fs/ceph/dir.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/ceph/ceph_debug.h>
   3
   4 #include <linux/spinlock.h>
   5 #include <linux/namei.h>
   6 #include <linux/slab.h>
   7 #include <linux/sched.h>
   8 #include <linux/xattr.h>
   9
  10 #include "super.h"
  11 #include "mds_client.h"
  12
  13 /*
  14  * Directory operations: readdir, lookup, create, link, unlink,
  15  * rename, etc.
  16  */
  17
  18 /*
  19  * Ceph MDS operations are specified in terms of a base ino and
  20  * relative path.  Thus, the client can specify an operation on a
  21  * specific inode (e.g., a getattr due to fstat(2)), or as a path
  22  * relative to, say, the root directory.
  23  *
  24  * Normally, we limit ourselves to strict inode ops (no path component)
  25  * or dentry operations (a single path component relative to an ino).  The
  26  * exception to this is open_root_dentry(), which will open the mount
  27  * point by name.
  28  */
  29
  30 const struct dentry_operations ceph_dentry_ops;
  31
  32 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di);
  33 static int __dir_lease_try_check(const struct dentry *dentry);
  34
  35 /*
  36  * Initialize ceph dentry state.
  37  */
  38 static int ceph_d_init(struct dentry *dentry)
  39 {
  40         struct ceph_dentry_info *di;
  41         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dentry->d_sb);
  42
  43         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  44         if (!di)
  45                 return -ENOMEM;          /* oh well */
  46
  47         di->dentry = dentry;
  48         di->lease_session = NULL;
  49         di->time = jiffies;
  50         dentry->d_fsdata = di;
  51         INIT_LIST_HEAD(&di->lease_list);
  52
  53         atomic64_inc(&mdsc->metric.total_dentries);
  54
  55         return 0;
  56 }
  57
  58 /*
  59  * for f_pos for readdir:
  60  * - hash order:
  61  *      (0xff << 52) | ((24 bits hash) << 28) |
  62  *      (the nth entry has hash collision);
  63  * - frag+name order;
  64  *      ((frag value) << 28) | (the nth entry in frag);
  65  */
  66 #define OFFSET_BITS     28
  67 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  68 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  69 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  70 {
  71         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  72         if (hash_order)
  73                 fpos |= HASH_ORDER;
  74         return fpos;
  75 }
  76
  77 static bool is_hash_order(loff_t p)
  78 {
  79         return (p & HASH_ORDER) == HASH_ORDER;
  80 }
  81
  82 static unsigned fpos_frag(loff_t p)
  83 {
  84         return p >> OFFSET_BITS;
  85 }
  86
  87 static unsigned fpos_hash(loff_t p)
  88 {
  89         return ceph_frag_value(fpos_frag(p));
  90 }
  91
  92 static unsigned fpos_off(loff_t p)
  93 {
  94         return p & OFFSET_MASK;
  95 }
  96
  97 static int fpos_cmp(loff_t l, loff_t r)
  98 {
  99         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
 100         if (v)
 101                 return v;
 102         return (int)(fpos_off(l) - fpos_off(r));
 103 }
 104
 105 /*
 106  * make note of the last dentry we read, so we can
 107  * continue at the same lexicographical point,
 108  * regardless of what dir changes take place on the
 109  * server.
 110  */
 111 static int note_last_dentry(struct ceph_dir_file_info *dfi, const char *name,
 112                             int len, unsigned next_offset)
 113 {
 114         char *buf = kmalloc(len+1, GFP_KERNEL);
 115         if (!buf)
 116                 return -ENOMEM;
 117         kfree(dfi->last_name);
 118         dfi->last_name = buf;
 119         memcpy(dfi->last_name, name, len);
 120         dfi->last_name[len] = 0;
 121         dfi->next_offset = next_offset;
 122         dout("note_last_dentry '%s'\n", dfi->last_name);
 123         return 0;
 124 }
 125
 126
 127 static struct dentry *
 128 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 129                         struct ceph_readdir_cache_control *cache_ctl)
 130 {
 131         struct inode *dir = d_inode(parent);
 132         struct dentry *dentry;
 133         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 134         loff_t ptr_pos = idx * sizeof(struct dentry *);
 135         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 136
 137         if (ptr_pos >= i_size_read(dir))
 138                 return NULL;
 139
 140         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 141                 ceph_readdir_cache_release(cache_ctl);
 142                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 143                 if (!cache_ctl->page) {
 144                         dout(" page %lu not found\n", ptr_pgoff);
 145                         return ERR_PTR(-EAGAIN);
 146                 }
 147                 /* reading/filling the cache are serialized by
 148                    i_mutex, no need to use page lock */
 149                 unlock_page(cache_ctl->page);
 150                 cache_ctl->dentries = kmap(cache_ctl->page);
 151         }
 152
 153         cache_ctl->index = idx & idx_mask;
 154
 155         rcu_read_lock();
 156         spin_lock(&parent->d_lock);
 157         /* check i_size again here, because empty directory can be
 158          * marked as complete while not holding the i_mutex. */
 159         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 160                 dentry = cache_ctl->dentries[cache_ctl->index];
 161         else
 162                 dentry = NULL;
 163         spin_unlock(&parent->d_lock);
 164         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 165                 dentry = NULL;
 166         rcu_read_unlock();
 167         return dentry ? : ERR_PTR(-EAGAIN);
 168 }
 169
 170 /*
 171  * When possible, we try to satisfy a readdir by peeking at the
 172  * dcache.  We make this work by carefully ordering dentries on
 173  * d_child when we initially get results back from the MDS, and
 174  * falling back to a "normal" sync readdir if any dentries in the dir
 175  * are dropped.
 176  *
 177  * Complete dir indicates that we have all dentries in the dir.  It is
 178  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 179  * the MDS if/when the directory is modified).
 180  */
 181 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 182                             int shared_gen)
 183 {
 184         struct ceph_dir_file_info *dfi = file->private_data;
 185         struct dentry *parent = file->f_path.dentry;
 186         struct inode *dir = d_inode(parent);
 187         struct dentry *dentry, *last = NULL;
 188         struct ceph_dentry_info *di;
 189         struct ceph_readdir_cache_control cache_ctl = {};
 190         u64 idx = 0;
 191         int err = 0;
 192
 193         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 194
 195         /* search start position */
 196         if (ctx->pos > 2) {
 197                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 198                 while (count > 0) {
 199                         u64 step = count >> 1;
 200                         dentry = __dcache_find_get_entry(parent, idx + step,
 201                                                          &cache_ctl);
 202                         if (!dentry) {
 203                                 /* use linar search */
 204                                 idx = 0;
 205                                 break;
 206                         }
 207                         if (IS_ERR(dentry)) {
 208                                 err = PTR_ERR(dentry);
 209                                 goto out;
 210                         }
 211                         di = ceph_dentry(dentry);
 212                         spin_lock(&dentry->d_lock);
 213                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 214                                 idx += step + 1;
 215                                 count -= step + 1;
 216                         } else {
 217                                 count = step;
 218                         }
 219                         spin_unlock(&dentry->d_lock);
 220                         dput(dentry);
 221                 }
 222
 223                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 224         }
 225
 226
 227         for (;;) {
 228                 bool emit_dentry = false;
 229                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 230                 if (!dentry) {
 231                         dfi->file_info.flags |= CEPH_F_ATEND;
 232                         err = 0;
 233                         break;
 234                 }
 235                 if (IS_ERR(dentry)) {
 236                         err = PTR_ERR(dentry);
 237                         goto out;
 238                 }
 239
 240                 spin_lock(&dentry->d_lock);
 241                 di = ceph_dentry(dentry);
 242                 if (d_unhashed(dentry) ||
 243                     d_really_is_negative(dentry) ||
 244                     di->lease_shared_gen != shared_gen) {
 245                         spin_unlock(&dentry->d_lock);
 246                         dput(dentry);
 247                         err = -EAGAIN;
 248                         goto out;
 249                 }
 250                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 251                         __ceph_dentry_dir_lease_touch(di);
 252                         emit_dentry = true;
 253                 }
 254                 spin_unlock(&dentry->d_lock);
 255
 256                 if (emit_dentry) {
 257                         dout(" %llx dentry %p %pd %p\n", di->offset,
 258                              dentry, dentry, d_inode(dentry));
 259                         ctx->pos = di->offset;
 260                         if (!dir_emit(ctx, dentry->d_name.name,
 261                                       dentry->d_name.len, ceph_present_inode(d_inode(dentry)),
 262                                       d_inode(dentry)->i_mode >> 12)) {
 263                                 dput(dentry);
 264                                 err = 0;
 265                                 break;
 266                         }
 267                         ctx->pos++;
 268
 269                         if (last)
 270                                 dput(last);
 271                         last = dentry;
 272                 } else {
 273                         dput(dentry);
 274                 }
 275         }
 276 out:
 277         ceph_readdir_cache_release(&cache_ctl);
 278         if (last) {
 279                 int ret;
 280                 di = ceph_dentry(last);
 281                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 282                                        fpos_off(di->offset) + 1);
 283                 if (ret < 0)
 284                         err = ret;
 285                 dput(last);
 286                 /* last_name no longer match cache index */
 287                 if (dfi->readdir_cache_idx >= 0) {
 288                         dfi->readdir_cache_idx = -1;
 289                         dfi->dir_release_count = 0;
 290                 }
 291         }
 292         return err;
 293 }
 294
 295 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 296 {
 297         if (!dfi->last_readdir)
 298                 return true;
 299         if (is_hash_order(pos))
 300                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 301         else
 302                 return dfi->frag != fpos_frag(pos);
 303 }
 304
 305 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 306 {
 307         struct ceph_dir_file_info *dfi = file->private_data;
 308         struct inode *inode = file_inode(file);
 309         struct ceph_inode_info *ci = ceph_inode(inode);
 310         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 311         struct ceph_mds_client *mdsc = fsc->mdsc;
 312         int i;
 313         int err;
 314         unsigned frag = -1;
 315         struct ceph_mds_reply_info_parsed *rinfo;
 316
 317         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 318         if (dfi->file_info.flags & CEPH_F_ATEND)
 319                 return 0;
 320
 321         /* always start with . and .. */
 322         if (ctx->pos == 0) {
 323                 dout("readdir off 0 -> '.'\n");
 324                 if (!dir_emit(ctx, ".", 1, ceph_present_inode(inode),
 325                             inode->i_mode >> 12))
 326                         return 0;
 327                 ctx->pos = 1;
 328         }
 329         if (ctx->pos == 1) {
 330                 u64 ino;
 331                 struct dentry *dentry = file->f_path.dentry;
 332
 333                 spin_lock(&dentry->d_lock);
 334                 ino = ceph_present_inode(dentry->d_parent->d_inode);
 335                 spin_unlock(&dentry->d_lock);
 336
 337                 dout("readdir off 1 -> '..'\n");
 338                 if (!dir_emit(ctx, "..", 2, ino, inode->i_mode >> 12))
 339                         return 0;
 340                 ctx->pos = 2;
 341         }
 342
 343         spin_lock(&ci->i_ceph_lock);
 344         /* request Fx cap. if have Fx, we don't need to release Fs cap
 345          * for later create/unlink. */
 346         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_WR);
 347         /* can we use the dcache? */
 348         if (ceph_test_mount_opt(fsc, DCACHE) &&
 349             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 350             ceph_snap(inode) != CEPH_SNAPDIR &&
 351             __ceph_dir_is_complete_ordered(ci) &&
 352             __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 353                 int shared_gen = atomic_read(&ci->i_shared_gen);
 354
 355                 spin_unlock(&ci->i_ceph_lock);
 356                 err = __dcache_readdir(file, ctx, shared_gen);
 357                 if (err != -EAGAIN)
 358                         return err;
 359         } else {
 360                 spin_unlock(&ci->i_ceph_lock);
 361         }
 362
 363         /* proceed with a normal readdir */
 364 more:
 365         /* do we have the correct frag content buffered? */
 366         if (need_send_readdir(dfi, ctx->pos)) {
 367                 struct ceph_mds_request *req;
 368                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 369                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 370
 371                 /* discard old result, if any */
 372                 if (dfi->last_readdir) {
 373                         ceph_mdsc_put_request(dfi->last_readdir);
 374                         dfi->last_readdir = NULL;
 375                 }
 376
 377                 if (is_hash_order(ctx->pos)) {
 378                         /* fragtree isn't always accurate. choose frag
 379                          * based on previous reply when possible. */
 380                         if (frag == (unsigned)-1)
 381                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 382                                                         NULL, NULL);
 383                 } else {
 384                         frag = fpos_frag(ctx->pos);
 385                 }
 386
 387                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 388                      ceph_vinop(inode), frag, dfi->last_name);
 389                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 390                 if (IS_ERR(req))
 391                         return PTR_ERR(req);
 392                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 393                 if (err) {
 394                         ceph_mdsc_put_request(req);
 395                         return err;
 396                 }
 397                 /* hints to request -> mds selection code */
 398                 req->r_direct_mode = USE_AUTH_MDS;
 399                 if (op == CEPH_MDS_OP_READDIR) {
 400                         req->r_direct_hash = ceph_frag_value(frag);
 401                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 402                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 403                 }
 404                 if (dfi->last_name) {
 405                         req->r_path2 = kstrdup(dfi->last_name, GFP_KERNEL);
 406                         if (!req->r_path2) {
 407                                 ceph_mdsc_put_request(req);
 408                                 return -ENOMEM;
 409                         }
 410                 } else if (is_hash_order(ctx->pos)) {
 411                         req->r_args.readdir.offset_hash =
 412                                 cpu_to_le32(fpos_hash(ctx->pos));
 413                 }
 414
 415                 req->r_dir_release_cnt = dfi->dir_release_count;
 416                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 417                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 418                 req->r_readdir_offset = dfi->next_offset;
 419                 req->r_args.readdir.frag = cpu_to_le32(frag);
 420                 req->r_args.readdir.flags =
 421                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 422
 423                 req->r_inode = inode;
 424                 ihold(inode);
 425                 req->r_dentry = dget(file->f_path.dentry);
 426                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 427                 if (err < 0) {
 428                         ceph_mdsc_put_request(req);
 429                         return err;
 430                 }
 431                 dout("readdir got and parsed readdir result=%d on "
 432                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 433                      err, frag,
 434                      (int)req->r_reply_info.dir_end,
 435                      (int)req->r_reply_info.dir_complete,
 436                      (int)req->r_reply_info.hash_order);
 437
 438                 rinfo = &req->r_reply_info;
 439                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 440                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 441                         if (!rinfo->hash_order) {
 442                                 dfi->next_offset = req->r_readdir_offset;
 443                                 /* adjust ctx->pos to beginning of frag */
 444                                 ctx->pos = ceph_make_fpos(frag,
 445                                                           dfi->next_offset,
 446                                                           false);
 447                         }
 448                 }
 449
 450                 dfi->frag = frag;
 451                 dfi->last_readdir = req;
 452
 453                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 454                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 455                         if (dfi->readdir_cache_idx < 0) {
 456                                 /* preclude from marking dir ordered */
 457                                 dfi->dir_ordered_count = 0;
 458                         } else if (ceph_frag_is_leftmost(frag) &&
 459                                    dfi->next_offset == 2) {
 460                                 /* note dir version at start of readdir so
 461                                  * we can tell if any dentries get dropped */
 462                                 dfi->dir_release_count = req->r_dir_release_cnt;
 463                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 464                         }
 465                 } else {
 466                         dout("readdir !did_prepopulate\n");
 467                         /* disable readdir cache */
 468                         dfi->readdir_cache_idx = -1;
 469                         /* preclude from marking dir complete */
 470                         dfi->dir_release_count = 0;
 471                 }
 472
 473                 /* note next offset and last dentry name */
 474                 if (rinfo->dir_nr > 0) {
 475                         struct ceph_mds_reply_dir_entry *rde =
 476                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 477                         unsigned next_offset = req->r_reply_info.dir_end ?
 478                                         2 : (fpos_off(rde->offset) + 1);
 479                         err = note_last_dentry(dfi, rde->name, rde->name_len,
 480                                                next_offset);
 481                         if (err)
 482                                 return err;
 483                 } else if (req->r_reply_info.dir_end) {
 484                         dfi->next_offset = 2;
 485                         /* keep last name */
 486                 }
 487         }
 488
 489         rinfo = &dfi->last_readdir->r_reply_info;
 490         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 491              dfi->frag, rinfo->dir_nr, ctx->pos,
 492              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 493
 494         i = 0;
 495         /* search start position */
 496         if (rinfo->dir_nr > 0) {
 497                 int step, nr = rinfo->dir_nr;
 498                 while (nr > 0) {
 499                         step = nr >> 1;
 500                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 501                                 i +=  step + 1;
 502                                 nr -= step + 1;
 503                         } else {
 504                                 nr = step;
 505                         }
 506                 }
 507         }
 508         for (; i < rinfo->dir_nr; i++) {
 509                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 510
 511                 BUG_ON(rde->offset < ctx->pos);
 512
 513                 ctx->pos = rde->offset;
 514                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 515                      i, rinfo->dir_nr, ctx->pos,
 516                      rde->name_len, rde->name, &rde->inode.in);
 517
 518                 BUG_ON(!rde->inode.in);
 519
 520                 if (!dir_emit(ctx, rde->name, rde->name_len,
 521                               ceph_present_ino(inode->i_sb, le64_to_cpu(rde->inode.in->ino)),
 522                               le32_to_cpu(rde->inode.in->mode) >> 12)) {
 523                         dout("filldir stopping us...\n");
 524                         return 0;
 525                 }
 526                 ctx->pos++;
 527         }
 528
 529         ceph_mdsc_put_request(dfi->last_readdir);
 530         dfi->last_readdir = NULL;
 531
 532         if (dfi->next_offset > 2) {
 533                 frag = dfi->frag;
 534                 goto more;
 535         }
 536
 537         /* more frags? */
 538         if (!ceph_frag_is_rightmost(dfi->frag)) {
 539                 frag = ceph_frag_next(dfi->frag);
 540                 if (is_hash_order(ctx->pos)) {
 541                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 542                                                         dfi->next_offset, true);
 543                         if (new_pos > ctx->pos)
 544                                 ctx->pos = new_pos;
 545                         /* keep last_name */
 546                 } else {
 547                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 548                                                         false);
 549                         kfree(dfi->last_name);
 550                         dfi->last_name = NULL;
 551                 }
 552                 dout("readdir next frag is %x\n", frag);
 553                 goto more;
 554         }
 555         dfi->file_info.flags |= CEPH_F_ATEND;
 556
 557         /*
 558          * if dir_release_count still matches the dir, no dentries
 559          * were released during the whole readdir, and we should have
 560          * the complete dir contents in our cache.
 561          */
 562         if (atomic64_read(&ci->i_release_count) ==
 563                                         dfi->dir_release_count) {
 564                 spin_lock(&ci->i_ceph_lock);
 565                 if (dfi->dir_ordered_count ==
 566                                 atomic64_read(&ci->i_ordered_count)) {
 567                         dout(" marking %p complete and ordered\n", inode);
 568                         /* use i_size to track number of entries in
 569                          * readdir cache */
 570                         BUG_ON(dfi->readdir_cache_idx < 0);
 571                         i_size_write(inode, dfi->readdir_cache_idx *
 572                                      sizeof(struct dentry*));
 573                 } else {
 574                         dout(" marking %p complete\n", inode);
 575                 }
 576                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 577                                         dfi->dir_ordered_count);
 578                 spin_unlock(&ci->i_ceph_lock);
 579         }
 580
 581         dout("readdir %p file %p done.\n", inode, file);
 582         return 0;
 583 }
 584
 585 static void reset_readdir(struct ceph_dir_file_info *dfi)
 586 {
 587         if (dfi->last_readdir) {
 588                 ceph_mdsc_put_request(dfi->last_readdir);
 589                 dfi->last_readdir = NULL;
 590         }
 591         kfree(dfi->last_name);
 592         dfi->last_name = NULL;
 593         dfi->dir_release_count = 0;
 594         dfi->readdir_cache_idx = -1;
 595         dfi->next_offset = 2;  /* compensate for . and .. */
 596         dfi->file_info.flags &= ~CEPH_F_ATEND;
 597 }
 598
 599 /*
 600  * discard buffered readdir content on seekdir(0), or seek to new frag,
 601  * or seek prior to current chunk
 602  */
 603 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 604 {
 605         struct ceph_mds_reply_info_parsed *rinfo;
 606         loff_t chunk_offset;
 607         if (new_pos == 0)
 608                 return true;
 609         if (is_hash_order(new_pos)) {
 610                 /* no need to reset last_name for a forward seek when
 611                  * dentries are sotred in hash order */
 612         } else if (dfi->frag != fpos_frag(new_pos)) {
 613                 return true;
 614         }
 615         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 616         if (!rinfo || !rinfo->dir_nr)
 617                 return true;
 618         chunk_offset = rinfo->dir_entries[0].offset;
 619         return new_pos < chunk_offset ||
 620                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 621 }
 622
 623 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 624 {
 625         struct ceph_dir_file_info *dfi = file->private_data;
 626         struct inode *inode = file->f_mapping->host;
 627         loff_t retval;
 628
 629         inode_lock(inode);
 630         retval = -EINVAL;
 631         switch (whence) {
 632         case SEEK_CUR:
 633                 offset += file->f_pos;
 634                 break;
 635         case SEEK_SET:
 636                 break;
 637         case SEEK_END:
 638                 retval = -EOPNOTSUPP;
 639                 goto out;
 640         default:
 641                 goto out;
 642         }
 643
 644         if (offset >= 0) {
 645                 if (need_reset_readdir(dfi, offset)) {
 646                         dout("dir_llseek dropping %p content\n", file);
 647                         reset_readdir(dfi);
 648                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 649                         /* for hash offset, we don't know if a forward seek
 650                          * is within same frag */
 651                         dfi->dir_release_count = 0;
 652                         dfi->readdir_cache_idx = -1;
 653                 }
 654
 655                 if (offset != file->f_pos) {
 656                         file->f_pos = offset;
 657                         file->f_version = 0;
 658                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 659                 }
 660                 retval = offset;
 661         }
 662 out:
 663         inode_unlock(inode);
 664         return retval;
 665 }
 666
 667 /*
 668  * Handle lookups for the hidden .snap directory.
 669  */
 670 struct dentry *ceph_handle_snapdir(struct ceph_mds_request *req,
 671                                    struct dentry *dentry)
 672 {
 673         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 674         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 675
 676         /* .snap dir? */
 677         if (ceph_snap(parent) == CEPH_NOSNAP &&
 678             strcmp(dentry->d_name.name, fsc->mount_options->snapdir_name) == 0) {
 679                 struct dentry *res;
 680                 struct inode *inode = ceph_get_snapdir(parent);
 681
 682                 res = d_splice_alias(inode, dentry);
 683                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p. Spliced dentry %p\n",
 684                      dentry, dentry, inode, res);
 685                 if (res)
 686                         dentry = res;
 687         }
 688         return dentry;
 689 }
 690
 691 /*
 692  * Figure out final result of a lookup/open request.
 693  *
 694  * Mainly, make sure we return the final req->r_dentry (if it already
 695  * existed) in place of the original VFS-provided dentry when they
 696  * differ.
 697  *
 698  * Gracefully handle the case where the MDS replies with -ENOENT and
 699  * no trace (which it may do, at its discretion, e.g., if it doesn't
 700  * care to issue a lease on the negative dentry).
 701  */
 702 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 703                                   struct dentry *dentry, int err)
 704 {
 705         if (err == -ENOENT) {
 706                 /* no trace? */
 707                 err = 0;
 708                 if (!req->r_reply_info.head->is_dentry) {
 709                         dout("ENOENT and no trace, dentry %p inode %p\n",
 710                              dentry, d_inode(dentry));
 711                         if (d_really_is_positive(dentry)) {
 712                                 d_drop(dentry);
 713                                 err = -ENOENT;
 714                         } else {
 715                                 d_add(dentry, NULL);
 716                         }
 717                 }
 718         }
 719         if (err)
 720                 dentry = ERR_PTR(err);
 721         else if (dentry != req->r_dentry)
 722                 dentry = dget(req->r_dentry);   /* we got spliced */
 723         else
 724                 dentry = NULL;
 725         return dentry;
 726 }
 727
 728 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 729 {
 730         return ceph_ino(inode) == CEPH_INO_ROOT &&
 731                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 732 }
 733
 734 /*
 735  * Look up a single dir entry.  If there is a lookup intent, inform
 736  * the MDS so that it gets our 'caps wanted' value in a single op.
 737  */
 738 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 739                                   unsigned int flags)
 740 {
 741         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 742         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 743         struct ceph_mds_request *req;
 744         int op;
 745         int mask;
 746         int err;
 747
 748         dout("lookup %p dentry %p '%pd'\n",
 749              dir, dentry, dentry);
 750
 751         if (dentry->d_name.len > NAME_MAX)
 752                 return ERR_PTR(-ENAMETOOLONG);
 753
 754         /* can we conclude ENOENT locally? */
 755         if (d_really_is_negative(dentry)) {
 756                 struct ceph_inode_info *ci = ceph_inode(dir);
 757                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 758
 759                 spin_lock(&ci->i_ceph_lock);
 760                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 761                 if (strncmp(dentry->d_name.name,
 762                             fsc->mount_options->snapdir_name,
 763                             dentry->d_name.len) &&
 764                     !is_root_ceph_dentry(dir, dentry) &&
 765                     ceph_test_mount_opt(fsc, DCACHE) &&
 766                     __ceph_dir_is_complete(ci) &&
 767                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 768                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 769                         spin_unlock(&ci->i_ceph_lock);
 770                         dout(" dir %p complete, -ENOENT\n", dir);
 771                         d_add(dentry, NULL);
 772                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 773                         return NULL;
 774                 }
 775                 spin_unlock(&ci->i_ceph_lock);
 776         }
 777
 778         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 779                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 780         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 781         if (IS_ERR(req))
 782                 return ERR_CAST(req);
 783         req->r_dentry = dget(dentry);
 784         req->r_num_caps = 2;
 785
 786         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 787         if (ceph_security_xattr_wanted(dir))
 788                 mask |= CEPH_CAP_XATTR_SHARED;
 789         req->r_args.getattr.mask = cpu_to_le32(mask);
 790
 791         ihold(dir);
 792         req->r_parent = dir;
 793         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 794         err = ceph_mdsc_do_request(mdsc, NULL, req);
 795         if (err == -ENOENT) {
 796                 struct dentry *res;
 797
 798                 res = ceph_handle_snapdir(req, dentry);
 799                 if (IS_ERR(res)) {
 800                         err = PTR_ERR(res);
 801                 } else {
 802                         dentry = res;
 803                         err = 0;
 804                 }
 805         }
 806         dentry = ceph_finish_lookup(req, dentry, err);
 807         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 808         dout("lookup result=%p\n", dentry);
 809         return dentry;
 810 }
 811
 812 /*
 813  * If we do a create but get no trace back from the MDS, follow up with
 814  * a lookup (the VFS expects us to link up the provided dentry).
 815  */
 816 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 817 {
 818         struct dentry *result = ceph_lookup(dir, dentry, 0);
 819
 820         if (result && !IS_ERR(result)) {
 821                 /*
 822                  * We created the item, then did a lookup, and found
 823                  * it was already linked to another inode we already
 824                  * had in our cache (and thus got spliced). To not
 825                  * confuse VFS (especially when inode is a directory),
 826                  * we don't link our dentry to that inode, return an
 827                  * error instead.
 828                  *
 829                  * This event should be rare and it happens only when
 830                  * we talk to old MDS. Recent MDS does not send traceless
 831                  * reply for request that creates new inode.
 832                  */
 833                 d_drop(result);
 834                 return -ESTALE;
 835         }
 836         return PTR_ERR(result);
 837 }
 838
 839 static int ceph_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 840                       struct dentry *dentry, umode_t mode, dev_t rdev)
 841 {
 842         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 843         struct ceph_mds_request *req;
 844         struct ceph_acl_sec_ctx as_ctx = {};
 845         int err;
 846
 847         if (ceph_snap(dir) != CEPH_NOSNAP)
 848                 return -EROFS;
 849
 850         if (ceph_quota_is_max_files_exceeded(dir)) {
 851                 err = -EDQUOT;
 852                 goto out;
 853         }
 854
 855         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 856         if (err < 0)
 857                 goto out;
 858         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 859         if (err < 0)
 860                 goto out;
 861
 862         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 863              dir, dentry, mode, rdev);
 864         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 865         if (IS_ERR(req)) {
 866                 err = PTR_ERR(req);
 867                 goto out;
 868         }
 869         req->r_dentry = dget(dentry);
 870         req->r_num_caps = 2;
 871         req->r_parent = dir;
 872         ihold(dir);
 873         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 874         req->r_args.mknod.mode = cpu_to_le32(mode);
 875         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 876         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 877         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 878         if (as_ctx.pagelist) {
 879                 req->r_pagelist = as_ctx.pagelist;
 880                 as_ctx.pagelist = NULL;
 881         }
 882         err = ceph_mdsc_do_request(mdsc, dir, req);
 883         if (!err && !req->r_reply_info.head->is_dentry)
 884                 err = ceph_handle_notrace_create(dir, dentry);
 885         ceph_mdsc_put_request(req);
 886 out:
 887         if (!err)
 888                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 889         else
 890                 d_drop(dentry);
 891         ceph_release_acl_sec_ctx(&as_ctx);
 892         return err;
 893 }
 894
 895 static int ceph_create(struct user_namespace *mnt_userns, struct inode *dir,
 896                        struct dentry *dentry, umode_t mode, bool excl)
 897 {
 898         return ceph_mknod(mnt_userns, dir, dentry, mode, 0);
 899 }
 900
 901 static int ceph_symlink(struct user_namespace *mnt_userns, struct inode *dir,
 902                         struct dentry *dentry, const char *dest)
 903 {
 904         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 905         struct ceph_mds_request *req;
 906         struct ceph_acl_sec_ctx as_ctx = {};
 907         int err;
 908
 909         if (ceph_snap(dir) != CEPH_NOSNAP)
 910                 return -EROFS;
 911
 912         if (ceph_quota_is_max_files_exceeded(dir)) {
 913                 err = -EDQUOT;
 914                 goto out;
 915         }
 916
 917         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 918         if (err < 0)
 919                 goto out;
 920
 921         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 922         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 923         if (IS_ERR(req)) {
 924                 err = PTR_ERR(req);
 925                 goto out;
 926         }
 927         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 928         if (!req->r_path2) {
 929                 err = -ENOMEM;
 930                 ceph_mdsc_put_request(req);
 931                 goto out;
 932         }
 933         req->r_parent = dir;
 934         ihold(dir);
 935
 936         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 937         req->r_dentry = dget(dentry);
 938         req->r_num_caps = 2;
 939         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 940         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 941         if (as_ctx.pagelist) {
 942                 req->r_pagelist = as_ctx.pagelist;
 943                 as_ctx.pagelist = NULL;
 944         }
 945         err = ceph_mdsc_do_request(mdsc, dir, req);
 946         if (!err && !req->r_reply_info.head->is_dentry)
 947                 err = ceph_handle_notrace_create(dir, dentry);
 948         ceph_mdsc_put_request(req);
 949 out:
 950         if (err)
 951                 d_drop(dentry);
 952         ceph_release_acl_sec_ctx(&as_ctx);
 953         return err;
 954 }
 955
 956 static int ceph_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 957                       struct dentry *dentry, umode_t mode)
 958 {
 959         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 960         struct ceph_mds_request *req;
 961         struct ceph_acl_sec_ctx as_ctx = {};
 962         int err = -EROFS;
 963         int op;
 964
 965         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 966                 /* mkdir .snap/foo is a MKSNAP */
 967                 op = CEPH_MDS_OP_MKSNAP;
 968                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 969                      dentry, dentry);
 970         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 971                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 972                 op = CEPH_MDS_OP_MKDIR;
 973         } else {
 974                 goto out;
 975         }
 976
 977         if (op == CEPH_MDS_OP_MKDIR &&
 978             ceph_quota_is_max_files_exceeded(dir)) {
 979                 err = -EDQUOT;
 980                 goto out;
 981         }
 982
 983         mode |= S_IFDIR;
 984         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 985         if (err < 0)
 986                 goto out;
 987         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 988         if (err < 0)
 989                 goto out;
 990
 991         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 992         if (IS_ERR(req)) {
 993                 err = PTR_ERR(req);
 994                 goto out;
 995         }
 996
 997         req->r_dentry = dget(dentry);
 998         req->r_num_caps = 2;
 999         req->r_parent = dir;
1000         ihold(dir);
1001         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1002         req->r_args.mkdir.mode = cpu_to_le32(mode);
1003         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
1004         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1005         if (as_ctx.pagelist) {
1006                 req->r_pagelist = as_ctx.pagelist;
1007                 as_ctx.pagelist = NULL;
1008         }
1009         err = ceph_mdsc_do_request(mdsc, dir, req);
1010         if (!err &&
1011             !req->r_reply_info.head->is_target &&
1012             !req->r_reply_info.head->is_dentry)
1013                 err = ceph_handle_notrace_create(dir, dentry);
1014         ceph_mdsc_put_request(req);
1015 out:
1016         if (!err)
1017                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1018         else
1019                 d_drop(dentry);
1020         ceph_release_acl_sec_ctx(&as_ctx);
1021         return err;
1022 }
1023
1024 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1025                      struct dentry *dentry)
1026 {
1027         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1028         struct ceph_mds_request *req;
1029         int err;
1030
1031         if (ceph_snap(dir) != CEPH_NOSNAP)
1032                 return -EROFS;
1033
1034         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1035              old_dentry, dentry);
1036         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1037         if (IS_ERR(req)) {
1038                 d_drop(dentry);
1039                 return PTR_ERR(req);
1040         }
1041         req->r_dentry = dget(dentry);
1042         req->r_num_caps = 2;
1043         req->r_old_dentry = dget(old_dentry);
1044         req->r_parent = dir;
1045         ihold(dir);
1046         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1047         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1048         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1049         /* release LINK_SHARED on source inode (mds will lock it) */
1050         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1051         err = ceph_mdsc_do_request(mdsc, dir, req);
1052         if (err) {
1053                 d_drop(dentry);
1054         } else if (!req->r_reply_info.head->is_dentry) {
1055                 ihold(d_inode(old_dentry));
1056                 d_instantiate(dentry, d_inode(old_dentry));
1057         }
1058         ceph_mdsc_put_request(req);
1059         return err;
1060 }
1061
1062 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1063                                  struct ceph_mds_request *req)
1064 {
1065         int result = req->r_err ? req->r_err :
1066                         le32_to_cpu(req->r_reply_info.head->result);
1067
1068         if (result == -EJUKEBOX)
1069                 goto out;
1070
1071         /* If op failed, mark everyone involved for errors */
1072         if (result) {
1073                 int pathlen = 0;
1074                 u64 base = 0;
1075                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1076                                                   &base, 0);
1077
1078                 /* mark error on parent + clear complete */
1079                 mapping_set_error(req->r_parent->i_mapping, result);
1080                 ceph_dir_clear_complete(req->r_parent);
1081
1082                 /* drop the dentry -- we don't know its status */
1083                 if (!d_unhashed(req->r_dentry))
1084                         d_drop(req->r_dentry);
1085
1086                 /* mark inode itself for an error (since metadata is bogus) */
1087                 mapping_set_error(req->r_old_inode->i_mapping, result);
1088
1089                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1090                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1091                 ceph_mdsc_free_path(path, pathlen);
1092         }
1093 out:
1094         iput(req->r_old_inode);
1095         ceph_mdsc_release_dir_caps(req);
1096 }
1097
1098 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1099 {
1100         struct ceph_inode_info *ci = ceph_inode(dir);
1101         struct ceph_dentry_info *di;
1102         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1103
1104         spin_lock(&ci->i_ceph_lock);
1105         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1106                 ceph_take_cap_refs(ci, want, false);
1107                 got = want;
1108         }
1109         spin_unlock(&ci->i_ceph_lock);
1110
1111         /* If we didn't get anything, return 0 */
1112         if (!got)
1113                 return 0;
1114
1115         spin_lock(&dentry->d_lock);
1116         di = ceph_dentry(dentry);
1117         /*
1118          * - We are holding Fx, which implies Fs caps.
1119          * - Only support async unlink for primary linkage
1120          */
1121         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1122             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1123                 want = 0;
1124         spin_unlock(&dentry->d_lock);
1125
1126         /* Do we still want what we've got? */
1127         if (want == got)
1128                 return got;
1129
1130         ceph_put_cap_refs(ci, got);
1131         return 0;
1132 }
1133
1134 /*
1135  * rmdir and unlink are differ only by the metadata op code
1136  */
1137 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1138 {
1139         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1140         struct ceph_mds_client *mdsc = fsc->mdsc;
1141         struct inode *inode = d_inode(dentry);
1142         struct ceph_mds_request *req;
1143         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1144         int err = -EROFS;
1145         int op;
1146
1147         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1148                 /* rmdir .snap/foo is RMSNAP */
1149                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1150                 op = CEPH_MDS_OP_RMSNAP;
1151         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1152                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1153                      dir, dentry, inode);
1154                 op = d_is_dir(dentry) ?
1155                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1156         } else
1157                 goto out;
1158 retry:
1159         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1160         if (IS_ERR(req)) {
1161                 err = PTR_ERR(req);
1162                 goto out;
1163         }
1164         req->r_dentry = dget(dentry);
1165         req->r_num_caps = 2;
1166         req->r_parent = dir;
1167         ihold(dir);
1168         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1169         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1170         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1171
1172         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1173             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1174                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1175                      dentry->d_name.len, dentry->d_name.name,
1176                      ceph_cap_string(req->r_dir_caps));
1177                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1178                 req->r_callback = ceph_async_unlink_cb;
1179                 req->r_old_inode = d_inode(dentry);
1180                 ihold(req->r_old_inode);
1181                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1182                 if (!err) {
1183                         /*
1184                          * We have enough caps, so we assume that the unlink
1185                          * will succeed. Fix up the target inode and dcache.
1186                          */
1187                         drop_nlink(inode);
1188                         d_delete(dentry);
1189                 } else if (err == -EJUKEBOX) {
1190                         try_async = false;
1191                         ceph_mdsc_put_request(req);
1192                         goto retry;
1193                 }
1194         } else {
1195                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1196                 err = ceph_mdsc_do_request(mdsc, dir, req);
1197                 if (!err && !req->r_reply_info.head->is_dentry)
1198                         d_delete(dentry);
1199         }
1200
1201         ceph_mdsc_put_request(req);
1202 out:
1203         return err;
1204 }
1205
1206 static int ceph_rename(struct user_namespace *mnt_userns, struct inode *old_dir,
1207                        struct dentry *old_dentry, struct inode *new_dir,
1208                        struct dentry *new_dentry, unsigned int flags)
1209 {
1210         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1211         struct ceph_mds_request *req;
1212         int op = CEPH_MDS_OP_RENAME;
1213         int err;
1214
1215         if (flags)
1216                 return -EINVAL;
1217
1218         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1219                 return -EXDEV;
1220         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1221                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1222                         op = CEPH_MDS_OP_RENAMESNAP;
1223                 else
1224                         return -EROFS;
1225         }
1226         /* don't allow cross-quota renames */
1227         if ((old_dir != new_dir) &&
1228             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1229                 return -EXDEV;
1230
1231         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1232              old_dir, old_dentry, new_dir, new_dentry);
1233         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1234         if (IS_ERR(req))
1235                 return PTR_ERR(req);
1236         ihold(old_dir);
1237         req->r_dentry = dget(new_dentry);
1238         req->r_num_caps = 2;
1239         req->r_old_dentry = dget(old_dentry);
1240         req->r_old_dentry_dir = old_dir;
1241         req->r_parent = new_dir;
1242         ihold(new_dir);
1243         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1244         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1245         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1246         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1247         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1248         /* release LINK_RDCACHE on source inode (mds will lock it) */
1249         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1250         if (d_really_is_positive(new_dentry)) {
1251                 req->r_inode_drop =
1252                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1253         }
1254         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1255         if (!err && !req->r_reply_info.head->is_dentry) {
1256                 /*
1257                  * Normally d_move() is done by fill_trace (called by
1258                  * do_request, above).  If there is no trace, we need
1259                  * to do it here.
1260                  */
1261                 d_move(old_dentry, new_dentry);
1262         }
1263         ceph_mdsc_put_request(req);
1264         return err;
1265 }
1266
1267 /*
1268  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1269  * Leases at front of the list will expire first. (Assume all leases have
1270  * similar duration)
1271  *
1272  * Called under dentry->d_lock.
1273  */
1274 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1275 {
1276         struct dentry *dn = di->dentry;
1277         struct ceph_mds_client *mdsc;
1278
1279         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1280
1281         di->flags |= CEPH_DENTRY_LEASE_LIST;
1282         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1283                 di->flags |= CEPH_DENTRY_REFERENCED;
1284                 return;
1285         }
1286
1287         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1288         spin_lock(&mdsc->dentry_list_lock);
1289         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1290         spin_unlock(&mdsc->dentry_list_lock);
1291 }
1292
1293 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1294                                      struct ceph_dentry_info *di)
1295 {
1296         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1297         di->lease_gen = 0;
1298         di->time = jiffies;
1299         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1300 }
1301
1302 /*
1303  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1304  * list if it's not in the list, otherwise set 'referenced' flag.
1305  *
1306  * Called under dentry->d_lock.
1307  */
1308 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1309 {
1310         struct dentry *dn = di->dentry;
1311         struct ceph_mds_client *mdsc;
1312
1313         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1314              di, dn, dn, di->offset);
1315
1316         if (!list_empty(&di->lease_list)) {
1317                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1318                         /* don't remove dentry from dentry lease list
1319                          * if its lease is valid */
1320                         if (__dentry_lease_is_valid(di))
1321                                 return;
1322                 } else {
1323                         di->flags |= CEPH_DENTRY_REFERENCED;
1324                         return;
1325                 }
1326         }
1327
1328         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1329                 di->flags |= CEPH_DENTRY_REFERENCED;
1330                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1331                 return;
1332         }
1333
1334         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1335         spin_lock(&mdsc->dentry_list_lock);
1336         __dentry_dir_lease_touch(mdsc, di),
1337         spin_unlock(&mdsc->dentry_list_lock);
1338 }
1339
1340 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1341 {
1342         struct ceph_mds_client *mdsc;
1343         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1344                 return;
1345         if (list_empty(&di->lease_list))
1346                 return;
1347
1348         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1349         spin_lock(&mdsc->dentry_list_lock);
1350         list_del_init(&di->lease_list);
1351         spin_unlock(&mdsc->dentry_list_lock);
1352 }
1353
1354 enum {
1355         KEEP    = 0,
1356         DELETE  = 1,
1357         TOUCH   = 2,
1358         STOP    = 4,
1359 };
1360
1361 struct ceph_lease_walk_control {
1362         bool dir_lease;
1363         bool expire_dir_lease;
1364         unsigned long nr_to_scan;
1365         unsigned long dir_lease_ttl;
1366 };
1367
1368 static unsigned long
1369 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1370                      struct ceph_lease_walk_control *lwc,
1371                      int (*check)(struct dentry*, void*))
1372 {
1373         struct ceph_dentry_info *di, *tmp;
1374         struct dentry *dentry, *last = NULL;
1375         struct list_head* list;
1376         LIST_HEAD(dispose);
1377         unsigned long freed = 0;
1378         int ret = 0;
1379
1380         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1381         spin_lock(&mdsc->dentry_list_lock);
1382         list_for_each_entry_safe(di, tmp, list, lease_list) {
1383                 if (!lwc->nr_to_scan)
1384                         break;
1385                 --lwc->nr_to_scan;
1386
1387                 dentry = di->dentry;
1388                 if (last == dentry)
1389                         break;
1390
1391                 if (!spin_trylock(&dentry->d_lock))
1392                         continue;
1393
1394                 if (__lockref_is_dead(&dentry->d_lockref)) {
1395                         list_del_init(&di->lease_list);
1396                         goto next;
1397                 }
1398
1399                 ret = check(dentry, lwc);
1400                 if (ret & TOUCH) {
1401                         /* move it into tail of dir lease list */
1402                         __dentry_dir_lease_touch(mdsc, di);
1403                         if (!last)
1404                                 last = dentry;
1405                 }
1406                 if (ret & DELETE) {
1407                         /* stale lease */
1408                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1409                         if (dentry->d_lockref.count > 0) {
1410                                 /* update_dentry_lease() will re-add
1411                                  * it to lease list, or
1412                                  * ceph_d_delete() will return 1 when
1413                                  * last reference is dropped */
1414                                 list_del_init(&di->lease_list);
1415                         } else {
1416                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1417                                 list_move_tail(&di->lease_list, &dispose);
1418                                 dget_dlock(dentry);
1419                         }
1420                 }
1421 next:
1422                 spin_unlock(&dentry->d_lock);
1423                 if (ret & STOP)
1424                         break;
1425         }
1426         spin_unlock(&mdsc->dentry_list_lock);
1427
1428         while (!list_empty(&dispose)) {
1429                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1430                                       lease_list);
1431                 dentry = di->dentry;
1432                 spin_lock(&dentry->d_lock);
1433
1434                 list_del_init(&di->lease_list);
1435                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1436                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1437                         spin_lock(&mdsc->dentry_list_lock);
1438                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1439                                 list_add_tail(&di->lease_list,
1440                                               &mdsc->dentry_leases);
1441                         } else {
1442                                 __dentry_dir_lease_touch(mdsc, di);
1443                         }
1444                         spin_unlock(&mdsc->dentry_list_lock);
1445                 } else {
1446                         freed++;
1447                 }
1448
1449                 spin_unlock(&dentry->d_lock);
1450                 /* ceph_d_delete() does the trick */
1451                 dput(dentry);
1452         }
1453         return freed;
1454 }
1455
1456 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1457 {
1458         struct ceph_dentry_info *di = ceph_dentry(dentry);
1459         int ret;
1460
1461         if (__dentry_lease_is_valid(di))
1462                 return STOP;
1463         ret = __dir_lease_try_check(dentry);
1464         if (ret == -EBUSY)
1465                 return KEEP;
1466         if (ret > 0)
1467                 return TOUCH;
1468         return DELETE;
1469 }
1470
1471 static int __dir_lease_check(struct dentry *dentry, void *arg)
1472 {
1473         struct ceph_lease_walk_control *lwc = arg;
1474         struct ceph_dentry_info *di = ceph_dentry(dentry);
1475
1476         int ret = __dir_lease_try_check(dentry);
1477         if (ret == -EBUSY)
1478                 return KEEP;
1479         if (ret > 0) {
1480                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1481                         return STOP;
1482                 /* Move dentry to tail of dir lease list if we don't want
1483                  * to delete it. So dentries in the list are checked in a
1484                  * round robin manner */
1485                 if (!lwc->expire_dir_lease)
1486                         return TOUCH;
1487                 if (dentry->d_lockref.count > 0 ||
1488                     (di->flags & CEPH_DENTRY_REFERENCED))
1489                         return TOUCH;
1490                 /* invalidate dir lease */
1491                 di->lease_shared_gen = 0;
1492         }
1493         return DELETE;
1494 }
1495
1496 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1497 {
1498         struct ceph_lease_walk_control lwc;
1499         unsigned long count;
1500         unsigned long freed;
1501
1502         spin_lock(&mdsc->caps_list_lock);
1503         if (mdsc->caps_use_max > 0 &&
1504             mdsc->caps_use_count > mdsc->caps_use_max)
1505                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1506         else
1507                 count = 0;
1508         spin_unlock(&mdsc->caps_list_lock);
1509
1510         lwc.dir_lease = false;
1511         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1512         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1513         if (!lwc.nr_to_scan) /* more invalid leases */
1514                 return -EAGAIN;
1515
1516         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1517                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1518
1519         lwc.dir_lease = true;
1520         lwc.expire_dir_lease = freed < count;
1521         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1522         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1523         if (!lwc.nr_to_scan) /* more to check */
1524                 return -EAGAIN;
1525
1526         return freed > 0 ? 1 : 0;
1527 }
1528
1529 /*
1530  * Ensure a dentry lease will no longer revalidate.
1531  */
1532 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1533 {
1534         struct ceph_dentry_info *di = ceph_dentry(dentry);
1535         spin_lock(&dentry->d_lock);
1536         di->time = jiffies;
1537         di->lease_shared_gen = 0;
1538         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1539         __dentry_lease_unlist(di);
1540         spin_unlock(&dentry->d_lock);
1541 }
1542
1543 /*
1544  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1545  * renew if the least is more than half up.
1546  */
1547 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1548 {
1549         struct ceph_mds_session *session;
1550
1551         if (!di->lease_gen)
1552                 return false;
1553
1554         session = di->lease_session;
1555         if (session) {
1556                 u32 gen;
1557                 unsigned long ttl;
1558
1559                 gen = atomic_read(&session->s_cap_gen);
1560                 ttl = session->s_cap_ttl;
1561
1562                 if (di->lease_gen == gen &&
1563                     time_before(jiffies, ttl) &&
1564                     time_before(jiffies, di->time))
1565                         return true;
1566         }
1567         di->lease_gen = 0;
1568         return false;
1569 }
1570
1571 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1572 {
1573         struct ceph_dentry_info *di;
1574         struct ceph_mds_session *session = NULL;
1575         u32 seq = 0;
1576         int valid = 0;
1577
1578         spin_lock(&dentry->d_lock);
1579         di = ceph_dentry(dentry);
1580         if (di && __dentry_lease_is_valid(di)) {
1581                 valid = 1;
1582
1583                 if (di->lease_renew_after &&
1584                     time_after(jiffies, di->lease_renew_after)) {
1585                         /*
1586                          * We should renew. If we're in RCU walk mode
1587                          * though, we can't do that so just return
1588                          * -ECHILD.
1589                          */
1590                         if (flags & LOOKUP_RCU) {
1591                                 valid = -ECHILD;
1592                         } else {
1593                                 session = ceph_get_mds_session(di->lease_session);
1594                                 seq = di->lease_seq;
1595                                 di->lease_renew_after = 0;
1596                                 di->lease_renew_from = jiffies;
1597                         }
1598                 }
1599         }
1600         spin_unlock(&dentry->d_lock);
1601
1602         if (session) {
1603                 ceph_mdsc_lease_send_msg(session, dentry,
1604                                          CEPH_MDS_LEASE_RENEW, seq);
1605                 ceph_put_mds_session(session);
1606         }
1607         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1608         return valid;
1609 }
1610
1611 /*
1612  * Called under dentry->d_lock.
1613  */
1614 static int __dir_lease_try_check(const struct dentry *dentry)
1615 {
1616         struct ceph_dentry_info *di = ceph_dentry(dentry);
1617         struct inode *dir;
1618         struct ceph_inode_info *ci;
1619         int valid = 0;
1620
1621         if (!di->lease_shared_gen)
1622                 return 0;
1623         if (IS_ROOT(dentry))
1624                 return 0;
1625
1626         dir = d_inode(dentry->d_parent);
1627         ci = ceph_inode(dir);
1628
1629         if (spin_trylock(&ci->i_ceph_lock)) {
1630                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1631                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1632                         valid = 1;
1633                 spin_unlock(&ci->i_ceph_lock);
1634         } else {
1635                 valid = -EBUSY;
1636         }
1637
1638         if (!valid)
1639                 di->lease_shared_gen = 0;
1640         return valid;
1641 }
1642
1643 /*
1644  * Check if directory-wide content lease/cap is valid.
1645  */
1646 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1647                               struct ceph_mds_client *mdsc)
1648 {
1649         struct ceph_inode_info *ci = ceph_inode(dir);
1650         int valid;
1651         int shared_gen;
1652
1653         spin_lock(&ci->i_ceph_lock);
1654         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1655         if (valid) {
1656                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1657                 shared_gen = atomic_read(&ci->i_shared_gen);
1658         }
1659         spin_unlock(&ci->i_ceph_lock);
1660         if (valid) {
1661                 struct ceph_dentry_info *di;
1662                 spin_lock(&dentry->d_lock);
1663                 di = ceph_dentry(dentry);
1664                 if (dir == d_inode(dentry->d_parent) &&
1665                     di && di->lease_shared_gen == shared_gen)
1666                         __ceph_dentry_dir_lease_touch(di);
1667                 else
1668                         valid = 0;
1669                 spin_unlock(&dentry->d_lock);
1670         }
1671         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1672              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1673         return valid;
1674 }
1675
1676 /*
1677  * Check if cached dentry can be trusted.
1678  */
1679 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1680 {
1681         int valid = 0;
1682         struct dentry *parent;
1683         struct inode *dir, *inode;
1684         struct ceph_mds_client *mdsc;
1685
1686         if (flags & LOOKUP_RCU) {
1687                 parent = READ_ONCE(dentry->d_parent);
1688                 dir = d_inode_rcu(parent);
1689                 if (!dir)
1690                         return -ECHILD;
1691                 inode = d_inode_rcu(dentry);
1692         } else {
1693                 parent = dget_parent(dentry);
1694                 dir = d_inode(parent);
1695                 inode = d_inode(dentry);
1696         }
1697
1698         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1699              dentry, inode, ceph_dentry(dentry)->offset);
1700
1701         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1702
1703         /* always trust cached snapped dentries, snapdir dentry */
1704         if (ceph_snap(dir) != CEPH_NOSNAP) {
1705                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1706                      dentry, inode);
1707                 valid = 1;
1708         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1709                 valid = 1;
1710         } else {
1711                 valid = dentry_lease_is_valid(dentry, flags);
1712                 if (valid == -ECHILD)
1713                         return valid;
1714                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1715                         if (inode)
1716                                 valid = ceph_is_any_caps(inode);
1717                         else
1718                                 valid = 1;
1719                 }
1720         }
1721
1722         if (!valid) {
1723                 struct ceph_mds_request *req;
1724                 int op, err;
1725                 u32 mask;
1726
1727                 if (flags & LOOKUP_RCU)
1728                         return -ECHILD;
1729
1730                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1731
1732                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1733                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1734                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1735                 if (!IS_ERR(req)) {
1736                         req->r_dentry = dget(dentry);
1737                         req->r_num_caps = 2;
1738                         req->r_parent = dir;
1739                         ihold(dir);
1740
1741                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1742                         if (ceph_security_xattr_wanted(dir))
1743                                 mask |= CEPH_CAP_XATTR_SHARED;
1744                         req->r_args.getattr.mask = cpu_to_le32(mask);
1745
1746                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1747                         switch (err) {
1748                         case 0:
1749                                 if (d_really_is_positive(dentry) &&
1750                                     d_inode(dentry) == req->r_target_inode)
1751                                         valid = 1;
1752                                 break;
1753                         case -ENOENT:
1754                                 if (d_really_is_negative(dentry))
1755                                         valid = 1;
1756                                 fallthrough;
1757                         default:
1758                                 break;
1759                         }
1760                         ceph_mdsc_put_request(req);
1761                         dout("d_revalidate %p lookup result=%d\n",
1762                              dentry, err);
1763                 }
1764         } else {
1765                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1766         }
1767
1768         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1769         if (!valid)
1770                 ceph_dir_clear_complete(dir);
1771
1772         if (!(flags & LOOKUP_RCU))
1773                 dput(parent);
1774         return valid;
1775 }
1776
1777 /*
1778  * Delete unused dentry that doesn't have valid lease
1779  *
1780  * Called under dentry->d_lock.
1781  */
1782 static int ceph_d_delete(const struct dentry *dentry)
1783 {
1784         struct ceph_dentry_info *di;
1785
1786         /* won't release caps */
1787         if (d_really_is_negative(dentry))
1788                 return 0;
1789         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1790                 return 0;
1791         /* vaild lease? */
1792         di = ceph_dentry(dentry);
1793         if (di) {
1794                 if (__dentry_lease_is_valid(di))
1795                         return 0;
1796                 if (__dir_lease_try_check(dentry))
1797                         return 0;
1798         }
1799         return 1;
1800 }
1801
1802 /*
1803  * Release our ceph_dentry_info.
1804  */
1805 static void ceph_d_release(struct dentry *dentry)
1806 {
1807         struct ceph_dentry_info *di = ceph_dentry(dentry);
1808         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1809
1810         dout("d_release %p\n", dentry);
1811
1812         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1813
1814         spin_lock(&dentry->d_lock);
1815         __dentry_lease_unlist(di);
1816         dentry->d_fsdata = NULL;
1817         spin_unlock(&dentry->d_lock);
1818
1819         ceph_put_mds_session(di->lease_session);
1820         kmem_cache_free(ceph_dentry_cachep, di);
1821 }
1822
1823 /*
1824  * When the VFS prunes a dentry from the cache, we need to clear the
1825  * complete flag on the parent directory.
1826  *
1827  * Called under dentry->d_lock.
1828  */
1829 static void ceph_d_prune(struct dentry *dentry)
1830 {
1831         struct ceph_inode_info *dir_ci;
1832         struct ceph_dentry_info *di;
1833
1834         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1835
1836         /* do we have a valid parent? */
1837         if (IS_ROOT(dentry))
1838                 return;
1839
1840         /* we hold d_lock, so d_parent is stable */
1841         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1842         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1843                 return;
1844
1845         /* who calls d_delete() should also disable dcache readdir */
1846         if (d_really_is_negative(dentry))
1847                 return;
1848
1849         /* d_fsdata does not get cleared until d_release */
1850         if (!d_unhashed(dentry)) {
1851                 __ceph_dir_clear_complete(dir_ci);
1852                 return;
1853         }
1854
1855         /* Disable dcache readdir just in case that someone called d_drop()
1856          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1857          * properly (dcache readdir is still enabled) */
1858         di = ceph_dentry(dentry);
1859         if (di->offset > 0 &&
1860             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1861                 __ceph_dir_clear_ordered(dir_ci);
1862 }
1863
1864 /*
1865  * read() on a dir.  This weird interface hack only works if mounted
1866  * with '-o dirstat'.
1867  */
1868 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1869                              loff_t *ppos)
1870 {
1871         struct ceph_dir_file_info *dfi = file->private_data;
1872         struct inode *inode = file_inode(file);
1873         struct ceph_inode_info *ci = ceph_inode(inode);
1874         int left;
1875         const int bufsize = 1024;
1876
1877         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1878                 return -EISDIR;
1879
1880         if (!dfi->dir_info) {
1881                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1882                 if (!dfi->dir_info)
1883                         return -ENOMEM;
1884                 dfi->dir_info_len =
1885                         snprintf(dfi->dir_info, bufsize,
1886                                 "entries:   %20lld\n"
1887                                 " files:    %20lld\n"
1888                                 " subdirs:  %20lld\n"
1889                                 "rentries:  %20lld\n"
1890                                 " rfiles:   %20lld\n"
1891                                 " rsubdirs: %20lld\n"
1892                                 "rbytes:    %20lld\n"
1893                                 "rctime:    %10lld.%09ld\n",
1894                                 ci->i_files + ci->i_subdirs,
1895                                 ci->i_files,
1896                                 ci->i_subdirs,
1897                                 ci->i_rfiles + ci->i_rsubdirs,
1898                                 ci->i_rfiles,
1899                                 ci->i_rsubdirs,
1900                                 ci->i_rbytes,
1901                                 ci->i_rctime.tv_sec,
1902                                 ci->i_rctime.tv_nsec);
1903         }
1904
1905         if (*ppos >= dfi->dir_info_len)
1906                 return 0;
1907         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1908         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1909         if (left == size)
1910                 return -EFAULT;
1911         *ppos += (size - left);
1912         return size - left;
1913 }
1914
1915
1916
1917 /*
1918  * Return name hash for a given dentry.  This is dependent on
1919  * the parent directory's hash function.
1920  */
1921 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1922 {
1923         struct ceph_inode_info *dci = ceph_inode(dir);
1924         unsigned hash;
1925
1926         switch (dci->i_dir_layout.dl_dir_hash) {
1927         case 0: /* for backward compat */
1928         case CEPH_STR_HASH_LINUX:
1929                 return dn->d_name.hash;
1930
1931         default:
1932                 spin_lock(&dn->d_lock);
1933                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1934                                      dn->d_name.name, dn->d_name.len);
1935                 spin_unlock(&dn->d_lock);
1936                 return hash;
1937         }
1938 }
1939
1940 const struct file_operations ceph_dir_fops = {
1941         .read = ceph_read_dir,
1942         .iterate = ceph_readdir,
1943         .llseek = ceph_dir_llseek,
1944         .open = ceph_open,
1945         .release = ceph_release,
1946         .unlocked_ioctl = ceph_ioctl,
1947         .compat_ioctl = compat_ptr_ioctl,
1948         .fsync = ceph_fsync,
1949         .lock = ceph_lock,
1950         .flock = ceph_flock,
1951 };
1952
1953 const struct file_operations ceph_snapdir_fops = {
1954         .iterate = ceph_readdir,
1955         .llseek = ceph_dir_llseek,
1956         .open = ceph_open,
1957         .release = ceph_release,
1958 };
1959
1960 const struct inode_operations ceph_dir_iops = {
1961         .lookup = ceph_lookup,
1962         .permission = ceph_permission,
1963         .getattr = ceph_getattr,
1964         .setattr = ceph_setattr,
1965         .listxattr = ceph_listxattr,
1966         .get_acl = ceph_get_acl,
1967         .set_acl = ceph_set_acl,
1968         .mknod = ceph_mknod,
1969         .symlink = ceph_symlink,
1970         .mkdir = ceph_mkdir,
1971         .link = ceph_link,
1972         .unlink = ceph_unlink,
1973         .rmdir = ceph_unlink,
1974         .rename = ceph_rename,
1975         .create = ceph_create,
1976         .atomic_open = ceph_atomic_open,
1977 };
1978
1979 const struct inode_operations ceph_snapdir_iops = {
1980         .lookup = ceph_lookup,
1981         .permission = ceph_permission,
1982         .getattr = ceph_getattr,
1983         .mkdir = ceph_mkdir,
1984         .rmdir = ceph_unlink,
1985         .rename = ceph_rename,
1986 };
1987
1988 const struct dentry_operations ceph_dentry_ops = {
1989         .d_revalidate = ceph_d_revalidate,
1990         .d_delete = ceph_d_delete,
1991         .d_release = ceph_d_release,
1992         .d_prune = ceph_d_prune,
1993         .d_init = ceph_d_init,
1994 };