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                 if (IS_ERR(inode))
 681                         return PTR_ERR(inode);
 682                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 683                      dentry, dentry, inode);
 684                 BUG_ON(!d_unhashed(dentry));
 685                 d_add(dentry, inode);
 686                 err = 0;
 687         }
 688         return err;
 689 }
 690
 691 /*
 692  * Figure out final result of a lookup/open request.
 693  *
 694  * Mainly, make sure we return the final req->r_dentry (if it already
 695  * existed) in place of the original VFS-provided dentry when they
 696  * differ.
 697  *
 698  * Gracefully handle the case where the MDS replies with -ENOENT and
 699  * no trace (which it may do, at its discretion, e.g., if it doesn't
 700  * care to issue a lease on the negative dentry).
 701  */
 702 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 703                                   struct dentry *dentry, int err)
 704 {
 705         if (err == -ENOENT) {
 706                 /* no trace? */
 707                 err = 0;
 708                 if (!req->r_reply_info.head->is_dentry) {
 709                         dout("ENOENT and no trace, dentry %p inode %p\n",
 710                              dentry, d_inode(dentry));
 711                         if (d_really_is_positive(dentry)) {
 712                                 d_drop(dentry);
 713                                 err = -ENOENT;
 714                         } else {
 715                                 d_add(dentry, NULL);
 716                         }
 717                 }
 718         }
 719         if (err)
 720                 dentry = ERR_PTR(err);
 721         else if (dentry != req->r_dentry)
 722                 dentry = dget(req->r_dentry);   /* we got spliced */
 723         else
 724                 dentry = NULL;
 725         return dentry;
 726 }
 727
 728 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 729 {
 730         return ceph_ino(inode) == CEPH_INO_ROOT &&
 731                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 732 }
 733
 734 /*
 735  * Look up a single dir entry.  If there is a lookup intent, inform
 736  * the MDS so that it gets our 'caps wanted' value in a single op.
 737  */
 738 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 739                                   unsigned int flags)
 740 {
 741         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 742         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 743         struct ceph_mds_request *req;
 744         int op;
 745         int mask;
 746         int err;
 747
 748         dout("lookup %p dentry %p '%pd'\n",
 749              dir, dentry, dentry);
 750
 751         if (dentry->d_name.len > NAME_MAX)
 752                 return ERR_PTR(-ENAMETOOLONG);
 753
 754         /* can we conclude ENOENT locally? */
 755         if (d_really_is_negative(dentry)) {
 756                 struct ceph_inode_info *ci = ceph_inode(dir);
 757                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 758
 759                 spin_lock(&ci->i_ceph_lock);
 760                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 761                 if (strncmp(dentry->d_name.name,
 762                             fsc->mount_options->snapdir_name,
 763                             dentry->d_name.len) &&
 764                     !is_root_ceph_dentry(dir, dentry) &&
 765                     ceph_test_mount_opt(fsc, DCACHE) &&
 766                     __ceph_dir_is_complete(ci) &&
 767                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 768                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 769                         spin_unlock(&ci->i_ceph_lock);
 770                         dout(" dir %p complete, -ENOENT\n", dir);
 771                         d_add(dentry, NULL);
 772                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 773                         return NULL;
 774                 }
 775                 spin_unlock(&ci->i_ceph_lock);
 776         }
 777
 778         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 779                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 780         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 781         if (IS_ERR(req))
 782                 return ERR_CAST(req);
 783         req->r_dentry = dget(dentry);
 784         req->r_num_caps = 2;
 785
 786         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 787         if (ceph_security_xattr_wanted(dir))
 788                 mask |= CEPH_CAP_XATTR_SHARED;
 789         req->r_args.getattr.mask = cpu_to_le32(mask);
 790
 791         req->r_parent = dir;
 792         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 793         err = ceph_mdsc_do_request(mdsc, NULL, req);
 794         err = ceph_handle_snapdir(req, dentry, err);
 795         dentry = ceph_finish_lookup(req, dentry, err);
 796         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 797         dout("lookup result=%p\n", dentry);
 798         return dentry;
 799 }
 800
 801 /*
 802  * If we do a create but get no trace back from the MDS, follow up with
 803  * a lookup (the VFS expects us to link up the provided dentry).
 804  */
 805 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 806 {
 807         struct dentry *result = ceph_lookup(dir, dentry, 0);
 808
 809         if (result && !IS_ERR(result)) {
 810                 /*
 811                  * We created the item, then did a lookup, and found
 812                  * it was already linked to another inode we already
 813                  * had in our cache (and thus got spliced). To not
 814                  * confuse VFS (especially when inode is a directory),
 815                  * we don't link our dentry to that inode, return an
 816                  * error instead.
 817                  *
 818                  * This event should be rare and it happens only when
 819                  * we talk to old MDS. Recent MDS does not send traceless
 820                  * reply for request that creates new inode.
 821                  */
 822                 d_drop(result);
 823                 return -ESTALE;
 824         }
 825         return PTR_ERR(result);
 826 }
 827
 828 static int ceph_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 829                       struct dentry *dentry, umode_t mode, dev_t rdev)
 830 {
 831         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 832         struct ceph_mds_request *req;
 833         struct ceph_acl_sec_ctx as_ctx = {};
 834         int err;
 835
 836         if (ceph_snap(dir) != CEPH_NOSNAP)
 837                 return -EROFS;
 838
 839         if (ceph_quota_is_max_files_exceeded(dir)) {
 840                 err = -EDQUOT;
 841                 goto out;
 842         }
 843
 844         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 845         if (err < 0)
 846                 goto out;
 847         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 848         if (err < 0)
 849                 goto out;
 850
 851         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 852              dir, dentry, mode, rdev);
 853         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 854         if (IS_ERR(req)) {
 855                 err = PTR_ERR(req);
 856                 goto out;
 857         }
 858         req->r_dentry = dget(dentry);
 859         req->r_num_caps = 2;
 860         req->r_parent = dir;
 861         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 862         req->r_args.mknod.mode = cpu_to_le32(mode);
 863         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 864         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 865         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 866         if (as_ctx.pagelist) {
 867                 req->r_pagelist = as_ctx.pagelist;
 868                 as_ctx.pagelist = NULL;
 869         }
 870         err = ceph_mdsc_do_request(mdsc, dir, req);
 871         if (!err && !req->r_reply_info.head->is_dentry)
 872                 err = ceph_handle_notrace_create(dir, dentry);
 873         ceph_mdsc_put_request(req);
 874 out:
 875         if (!err)
 876                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 877         else
 878                 d_drop(dentry);
 879         ceph_release_acl_sec_ctx(&as_ctx);
 880         return err;
 881 }
 882
 883 static int ceph_create(struct user_namespace *mnt_userns, struct inode *dir,
 884                        struct dentry *dentry, umode_t mode, bool excl)
 885 {
 886         return ceph_mknod(mnt_userns, dir, dentry, mode, 0);
 887 }
 888
 889 static int ceph_symlink(struct user_namespace *mnt_userns, struct inode *dir,
 890                         struct dentry *dentry, const char *dest)
 891 {
 892         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 893         struct ceph_mds_request *req;
 894         struct ceph_acl_sec_ctx as_ctx = {};
 895         int err;
 896
 897         if (ceph_snap(dir) != CEPH_NOSNAP)
 898                 return -EROFS;
 899
 900         if (ceph_quota_is_max_files_exceeded(dir)) {
 901                 err = -EDQUOT;
 902                 goto out;
 903         }
 904
 905         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 906         if (err < 0)
 907                 goto out;
 908
 909         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 910         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 911         if (IS_ERR(req)) {
 912                 err = PTR_ERR(req);
 913                 goto out;
 914         }
 915         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 916         if (!req->r_path2) {
 917                 err = -ENOMEM;
 918                 ceph_mdsc_put_request(req);
 919                 goto out;
 920         }
 921         req->r_parent = dir;
 922         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 923         req->r_dentry = dget(dentry);
 924         req->r_num_caps = 2;
 925         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 926         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 927         if (as_ctx.pagelist) {
 928                 req->r_pagelist = as_ctx.pagelist;
 929                 as_ctx.pagelist = NULL;
 930         }
 931         err = ceph_mdsc_do_request(mdsc, dir, req);
 932         if (!err && !req->r_reply_info.head->is_dentry)
 933                 err = ceph_handle_notrace_create(dir, dentry);
 934         ceph_mdsc_put_request(req);
 935 out:
 936         if (err)
 937                 d_drop(dentry);
 938         ceph_release_acl_sec_ctx(&as_ctx);
 939         return err;
 940 }
 941
 942 static int ceph_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 943                       struct dentry *dentry, umode_t mode)
 944 {
 945         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 946         struct ceph_mds_request *req;
 947         struct ceph_acl_sec_ctx as_ctx = {};
 948         int err = -EROFS;
 949         int op;
 950
 951         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 952                 /* mkdir .snap/foo is a MKSNAP */
 953                 op = CEPH_MDS_OP_MKSNAP;
 954                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 955                      dentry, dentry);
 956         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 957                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 958                 op = CEPH_MDS_OP_MKDIR;
 959         } else {
 960                 goto out;
 961         }
 962
 963         if (op == CEPH_MDS_OP_MKDIR &&
 964             ceph_quota_is_max_files_exceeded(dir)) {
 965                 err = -EDQUOT;
 966                 goto out;
 967         }
 968
 969         mode |= S_IFDIR;
 970         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 971         if (err < 0)
 972                 goto out;
 973         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 974         if (err < 0)
 975                 goto out;
 976
 977         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 978         if (IS_ERR(req)) {
 979                 err = PTR_ERR(req);
 980                 goto out;
 981         }
 982
 983         req->r_dentry = dget(dentry);
 984         req->r_num_caps = 2;
 985         req->r_parent = dir;
 986         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 987         req->r_args.mkdir.mode = cpu_to_le32(mode);
 988         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 989         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 990         if (as_ctx.pagelist) {
 991                 req->r_pagelist = as_ctx.pagelist;
 992                 as_ctx.pagelist = NULL;
 993         }
 994         err = ceph_mdsc_do_request(mdsc, dir, req);
 995         if (!err &&
 996             !req->r_reply_info.head->is_target &&
 997             !req->r_reply_info.head->is_dentry)
 998                 err = ceph_handle_notrace_create(dir, dentry);
 999         ceph_mdsc_put_request(req);
1000 out:
1001         if (!err)
1002                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1003         else
1004                 d_drop(dentry);
1005         ceph_release_acl_sec_ctx(&as_ctx);
1006         return err;
1007 }
1008
1009 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1010                      struct dentry *dentry)
1011 {
1012         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1013         struct ceph_mds_request *req;
1014         int err;
1015
1016         if (ceph_snap(dir) != CEPH_NOSNAP)
1017                 return -EROFS;
1018
1019         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1020              old_dentry, dentry);
1021         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1022         if (IS_ERR(req)) {
1023                 d_drop(dentry);
1024                 return PTR_ERR(req);
1025         }
1026         req->r_dentry = dget(dentry);
1027         req->r_num_caps = 2;
1028         req->r_old_dentry = dget(old_dentry);
1029         req->r_parent = dir;
1030         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1031         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1032         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1033         /* release LINK_SHARED on source inode (mds will lock it) */
1034         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1035         err = ceph_mdsc_do_request(mdsc, dir, req);
1036         if (err) {
1037                 d_drop(dentry);
1038         } else if (!req->r_reply_info.head->is_dentry) {
1039                 ihold(d_inode(old_dentry));
1040                 d_instantiate(dentry, d_inode(old_dentry));
1041         }
1042         ceph_mdsc_put_request(req);
1043         return err;
1044 }
1045
1046 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1047                                  struct ceph_mds_request *req)
1048 {
1049         int result = req->r_err ? req->r_err :
1050                         le32_to_cpu(req->r_reply_info.head->result);
1051
1052         if (result == -EJUKEBOX)
1053                 goto out;
1054
1055         /* If op failed, mark everyone involved for errors */
1056         if (result) {
1057                 int pathlen = 0;
1058                 u64 base = 0;
1059                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1060                                                   &base, 0);
1061
1062                 /* mark error on parent + clear complete */
1063                 mapping_set_error(req->r_parent->i_mapping, result);
1064                 ceph_dir_clear_complete(req->r_parent);
1065
1066                 /* drop the dentry -- we don't know its status */
1067                 if (!d_unhashed(req->r_dentry))
1068                         d_drop(req->r_dentry);
1069
1070                 /* mark inode itself for an error (since metadata is bogus) */
1071                 mapping_set_error(req->r_old_inode->i_mapping, result);
1072
1073                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1074                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1075                 ceph_mdsc_free_path(path, pathlen);
1076         }
1077 out:
1078         iput(req->r_old_inode);
1079         ceph_mdsc_release_dir_caps(req);
1080 }
1081
1082 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1083 {
1084         struct ceph_inode_info *ci = ceph_inode(dir);
1085         struct ceph_dentry_info *di;
1086         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1087
1088         spin_lock(&ci->i_ceph_lock);
1089         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1090                 ceph_take_cap_refs(ci, want, false);
1091                 got = want;
1092         }
1093         spin_unlock(&ci->i_ceph_lock);
1094
1095         /* If we didn't get anything, return 0 */
1096         if (!got)
1097                 return 0;
1098
1099         spin_lock(&dentry->d_lock);
1100         di = ceph_dentry(dentry);
1101         /*
1102          * - We are holding Fx, which implies Fs caps.
1103          * - Only support async unlink for primary linkage
1104          */
1105         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1106             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1107                 want = 0;
1108         spin_unlock(&dentry->d_lock);
1109
1110         /* Do we still want what we've got? */
1111         if (want == got)
1112                 return got;
1113
1114         ceph_put_cap_refs(ci, got);
1115         return 0;
1116 }
1117
1118 /*
1119  * rmdir and unlink are differ only by the metadata op code
1120  */
1121 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1122 {
1123         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1124         struct ceph_mds_client *mdsc = fsc->mdsc;
1125         struct inode *inode = d_inode(dentry);
1126         struct ceph_mds_request *req;
1127         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1128         int err = -EROFS;
1129         int op;
1130
1131         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1132                 /* rmdir .snap/foo is RMSNAP */
1133                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1134                 op = CEPH_MDS_OP_RMSNAP;
1135         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1136                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1137                      dir, dentry, inode);
1138                 op = d_is_dir(dentry) ?
1139                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1140         } else
1141                 goto out;
1142 retry:
1143         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1144         if (IS_ERR(req)) {
1145                 err = PTR_ERR(req);
1146                 goto out;
1147         }
1148         req->r_dentry = dget(dentry);
1149         req->r_num_caps = 2;
1150         req->r_parent = dir;
1151         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1152         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1153         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1154
1155         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1156             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1157                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1158                      dentry->d_name.len, dentry->d_name.name,
1159                      ceph_cap_string(req->r_dir_caps));
1160                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1161                 req->r_callback = ceph_async_unlink_cb;
1162                 req->r_old_inode = d_inode(dentry);
1163                 ihold(req->r_old_inode);
1164                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1165                 if (!err) {
1166                         /*
1167                          * We have enough caps, so we assume that the unlink
1168                          * will succeed. Fix up the target inode and dcache.
1169                          */
1170                         drop_nlink(inode);
1171                         d_delete(dentry);
1172                 } else if (err == -EJUKEBOX) {
1173                         try_async = false;
1174                         ceph_mdsc_put_request(req);
1175                         goto retry;
1176                 }
1177         } else {
1178                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1179                 err = ceph_mdsc_do_request(mdsc, dir, req);
1180                 if (!err && !req->r_reply_info.head->is_dentry)
1181                         d_delete(dentry);
1182         }
1183
1184         ceph_mdsc_put_request(req);
1185 out:
1186         return err;
1187 }
1188
1189 static int ceph_rename(struct user_namespace *mnt_userns, struct inode *old_dir,
1190                        struct dentry *old_dentry, struct inode *new_dir,
1191                        struct dentry *new_dentry, unsigned int flags)
1192 {
1193         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1194         struct ceph_mds_request *req;
1195         int op = CEPH_MDS_OP_RENAME;
1196         int err;
1197
1198         if (flags)
1199                 return -EINVAL;
1200
1201         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1202                 return -EXDEV;
1203         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1204                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1205                         op = CEPH_MDS_OP_RENAMESNAP;
1206                 else
1207                         return -EROFS;
1208         }
1209         /* don't allow cross-quota renames */
1210         if ((old_dir != new_dir) &&
1211             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1212                 return -EXDEV;
1213
1214         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1215              old_dir, old_dentry, new_dir, new_dentry);
1216         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1217         if (IS_ERR(req))
1218                 return PTR_ERR(req);
1219         ihold(old_dir);
1220         req->r_dentry = dget(new_dentry);
1221         req->r_num_caps = 2;
1222         req->r_old_dentry = dget(old_dentry);
1223         req->r_old_dentry_dir = old_dir;
1224         req->r_parent = new_dir;
1225         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1226         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1227         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1228         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1229         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1230         /* release LINK_RDCACHE on source inode (mds will lock it) */
1231         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1232         if (d_really_is_positive(new_dentry)) {
1233                 req->r_inode_drop =
1234                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1235         }
1236         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1237         if (!err && !req->r_reply_info.head->is_dentry) {
1238                 /*
1239                  * Normally d_move() is done by fill_trace (called by
1240                  * do_request, above).  If there is no trace, we need
1241                  * to do it here.
1242                  */
1243                 d_move(old_dentry, new_dentry);
1244         }
1245         ceph_mdsc_put_request(req);
1246         return err;
1247 }
1248
1249 /*
1250  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1251  * Leases at front of the list will expire first. (Assume all leases have
1252  * similar duration)
1253  *
1254  * Called under dentry->d_lock.
1255  */
1256 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1257 {
1258         struct dentry *dn = di->dentry;
1259         struct ceph_mds_client *mdsc;
1260
1261         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1262
1263         di->flags |= CEPH_DENTRY_LEASE_LIST;
1264         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1265                 di->flags |= CEPH_DENTRY_REFERENCED;
1266                 return;
1267         }
1268
1269         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1270         spin_lock(&mdsc->dentry_list_lock);
1271         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1272         spin_unlock(&mdsc->dentry_list_lock);
1273 }
1274
1275 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1276                                      struct ceph_dentry_info *di)
1277 {
1278         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1279         di->lease_gen = 0;
1280         di->time = jiffies;
1281         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1282 }
1283
1284 /*
1285  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1286  * list if it's not in the list, otherwise set 'referenced' flag.
1287  *
1288  * Called under dentry->d_lock.
1289  */
1290 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1291 {
1292         struct dentry *dn = di->dentry;
1293         struct ceph_mds_client *mdsc;
1294
1295         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1296              di, dn, dn, di->offset);
1297
1298         if (!list_empty(&di->lease_list)) {
1299                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1300                         /* don't remove dentry from dentry lease list
1301                          * if its lease is valid */
1302                         if (__dentry_lease_is_valid(di))
1303                                 return;
1304                 } else {
1305                         di->flags |= CEPH_DENTRY_REFERENCED;
1306                         return;
1307                 }
1308         }
1309
1310         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1311                 di->flags |= CEPH_DENTRY_REFERENCED;
1312                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1313                 return;
1314         }
1315
1316         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1317         spin_lock(&mdsc->dentry_list_lock);
1318         __dentry_dir_lease_touch(mdsc, di),
1319         spin_unlock(&mdsc->dentry_list_lock);
1320 }
1321
1322 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1323 {
1324         struct ceph_mds_client *mdsc;
1325         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1326                 return;
1327         if (list_empty(&di->lease_list))
1328                 return;
1329
1330         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1331         spin_lock(&mdsc->dentry_list_lock);
1332         list_del_init(&di->lease_list);
1333         spin_unlock(&mdsc->dentry_list_lock);
1334 }
1335
1336 enum {
1337         KEEP    = 0,
1338         DELETE  = 1,
1339         TOUCH   = 2,
1340         STOP    = 4,
1341 };
1342
1343 struct ceph_lease_walk_control {
1344         bool dir_lease;
1345         bool expire_dir_lease;
1346         unsigned long nr_to_scan;
1347         unsigned long dir_lease_ttl;
1348 };
1349
1350 static unsigned long
1351 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1352                      struct ceph_lease_walk_control *lwc,
1353                      int (*check)(struct dentry*, void*))
1354 {
1355         struct ceph_dentry_info *di, *tmp;
1356         struct dentry *dentry, *last = NULL;
1357         struct list_head* list;
1358         LIST_HEAD(dispose);
1359         unsigned long freed = 0;
1360         int ret = 0;
1361
1362         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1363         spin_lock(&mdsc->dentry_list_lock);
1364         list_for_each_entry_safe(di, tmp, list, lease_list) {
1365                 if (!lwc->nr_to_scan)
1366                         break;
1367                 --lwc->nr_to_scan;
1368
1369                 dentry = di->dentry;
1370                 if (last == dentry)
1371                         break;
1372
1373                 if (!spin_trylock(&dentry->d_lock))
1374                         continue;
1375
1376                 if (__lockref_is_dead(&dentry->d_lockref)) {
1377                         list_del_init(&di->lease_list);
1378                         goto next;
1379                 }
1380
1381                 ret = check(dentry, lwc);
1382                 if (ret & TOUCH) {
1383                         /* move it into tail of dir lease list */
1384                         __dentry_dir_lease_touch(mdsc, di);
1385                         if (!last)
1386                                 last = dentry;
1387                 }
1388                 if (ret & DELETE) {
1389                         /* stale lease */
1390                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1391                         if (dentry->d_lockref.count > 0) {
1392                                 /* update_dentry_lease() will re-add
1393                                  * it to lease list, or
1394                                  * ceph_d_delete() will return 1 when
1395                                  * last reference is dropped */
1396                                 list_del_init(&di->lease_list);
1397                         } else {
1398                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1399                                 list_move_tail(&di->lease_list, &dispose);
1400                                 dget_dlock(dentry);
1401                         }
1402                 }
1403 next:
1404                 spin_unlock(&dentry->d_lock);
1405                 if (ret & STOP)
1406                         break;
1407         }
1408         spin_unlock(&mdsc->dentry_list_lock);
1409
1410         while (!list_empty(&dispose)) {
1411                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1412                                       lease_list);
1413                 dentry = di->dentry;
1414                 spin_lock(&dentry->d_lock);
1415
1416                 list_del_init(&di->lease_list);
1417                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1418                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1419                         spin_lock(&mdsc->dentry_list_lock);
1420                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1421                                 list_add_tail(&di->lease_list,
1422                                               &mdsc->dentry_leases);
1423                         } else {
1424                                 __dentry_dir_lease_touch(mdsc, di);
1425                         }
1426                         spin_unlock(&mdsc->dentry_list_lock);
1427                 } else {
1428                         freed++;
1429                 }
1430
1431                 spin_unlock(&dentry->d_lock);
1432                 /* ceph_d_delete() does the trick */
1433                 dput(dentry);
1434         }
1435         return freed;
1436 }
1437
1438 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1439 {
1440         struct ceph_dentry_info *di = ceph_dentry(dentry);
1441         int ret;
1442
1443         if (__dentry_lease_is_valid(di))
1444                 return STOP;
1445         ret = __dir_lease_try_check(dentry);
1446         if (ret == -EBUSY)
1447                 return KEEP;
1448         if (ret > 0)
1449                 return TOUCH;
1450         return DELETE;
1451 }
1452
1453 static int __dir_lease_check(struct dentry *dentry, void *arg)
1454 {
1455         struct ceph_lease_walk_control *lwc = arg;
1456         struct ceph_dentry_info *di = ceph_dentry(dentry);
1457
1458         int ret = __dir_lease_try_check(dentry);
1459         if (ret == -EBUSY)
1460                 return KEEP;
1461         if (ret > 0) {
1462                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1463                         return STOP;
1464                 /* Move dentry to tail of dir lease list if we don't want
1465                  * to delete it. So dentries in the list are checked in a
1466                  * round robin manner */
1467                 if (!lwc->expire_dir_lease)
1468                         return TOUCH;
1469                 if (dentry->d_lockref.count > 0 ||
1470                     (di->flags & CEPH_DENTRY_REFERENCED))
1471                         return TOUCH;
1472                 /* invalidate dir lease */
1473                 di->lease_shared_gen = 0;
1474         }
1475         return DELETE;
1476 }
1477
1478 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1479 {
1480         struct ceph_lease_walk_control lwc;
1481         unsigned long count;
1482         unsigned long freed;
1483
1484         spin_lock(&mdsc->caps_list_lock);
1485         if (mdsc->caps_use_max > 0 &&
1486             mdsc->caps_use_count > mdsc->caps_use_max)
1487                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1488         else
1489                 count = 0;
1490         spin_unlock(&mdsc->caps_list_lock);
1491
1492         lwc.dir_lease = false;
1493         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1494         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1495         if (!lwc.nr_to_scan) /* more invalid leases */
1496                 return -EAGAIN;
1497
1498         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1499                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1500
1501         lwc.dir_lease = true;
1502         lwc.expire_dir_lease = freed < count;
1503         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1504         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1505         if (!lwc.nr_to_scan) /* more to check */
1506                 return -EAGAIN;
1507
1508         return freed > 0 ? 1 : 0;
1509 }
1510
1511 /*
1512  * Ensure a dentry lease will no longer revalidate.
1513  */
1514 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1515 {
1516         struct ceph_dentry_info *di = ceph_dentry(dentry);
1517         spin_lock(&dentry->d_lock);
1518         di->time = jiffies;
1519         di->lease_shared_gen = 0;
1520         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1521         __dentry_lease_unlist(di);
1522         spin_unlock(&dentry->d_lock);
1523 }
1524
1525 /*
1526  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1527  * renew if the least is more than half up.
1528  */
1529 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1530 {
1531         struct ceph_mds_session *session;
1532
1533         if (!di->lease_gen)
1534                 return false;
1535
1536         session = di->lease_session;
1537         if (session) {
1538                 u32 gen;
1539                 unsigned long ttl;
1540
1541                 spin_lock(&session->s_gen_ttl_lock);
1542                 gen = session->s_cap_gen;
1543                 ttl = session->s_cap_ttl;
1544                 spin_unlock(&session->s_gen_ttl_lock);
1545
1546                 if (di->lease_gen == gen &&
1547                     time_before(jiffies, ttl) &&
1548                     time_before(jiffies, di->time))
1549                         return true;
1550         }
1551         di->lease_gen = 0;
1552         return false;
1553 }
1554
1555 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1556 {
1557         struct ceph_dentry_info *di;
1558         struct ceph_mds_session *session = NULL;
1559         u32 seq = 0;
1560         int valid = 0;
1561
1562         spin_lock(&dentry->d_lock);
1563         di = ceph_dentry(dentry);
1564         if (di && __dentry_lease_is_valid(di)) {
1565                 valid = 1;
1566
1567                 if (di->lease_renew_after &&
1568                     time_after(jiffies, di->lease_renew_after)) {
1569                         /*
1570                          * We should renew. If we're in RCU walk mode
1571                          * though, we can't do that so just return
1572                          * -ECHILD.
1573                          */
1574                         if (flags & LOOKUP_RCU) {
1575                                 valid = -ECHILD;
1576                         } else {
1577                                 session = ceph_get_mds_session(di->lease_session);
1578                                 seq = di->lease_seq;
1579                                 di->lease_renew_after = 0;
1580                                 di->lease_renew_from = jiffies;
1581                         }
1582                 }
1583         }
1584         spin_unlock(&dentry->d_lock);
1585
1586         if (session) {
1587                 ceph_mdsc_lease_send_msg(session, dentry,
1588                                          CEPH_MDS_LEASE_RENEW, seq);
1589                 ceph_put_mds_session(session);
1590         }
1591         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1592         return valid;
1593 }
1594
1595 /*
1596  * Called under dentry->d_lock.
1597  */
1598 static int __dir_lease_try_check(const struct dentry *dentry)
1599 {
1600         struct ceph_dentry_info *di = ceph_dentry(dentry);
1601         struct inode *dir;
1602         struct ceph_inode_info *ci;
1603         int valid = 0;
1604
1605         if (!di->lease_shared_gen)
1606                 return 0;
1607         if (IS_ROOT(dentry))
1608                 return 0;
1609
1610         dir = d_inode(dentry->d_parent);
1611         ci = ceph_inode(dir);
1612
1613         if (spin_trylock(&ci->i_ceph_lock)) {
1614                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1615                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1616                         valid = 1;
1617                 spin_unlock(&ci->i_ceph_lock);
1618         } else {
1619                 valid = -EBUSY;
1620         }
1621
1622         if (!valid)
1623                 di->lease_shared_gen = 0;
1624         return valid;
1625 }
1626
1627 /*
1628  * Check if directory-wide content lease/cap is valid.
1629  */
1630 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1631                               struct ceph_mds_client *mdsc)
1632 {
1633         struct ceph_inode_info *ci = ceph_inode(dir);
1634         int valid;
1635         int shared_gen;
1636
1637         spin_lock(&ci->i_ceph_lock);
1638         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1639         if (valid) {
1640                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1641                 shared_gen = atomic_read(&ci->i_shared_gen);
1642         }
1643         spin_unlock(&ci->i_ceph_lock);
1644         if (valid) {
1645                 struct ceph_dentry_info *di;
1646                 spin_lock(&dentry->d_lock);
1647                 di = ceph_dentry(dentry);
1648                 if (dir == d_inode(dentry->d_parent) &&
1649                     di && di->lease_shared_gen == shared_gen)
1650                         __ceph_dentry_dir_lease_touch(di);
1651                 else
1652                         valid = 0;
1653                 spin_unlock(&dentry->d_lock);
1654         }
1655         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1656              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1657         return valid;
1658 }
1659
1660 /*
1661  * Check if cached dentry can be trusted.
1662  */
1663 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1664 {
1665         int valid = 0;
1666         struct dentry *parent;
1667         struct inode *dir, *inode;
1668         struct ceph_mds_client *mdsc;
1669
1670         if (flags & LOOKUP_RCU) {
1671                 parent = READ_ONCE(dentry->d_parent);
1672                 dir = d_inode_rcu(parent);
1673                 if (!dir)
1674                         return -ECHILD;
1675                 inode = d_inode_rcu(dentry);
1676         } else {
1677                 parent = dget_parent(dentry);
1678                 dir = d_inode(parent);
1679                 inode = d_inode(dentry);
1680         }
1681
1682         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1683              dentry, inode, ceph_dentry(dentry)->offset);
1684
1685         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1686
1687         /* always trust cached snapped dentries, snapdir dentry */
1688         if (ceph_snap(dir) != CEPH_NOSNAP) {
1689                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1690                      dentry, inode);
1691                 valid = 1;
1692         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1693                 valid = 1;
1694         } else {
1695                 valid = dentry_lease_is_valid(dentry, flags);
1696                 if (valid == -ECHILD)
1697                         return valid;
1698                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1699                         if (inode)
1700                                 valid = ceph_is_any_caps(inode);
1701                         else
1702                                 valid = 1;
1703                 }
1704         }
1705
1706         if (!valid) {
1707                 struct ceph_mds_request *req;
1708                 int op, err;
1709                 u32 mask;
1710
1711                 if (flags & LOOKUP_RCU)
1712                         return -ECHILD;
1713
1714                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1715
1716                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1717                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1718                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1719                 if (!IS_ERR(req)) {
1720                         req->r_dentry = dget(dentry);
1721                         req->r_num_caps = 2;
1722                         req->r_parent = dir;
1723
1724                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1725                         if (ceph_security_xattr_wanted(dir))
1726                                 mask |= CEPH_CAP_XATTR_SHARED;
1727                         req->r_args.getattr.mask = cpu_to_le32(mask);
1728
1729                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1730                         switch (err) {
1731                         case 0:
1732                                 if (d_really_is_positive(dentry) &&
1733                                     d_inode(dentry) == req->r_target_inode)
1734                                         valid = 1;
1735                                 break;
1736                         case -ENOENT:
1737                                 if (d_really_is_negative(dentry))
1738                                         valid = 1;
1739                                 fallthrough;
1740                         default:
1741                                 break;
1742                         }
1743                         ceph_mdsc_put_request(req);
1744                         dout("d_revalidate %p lookup result=%d\n",
1745                              dentry, err);
1746                 }
1747         } else {
1748                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1749         }
1750
1751         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1752         if (!valid)
1753                 ceph_dir_clear_complete(dir);
1754
1755         if (!(flags & LOOKUP_RCU))
1756                 dput(parent);
1757         return valid;
1758 }
1759
1760 /*
1761  * Delete unused dentry that doesn't have valid lease
1762  *
1763  * Called under dentry->d_lock.
1764  */
1765 static int ceph_d_delete(const struct dentry *dentry)
1766 {
1767         struct ceph_dentry_info *di;
1768
1769         /* won't release caps */
1770         if (d_really_is_negative(dentry))
1771                 return 0;
1772         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1773                 return 0;
1774         /* vaild lease? */
1775         di = ceph_dentry(dentry);
1776         if (di) {
1777                 if (__dentry_lease_is_valid(di))
1778                         return 0;
1779                 if (__dir_lease_try_check(dentry))
1780                         return 0;
1781         }
1782         return 1;
1783 }
1784
1785 /*
1786  * Release our ceph_dentry_info.
1787  */
1788 static void ceph_d_release(struct dentry *dentry)
1789 {
1790         struct ceph_dentry_info *di = ceph_dentry(dentry);
1791         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1792
1793         dout("d_release %p\n", dentry);
1794
1795         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1796
1797         spin_lock(&dentry->d_lock);
1798         __dentry_lease_unlist(di);
1799         dentry->d_fsdata = NULL;
1800         spin_unlock(&dentry->d_lock);
1801
1802         if (di->lease_session)
1803                 ceph_put_mds_session(di->lease_session);
1804         kmem_cache_free(ceph_dentry_cachep, di);
1805 }
1806
1807 /*
1808  * When the VFS prunes a dentry from the cache, we need to clear the
1809  * complete flag on the parent directory.
1810  *
1811  * Called under dentry->d_lock.
1812  */
1813 static void ceph_d_prune(struct dentry *dentry)
1814 {
1815         struct ceph_inode_info *dir_ci;
1816         struct ceph_dentry_info *di;
1817
1818         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1819
1820         /* do we have a valid parent? */
1821         if (IS_ROOT(dentry))
1822                 return;
1823
1824         /* we hold d_lock, so d_parent is stable */
1825         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1826         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1827                 return;
1828
1829         /* who calls d_delete() should also disable dcache readdir */
1830         if (d_really_is_negative(dentry))
1831                 return;
1832
1833         /* d_fsdata does not get cleared until d_release */
1834         if (!d_unhashed(dentry)) {
1835                 __ceph_dir_clear_complete(dir_ci);
1836                 return;
1837         }
1838
1839         /* Disable dcache readdir just in case that someone called d_drop()
1840          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1841          * properly (dcache readdir is still enabled) */
1842         di = ceph_dentry(dentry);
1843         if (di->offset > 0 &&
1844             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1845                 __ceph_dir_clear_ordered(dir_ci);
1846 }
1847
1848 /*
1849  * read() on a dir.  This weird interface hack only works if mounted
1850  * with '-o dirstat'.
1851  */
1852 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1853                              loff_t *ppos)
1854 {
1855         struct ceph_dir_file_info *dfi = file->private_data;
1856         struct inode *inode = file_inode(file);
1857         struct ceph_inode_info *ci = ceph_inode(inode);
1858         int left;
1859         const int bufsize = 1024;
1860
1861         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1862                 return -EISDIR;
1863
1864         if (!dfi->dir_info) {
1865                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1866                 if (!dfi->dir_info)
1867                         return -ENOMEM;
1868                 dfi->dir_info_len =
1869                         snprintf(dfi->dir_info, bufsize,
1870                                 "entries:   %20lld\n"
1871                                 " files:    %20lld\n"
1872                                 " subdirs:  %20lld\n"
1873                                 "rentries:  %20lld\n"
1874                                 " rfiles:   %20lld\n"
1875                                 " rsubdirs: %20lld\n"
1876                                 "rbytes:    %20lld\n"
1877                                 "rctime:    %10lld.%09ld\n",
1878                                 ci->i_files + ci->i_subdirs,
1879                                 ci->i_files,
1880                                 ci->i_subdirs,
1881                                 ci->i_rfiles + ci->i_rsubdirs,
1882                                 ci->i_rfiles,
1883                                 ci->i_rsubdirs,
1884                                 ci->i_rbytes,
1885                                 ci->i_rctime.tv_sec,
1886                                 ci->i_rctime.tv_nsec);
1887         }
1888
1889         if (*ppos >= dfi->dir_info_len)
1890                 return 0;
1891         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1892         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1893         if (left == size)
1894                 return -EFAULT;
1895         *ppos += (size - left);
1896         return size - left;
1897 }
1898
1899
1900
1901 /*
1902  * Return name hash for a given dentry.  This is dependent on
1903  * the parent directory's hash function.
1904  */
1905 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1906 {
1907         struct ceph_inode_info *dci = ceph_inode(dir);
1908         unsigned hash;
1909
1910         switch (dci->i_dir_layout.dl_dir_hash) {
1911         case 0: /* for backward compat */
1912         case CEPH_STR_HASH_LINUX:
1913                 return dn->d_name.hash;
1914
1915         default:
1916                 spin_lock(&dn->d_lock);
1917                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1918                                      dn->d_name.name, dn->d_name.len);
1919                 spin_unlock(&dn->d_lock);
1920                 return hash;
1921         }
1922 }
1923
1924 const struct file_operations ceph_dir_fops = {
1925         .read = ceph_read_dir,
1926         .iterate = ceph_readdir,
1927         .llseek = ceph_dir_llseek,
1928         .open = ceph_open,
1929         .release = ceph_release,
1930         .unlocked_ioctl = ceph_ioctl,
1931         .compat_ioctl = compat_ptr_ioctl,
1932         .fsync = ceph_fsync,
1933         .lock = ceph_lock,
1934         .flock = ceph_flock,
1935 };
1936
1937 const struct file_operations ceph_snapdir_fops = {
1938         .iterate = ceph_readdir,
1939         .llseek = ceph_dir_llseek,
1940         .open = ceph_open,
1941         .release = ceph_release,
1942 };
1943
1944 const struct inode_operations ceph_dir_iops = {
1945         .lookup = ceph_lookup,
1946         .permission = ceph_permission,
1947         .getattr = ceph_getattr,
1948         .setattr = ceph_setattr,
1949         .listxattr = ceph_listxattr,
1950         .get_acl = ceph_get_acl,
1951         .set_acl = ceph_set_acl,
1952         .mknod = ceph_mknod,
1953         .symlink = ceph_symlink,
1954         .mkdir = ceph_mkdir,
1955         .link = ceph_link,
1956         .unlink = ceph_unlink,
1957         .rmdir = ceph_unlink,
1958         .rename = ceph_rename,
1959         .create = ceph_create,
1960         .atomic_open = ceph_atomic_open,
1961 };
1962
1963 const struct inode_operations ceph_snapdir_iops = {
1964         .lookup = ceph_lookup,
1965         .permission = ceph_permission,
1966         .getattr = ceph_getattr,
1967         .mkdir = ceph_mkdir,
1968         .rmdir = ceph_unlink,
1969         .rename = ceph_rename,
1970 };
1971
1972 const struct dentry_operations ceph_dentry_ops = {
1973         .d_revalidate = ceph_d_revalidate,
1974         .d_delete = ceph_d_delete,
1975         .d_release = ceph_d_release,
1976         .d_prune = ceph_d_prune,
1977         .d_init = ceph_d_init,
1978 };