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_rwsem, 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_rwsem. */
 159         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 160                 dentry = cache_ctl->dentries[cache_ctl->index];
 161         else
 162                 dentry = NULL;
 163         spin_unlock(&parent->d_lock);
 164         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 165                 dentry = NULL;
 166         rcu_read_unlock();
 167         return dentry ? : ERR_PTR(-EAGAIN);
 168 }
 169
 170 /*
 171  * When possible, we try to satisfy a readdir by peeking at the
 172  * dcache.  We make this work by carefully ordering dentries on
 173  * d_child when we initially get results back from the MDS, and
 174  * falling back to a "normal" sync readdir if any dentries in the dir
 175  * are dropped.
 176  *
 177  * Complete dir indicates that we have all dentries in the dir.  It is
 178  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 179  * the MDS if/when the directory is modified).
 180  */
 181 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 182                             int shared_gen)
 183 {
 184         struct ceph_dir_file_info *dfi = file->private_data;
 185         struct dentry *parent = file->f_path.dentry;
 186         struct inode *dir = d_inode(parent);
 187         struct dentry *dentry, *last = NULL;
 188         struct ceph_dentry_info *di;
 189         struct ceph_readdir_cache_control cache_ctl = {};
 190         u64 idx = 0;
 191         int err = 0;
 192
 193         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 194
 195         /* search start position */
 196         if (ctx->pos > 2) {
 197                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 198                 while (count > 0) {
 199                         u64 step = count >> 1;
 200                         dentry = __dcache_find_get_entry(parent, idx + step,
 201                                                          &cache_ctl);
 202                         if (!dentry) {
 203                                 /* use linar search */
 204                                 idx = 0;
 205                                 break;
 206                         }
 207                         if (IS_ERR(dentry)) {
 208                                 err = PTR_ERR(dentry);
 209                                 goto out;
 210                         }
 211                         di = ceph_dentry(dentry);
 212                         spin_lock(&dentry->d_lock);
 213                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 214                                 idx += step + 1;
 215                                 count -= step + 1;
 216                         } else {
 217                                 count = step;
 218                         }
 219                         spin_unlock(&dentry->d_lock);
 220                         dput(dentry);
 221                 }
 222
 223                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 224         }
 225
 226
 227         for (;;) {
 228                 bool emit_dentry = false;
 229                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 230                 if (!dentry) {
 231                         dfi->file_info.flags |= CEPH_F_ATEND;
 232                         err = 0;
 233                         break;
 234                 }
 235                 if (IS_ERR(dentry)) {
 236                         err = PTR_ERR(dentry);
 237                         goto out;
 238                 }
 239
 240                 spin_lock(&dentry->d_lock);
 241                 di = ceph_dentry(dentry);
 242                 if (d_unhashed(dentry) ||
 243                     d_really_is_negative(dentry) ||
 244                     di->lease_shared_gen != shared_gen) {
 245                         spin_unlock(&dentry->d_lock);
 246                         dput(dentry);
 247                         err = -EAGAIN;
 248                         goto out;
 249                 }
 250                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 251                         __ceph_dentry_dir_lease_touch(di);
 252                         emit_dentry = true;
 253                 }
 254                 spin_unlock(&dentry->d_lock);
 255
 256                 if (emit_dentry) {
 257                         dout(" %llx dentry %p %pd %p\n", di->offset,
 258                              dentry, dentry, d_inode(dentry));
 259                         ctx->pos = di->offset;
 260                         if (!dir_emit(ctx, dentry->d_name.name,
 261                                       dentry->d_name.len, ceph_present_inode(d_inode(dentry)),
 262                                       d_inode(dentry)->i_mode >> 12)) {
 263                                 dput(dentry);
 264                                 err = 0;
 265                                 break;
 266                         }
 267                         ctx->pos++;
 268
 269                         if (last)
 270                                 dput(last);
 271                         last = dentry;
 272                 } else {
 273                         dput(dentry);
 274                 }
 275         }
 276 out:
 277         ceph_readdir_cache_release(&cache_ctl);
 278         if (last) {
 279                 int ret;
 280                 di = ceph_dentry(last);
 281                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 282                                        fpos_off(di->offset) + 1);
 283                 if (ret < 0)
 284                         err = ret;
 285                 dput(last);
 286                 /* last_name no longer match cache index */
 287                 if (dfi->readdir_cache_idx >= 0) {
 288                         dfi->readdir_cache_idx = -1;
 289                         dfi->dir_release_count = 0;
 290                 }
 291         }
 292         return err;
 293 }
 294
 295 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 296 {
 297         if (!dfi->last_readdir)
 298                 return true;
 299         if (is_hash_order(pos))
 300                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 301         else
 302                 return dfi->frag != fpos_frag(pos);
 303 }
 304
 305 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 306 {
 307         struct ceph_dir_file_info *dfi = file->private_data;
 308         struct inode *inode = file_inode(file);
 309         struct ceph_inode_info *ci = ceph_inode(inode);
 310         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 311         struct ceph_mds_client *mdsc = fsc->mdsc;
 312         int i;
 313         int err;
 314         unsigned frag = -1;
 315         struct ceph_mds_reply_info_parsed *rinfo;
 316
 317         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 318         if (dfi->file_info.flags & CEPH_F_ATEND)
 319                 return 0;
 320
 321         /* always start with . and .. */
 322         if (ctx->pos == 0) {
 323                 dout("readdir off 0 -> '.'\n");
 324                 if (!dir_emit(ctx, ".", 1, ceph_present_inode(inode),
 325                             inode->i_mode >> 12))
 326                         return 0;
 327                 ctx->pos = 1;
 328         }
 329         if (ctx->pos == 1) {
 330                 u64 ino;
 331                 struct dentry *dentry = file->f_path.dentry;
 332
 333                 spin_lock(&dentry->d_lock);
 334                 ino = ceph_present_inode(dentry->d_parent->d_inode);
 335                 spin_unlock(&dentry->d_lock);
 336
 337                 dout("readdir off 1 -> '..'\n");
 338                 if (!dir_emit(ctx, "..", 2, ino, inode->i_mode >> 12))
 339                         return 0;
 340                 ctx->pos = 2;
 341         }
 342
 343         spin_lock(&ci->i_ceph_lock);
 344         /* request Fx cap. if have Fx, we don't need to release Fs cap
 345          * for later create/unlink. */
 346         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_WR);
 347         /* can we use the dcache? */
 348         if (ceph_test_mount_opt(fsc, DCACHE) &&
 349             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 350             ceph_snap(inode) != CEPH_SNAPDIR &&
 351             __ceph_dir_is_complete_ordered(ci) &&
 352             __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 353                 int shared_gen = atomic_read(&ci->i_shared_gen);
 354
 355                 spin_unlock(&ci->i_ceph_lock);
 356                 err = __dcache_readdir(file, ctx, shared_gen);
 357                 if (err != -EAGAIN)
 358                         return err;
 359         } else {
 360                 spin_unlock(&ci->i_ceph_lock);
 361         }
 362
 363         /* proceed with a normal readdir */
 364 more:
 365         /* do we have the correct frag content buffered? */
 366         if (need_send_readdir(dfi, ctx->pos)) {
 367                 struct ceph_mds_request *req;
 368                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 369                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 370
 371                 /* discard old result, if any */
 372                 if (dfi->last_readdir) {
 373                         ceph_mdsc_put_request(dfi->last_readdir);
 374                         dfi->last_readdir = NULL;
 375                 }
 376
 377                 if (is_hash_order(ctx->pos)) {
 378                         /* fragtree isn't always accurate. choose frag
 379                          * based on previous reply when possible. */
 380                         if (frag == (unsigned)-1)
 381                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 382                                                         NULL, NULL);
 383                 } else {
 384                         frag = fpos_frag(ctx->pos);
 385                 }
 386
 387                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 388                      ceph_vinop(inode), frag, dfi->last_name);
 389                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 390                 if (IS_ERR(req))
 391                         return PTR_ERR(req);
 392                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 393                 if (err) {
 394                         ceph_mdsc_put_request(req);
 395                         return err;
 396                 }
 397                 /* hints to request -> mds selection code */
 398                 req->r_direct_mode = USE_AUTH_MDS;
 399                 if (op == CEPH_MDS_OP_READDIR) {
 400                         req->r_direct_hash = ceph_frag_value(frag);
 401                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 402                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 403                 }
 404                 if (dfi->last_name) {
 405                         req->r_path2 = kstrdup(dfi->last_name, GFP_KERNEL);
 406                         if (!req->r_path2) {
 407                                 ceph_mdsc_put_request(req);
 408                                 return -ENOMEM;
 409                         }
 410                 } else if (is_hash_order(ctx->pos)) {
 411                         req->r_args.readdir.offset_hash =
 412                                 cpu_to_le32(fpos_hash(ctx->pos));
 413                 }
 414
 415                 req->r_dir_release_cnt = dfi->dir_release_count;
 416                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 417                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 418                 req->r_readdir_offset = dfi->next_offset;
 419                 req->r_args.readdir.frag = cpu_to_le32(frag);
 420                 req->r_args.readdir.flags =
 421                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 422
 423                 req->r_inode = inode;
 424                 ihold(inode);
 425                 req->r_dentry = dget(file->f_path.dentry);
 426                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 427                 if (err < 0) {
 428                         ceph_mdsc_put_request(req);
 429                         return err;
 430                 }
 431                 dout("readdir got and parsed readdir result=%d on "
 432                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 433                      err, frag,
 434                      (int)req->r_reply_info.dir_end,
 435                      (int)req->r_reply_info.dir_complete,
 436                      (int)req->r_reply_info.hash_order);
 437
 438                 rinfo = &req->r_reply_info;
 439                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 440                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 441                         if (!rinfo->hash_order) {
 442                                 dfi->next_offset = req->r_readdir_offset;
 443                                 /* adjust ctx->pos to beginning of frag */
 444                                 ctx->pos = ceph_make_fpos(frag,
 445                                                           dfi->next_offset,
 446                                                           false);
 447                         }
 448                 }
 449
 450                 dfi->frag = frag;
 451                 dfi->last_readdir = req;
 452
 453                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 454                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 455                         if (dfi->readdir_cache_idx < 0) {
 456                                 /* preclude from marking dir ordered */
 457                                 dfi->dir_ordered_count = 0;
 458                         } else if (ceph_frag_is_leftmost(frag) &&
 459                                    dfi->next_offset == 2) {
 460                                 /* note dir version at start of readdir so
 461                                  * we can tell if any dentries get dropped */
 462                                 dfi->dir_release_count = req->r_dir_release_cnt;
 463                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 464                         }
 465                 } else {
 466                         dout("readdir !did_prepopulate\n");
 467                         /* disable readdir cache */
 468                         dfi->readdir_cache_idx = -1;
 469                         /* preclude from marking dir complete */
 470                         dfi->dir_release_count = 0;
 471                 }
 472
 473                 /* note next offset and last dentry name */
 474                 if (rinfo->dir_nr > 0) {
 475                         struct ceph_mds_reply_dir_entry *rde =
 476                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 477                         unsigned next_offset = req->r_reply_info.dir_end ?
 478                                         2 : (fpos_off(rde->offset) + 1);
 479                         err = note_last_dentry(dfi, rde->name, rde->name_len,
 480                                                next_offset);
 481                         if (err) {
 482                                 ceph_mdsc_put_request(dfi->last_readdir);
 483                                 dfi->last_readdir = NULL;
 484                                 return err;
 485                         }
 486                 } else if (req->r_reply_info.dir_end) {
 487                         dfi->next_offset = 2;
 488                         /* keep last name */
 489                 }
 490         }
 491
 492         rinfo = &dfi->last_readdir->r_reply_info;
 493         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 494              dfi->frag, rinfo->dir_nr, ctx->pos,
 495              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 496
 497         i = 0;
 498         /* search start position */
 499         if (rinfo->dir_nr > 0) {
 500                 int step, nr = rinfo->dir_nr;
 501                 while (nr > 0) {
 502                         step = nr >> 1;
 503                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 504                                 i +=  step + 1;
 505                                 nr -= step + 1;
 506                         } else {
 507                                 nr = step;
 508                         }
 509                 }
 510         }
 511         for (; i < rinfo->dir_nr; i++) {
 512                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 513
 514                 BUG_ON(rde->offset < ctx->pos);
 515
 516                 ctx->pos = rde->offset;
 517                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 518                      i, rinfo->dir_nr, ctx->pos,
 519                      rde->name_len, rde->name, &rde->inode.in);
 520
 521                 BUG_ON(!rde->inode.in);
 522
 523                 if (!dir_emit(ctx, rde->name, rde->name_len,
 524                               ceph_present_ino(inode->i_sb, le64_to_cpu(rde->inode.in->ino)),
 525                               le32_to_cpu(rde->inode.in->mode) >> 12)) {
 526                         /*
 527                          * NOTE: Here no need to put the 'dfi->last_readdir',
 528                          * because when dir_emit stops us it's most likely
 529                          * doesn't have enough memory, etc. So for next readdir
 530                          * it will continue.
 531                          */
 532                         dout("filldir stopping us...\n");
 533                         return 0;
 534                 }
 535                 ctx->pos++;
 536         }
 537
 538         ceph_mdsc_put_request(dfi->last_readdir);
 539         dfi->last_readdir = NULL;
 540
 541         if (dfi->next_offset > 2) {
 542                 frag = dfi->frag;
 543                 goto more;
 544         }
 545
 546         /* more frags? */
 547         if (!ceph_frag_is_rightmost(dfi->frag)) {
 548                 frag = ceph_frag_next(dfi->frag);
 549                 if (is_hash_order(ctx->pos)) {
 550                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 551                                                         dfi->next_offset, true);
 552                         if (new_pos > ctx->pos)
 553                                 ctx->pos = new_pos;
 554                         /* keep last_name */
 555                 } else {
 556                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 557                                                         false);
 558                         kfree(dfi->last_name);
 559                         dfi->last_name = NULL;
 560                 }
 561                 dout("readdir next frag is %x\n", frag);
 562                 goto more;
 563         }
 564         dfi->file_info.flags |= CEPH_F_ATEND;
 565
 566         /*
 567          * if dir_release_count still matches the dir, no dentries
 568          * were released during the whole readdir, and we should have
 569          * the complete dir contents in our cache.
 570          */
 571         if (atomic64_read(&ci->i_release_count) ==
 572                                         dfi->dir_release_count) {
 573                 spin_lock(&ci->i_ceph_lock);
 574                 if (dfi->dir_ordered_count ==
 575                                 atomic64_read(&ci->i_ordered_count)) {
 576                         dout(" marking %p complete and ordered\n", inode);
 577                         /* use i_size to track number of entries in
 578                          * readdir cache */
 579                         BUG_ON(dfi->readdir_cache_idx < 0);
 580                         i_size_write(inode, dfi->readdir_cache_idx *
 581                                      sizeof(struct dentry*));
 582                 } else {
 583                         dout(" marking %p complete\n", inode);
 584                 }
 585                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 586                                         dfi->dir_ordered_count);
 587                 spin_unlock(&ci->i_ceph_lock);
 588         }
 589
 590         dout("readdir %p file %p done.\n", inode, file);
 591         return 0;
 592 }
 593
 594 static void reset_readdir(struct ceph_dir_file_info *dfi)
 595 {
 596         if (dfi->last_readdir) {
 597                 ceph_mdsc_put_request(dfi->last_readdir);
 598                 dfi->last_readdir = NULL;
 599         }
 600         kfree(dfi->last_name);
 601         dfi->last_name = NULL;
 602         dfi->dir_release_count = 0;
 603         dfi->readdir_cache_idx = -1;
 604         dfi->next_offset = 2;  /* compensate for . and .. */
 605         dfi->file_info.flags &= ~CEPH_F_ATEND;
 606 }
 607
 608 /*
 609  * discard buffered readdir content on seekdir(0), or seek to new frag,
 610  * or seek prior to current chunk
 611  */
 612 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 613 {
 614         struct ceph_mds_reply_info_parsed *rinfo;
 615         loff_t chunk_offset;
 616         if (new_pos == 0)
 617                 return true;
 618         if (is_hash_order(new_pos)) {
 619                 /* no need to reset last_name for a forward seek when
 620                  * dentries are sotred in hash order */
 621         } else if (dfi->frag != fpos_frag(new_pos)) {
 622                 return true;
 623         }
 624         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 625         if (!rinfo || !rinfo->dir_nr)
 626                 return true;
 627         chunk_offset = rinfo->dir_entries[0].offset;
 628         return new_pos < chunk_offset ||
 629                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 630 }
 631
 632 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 633 {
 634         struct ceph_dir_file_info *dfi = file->private_data;
 635         struct inode *inode = file->f_mapping->host;
 636         loff_t retval;
 637
 638         inode_lock(inode);
 639         retval = -EINVAL;
 640         switch (whence) {
 641         case SEEK_CUR:
 642                 offset += file->f_pos;
 643                 break;
 644         case SEEK_SET:
 645                 break;
 646         case SEEK_END:
 647                 retval = -EOPNOTSUPP;
 648                 goto out;
 649         default:
 650                 goto out;
 651         }
 652
 653         if (offset >= 0) {
 654                 if (need_reset_readdir(dfi, offset)) {
 655                         dout("dir_llseek dropping %p content\n", file);
 656                         reset_readdir(dfi);
 657                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 658                         /* for hash offset, we don't know if a forward seek
 659                          * is within same frag */
 660                         dfi->dir_release_count = 0;
 661                         dfi->readdir_cache_idx = -1;
 662                 }
 663
 664                 if (offset != file->f_pos) {
 665                         file->f_pos = offset;
 666                         file->f_version = 0;
 667                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 668                 }
 669                 retval = offset;
 670         }
 671 out:
 672         inode_unlock(inode);
 673         return retval;
 674 }
 675
 676 /*
 677  * Handle lookups for the hidden .snap directory.
 678  */
 679 struct dentry *ceph_handle_snapdir(struct ceph_mds_request *req,
 680                                    struct dentry *dentry)
 681 {
 682         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 683         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_rwsem */
 684
 685         /* .snap dir? */
 686         if (ceph_snap(parent) == CEPH_NOSNAP &&
 687             strcmp(dentry->d_name.name, fsc->mount_options->snapdir_name) == 0) {
 688                 struct dentry *res;
 689                 struct inode *inode = ceph_get_snapdir(parent);
 690
 691                 res = d_splice_alias(inode, dentry);
 692                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p. Spliced dentry %p\n",
 693                      dentry, dentry, inode, res);
 694                 if (res)
 695                         dentry = res;
 696         }
 697         return dentry;
 698 }
 699
 700 /*
 701  * Figure out final result of a lookup/open request.
 702  *
 703  * Mainly, make sure we return the final req->r_dentry (if it already
 704  * existed) in place of the original VFS-provided dentry when they
 705  * differ.
 706  *
 707  * Gracefully handle the case where the MDS replies with -ENOENT and
 708  * no trace (which it may do, at its discretion, e.g., if it doesn't
 709  * care to issue a lease on the negative dentry).
 710  */
 711 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 712                                   struct dentry *dentry, int err)
 713 {
 714         if (err == -ENOENT) {
 715                 /* no trace? */
 716                 err = 0;
 717                 if (!req->r_reply_info.head->is_dentry) {
 718                         dout("ENOENT and no trace, dentry %p inode %p\n",
 719                              dentry, d_inode(dentry));
 720                         if (d_really_is_positive(dentry)) {
 721                                 d_drop(dentry);
 722                                 err = -ENOENT;
 723                         } else {
 724                                 d_add(dentry, NULL);
 725                         }
 726                 }
 727         }
 728         if (err)
 729                 dentry = ERR_PTR(err);
 730         else if (dentry != req->r_dentry)
 731                 dentry = dget(req->r_dentry);   /* we got spliced */
 732         else
 733                 dentry = NULL;
 734         return dentry;
 735 }
 736
 737 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 738 {
 739         return ceph_ino(inode) == CEPH_INO_ROOT &&
 740                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 741 }
 742
 743 /*
 744  * Look up a single dir entry.  If there is a lookup intent, inform
 745  * the MDS so that it gets our 'caps wanted' value in a single op.
 746  */
 747 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 748                                   unsigned int flags)
 749 {
 750         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 751         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 752         struct ceph_mds_request *req;
 753         int op;
 754         int mask;
 755         int err;
 756
 757         dout("lookup %p dentry %p '%pd'\n",
 758              dir, dentry, dentry);
 759
 760         if (dentry->d_name.len > NAME_MAX)
 761                 return ERR_PTR(-ENAMETOOLONG);
 762
 763         /* can we conclude ENOENT locally? */
 764         if (d_really_is_negative(dentry)) {
 765                 struct ceph_inode_info *ci = ceph_inode(dir);
 766                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 767
 768                 spin_lock(&ci->i_ceph_lock);
 769                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 770                 if (strncmp(dentry->d_name.name,
 771                             fsc->mount_options->snapdir_name,
 772                             dentry->d_name.len) &&
 773                     !is_root_ceph_dentry(dir, dentry) &&
 774                     ceph_test_mount_opt(fsc, DCACHE) &&
 775                     __ceph_dir_is_complete(ci) &&
 776                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 777                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 778                         spin_unlock(&ci->i_ceph_lock);
 779                         dout(" dir %p complete, -ENOENT\n", dir);
 780                         d_add(dentry, NULL);
 781                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 782                         return NULL;
 783                 }
 784                 spin_unlock(&ci->i_ceph_lock);
 785         }
 786
 787         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 788                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 789         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 790         if (IS_ERR(req))
 791                 return ERR_CAST(req);
 792         req->r_dentry = dget(dentry);
 793         req->r_num_caps = 2;
 794
 795         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 796         if (ceph_security_xattr_wanted(dir))
 797                 mask |= CEPH_CAP_XATTR_SHARED;
 798         req->r_args.getattr.mask = cpu_to_le32(mask);
 799
 800         ihold(dir);
 801         req->r_parent = dir;
 802         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 803         err = ceph_mdsc_do_request(mdsc, NULL, req);
 804         if (err == -ENOENT) {
 805                 struct dentry *res;
 806
 807                 res = ceph_handle_snapdir(req, dentry);
 808                 if (IS_ERR(res)) {
 809                         err = PTR_ERR(res);
 810                 } else {
 811                         dentry = res;
 812                         err = 0;
 813                 }
 814         }
 815         dentry = ceph_finish_lookup(req, dentry, err);
 816         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 817         dout("lookup result=%p\n", dentry);
 818         return dentry;
 819 }
 820
 821 /*
 822  * If we do a create but get no trace back from the MDS, follow up with
 823  * a lookup (the VFS expects us to link up the provided dentry).
 824  */
 825 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 826 {
 827         struct dentry *result = ceph_lookup(dir, dentry, 0);
 828
 829         if (result && !IS_ERR(result)) {
 830                 /*
 831                  * We created the item, then did a lookup, and found
 832                  * it was already linked to another inode we already
 833                  * had in our cache (and thus got spliced). To not
 834                  * confuse VFS (especially when inode is a directory),
 835                  * we don't link our dentry to that inode, return an
 836                  * error instead.
 837                  *
 838                  * This event should be rare and it happens only when
 839                  * we talk to old MDS. Recent MDS does not send traceless
 840                  * reply for request that creates new inode.
 841                  */
 842                 d_drop(result);
 843                 return -ESTALE;
 844         }
 845         return PTR_ERR(result);
 846 }
 847
 848 static int ceph_mknod(struct mnt_idmap *idmap, struct inode *dir,
 849                       struct dentry *dentry, umode_t mode, dev_t rdev)
 850 {
 851         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 852         struct ceph_mds_request *req;
 853         struct ceph_acl_sec_ctx as_ctx = {};
 854         int err;
 855
 856         if (ceph_snap(dir) != CEPH_NOSNAP)
 857                 return -EROFS;
 858
 859         err = ceph_wait_on_conflict_unlink(dentry);
 860         if (err)
 861                 return err;
 862
 863         if (ceph_quota_is_max_files_exceeded(dir)) {
 864                 err = -EDQUOT;
 865                 goto out;
 866         }
 867
 868         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 869         if (err < 0)
 870                 goto out;
 871         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 872         if (err < 0)
 873                 goto out;
 874
 875         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 876              dir, dentry, mode, rdev);
 877         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 878         if (IS_ERR(req)) {
 879                 err = PTR_ERR(req);
 880                 goto out;
 881         }
 882         req->r_dentry = dget(dentry);
 883         req->r_num_caps = 2;
 884         req->r_parent = dir;
 885         ihold(dir);
 886         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 887         req->r_args.mknod.mode = cpu_to_le32(mode);
 888         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 889         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 890         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 891         if (as_ctx.pagelist) {
 892                 req->r_pagelist = as_ctx.pagelist;
 893                 as_ctx.pagelist = NULL;
 894         }
 895         err = ceph_mdsc_do_request(mdsc, dir, req);
 896         if (!err && !req->r_reply_info.head->is_dentry)
 897                 err = ceph_handle_notrace_create(dir, dentry);
 898         ceph_mdsc_put_request(req);
 899 out:
 900         if (!err)
 901                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 902         else
 903                 d_drop(dentry);
 904         ceph_release_acl_sec_ctx(&as_ctx);
 905         return err;
 906 }
 907
 908 static int ceph_create(struct mnt_idmap *idmap, struct inode *dir,
 909                        struct dentry *dentry, umode_t mode, bool excl)
 910 {
 911         return ceph_mknod(idmap, dir, dentry, mode, 0);
 912 }
 913
 914 static int ceph_symlink(struct mnt_idmap *idmap, struct inode *dir,
 915                         struct dentry *dentry, const char *dest)
 916 {
 917         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 918         struct ceph_mds_request *req;
 919         struct ceph_acl_sec_ctx as_ctx = {};
 920         int err;
 921
 922         if (ceph_snap(dir) != CEPH_NOSNAP)
 923                 return -EROFS;
 924
 925         err = ceph_wait_on_conflict_unlink(dentry);
 926         if (err)
 927                 return err;
 928
 929         if (ceph_quota_is_max_files_exceeded(dir)) {
 930                 err = -EDQUOT;
 931                 goto out;
 932         }
 933
 934         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 935         if (err < 0)
 936                 goto out;
 937
 938         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 939         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 940         if (IS_ERR(req)) {
 941                 err = PTR_ERR(req);
 942                 goto out;
 943         }
 944         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 945         if (!req->r_path2) {
 946                 err = -ENOMEM;
 947                 ceph_mdsc_put_request(req);
 948                 goto out;
 949         }
 950         req->r_parent = dir;
 951         ihold(dir);
 952
 953         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 954         req->r_dentry = dget(dentry);
 955         req->r_num_caps = 2;
 956         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 957         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 958         if (as_ctx.pagelist) {
 959                 req->r_pagelist = as_ctx.pagelist;
 960                 as_ctx.pagelist = NULL;
 961         }
 962         err = ceph_mdsc_do_request(mdsc, dir, req);
 963         if (!err && !req->r_reply_info.head->is_dentry)
 964                 err = ceph_handle_notrace_create(dir, dentry);
 965         ceph_mdsc_put_request(req);
 966 out:
 967         if (err)
 968                 d_drop(dentry);
 969         ceph_release_acl_sec_ctx(&as_ctx);
 970         return err;
 971 }
 972
 973 static int ceph_mkdir(struct mnt_idmap *idmap, struct inode *dir,
 974                       struct dentry *dentry, umode_t mode)
 975 {
 976         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 977         struct ceph_mds_request *req;
 978         struct ceph_acl_sec_ctx as_ctx = {};
 979         int err;
 980         int op;
 981
 982         err = ceph_wait_on_conflict_unlink(dentry);
 983         if (err)
 984                 return err;
 985
 986         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 987                 /* mkdir .snap/foo is a MKSNAP */
 988                 op = CEPH_MDS_OP_MKSNAP;
 989                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 990                      dentry, dentry);
 991         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 992                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 993                 op = CEPH_MDS_OP_MKDIR;
 994         } else {
 995                 err = -EROFS;
 996                 goto out;
 997         }
 998
 999         if (op == CEPH_MDS_OP_MKDIR &&
1000             ceph_quota_is_max_files_exceeded(dir)) {
1001                 err = -EDQUOT;
1002                 goto out;
1003         }
1004
1005         mode |= S_IFDIR;
1006         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
1007         if (err < 0)
1008                 goto out;
1009         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
1010         if (err < 0)
1011                 goto out;
1012
1013         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1014         if (IS_ERR(req)) {
1015                 err = PTR_ERR(req);
1016                 goto out;
1017         }
1018
1019         req->r_dentry = dget(dentry);
1020         req->r_num_caps = 2;
1021         req->r_parent = dir;
1022         ihold(dir);
1023         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1024         req->r_args.mkdir.mode = cpu_to_le32(mode);
1025         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
1026         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1027         if (as_ctx.pagelist) {
1028                 req->r_pagelist = as_ctx.pagelist;
1029                 as_ctx.pagelist = NULL;
1030         }
1031         err = ceph_mdsc_do_request(mdsc, dir, req);
1032         if (!err &&
1033             !req->r_reply_info.head->is_target &&
1034             !req->r_reply_info.head->is_dentry)
1035                 err = ceph_handle_notrace_create(dir, dentry);
1036         ceph_mdsc_put_request(req);
1037 out:
1038         if (!err)
1039                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1040         else
1041                 d_drop(dentry);
1042         ceph_release_acl_sec_ctx(&as_ctx);
1043         return err;
1044 }
1045
1046 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1047                      struct dentry *dentry)
1048 {
1049         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1050         struct ceph_mds_request *req;
1051         int err;
1052
1053         if (dentry->d_flags & DCACHE_DISCONNECTED)
1054                 return -EINVAL;
1055
1056         err = ceph_wait_on_conflict_unlink(dentry);
1057         if (err)
1058                 return err;
1059
1060         if (ceph_snap(dir) != CEPH_NOSNAP)
1061                 return -EROFS;
1062
1063         dout("link in dir %p %llx.%llx old_dentry %p:'%pd' dentry %p:'%pd'\n",
1064              dir, ceph_vinop(dir), old_dentry, old_dentry, dentry, dentry);
1065         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1066         if (IS_ERR(req)) {
1067                 d_drop(dentry);
1068                 return PTR_ERR(req);
1069         }
1070         req->r_dentry = dget(dentry);
1071         req->r_num_caps = 2;
1072         req->r_old_dentry = dget(old_dentry);
1073         /*
1074          * The old_dentry maybe a DCACHE_DISCONNECTED dentry, then we
1075          * will just pass the ino# to MDSs.
1076          */
1077         if (old_dentry->d_flags & DCACHE_DISCONNECTED)
1078                 req->r_ino2 = ceph_vino(d_inode(old_dentry));
1079         req->r_parent = dir;
1080         ihold(dir);
1081         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1082         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1083         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1084         /* release LINK_SHARED on source inode (mds will lock it) */
1085         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1086         err = ceph_mdsc_do_request(mdsc, dir, req);
1087         if (err) {
1088                 d_drop(dentry);
1089         } else if (!req->r_reply_info.head->is_dentry) {
1090                 ihold(d_inode(old_dentry));
1091                 d_instantiate(dentry, d_inode(old_dentry));
1092         }
1093         ceph_mdsc_put_request(req);
1094         return err;
1095 }
1096
1097 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1098                                  struct ceph_mds_request *req)
1099 {
1100         struct dentry *dentry = req->r_dentry;
1101         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1102         struct ceph_dentry_info *di = ceph_dentry(dentry);
1103         int result = req->r_err ? req->r_err :
1104                         le32_to_cpu(req->r_reply_info.head->result);
1105
1106         if (!test_bit(CEPH_DENTRY_ASYNC_UNLINK_BIT, &di->flags))
1107                 pr_warn("%s dentry %p:%pd async unlink bit is not set\n",
1108                         __func__, dentry, dentry);
1109
1110         spin_lock(&fsc->async_unlink_conflict_lock);
1111         hash_del_rcu(&di->hnode);
1112         spin_unlock(&fsc->async_unlink_conflict_lock);
1113
1114         spin_lock(&dentry->d_lock);
1115         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1116         wake_up_bit(&di->flags, CEPH_DENTRY_ASYNC_UNLINK_BIT);
1117         spin_unlock(&dentry->d_lock);
1118
1119         synchronize_rcu();
1120
1121         if (result == -EJUKEBOX)
1122                 goto out;
1123
1124         /* If op failed, mark everyone involved for errors */
1125         if (result) {
1126                 int pathlen = 0;
1127                 u64 base = 0;
1128                 char *path = ceph_mdsc_build_path(dentry, &pathlen,
1129                                                   &base, 0);
1130
1131                 /* mark error on parent + clear complete */
1132                 mapping_set_error(req->r_parent->i_mapping, result);
1133                 ceph_dir_clear_complete(req->r_parent);
1134
1135                 /* drop the dentry -- we don't know its status */
1136                 if (!d_unhashed(dentry))
1137                         d_drop(dentry);
1138
1139                 /* mark inode itself for an error (since metadata is bogus) */
1140                 mapping_set_error(req->r_old_inode->i_mapping, result);
1141
1142                 pr_warn("async unlink failure path=(%llx)%s result=%d!\n",
1143                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1144                 ceph_mdsc_free_path(path, pathlen);
1145         }
1146 out:
1147         iput(req->r_old_inode);
1148         ceph_mdsc_release_dir_caps(req);
1149 }
1150
1151 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1152 {
1153         struct ceph_inode_info *ci = ceph_inode(dir);
1154         struct ceph_dentry_info *di;
1155         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1156
1157         spin_lock(&ci->i_ceph_lock);
1158         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1159                 ceph_take_cap_refs(ci, want, false);
1160                 got = want;
1161         }
1162         spin_unlock(&ci->i_ceph_lock);
1163
1164         /* If we didn't get anything, return 0 */
1165         if (!got)
1166                 return 0;
1167
1168         spin_lock(&dentry->d_lock);
1169         di = ceph_dentry(dentry);
1170         /*
1171          * - We are holding Fx, which implies Fs caps.
1172          * - Only support async unlink for primary linkage
1173          */
1174         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1175             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1176                 want = 0;
1177         spin_unlock(&dentry->d_lock);
1178
1179         /* Do we still want what we've got? */
1180         if (want == got)
1181                 return got;
1182
1183         ceph_put_cap_refs(ci, got);
1184         return 0;
1185 }
1186
1187 /*
1188  * rmdir and unlink are differ only by the metadata op code
1189  */
1190 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1191 {
1192         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1193         struct ceph_mds_client *mdsc = fsc->mdsc;
1194         struct inode *inode = d_inode(dentry);
1195         struct ceph_mds_request *req;
1196         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1197         int err = -EROFS;
1198         int op;
1199
1200         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1201                 /* rmdir .snap/foo is RMSNAP */
1202                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1203                 op = CEPH_MDS_OP_RMSNAP;
1204         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1205                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1206                      dir, dentry, inode);
1207                 op = d_is_dir(dentry) ?
1208                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1209         } else
1210                 goto out;
1211 retry:
1212         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1213         if (IS_ERR(req)) {
1214                 err = PTR_ERR(req);
1215                 goto out;
1216         }
1217         req->r_dentry = dget(dentry);
1218         req->r_num_caps = 2;
1219         req->r_parent = dir;
1220         ihold(dir);
1221         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1222         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1223         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1224
1225         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1226             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1227                 struct ceph_dentry_info *di = ceph_dentry(dentry);
1228
1229                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1230                      dentry->d_name.len, dentry->d_name.name,
1231                      ceph_cap_string(req->r_dir_caps));
1232                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1233                 req->r_callback = ceph_async_unlink_cb;
1234                 req->r_old_inode = d_inode(dentry);
1235                 ihold(req->r_old_inode);
1236
1237                 spin_lock(&dentry->d_lock);
1238                 di->flags |= CEPH_DENTRY_ASYNC_UNLINK;
1239                 spin_unlock(&dentry->d_lock);
1240
1241                 spin_lock(&fsc->async_unlink_conflict_lock);
1242                 hash_add_rcu(fsc->async_unlink_conflict, &di->hnode,
1243                              dentry->d_name.hash);
1244                 spin_unlock(&fsc->async_unlink_conflict_lock);
1245
1246                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1247                 if (!err) {
1248                         /*
1249                          * We have enough caps, so we assume that the unlink
1250                          * will succeed. Fix up the target inode and dcache.
1251                          */
1252                         drop_nlink(inode);
1253                         d_delete(dentry);
1254                 } else {
1255                         spin_lock(&fsc->async_unlink_conflict_lock);
1256                         hash_del_rcu(&di->hnode);
1257                         spin_unlock(&fsc->async_unlink_conflict_lock);
1258
1259                         spin_lock(&dentry->d_lock);
1260                         di->flags &= ~CEPH_DENTRY_ASYNC_UNLINK;
1261                         spin_unlock(&dentry->d_lock);
1262
1263                         if (err == -EJUKEBOX) {
1264                                 try_async = false;
1265                                 ceph_mdsc_put_request(req);
1266                                 goto retry;
1267                         }
1268                 }
1269         } else {
1270                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1271                 err = ceph_mdsc_do_request(mdsc, dir, req);
1272                 if (!err && !req->r_reply_info.head->is_dentry)
1273                         d_delete(dentry);
1274         }
1275
1276         ceph_mdsc_put_request(req);
1277 out:
1278         return err;
1279 }
1280
1281 static int ceph_rename(struct mnt_idmap *idmap, struct inode *old_dir,
1282                        struct dentry *old_dentry, struct inode *new_dir,
1283                        struct dentry *new_dentry, unsigned int flags)
1284 {
1285         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1286         struct ceph_mds_request *req;
1287         int op = CEPH_MDS_OP_RENAME;
1288         int err;
1289
1290         if (flags)
1291                 return -EINVAL;
1292
1293         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1294                 return -EXDEV;
1295         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1296                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1297                         op = CEPH_MDS_OP_RENAMESNAP;
1298                 else
1299                         return -EROFS;
1300         }
1301         /* don't allow cross-quota renames */
1302         if ((old_dir != new_dir) &&
1303             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1304                 return -EXDEV;
1305
1306         err = ceph_wait_on_conflict_unlink(new_dentry);
1307         if (err)
1308                 return err;
1309
1310         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1311              old_dir, old_dentry, new_dir, new_dentry);
1312         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1313         if (IS_ERR(req))
1314                 return PTR_ERR(req);
1315         ihold(old_dir);
1316         req->r_dentry = dget(new_dentry);
1317         req->r_num_caps = 2;
1318         req->r_old_dentry = dget(old_dentry);
1319         req->r_old_dentry_dir = old_dir;
1320         req->r_parent = new_dir;
1321         ihold(new_dir);
1322         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1323         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1324         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1325         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1326         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1327         /* release LINK_RDCACHE on source inode (mds will lock it) */
1328         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1329         if (d_really_is_positive(new_dentry)) {
1330                 req->r_inode_drop =
1331                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1332         }
1333         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1334         if (!err && !req->r_reply_info.head->is_dentry) {
1335                 /*
1336                  * Normally d_move() is done by fill_trace (called by
1337                  * do_request, above).  If there is no trace, we need
1338                  * to do it here.
1339                  */
1340                 d_move(old_dentry, new_dentry);
1341         }
1342         ceph_mdsc_put_request(req);
1343         return err;
1344 }
1345
1346 /*
1347  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1348  * Leases at front of the list will expire first. (Assume all leases have
1349  * similar duration)
1350  *
1351  * Called under dentry->d_lock.
1352  */
1353 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1354 {
1355         struct dentry *dn = di->dentry;
1356         struct ceph_mds_client *mdsc;
1357
1358         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1359
1360         di->flags |= CEPH_DENTRY_LEASE_LIST;
1361         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1362                 di->flags |= CEPH_DENTRY_REFERENCED;
1363                 return;
1364         }
1365
1366         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1367         spin_lock(&mdsc->dentry_list_lock);
1368         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1369         spin_unlock(&mdsc->dentry_list_lock);
1370 }
1371
1372 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1373                                      struct ceph_dentry_info *di)
1374 {
1375         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1376         di->lease_gen = 0;
1377         di->time = jiffies;
1378         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1379 }
1380
1381 /*
1382  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1383  * list if it's not in the list, otherwise set 'referenced' flag.
1384  *
1385  * Called under dentry->d_lock.
1386  */
1387 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1388 {
1389         struct dentry *dn = di->dentry;
1390         struct ceph_mds_client *mdsc;
1391
1392         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1393              di, dn, dn, di->offset);
1394
1395         if (!list_empty(&di->lease_list)) {
1396                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1397                         /* don't remove dentry from dentry lease list
1398                          * if its lease is valid */
1399                         if (__dentry_lease_is_valid(di))
1400                                 return;
1401                 } else {
1402                         di->flags |= CEPH_DENTRY_REFERENCED;
1403                         return;
1404                 }
1405         }
1406
1407         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1408                 di->flags |= CEPH_DENTRY_REFERENCED;
1409                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1410                 return;
1411         }
1412
1413         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1414         spin_lock(&mdsc->dentry_list_lock);
1415         __dentry_dir_lease_touch(mdsc, di),
1416         spin_unlock(&mdsc->dentry_list_lock);
1417 }
1418
1419 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1420 {
1421         struct ceph_mds_client *mdsc;
1422         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1423                 return;
1424         if (list_empty(&di->lease_list))
1425                 return;
1426
1427         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1428         spin_lock(&mdsc->dentry_list_lock);
1429         list_del_init(&di->lease_list);
1430         spin_unlock(&mdsc->dentry_list_lock);
1431 }
1432
1433 enum {
1434         KEEP    = 0,
1435         DELETE  = 1,
1436         TOUCH   = 2,
1437         STOP    = 4,
1438 };
1439
1440 struct ceph_lease_walk_control {
1441         bool dir_lease;
1442         bool expire_dir_lease;
1443         unsigned long nr_to_scan;
1444         unsigned long dir_lease_ttl;
1445 };
1446
1447 static unsigned long
1448 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1449                      struct ceph_lease_walk_control *lwc,
1450                      int (*check)(struct dentry*, void*))
1451 {
1452         struct ceph_dentry_info *di, *tmp;
1453         struct dentry *dentry, *last = NULL;
1454         struct list_head* list;
1455         LIST_HEAD(dispose);
1456         unsigned long freed = 0;
1457         int ret = 0;
1458
1459         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1460         spin_lock(&mdsc->dentry_list_lock);
1461         list_for_each_entry_safe(di, tmp, list, lease_list) {
1462                 if (!lwc->nr_to_scan)
1463                         break;
1464                 --lwc->nr_to_scan;
1465
1466                 dentry = di->dentry;
1467                 if (last == dentry)
1468                         break;
1469
1470                 if (!spin_trylock(&dentry->d_lock))
1471                         continue;
1472
1473                 if (__lockref_is_dead(&dentry->d_lockref)) {
1474                         list_del_init(&di->lease_list);
1475                         goto next;
1476                 }
1477
1478                 ret = check(dentry, lwc);
1479                 if (ret & TOUCH) {
1480                         /* move it into tail of dir lease list */
1481                         __dentry_dir_lease_touch(mdsc, di);
1482                         if (!last)
1483                                 last = dentry;
1484                 }
1485                 if (ret & DELETE) {
1486                         /* stale lease */
1487                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1488                         if (dentry->d_lockref.count > 0) {
1489                                 /* update_dentry_lease() will re-add
1490                                  * it to lease list, or
1491                                  * ceph_d_delete() will return 1 when
1492                                  * last reference is dropped */
1493                                 list_del_init(&di->lease_list);
1494                         } else {
1495                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1496                                 list_move_tail(&di->lease_list, &dispose);
1497                                 dget_dlock(dentry);
1498                         }
1499                 }
1500 next:
1501                 spin_unlock(&dentry->d_lock);
1502                 if (ret & STOP)
1503                         break;
1504         }
1505         spin_unlock(&mdsc->dentry_list_lock);
1506
1507         while (!list_empty(&dispose)) {
1508                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1509                                       lease_list);
1510                 dentry = di->dentry;
1511                 spin_lock(&dentry->d_lock);
1512
1513                 list_del_init(&di->lease_list);
1514                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1515                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1516                         spin_lock(&mdsc->dentry_list_lock);
1517                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1518                                 list_add_tail(&di->lease_list,
1519                                               &mdsc->dentry_leases);
1520                         } else {
1521                                 __dentry_dir_lease_touch(mdsc, di);
1522                         }
1523                         spin_unlock(&mdsc->dentry_list_lock);
1524                 } else {
1525                         freed++;
1526                 }
1527
1528                 spin_unlock(&dentry->d_lock);
1529                 /* ceph_d_delete() does the trick */
1530                 dput(dentry);
1531         }
1532         return freed;
1533 }
1534
1535 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1536 {
1537         struct ceph_dentry_info *di = ceph_dentry(dentry);
1538         int ret;
1539
1540         if (__dentry_lease_is_valid(di))
1541                 return STOP;
1542         ret = __dir_lease_try_check(dentry);
1543         if (ret == -EBUSY)
1544                 return KEEP;
1545         if (ret > 0)
1546                 return TOUCH;
1547         return DELETE;
1548 }
1549
1550 static int __dir_lease_check(struct dentry *dentry, void *arg)
1551 {
1552         struct ceph_lease_walk_control *lwc = arg;
1553         struct ceph_dentry_info *di = ceph_dentry(dentry);
1554
1555         int ret = __dir_lease_try_check(dentry);
1556         if (ret == -EBUSY)
1557                 return KEEP;
1558         if (ret > 0) {
1559                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1560                         return STOP;
1561                 /* Move dentry to tail of dir lease list if we don't want
1562                  * to delete it. So dentries in the list are checked in a
1563                  * round robin manner */
1564                 if (!lwc->expire_dir_lease)
1565                         return TOUCH;
1566                 if (dentry->d_lockref.count > 0 ||
1567                     (di->flags & CEPH_DENTRY_REFERENCED))
1568                         return TOUCH;
1569                 /* invalidate dir lease */
1570                 di->lease_shared_gen = 0;
1571         }
1572         return DELETE;
1573 }
1574
1575 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1576 {
1577         struct ceph_lease_walk_control lwc;
1578         unsigned long count;
1579         unsigned long freed;
1580
1581         spin_lock(&mdsc->caps_list_lock);
1582         if (mdsc->caps_use_max > 0 &&
1583             mdsc->caps_use_count > mdsc->caps_use_max)
1584                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1585         else
1586                 count = 0;
1587         spin_unlock(&mdsc->caps_list_lock);
1588
1589         lwc.dir_lease = false;
1590         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1591         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1592         if (!lwc.nr_to_scan) /* more invalid leases */
1593                 return -EAGAIN;
1594
1595         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1596                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1597
1598         lwc.dir_lease = true;
1599         lwc.expire_dir_lease = freed < count;
1600         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1601         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1602         if (!lwc.nr_to_scan) /* more to check */
1603                 return -EAGAIN;
1604
1605         return freed > 0 ? 1 : 0;
1606 }
1607
1608 /*
1609  * Ensure a dentry lease will no longer revalidate.
1610  */
1611 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1612 {
1613         struct ceph_dentry_info *di = ceph_dentry(dentry);
1614         spin_lock(&dentry->d_lock);
1615         di->time = jiffies;
1616         di->lease_shared_gen = 0;
1617         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1618         __dentry_lease_unlist(di);
1619         spin_unlock(&dentry->d_lock);
1620 }
1621
1622 /*
1623  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1624  * renew if the least is more than half up.
1625  */
1626 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1627 {
1628         struct ceph_mds_session *session;
1629
1630         if (!di->lease_gen)
1631                 return false;
1632
1633         session = di->lease_session;
1634         if (session) {
1635                 u32 gen;
1636                 unsigned long ttl;
1637
1638                 gen = atomic_read(&session->s_cap_gen);
1639                 ttl = session->s_cap_ttl;
1640
1641                 if (di->lease_gen == gen &&
1642                     time_before(jiffies, ttl) &&
1643                     time_before(jiffies, di->time))
1644                         return true;
1645         }
1646         di->lease_gen = 0;
1647         return false;
1648 }
1649
1650 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1651 {
1652         struct ceph_dentry_info *di;
1653         struct ceph_mds_session *session = NULL;
1654         u32 seq = 0;
1655         int valid = 0;
1656
1657         spin_lock(&dentry->d_lock);
1658         di = ceph_dentry(dentry);
1659         if (di && __dentry_lease_is_valid(di)) {
1660                 valid = 1;
1661
1662                 if (di->lease_renew_after &&
1663                     time_after(jiffies, di->lease_renew_after)) {
1664                         /*
1665                          * We should renew. If we're in RCU walk mode
1666                          * though, we can't do that so just return
1667                          * -ECHILD.
1668                          */
1669                         if (flags & LOOKUP_RCU) {
1670                                 valid = -ECHILD;
1671                         } else {
1672                                 session = ceph_get_mds_session(di->lease_session);
1673                                 seq = di->lease_seq;
1674                                 di->lease_renew_after = 0;
1675                                 di->lease_renew_from = jiffies;
1676                         }
1677                 }
1678         }
1679         spin_unlock(&dentry->d_lock);
1680
1681         if (session) {
1682                 ceph_mdsc_lease_send_msg(session, dentry,
1683                                          CEPH_MDS_LEASE_RENEW, seq);
1684                 ceph_put_mds_session(session);
1685         }
1686         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1687         return valid;
1688 }
1689
1690 /*
1691  * Called under dentry->d_lock.
1692  */
1693 static int __dir_lease_try_check(const struct dentry *dentry)
1694 {
1695         struct ceph_dentry_info *di = ceph_dentry(dentry);
1696         struct inode *dir;
1697         struct ceph_inode_info *ci;
1698         int valid = 0;
1699
1700         if (!di->lease_shared_gen)
1701                 return 0;
1702         if (IS_ROOT(dentry))
1703                 return 0;
1704
1705         dir = d_inode(dentry->d_parent);
1706         ci = ceph_inode(dir);
1707
1708         if (spin_trylock(&ci->i_ceph_lock)) {
1709                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1710                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1711                         valid = 1;
1712                 spin_unlock(&ci->i_ceph_lock);
1713         } else {
1714                 valid = -EBUSY;
1715         }
1716
1717         if (!valid)
1718                 di->lease_shared_gen = 0;
1719         return valid;
1720 }
1721
1722 /*
1723  * Check if directory-wide content lease/cap is valid.
1724  */
1725 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1726                               struct ceph_mds_client *mdsc)
1727 {
1728         struct ceph_inode_info *ci = ceph_inode(dir);
1729         int valid;
1730         int shared_gen;
1731
1732         spin_lock(&ci->i_ceph_lock);
1733         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1734         if (valid) {
1735                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1736                 shared_gen = atomic_read(&ci->i_shared_gen);
1737         }
1738         spin_unlock(&ci->i_ceph_lock);
1739         if (valid) {
1740                 struct ceph_dentry_info *di;
1741                 spin_lock(&dentry->d_lock);
1742                 di = ceph_dentry(dentry);
1743                 if (dir == d_inode(dentry->d_parent) &&
1744                     di && di->lease_shared_gen == shared_gen)
1745                         __ceph_dentry_dir_lease_touch(di);
1746                 else
1747                         valid = 0;
1748                 spin_unlock(&dentry->d_lock);
1749         }
1750         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1751              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1752         return valid;
1753 }
1754
1755 /*
1756  * Check if cached dentry can be trusted.
1757  */
1758 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1759 {
1760         int valid = 0;
1761         struct dentry *parent;
1762         struct inode *dir, *inode;
1763         struct ceph_mds_client *mdsc;
1764
1765         if (flags & LOOKUP_RCU) {
1766                 parent = READ_ONCE(dentry->d_parent);
1767                 dir = d_inode_rcu(parent);
1768                 if (!dir)
1769                         return -ECHILD;
1770                 inode = d_inode_rcu(dentry);
1771         } else {
1772                 parent = dget_parent(dentry);
1773                 dir = d_inode(parent);
1774                 inode = d_inode(dentry);
1775         }
1776
1777         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1778              dentry, inode, ceph_dentry(dentry)->offset);
1779
1780         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1781
1782         /* always trust cached snapped dentries, snapdir dentry */
1783         if (ceph_snap(dir) != CEPH_NOSNAP) {
1784                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1785                      dentry, inode);
1786                 valid = 1;
1787         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1788                 valid = 1;
1789         } else {
1790                 valid = dentry_lease_is_valid(dentry, flags);
1791                 if (valid == -ECHILD)
1792                         return valid;
1793                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1794                         if (inode)
1795                                 valid = ceph_is_any_caps(inode);
1796                         else
1797                                 valid = 1;
1798                 }
1799         }
1800
1801         if (!valid) {
1802                 struct ceph_mds_request *req;
1803                 int op, err;
1804                 u32 mask;
1805
1806                 if (flags & LOOKUP_RCU)
1807                         return -ECHILD;
1808
1809                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1810
1811                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1812                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1813                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1814                 if (!IS_ERR(req)) {
1815                         req->r_dentry = dget(dentry);
1816                         req->r_num_caps = 2;
1817                         req->r_parent = dir;
1818                         ihold(dir);
1819
1820                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1821                         if (ceph_security_xattr_wanted(dir))
1822                                 mask |= CEPH_CAP_XATTR_SHARED;
1823                         req->r_args.getattr.mask = cpu_to_le32(mask);
1824
1825                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1826                         switch (err) {
1827                         case 0:
1828                                 if (d_really_is_positive(dentry) &&
1829                                     d_inode(dentry) == req->r_target_inode)
1830                                         valid = 1;
1831                                 break;
1832                         case -ENOENT:
1833                                 if (d_really_is_negative(dentry))
1834                                         valid = 1;
1835                                 fallthrough;
1836                         default:
1837                                 break;
1838                         }
1839                         ceph_mdsc_put_request(req);
1840                         dout("d_revalidate %p lookup result=%d\n",
1841                              dentry, err);
1842                 }
1843         } else {
1844                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1845         }
1846
1847         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1848         if (!valid)
1849                 ceph_dir_clear_complete(dir);
1850
1851         if (!(flags & LOOKUP_RCU))
1852                 dput(parent);
1853         return valid;
1854 }
1855
1856 /*
1857  * Delete unused dentry that doesn't have valid lease
1858  *
1859  * Called under dentry->d_lock.
1860  */
1861 static int ceph_d_delete(const struct dentry *dentry)
1862 {
1863         struct ceph_dentry_info *di;
1864
1865         /* won't release caps */
1866         if (d_really_is_negative(dentry))
1867                 return 0;
1868         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1869                 return 0;
1870         /* vaild lease? */
1871         di = ceph_dentry(dentry);
1872         if (di) {
1873                 if (__dentry_lease_is_valid(di))
1874                         return 0;
1875                 if (__dir_lease_try_check(dentry))
1876                         return 0;
1877         }
1878         return 1;
1879 }
1880
1881 /*
1882  * Release our ceph_dentry_info.
1883  */
1884 static void ceph_d_release(struct dentry *dentry)
1885 {
1886         struct ceph_dentry_info *di = ceph_dentry(dentry);
1887         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1888
1889         dout("d_release %p\n", dentry);
1890
1891         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1892
1893         spin_lock(&dentry->d_lock);
1894         __dentry_lease_unlist(di);
1895         dentry->d_fsdata = NULL;
1896         spin_unlock(&dentry->d_lock);
1897
1898         ceph_put_mds_session(di->lease_session);
1899         kmem_cache_free(ceph_dentry_cachep, di);
1900 }
1901
1902 /*
1903  * When the VFS prunes a dentry from the cache, we need to clear the
1904  * complete flag on the parent directory.
1905  *
1906  * Called under dentry->d_lock.
1907  */
1908 static void ceph_d_prune(struct dentry *dentry)
1909 {
1910         struct ceph_inode_info *dir_ci;
1911         struct ceph_dentry_info *di;
1912
1913         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1914
1915         /* do we have a valid parent? */
1916         if (IS_ROOT(dentry))
1917                 return;
1918
1919         /* we hold d_lock, so d_parent is stable */
1920         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1921         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1922                 return;
1923
1924         /* who calls d_delete() should also disable dcache readdir */
1925         if (d_really_is_negative(dentry))
1926                 return;
1927
1928         /* d_fsdata does not get cleared until d_release */
1929         if (!d_unhashed(dentry)) {
1930                 __ceph_dir_clear_complete(dir_ci);
1931                 return;
1932         }
1933
1934         /* Disable dcache readdir just in case that someone called d_drop()
1935          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1936          * properly (dcache readdir is still enabled) */
1937         di = ceph_dentry(dentry);
1938         if (di->offset > 0 &&
1939             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1940                 __ceph_dir_clear_ordered(dir_ci);
1941 }
1942
1943 /*
1944  * read() on a dir.  This weird interface hack only works if mounted
1945  * with '-o dirstat'.
1946  */
1947 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1948                              loff_t *ppos)
1949 {
1950         struct ceph_dir_file_info *dfi = file->private_data;
1951         struct inode *inode = file_inode(file);
1952         struct ceph_inode_info *ci = ceph_inode(inode);
1953         int left;
1954         const int bufsize = 1024;
1955
1956         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1957                 return -EISDIR;
1958
1959         if (!dfi->dir_info) {
1960                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1961                 if (!dfi->dir_info)
1962                         return -ENOMEM;
1963                 dfi->dir_info_len =
1964                         snprintf(dfi->dir_info, bufsize,
1965                                 "entries:   %20lld\n"
1966                                 " files:    %20lld\n"
1967                                 " subdirs:  %20lld\n"
1968                                 "rentries:  %20lld\n"
1969                                 " rfiles:   %20lld\n"
1970                                 " rsubdirs: %20lld\n"
1971                                 "rbytes:    %20lld\n"
1972                                 "rctime:    %10lld.%09ld\n",
1973                                 ci->i_files + ci->i_subdirs,
1974                                 ci->i_files,
1975                                 ci->i_subdirs,
1976                                 ci->i_rfiles + ci->i_rsubdirs,
1977                                 ci->i_rfiles,
1978                                 ci->i_rsubdirs,
1979                                 ci->i_rbytes,
1980                                 ci->i_rctime.tv_sec,
1981                                 ci->i_rctime.tv_nsec);
1982         }
1983
1984         if (*ppos >= dfi->dir_info_len)
1985                 return 0;
1986         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1987         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1988         if (left == size)
1989                 return -EFAULT;
1990         *ppos += (size - left);
1991         return size - left;
1992 }
1993
1994
1995
1996 /*
1997  * Return name hash for a given dentry.  This is dependent on
1998  * the parent directory's hash function.
1999  */
2000 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
2001 {
2002         struct ceph_inode_info *dci = ceph_inode(dir);
2003         unsigned hash;
2004
2005         switch (dci->i_dir_layout.dl_dir_hash) {
2006         case 0: /* for backward compat */
2007         case CEPH_STR_HASH_LINUX:
2008                 return dn->d_name.hash;
2009
2010         default:
2011                 spin_lock(&dn->d_lock);
2012                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
2013                                      dn->d_name.name, dn->d_name.len);
2014                 spin_unlock(&dn->d_lock);
2015                 return hash;
2016         }
2017 }
2018
2019 const struct file_operations ceph_dir_fops = {
2020         .read = ceph_read_dir,
2021         .iterate = ceph_readdir,
2022         .llseek = ceph_dir_llseek,
2023         .open = ceph_open,
2024         .release = ceph_release,
2025         .unlocked_ioctl = ceph_ioctl,
2026         .compat_ioctl = compat_ptr_ioctl,
2027         .fsync = ceph_fsync,
2028         .lock = ceph_lock,
2029         .flock = ceph_flock,
2030 };
2031
2032 const struct file_operations ceph_snapdir_fops = {
2033         .iterate = ceph_readdir,
2034         .llseek = ceph_dir_llseek,
2035         .open = ceph_open,
2036         .release = ceph_release,
2037 };
2038
2039 const struct inode_operations ceph_dir_iops = {
2040         .lookup = ceph_lookup,
2041         .permission = ceph_permission,
2042         .getattr = ceph_getattr,
2043         .setattr = ceph_setattr,
2044         .listxattr = ceph_listxattr,
2045         .get_inode_acl = ceph_get_acl,
2046         .set_acl = ceph_set_acl,
2047         .mknod = ceph_mknod,
2048         .symlink = ceph_symlink,
2049         .mkdir = ceph_mkdir,
2050         .link = ceph_link,
2051         .unlink = ceph_unlink,
2052         .rmdir = ceph_unlink,
2053         .rename = ceph_rename,
2054         .create = ceph_create,
2055         .atomic_open = ceph_atomic_open,
2056 };
2057
2058 const struct inode_operations ceph_snapdir_iops = {
2059         .lookup = ceph_lookup,
2060         .permission = ceph_permission,
2061         .getattr = ceph_getattr,
2062         .mkdir = ceph_mkdir,
2063         .rmdir = ceph_unlink,
2064         .rename = ceph_rename,
2065 };
2066
2067 const struct dentry_operations ceph_dentry_ops = {
2068         .d_revalidate = ceph_d_revalidate,
2069         .d_delete = ceph_d_delete,
2070         .d_release = ceph_d_release,
2071         .d_prune = ceph_d_prune,
2072         .d_init = ceph_d_init,
2073 };