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 /*
  33  * Initialize ceph dentry state.
  34  */
  35 static int ceph_d_init(struct dentry *dentry)
  36 {
  37         struct ceph_dentry_info *di;
  38
  39         di = kmem_cache_zalloc(ceph_dentry_cachep, GFP_KERNEL);
  40         if (!di)
  41                 return -ENOMEM;          /* oh well */
  42
  43         di->dentry = dentry;
  44         di->lease_session = NULL;
  45         di->time = jiffies;
  46         dentry->d_fsdata = di;
  47         ceph_dentry_lru_add(dentry);
  48         return 0;
  49 }
  50
  51 /*
  52  * for f_pos for readdir:
  53  * - hash order:
  54  *      (0xff << 52) | ((24 bits hash) << 28) |
  55  *      (the nth entry has hash collision);
  56  * - frag+name order;
  57  *      ((frag value) << 28) | (the nth entry in frag);
  58  */
  59 #define OFFSET_BITS     28
  60 #define OFFSET_MASK     ((1 << OFFSET_BITS) - 1)
  61 #define HASH_ORDER      (0xffull << (OFFSET_BITS + 24))
  62 loff_t ceph_make_fpos(unsigned high, unsigned off, bool hash_order)
  63 {
  64         loff_t fpos = ((loff_t)high << 28) | (loff_t)off;
  65         if (hash_order)
  66                 fpos |= HASH_ORDER;
  67         return fpos;
  68 }
  69
  70 static bool is_hash_order(loff_t p)
  71 {
  72         return (p & HASH_ORDER) == HASH_ORDER;
  73 }
  74
  75 static unsigned fpos_frag(loff_t p)
  76 {
  77         return p >> OFFSET_BITS;
  78 }
  79
  80 static unsigned fpos_hash(loff_t p)
  81 {
  82         return ceph_frag_value(fpos_frag(p));
  83 }
  84
  85 static unsigned fpos_off(loff_t p)
  86 {
  87         return p & OFFSET_MASK;
  88 }
  89
  90 static int fpos_cmp(loff_t l, loff_t r)
  91 {
  92         int v = ceph_frag_compare(fpos_frag(l), fpos_frag(r));
  93         if (v)
  94                 return v;
  95         return (int)(fpos_off(l) - fpos_off(r));
  96 }
  97
  98 /*
  99  * make note of the last dentry we read, so we can
 100  * continue at the same lexicographical point,
 101  * regardless of what dir changes take place on the
 102  * server.
 103  */
 104 static int note_last_dentry(struct ceph_dir_file_info *dfi, const char *name,
 105                             int len, unsigned next_offset)
 106 {
 107         char *buf = kmalloc(len+1, GFP_KERNEL);
 108         if (!buf)
 109                 return -ENOMEM;
 110         kfree(dfi->last_name);
 111         dfi->last_name = buf;
 112         memcpy(dfi->last_name, name, len);
 113         dfi->last_name[len] = 0;
 114         dfi->next_offset = next_offset;
 115         dout("note_last_dentry '%s'\n", dfi->last_name);
 116         return 0;
 117 }
 118
 119
 120 static struct dentry *
 121 __dcache_find_get_entry(struct dentry *parent, u64 idx,
 122                         struct ceph_readdir_cache_control *cache_ctl)
 123 {
 124         struct inode *dir = d_inode(parent);
 125         struct dentry *dentry;
 126         unsigned idx_mask = (PAGE_SIZE / sizeof(struct dentry *)) - 1;
 127         loff_t ptr_pos = idx * sizeof(struct dentry *);
 128         pgoff_t ptr_pgoff = ptr_pos >> PAGE_SHIFT;
 129
 130         if (ptr_pos >= i_size_read(dir))
 131                 return NULL;
 132
 133         if (!cache_ctl->page || ptr_pgoff != page_index(cache_ctl->page)) {
 134                 ceph_readdir_cache_release(cache_ctl);
 135                 cache_ctl->page = find_lock_page(&dir->i_data, ptr_pgoff);
 136                 if (!cache_ctl->page) {
 137                         dout(" page %lu not found\n", ptr_pgoff);
 138                         return ERR_PTR(-EAGAIN);
 139                 }
 140                 /* reading/filling the cache are serialized by
 141                    i_mutex, no need to use page lock */
 142                 unlock_page(cache_ctl->page);
 143                 cache_ctl->dentries = kmap(cache_ctl->page);
 144         }
 145
 146         cache_ctl->index = idx & idx_mask;
 147
 148         rcu_read_lock();
 149         spin_lock(&parent->d_lock);
 150         /* check i_size again here, because empty directory can be
 151          * marked as complete while not holding the i_mutex. */
 152         if (ceph_dir_is_complete_ordered(dir) && ptr_pos < i_size_read(dir))
 153                 dentry = cache_ctl->dentries[cache_ctl->index];
 154         else
 155                 dentry = NULL;
 156         spin_unlock(&parent->d_lock);
 157         if (dentry && !lockref_get_not_dead(&dentry->d_lockref))
 158                 dentry = NULL;
 159         rcu_read_unlock();
 160         return dentry ? : ERR_PTR(-EAGAIN);
 161 }
 162
 163 /*
 164  * When possible, we try to satisfy a readdir by peeking at the
 165  * dcache.  We make this work by carefully ordering dentries on
 166  * d_child when we initially get results back from the MDS, and
 167  * falling back to a "normal" sync readdir if any dentries in the dir
 168  * are dropped.
 169  *
 170  * Complete dir indicates that we have all dentries in the dir.  It is
 171  * defined IFF we hold CEPH_CAP_FILE_SHARED (which will be revoked by
 172  * the MDS if/when the directory is modified).
 173  */
 174 static int __dcache_readdir(struct file *file,  struct dir_context *ctx,
 175                             int shared_gen)
 176 {
 177         struct ceph_dir_file_info *dfi = file->private_data;
 178         struct dentry *parent = file->f_path.dentry;
 179         struct inode *dir = d_inode(parent);
 180         struct dentry *dentry, *last = NULL;
 181         struct ceph_dentry_info *di;
 182         struct ceph_readdir_cache_control cache_ctl = {};
 183         u64 idx = 0;
 184         int err = 0;
 185
 186         dout("__dcache_readdir %p v%u at %llx\n", dir, (unsigned)shared_gen, ctx->pos);
 187
 188         /* search start position */
 189         if (ctx->pos > 2) {
 190                 u64 count = div_u64(i_size_read(dir), sizeof(struct dentry *));
 191                 while (count > 0) {
 192                         u64 step = count >> 1;
 193                         dentry = __dcache_find_get_entry(parent, idx + step,
 194                                                          &cache_ctl);
 195                         if (!dentry) {
 196                                 /* use linar search */
 197                                 idx = 0;
 198                                 break;
 199                         }
 200                         if (IS_ERR(dentry)) {
 201                                 err = PTR_ERR(dentry);
 202                                 goto out;
 203                         }
 204                         di = ceph_dentry(dentry);
 205                         spin_lock(&dentry->d_lock);
 206                         if (fpos_cmp(di->offset, ctx->pos) < 0) {
 207                                 idx += step + 1;
 208                                 count -= step + 1;
 209                         } else {
 210                                 count = step;
 211                         }
 212                         spin_unlock(&dentry->d_lock);
 213                         dput(dentry);
 214                 }
 215
 216                 dout("__dcache_readdir %p cache idx %llu\n", dir, idx);
 217         }
 218
 219
 220         for (;;) {
 221                 bool emit_dentry = false;
 222                 dentry = __dcache_find_get_entry(parent, idx++, &cache_ctl);
 223                 if (!dentry) {
 224                         dfi->file_info.flags |= CEPH_F_ATEND;
 225                         err = 0;
 226                         break;
 227                 }
 228                 if (IS_ERR(dentry)) {
 229                         err = PTR_ERR(dentry);
 230                         goto out;
 231                 }
 232
 233                 spin_lock(&dentry->d_lock);
 234                 di = ceph_dentry(dentry);
 235                 if (d_unhashed(dentry) ||
 236                     d_really_is_negative(dentry) ||
 237                     di->lease_shared_gen != shared_gen) {
 238                         spin_unlock(&dentry->d_lock);
 239                         dput(dentry);
 240                         err = -EAGAIN;
 241                         goto out;
 242                 }
 243                 if (fpos_cmp(ctx->pos, di->offset) <= 0) {
 244                         emit_dentry = true;
 245                 }
 246                 spin_unlock(&dentry->d_lock);
 247
 248                 if (emit_dentry) {
 249                         dout(" %llx dentry %p %pd %p\n", di->offset,
 250                              dentry, dentry, d_inode(dentry));
 251                         ctx->pos = di->offset;
 252                         if (!dir_emit(ctx, dentry->d_name.name,
 253                                       dentry->d_name.len,
 254                                       ceph_translate_ino(dentry->d_sb,
 255                                                          d_inode(dentry)->i_ino),
 256                                       d_inode(dentry)->i_mode >> 12)) {
 257                                 dput(dentry);
 258                                 err = 0;
 259                                 break;
 260                         }
 261                         ctx->pos++;
 262
 263                         if (last)
 264                                 dput(last);
 265                         last = dentry;
 266                 } else {
 267                         dput(dentry);
 268                 }
 269         }
 270 out:
 271         ceph_readdir_cache_release(&cache_ctl);
 272         if (last) {
 273                 int ret;
 274                 di = ceph_dentry(last);
 275                 ret = note_last_dentry(dfi, last->d_name.name, last->d_name.len,
 276                                        fpos_off(di->offset) + 1);
 277                 if (ret < 0)
 278                         err = ret;
 279                 dput(last);
 280                 /* last_name no longer match cache index */
 281                 if (dfi->readdir_cache_idx >= 0) {
 282                         dfi->readdir_cache_idx = -1;
 283                         dfi->dir_release_count = 0;
 284                 }
 285         }
 286         return err;
 287 }
 288
 289 static bool need_send_readdir(struct ceph_dir_file_info *dfi, loff_t pos)
 290 {
 291         if (!dfi->last_readdir)
 292                 return true;
 293         if (is_hash_order(pos))
 294                 return !ceph_frag_contains_value(dfi->frag, fpos_hash(pos));
 295         else
 296                 return dfi->frag != fpos_frag(pos);
 297 }
 298
 299 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 300 {
 301         struct ceph_dir_file_info *dfi = file->private_data;
 302         struct inode *inode = file_inode(file);
 303         struct ceph_inode_info *ci = ceph_inode(inode);
 304         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 305         struct ceph_mds_client *mdsc = fsc->mdsc;
 306         int i;
 307         int err;
 308         unsigned frag = -1;
 309         struct ceph_mds_reply_info_parsed *rinfo;
 310
 311         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 312         if (dfi->file_info.flags & CEPH_F_ATEND)
 313                 return 0;
 314
 315         /* always start with . and .. */
 316         if (ctx->pos == 0) {
 317                 dout("readdir off 0 -> '.'\n");
 318                 if (!dir_emit(ctx, ".", 1,
 319                             ceph_translate_ino(inode->i_sb, inode->i_ino),
 320                             inode->i_mode >> 12))
 321                         return 0;
 322                 ctx->pos = 1;
 323         }
 324         if (ctx->pos == 1) {
 325                 ino_t ino = parent_ino(file->f_path.dentry);
 326                 dout("readdir off 1 -> '..'\n");
 327                 if (!dir_emit(ctx, "..", 2,
 328                             ceph_translate_ino(inode->i_sb, ino),
 329                             inode->i_mode >> 12))
 330                         return 0;
 331                 ctx->pos = 2;
 332         }
 333
 334         /* can we use the dcache? */
 335         spin_lock(&ci->i_ceph_lock);
 336         if (ceph_test_mount_opt(fsc, DCACHE) &&
 337             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 338             ceph_snap(inode) != CEPH_SNAPDIR &&
 339             __ceph_dir_is_complete_ordered(ci) &&
 340             __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
 341                 int shared_gen = atomic_read(&ci->i_shared_gen);
 342                 spin_unlock(&ci->i_ceph_lock);
 343                 err = __dcache_readdir(file, ctx, shared_gen);
 344                 if (err != -EAGAIN)
 345                         return err;
 346         } else {
 347                 spin_unlock(&ci->i_ceph_lock);
 348         }
 349
 350         /* proceed with a normal readdir */
 351 more:
 352         /* do we have the correct frag content buffered? */
 353         if (need_send_readdir(dfi, ctx->pos)) {
 354                 struct ceph_mds_request *req;
 355                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 356                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 357
 358                 /* discard old result, if any */
 359                 if (dfi->last_readdir) {
 360                         ceph_mdsc_put_request(dfi->last_readdir);
 361                         dfi->last_readdir = NULL;
 362                 }
 363
 364                 if (is_hash_order(ctx->pos)) {
 365                         /* fragtree isn't always accurate. choose frag
 366                          * based on previous reply when possible. */
 367                         if (frag == (unsigned)-1)
 368                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 369                                                         NULL, NULL);
 370                 } else {
 371                         frag = fpos_frag(ctx->pos);
 372                 }
 373
 374                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 375                      ceph_vinop(inode), frag, dfi->last_name);
 376                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 377                 if (IS_ERR(req))
 378                         return PTR_ERR(req);
 379                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 380                 if (err) {
 381                         ceph_mdsc_put_request(req);
 382                         return err;
 383                 }
 384                 /* hints to request -> mds selection code */
 385                 req->r_direct_mode = USE_AUTH_MDS;
 386                 if (op == CEPH_MDS_OP_READDIR) {
 387                         req->r_direct_hash = ceph_frag_value(frag);
 388                         __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 389                         req->r_inode_drop = CEPH_CAP_FILE_EXCL;
 390                 }
 391                 if (dfi->last_name) {
 392                         req->r_path2 = kstrdup(dfi->last_name, GFP_KERNEL);
 393                         if (!req->r_path2) {
 394                                 ceph_mdsc_put_request(req);
 395                                 return -ENOMEM;
 396                         }
 397                 } else if (is_hash_order(ctx->pos)) {
 398                         req->r_args.readdir.offset_hash =
 399                                 cpu_to_le32(fpos_hash(ctx->pos));
 400                 }
 401
 402                 req->r_dir_release_cnt = dfi->dir_release_count;
 403                 req->r_dir_ordered_cnt = dfi->dir_ordered_count;
 404                 req->r_readdir_cache_idx = dfi->readdir_cache_idx;
 405                 req->r_readdir_offset = dfi->next_offset;
 406                 req->r_args.readdir.frag = cpu_to_le32(frag);
 407                 req->r_args.readdir.flags =
 408                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 409
 410                 req->r_inode = inode;
 411                 ihold(inode);
 412                 req->r_dentry = dget(file->f_path.dentry);
 413                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 414                 if (err < 0) {
 415                         ceph_mdsc_put_request(req);
 416                         return err;
 417                 }
 418                 dout("readdir got and parsed readdir result=%d on "
 419                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 420                      err, frag,
 421                      (int)req->r_reply_info.dir_end,
 422                      (int)req->r_reply_info.dir_complete,
 423                      (int)req->r_reply_info.hash_order);
 424
 425                 rinfo = &req->r_reply_info;
 426                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 427                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 428                         if (!rinfo->hash_order) {
 429                                 dfi->next_offset = req->r_readdir_offset;
 430                                 /* adjust ctx->pos to beginning of frag */
 431                                 ctx->pos = ceph_make_fpos(frag,
 432                                                           dfi->next_offset,
 433                                                           false);
 434                         }
 435                 }
 436
 437                 dfi->frag = frag;
 438                 dfi->last_readdir = req;
 439
 440                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 441                         dfi->readdir_cache_idx = req->r_readdir_cache_idx;
 442                         if (dfi->readdir_cache_idx < 0) {
 443                                 /* preclude from marking dir ordered */
 444                                 dfi->dir_ordered_count = 0;
 445                         } else if (ceph_frag_is_leftmost(frag) &&
 446                                    dfi->next_offset == 2) {
 447                                 /* note dir version at start of readdir so
 448                                  * we can tell if any dentries get dropped */
 449                                 dfi->dir_release_count = req->r_dir_release_cnt;
 450                                 dfi->dir_ordered_count = req->r_dir_ordered_cnt;
 451                         }
 452                 } else {
 453                         dout("readdir !did_prepopulate\n");
 454                         /* disable readdir cache */
 455                         dfi->readdir_cache_idx = -1;
 456                         /* preclude from marking dir complete */
 457                         dfi->dir_release_count = 0;
 458                 }
 459
 460                 /* note next offset and last dentry name */
 461                 if (rinfo->dir_nr > 0) {
 462                         struct ceph_mds_reply_dir_entry *rde =
 463                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 464                         unsigned next_offset = req->r_reply_info.dir_end ?
 465                                         2 : (fpos_off(rde->offset) + 1);
 466                         err = note_last_dentry(dfi, rde->name, rde->name_len,
 467                                                next_offset);
 468                         if (err)
 469                                 return err;
 470                 } else if (req->r_reply_info.dir_end) {
 471                         dfi->next_offset = 2;
 472                         /* keep last name */
 473                 }
 474         }
 475
 476         rinfo = &dfi->last_readdir->r_reply_info;
 477         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 478              dfi->frag, rinfo->dir_nr, ctx->pos,
 479              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 480
 481         i = 0;
 482         /* search start position */
 483         if (rinfo->dir_nr > 0) {
 484                 int step, nr = rinfo->dir_nr;
 485                 while (nr > 0) {
 486                         step = nr >> 1;
 487                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 488                                 i +=  step + 1;
 489                                 nr -= step + 1;
 490                         } else {
 491                                 nr = step;
 492                         }
 493                 }
 494         }
 495         for (; i < rinfo->dir_nr; i++) {
 496                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 497                 struct ceph_vino vino;
 498                 ino_t ino;
 499                 u32 ftype;
 500
 501                 BUG_ON(rde->offset < ctx->pos);
 502
 503                 ctx->pos = rde->offset;
 504                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 505                      i, rinfo->dir_nr, ctx->pos,
 506                      rde->name_len, rde->name, &rde->inode.in);
 507
 508                 BUG_ON(!rde->inode.in);
 509                 ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
 510                 vino.ino = le64_to_cpu(rde->inode.in->ino);
 511                 vino.snap = le64_to_cpu(rde->inode.in->snapid);
 512                 ino = ceph_vino_to_ino(vino);
 513
 514                 if (!dir_emit(ctx, rde->name, rde->name_len,
 515                               ceph_translate_ino(inode->i_sb, ino), ftype)) {
 516                         dout("filldir stopping us...\n");
 517                         return 0;
 518                 }
 519                 ctx->pos++;
 520         }
 521
 522         ceph_mdsc_put_request(dfi->last_readdir);
 523         dfi->last_readdir = NULL;
 524
 525         if (dfi->next_offset > 2) {
 526                 frag = dfi->frag;
 527                 goto more;
 528         }
 529
 530         /* more frags? */
 531         if (!ceph_frag_is_rightmost(dfi->frag)) {
 532                 frag = ceph_frag_next(dfi->frag);
 533                 if (is_hash_order(ctx->pos)) {
 534                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 535                                                         dfi->next_offset, true);
 536                         if (new_pos > ctx->pos)
 537                                 ctx->pos = new_pos;
 538                         /* keep last_name */
 539                 } else {
 540                         ctx->pos = ceph_make_fpos(frag, dfi->next_offset,
 541                                                         false);
 542                         kfree(dfi->last_name);
 543                         dfi->last_name = NULL;
 544                 }
 545                 dout("readdir next frag is %x\n", frag);
 546                 goto more;
 547         }
 548         dfi->file_info.flags |= CEPH_F_ATEND;
 549
 550         /*
 551          * if dir_release_count still matches the dir, no dentries
 552          * were released during the whole readdir, and we should have
 553          * the complete dir contents in our cache.
 554          */
 555         if (atomic64_read(&ci->i_release_count) ==
 556                                         dfi->dir_release_count) {
 557                 spin_lock(&ci->i_ceph_lock);
 558                 if (dfi->dir_ordered_count ==
 559                                 atomic64_read(&ci->i_ordered_count)) {
 560                         dout(" marking %p complete and ordered\n", inode);
 561                         /* use i_size to track number of entries in
 562                          * readdir cache */
 563                         BUG_ON(dfi->readdir_cache_idx < 0);
 564                         i_size_write(inode, dfi->readdir_cache_idx *
 565                                      sizeof(struct dentry*));
 566                 } else {
 567                         dout(" marking %p complete\n", inode);
 568                 }
 569                 __ceph_dir_set_complete(ci, dfi->dir_release_count,
 570                                         dfi->dir_ordered_count);
 571                 spin_unlock(&ci->i_ceph_lock);
 572         }
 573
 574         dout("readdir %p file %p done.\n", inode, file);
 575         return 0;
 576 }
 577
 578 static void reset_readdir(struct ceph_dir_file_info *dfi)
 579 {
 580         if (dfi->last_readdir) {
 581                 ceph_mdsc_put_request(dfi->last_readdir);
 582                 dfi->last_readdir = NULL;
 583         }
 584         kfree(dfi->last_name);
 585         dfi->last_name = NULL;
 586         dfi->dir_release_count = 0;
 587         dfi->readdir_cache_idx = -1;
 588         dfi->next_offset = 2;  /* compensate for . and .. */
 589         dfi->file_info.flags &= ~CEPH_F_ATEND;
 590 }
 591
 592 /*
 593  * discard buffered readdir content on seekdir(0), or seek to new frag,
 594  * or seek prior to current chunk
 595  */
 596 static bool need_reset_readdir(struct ceph_dir_file_info *dfi, loff_t new_pos)
 597 {
 598         struct ceph_mds_reply_info_parsed *rinfo;
 599         loff_t chunk_offset;
 600         if (new_pos == 0)
 601                 return true;
 602         if (is_hash_order(new_pos)) {
 603                 /* no need to reset last_name for a forward seek when
 604                  * dentries are sotred in hash order */
 605         } else if (dfi->frag != fpos_frag(new_pos)) {
 606                 return true;
 607         }
 608         rinfo = dfi->last_readdir ? &dfi->last_readdir->r_reply_info : NULL;
 609         if (!rinfo || !rinfo->dir_nr)
 610                 return true;
 611         chunk_offset = rinfo->dir_entries[0].offset;
 612         return new_pos < chunk_offset ||
 613                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 614 }
 615
 616 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 617 {
 618         struct ceph_dir_file_info *dfi = file->private_data;
 619         struct inode *inode = file->f_mapping->host;
 620         loff_t retval;
 621
 622         inode_lock(inode);
 623         retval = -EINVAL;
 624         switch (whence) {
 625         case SEEK_CUR:
 626                 offset += file->f_pos;
 627         case SEEK_SET:
 628                 break;
 629         case SEEK_END:
 630                 retval = -EOPNOTSUPP;
 631         default:
 632                 goto out;
 633         }
 634
 635         if (offset >= 0) {
 636                 if (need_reset_readdir(dfi, offset)) {
 637                         dout("dir_llseek dropping %p content\n", file);
 638                         reset_readdir(dfi);
 639                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 640                         /* for hash offset, we don't know if a forward seek
 641                          * is within same frag */
 642                         dfi->dir_release_count = 0;
 643                         dfi->readdir_cache_idx = -1;
 644                 }
 645
 646                 if (offset != file->f_pos) {
 647                         file->f_pos = offset;
 648                         file->f_version = 0;
 649                         dfi->file_info.flags &= ~CEPH_F_ATEND;
 650                 }
 651                 retval = offset;
 652         }
 653 out:
 654         inode_unlock(inode);
 655         return retval;
 656 }
 657
 658 /*
 659  * Handle lookups for the hidden .snap directory.
 660  */
 661 int ceph_handle_snapdir(struct ceph_mds_request *req,
 662                         struct dentry *dentry, int err)
 663 {
 664         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 665         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 666
 667         /* .snap dir? */
 668         if (err == -ENOENT &&
 669             ceph_snap(parent) == CEPH_NOSNAP &&
 670             strcmp(dentry->d_name.name,
 671                    fsc->mount_options->snapdir_name) == 0) {
 672                 struct inode *inode = ceph_get_snapdir(parent);
 673                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 674                      dentry, dentry, inode);
 675                 BUG_ON(!d_unhashed(dentry));
 676                 d_add(dentry, inode);
 677                 err = 0;
 678         }
 679         return err;
 680 }
 681
 682 /*
 683  * Figure out final result of a lookup/open request.
 684  *
 685  * Mainly, make sure we return the final req->r_dentry (if it already
 686  * existed) in place of the original VFS-provided dentry when they
 687  * differ.
 688  *
 689  * Gracefully handle the case where the MDS replies with -ENOENT and
 690  * no trace (which it may do, at its discretion, e.g., if it doesn't
 691  * care to issue a lease on the negative dentry).
 692  */
 693 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 694                                   struct dentry *dentry, int err)
 695 {
 696         if (err == -ENOENT) {
 697                 /* no trace? */
 698                 err = 0;
 699                 if (!req->r_reply_info.head->is_dentry) {
 700                         dout("ENOENT and no trace, dentry %p inode %p\n",
 701                              dentry, d_inode(dentry));
 702                         if (d_really_is_positive(dentry)) {
 703                                 d_drop(dentry);
 704                                 err = -ENOENT;
 705                         } else {
 706                                 d_add(dentry, NULL);
 707                         }
 708                 }
 709         }
 710         if (err)
 711                 dentry = ERR_PTR(err);
 712         else if (dentry != req->r_dentry)
 713                 dentry = dget(req->r_dentry);   /* we got spliced */
 714         else
 715                 dentry = NULL;
 716         return dentry;
 717 }
 718
 719 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 720 {
 721         return ceph_ino(inode) == CEPH_INO_ROOT &&
 722                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 723 }
 724
 725 /*
 726  * Look up a single dir entry.  If there is a lookup intent, inform
 727  * the MDS so that it gets our 'caps wanted' value in a single op.
 728  */
 729 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 730                                   unsigned int flags)
 731 {
 732         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 733         struct ceph_mds_client *mdsc = fsc->mdsc;
 734         struct ceph_mds_request *req;
 735         int op;
 736         int mask;
 737         int err;
 738
 739         dout("lookup %p dentry %p '%pd'\n",
 740              dir, dentry, dentry);
 741
 742         if (dentry->d_name.len > NAME_MAX)
 743                 return ERR_PTR(-ENAMETOOLONG);
 744
 745         /* can we conclude ENOENT locally? */
 746         if (d_really_is_negative(dentry)) {
 747                 struct ceph_inode_info *ci = ceph_inode(dir);
 748                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 749
 750                 spin_lock(&ci->i_ceph_lock);
 751                 dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
 752                 if (strncmp(dentry->d_name.name,
 753                             fsc->mount_options->snapdir_name,
 754                             dentry->d_name.len) &&
 755                     !is_root_ceph_dentry(dir, dentry) &&
 756                     ceph_test_mount_opt(fsc, DCACHE) &&
 757                     __ceph_dir_is_complete(ci) &&
 758                     (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
 759                         spin_unlock(&ci->i_ceph_lock);
 760                         dout(" dir %p complete, -ENOENT\n", dir);
 761                         d_add(dentry, NULL);
 762                         di->lease_shared_gen = atomic_read(&ci->i_shared_gen);
 763                         return NULL;
 764                 }
 765                 spin_unlock(&ci->i_ceph_lock);
 766         }
 767
 768         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 769                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 770         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 771         if (IS_ERR(req))
 772                 return ERR_CAST(req);
 773         req->r_dentry = dget(dentry);
 774         req->r_num_caps = 2;
 775
 776         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 777         if (ceph_security_xattr_wanted(dir))
 778                 mask |= CEPH_CAP_XATTR_SHARED;
 779         req->r_args.getattr.mask = cpu_to_le32(mask);
 780
 781         req->r_parent = dir;
 782         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 783         err = ceph_mdsc_do_request(mdsc, NULL, req);
 784         err = ceph_handle_snapdir(req, dentry, err);
 785         dentry = ceph_finish_lookup(req, dentry, err);
 786         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 787         dout("lookup result=%p\n", dentry);
 788         return dentry;
 789 }
 790
 791 /*
 792  * If we do a create but get no trace back from the MDS, follow up with
 793  * a lookup (the VFS expects us to link up the provided dentry).
 794  */
 795 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 796 {
 797         struct dentry *result = ceph_lookup(dir, dentry, 0);
 798
 799         if (result && !IS_ERR(result)) {
 800                 /*
 801                  * We created the item, then did a lookup, and found
 802                  * it was already linked to another inode we already
 803                  * had in our cache (and thus got spliced). To not
 804                  * confuse VFS (especially when inode is a directory),
 805                  * we don't link our dentry to that inode, return an
 806                  * error instead.
 807                  *
 808                  * This event should be rare and it happens only when
 809                  * we talk to old MDS. Recent MDS does not send traceless
 810                  * reply for request that creates new inode.
 811                  */
 812                 d_drop(result);
 813                 return -ESTALE;
 814         }
 815         return PTR_ERR(result);
 816 }
 817
 818 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 819                       umode_t mode, dev_t rdev)
 820 {
 821         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 822         struct ceph_mds_client *mdsc = fsc->mdsc;
 823         struct ceph_mds_request *req;
 824         struct ceph_acls_info acls = {};
 825         int err;
 826
 827         if (ceph_snap(dir) != CEPH_NOSNAP)
 828                 return -EROFS;
 829
 830         if (ceph_quota_is_max_files_exceeded(dir)) {
 831                 err = -EDQUOT;
 832                 goto out;
 833         }
 834
 835         err = ceph_pre_init_acls(dir, &mode, &acls);
 836         if (err < 0)
 837                 goto out;
 838
 839         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 840              dir, dentry, mode, rdev);
 841         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 842         if (IS_ERR(req)) {
 843                 err = PTR_ERR(req);
 844                 goto out;
 845         }
 846         req->r_dentry = dget(dentry);
 847         req->r_num_caps = 2;
 848         req->r_parent = dir;
 849         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 850         req->r_args.mknod.mode = cpu_to_le32(mode);
 851         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 852         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 853         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 854         if (acls.pagelist) {
 855                 req->r_pagelist = acls.pagelist;
 856                 acls.pagelist = NULL;
 857         }
 858         err = ceph_mdsc_do_request(mdsc, dir, req);
 859         if (!err && !req->r_reply_info.head->is_dentry)
 860                 err = ceph_handle_notrace_create(dir, dentry);
 861         ceph_mdsc_put_request(req);
 862 out:
 863         if (!err)
 864                 ceph_init_inode_acls(d_inode(dentry), &acls);
 865         else
 866                 d_drop(dentry);
 867         ceph_release_acls_info(&acls);
 868         return err;
 869 }
 870
 871 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 872                        bool excl)
 873 {
 874         return ceph_mknod(dir, dentry, mode, 0);
 875 }
 876
 877 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 878                             const char *dest)
 879 {
 880         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 881         struct ceph_mds_client *mdsc = fsc->mdsc;
 882         struct ceph_mds_request *req;
 883         int err;
 884
 885         if (ceph_snap(dir) != CEPH_NOSNAP)
 886                 return -EROFS;
 887
 888         if (ceph_quota_is_max_files_exceeded(dir)) {
 889                 err = -EDQUOT;
 890                 goto out;
 891         }
 892
 893         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 894         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 895         if (IS_ERR(req)) {
 896                 err = PTR_ERR(req);
 897                 goto out;
 898         }
 899         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 900         if (!req->r_path2) {
 901                 err = -ENOMEM;
 902                 ceph_mdsc_put_request(req);
 903                 goto out;
 904         }
 905         req->r_parent = dir;
 906         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 907         req->r_dentry = dget(dentry);
 908         req->r_num_caps = 2;
 909         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 910         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 911         err = ceph_mdsc_do_request(mdsc, dir, req);
 912         if (!err && !req->r_reply_info.head->is_dentry)
 913                 err = ceph_handle_notrace_create(dir, dentry);
 914         ceph_mdsc_put_request(req);
 915 out:
 916         if (err)
 917                 d_drop(dentry);
 918         return err;
 919 }
 920
 921 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 922 {
 923         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 924         struct ceph_mds_client *mdsc = fsc->mdsc;
 925         struct ceph_mds_request *req;
 926         struct ceph_acls_info acls = {};
 927         int err = -EROFS;
 928         int op;
 929
 930         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 931                 /* mkdir .snap/foo is a MKSNAP */
 932                 op = CEPH_MDS_OP_MKSNAP;
 933                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 934                      dentry, dentry);
 935         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 936                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 937                 op = CEPH_MDS_OP_MKDIR;
 938         } else {
 939                 goto out;
 940         }
 941
 942         if (op == CEPH_MDS_OP_MKDIR &&
 943             ceph_quota_is_max_files_exceeded(dir)) {
 944                 err = -EDQUOT;
 945                 goto out;
 946         }
 947
 948         mode |= S_IFDIR;
 949         err = ceph_pre_init_acls(dir, &mode, &acls);
 950         if (err < 0)
 951                 goto out;
 952
 953         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 954         if (IS_ERR(req)) {
 955                 err = PTR_ERR(req);
 956                 goto out;
 957         }
 958
 959         req->r_dentry = dget(dentry);
 960         req->r_num_caps = 2;
 961         req->r_parent = dir;
 962         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 963         req->r_args.mkdir.mode = cpu_to_le32(mode);
 964         req->r_dentry_drop = CEPH_CAP_FILE_SHARED | CEPH_CAP_AUTH_EXCL;
 965         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 966         if (acls.pagelist) {
 967                 req->r_pagelist = acls.pagelist;
 968                 acls.pagelist = NULL;
 969         }
 970         err = ceph_mdsc_do_request(mdsc, dir, req);
 971         if (!err &&
 972             !req->r_reply_info.head->is_target &&
 973             !req->r_reply_info.head->is_dentry)
 974                 err = ceph_handle_notrace_create(dir, dentry);
 975         ceph_mdsc_put_request(req);
 976 out:
 977         if (!err)
 978                 ceph_init_inode_acls(d_inode(dentry), &acls);
 979         else
 980                 d_drop(dentry);
 981         ceph_release_acls_info(&acls);
 982         return err;
 983 }
 984
 985 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
 986                      struct dentry *dentry)
 987 {
 988         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 989         struct ceph_mds_client *mdsc = fsc->mdsc;
 990         struct ceph_mds_request *req;
 991         int err;
 992
 993         if (ceph_snap(dir) != CEPH_NOSNAP)
 994                 return -EROFS;
 995
 996         dout("link in dir %p old_dentry %p dentry %p\n", dir,
 997              old_dentry, dentry);
 998         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
 999         if (IS_ERR(req)) {
1000                 d_drop(dentry);
1001                 return PTR_ERR(req);
1002         }
1003         req->r_dentry = dget(dentry);
1004         req->r_num_caps = 2;
1005         req->r_old_dentry = dget(old_dentry);
1006         req->r_parent = dir;
1007         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1008         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1009         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1010         /* release LINK_SHARED on source inode (mds will lock it) */
1011         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1012         err = ceph_mdsc_do_request(mdsc, dir, req);
1013         if (err) {
1014                 d_drop(dentry);
1015         } else if (!req->r_reply_info.head->is_dentry) {
1016                 ihold(d_inode(old_dentry));
1017                 d_instantiate(dentry, d_inode(old_dentry));
1018         }
1019         ceph_mdsc_put_request(req);
1020         return err;
1021 }
1022
1023 /*
1024  * rmdir and unlink are differ only by the metadata op code
1025  */
1026 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1027 {
1028         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1029         struct ceph_mds_client *mdsc = fsc->mdsc;
1030         struct inode *inode = d_inode(dentry);
1031         struct ceph_mds_request *req;
1032         int err = -EROFS;
1033         int op;
1034
1035         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1036                 /* rmdir .snap/foo is RMSNAP */
1037                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1038                 op = CEPH_MDS_OP_RMSNAP;
1039         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1040                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1041                      dir, dentry, inode);
1042                 op = d_is_dir(dentry) ?
1043                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1044         } else
1045                 goto out;
1046         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1047         if (IS_ERR(req)) {
1048                 err = PTR_ERR(req);
1049                 goto out;
1050         }
1051         req->r_dentry = dget(dentry);
1052         req->r_num_caps = 2;
1053         req->r_parent = dir;
1054         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1055         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1056         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1057         req->r_inode_drop = ceph_drop_caps_for_unlink(inode);
1058         err = ceph_mdsc_do_request(mdsc, dir, req);
1059         if (!err && !req->r_reply_info.head->is_dentry)
1060                 d_delete(dentry);
1061         ceph_mdsc_put_request(req);
1062 out:
1063         return err;
1064 }
1065
1066 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1067                        struct inode *new_dir, struct dentry *new_dentry,
1068                        unsigned int flags)
1069 {
1070         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1071         struct ceph_mds_client *mdsc = fsc->mdsc;
1072         struct ceph_mds_request *req;
1073         int op = CEPH_MDS_OP_RENAME;
1074         int err;
1075
1076         if (flags)
1077                 return -EINVAL;
1078
1079         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1080                 return -EXDEV;
1081         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1082                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1083                         op = CEPH_MDS_OP_RENAMESNAP;
1084                 else
1085                         return -EROFS;
1086         }
1087         /* don't allow cross-quota renames */
1088         if ((old_dir != new_dir) &&
1089             (!ceph_quota_is_same_realm(old_dir, new_dir)))
1090                 return -EXDEV;
1091
1092         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1093              old_dir, old_dentry, new_dir, new_dentry);
1094         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1095         if (IS_ERR(req))
1096                 return PTR_ERR(req);
1097         ihold(old_dir);
1098         req->r_dentry = dget(new_dentry);
1099         req->r_num_caps = 2;
1100         req->r_old_dentry = dget(old_dentry);
1101         req->r_old_dentry_dir = old_dir;
1102         req->r_parent = new_dir;
1103         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1104         req->r_old_dentry_drop = CEPH_CAP_FILE_SHARED;
1105         req->r_old_dentry_unless = CEPH_CAP_FILE_EXCL;
1106         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1107         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1108         /* release LINK_RDCACHE on source inode (mds will lock it) */
1109         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1110         if (d_really_is_positive(new_dentry)) {
1111                 req->r_inode_drop =
1112                         ceph_drop_caps_for_unlink(d_inode(new_dentry));
1113         }
1114         err = ceph_mdsc_do_request(mdsc, old_dir, req);
1115         if (!err && !req->r_reply_info.head->is_dentry) {
1116                 /*
1117                  * Normally d_move() is done by fill_trace (called by
1118                  * do_request, above).  If there is no trace, we need
1119                  * to do it here.
1120                  */
1121                 d_move(old_dentry, new_dentry);
1122         }
1123         ceph_mdsc_put_request(req);
1124         return err;
1125 }
1126
1127 /*
1128  * Ensure a dentry lease will no longer revalidate.
1129  */
1130 void ceph_invalidate_dentry_lease(struct dentry *dentry)
1131 {
1132         spin_lock(&dentry->d_lock);
1133         ceph_dentry(dentry)->time = jiffies;
1134         ceph_dentry(dentry)->lease_shared_gen = 0;
1135         spin_unlock(&dentry->d_lock);
1136 }
1137
1138 /*
1139  * Check if dentry lease is valid.  If not, delete the lease.  Try to
1140  * renew if the least is more than half up.
1141  */
1142 static int dentry_lease_is_valid(struct dentry *dentry, unsigned int flags,
1143                                  struct inode *dir)
1144 {
1145         struct ceph_dentry_info *di;
1146         struct ceph_mds_session *s;
1147         int valid = 0;
1148         u32 gen;
1149         unsigned long ttl;
1150         struct ceph_mds_session *session = NULL;
1151         u32 seq = 0;
1152
1153         spin_lock(&dentry->d_lock);
1154         di = ceph_dentry(dentry);
1155         if (di && di->lease_session) {
1156                 s = di->lease_session;
1157                 spin_lock(&s->s_gen_ttl_lock);
1158                 gen = s->s_cap_gen;
1159                 ttl = s->s_cap_ttl;
1160                 spin_unlock(&s->s_gen_ttl_lock);
1161
1162                 if (di->lease_gen == gen &&
1163                     time_before(jiffies, di->time) &&
1164                     time_before(jiffies, ttl)) {
1165                         valid = 1;
1166                         if (di->lease_renew_after &&
1167                             time_after(jiffies, di->lease_renew_after)) {
1168                                 /*
1169                                  * We should renew. If we're in RCU walk mode
1170                                  * though, we can't do that so just return
1171                                  * -ECHILD.
1172                                  */
1173                                 if (flags & LOOKUP_RCU) {
1174                                         valid = -ECHILD;
1175                                 } else {
1176                                         session = ceph_get_mds_session(s);
1177                                         seq = di->lease_seq;
1178                                         di->lease_renew_after = 0;
1179                                         di->lease_renew_from = jiffies;
1180                                 }
1181                         }
1182                 }
1183         }
1184         spin_unlock(&dentry->d_lock);
1185
1186         if (session) {
1187                 ceph_mdsc_lease_send_msg(session, dir, dentry,
1188                                          CEPH_MDS_LEASE_RENEW, seq);
1189                 ceph_put_mds_session(session);
1190         }
1191         dout("dentry_lease_is_valid - dentry %p = %d\n", dentry, valid);
1192         return valid;
1193 }
1194
1195 /*
1196  * Check if directory-wide content lease/cap is valid.
1197  */
1198 static int dir_lease_is_valid(struct inode *dir, struct dentry *dentry)
1199 {
1200         struct ceph_inode_info *ci = ceph_inode(dir);
1201         struct ceph_dentry_info *di = ceph_dentry(dentry);
1202         int valid = 0;
1203
1204         spin_lock(&ci->i_ceph_lock);
1205         if (atomic_read(&ci->i_shared_gen) == di->lease_shared_gen)
1206                 valid = __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1);
1207         spin_unlock(&ci->i_ceph_lock);
1208         dout("dir_lease_is_valid dir %p v%u dentry %p v%u = %d\n",
1209              dir, (unsigned)atomic_read(&ci->i_shared_gen),
1210              dentry, (unsigned)di->lease_shared_gen, valid);
1211         return valid;
1212 }
1213
1214 /*
1215  * Check if cached dentry can be trusted.
1216  */
1217 static int ceph_d_revalidate(struct dentry *dentry, unsigned int flags)
1218 {
1219         int valid = 0;
1220         struct dentry *parent;
1221         struct inode *dir;
1222
1223         if (flags & LOOKUP_RCU) {
1224                 parent = READ_ONCE(dentry->d_parent);
1225                 dir = d_inode_rcu(parent);
1226                 if (!dir)
1227                         return -ECHILD;
1228         } else {
1229                 parent = dget_parent(dentry);
1230                 dir = d_inode(parent);
1231         }
1232
1233         dout("d_revalidate %p '%pd' inode %p offset %lld\n", dentry,
1234              dentry, d_inode(dentry), ceph_dentry(dentry)->offset);
1235
1236         /* always trust cached snapped dentries, snapdir dentry */
1237         if (ceph_snap(dir) != CEPH_NOSNAP) {
1238                 dout("d_revalidate %p '%pd' inode %p is SNAPPED\n", dentry,
1239                      dentry, d_inode(dentry));
1240                 valid = 1;
1241         } else if (d_really_is_positive(dentry) &&
1242                    ceph_snap(d_inode(dentry)) == CEPH_SNAPDIR) {
1243                 valid = 1;
1244         } else {
1245                 valid = dentry_lease_is_valid(dentry, flags, dir);
1246                 if (valid == -ECHILD)
1247                         return valid;
1248                 if (valid || dir_lease_is_valid(dir, dentry)) {
1249                         if (d_really_is_positive(dentry))
1250                                 valid = ceph_is_any_caps(d_inode(dentry));
1251                         else
1252                                 valid = 1;
1253                 }
1254         }
1255
1256         if (!valid) {
1257                 struct ceph_mds_client *mdsc =
1258                         ceph_sb_to_client(dir->i_sb)->mdsc;
1259                 struct ceph_mds_request *req;
1260                 int op, err;
1261                 u32 mask;
1262
1263                 if (flags & LOOKUP_RCU)
1264                         return -ECHILD;
1265
1266                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1267                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1268                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1269                 if (!IS_ERR(req)) {
1270                         req->r_dentry = dget(dentry);
1271                         req->r_num_caps = 2;
1272                         req->r_parent = dir;
1273
1274                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1275                         if (ceph_security_xattr_wanted(dir))
1276                                 mask |= CEPH_CAP_XATTR_SHARED;
1277                         req->r_args.getattr.mask = cpu_to_le32(mask);
1278
1279                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1280                         switch (err) {
1281                         case 0:
1282                                 if (d_really_is_positive(dentry) &&
1283                                     d_inode(dentry) == req->r_target_inode)
1284                                         valid = 1;
1285                                 break;
1286                         case -ENOENT:
1287                                 if (d_really_is_negative(dentry))
1288                                         valid = 1;
1289                                 /* Fallthrough */
1290                         default:
1291                                 break;
1292                         }
1293                         ceph_mdsc_put_request(req);
1294                         dout("d_revalidate %p lookup result=%d\n",
1295                              dentry, err);
1296                 }
1297         }
1298
1299         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1300         if (valid) {
1301                 ceph_dentry_lru_touch(dentry);
1302         } else {
1303                 ceph_dir_clear_complete(dir);
1304         }
1305
1306         if (!(flags & LOOKUP_RCU))
1307                 dput(parent);
1308         return valid;
1309 }
1310
1311 /*
1312  * Release our ceph_dentry_info.
1313  */
1314 static void ceph_d_release(struct dentry *dentry)
1315 {
1316         struct ceph_dentry_info *di = ceph_dentry(dentry);
1317
1318         dout("d_release %p\n", dentry);
1319         ceph_dentry_lru_del(dentry);
1320
1321         spin_lock(&dentry->d_lock);
1322         dentry->d_fsdata = NULL;
1323         spin_unlock(&dentry->d_lock);
1324
1325         if (di->lease_session)
1326                 ceph_put_mds_session(di->lease_session);
1327         kmem_cache_free(ceph_dentry_cachep, di);
1328 }
1329
1330 /*
1331  * When the VFS prunes a dentry from the cache, we need to clear the
1332  * complete flag on the parent directory.
1333  *
1334  * Called under dentry->d_lock.
1335  */
1336 static void ceph_d_prune(struct dentry *dentry)
1337 {
1338         struct ceph_inode_info *dir_ci;
1339         struct ceph_dentry_info *di;
1340
1341         dout("ceph_d_prune %pd %p\n", dentry, dentry);
1342
1343         /* do we have a valid parent? */
1344         if (IS_ROOT(dentry))
1345                 return;
1346
1347         /* we hold d_lock, so d_parent is stable */
1348         dir_ci = ceph_inode(d_inode(dentry->d_parent));
1349         if (dir_ci->i_vino.snap == CEPH_SNAPDIR)
1350                 return;
1351
1352         /* who calls d_delete() should also disable dcache readdir */
1353         if (d_really_is_negative(dentry))
1354                 return;
1355
1356         /* d_fsdata does not get cleared until d_release */
1357         if (!d_unhashed(dentry)) {
1358                 __ceph_dir_clear_complete(dir_ci);
1359                 return;
1360         }
1361
1362         /* Disable dcache readdir just in case that someone called d_drop()
1363          * or d_invalidate(), but MDS didn't revoke CEPH_CAP_FILE_SHARED
1364          * properly (dcache readdir is still enabled) */
1365         di = ceph_dentry(dentry);
1366         if (di->offset > 0 &&
1367             di->lease_shared_gen == atomic_read(&dir_ci->i_shared_gen))
1368                 __ceph_dir_clear_ordered(dir_ci);
1369 }
1370
1371 /*
1372  * read() on a dir.  This weird interface hack only works if mounted
1373  * with '-o dirstat'.
1374  */
1375 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1376                              loff_t *ppos)
1377 {
1378         struct ceph_dir_file_info *dfi = file->private_data;
1379         struct inode *inode = file_inode(file);
1380         struct ceph_inode_info *ci = ceph_inode(inode);
1381         int left;
1382         const int bufsize = 1024;
1383
1384         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1385                 return -EISDIR;
1386
1387         if (!dfi->dir_info) {
1388                 dfi->dir_info = kmalloc(bufsize, GFP_KERNEL);
1389                 if (!dfi->dir_info)
1390                         return -ENOMEM;
1391                 dfi->dir_info_len =
1392                         snprintf(dfi->dir_info, bufsize,
1393                                 "entries:   %20lld\n"
1394                                 " files:    %20lld\n"
1395                                 " subdirs:  %20lld\n"
1396                                 "rentries:  %20lld\n"
1397                                 " rfiles:   %20lld\n"
1398                                 " rsubdirs: %20lld\n"
1399                                 "rbytes:    %20lld\n"
1400                                 "rctime:    %10lld.%09ld\n",
1401                                 ci->i_files + ci->i_subdirs,
1402                                 ci->i_files,
1403                                 ci->i_subdirs,
1404                                 ci->i_rfiles + ci->i_rsubdirs,
1405                                 ci->i_rfiles,
1406                                 ci->i_rsubdirs,
1407                                 ci->i_rbytes,
1408                                 ci->i_rctime.tv_sec,
1409                                 ci->i_rctime.tv_nsec);
1410         }
1411
1412         if (*ppos >= dfi->dir_info_len)
1413                 return 0;
1414         size = min_t(unsigned, size, dfi->dir_info_len-*ppos);
1415         left = copy_to_user(buf, dfi->dir_info + *ppos, size);
1416         if (left == size)
1417                 return -EFAULT;
1418         *ppos += (size - left);
1419         return size - left;
1420 }
1421
1422 /*
1423  * We maintain a private dentry LRU.
1424  *
1425  * FIXME: this needs to be changed to a per-mds lru to be useful.
1426  */
1427 void ceph_dentry_lru_add(struct dentry *dn)
1428 {
1429         struct ceph_dentry_info *di = ceph_dentry(dn);
1430         struct ceph_mds_client *mdsc;
1431
1432         dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1433         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1434         spin_lock(&mdsc->dentry_lru_lock);
1435         list_add_tail(&di->lru, &mdsc->dentry_lru);
1436         mdsc->num_dentry++;
1437         spin_unlock(&mdsc->dentry_lru_lock);
1438 }
1439
1440 void ceph_dentry_lru_touch(struct dentry *dn)
1441 {
1442         struct ceph_dentry_info *di = ceph_dentry(dn);
1443         struct ceph_mds_client *mdsc;
1444
1445         dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1446              di->offset);
1447         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1448         spin_lock(&mdsc->dentry_lru_lock);
1449         list_move_tail(&di->lru, &mdsc->dentry_lru);
1450         spin_unlock(&mdsc->dentry_lru_lock);
1451 }
1452
1453 void ceph_dentry_lru_del(struct dentry *dn)
1454 {
1455         struct ceph_dentry_info *di = ceph_dentry(dn);
1456         struct ceph_mds_client *mdsc;
1457
1458         dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1459         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1460         spin_lock(&mdsc->dentry_lru_lock);
1461         list_del_init(&di->lru);
1462         mdsc->num_dentry--;
1463         spin_unlock(&mdsc->dentry_lru_lock);
1464 }
1465
1466 /*
1467  * Return name hash for a given dentry.  This is dependent on
1468  * the parent directory's hash function.
1469  */
1470 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1471 {
1472         struct ceph_inode_info *dci = ceph_inode(dir);
1473
1474         switch (dci->i_dir_layout.dl_dir_hash) {
1475         case 0: /* for backward compat */
1476         case CEPH_STR_HASH_LINUX:
1477                 return dn->d_name.hash;
1478
1479         default:
1480                 return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1481                                      dn->d_name.name, dn->d_name.len);
1482         }
1483 }
1484
1485 const struct file_operations ceph_dir_fops = {
1486         .read = ceph_read_dir,
1487         .iterate = ceph_readdir,
1488         .llseek = ceph_dir_llseek,
1489         .open = ceph_open,
1490         .release = ceph_release,
1491         .unlocked_ioctl = ceph_ioctl,
1492         .fsync = ceph_fsync,
1493         .lock = ceph_lock,
1494         .flock = ceph_flock,
1495 };
1496
1497 const struct file_operations ceph_snapdir_fops = {
1498         .iterate = ceph_readdir,
1499         .llseek = ceph_dir_llseek,
1500         .open = ceph_open,
1501         .release = ceph_release,
1502 };
1503
1504 const struct inode_operations ceph_dir_iops = {
1505         .lookup = ceph_lookup,
1506         .permission = ceph_permission,
1507         .getattr = ceph_getattr,
1508         .setattr = ceph_setattr,
1509         .listxattr = ceph_listxattr,
1510         .get_acl = ceph_get_acl,
1511         .set_acl = ceph_set_acl,
1512         .mknod = ceph_mknod,
1513         .symlink = ceph_symlink,
1514         .mkdir = ceph_mkdir,
1515         .link = ceph_link,
1516         .unlink = ceph_unlink,
1517         .rmdir = ceph_unlink,
1518         .rename = ceph_rename,
1519         .create = ceph_create,
1520         .atomic_open = ceph_atomic_open,
1521 };
1522
1523 const struct inode_operations ceph_snapdir_iops = {
1524         .lookup = ceph_lookup,
1525         .permission = ceph_permission,
1526         .getattr = ceph_getattr,
1527         .mkdir = ceph_mkdir,
1528         .rmdir = ceph_unlink,
1529         .rename = ceph_rename,
1530 };
1531
1532 const struct dentry_operations ceph_dentry_ops = {
1533         .d_revalidate = ceph_d_revalidate,
1534         .d_release = ceph_d_release,
1535         .d_prune = ceph_d_prune,
1536         .d_init = ceph_d_init,
1537 };