s_gen_ttl_lock);
1540                 gen = session->s_cap_gen;
1541                 ttl = session->s_cap_ttl;
1542                 spin_unlock(&session->s_gen_ttl_lock);
1543
1544                 if (di->lease_gen == gen &&
1545                     time_before(jiffies, ttl) &&
1546                     time_before(jiffies, di->time))
1547                         return true;
1548         }
1549         di->lease_gen = 0;
1550         return false;
1551 }
1552
1553 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1554 {
1555         struct ceph_dentry_info *di;
1556         struct ceph_mds_session *session = NULL;
1557         u32 seq = 0;
1558         int valid = 0;
1559
1560         spin_lock(&dentry->d_lock);
1561         di = ceph_dentry(dentry);
1562         if (di && __dentry_lease_is_valid(di)) {
1563                 valid = 1;
1564
1565                 if (di->lease_renew_after &&
1566                     time_after(jiffies, di->lease_renew_after)) {
1567                         /*
1568                          * We should renew. If we're in RCU walk mode
1569                          * though, we can't do that so just return
1570                          * -ECHILD.
1571                          */
1572                         if (flags & LOOKUP_RCU) {
1573                                 valid = -ECHILD;
1574                         } else {
1575                                 session = ceph_get_mds_session(di->lease_session);
1576                                 seq = di->lease_seq;
1577                                 di->lease_renew_after = 0;
1578                                 di->lease_renew_from = jiffies;
1579                         }
1580                 }
1581         }
1582         spin_unlock(&dentry->d_lock);
1583
1584         if (session) {
1585                 ceph_mdsc_lease_send_msg(session, dentry,
1586                                          CEPH_MDS_LEASE_RENEW, seq);
1587                 ceph_put_mds_session(session);
1588         }
1589         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1590         return valid;
1591 }
1592
1593 /*
1594  * Called under dentry->d_lock.
1595  */
1596 static int __dir_lease_try_check(const struct dentry *dentry)
1597 {
1598         struct ceph_dentry_info *di = ceph_dentry(dentry);
1599         struct inode *dir;
1600         struct ceph_inode_info *ci;
1601         int valid = 0;
1602
1603         if (!di->lease_shared_gen)
1604                 return 0;
1605         if (IS_ROOT(dentry))
1606                 return 0;
1607
1608         dir = d_inode(dentry->d_parent);
1609         ci = ceph_inode(dir);
1610
1611         if (spin_trylock(&ci->i_ceph_lock)) {
1612                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1613                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1614                         valid = 1;
1615                 spin_unlock(&ci->i_ceph_lock);
1616         } else {
1617                 valid = -EBUSY;
1618         }
1619
1620         if (!valid)
1621                 di->lease_shared_gen = 0;
1622         return valid;
1623 }
1624
1625 /*
1626  * Check if directory-wide content lease/cap is valid.
1627  */
1628 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1629                               struct ceph_mds_client *mdsc)
1630 {
1631         struct ceph_inode_info *ci = ceph_inode(dir);
1632         int valid;
1633         int shared_gen;
1634
1635         spin_lock(&ci->i_ceph_lock);
1636         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1637         if (valid) {
1638                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1639                 shared_gen = atomic_read(&ci->i_shared_gen);
1640         }
1641         spin_unlock(&ci->i_ceph_lock);
1642         if (valid) {
1643                 struct ceph_dentry_info *di;
1644                 spin_lock(&dentry->d_lock);
1645                 di = ceph_dentry(dentry);
1646                 if (dir == d_inode(dentry->d_parent) &&
1647                     di && di->lease_shared_gen == shared_gen)
1648                         __ceph_dentry_dir_lease_touch(di);
1649                 else
1650                         valid = 0;
1651                 spin_unlock(&dentry->d_lock);
1652         }
1653         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1654              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1655         return valid;
1656 }
1657
1658 /*
1659  * Check if cached dentry can be trusted.
1660  */
1661 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1662 {
1663         int valid = 0;
1664         struct dentry *parent;
1665         struct inode *dir, *inode;
1666         struct ceph_mds_client *mdsc;
1667
1668         if (flags & LOOKUP_RCU) {
1669                 parent = READ_ONCE(dentry->d_parent);
1670                 dir = d_inode_rcu(parent);
1671                 if (!dir)
1672                         return -ECHILD;
1673                 inode = d_inode_rcu(dentry);
1674         } else {
1675                 parent = dget_parent(dentry);
1676                 dir = d_inode(parent);
1677                 inode = d_inode(dentry);
1678         }
1679
1680         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1681              dentry, inode, ceph_dentry(dentry)->offset);
1682
1683         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1684
1685         /* always trust cached snapped dentries, snapdir dentry */
1686         if (ceph_snap(dir) != CEPH_NOSNAP) {
1687                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1688                      dentry, inode);
1689                 valid = 1;
1690         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1691                 valid = 1;
1692         } else {
1693                 valid = dentry_lease_is_valid(dentry, flags);
1694                 if (valid == -ECHILD)
1695                         return valid;
1696                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1697                         if (inode)
1698                                 valid = ceph_is_any_caps(inode);
1699                         else
1700                                 valid = 1;
1701                 }
1702         }
1703
1704         if (!valid) {
1705                 struct ceph_mds_request *req;
1706                 int op, err;
1707                 u32 mask;
1708
1709                 if (flags & LOOKUP_RCU)
1710                         return -ECHILD;
1711
1712                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1713
1714                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1715                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1716                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1717                 if (!IS_ERR(req)) {
1718                         req->r_dentry = dget(dentry);
1719                         req->r_num_caps = 2;
1720                         req->r_parent = dir;
1721
1722                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1723                         if (ceph_security_xattr_wanted(dir))
1724                                 mask |= CEPH_CAP_XATTR_SHARED;
1725                         req->r_args.getattr.mask = cpu_to_le32(mask);
1726
1727                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1728                         switch (err) {
1729                         case 0:
1730                                 if (d_really_is_positive(dentry) &&
1731                                     d_inode(dentry) == req->r_target_inode)
1732                                         valid = 1;
1733                                 break;
1734                         case -ENOENT:
1735                                 if (d_really_is_negative(dentry))
1736                                         valid = 1;
1737                                 fallthrough;
1738                         default:
1739                                 break;
1740                         }
1741                         ceph_mdsc_put_request(req);
1742                         dout("d_revalidate %p lookup result=%d\n",
1743                              dentry, err);
1744                 }
1745         } else {
1746                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1747         }
1748
1749         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1750         if (!valid)
1751                 ceph_dir_clear_complete(dir);
1752
1753         if (!(flags & LOOKUP_RCU))
1754                 dput(parent);
1755         return valid;
1756 }
1757
1758 /*
1759  * Delete unused dentry that doesn't have valid lease
1760  *
1761  * Called under dentry->d_lock.
1762  */
1763 static int ceph_d_delete(const struct dentry *dentry)
1764 {
1765         struct ceph_dentry_info *di;
1766
1767         /* won't release caps */
1768         if (d_really_is_negative(dentry))
1769                 return 0;
1770         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1771                 return 0;
1772         /* vaild lease? */
1773         di = ceph_dentry(dentry);
1774         if (di) {
1775                 if (__dentry_lease_is_valid(di))
1776                         return 0;
1777                 if (__dir_lease_try_check(dentry))
1778                         return 0;
1779         }
1780         return 1;
1781 }
1782
1783 /*
1784  * Release our ceph_dentry_info.
1785  */
1786 static void ceph_d_release(struct dentry *dentry)
1787 {
1788         struct ceph_dentry_info *di = ceph_dentry(dentry);
1789         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1790
1791         dout("d_release %p\n", dentry);
1792
1793         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1794
1795         spin_lock(&dentry->d_lock);
1796         __dentry_lease_unlist(di);
1797         dentry->d_fsdata = NULL;
1798         spin_unlock(&dentry->d_lock);
1799
1800         if (di->lease_session)
1801                 ceph_put_mds_session(di->lease_session);
1802         kmem_cache_free(ceph_dentry_cachep, di);
1803 }
1804
1805 /*
1806  * When the VFS prunes a dentry from the cache, we need to clear the
1807  * complete flag on the parent directory.
1808  *
1809  * Called under dentry->d_lock.
1810  */
1811 static void ceph_d_prune(struct dentry *dentry)
1812 {
1813         struct ceph_inode_info *dir_ci;
1814         struct ceph_dentry_info *di;
1815
1816         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1817
1818         /* do we have a valid parent? */
1819         if (IS_ROOT(dentry))
1820                 return;
1821
1822         /* we hold d_lock, so d_parent is stable */
1823         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1824         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1825                 return;
1826
1827         /* who calls d_delete() should also disable dcache readdir */
1828         if (d_really_is_negative(dentry))
1829                 return;
1830
1831         /* d_fsdata does not get cleared until d_release */
1832         if (!d_unhashed(dentry)) {
1833                 __ceph_dir_clear_complete(dir_ci);
1834                 return;
1835         }
1836
1837         /* Disable dcache readdir just in case that someone called d_drop()
1838          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1839          * properly (dcache readdir is still enabled) */
1840         di = ceph_dentry(dentry);
1841         if (di->offset > 0 &&
1842             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1843                 __ceph_dir_clear_ordered(dir_ci);
1844 }
1845
1846 /*
1847  * read() on a dir.  This weird interface hack only works if mounted
1848  * with '-o dirstat'.
1849  */
1850 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1851                              loff_t *ppos)
1852 {
1853         struct ceph_dir_file_info *dfi = file->private_data;
1854         struct inode *inode = file_inode(file);
1855         struct ceph_inode_info *ci = ceph_inode(inode);
1856         int left;
1857         const int bufsize = 1024;
1858
1859         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1860                 return -EISDIR;
1861
1862         if (!dfi->dir_info) {
1863                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1864                 if (!dfi->dir_info)
1865                         return -ENOMEM;
1866                 dfi->dir_info_len =
1867                         snprintf(dfi->dir_info, bufsize,
1868                                 "entries:   %20lld\n"
1869                                 " files:    %20lld\n"
1870                                 " subdirs:  %20lld\n"
1871                                 "rentries:  %20lld\n"
1872                                 " rfiles:   %20lld\n"
1873                                 " rsubdirs: %20lld\n"
1874                                 "rbytes:    %20lld\n"
1875                                 "rctime:    %10lld.%09ld\n",
1876                                 ci->i_files + ci->i_subdirs,
1877                                 ci->i_files,
1878                                 ci->i_subdirs,
1879                                 ci->i_rfiles + ci->i_rsubdirs,
1880                                 ci->i_rfiles,
1881                                 ci->i_rsubdirs,
1882                                 ci->i_rbytes,
1883                                 ci->i_rctime.tv_sec,
1884                                 ci->i_rctime.tv_nsec);
1885         }
1886
1887         if (*ppos >= dfi->dir_info_len)
1888                 return 0;
1889         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1890         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1891         if (left == size)
1892                 return -EFAULT;
1893         *ppos += (size - left);
1894         return size - left;
1895 }
1896
1897
1898
1899 /*
1900  * Return name hash for a given dentry.  This is dependent on
1901  * the parent directory's hash function.
1902  */
1903 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1904 {
1905         struct ceph_inode_info *dci = ceph_inode(dir);
1906         unsigned hash;
1907
1908         switch (dci->i_dir_layout.dl_dir_hash) {
1909         case 0: /* for backward compat */
1910         case CEPH_STR_HASH_LINUX:
1911                 return dn->d_name.hash;
1912
1913         default:
1914                 spin_lock(&dn->d_lock);
1915                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1916                                      dn->d_name.name, dn->d_name.len);
1917                 spin_unlock(&dn->d_lock);
1918                 return hash;
1919         }
1920 }
1921
1922 const struct file_operations ceph_dir_fops = {
1923         .read = ceph_read_dir,
1924         .iterate = ceph_readdir,
1925         .llseek = ceph_dir_llseek,
1926         .open = ceph_open,
1927         .release = ceph_release,
1928         .unlocked_ioctl = ceph_ioctl,
1929         .compat_ioctl = compat_ptr_ioctl,
1930         .fsync = ceph_fsync,
1931         .lock = ceph_lock,
1932         .flock = ceph_flock,
1933 };
1934
1935 const struct file_operations ceph_snapdir_fops = {
1936         .iterate = ceph_readdir,
1937         .llseek = ceph_dir_llseek,
1938         .open = ceph_open,
1939         .release = ceph_release,
1940 };
1941
1942 const struct inode_operations ceph_dir_iops = {
1943         .lookup = ceph_lookup,
1944         .permission = ceph_permission,
1945         .getattr = ceph_getattr,
1946         .setattr = ceph_setattr,
1947         .listxattr = ceph_listxattr,
1948         .get_acl = ceph_get_acl,
1949         .set_acl = ceph_set_acl,
1950         .mknod = ceph_mknod,
1951         .symlink = ceph_symlink,
1952         .mkdir = ceph_mkdir,
1953         .link = ceph_link,
1954         .unlink = ceph_unlink,
1955         .rmdir = ceph_unlink,
1956         .rename = ceph_rename,
1957         .create = ceph_create,
1958         .atomic_open = ceph_atomic_open,
1959 };
1960
1961 const struct inode_operations ceph_snapdir_iops = {
1962         .lookup = ceph_lookup,
1963         .permission = ceph_permission,
1964         .getattr = ceph_getattr,
1965         .mkdir = ceph_mkdir,
1966         .rmdir = ceph_unlink,
1967         .rename = ceph_rename,
1968 };
1969
1970 const struct dentry_operations ceph_dentry_ops = {
1971         .d_revalidate = ceph_d_revalidate,
1972         .d_delete = ceph_d_delete,
1973         .d_release = ceph_d_release,
1974         .d_prune = ceph_d_prune,
1975         .d_init = ceph_d_init,
1976 };