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, int err)
 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 (err == -ENOENT &&
 678             ceph_snap(parent) == CEPH_NOSNAP &&
 679             strcmp(dentry->d_name.name, fsc->mount_options->snapdir_name) == 0) {
 680                 struct dentry *res;
 681                 struct inode *inode = ceph_get_snapdir(parent);
 682
 683                 res = d_splice_alias(inode, dentry);
 684                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p. Spliced dentry %p\n",
 685                      dentry, dentry, inode, res);
 686                 if (res)
 687                         dentry = res;
 688         }
 689         return dentry;
 690 }
 691
 692 /*
 693  * Figure out final result of a lookup/open request.
 694  *
 695  * Mainly, make sure we return the final req->r_dentry (if it already
 696  * existed) in place of the original VFS-provided dentry when they
 697  * differ.
 698  *
 699  * Gracefully handle the case where the MDS replies with -ENOENT and
 700  * no trace (which it may do, at its discretion, e.g., if it doesn't
 701  * care to issue a lease on the negative dentry).
 702  */
 703 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 704                                   struct dentry *dentry, int err)
 705 {
 706         if (err == -ENOENT) {
 707                 /* no trace? */
 708                 err = 0;
 709                 if (!req->r_reply_info.head->is_dentry) {
 710                         dout("ENOENT and no trace, dentry %p inode %p\n",
 711                              dentry, d_inode(dentry));
 712                         if (d_really_is_positive(dentry)) {
 713                                 d_drop(dentry);
 714                                 err = -ENOENT;
 715                         } else {
 716                                 d_add(dentry, NULL);
 717                         }
 718                 }
 719         }
 720         if (err)
 721                 dentry = ERR_PTR(err);
 722         else if (dentry != req->r_dentry)
 723                 dentry = dget(req->r_dentry);   /* we got spliced */
 724         else
 725                 dentry = NULL;
 726         return dentry;
 727 }
 728
 729 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 730 {
 731         return ceph_ino(inode) == CEPH_INO_ROOT &&
 732                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 733 }
 734
 735 /*
 736  * Look up a single dir entry.  If there is a lookup intent, inform
 737  * the MDS so that it gets our 'caps wanted' value in a single op.
 738  */
 739 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 740                                   unsigned int flags)
 741 {
 742         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 743         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 744         struct ceph_mds_request *req;
 745         struct dentry *res;
 746         int op;
 747         int mask;
 748         int err;
 749
 750         dout("lookup %p dentry %p '%pd'\n",
 751              dir, dentry, dentry);
 752
 753         if (dentry->d_name.len > NAME_MAX)
 754                 return ERR_PTR(-ENAMETOOLONG);
 755
 756         /* can we conclude ENOENT locally? */
 757         if (d_really_is_negative(dentry)) {
 758                 struct ceph_inode_info *ci = ceph_inode(dir);
 759                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 760
 761                 spin_lock(&ci->i_ceph_lock);
 762                 dout(" dir %p flags are 0x%lx\n", dir, ci->i_ceph_flags);
 763                 if (strncmp(dentry->d_name.name,
 764                             fsc->mount_options->snapdir_name,
 765                             dentry->d_name.len) &&
 766                     !is_root_ceph_dentry(dir, dentry) &&
 767                     ceph_test_mount_opt(fsc, DCACHE) &&
 768                     __ceph_dir_is_complete(ci) &&
 769                     __ceph_caps_issued_mask_metric(ci, CEPH_CAP_FILE_SHARED, 1)) {
 770                         __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
 771                         spin_unlock(&ci->i_ceph_lock);
 772                         dout(" dir %p complete, -ENOENT\n", dir);
 773                         d_add(dentry, NULL);
 774                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 775                         return NULL;
 776                 }
 777                 spin_unlock(&ci->i_ceph_lock);
 778         }
 779
 780         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 781                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 782         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 783         if (IS_ERR(req))
 784                 return ERR_CAST(req);
 785         req->r_dentry = dget(dentry);
 786         req->r_num_caps = 2;
 787
 788         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 789         if (ceph_security_xattr_wanted(dir))
 790                 mask |= CEPH_CAP_XATTR_SHARED;
 791         req->r_args.getattr.mask = cpu_to_le32(mask);
 792
 793         req->r_parent = dir;
 794         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 795         err = ceph_mdsc_do_request(mdsc, NULL, req);
 796         res = ceph_handle_snapdir(req, dentry, err);
 797         if (IS_ERR(res)) {
 798                 err = PTR_ERR(res);
 799         } else {
 800                 dentry = res;
 801                 err = 0;
 802         }
 803         dentry = ceph_finish_lookup(req, dentry, err);
 804         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 805         dout("lookup result=%p\n", dentry);
 806         return dentry;
 807 }
 808
 809 /*
 810  * If we do a create but get no trace back from the MDS, follow up with
 811  * a lookup (the VFS expects us to link up the provided dentry).
 812  */
 813 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 814 {
 815         struct dentry *result = ceph_lookup(dir, dentry, 0);
 816
 817         if (result && !IS_ERR(result)) {
 818                 /*
 819                  * We created the item, then did a lookup, and found
 820                  * it was already linked to another inode we already
 821                  * had in our cache (and thus got spliced). To not
 822                  * confuse VFS (especially when inode is a directory),
 823                  * we don't link our dentry to that inode, return an
 824                  * error instead.
 825                  *
 826                  * This event should be rare and it happens only when
 827                  * we talk to old MDS. Recent MDS does not send traceless
 828                  * reply for request that creates new inode.
 829                  */
 830                 d_drop(result);
 831                 return -ESTALE;
 832         }
 833         return PTR_ERR(result);
 834 }
 835
 836 static int ceph_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 837                       struct dentry *dentry, umode_t mode, dev_t rdev)
 838 {
 839         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 840         struct ceph_mds_request *req;
 841         struct ceph_acl_sec_ctx as_ctx = {};
 842         int err;
 843
 844         if (ceph_snap(dir) != CEPH_NOSNAP)
 845                 return -EROFS;
 846
 847         if (ceph_quota_is_max_files_exceeded(dir)) {
 848                 err = -EDQUOT;
 849                 goto out;
 850         }
 851
 852         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 853         if (err < 0)
 854                 goto out;
 855         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 856         if (err < 0)
 857                 goto out;
 858
 859         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 860              dir, dentry, mode, rdev);
 861         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 862         if (IS_ERR(req)) {
 863                 err = PTR_ERR(req);
 864                 goto out;
 865         }
 866         req->r_dentry = dget(dentry);
 867         req->r_num_caps = 2;
 868         req->r_parent = dir;
 869         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 870         req->r_args.mknod.mode = cpu_to_le32(mode);
 871         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 872         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 873         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 874         if (as_ctx.pagelist) {
 875                 req->r_pagelist = as_ctx.pagelist;
 876                 as_ctx.pagelist = NULL;
 877         }
 878         err = ceph_mdsc_do_request(mdsc, dir, req);
 879         if (!err && !req->r_reply_info.head->is_dentry)
 880                 err = ceph_handle_notrace_create(dir, dentry);
 881         ceph_mdsc_put_request(req);
 882 out:
 883         if (!err)
 884                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
 885         else
 886                 d_drop(dentry);
 887         ceph_release_acl_sec_ctx(&as_ctx);
 888         return err;
 889 }
 890
 891 static int ceph_create(struct user_namespace *mnt_userns, struct inode *dir,
 892                        struct dentry *dentry, umode_t mode, bool excl)
 893 {
 894         return ceph_mknod(mnt_userns, dir, dentry, mode, 0);
 895 }
 896
 897 static int ceph_symlink(struct user_namespace *mnt_userns, struct inode *dir,
 898                         struct dentry *dentry, const char *dest)
 899 {
 900         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 901         struct ceph_mds_request *req;
 902         struct ceph_acl_sec_ctx as_ctx = {};
 903         int err;
 904
 905         if (ceph_snap(dir) != CEPH_NOSNAP)
 906                 return -EROFS;
 907
 908         if (ceph_quota_is_max_files_exceeded(dir)) {
 909                 err = -EDQUOT;
 910                 goto out;
 911         }
 912
 913         err = ceph_security_init_secctx(dentry, S_IFLNK | 0777, &as_ctx);
 914         if (err < 0)
 915                 goto out;
 916
 917         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 918         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 919         if (IS_ERR(req)) {
 920                 err = PTR_ERR(req);
 921                 goto out;
 922         }
 923         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 924         if (!req->r_path2) {
 925                 err = -ENOMEM;
 926                 ceph_mdsc_put_request(req);
 927                 goto out;
 928         }
 929         req->r_parent = dir;
 930         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 931         req->r_dentry = dget(dentry);
 932         req->r_num_caps = 2;
 933         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 934         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 935         if (as_ctx.pagelist) {
 936                 req->r_pagelist = as_ctx.pagelist;
 937                 as_ctx.pagelist = NULL;
 938         }
 939         err = ceph_mdsc_do_request(mdsc, dir, req);
 940         if (!err && !req->r_reply_info.head->is_dentry)
 941                 err = ceph_handle_notrace_create(dir, dentry);
 942         ceph_mdsc_put_request(req);
 943 out:
 944         if (err)
 945                 d_drop(dentry);
 946         ceph_release_acl_sec_ctx(&as_ctx);
 947         return err;
 948 }
 949
 950 static int ceph_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 951                       struct dentry *dentry, umode_t mode)
 952 {
 953         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
 954         struct ceph_mds_request *req;
 955         struct ceph_acl_sec_ctx as_ctx = {};
 956         int err = -EROFS;
 957         int op;
 958
 959         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 960                 /* mkdir .snap/foo is a MKSNAP */
 961                 op = CEPH_MDS_OP_MKSNAP;
 962                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 963                      dentry, dentry);
 964         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 965                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 966                 op = CEPH_MDS_OP_MKDIR;
 967         } else {
 968                 goto out;
 969         }
 970
 971         if (op == CEPH_MDS_OP_MKDIR &&
 972             ceph_quota_is_max_files_exceeded(dir)) {
 973                 err = -EDQUOT;
 974                 goto out;
 975         }
 976
 977         mode |= S_IFDIR;
 978         err = ceph_pre_init_acls(dir, &mode, &as_ctx);
 979         if (err < 0)
 980                 goto out;
 981         err = ceph_security_init_secctx(dentry, mode, &as_ctx);
 982         if (err < 0)
 983                 goto out;
 984
 985         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 986         if (IS_ERR(req)) {
 987                 err = PTR_ERR(req);
 988                 goto out;
 989         }
 990
 991         req->r_dentry = dget(dentry);
 992         req->r_num_caps = 2;
 993         req->r_parent = dir;
 994         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 995         req->r_args.mkdir.mode = cpu_to_le32(mode);
 996         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 997         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 998         if (as_ctx.pagelist) {
 999                 req->r_pagelist = as_ctx.pagelist;
1000                 as_ctx.pagelist = NULL;
1001         }
1002         err = ceph_mdsc_do_request(mdsc, dir, req);
1003         if (!err &&
1004             !req->r_reply_info.head->is_target &&
1005             !req->r_reply_info.head->is_dentry)
1006                 err = ceph_handle_notrace_create(dir, dentry);
1007         ceph_mdsc_put_request(req);
1008 out:
1009         if (!err)
1010                 ceph_init_inode_acls(d_inode(dentry), &as_ctx);
1011         else
1012                 d_drop(dentry);
1013         ceph_release_acl_sec_ctx(&as_ctx);
1014         return err;
1015 }
1016
1017 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
1018                      struct dentry *dentry)
1019 {
1020         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
1021         struct ceph_mds_request *req;
1022         int err;
1023
1024         if (ceph_snap(dir) != CEPH_NOSNAP)
1025                 return -EROFS;
1026
1027         dout("link in dir %p old_dentry %p dentry %p\n", dir,
1028              old_dentry, dentry);
1029         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
1030         if (IS_ERR(req)) {
1031                 d_drop(dentry);
1032                 return PTR_ERR(req);
1033         }
1034         req->r_dentry = dget(dentry);
1035         req->r_num_caps = 2;
1036         req->r_old_dentry = dget(old_dentry);
1037         req->r_parent = dir;
1038         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1039         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1040         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1041         /* release LINK_SHARED on source inode (mds will lock it) */
1042         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1043         err = ceph_mdsc_do_request(mdsc, dir, req);
1044         if (err) {
1045                 d_drop(dentry);
1046         } else if (!req->r_reply_info.head->is_dentry) {
1047                 ihold(d_inode(old_dentry));
1048                 d_instantiate(dentry, d_inode(old_dentry));
1049         }
1050         ceph_mdsc_put_request(req);
1051         return err;
1052 }
1053
1054 static void ceph_async_unlink_cb(struct ceph_mds_client *mdsc,
1055                                  struct ceph_mds_request *req)
1056 {
1057         int result = req->r_err ? req->r_err :
1058                         le32_to_cpu(req->r_reply_info.head->result);
1059
1060         if (result == -EJUKEBOX)
1061                 goto out;
1062
1063         /* If op failed, mark everyone involved for errors */
1064         if (result) {
1065                 int pathlen = 0;
1066                 u64 base = 0;
1067                 char *path = ceph_mdsc_build_path(req->r_dentry, &pathlen,
1068                                                   &base, 0);
1069
1070                 /* mark error on parent + clear complete */
1071                 mapping_set_error(req->r_parent->i_mapping, result);
1072                 ceph_dir_clear_complete(req->r_parent);
1073
1074                 /* drop the dentry -- we don't know its status */
1075                 if (!d_unhashed(req->r_dentry))
1076                         d_drop(req->r_dentry);
1077
1078                 /* mark inode itself for an error (since metadata is bogus) */
1079                 mapping_set_error(req->r_old_inode->i_mapping, result);
1080
1081                 pr_warn("ceph: async unlink failure path=(%llx)%s result=%d!\n",
1082                         base, IS_ERR(path) ? "<<bad>>" : path, result);
1083                 ceph_mdsc_free_path(path, pathlen);
1084         }
1085 out:
1086         iput(req->r_old_inode);
1087         ceph_mdsc_release_dir_caps(req);
1088 }
1089
1090 static int get_caps_for_async_unlink(struct inode *dir, struct dentry *dentry)
1091 {
1092         struct ceph_inode_info *ci = ceph_inode(dir);
1093         struct ceph_dentry_info *di;
1094         int got = 0, want = CEPH_CAP_FILE_EXCL | CEPH_CAP_DIR_UNLINK;
1095
1096         spin_lock(&ci->i_ceph_lock);
1097         if ((__ceph_caps_issued(ci, NULL) & want) == want) {
1098                 ceph_take_cap_refs(ci, want, false);
1099                 got = want;
1100         }
1101         spin_unlock(&ci->i_ceph_lock);
1102
1103         /* If we didn't get anything, return 0 */
1104         if (!got)
1105                 return 0;
1106
1107         spin_lock(&dentry->d_lock);
1108         di = ceph_dentry(dentry);
1109         /*
1110          * - We are holding Fx, which implies Fs caps.
1111          * - Only support async unlink for primary linkage
1112          */
1113         if (atomic_read(&ci->i_shared_gen) != di->lease_shared_gen ||
1114             !(di->flags & CEPH_DENTRY_PRIMARY_LINK))
1115                 want = 0;
1116         spin_unlock(&dentry->d_lock);
1117
1118         /* Do we still want what we've got? */
1119         if (want == got)
1120                 return got;
1121
1122         ceph_put_cap_refs(ci, got);
1123         return 0;
1124 }
1125
1126 /*
1127  * rmdir and unlink are differ only by the metadata op code
1128  */
1129 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1130 {
1131         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1132         struct ceph_mds_client *mdsc = fsc->mdsc;
1133         struct inode *inode = d_inode(dentry);
1134         struct ceph_mds_request *req;
1135         bool try_async = ceph_test_mount_opt(fsc, ASYNC_DIROPS);
1136         int err = -EROFS;
1137         int op;
1138
1139         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1140                 /* rmdir .snap/foo is RMSNAP */
1141                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1142                 op = CEPH_MDS_OP_RMSNAP;
1143         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1144                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1145                      dir, dentry, inode);
1146                 op = d_is_dir(dentry) ?
1147                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1148         } else
1149                 goto out;
1150 retry:
1151         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1152         if (IS_ERR(req)) {
1153                 err = PTR_ERR(req);
1154                 goto out;
1155         }
1156         req->r_dentry = dget(dentry);
1157         req->r_num_caps = 2;
1158         req->r_parent = dir;
1159         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1160         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1161         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1162
1163         if (try_async && op == CEPH_MDS_OP_UNLINK &&
1164             (req->r_dir_caps = get_caps_for_async_unlink(dir, dentry))) {
1165                 dout("async unlink on %llu/%.*s caps=%s", ceph_ino(dir),
1166                      dentry->d_name.len, dentry->d_name.name,
1167                      ceph_cap_string(req->r_dir_caps));
1168                 set_bit(CEPH_MDS_R_ASYNC, &req->r_req_flags);
1169                 req->r_callback = ceph_async_unlink_cb;
1170                 req->r_old_inode = d_inode(dentry);
1171                 ihold(req->r_old_inode);
1172                 err = ceph_mdsc_submit_request(mdsc, dir, req);
1173                 if (!err) {
1174                         /*
1175                          * We have enough caps, so we assume that the unlink
1176                          * will succeed. Fix up the target inode and dcache.
1177                          */
1178                         drop_nlink(inode);
1179                         d_delete(dentry);
1180                 } else if (err == -EJUKEBOX) {
1181                         try_async = false;
1182                         ceph_mdsc_put_request(req);
1183                         goto retry;
1184                 }
1185         } else {
1186                 set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1187                 err = ceph_mdsc_do_request(mdsc, dir, req);
1188                 if (!err && !req->r_reply_info.head->is_dentry)
1189                         d_delete(dentry);
1190         }
1191
1192         ceph_mdsc_put_request(req);
1193 out:
1194         return err;
1195 }
1196
1197 static int ceph_rename(struct user_namespace *mnt_userns, struct inode *old_dir,
1198                        struct dentry *old_dentry, struct inode *new_dir,
1199                        struct dentry *new_dentry, unsigned int flags)
1200 {
1201         struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(old_dir->i_sb);
1202         struct ceph_mds_request *req;
1203         int op = CEPH_MDS_OP_RENAME;
1204         int err;
1205
1206         if (flags)
1207                 return -EINVAL;
1208
1209         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1210                 return -EXDEV;
1211         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1212                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1213                         op = CEPH_MDS_OP_RENAMESNAP;
1214                 else
1215                         return -EROFS;
1216         }
1217         /* don't allow cross-quota renames */
1218         if ((old_dir != new_dir) &&
1219             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1220                 return -EXDEV;
1221
1222         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1223              old_dir, old_dentry, new_dir, new_dentry);
1224         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1225         if (IS_ERR(req))
1226                 return PTR_ERR(req);
1227         ihold(old_dir);
1228         req->r_dentry = dget(new_dentry);
1229         req->r_num_caps = 2;
1230         req->r_old_dentry = dget(old_dentry);
1231         req->r_old_dentry_dir = old_dir;
1232         req->r_parent = new_dir;
1233         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1234         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1235         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1236         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1237         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1238         /* release LINK_RDCACHE on source inode (mds will lock it) */
1239         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1240         if (d_really_is_positive(new_dentry)) {
1241                 req->r_inode_drop =
1242                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1243         }
1244         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1245         if (!err && !req->r_reply_info.head->is_dentry) {
1246                 /*
1247                  * Normally d_move() is done by fill_trace (called by
1248                  * do_request, above).  If there is no trace, we need
1249                  * to do it here.
1250                  */
1251                 d_move(old_dentry, new_dentry);
1252         }
1253         ceph_mdsc_put_request(req);
1254         return err;
1255 }
1256
1257 /*
1258  * Move dentry to tail of mdsc->dentry_leases list when lease is updated.
1259  * Leases at front of the list will expire first. (Assume all leases have
1260  * similar duration)
1261  *
1262  * Called under dentry->d_lock.
1263  */
1264 void __ceph_dentry_lease_touch(struct ceph_dentry_info *di)
1265 {
1266         struct dentry *dn = di->dentry;
1267         struct ceph_mds_client *mdsc;
1268
1269         dout("dentry_lease_touch %p %p '%pd'\n", di, dn, dn);
1270
1271         di->flags |= CEPH_DENTRY_LEASE_LIST;
1272         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1273                 di->flags |= CEPH_DENTRY_REFERENCED;
1274                 return;
1275         }
1276
1277         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1278         spin_lock(&mdsc->dentry_list_lock);
1279         list_move_tail(&di->lease_list, &mdsc->dentry_leases);
1280         spin_unlock(&mdsc->dentry_list_lock);
1281 }
1282
1283 static void __dentry_dir_lease_touch(struct ceph_mds_client* mdsc,
1284                                      struct ceph_dentry_info *di)
1285 {
1286         di->flags &= ~(CEPH_DENTRY_LEASE_LIST | CEPH_DENTRY_REFERENCED);
1287         di->lease_gen = 0;
1288         di->time = jiffies;
1289         list_move_tail(&di->lease_list, &mdsc->dentry_dir_leases);
1290 }
1291
1292 /*
1293  * When dir lease is used, add dentry to tail of mdsc->dentry_dir_leases
1294  * list if it's not in the list, otherwise set 'referenced' flag.
1295  *
1296  * Called under dentry->d_lock.
1297  */
1298 void __ceph_dentry_dir_lease_touch(struct ceph_dentry_info *di)
1299 {
1300         struct dentry *dn = di->dentry;
1301         struct ceph_mds_client *mdsc;
1302
1303         dout("dentry_dir_lease_touch %p %p '%pd' (offset 0x%llx)\n",
1304              di, dn, dn, di->offset);
1305
1306         if (!list_empty(&di->lease_list)) {
1307                 if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1308                         /* don't remove dentry from dentry lease list
1309                          * if its lease is valid */
1310                         if (__dentry_lease_is_valid(di))
1311                                 return;
1312                 } else {
1313                         di->flags |= CEPH_DENTRY_REFERENCED;
1314                         return;
1315                 }
1316         }
1317
1318         if (di->flags & CEPH_DENTRY_SHRINK_LIST) {
1319                 di->flags |= CEPH_DENTRY_REFERENCED;
1320                 di->flags &= ~CEPH_DENTRY_LEASE_LIST;
1321                 return;
1322         }
1323
1324         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1325         spin_lock(&mdsc->dentry_list_lock);
1326         __dentry_dir_lease_touch(mdsc, di),
1327         spin_unlock(&mdsc->dentry_list_lock);
1328 }
1329
1330 static void __dentry_lease_unlist(struct ceph_dentry_info *di)
1331 {
1332         struct ceph_mds_client *mdsc;
1333         if (di->flags & CEPH_DENTRY_SHRINK_LIST)
1334                 return;
1335         if (list_empty(&di->lease_list))
1336                 return;
1337
1338         mdsc = ceph_sb_to_client(di->dentry->d_sb)->mdsc;
1339         spin_lock(&mdsc->dentry_list_lock);
1340         list_del_init(&di->lease_list);
1341         spin_unlock(&mdsc->dentry_list_lock);
1342 }
1343
1344 enum {
1345         KEEP    = 0,
1346         DELETE  = 1,
1347         TOUCH   = 2,
1348         STOP    = 4,
1349 };
1350
1351 struct ceph_lease_walk_control {
1352         bool dir_lease;
1353         bool expire_dir_lease;
1354         unsigned long nr_to_scan;
1355         unsigned long dir_lease_ttl;
1356 };
1357
1358 static unsigned long
1359 __dentry_leases_walk(struct ceph_mds_client *mdsc,
1360                      struct ceph_lease_walk_control *lwc,
1361                      int (*check)(struct dentry*, void*))
1362 {
1363         struct ceph_dentry_info *di, *tmp;
1364         struct dentry *dentry, *last = NULL;
1365         struct list_head* list;
1366         LIST_HEAD(dispose);
1367         unsigned long freed = 0;
1368         int ret = 0;
1369
1370         list = lwc->dir_lease ? &mdsc->dentry_dir_leases : &mdsc->dentry_leases;
1371         spin_lock(&mdsc->dentry_list_lock);
1372         list_for_each_entry_safe(di, tmp, list, lease_list) {
1373                 if (!lwc->nr_to_scan)
1374                         break;
1375                 --lwc->nr_to_scan;
1376
1377                 dentry = di->dentry;
1378                 if (last == dentry)
1379                         break;
1380
1381                 if (!spin_trylock(&dentry->d_lock))
1382                         continue;
1383
1384                 if (__lockref_is_dead(&dentry->d_lockref)) {
1385                         list_del_init(&di->lease_list);
1386                         goto next;
1387                 }
1388
1389                 ret = check(dentry, lwc);
1390                 if (ret & TOUCH) {
1391                         /* move it into tail of dir lease list */
1392                         __dentry_dir_lease_touch(mdsc, di);
1393                         if (!last)
1394                                 last = dentry;
1395                 }
1396                 if (ret & DELETE) {
1397                         /* stale lease */
1398                         di->flags &= ~CEPH_DENTRY_REFERENCED;
1399                         if (dentry->d_lockref.count > 0) {
1400                                 /* update_dentry_lease() will re-add
1401                                  * it to lease list, or
1402                                  * ceph_d_delete() will return 1 when
1403                                  * last reference is dropped */
1404                                 list_del_init(&di->lease_list);
1405                         } else {
1406                                 di->flags |= CEPH_DENTRY_SHRINK_LIST;
1407                                 list_move_tail(&di->lease_list, &dispose);
1408                                 dget_dlock(dentry);
1409                         }
1410                 }
1411 next:
1412                 spin_unlock(&dentry->d_lock);
1413                 if (ret & STOP)
1414                         break;
1415         }
1416         spin_unlock(&mdsc->dentry_list_lock);
1417
1418         while (!list_empty(&dispose)) {
1419                 di = list_first_entry(&dispose, struct ceph_dentry_info,
1420                                       lease_list);
1421                 dentry = di->dentry;
1422                 spin_lock(&dentry->d_lock);
1423
1424                 list_del_init(&di->lease_list);
1425                 di->flags &= ~CEPH_DENTRY_SHRINK_LIST;
1426                 if (di->flags & CEPH_DENTRY_REFERENCED) {
1427                         spin_lock(&mdsc->dentry_list_lock);
1428                         if (di->flags & CEPH_DENTRY_LEASE_LIST) {
1429                                 list_add_tail(&di->lease_list,
1430                                               &mdsc->dentry_leases);
1431                         } else {
1432                                 __dentry_dir_lease_touch(mdsc, di);
1433                         }
1434                         spin_unlock(&mdsc->dentry_list_lock);
1435                 } else {
1436                         freed++;
1437                 }
1438
1439                 spin_unlock(&dentry->d_lock);
1440                 /* ceph_d_delete() does the trick */
1441                 dput(dentry);
1442         }
1443         return freed;
1444 }
1445
1446 static int __dentry_lease_check(struct dentry *dentry, void *arg)
1447 {
1448         struct ceph_dentry_info *di = ceph_dentry(dentry);
1449         int ret;
1450
1451         if (__dentry_lease_is_valid(di))
1452                 return STOP;
1453         ret = __dir_lease_try_check(dentry);
1454         if (ret == -EBUSY)
1455                 return KEEP;
1456         if (ret > 0)
1457                 return TOUCH;
1458         return DELETE;
1459 }
1460
1461 static int __dir_lease_check(struct dentry *dentry, void *arg)
1462 {
1463         struct ceph_lease_walk_control *lwc = arg;
1464         struct ceph_dentry_info *di = ceph_dentry(dentry);
1465
1466         int ret = __dir_lease_try_check(dentry);
1467         if (ret == -EBUSY)
1468                 return KEEP;
1469         if (ret > 0) {
1470                 if (time_before(jiffies, di->time + lwc->dir_lease_ttl))
1471                         return STOP;
1472                 /* Move dentry to tail of dir lease list if we don't want
1473                  * to delete it. So dentries in the list are checked in a
1474                  * round robin manner */
1475                 if (!lwc->expire_dir_lease)
1476                         return TOUCH;
1477                 if (dentry->d_lockref.count > 0 ||
1478                     (di->flags & CEPH_DENTRY_REFERENCED))
1479                         return TOUCH;
1480                 /* invalidate dir lease */
1481                 di->lease_shared_gen = 0;
1482         }
1483         return DELETE;
1484 }
1485
1486 int ceph_trim_dentries(struct ceph_mds_client *mdsc)
1487 {
1488         struct ceph_lease_walk_control lwc;
1489         unsigned long count;
1490         unsigned long freed;
1491
1492         spin_lock(&mdsc->caps_list_lock);
1493         if (mdsc->caps_use_max > 0 &&
1494             mdsc->caps_use_count > mdsc->caps_use_max)
1495                 count = mdsc->caps_use_count - mdsc->caps_use_max;
1496         else
1497                 count = 0;
1498         spin_unlock(&mdsc->caps_list_lock);
1499
1500         lwc.dir_lease = false;
1501         lwc.nr_to_scan  = CEPH_CAPS_PER_RELEASE * 2;
1502         freed = __dentry_leases_walk(mdsc, &lwc, __dentry_lease_check);
1503         if (!lwc.nr_to_scan) /* more invalid leases */
1504                 return -EAGAIN;
1505
1506         if (lwc.nr_to_scan < CEPH_CAPS_PER_RELEASE)
1507                 lwc.nr_to_scan = CEPH_CAPS_PER_RELEASE;
1508
1509         lwc.dir_lease = true;
1510         lwc.expire_dir_lease = freed < count;
1511         lwc.dir_lease_ttl = mdsc->fsc->mount_options->caps_wanted_delay_max * HZ;
1512         freed +=__dentry_leases_walk(mdsc, &lwc, __dir_lease_check);
1513         if (!lwc.nr_to_scan) /* more to check */
1514                 return -EAGAIN;
1515
1516         return freed > 0 ? 1 : 0;
1517 }
1518
1519 /*
1520  * Ensure a dentry lease will no longer revalidate.
1521  */
1522 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1523 {
1524         struct ceph_dentry_info *di = ceph_dentry(dentry);
1525         spin_lock(&dentry->d_lock);
1526         di->time = jiffies;
1527         di->lease_shared_gen = 0;
1528         di->flags &= ~CEPH_DENTRY_PRIMARY_LINK;
1529         __dentry_lease_unlist(di);
1530         spin_unlock(&dentry->d_lock);
1531 }
1532
1533 /*
1534  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1535  * renew if the least is more than half up.
1536  */
1537 static bool __dentry_lease_is_valid(struct ceph_dentry_info *di)
1538 {
1539         struct ceph_mds_session *session;
1540
1541         if (!di->lease_gen)
1542                 return false;
1543
1544         session = di->lease_session;
1545         if (session) {
1546                 u32 gen;
1547                 unsigned long ttl;
1548
1549                 spin_lock(&session->s_gen_ttl_lock);
1550                 gen = session->s_cap_gen;
1551                 ttl = session->s_cap_ttl;
1552                 spin_unlock(&session->s_gen_ttl_lock);
1553
1554                 if (di->lease_gen == gen &&
1555                     time_before(jiffies, ttl) &&
1556                     time_before(jiffies, di->time))
1557                         return true;
1558         }
1559         di->lease_gen = 0;
1560         return false;
1561 }
1562
1563 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags)
1564 {
1565         struct ceph_dentry_info *di;
1566         struct ceph_mds_session *session = NULL;
1567         u32 seq = 0;
1568         int valid = 0;
1569
1570         spin_lock(&dentry->d_lock);
1571         di = ceph_dentry(dentry);
1572         if (di && __dentry_lease_is_valid(di)) {
1573                 valid = 1;
1574
1575                 if (di->lease_renew_after &&
1576                     time_after(jiffies, di->lease_renew_after)) {
1577                         /*
1578                          * We should renew. If we're in RCU walk mode
1579                          * though, we can't do that so just return
1580                          * -ECHILD.
1581                          */
1582                         if (flags & LOOKUP_RCU) {
1583                                 valid = -ECHILD;
1584                         } else {
1585                                 session = ceph_get_mds_session(di->lease_session);
1586                                 seq = di->lease_seq;
1587                                 di->lease_renew_after = 0;
1588                                 di->lease_renew_from = jiffies;
1589                         }
1590                 }
1591         }
1592         spin_unlock(&dentry->d_lock);
1593
1594         if (session) {
1595                 ceph_mdsc_lease_send_msg(session, dentry,
1596                                          CEPH_MDS_LEASE_RENEW, seq);
1597                 ceph_put_mds_session(session);
1598         }
1599         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1600         return valid;
1601 }
1602
1603 /*
1604  * Called under dentry->d_lock.
1605  */
1606 static int __dir_lease_try_check(const struct dentry *dentry)
1607 {
1608         struct ceph_dentry_info *di = ceph_dentry(dentry);
1609         struct inode *dir;
1610         struct ceph_inode_info *ci;
1611         int valid = 0;
1612
1613         if (!di->lease_shared_gen)
1614                 return 0;
1615         if (IS_ROOT(dentry))
1616                 return 0;
1617
1618         dir = d_inode(dentry->d_parent);
1619         ci = ceph_inode(dir);
1620
1621         if (spin_trylock(&ci->i_ceph_lock)) {
1622                 if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen &&
1623                     __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 0))
1624                         valid = 1;
1625                 spin_unlock(&ci->i_ceph_lock);
1626         } else {
1627                 valid = -EBUSY;
1628         }
1629
1630         if (!valid)
1631                 di->lease_shared_gen = 0;
1632         return valid;
1633 }
1634
1635 /*
1636  * Check if directory-wide content lease/cap is valid.
1637  */
1638 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry,
1639                               struct ceph_mds_client *mdsc)
1640 {
1641         struct ceph_inode_info *ci = ceph_inode(dir);
1642         int valid;
1643         int shared_gen;
1644
1645         spin_lock(&ci->i_ceph_lock);
1646         valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1647         if (valid) {
1648                 __ceph_touch_fmode(ci, mdsc, CEPH_FILE_MODE_RD);
1649                 shared_gen = atomic_read(&ci->i_shared_gen);
1650         }
1651         spin_unlock(&ci->i_ceph_lock);
1652         if (valid) {
1653                 struct ceph_dentry_info *di;
1654                 spin_lock(&dentry->d_lock);
1655                 di = ceph_dentry(dentry);
1656                 if (dir == d_inode(dentry->d_parent) &&
1657                     di && di->lease_shared_gen == shared_gen)
1658                         __ceph_dentry_dir_lease_touch(di);
1659                 else
1660                         valid = 0;
1661                 spin_unlock(&dentry->d_lock);
1662         }
1663         dout("dir_lease_is_valid dir %p v%u dentry %p = %d\n",
1664              dir, (unsigned)atomic_read(&ci->i_shared_gen), dentry, valid);
1665         return valid;
1666 }
1667
1668 /*
1669  * Check if cached dentry can be trusted.
1670  */
1671 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1672 {
1673         int valid = 0;
1674         struct dentry *parent;
1675         struct inode *dir, *inode;
1676         struct ceph_mds_client *mdsc;
1677
1678         if (flags & LOOKUP_RCU) {
1679                 parent = READ_ONCE(dentry->d_parent);
1680                 dir = d_inode_rcu(parent);
1681                 if (!dir)
1682                         return -ECHILD;
1683                 inode = d_inode_rcu(dentry);
1684         } else {
1685                 parent = dget_parent(dentry);
1686                 dir = d_inode(parent);
1687                 inode = d_inode(dentry);
1688         }
1689
1690         dout("d_revalidate %p '%pd' inode %p offset 0x%llx\n", dentry,
1691              dentry, inode, ceph_dentry(dentry)->offset);
1692
1693         mdsc = ceph_sb_to_client(dir->i_sb)->mdsc;
1694
1695         /* always trust cached snapped dentries, snapdir dentry */
1696         if (ceph_snap(dir) != CEPH_NOSNAP) {
1697                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1698                      dentry, inode);
1699                 valid = 1;
1700         } else if (inode && ceph_snap(inode) == CEPH_SNAPDIR) {
1701                 valid = 1;
1702         } else {
1703                 valid = dentry_lease_is_valid(dentry, flags);
1704                 if (valid == -ECHILD)
1705                         return valid;
1706                 if (valid || dir_lease_is_valid(dir, dentry, mdsc)) {
1707                         if (inode)
1708                                 valid = ceph_is_any_caps(inode);
1709                         else
1710                                 valid = 1;
1711                 }
1712         }
1713
1714         if (!valid) {
1715                 struct ceph_mds_request *req;
1716                 int op, err;
1717                 u32 mask;
1718
1719                 if (flags & LOOKUP_RCU)
1720                         return -ECHILD;
1721
1722                 percpu_counter_inc(&mdsc->metric.d_lease_mis);
1723
1724                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1725                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1726                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1727                 if (!IS_ERR(req)) {
1728                         req->r_dentry = dget(dentry);
1729                         req->r_num_caps = 2;
1730                         req->r_parent = dir;
1731
1732                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1733                         if (ceph_security_xattr_wanted(dir))
1734                                 mask |= CEPH_CAP_XATTR_SHARED;
1735                         req->r_args.getattr.mask = cpu_to_le32(mask);
1736
1737                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1738                         switch (err) {
1739                         case 0:
1740                                 if (d_really_is_positive(dentry) &&
1741                                     d_inode(dentry) == req->r_target_inode)
1742                                         valid = 1;
1743                                 break;
1744                         case -ENOENT:
1745                                 if (d_really_is_negative(dentry))
1746                                         valid = 1;
1747                                 fallthrough;
1748                         default:
1749                                 break;
1750                         }
1751                         ceph_mdsc_put_request(req);
1752                         dout("d_revalidate %p lookup result=%d\n",
1753                              dentry, err);
1754                 }
1755         } else {
1756                 percpu_counter_inc(&mdsc->metric.d_lease_hit);
1757         }
1758
1759         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1760         if (!valid)
1761                 ceph_dir_clear_complete(dir);
1762
1763         if (!(flags & LOOKUP_RCU))
1764                 dput(parent);
1765         return valid;
1766 }
1767
1768 /*
1769  * Delete unused dentry that doesn't have valid lease
1770  *
1771  * Called under dentry->d_lock.
1772  */
1773 static int ceph_d_delete(const struct dentry *dentry)
1774 {
1775         struct ceph_dentry_info *di;
1776
1777         /* won't release caps */
1778         if (d_really_is_negative(dentry))
1779                 return 0;
1780         if (ceph_snap(d_inode(dentry)) != CEPH_NOSNAP)
1781                 return 0;
1782         /* vaild lease? */
1783         di = ceph_dentry(dentry);
1784         if (di) {
1785                 if (__dentry_lease_is_valid(di))
1786                         return 0;
1787                 if (__dir_lease_try_check(dentry))
1788                         return 0;
1789         }
1790         return 1;
1791 }
1792
1793 /*
1794  * Release our ceph_dentry_info.
1795  */
1796 static void ceph_d_release(struct dentry *dentry)
1797 {
1798         struct ceph_dentry_info *di = ceph_dentry(dentry);
1799         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
1800
1801         dout("d_release %p\n", dentry);
1802
1803         atomic64_dec(&fsc->mdsc->metric.total_dentries);
1804
1805         spin_lock(&dentry->d_lock);
1806         __dentry_lease_unlist(di);
1807         dentry->d_fsdata = NULL;
1808         spin_unlock(&dentry->d_lock);
1809
1810         if (di->lease_session)
1811                 ceph_put_mds_session(di->lease_session);
1812         kmem_cache_free(ceph_dentry_cachep, di);
1813 }
1814
1815 /*
1816  * When the VFS prunes a dentry from the cache, we need to clear the
1817  * complete flag on the parent directory.
1818  *
1819  * Called under dentry->d_lock.
1820  */
1821 static void ceph_d_prune(struct dentry *dentry)
1822 {
1823         struct ceph_inode_info *dir_ci;
1824         struct ceph_dentry_info *di;
1825
1826         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1827
1828         /* do we have a valid parent? */
1829         if (IS_ROOT(dentry))
1830                 return;
1831
1832         /* we hold d_lock, so d_parent is stable */
1833         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1834         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1835                 return;
1836
1837         /* who calls d_delete() should also disable dcache readdir */
1838         if (d_really_is_negative(dentry))
1839                 return;
1840
1841         /* d_fsdata does not get cleared until d_release */
1842         if (!d_unhashed(dentry)) {
1843                 __ceph_dir_clear_complete(dir_ci);
1844                 return;
1845         }
1846
1847         /* Disable dcache readdir just in case that someone called d_drop()
1848          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1849          * properly (dcache readdir is still enabled) */
1850         di = ceph_dentry(dentry);
1851         if (di->offset > 0 &&
1852             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1853                 __ceph_dir_clear_ordered(dir_ci);
1854 }
1855
1856 /*
1857  * read() on a dir.  This weird interface hack only works if mounted
1858  * with '-o dirstat'.
1859  */
1860 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1861                              loff_t *ppos)
1862 {
1863         struct ceph_dir_file_info *dfi = file->private_data;
1864         struct inode *inode = file_inode(file);
1865         struct ceph_inode_info *ci = ceph_inode(inode);
1866         int left;
1867         const int bufsize = 1024;
1868
1869         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1870                 return -EISDIR;
1871
1872         if (!dfi->dir_info) {
1873                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1874                 if (!dfi->dir_info)
1875                         return -ENOMEM;
1876                 dfi->dir_info_len =
1877                         snprintf(dfi->dir_info, bufsize,
1878                                 "entries:   %20lld\n"
1879                                 " files:    %20lld\n"
1880                                 " subdirs:  %20lld\n"
1881                                 "rentries:  %20lld\n"
1882                                 " rfiles:   %20lld\n"
1883                                 " rsubdirs: %20lld\n"
1884                                 "rbytes:    %20lld\n"
1885                                 "rctime:    %10lld.%09ld\n",
1886                                 ci->i_files + ci->i_subdirs,
1887                                 ci->i_files,
1888                                 ci->i_subdirs,
1889                                 ci->i_rfiles + ci->i_rsubdirs,
1890                                 ci->i_rfiles,
1891                                 ci->i_rsubdirs,
1892                                 ci->i_rbytes,
1893                                 ci->i_rctime.tv_sec,
1894                                 ci->i_rctime.tv_nsec);
1895         }
1896
1897         if (*ppos >= dfi->dir_info_len)
1898                 return 0;
1899         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1900         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1901         if (left == size)
1902                 return -EFAULT;
1903         *ppos += (size - left);
1904         return size - left;
1905 }
1906
1907
1908
1909 /*
1910  * Return name hash for a given dentry.  This is dependent on
1911  * the parent directory's hash function.
1912  */
1913 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1914 {
1915         struct ceph_inode_info *dci = ceph_inode(dir);
1916         unsigned hash;
1917
1918         switch (dci->i_dir_layout.dl_dir_hash) {
1919         case 0: /* for backward compat */
1920         case CEPH_STR_HASH_LINUX:
1921                 return dn->d_name.hash;
1922
1923         default:
1924                 spin_lock(&dn->d_lock);
1925                 hash = ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1926                                      dn->d_name.name, dn->d_name.len);
1927                 spin_unlock(&dn->d_lock);
1928                 return hash;
1929         }
1930 }
1931
1932 const struct file_operations ceph_dir_fops = {
1933         .read = ceph_read_dir,
1934         .iterate = ceph_readdir,
1935         .llseek = ceph_dir_llseek,
1936         .open = ceph_open,
1937         .release = ceph_release,
1938         .unlocked_ioctl = ceph_ioctl,
1939         .compat_ioctl = compat_ptr_ioctl,
1940         .fsync = ceph_fsync,
1941         .lock = ceph_lock,
1942         .flock = ceph_flock,
1943 };
1944
1945 const struct file_operations ceph_snapdir_fops = {
1946         .iterate = ceph_readdir,
1947         .llseek = ceph_dir_llseek,
1948         .open = ceph_open,
1949         .release = ceph_release,
1950 };
1951
1952 const struct inode_operations ceph_dir_iops = {
1953         .lookup = ceph_lookup,
1954         .permission = ceph_permission,
1955         .getattr = ceph_getattr,
1956         .setattr = ceph_setattr,
1957         .listxattr = ceph_listxattr,
1958         .get_acl = ceph_get_acl,
1959         .set_acl = ceph_set_acl,
1960         .mknod = ceph_mknod,
1961         .symlink = ceph_symlink,
1962         .mkdir = ceph_mkdir,
1963         .link = ceph_link,
1964         .unlink = ceph_unlink,
1965         .rmdir = ceph_unlink,
1966         .rename = ceph_rename,
1967         .create = ceph_create,
1968         .atomic_open = ceph_atomic_open,
1969 };
1970
1971 const struct inode_operations ceph_snapdir_iops = {
1972         .lookup = ceph_lookup,
1973         .permission = ceph_permission,
1974         .getattr = ceph_getattr,
1975         .mkdir = ceph_mkdir,
1976         .rmdir = ceph_unlink,
1977         .rename = ceph_rename,
1978 };
1979
1980 const struct dentry_operations ceph_dentry_ops = {
1981         .d_revalidate = ceph_d_revalidate,
1982         .d_delete = ceph_d_delete,
1983         .d_release = ceph_d_release,
1984         .d_prune = ceph_d_prune,
1985         .d_init = ceph_d_init,
1986 };