fs/ceph/dir.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/spinlock.h>
   5 #include <linux/namei.h>
   6 #include <linux/slab.h>
   7 #include <linux/sched.h>
   8 #include <linux/xattr.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12
  13 /*
  14  * Directory operations: readdir, lookup, create, link, unlink,
  15  * rename, etc.
  16  */
  17
  18 /*
  19  * Ceph MDS operations are specified in terms of a base ino and
  20  * relative path.  Thus, the client can specify an operation on a
  21  * specific inode (e.g., a getattr due to fstat(2)), or as a path
  22  * relative to, say, the root directory.
  23  *
  24  * Normally, we limit ourselves to strict inode ops (no path component)
  25  * or dentry operations (a single path component relative to an ino).  The
  26  * exception to this is open_root_dentry(), which will open the mount
  27  * point by name.
  28  */
  29
  30 const struct dentry_operations ceph_dentry_ops;
  31
  32 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di);
  33 static int __dir_lease_try_check(const struct dentry *dentry);
  34
  35 /*
  36  * Initialize ceph dentry state.
  37  */
  38 static int ceph_d_init(struct dentry *dentry)
  39 {
  40         struct ceph_dentry_info *di;
  41         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
  42         struct ceph_mds_client *mdsc = fsc->mdsc;
  43
  44         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  45         if (!di)
  46                 return -ENOMEM;          /* oh well */
  47
  48         di->dentry = dentry;
  49         di->lease_session = NULL;
  50         di->time = jiffies;
  51         dentry->d_fsdata = di;
  52         INIT_LIST_HEAD(&di->lease_list);
  53
  54         atomic64_inc(&mdsc->metric.total_dentries);
  55
  56         return 0;
  57 }
  58
  59 /*
  60  * for f_pos for readdir:
  61  * - hash order:
  62  *      (0xff << 52) | ((24 bits hash) << 28) |
  63  *      (the nth entry has hash collision);
  64  * - frag+name order;
  65  *      ((frag value) << 28) | (the nth entry in frag);
  66  */
  67 #define OFFSET_BITS     28
  68 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  69 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  70 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  71 {
  72         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  73         if (hash_order)
  74                 fpos |= HASH_ORDER;
  75         return fpos;
  76 }
  77
  78 static bool is_hash_order(loff_t p)
  79 {
  80         return (p & HASH_ORDER) == HASH_ORDER;
  81 }
  82
  83 static unsigned fpos_frag(loff_t p)
  84 {
  85         return p >> OFFSET_BITS;
  86 }
  87
  88 static unsigned fpos_hash(loff_t p)
  89 {
  90         return ceph_frag_value(fpos_frag(p));
  91 }
  92
  93 static unsigned fpos_off(loff_t p)
  94 {
  95         return p & OFFSET_MASK;
  96 }
  97
  98 static int fpos_cmp(loff_t l, loff_t r)
  99 {
 100         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
 101         if (v)
 102                 return v;
 103         return (int)(fpos_off(l) - fpos_off(r));
 104 }
 105
 106 /*
 107  * make note of the last dentry we read, so we can
 108  * continue at the same lexicographical point,
 109  * regardless of what dir changes take place on the
 110  * server.
 111  */
 112 static int note_last_dentry(struct ceph_dir_file_info *dfi, const char *name,
 113                             int len, unsigned next_offset)
 114 {
 115         char *buf = kmalloc(len+1, GFP_KERNEL);
 116         if (!buf)
 117                 return -ENOMEM;
 118         kfree(dfi->last_name);
 119         dfi->last_name = buf;
 120         memcpy(dfi->last_name, name, len);
 121         dfi->last_name[len] = 0;
 122         dfi->next_offset = next_offset;
 123         dout("note_last_dentry '%s'\n", dfi->last_name);
 124         return 0;
 125 }
 126
 127
 128 static struct dentry *
 129 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 130                         struct ceph_readdir_cache_control *cache_ctl)
 131 {
 132         struct inode *dir = d_inode(parent);
 133         struct dentry *dentry;
 134         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 135         loff_t ptr_pos = idx * sizeof(struct dentry *);
 136         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 137
 138         if (ptr_pos >= i_size_read(dir))
 139                 return NULL;
 140
 141         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 142                 ceph_readdir_cache_release(cache_ctl);
 143                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 144                 if (!cache_ctl->page) {
 145                         dout(" page %lu not found\n", ptr_pgoff);
 146                         return ERR_PTR(-EAGAIN);
 147                 }
 148                 /* reading/filling the cache are serialized by
 149                    i_mutex, no need to use page lock */
 150                 unlock_page(cache_ctl->page);
 151                 cache_ctl->dentries = kmap(cache_ctl->page);
 152         }
 153
 154         cache_ctl->index = idx & idx_mask;
 155
 156         rcu_read_lock();
 157         spin_lock(&parent->d_lock);
 158         /* check i_size again here, because empty directory can be
 159          * marked as complete while not holding the i_mutex. */
 160         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 161                 dentry = cache_ctl->dentries[cache_ctl->index];
 162         else
 163                 dentry = NULL;
 164         spin_unlock(&parent->d_lock);
 165         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 166                 dentry = NULL;
 167         rcu_read_unlock();
 168         return dentry ? : ERR_PTR(-EAGAIN);
 169 }
 170
 171 /*
 172  * When possible, we try to satisfy a readdir by peeking at the
 173  * dcache.  We make this work by carefully ordering dentries on
 174  * d_child when we initially get results back from the MDS, and
 175  * falling back to a "normal" sync readdir if any dentries in the dir
 176  * are dropped.
 177  *
 178  * Complete dir indicates that we have all dentries in the dir.  It is
 179  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 180  * the MDS if/when the directory is modified).
 181  */
 182 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 183                             int shared_gen)
 184 {
 185         struct ceph_dir_file_info *dfi = file->private_data;
 186         struct dentry *parent = file->f_path.dentry;
 187         struct inode *dir = d_inode(parent);
 188         struct dentry *dentry, *last = NULL;
 189         struct ceph_dentry_info *di;
 190         struct ceph_readdir_cache_control cache_ctl = {};
 191         u64 idx = 0;
 192         int err = 0;
 193
 194         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 195
 196         /* search start position */
 197         if (ctx->pos > 2) {
 198                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 199                 while (count > 0) {
 200                         u64 step = count >> 1;
 201                         dentry = __dcache_find_get_entry(parent, idx + step,
 202                                                          &cache_ctl);
 203                         if (!dentry) {
 204                                 /* use linar search */
 205                                 idx = 0;
 206                                 break;
 207                         }
 208                         if (IS_ERR(dentry)) {
 209                                 err = PTR_ERR(dentry);
 210                                 goto out;
 211                         }
 212                         di = ceph_dentry(dentry);
 213                         spin_lock(&dentry->d_lock);
 214                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 215                                 idx += step + 1;
 216                                 count -= step + 1;
 217                         } else {
 218                                 count = step;
 219                         }
 220                         spin_unlock(&dentry->d_lock);
 221                         dput(dentry);
 222                 }
 223
 224                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 225         }
 226
 227
 228         for (;;) {
 229                 bool emit_dentry = false;
 230                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 231                 if (!dentry) {
 232                         dfi->file_info.flags |= CEPH_F_ATEND;
 233                         err = 0;
 234                         break;
 235                 }
 236                 if (IS_ERR(dentry)) {
 237                         err = PTR_ERR(dentry);
 238                         goto out;
 239                 }
 240
 241                 spin_lock(&dentry->d_lock);
 242                 di = ceph_dentry(dentry);
 243                 if (d_unhashed(dentry) ||
 244                     d_really_is_negative(dentry) ||
 245                     di->lease_shared_gen != shared_gen) {
 246                         spin_unlock(&dentry->d_lock);
 247                         dput(dentry);
 248                         err = -EAGAIN;
 249                         goto out;
 250                 }
 251                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 252                         __ceph_dentry_dir_lease_touch(di);
 253                         emit_dentry = true;
 254                 }
 255                 spin_unlock(&dentry->d_lock);
 256
 257                 if (emit_dentry) {
 258                         dout(" %llx dentry %p %pd %p\n", di->offset,
 259                              dentry, dentry, d_inode(dentry));
 260                         ctx->pos = di->offset;
 261                         if (!dir_emit(ctx, dentry->d_name.name,
 262                                       dentry->d_name.len, ceph_present_inode(d_inode(dentry)),
 263                                       d_inode(dentry)->i_mode >> 12)) {
 264                                 dput(dentry);
 265                                 err = 0;
 266                                 break;
 267                         }
 268                         ctx->pos++;
 269
 270                         if (last)
 271                                 dput(last);
 272                         last = dentry;
 273                 } else {
 274                         dput(dentry);
 275                 }
 276         }
 277 out:
 278         ceph_readdir_cache_release(&cache_ctl);
 279         if (last) {
 280                 int ret;
 281                 di = ceph_dentry(last);
 282                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 283                                        fpos_off(di->offset) + 1);
 284                 if (ret < 0)
 285                         err = ret;
 286                 dput(last);
 287                 /* last_name no longer match cache index */
 288                 if (dfi->readdir_cache_idx >= 0) {
 289                         dfi->readdir_cache_idx = -1;
 290                         dfi->dir_release_count = 0;
 291                 }
 292         }
 293         return err;
 294 }
 295
 296 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 297 {
 298         if (!dfi->last_readdir)
 299                 return true;
 300         if (is_hash_order(pos))
 301                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 302         else
 303                 return dfi->frag != fpos_frag(pos);
 304 }
 305
 306 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 307 {
 308         struct ceph_dir_file_info *dfi = file->private_data;
 309         struct inode *inode = file_inode(file);
 310         struct ceph_inode_info *ci = ceph_inode(inode);
 311         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 312         struct ceph_mds_client *mdsc = fsc->mdsc;
 313         int i;
 314         int err;
 315         unsigned frag = -1;
 316         struct ceph_mds_reply_info_parsed *rinfo;
 317
 318         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 319         if (dfi->file_info.flags & CEPH_F_ATEND)
 320                 return 0;
 321
 322         /* always start with . and .. */
 323         if (ctx->pos == 0) {
 324                 dout("readdir off 0 -> '.'\n");
 325                 if (!dir_emit(ctx, ".", 1, ceph_present_inode(inode),
 326                             inode->i_mode >> 12))
 327                         return 0;
 328                 ctx->pos = 1;
 329         }
 330         if (ctx->pos == 1) {
 331                 u64 ino;
 332                 struct dentry *dentry = file->f_path.dentry;
 333
 334                 spin_lock(&dentry->d_lock);
 335                 ino = ceph_present_inode(dentry->d_parent->d_inode);
 336                 spin_unlock(&dentry->d_lock);
 337
 338                 dout("readdir off 1 -> '..'\n");
 339                 if (!dir_emit(ctx, "..", 2, ino, inode->i_mode >> 12))
 340                         return 0;
 341                 ctx->pos = 2;
 342         }
 343
 344         spin_lock(&ci->i_ceph_lock);
 345         /* request Fx cap. if have Fx, we don't need to release Fs cap
 346          * for later create/unlink. */
 347         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_WR);
 348         /* can we use the dcache? */
 349         if (ceph_test_mount_opt(fsc, DCACHE) &&
 350             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 351             ceph_snap(inode) != CEPH_SNAPDIR &&
 352             __ceph_dir_is_complete_ordered(ci) &&
 353             __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 354                 int shared_gen = atomic_read(&ci->i_shared_gen);
 355
 356                 spin_unlock(&ci->i_ceph_lock);
 357                 err = __dcache_readdir(file, ctx, shared_gen);
 358                 if (err != -EAGAIN)
 359                         return err;
 360         } else {
 361                 spin_unlock(&ci->i_ceph_lock);
 362         }
 363
 364         /* proceed with a normal readdir */
 365 more:
 366         /* do we have the correct frag content buffered? */
 367         if (need_send_readdir(dfi, ctx->pos)) {
 368                 struct ceph_mds_request *req;
 369                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 370                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 371
 372                 /* discard old result, if any */
 373                 if (dfi->last_readdir) {
 374                         ceph_mdsc_put_request(dfi->last_readdir);
 375                         dfi->last_readdir = NULL;
 376                 }
 377
 378                 if (is_hash_order(ctx->pos)) {
 379                         /* fragtree isn't always accurate. choose frag
 380                          * based on previous reply when possible. */
 381                         if (frag == (unsigned)-1)
 382                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 383                                                         NULL, NULL);
 384                 } else {
 385                         frag = fpos_frag(ctx->pos);
 386                 }
 387
 388                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 389                      ceph_vinop(inode), frag, dfi->last_name);
 390                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 391                 if (IS_ERR(req))
 392                         return PTR_ERR(req);
 393                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 394                 if (err) {
 395                         ceph_mdsc_put_request(req);
 396                         return err;
 397                 }
 398                 /* hints to request -> mds selection code */
 399                 req->r_direct_mode = USE_AUTH_MDS;
 400                 if (op == CEPH_MDS_OP_READDIR) {
 401                         req->r_direct_hash = ceph_frag_value(frag);
 402                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 403                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 404                 }
 405                 if (dfi->last_name) {
 406                         req->r_path2 = kstrdup(dfi->last_name, GFP_KERNEL);
 407                         if (!req->r_path2) {
 408                                 ceph_mdsc_put_request(req);
 409                                 return -ENOMEM;
 410                         }
 411                 } else if (is_hash_order(ctx->pos)) {
 412                         req->r_args.readdir.offset_hash =
 413                                 cpu_to_le32(fpos_hash(ctx->pos));
 414                 }
 415
 416                 req->r_dir_release_cnt = dfi->dir_release_count;
 417                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 418                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 419                 req->r_readdir_offset = dfi->next_offset;
 420                 req->r_args.readdir.frag = cpu_to_le32(frag);
 421                 req->r_args.readdir.flags =
 422                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 423
 424                 req->r_inode = inode;
 425                 ihold(inode);
 426                 req->r_dentry = dget(file->f_path.dentry);
 427                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 428                 if (err < 0) {
 429                         ceph_mdsc_put_request(req);
 430                         return err;
 431                 }
 432                 dout("readdir got and parsed readdir result=%d on "
 433                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 434                      err, frag,
 435                      (int)req->r_reply_info.dir_end,
 436                      (int)req->r_reply_info.dir_complete,
 437                      (int)req->r_reply_info.hash_order);
 438
 439                 rinfo = &req->r_reply_info;
 440                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 441                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 442                         if (!rinfo->hash_order) {
 443                                 dfi->next_offset = req->r_readdir_offset;
 444                                 /* adjust ctx->pos to beginning of frag */
 445                                 ctx->pos = ceph_make_fpos(frag,
 446                                                           dfi->next_offset,
 447                                                           false);
 448                         }
 449                 }
 450
 451                 dfi->frag = frag;
 452                 dfi->last_readdir = req;
 453
 454                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 455                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 456                         if (dfi->readdir_cache_idx < 0) {
 457                                 /* preclude from marking dir ordered */
 458                                 dfi->dir_ordered_count = 0;
 459                         } else if (ceph_frag_is_leftmost(frag) &&
 460                                    dfi->next_offset == 2) {
 461                                 /* note dir version at start of readdir so
 462                                  * we can tell if any dentries get dropped */
 463                                 dfi->dir_release_count = req->r_dir_release_cnt;
 464                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 465                         }
 466                 } else {
 467                         dout("readdir !did_prepopulate\n");
 468                         /* disable readdir cache */
 469                         dfi->readdir_cache_idx = -1;
 470                         /* preclude from marking dir complete */
 471                         dfi->dir_release_count = 0;
 472                 }
 473
 474                 /* note next offset and last dentry name */
 475                 if (rinfo->dir_nr > 0) {
 476                         struct ceph_mds_reply_dir_entry *rde =
 477                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 478                         unsigned next_offset = req->r_reply_info.dir_end ?
 479                                         2 : (fpos_off(rde->offset) + 1);
 480                         err = note_last_dentry(dfi, rde->name, rde->name_len,
 481                                                next_offset);
 482                         if (err)
 483                                 return err;
 484                 } else if (req->r_reply_info.dir_end) {
 485                         dfi->next_offset = 2;
 486                         /* keep last name */
 487                 }
 488         }
 489
 490         rinfo = &dfi->last_readdir->r_reply_info;
 491         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 492              dfi->frag, rinfo->dir_nr, ctx->pos,
 493              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 494
 495         i = 0;
 496         /* search start position */
 497         if (rinfo->dir_nr > 0) {
 498                 int step, nr = rinfo->dir_nr;
 499                 while (nr > 0) {
 500                         step = nr >> 1;
 501                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 502                                 i +=  step + 1;
 503                                 nr -= step + 1;
 504                         } else {
 505                                 nr = step;
 506                         }
 507                 }
 508         }
 509         for (; i < rinfo->dir_nr; i++) {
 510                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 511
 512                 BUG_ON(rde->offset < ctx->pos);
 513
 514                 ctx->pos = rde->offset;
 515                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 516                      i, rinfo->dir_nr, ctx->pos,
 517                      rde->name_len, rde->name, &rde->inode.in);
 518
 519                 BUG_ON(!rde->inode.in);
 520
 521                 if (!dir_emit(ctx, rde->name, rde->name_len,
 522                               ceph_present_ino(inode->i_sb, le64_to_cpu(rde->inode.in->ino)),
 523                               le32_to_cpu(rde->inode.in->mode) >> 12)) {
 524                         dout("filldir stopping us...\n");
 525                         return 0;
 526                 }
 527                 ctx->pos++;
 528         }
 529
 530         ceph_mdsc_put_request(dfi->last_readdir);
 531         dfi->last_readdir = NULL;
 532
 533         if (dfi->next_offset > 2) {
 534                 frag = dfi->frag;
 535                 goto more;
 536         }
 537
 538         /* more frags? */
 539         if (!ceph_frag_is_rightmost(dfi->frag)) {
 540                 frag = ceph_frag_next(dfi->frag);
 541                 if (is_hash_order(ctx->pos)) {
 542                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 543                                                         dfi->next_offset, true);
 544                         if (new_pos > ctx->pos)
 545                                 ctx->pos = new_pos;
 546                         /* keep last_name */
 547                 } else {
 548                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 549                                                         false);
 550                         kfree(dfi->last_name);
 551                         dfi->last_name = NULL;
 552                 }
 553                 dout("readdir next frag is %x\n", frag);
 554                 goto more;
 555         }
 556         dfi->file_info.flags |= CEPH_F_ATEND;
 557
 558         /*
 559          * if dir_release_count still matches the dir, no dentries
 560          * were released during the whole readdir, and we should have
 561          * the complete dir contents in our cache.
 562          */
 563         if (atomic64_read(&ci->i_release_count) ==
 564                                         dfi->dir_release_count) {
 565                 spin_lock(&ci->i_ceph_lock);
 566                 if (dfi->dir_ordered_count ==
 567                                 atomic64_read(&ci->i_ordered_count)) {
 568                         dout(" marking %p complete and ordered\n", inode);
 569                         /* use i_size to track number of entries in
 570                          * readdir cache */
 571                         BUG_ON(dfi->readdir_cache_idx < 0);
 572                         i_size_write(inode, dfi->readdir_cache_idx *
 573                                      sizeof(struct dentry*));
 574                 } else {
 575                         dout(" marking %p complete\n", inode);
 576                 }
 577                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 578                                         dfi->dir_ordered_count);
 579                 spin_unlock(&ci->i_ceph_lock);
 580         }
 581
 582         dout("readdir %p file %p done.\n", inode, file);
 583         return 0;
 584 }
 585
 586 static void reset_readdir(struct ceph_dir_file_info *dfi)
 587 {
 588         if (dfi->last_readdir) {
 589                 ceph_mdsc_put_request(dfi->last_readdir);
 590                 dfi->last_readdir = NULL;
 591         }
 592         kfree(dfi->last_name);
 593         dfi->last_name = NULL;
 594         dfi->dir_release_count = 0;
 595         dfi->readdir_cache_idx = -1;
 596         dfi->next_offset = 2;  /* compensate for . and .. */
 597         dfi->file_info.flags &= ~CEPH_F_ATEND;
 598 }
 599
 600 /*
 601  * discard buffered readdir content on seekdir(0), or seek to new frag,
 602  * or seek prior to current chunk
 603  */
 604 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 605 {
 606         struct ceph_mds_reply_info_parsed *rinfo;
 607         loff_t chunk_offset;
 608         if (new_pos == 0)
 609                 return true;
 610         if (is_hash_order(new_pos)) {
 611                 /* no need to reset last_name for a forward seek when
 612                  * dentries are sotred in hash order */
 613         } else if (dfi->frag != fpos_frag(new_pos)) {
 614                 return true;
 615         }
 616         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 617         if (!rinfo || !rinfo->dir_nr)
 618                 return true;
 619         chunk_offset = rinfo->dir_entries[0].offset;
 620         return new_pos < chunk_offset ||
 621                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 622 }
 623
 624 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 625 {
 626         struct ceph_dir_file_info *dfi = file->private_data;
 627         struct inode *inode = file->f_mapping->host;
 628         loff_t retval;
 629
 630         inode_lock(inode);
 631         retval = -EINVAL;
 632         switch (whence) {
 633         case SEEK_CUR:
 634                 offset += file->f_pos;
 635         case SEEK_SET:
 636                 break;
 637         case SEEK_END:
 638                 retval = -EOPNOTSUPP;
 639         default:
 640                 goto out;
 641         }
 642
 643         if (offset >= 0) {
 644                 if (need_reset_readdir(dfi, offset)) {
 645                         dout("dir_llseek dropping %p content\n", file);
 646                         reset_readdir(dfi);
 647                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 648                         /* for hash offset, we don't know if a forward seek
 649                          * is within same frag */
 650                         dfi->dir_release_count = 0;
 651                         dfi->readdir_cache_idx = -1;
 652                 }
 653
 654                 if (offset != file->f_pos) {
 655                         file->f_pos = offset;
 656                         file->f_version = 0;
 657                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 658                 }
 659                 retval = offset;
 660         }
 661 out:
 662         inode_unlock(inode);
 663         return retval;
 664 }
 665
 666 /*
 667  * Handle lookups for the hidden .snap directory.
 668  */
 669 int ceph_handle_snapdir(struct ceph_mds_request *req,
 670                         struct dentry *dentry, int err)
 671 {
 672         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 673         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 674
 675         /* .snap dir? */
 676         if (err == -ENOENT &&
 677             ceph_snap(parent) == CEPH_NOSNAP &&
 678             strcmp(dentry->d_name.name,
 679                    fsc->mount_options->snapdir_name) == 0) {
 680                 struct inode *inode = ceph_get_snapdir(parent);
 681                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 682                      dentry, dentry, inode);
 683                 BUG_ON(!d_unhashed(dentry));
 684                 d_add(dentry, inode);
 685                 err = 0;
 686         }
 687         return err;
 688 }
 689
 690 /*
 691  * Figure out final result of a lookup/open request.
 692  *
 693  * Mainly, make sure we return the final req->r_dentry (if it already
 694  * existed) in place of the original VFS-provided dentry when they
 695  * differ.
 696  *
 697  * Gracefully handle the case where the MDS replies with -ENOENT and
 698  * no trace (which it may do, at its discretion, e.g., if it doesn't
 699  * care to issue a lease on the negative dentry).
 700  */
 701 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 702                                   struct dentry *dentry, int err)
 703 {
 704         if (err == -ENOENT) {
 705                 /* no trace? */
 706                 err = 0;
 707                 if (!req->r_reply_info.head->is_dentry) {
 708                         dout("ENOENT and no trace, dentry %p inode %p\n",
 709                              dentry, d_inode(dentry));
 710                         if (d_really_is_positive(dentry)) {
 711                                 d_drop(dentry);
 712                                 err = -ENOENT;
 713                         } else {
 714                                 d_add(dentry, NULL);
 715                         }
 716                 }
 717         }
 718         if (err)
 719                 dentry = ERR_PTR(err);
 720         else if (dentry != req->r_dentry)
 721                 dentry = dget(req->r_dentry);   /* we got spliced */
 722         else
 723                 dentry = NULL;
 724         return dentry;
 725 }
 726
 727 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 728 {
 729         return ceph_ino(inode) == CEPH_INO_ROOT &&
 730                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 731 }
 732
 733 /*
 734  * Look up a single dir entry.  If there is a lookup intent, inform
 735  * the MDS so that it gets our 'caps wanted' value in a single op.
 736  */
 737 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 738                                   unsigned int flags)
 739 {
 740         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 741         struct ceph_mds_client *mdsc = fsc->mdsc;
 742         struct ceph_mds_request *req;
 743         int op;
 744         int mask;
 745         int err;
 746
 747         dout("lookup %p dentry %p '%pd'\n",
 748              dir, dentry, dentry);
 749
 750         if (dentry->d_name.len > NAME_MAX)
 751                 return ERR_PTR(-ENAMETOOLONG);
 752
 753         /* can we conclude ENOENT locally? */
 754         if (d_really_is_negative(dentry)) {
 755                 struct ceph_inode_info *ci = ceph_inode(dir);
 756                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 757
 758                 spin_lock(&ci->i_ceph_lock);
 759                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 760                 if (strncmp(dentry->d_name.name,
 761                             fsc->mount_options->snapdir_name,
 762                             dentry->d_name.len) &&
 763                     !is_root_ceph_dentry(dir, dentry) &&
 764                     ceph_test_mount_opt(fsc, DCACHE) &&
 765                     __ceph_dir_is_complete(ci) &&
 766                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 767                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 768                         spin_unlock(&ci->i_ceph_lock);
 769                         dout(" dir %p complete, -ENOENT\n", dir);
 770                         d_add(dentry, NULL);
 771                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 772                         return NULL;
 773                 }
 774                 spin_unlock(&ci->i_ceph_lock);
 775         }
 776
 777         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 778                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 779         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 780         if (IS_ERR(req))
 781                 return ERR_CAST(req);
 782         req->r_dentry = dget(dentry);
 783         req->r_num_caps = 2;
 784
 785         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 786         if (ceph_security_xattr_wanted(dir))
 787                 mask |= CEPH_CAP_XATTR_SHARED;
 788         req->r_args.getattr.mask = cpu_to_le32(mask);
 789
 790         req->r_parent = dir;
 791         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 792         err = ceph_mdsc_do_request(mdsc, NULL, req);
 793         err = ceph_handle_snapdir(req, dentry, err);
 794         dentry = ceph_finish_lookup(req, dentry, err);
 795         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 796         dout("lookup result=%p\n", dentry);
 797         return dentry;
 798 }
 799
 800 /*
 801  * If we do a create but get no trace back from the MDS, follow up with
 802  * a lookup (the VFS expects us to link up the provided dentry).
 803  */
 804 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 805 {
 806         struct dentry *result = ceph_lookup(dir, dentry, 0);
 807
 808         if (result && !IS_ERR(result)) {
 809                 /*
 810                  * We created the item, then did a lookup, and found
 811                  * it was already linked to another inode we already
 812                  * had in our cache (and thus got spliced). To not
 813                  * confuse VFS (especially when inode is a directory),
 814                  * we don't link our dentry to that inode, return an
 815                  * error instead.
 816                  *
 817                  * This event should be rare and it happens only when
 818                  * we talk to old MDS. Recent MDS does not send traceless
 819                  * reply for request that creates new inode.
 820                  */
 821                 d_drop(result);
 822                 return -ESTALE;
 823         }
 824         return PTR_ERR(result);
 825 }
 826
 827 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 828                       umode_t mode, dev_t rdev)
 829 {
 830         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 831         struct ceph_mds_client *mdsc = fsc->mdsc;
 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 inode *dir, struct dentry *dentry, umode_t mode,
 884                        bool excl)
 885 {
 886         return ceph_mknod(dir, dentry, mode, 0);
 887 }
 888
 889 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 890                             const char *dest)
 891 {
 892         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 893         struct ceph_mds_client *mdsc = fsc->mdsc;
 894         struct ceph_mds_request *req;
 895         struct ceph_acl_sec_ctx as_ctx = {};
 896         int err;
 897
 898         if (ceph_snap(dir) != CEPH_NOSNAP)
 899                 return -EROFS;
 900
 901         if (ceph_quota_is_max_files_exceeded(dir)) {
 902                 err = -EDQUOT;
 903                 goto out;
 904         }
 905
 906         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 907         if (err < 0)
 908                 goto out;
 909
 910         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 911         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 912         if (IS_ERR(req)) {
 913                 err = PTR_ERR(req);
 914                 goto out;
 915         }
 916         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 917         if (!req->r_path2) {
 918                 err = -ENOMEM;
 919                 ceph_mdsc_put_request(req);
 920                 goto out;
 921         }
 922         req->r_parent = dir;
 923         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 924         req->r_dentry = dget(dentry);
 925         req->r_num_caps = 2;
 926         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 927         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 928         if (as_ctx.pagelist) {
 929                 req->r_pagelist = as_ctx.pagelist;
 930                 as_ctx.pagelist = NULL;
 931         }
 932         err = ceph_mdsc_do_request(mdsc, dir, req);
 933         if (!err && !req->r_reply_info.head->is_dentry)
 934                 err = ceph_handle_notrace_create(dir, dentry);
 935         ceph_mdsc_put_request(req);
 936 out:
 937         if (err)
 938                 d_drop(dentry);
 939         ceph_release_acl_sec_ctx(&as_ctx);
 940         return err;
 941 }
 942
 943 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 944 {
 945         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 946         struct ceph_mds_client *mdsc = fsc->mdsc;
 947         struct ceph_mds_request *req;
 948         struct ceph_acl_sec_ctx as_ctx = {};
 949         int err = -EROFS;
 950         int op;
 951
 952         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 953                 /* mkdir .snap/foo is a MKSNAP */
 954                 op = CEPH_MDS_OP_MKSNAP;
 955                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 956                      dentry, dentry);
 957         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 958                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 959                 op = CEPH_MDS_OP_MKDIR;
 960         } else {
 961                 goto out;
 962         }
 963
 964         if (op == CEPH_MDS_OP_MKDIR &&
 965             ceph_quota_is_max_files_exceeded(dir)) {
 966                 err = -EDQUOT;
 967                 goto out;
 968         }
 969
 970         mode |= S_IFDIR;
 971         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 972         if (err < 0)
 973                 goto out;
 974         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 975         if (err < 0)
 976                 goto out;
 977
 978         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 979         if (IS_ERR(req)) {
 980                 err = PTR_ERR(req);
 981                 goto out;
 982         }
 983
 984         req->r_dentry = dget(dentry);
 985         req->r_num_caps = 2;
 986         req->r_parent = dir;
 987         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 988         req->r_args.mkdir.mode = cpu_to_le32(mode);
 989         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 990         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 991         if (as_ctx.pagelist) {
 992                 req->r_pagelist = as_ctx.pagelist;
 993                 as_ctx.pagelist = NULL;
 994         }
 995         err = ceph_mdsc_do_request(mdsc, dir, req);
 996         if (!err &&
 997             !req->r_reply_info.head->is_target &&
 998             !req->r_reply_info.head->is_dentry)
 999                 err = ceph_handle_notrace_create(dir, dentry);
1000         ceph_mdsc_put_request(req);
1001 out:
1002         if (!err)
1003                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1004         else
1005                 d_drop(dentry);
1006         ceph_release_acl_sec_ctx(&as_ctx);
1007         return err;
1008 }
1009
1010 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1011                      struct dentry *dentry)
1012 {
1013         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1014         struct ceph_mds_client *mdsc = fsc->mdsc;
1015         struct ceph_mds_request *req;
1016         int err;
1017
1018         if (ceph_snap(dir) != CEPH_NOSNAP)
1019                 return -EROFS;
1020
1021         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1022              old_dentry, dentry);
1023         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1024         if (IS_ERR(req)) {
1025                 d_drop(dentry);
1026                 return PTR_ERR(req);
1027         }
1028         req->r_dentry = dget(dentry);
1029         req->r_num_caps = 2;
1030         req->r_old_dentry = dget(old_dentry);
1031         req->r_parent = dir;
1032         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1033         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1034         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1035         /* release LINK_SHARED on source inode (mds will lock it) */
1036         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1037         err = ceph_mdsc_do_request(mdsc, dir, req);
1038         if (err) {
1039                 d_drop(dentry);
1040         } else if (!req->r_reply_info.head->is_dentry) {
1041                 ihold(d_inode(old_dentry));
1042                 d_instantiate(dentry, d_inode(old_dentry));
1043         }
1044         ceph_mdsc_put_request(req);
1045         return err;
1046 }
1047
1048 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1049                                  struct ceph_mds_request *req)
1050 {
1051         int result = req->r_err ? req->r_err :
1052                         le32_to_cpu(req->r_reply_info.head->result);
1053
1054         if (result == -EJUKEBOX)
1055                 goto out;
1056
1057         /* If op failed, mark everyone involved for errors */
1058         if (result) {
1059                 int pathlen = 0;
1060                 u64 base = 0;
1061                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1062                                                   &base, 0);
1063
1064                 /* mark error on parent + clear complete */
1065                 mapping_set_error(req->r_parent->i_mapping, result);
1066                 ceph_dir_clear_complete(req->r_parent);
1067
1068                 /* drop the dentry -- we don't know its status */
1069                 if (!d_unhashed(req->r_dentry))
1070                         d_drop(req->r_dentry);
1071
1072                 /* mark inode itself for an error (since metadata is bogus) */
1073                 mapping_set_error(req->r_old_inode->i_mapping, result);
1074
1075                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1076                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1077                 ceph_mdsc_free_path(path, pathlen);
1078         }
1079 out:
1080         iput(req->r_old_inode);
1081         ceph_mdsc_release_dir_caps(req);
1082 }
1083
1084 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1085 {
1086         struct ceph_inode_info *ci = ceph_inode(dir);
1087         struct ceph_dentry_info *di;
1088         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1089
1090         spin_lock(&ci->i_ceph_lock);
1091         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1092                 ceph_take_cap_refs(ci, want, false);
1093                 got = want;
1094         }
1095         spin_unlock(&ci->i_ceph_lock);
1096
1097         /* If we didn't get anything, return 0 */
1098         if (!got)
1099                 return 0;
1100
1101         spin_lock(&dentry->d_lock);
1102         di = ceph_dentry(dentry);
1103         /*
1104          * - We are holding Fx, which implies Fs caps.
1105          * - Only support async unlink for primary linkage
1106          */
1107         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1108             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1109                 want = 0;
1110         spin_unlock(&dentry->d_lock);
1111
1112         /* Do we still want what we've got? */
1113         if (want == got)
1114                 return got;
1115
1116         ceph_put_cap_refs(ci, got);
1117         return 0;
1118 }
1119
1120 /*
1121  * rmdir and unlink are differ only by the metadata op code
1122  */
1123 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1124 {
1125         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1126         struct ceph_mds_client *mdsc = fsc->mdsc;
1127         struct inode *inode = d_inode(dentry);
1128         struct ceph_mds_request *req;
1129         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1130         int err = -EROFS;
1131         int op;
1132
1133         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1134                 /* rmdir .snap/foo is RMSNAP */
1135                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1136                 op = CEPH_MDS_OP_RMSNAP;
1137         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1138                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1139                      dir, dentry, inode);
1140                 op = d_is_dir(dentry) ?
1141                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1142         } else
1143                 goto out;
1144 retry:
1145         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1146         if (IS_ERR(req)) {
1147                 err = PTR_ERR(req);
1148                 goto out;
1149         }
1150         req->r_dentry = dget(dentry);
1151         req->r_num_caps = 2;
1152         req->r_parent = dir;
1153         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1154         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1155         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1156
1157         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1158             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1159                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1160                      dentry->d_name.len, dentry->d_name.name,
1161                      ceph_cap_string(req->r_dir_caps));
1162                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1163                 req->r_callback = ceph_async_unlink_cb;
1164                 req->r_old_inode = d_inode(dentry);
1165                 ihold(req->r_old_inode);
1166                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1167                 if (!err) {
1168                         /*
1169                          * We have enough caps, so we assume that the unlink
1170                          * will succeed. Fix up the target inode and dcache.
1171                          */
1172                         drop_nlink(inode);
1173                         d_delete(dentry);
1174                 } else if (err == -EJUKEBOX) {
1175                         try_async = false;
1176                         ceph_mdsc_put_request(req);
1177                         goto retry;
1178                 }
1179         } else {
1180                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1181                 err = ceph_mdsc_do_request(mdsc, dir, req);
1182                 if (!err && !req->r_reply_info.head->is_dentry)
1183                         d_delete(dentry);
1184         }
1185
1186         ceph_mdsc_put_request(req);
1187 out:
1188         return err;
1189 }
1190
1191 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1192                        struct inode *new_dir, struct dentry *new_dentry,
1193                        unsigned int flags)
1194 {
1195         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1196         struct ceph_mds_client *mdsc = fsc->mdsc;
1197         struct ceph_mds_request *req;
1198         int op = CEPH_MDS_OP_RENAME;
1199         int err;
1200
1201         if (flags)
1202                 return -EINVAL;
1203
1204         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1205                 return -EXDEV;
1206         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1207                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1208                         op = CEPH_MDS_OP_RENAMESNAP;
1209                 else
1210                         return -EROFS;
1211         } else if (old_dir != new_dir) {
1212                 err = ceph_quota_check_rename(mdsc, d_inode(old_dentry),
1213                                               new_dir);
1214                 if (err)
1215                         return err;
1216         }
1217
1218         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1219              old_dir, old_dentry, new_dir, new_dentry);
1220         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1221         if (IS_ERR(req))
1222                 return PTR_ERR(req);
1223         ihold(old_dir);
1224         req->r_dentry = dget(new_dentry);
1225         req->r_num_caps = 2;
1226         req->r_old_dentry = dget(old_dentry);
1227         req->r_old_dentry_dir = old_dir;
1228         req->r_parent = new_dir;
1229         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1230         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1231         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1232         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1233         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1234         /* release LINK_RDCACHE on source inode (mds will lock it) */
1235         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1236         if (d_really_is_positive(new_dentry)) {
1237                 req->r_inode_drop =
1238                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1239         }
1240         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1241         if (!err && !req->r_reply_info.head->is_dentry) {
1242                 /*
1243                  * Normally d_move() is done by fill_trace (called by
1244                  * do_request, above).  If there is no trace, we need
1245                  * to do it here.
1246                  */
1247                 d_move(old_dentry, new_dentry);
1248         }
1249         ceph_mdsc_put_request(req);
1250         return err;
1251 }
1252
1253 /*
1254  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1255  * Leases at front of the list will expire first. (Assume all leases have
1256  * similar duration)
1257  *
1258  * Called under dentry->d_lock.
1259  */
1260 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1261 {
1262         struct dentry *dn = di->dentry;
1263         struct ceph_mds_client *mdsc;
1264
1265         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1266
1267         di->flags |= CEPH_DENTRY_LEASE_LIST;
1268         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1269                 di->flags |= CEPH_DENTRY_REFERENCED;
1270                 return;
1271         }
1272
1273         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1274         spin_lock(&mdsc->dentry_list_lock);
1275         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1276         spin_unlock(&mdsc->dentry_list_lock);
1277 }
1278
1279 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1280                                      struct ceph_dentry_info *di)
1281 {
1282         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1283         di->lease_gen = 0;
1284         di->time = jiffies;
1285         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1286 }
1287
1288 /*
1289  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1290  * list if it's not in the list, otherwise set 'referenced' flag.
1291  *
1292  * Called under dentry->d_lock.
1293  */
1294 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1295 {
1296         struct dentry *dn = di->dentry;
1297         struct ceph_mds_client *mdsc;
1298
1299         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1300              di, dn, dn, di->offset);
1301
1302         if (!list_empty(&di->lease_list)) {
1303                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1304                         /* don't remove dentry from dentry lease list
1305                          * if its lease is valid */
1306                         if (__dentry_lease_is_valid(di))
1307                                 return;
1308                 } else {
1309                         di->flags |= CEPH_DENTRY_REFERENCED;
1310                         return;
1311                 }
1312         }
1313
1314         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1315                 di->flags |= CEPH_DENTRY_REFERENCED;
1316                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1317                 return;
1318         }
1319
1320         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1321         spin_lock(&mdsc->dentry_list_lock);
1322         __dentry_dir_lease_touch(mdsc, di),
1323         spin_unlock(&mdsc->dentry_list_lock);
1324 }
1325
1326 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1327 {
1328         struct ceph_mds_client *mdsc;
1329         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1330                 return;
1331         if (list_empty(&di->lease_list))
1332                 return;
1333
1334         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1335         spin_lock(&mdsc->dentry_list_lock);
1336         list_del_init(&di->lease_list);
1337         spin_unlock(&mdsc->dentry_list_lock);
1338 }
1339
1340 enum {
1341         KEEP    = 0,
1342         DELETE  = 1,
1343         TOUCH   = 2,
1344         STOP    = 4,
1345 };
1346
1347 struct ceph_lease_walk_control {
1348         bool dir_lease;
1349         bool expire_dir_lease;
1350         unsigned long nr_to_scan;
1351         unsigned long dir_lease_ttl;
1352 };
1353
1354 static unsigned long
1355 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1356                      struct ceph_lease_walk_control *lwc,
1357                      int (*check)(struct dentry*, void*))
1358 {
1359         struct ceph_dentry_info *di, *tmp;
1360         struct dentry *dentry, *last = NULL;
1361         struct list_head* list;
1362         LIST_HEAD(dispose);
1363         unsigned long freed = 0;
1364         int ret = 0;
1365
1366         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1367         spin_lock(&mdsc->dentry_list_lock);
1368         list_for_each_entry_safe(di, tmp, list, lease_list) {
1369                 if (!lwc->nr_to_scan)
1370                         break;
1371                 --lwc->nr_to_scan;
1372
1373                 dentry = di->dentry;
1374                 if (last == dentry)
1375                         break;
1376
1377                 if (!spin_trylock(&dentry->d_lock))
1378                         continue;
1379
1380                 if (__lockref_is_dead(&dentry->d_lockref)) {
1381                         list_del_init(&di->lease_list);
1382                         goto next;
1383                 }
1384
1385                 ret = check(dentry, lwc);
1386                 if (ret & TOUCH) {
1387                         /* move it into tail of dir lease list */
1388                         __dentry_dir_lease_touch(mdsc, di);
1389                         if (!last)
1390                                 last = dentry;
1391                 }
1392                 if (ret & DELETE) {
1393                         /* stale lease */
1394                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1395                         if (dentry->d_lockref.count > 0) {
1396                                 /* update_dentry_lease() will re-add
1397                                  * it to lease list, or
1398                                  * ceph_d_delete() will return 1 when
1399                                  * last reference is dropped */
1400                                 list_del_init(&di->lease_list);
1401                         } else {
1402                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1403                                 list_move_tail(&di->lease_list, &dispose);
1404                                 dget_dlock(dentry);
1405                         }
1406                 }
1407 next:
1408                 spin_unlock(&dentry->d_lock);
1409                 if (ret & STOP)
1410                         break;
1411         }
1412         spin_unlock(&mdsc->dentry_list_lock);
1413
1414         while (!list_empty(&dispose)) {
1415                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1416                                       lease_list);
1417                 dentry = di->dentry;
1418                 spin_lock(&dentry->d_lock);
1419
1420                 list_del_init(&di->lease_list);
1421                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1422                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1423                         spin_lock(&mdsc->dentry_list_lock);
1424                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1425                                 list_add_tail(&di->lease_list,
1426                                               &mdsc->dentry_leases);
1427                         } else {
1428                                 __dentry_dir_lease_touch(mdsc, di);
1429                         }
1430                         spin_unlock(&mdsc->dentry_list_lock);
1431                 } else {
1432                         freed++;
1433                 }
1434
1435                 spin_unlock(&dentry->d_lock);
1436                 /* ceph_d_delete() does the trick */
1437                 dput(dentry);
1438         }
1439         return freed;
1440 }
1441
1442 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1443 {
1444         struct ceph_dentry_info *di = ceph_dentry(dentry);
1445         int ret;
1446
1447         if (__dentry_lease_is_valid(di))
1448                 return STOP;
1449         ret = __dir_lease_try_check(dentry);
1450         if (ret == -EBUSY)
1451                 return KEEP;
1452         if (ret > 0)
1453                 return TOUCH;
1454         return DELETE;
1455 }
1456
1457 static int __dir_lease_check(struct dentry *dentry, void *arg)
1458 {
1459         struct ceph_lease_walk_control *lwc = arg;
1460         struct ceph_dentry_info *di = ceph_dentry(dentry);
1461
1462         int ret = __dir_lease_try_check(dentry);
1463         if (ret == -EBUSY)
1464                 return KEEP;
1465         if (ret > 0) {
1466                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1467                         return STOP;
1468                 /* Move dentry to tail of dir lease list if we don't want
1469                  * to delete it. So dentries in the list are checked in a
1470                  * round robin manner */
1471                 if (!lwc->expire_dir_lease)
1472                         return TOUCH;
1473                 if (dentry->d_lockref.count > 0 ||
1474                     (di->flags & CEPH_DENTRY_REFERENCED))
1475                         return TOUCH;
1476                 /* invalidate dir lease */
1477                 di->lease_shared_gen = 0;
1478         }
1479         return DELETE;
1480 }
1481
1482 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1483 {
1484         struct ceph_lease_walk_control lwc;
1485         unsigned long count;
1486         unsigned long freed;
1487
1488         spin_lock(&mdsc->caps_list_lock);
1489         if (mdsc->caps_use_max > 0 &&
1490             mdsc->caps_use_count > mdsc->caps_use_max)
1491                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1492         else
1493                 count = 0;
1494         spin_unlock(&mdsc->caps_list_lock);
1495
1496         lwc.dir_lease = false;
1497         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1498         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1499         if (!lwc.nr_to_scan) /* more invalid leases */
1500                 return -EAGAIN;
1501
1502         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1503                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1504
1505         lwc.dir_lease = true;
1506         lwc.expire_dir_lease = freed < count;
1507         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1508         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1509         if (!lwc.nr_to_scan) /* more to check */
1510                 return -EAGAIN;
1511
1512         return freed > 0 ? 1 : 0;
1513 }
1514
1515 /*
1516  * Ensure a dentry lease will no longer revalidate.
1517  */
1518 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1519 {
1520         struct ceph_dentry_info *di = ceph_dentry(dentry);
1521         spin_lock(&dentry->d_lock);
1522         di->time = jiffies;
1523         di->lease_shared_gen = 0;
1524         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1525         __dentry_lease_unlist(di);
1526         spin_unlock(&dentry->d_lock);
1527 }
1528
1529 /*
1530  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1531  * renew if the least is more than half up.
1532  */
1533 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1534 {
1535         struct ceph_mds_session *session;
1536
1537         if (!di->lease_gen)
1538                 return false;
1539
1540         session = di->lease_session;
1541         if (session) {
1542                 u32 gen;
1543                 unsigned long ttl;
1544
1545                 spin_lock(&session->s_gen_ttl_lock);
1546                 gen = session->s_cap_gen;
1547                 ttl = session->s_cap_ttl;
1548                 spin_unlock(&session->s_gen_ttl_lock);
1549
1550                 if (di->lease_gen == gen &&
1551                     time_before(jiffies, ttl) &&
1552                     time_before(jiffies, di->time))
1553                         return true;
1554         }
1555         di->lease_gen = 0;
1556         return false;
1557 }
1558
1559 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1560 {
1561         struct ceph_dentry_info *di;
1562         struct ceph_mds_session *session = NULL;
1563         u32 seq = 0;
1564         int valid = 0;
1565
1566         spin_lock(&dentry->d_lock);
1567         di = ceph_dentry(dentry);
1568         if (di && __dentry_lease_is_valid(di)) {
1569                 valid = 1;
1570
1571                 if (di->lease_renew_after &&
1572                     time_after(jiffies, di->lease_renew_after)) {
1573                         /*
1574                          * We should renew. If we're in RCU walk mode
1575                          * though, we can't do that so just return
1576                          * -ECHILD.
1577                          */
1578                         if (flags & LOOKUP_RCU) {
1579                                 valid = -ECHILD;
1580                         } else {
1581                                 session = ceph_get_mds_session(di->lease_session);
1582                                 seq = di->lease_seq;
1583                                 di->lease_renew_after = 0;
1584                                 di->lease_renew_from = jiffies;
1585                         }
1586                 }
1587         }
1588         spin_unlock(&dentry->d_lock);
1589
1590         if (session) {
1591                 ceph_mdsc_lease_send_msg(session, dentry,
1592                                          CEPH_MDS_LEASE_RENEW, seq);
1593                 ceph_put_mds_session(session);
1594         }
1595         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1596         return valid;
1597 }
1598
1599 /*
1600  * Called under dentry->d_lock.
1601  */
1602 static int __dir_lease_try_check(const struct dentry *dentry)
1603 {
1604         struct ceph_dentry_info *di = ceph_dentry(dentry);
1605         struct inode *dir;
1606         struct ceph_inode_info *ci;
1607         int valid = 0;
1608
1609         if (!di->lease_shared_gen)
1610                 return 0;
1611         if (IS_ROOT(dentry))
1612                 return 0;
1613
1614         dir = d_inode(dentry->d_parent);
1615         ci = ceph_inode(dir);
1616
1617         if (spin_trylock(&ci->i_ceph_lock)) {
1618                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1619                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1620                         valid = 1;
1621                 spin_unlock(&ci->i_ceph_lock);
1622         } else {
1623                 valid = -EBUSY;
1624         }
1625
1626         if (!valid)
1627                 di->lease_shared_gen = 0;
1628         return valid;
1629 }
1630
1631 /*
1632  * Check if directory-wide content lease/cap is valid.
1633  */
1634 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1635                               struct ceph_mds_client *mdsc)
1636 {
1637         struct ceph_inode_info *ci = ceph_inode(dir);
1638         int valid;
1639         int shared_gen;
1640
1641         spin_lock(&ci->i_ceph_lock);
1642         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1643         if (valid) {
1644                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1645                 shared_gen = atomic_read(&ci->i_shared_gen);
1646         }
1647         spin_unlock(&ci->i_ceph_lock);
1648         if (valid) {
1649                 struct ceph_dentry_info *di;
1650                 spin_lock(&dentry->d_lock);
1651                 di = ceph_dentry(dentry);
1652                 if (dir == d_inode(dentry->d_parent) &&
1653                     di && di->lease_shared_gen == shared_gen)
1654                         __ceph_dentry_dir_lease_touch(di);
1655                 else
1656                         valid = 0;
1657                 spin_unlock(&dentry->d_lock);
1658         }
1659         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1660              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1661         return valid;
1662 }
1663
1664 /*
1665  * Check if cached dentry can be trusted.
1666  */
1667 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1668 {
1669         int valid = 0;
1670         struct dentry *parent;
1671         struct inode *dir, *inode;
1672         struct ceph_mds_client *mdsc;
1673
1674         if (flags & LOOKUP_RCU) {
1675                 parent = READ_ONCE(dentry->d_parent);
1676                 dir = d_inode_rcu(parent);
1677                 if (!dir)
1678                         return -ECHILD;
1679                 inode = d_inode_rcu(dentry);
1680         } else {
1681                 parent = dget_parent(dentry);
1682                 dir = d_inode(parent);
1683                 inode = d_inode(dentry);
1684         }
1685
1686         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1687              dentry, inode, ceph_dentry(dentry)->offset);
1688
1689         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1690
1691         /* always trust cached snapped dentries, snapdir dentry */
1692         if (ceph_snap(dir) != CEPH_NOSNAP) {
1693                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1694                      dentry, inode);
1695                 valid = 1;
1696         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1697                 valid = 1;
1698         } else {
1699                 valid = dentry_lease_is_valid(dentry, flags);
1700                 if (valid == -ECHILD)
1701                         return valid;
1702                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1703                         if (inode)
1704                                 valid = ceph_is_any_caps(inode);
1705                         else
1706                                 valid = 1;
1707                 }
1708         }
1709
1710         if (!valid) {
1711                 struct ceph_mds_request *req;
1712                 int op, err;
1713                 u32 mask;
1714
1715                 if (flags & LOOKUP_RCU)
1716                         return -ECHILD;
1717
1718                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1719
1720                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1721                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1722                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1723                 if (!IS_ERR(req)) {
1724                         req->r_dentry = dget(dentry);
1725                         req->r_num_caps = 2;
1726                         req->r_parent = dir;
1727
1728                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1729                         if (ceph_security_xattr_wanted(dir))
1730                                 mask |= CEPH_CAP_XATTR_SHARED;
1731                         req->r_args.getattr.mask = cpu_to_le32(mask);
1732
1733                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1734                         switch (err) {
1735                         case 0:
1736                                 if (d_really_is_positive(dentry) &&
1737                                     d_inode(dentry) == req->r_target_inode)
1738                                         valid = 1;
1739                                 break;
1740                         case -ENOENT:
1741                                 if (d_really_is_negative(dentry))
1742                                         valid = 1;
1743                                 fallthrough;
1744                         default:
1745                                 break;
1746                         }
1747                         ceph_mdsc_put_request(req);
1748                         dout("d_revalidate %p lookup result=%d\n",
1749                              dentry, err);
1750                 }
1751         } else {
1752                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1753         }
1754
1755         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1756         if (!valid)
1757                 ceph_dir_clear_complete(dir);
1758
1759         if (!(flags & LOOKUP_RCU))
1760                 dput(parent);
1761         return valid;
1762 }
1763
1764 /*
1765  * Delete unused dentry that doesn't have valid lease
1766  *
1767  * Called under dentry->d_lock.
1768  */
1769 static int ceph_d_delete(const struct dentry *dentry)
1770 {
1771         struct ceph_dentry_info *di;
1772
1773         /* won't release caps */
1774         if (d_really_is_negative(dentry))
1775                 return 0;
1776         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1777                 return 0;
1778         /* vaild lease? */
1779         di = ceph_dentry(dentry);
1780         if (di) {
1781                 if (__dentry_lease_is_valid(di))
1782                         return 0;
1783                 if (__dir_lease_try_check(dentry))
1784                         return 0;
1785         }
1786         return 1;
1787 }
1788
1789 /*
1790  * Release our ceph_dentry_info.
1791  */
1792 static void ceph_d_release(struct dentry *dentry)
1793 {
1794         struct ceph_dentry_info *di = ceph_dentry(dentry);
1795         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1796
1797         dout("d_release %p\n", dentry);
1798
1799         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1800
1801         spin_lock(&dentry->d_lock);
1802         __dentry_lease_unlist(di);
1803         dentry->d_fsdata = NULL;
1804         spin_unlock(&dentry->d_lock);
1805
1806         if (di->lease_session)
1807                 ceph_put_mds_session(di->lease_session);
1808         kmem_cache_free(ceph_dentry_cachep, di);
1809 }
1810
1811 /*
1812  * When the VFS prunes a dentry from the cache, we need to clear the
1813  * complete flag on the parent directory.
1814  *
1815  * Called under dentry->d_lock.
1816  */
1817 static void ceph_d_prune(struct dentry *dentry)
1818 {
1819         struct ceph_inode_info *dir_ci;
1820         struct ceph_dentry_info *di;
1821
1822         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1823
1824         /* do we have a valid parent? */
1825         if (IS_ROOT(dentry))
1826                 return;
1827
1828         /* we hold d_lock, so d_parent is stable */
1829         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1830         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1831                 return;
1832
1833         /* who calls d_delete() should also disable dcache readdir */
1834         if (d_really_is_negative(dentry))
1835                 return;
1836
1837         /* d_fsdata does not get cleared until d_release */
1838         if (!d_unhashed(dentry)) {
1839                 __ceph_dir_clear_complete(dir_ci);
1840                 return;
1841         }
1842
1843         /* Disable dcache readdir just in case that someone called d_drop()
1844          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1845          * properly (dcache readdir is still enabled) */
1846         di = ceph_dentry(dentry);
1847         if (di->offset > 0 &&
1848             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1849                 __ceph_dir_clear_ordered(dir_ci);
1850 }
1851
1852 /*
1853  * read() on a dir.  This weird interface hack only works if mounted
1854  * with '-o dirstat'.
1855  */
1856 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1857                              loff_t *ppos)
1858 {
1859         struct ceph_dir_file_info *dfi = file->private_data;
1860         struct inode *inode = file_inode(file);
1861         struct ceph_inode_info *ci = ceph_inode(inode);
1862         int left;
1863         const int bufsize = 1024;
1864
1865         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1866                 return -EISDIR;
1867
1868         if (!dfi->dir_info) {
1869                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1870                 if (!dfi->dir_info)
1871                         return -ENOMEM;
1872                 dfi->dir_info_len =
1873                         snprintf(dfi->dir_info, bufsize,
1874                                 "entries:   %20lld\n"
1875                                 " files:    %20lld\n"
1876                                 " subdirs:  %20lld\n"
1877                                 "rentries:  %20lld\n"
1878                                 " rfiles:   %20lld\n"
1879                                 " rsubdirs: %20lld\n"
1880                                 "rbytes:    %20lld\n"
1881                                 "rctime:    %10lld.%09ld\n",
1882                                 ci->i_files + ci->i_subdirs,
1883                                 ci->i_files,
1884                                 ci->i_subdirs,
1885                                 ci->i_rfiles + ci->i_rsubdirs,
1886                                 ci->i_rfiles,
1887                                 ci->i_rsubdirs,
1888                                 ci->i_rbytes,
1889                                 ci->i_rctime.tv_sec,
1890                                 ci->i_rctime.tv_nsec);
1891         }
1892
1893         if (*ppos >= dfi->dir_info_len)
1894                 return 0;
1895         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1896         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1897         if (left == size)
1898                 return -EFAULT;
1899         *ppos += (size - left);
1900         return size - left;
1901 }
1902
1903
1904
1905 /*
1906  * Return name hash for a given dentry.  This is dependent on
1907  * the parent directory's hash function.
1908  */
1909 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1910 {
1911         struct ceph_inode_info *dci = ceph_inode(dir);
1912         unsigned hash;
1913
1914         switch (dci->i_dir_layout.dl_dir_hash) {
1915         case 0: /* for backward compat */
1916         case CEPH_STR_HASH_LINUX:
1917                 return dn->d_name.hash;
1918
1919         default:
1920                 spin_lock(&dn->d_lock);
1921                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1922                                      dn->d_name.name, dn->d_name.len);
1923                 spin_unlock(&dn->d_lock);
1924                 return hash;
1925         }
1926 }
1927
1928 const struct file_operations ceph_dir_fops = {
1929         .read = ceph_read_dir,
1930         .iterate = ceph_readdir,
1931         .llseek = ceph_dir_llseek,
1932         .open = ceph_open,
1933         .release = ceph_release,
1934         .unlocked_ioctl = ceph_ioctl,
1935         .compat_ioctl = compat_ptr_ioctl,
1936         .fsync = ceph_fsync,
1937         .lock = ceph_lock,
1938         .flock = ceph_flock,
1939 };
1940
1941 const struct file_operations ceph_snapdir_fops = {
1942         .iterate = ceph_readdir,
1943         .llseek = ceph_dir_llseek,
1944         .open = ceph_open,
1945         .release = ceph_release,
1946 };
1947
1948 const struct inode_operations ceph_dir_iops = {
1949         .lookup = ceph_lookup,
1950         .permission = ceph_permission,
1951         .getattr = ceph_getattr,
1952         .setattr = ceph_setattr,
1953         .listxattr = ceph_listxattr,
1954         .get_acl = ceph_get_acl,
1955         .set_acl = ceph_set_acl,
1956         .mknod = ceph_mknod,
1957         .symlink = ceph_symlink,
1958         .mkdir = ceph_mkdir,
1959         .link = ceph_link,
1960         .unlink = ceph_unlink,
1961         .rmdir = ceph_unlink,
1962         .rename = ceph_rename,
1963         .create = ceph_create,
1964         .atomic_open = ceph_atomic_open,
1965 };
1966
1967 const struct inode_operations ceph_snapdir_iops = {
1968         .lookup = ceph_lookup,
1969         .permission = ceph_permission,
1970         .getattr = ceph_getattr,
1971         .mkdir = ceph_mkdir,
1972         .rmdir = ceph_unlink,
1973         .rename = ceph_rename,
1974 };
1975
1976 const struct dentry_operations ceph_dentry_ops = {
1977         .d_revalidate = ceph_d_revalidate,
1978         .d_delete = ceph_d_delete,
1979         .d_release = ceph_d_release,
1980         .d_prune = ceph_d_prune,
1981         .d_init = ceph_d_init,
1982 };