fs/ceph/dir.c

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