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 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_file_info *fi, 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(fi->last_name);
 111         fi->last_name = buf;
 112         memcpy(fi->last_name, name, len);
 113         fi->last_name[len] = 0;
 114         fi->next_offset = next_offset;
 115         dout("note_last_dentry '%s'\n", fi->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                             u32 shared_gen)
 176 {
 177         struct ceph_file_info *fi = 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, 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                         fi->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                 di = ceph_dentry(dentry);
 234                 spin_lock(&dentry->d_lock);
 235                 if (di->lease_shared_gen == shared_gen &&
 236                     d_really_is_positive(dentry) &&
 237                     fpos_cmp(ctx->pos, di->offset) <= 0) {
 238                         emit_dentry = true;
 239                 }
 240                 spin_unlock(&dentry->d_lock);
 241
 242                 if (emit_dentry) {
 243                         dout(" %llx dentry %p %pd %p\n", di->offset,
 244                              dentry, dentry, d_inode(dentry));
 245                         ctx->pos = di->offset;
 246                         if (!dir_emit(ctx, dentry->d_name.name,
 247                                       dentry->d_name.len,
 248                                       ceph_translate_ino(dentry->d_sb,
 249                                                          d_inode(dentry)->i_ino),
 250                                       d_inode(dentry)->i_mode >> 12)) {
 251                                 dput(dentry);
 252                                 err = 0;
 253                                 break;
 254                         }
 255                         ctx->pos++;
 256
 257                         if (last)
 258                                 dput(last);
 259                         last = dentry;
 260                 } else {
 261                         dput(dentry);
 262                 }
 263         }
 264 out:
 265         ceph_readdir_cache_release(&cache_ctl);
 266         if (last) {
 267                 int ret;
 268                 di = ceph_dentry(last);
 269                 ret = note_last_dentry(fi, last->d_name.name, last->d_name.len,
 270                                        fpos_off(di->offset) + 1);
 271                 if (ret < 0)
 272                         err = ret;
 273                 dput(last);
 274                 /* last_name no longer match cache index */
 275                 if (fi->readdir_cache_idx >= 0) {
 276                         fi->readdir_cache_idx = -1;
 277                         fi->dir_release_count = 0;
 278                 }
 279         }
 280         return err;
 281 }
 282
 283 static bool need_send_readdir(struct ceph_file_info *fi, loff_t pos)
 284 {
 285         if (!fi->last_readdir)
 286                 return true;
 287         if (is_hash_order(pos))
 288                 return !ceph_frag_contains_value(fi->frag, fpos_hash(pos));
 289         else
 290                 return fi->frag != fpos_frag(pos);
 291 }
 292
 293 static int ceph_readdir(struct file *file, struct dir_context *ctx)
 294 {
 295         struct ceph_file_info *fi = file->private_data;
 296         struct inode *inode = file_inode(file);
 297         struct ceph_inode_info *ci = ceph_inode(inode);
 298         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
 299         struct ceph_mds_client *mdsc = fsc->mdsc;
 300         int i;
 301         int err;
 302         unsigned frag = -1;
 303         struct ceph_mds_reply_info_parsed *rinfo;
 304
 305         dout("readdir %p file %p pos %llx\n", inode, file, ctx->pos);
 306         if (fi->flags & CEPH_F_ATEND)
 307                 return 0;
 308
 309         /* always start with . and .. */
 310         if (ctx->pos == 0) {
 311                 dout("readdir off 0 -> '.'\n");
 312                 if (!dir_emit(ctx, ".", 1,
 313                             ceph_translate_ino(inode->i_sb, inode->i_ino),
 314                             inode->i_mode >> 12))
 315                         return 0;
 316                 ctx->pos = 1;
 317         }
 318         if (ctx->pos == 1) {
 319                 ino_t ino = parent_ino(file->f_path.dentry);
 320                 dout("readdir off 1 -> '..'\n");
 321                 if (!dir_emit(ctx, "..", 2,
 322                             ceph_translate_ino(inode->i_sb, ino),
 323                             inode->i_mode >> 12))
 324                         return 0;
 325                 ctx->pos = 2;
 326         }
 327
 328         /* can we use the dcache? */
 329         spin_lock(&ci->i_ceph_lock);
 330         if (ceph_test_mount_opt(fsc, DCACHE) &&
 331             !ceph_test_mount_opt(fsc, NOASYNCREADDIR) &&
 332             ceph_snap(inode) != CEPH_SNAPDIR &&
 333             __ceph_dir_is_complete_ordered(ci) &&
 334             __ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1)) {
 335                 u32 shared_gen = ci->i_shared_gen;
 336                 spin_unlock(&ci->i_ceph_lock);
 337                 err = __dcache_readdir(file, ctx, shared_gen);
 338                 if (err != -EAGAIN)
 339                         return err;
 340         } else {
 341                 spin_unlock(&ci->i_ceph_lock);
 342         }
 343
 344         /* proceed with a normal readdir */
 345 more:
 346         /* do we have the correct frag content buffered? */
 347         if (need_send_readdir(fi, ctx->pos)) {
 348                 struct ceph_mds_request *req;
 349                 int op = ceph_snap(inode) == CEPH_SNAPDIR ?
 350                         CEPH_MDS_OP_LSSNAP : CEPH_MDS_OP_READDIR;
 351
 352                 /* discard old result, if any */
 353                 if (fi->last_readdir) {
 354                         ceph_mdsc_put_request(fi->last_readdir);
 355                         fi->last_readdir = NULL;
 356                 }
 357
 358                 if (is_hash_order(ctx->pos)) {
 359                         /* fragtree isn't always accurate. choose frag
 360                          * based on previous reply when possible. */
 361                         if (frag == (unsigned)-1)
 362                                 frag = ceph_choose_frag(ci, fpos_hash(ctx->pos),
 363                                                         NULL, NULL);
 364                 } else {
 365                         frag = fpos_frag(ctx->pos);
 366                 }
 367
 368                 dout("readdir fetching %llx.%llx frag %x offset '%s'\n",
 369                      ceph_vinop(inode), frag, fi->last_name);
 370                 req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 371                 if (IS_ERR(req))
 372                         return PTR_ERR(req);
 373                 err = ceph_alloc_readdir_reply_buffer(req, inode);
 374                 if (err) {
 375                         ceph_mdsc_put_request(req);
 376                         return err;
 377                 }
 378                 /* hints to request -> mds selection code */
 379                 req->r_direct_mode = USE_AUTH_MDS;
 380                 req->r_direct_hash = ceph_frag_value(frag);
 381                 __set_bit(CEPH_MDS_R_DIRECT_IS_HASH, &req->r_req_flags);
 382                 if (fi->last_name) {
 383                         req->r_path2 = kstrdup(fi->last_name, GFP_KERNEL);
 384                         if (!req->r_path2) {
 385                                 ceph_mdsc_put_request(req);
 386                                 return -ENOMEM;
 387                         }
 388                 } else if (is_hash_order(ctx->pos)) {
 389                         req->r_args.readdir.offset_hash =
 390                                 cpu_to_le32(fpos_hash(ctx->pos));
 391                 }
 392
 393                 req->r_dir_release_cnt = fi->dir_release_count;
 394                 req->r_dir_ordered_cnt = fi->dir_ordered_count;
 395                 req->r_readdir_cache_idx = fi->readdir_cache_idx;
 396                 req->r_readdir_offset = fi->next_offset;
 397                 req->r_args.readdir.frag = cpu_to_le32(frag);
 398                 req->r_args.readdir.flags =
 399                                 cpu_to_le16(CEPH_READDIR_REPLY_BITFLAGS);
 400
 401                 req->r_inode = inode;
 402                 ihold(inode);
 403                 req->r_dentry = dget(file->f_path.dentry);
 404                 err = ceph_mdsc_do_request(mdsc, NULL, req);
 405                 if (err < 0) {
 406                         ceph_mdsc_put_request(req);
 407                         return err;
 408                 }
 409                 dout("readdir got and parsed readdir result=%d on "
 410                      "frag %x, end=%d, complete=%d, hash_order=%d\n",
 411                      err, frag,
 412                      (int)req->r_reply_info.dir_end,
 413                      (int)req->r_reply_info.dir_complete,
 414                      (int)req->r_reply_info.hash_order);
 415
 416                 rinfo = &req->r_reply_info;
 417                 if (le32_to_cpu(rinfo->dir_dir->frag) != frag) {
 418                         frag = le32_to_cpu(rinfo->dir_dir->frag);
 419                         if (!rinfo->hash_order) {
 420                                 fi->next_offset = req->r_readdir_offset;
 421                                 /* adjust ctx->pos to beginning of frag */
 422                                 ctx->pos = ceph_make_fpos(frag,
 423                                                           fi->next_offset,
 424                                                           false);
 425                         }
 426                 }
 427
 428                 fi->frag = frag;
 429                 fi->last_readdir = req;
 430
 431                 if (test_bit(CEPH_MDS_R_DID_PREPOPULATE, &req->r_req_flags)) {
 432                         fi->readdir_cache_idx = req->r_readdir_cache_idx;
 433                         if (fi->readdir_cache_idx < 0) {
 434                                 /* preclude from marking dir ordered */
 435                                 fi->dir_ordered_count = 0;
 436                         } else if (ceph_frag_is_leftmost(frag) &&
 437                                    fi->next_offset == 2) {
 438                                 /* note dir version at start of readdir so
 439                                  * we can tell if any dentries get dropped */
 440                                 fi->dir_release_count = req->r_dir_release_cnt;
 441                                 fi->dir_ordered_count = req->r_dir_ordered_cnt;
 442                         }
 443                 } else {
 444                         dout("readdir !did_prepopulate");
 445                         /* disable readdir cache */
 446                         fi->readdir_cache_idx = -1;
 447                         /* preclude from marking dir complete */
 448                         fi->dir_release_count = 0;
 449                 }
 450
 451                 /* note next offset and last dentry name */
 452                 if (rinfo->dir_nr > 0) {
 453                         struct ceph_mds_reply_dir_entry *rde =
 454                                         rinfo->dir_entries + (rinfo->dir_nr-1);
 455                         unsigned next_offset = req->r_reply_info.dir_end ?
 456                                         2 : (fpos_off(rde->offset) + 1);
 457                         err = note_last_dentry(fi, rde->name, rde->name_len,
 458                                                next_offset);
 459                         if (err)
 460                                 return err;
 461                 } else if (req->r_reply_info.dir_end) {
 462                         fi->next_offset = 2;
 463                         /* keep last name */
 464                 }
 465         }
 466
 467         rinfo = &fi->last_readdir->r_reply_info;
 468         dout("readdir frag %x num %d pos %llx chunk first %llx\n",
 469              fi->frag, rinfo->dir_nr, ctx->pos,
 470              rinfo->dir_nr ? rinfo->dir_entries[0].offset : 0LL);
 471
 472         i = 0;
 473         /* search start position */
 474         if (rinfo->dir_nr > 0) {
 475                 int step, nr = rinfo->dir_nr;
 476                 while (nr > 0) {
 477                         step = nr >> 1;
 478                         if (rinfo->dir_entries[i + step].offset < ctx->pos) {
 479                                 i +=  step + 1;
 480                                 nr -= step + 1;
 481                         } else {
 482                                 nr = step;
 483                         }
 484                 }
 485         }
 486         for (; i < rinfo->dir_nr; i++) {
 487                 struct ceph_mds_reply_dir_entry *rde = rinfo->dir_entries + i;
 488                 struct ceph_vino vino;
 489                 ino_t ino;
 490                 u32 ftype;
 491
 492                 BUG_ON(rde->offset < ctx->pos);
 493
 494                 ctx->pos = rde->offset;
 495                 dout("readdir (%d/%d) -> %llx '%.*s' %p\n",
 496                      i, rinfo->dir_nr, ctx->pos,
 497                      rde->name_len, rde->name, &rde->inode.in);
 498
 499                 BUG_ON(!rde->inode.in);
 500                 ftype = le32_to_cpu(rde->inode.in->mode) >> 12;
 501                 vino.ino = le64_to_cpu(rde->inode.in->ino);
 502                 vino.snap = le64_to_cpu(rde->inode.in->snapid);
 503                 ino = ceph_vino_to_ino(vino);
 504
 505                 if (!dir_emit(ctx, rde->name, rde->name_len,
 506                               ceph_translate_ino(inode->i_sb, ino), ftype)) {
 507                         dout("filldir stopping us...\n");
 508                         return 0;
 509                 }
 510                 ctx->pos++;
 511         }
 512
 513         ceph_mdsc_put_request(fi->last_readdir);
 514         fi->last_readdir = NULL;
 515
 516         if (fi->next_offset > 2) {
 517                 frag = fi->frag;
 518                 goto more;
 519         }
 520
 521         /* more frags? */
 522         if (!ceph_frag_is_rightmost(fi->frag)) {
 523                 frag = ceph_frag_next(fi->frag);
 524                 if (is_hash_order(ctx->pos)) {
 525                         loff_t new_pos = ceph_make_fpos(ceph_frag_value(frag),
 526                                                         fi->next_offset, true);
 527                         if (new_pos > ctx->pos)
 528                                 ctx->pos = new_pos;
 529                         /* keep last_name */
 530                 } else {
 531                         ctx->pos = ceph_make_fpos(frag, fi->next_offset, false);
 532                         kfree(fi->last_name);
 533                         fi->last_name = NULL;
 534                 }
 535                 dout("readdir next frag is %x\n", frag);
 536                 goto more;
 537         }
 538         fi->flags |= CEPH_F_ATEND;
 539
 540         /*
 541          * if dir_release_count still matches the dir, no dentries
 542          * were released during the whole readdir, and we should have
 543          * the complete dir contents in our cache.
 544          */
 545         if (atomic64_read(&ci->i_release_count) == fi->dir_release_count) {
 546                 spin_lock(&ci->i_ceph_lock);
 547                 if (fi->dir_ordered_count == atomic64_read(&ci->i_ordered_count)) {
 548                         dout(" marking %p complete and ordered\n", inode);
 549                         /* use i_size to track number of entries in
 550                          * readdir cache */
 551                         BUG_ON(fi->readdir_cache_idx < 0);
 552                         i_size_write(inode, fi->readdir_cache_idx *
 553                                      sizeof(struct dentry*));
 554                 } else {
 555                         dout(" marking %p complete\n", inode);
 556                 }
 557                 __ceph_dir_set_complete(ci, fi->dir_release_count,
 558                                         fi->dir_ordered_count);
 559                 spin_unlock(&ci->i_ceph_lock);
 560         }
 561
 562         dout("readdir %p file %p done.\n", inode, file);
 563         return 0;
 564 }
 565
 566 static void reset_readdir(struct ceph_file_info *fi)
 567 {
 568         if (fi->last_readdir) {
 569                 ceph_mdsc_put_request(fi->last_readdir);
 570                 fi->last_readdir = NULL;
 571         }
 572         kfree(fi->last_name);
 573         fi->last_name = NULL;
 574         fi->dir_release_count = 0;
 575         fi->readdir_cache_idx = -1;
 576         fi->next_offset = 2;  /* compensate for . and .. */
 577         fi->flags &= ~CEPH_F_ATEND;
 578 }
 579
 580 /*
 581  * discard buffered readdir content on seekdir(0), or seek to new frag,
 582  * or seek prior to current chunk
 583  */
 584 static bool need_reset_readdir(struct ceph_file_info *fi, loff_t new_pos)
 585 {
 586         struct ceph_mds_reply_info_parsed *rinfo;
 587         loff_t chunk_offset;
 588         if (new_pos == 0)
 589                 return true;
 590         if (is_hash_order(new_pos)) {
 591                 /* no need to reset last_name for a forward seek when
 592                  * dentries are sotred in hash order */
 593         } else if (fi->frag != fpos_frag(new_pos)) {
 594                 return true;
 595         }
 596         rinfo = fi->last_readdir ? &fi->last_readdir->r_reply_info : NULL;
 597         if (!rinfo || !rinfo->dir_nr)
 598                 return true;
 599         chunk_offset = rinfo->dir_entries[0].offset;
 600         return new_pos < chunk_offset ||
 601                is_hash_order(new_pos) != is_hash_order(chunk_offset);
 602 }
 603
 604 static loff_t ceph_dir_llseek(struct file *file, loff_t offset, int whence)
 605 {
 606         struct ceph_file_info *fi = file->private_data;
 607         struct inode *inode = file->f_mapping->host;
 608         loff_t retval;
 609
 610         inode_lock(inode);
 611         retval = -EINVAL;
 612         switch (whence) {
 613         case SEEK_CUR:
 614                 offset += file->f_pos;
 615         case SEEK_SET:
 616                 break;
 617         case SEEK_END:
 618                 retval = -EOPNOTSUPP;
 619         default:
 620                 goto out;
 621         }
 622
 623         if (offset >= 0) {
 624                 if (need_reset_readdir(fi, offset)) {
 625                         dout("dir_llseek dropping %p content\n", file);
 626                         reset_readdir(fi);
 627                 } else if (is_hash_order(offset) && offset > file->f_pos) {
 628                         /* for hash offset, we don't know if a forward seek
 629                          * is within same frag */
 630                         fi->dir_release_count = 0;
 631                         fi->readdir_cache_idx = -1;
 632                 }
 633
 634                 if (offset != file->f_pos) {
 635                         file->f_pos = offset;
 636                         file->f_version = 0;
 637                         fi->flags &= ~CEPH_F_ATEND;
 638                 }
 639                 retval = offset;
 640         }
 641 out:
 642         inode_unlock(inode);
 643         return retval;
 644 }
 645
 646 /*
 647  * Handle lookups for the hidden .snap directory.
 648  */
 649 int ceph_handle_snapdir(struct ceph_mds_request *req,
 650                         struct dentry *dentry, int err)
 651 {
 652         struct ceph_fs_client *fsc = ceph_sb_to_client(dentry->d_sb);
 653         struct inode *parent = d_inode(dentry->d_parent); /* we hold i_mutex */
 654
 655         /* .snap dir? */
 656         if (err == -ENOENT &&
 657             ceph_snap(parent) == CEPH_NOSNAP &&
 658             strcmp(dentry->d_name.name,
 659                    fsc->mount_options->snapdir_name) == 0) {
 660                 struct inode *inode = ceph_get_snapdir(parent);
 661                 dout("ENOENT on snapdir %p '%pd', linking to snapdir %p\n",
 662                      dentry, dentry, inode);
 663                 BUG_ON(!d_unhashed(dentry));
 664                 d_add(dentry, inode);
 665                 err = 0;
 666         }
 667         return err;
 668 }
 669
 670 /*
 671  * Figure out final result of a lookup/open request.
 672  *
 673  * Mainly, make sure we return the final req->r_dentry (if it already
 674  * existed) in place of the original VFS-provided dentry when they
 675  * differ.
 676  *
 677  * Gracefully handle the case where the MDS replies with -ENOENT and
 678  * no trace (which it may do, at its discretion, e.g., if it doesn't
 679  * care to issue a lease on the negative dentry).
 680  */
 681 struct dentry *ceph_finish_lookup(struct ceph_mds_request *req,
 682                                   struct dentry *dentry, int err)
 683 {
 684         if (err == -ENOENT) {
 685                 /* no trace? */
 686                 err = 0;
 687                 if (!req->r_reply_info.head->is_dentry) {
 688                         dout("ENOENT and no trace, dentry %p inode %p\n",
 689                              dentry, d_inode(dentry));
 690                         if (d_really_is_positive(dentry)) {
 691                                 d_drop(dentry);
 692                                 err = -ENOENT;
 693                         } else {
 694                                 d_add(dentry, NULL);
 695                         }
 696                 }
 697         }
 698         if (err)
 699                 dentry = ERR_PTR(err);
 700         else if (dentry != req->r_dentry)
 701                 dentry = dget(req->r_dentry);   /* we got spliced */
 702         else
 703                 dentry = NULL;
 704         return dentry;
 705 }
 706
 707 static bool is_root_ceph_dentry(struct inode *inode, struct dentry *dentry)
 708 {
 709         return ceph_ino(inode) == CEPH_INO_ROOT &&
 710                 strncmp(dentry->d_name.name, ".ceph", 5) == 0;
 711 }
 712
 713 /*
 714  * Look up a single dir entry.  If there is a lookup intent, inform
 715  * the MDS so that it gets our 'caps wanted' value in a single op.
 716  */
 717 static struct dentry *ceph_lookup(struct inode *dir, struct dentry *dentry,
 718                                   unsigned int flags)
 719 {
 720         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 721         struct ceph_mds_client *mdsc = fsc->mdsc;
 722         struct ceph_mds_request *req;
 723         int op;
 724         int mask;
 725         int err;
 726
 727         dout("lookup %p dentry %p '%pd'\n",
 728              dir, dentry, dentry);
 729
 730         if (dentry->d_name.len > NAME_MAX)
 731                 return ERR_PTR(-ENAMETOOLONG);
 732
 733         /* can we conclude ENOENT locally? */
 734         if (d_really_is_negative(dentry)) {
 735                 struct ceph_inode_info *ci = ceph_inode(dir);
 736                 struct ceph_dentry_info *di = ceph_dentry(dentry);
 737
 738                 spin_lock(&ci->i_ceph_lock);
 739                 dout(" dir %p flags are %d\n", dir, ci->i_ceph_flags);
 740                 if (strncmp(dentry->d_name.name,
 741                             fsc->mount_options->snapdir_name,
 742                             dentry->d_name.len) &&
 743                     !is_root_ceph_dentry(dir, dentry) &&
 744                     ceph_test_mount_opt(fsc, DCACHE) &&
 745                     __ceph_dir_is_complete(ci) &&
 746                     (__ceph_caps_issued_mask(ci, CEPH_CAP_FILE_SHARED, 1))) {
 747                         spin_unlock(&ci->i_ceph_lock);
 748                         dout(" dir %p complete, -ENOENT\n", dir);
 749                         d_add(dentry, NULL);
 750                         di->lease_shared_gen = ci->i_shared_gen;
 751                         return NULL;
 752                 }
 753                 spin_unlock(&ci->i_ceph_lock);
 754         }
 755
 756         op = ceph_snap(dir) == CEPH_SNAPDIR ?
 757                 CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
 758         req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
 759         if (IS_ERR(req))
 760                 return ERR_CAST(req);
 761         req->r_dentry = dget(dentry);
 762         req->r_num_caps = 2;
 763
 764         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
 765         if (ceph_security_xattr_wanted(dir))
 766                 mask |= CEPH_CAP_XATTR_SHARED;
 767         req->r_args.getattr.mask = cpu_to_le32(mask);
 768
 769         req->r_parent = dir;
 770         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 771         err = ceph_mdsc_do_request(mdsc, NULL, req);
 772         err = ceph_handle_snapdir(req, dentry, err);
 773         dentry = ceph_finish_lookup(req, dentry, err);
 774         ceph_mdsc_put_request(req);  /* will dput(dentry) */
 775         dout("lookup result=%p\n", dentry);
 776         return dentry;
 777 }
 778
 779 /*
 780  * If we do a create but get no trace back from the MDS, follow up with
 781  * a lookup (the VFS expects us to link up the provided dentry).
 782  */
 783 int ceph_handle_notrace_create(struct inode *dir, struct dentry *dentry)
 784 {
 785         struct dentry *result = ceph_lookup(dir, dentry, 0);
 786
 787         if (result && !IS_ERR(result)) {
 788                 /*
 789                  * We created the item, then did a lookup, and found
 790                  * it was already linked to another inode we already
 791                  * had in our cache (and thus got spliced). To not
 792                  * confuse VFS (especially when inode is a directory),
 793                  * we don't link our dentry to that inode, return an
 794                  * error instead.
 795                  *
 796                  * This event should be rare and it happens only when
 797                  * we talk to old MDS. Recent MDS does not send traceless
 798                  * reply for request that creates new inode.
 799                  */
 800                 d_drop(result);
 801                 return -ESTALE;
 802         }
 803         return PTR_ERR(result);
 804 }
 805
 806 static int ceph_mknod(struct inode *dir, struct dentry *dentry,
 807                       umode_t mode, dev_t rdev)
 808 {
 809         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 810         struct ceph_mds_client *mdsc = fsc->mdsc;
 811         struct ceph_mds_request *req;
 812         struct ceph_acls_info acls = {};
 813         int err;
 814
 815         if (ceph_snap(dir) != CEPH_NOSNAP)
 816                 return -EROFS;
 817
 818         err = ceph_pre_init_acls(dir, &mode, &acls);
 819         if (err < 0)
 820                 return err;
 821
 822         dout("mknod in dir %p dentry %p mode 0%ho rdev %d\n",
 823              dir, dentry, mode, rdev);
 824         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_MKNOD, USE_AUTH_MDS);
 825         if (IS_ERR(req)) {
 826                 err = PTR_ERR(req);
 827                 goto out;
 828         }
 829         req->r_dentry = dget(dentry);
 830         req->r_num_caps = 2;
 831         req->r_parent = dir;
 832         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 833         req->r_args.mknod.mode = cpu_to_le32(mode);
 834         req->r_args.mknod.rdev = cpu_to_le32(rdev);
 835         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 836         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 837         if (acls.pagelist) {
 838                 req->r_pagelist = acls.pagelist;
 839                 acls.pagelist = NULL;
 840         }
 841         err = ceph_mdsc_do_request(mdsc, dir, req);
 842         if (!err && !req->r_reply_info.head->is_dentry)
 843                 err = ceph_handle_notrace_create(dir, dentry);
 844         ceph_mdsc_put_request(req);
 845 out:
 846         if (!err)
 847                 ceph_init_inode_acls(d_inode(dentry), &acls);
 848         else
 849                 d_drop(dentry);
 850         ceph_release_acls_info(&acls);
 851         return err;
 852 }
 853
 854 static int ceph_create(struct inode *dir, struct dentry *dentry, umode_t mode,
 855                        bool excl)
 856 {
 857         return ceph_mknod(dir, dentry, mode, 0);
 858 }
 859
 860 static int ceph_symlink(struct inode *dir, struct dentry *dentry,
 861                             const char *dest)
 862 {
 863         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 864         struct ceph_mds_client *mdsc = fsc->mdsc;
 865         struct ceph_mds_request *req;
 866         int err;
 867
 868         if (ceph_snap(dir) != CEPH_NOSNAP)
 869                 return -EROFS;
 870
 871         dout("symlink in dir %p dentry %p to '%s'\n", dir, dentry, dest);
 872         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_SYMLINK, USE_AUTH_MDS);
 873         if (IS_ERR(req)) {
 874                 err = PTR_ERR(req);
 875                 goto out;
 876         }
 877         req->r_path2 = kstrdup(dest, GFP_KERNEL);
 878         if (!req->r_path2) {
 879                 err = -ENOMEM;
 880                 ceph_mdsc_put_request(req);
 881                 goto out;
 882         }
 883         req->r_parent = dir;
 884         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 885         req->r_dentry = dget(dentry);
 886         req->r_num_caps = 2;
 887         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 888         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 889         err = ceph_mdsc_do_request(mdsc, dir, req);
 890         if (!err && !req->r_reply_info.head->is_dentry)
 891                 err = ceph_handle_notrace_create(dir, dentry);
 892         ceph_mdsc_put_request(req);
 893 out:
 894         if (err)
 895                 d_drop(dentry);
 896         return err;
 897 }
 898
 899 static int ceph_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 900 {
 901         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 902         struct ceph_mds_client *mdsc = fsc->mdsc;
 903         struct ceph_mds_request *req;
 904         struct ceph_acls_info acls = {};
 905         int err = -EROFS;
 906         int op;
 907
 908         if (ceph_snap(dir) == CEPH_SNAPDIR) {
 909                 /* mkdir .snap/foo is a MKSNAP */
 910                 op = CEPH_MDS_OP_MKSNAP;
 911                 dout("mksnap dir %p snap '%pd' dn %p\n", dir,
 912                      dentry, dentry);
 913         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
 914                 dout("mkdir dir %p dn %p mode 0%ho\n", dir, dentry, mode);
 915                 op = CEPH_MDS_OP_MKDIR;
 916         } else {
 917                 goto out;
 918         }
 919
 920         mode |= S_IFDIR;
 921         err = ceph_pre_init_acls(dir, &mode, &acls);
 922         if (err < 0)
 923                 goto out;
 924
 925         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
 926         if (IS_ERR(req)) {
 927                 err = PTR_ERR(req);
 928                 goto out;
 929         }
 930
 931         req->r_dentry = dget(dentry);
 932         req->r_num_caps = 2;
 933         req->r_parent = dir;
 934         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 935         req->r_args.mkdir.mode = cpu_to_le32(mode);
 936         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 937         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 938         if (acls.pagelist) {
 939                 req->r_pagelist = acls.pagelist;
 940                 acls.pagelist = NULL;
 941         }
 942         err = ceph_mdsc_do_request(mdsc, dir, req);
 943         if (!err &&
 944             !req->r_reply_info.head->is_target &&
 945             !req->r_reply_info.head->is_dentry)
 946                 err = ceph_handle_notrace_create(dir, dentry);
 947         ceph_mdsc_put_request(req);
 948 out:
 949         if (!err)
 950                 ceph_init_inode_acls(d_inode(dentry), &acls);
 951         else
 952                 d_drop(dentry);
 953         ceph_release_acls_info(&acls);
 954         return err;
 955 }
 956
 957 static int ceph_link(struct dentry *old_dentry, struct inode *dir,
 958                      struct dentry *dentry)
 959 {
 960         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
 961         struct ceph_mds_client *mdsc = fsc->mdsc;
 962         struct ceph_mds_request *req;
 963         int err;
 964
 965         if (ceph_snap(dir) != CEPH_NOSNAP)
 966                 return -EROFS;
 967
 968         dout("link in dir %p old_dentry %p dentry %p\n", dir,
 969              old_dentry, dentry);
 970         req = ceph_mdsc_create_request(mdsc, CEPH_MDS_OP_LINK, USE_AUTH_MDS);
 971         if (IS_ERR(req)) {
 972                 d_drop(dentry);
 973                 return PTR_ERR(req);
 974         }
 975         req->r_dentry = dget(dentry);
 976         req->r_num_caps = 2;
 977         req->r_old_dentry = dget(old_dentry);
 978         req->r_parent = dir;
 979         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
 980         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
 981         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
 982         /* release LINK_SHARED on source inode (mds will lock it) */
 983         req->r_old_inode_drop = CEPH_CAP_LINK_SHARED;
 984         err = ceph_mdsc_do_request(mdsc, dir, req);
 985         if (err) {
 986                 d_drop(dentry);
 987         } else if (!req->r_reply_info.head->is_dentry) {
 988                 ihold(d_inode(old_dentry));
 989                 d_instantiate(dentry, d_inode(old_dentry));
 990         }
 991         ceph_mdsc_put_request(req);
 992         return err;
 993 }
 994
 995 /*
 996  * For a soon-to-be unlinked file, drop the AUTH_RDCACHE caps.  If it
 997  * looks like the link count will hit 0, drop any other caps (other
 998  * than PIN) we don't specifically want (due to the file still being
 999  * open).
1000  */
1001 static int drop_caps_for_unlink(struct inode *inode)
1002 {
1003         struct ceph_inode_info *ci = ceph_inode(inode);
1004         int drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
1005
1006         spin_lock(&ci->i_ceph_lock);
1007         if (inode->i_nlink == 1) {
1008                 drop |= ~(__ceph_caps_wanted(ci) | CEPH_CAP_PIN);
1009                 ci->i_ceph_flags |= CEPH_I_NODELAY;
1010         }
1011         spin_unlock(&ci->i_ceph_lock);
1012         return drop;
1013 }
1014
1015 /*
1016  * rmdir and unlink are differ only by the metadata op code
1017  */
1018 static int ceph_unlink(struct inode *dir, struct dentry *dentry)
1019 {
1020         struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
1021         struct ceph_mds_client *mdsc = fsc->mdsc;
1022         struct inode *inode = d_inode(dentry);
1023         struct ceph_mds_request *req;
1024         int err = -EROFS;
1025         int op;
1026
1027         if (ceph_snap(dir) == CEPH_SNAPDIR) {
1028                 /* rmdir .snap/foo is RMSNAP */
1029                 dout("rmsnap dir %p '%pd' dn %p\n", dir, dentry, dentry);
1030                 op = CEPH_MDS_OP_RMSNAP;
1031         } else if (ceph_snap(dir) == CEPH_NOSNAP) {
1032                 dout("unlink/rmdir dir %p dn %p inode %p\n",
1033                      dir, dentry, inode);
1034                 op = d_is_dir(dentry) ?
1035                         CEPH_MDS_OP_RMDIR : CEPH_MDS_OP_UNLINK;
1036         } else
1037                 goto out;
1038         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1039         if (IS_ERR(req)) {
1040                 err = PTR_ERR(req);
1041                 goto out;
1042         }
1043         req->r_dentry = dget(dentry);
1044         req->r_num_caps = 2;
1045         req->r_parent = dir;
1046         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
1047         req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
1048         req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
1049         req->r_inode_drop = drop_caps_for_unlink(inode);
1050         err = ceph_mdsc_do_request(mdsc, dir, req);
1051         if (!err && !req->r_reply_info.head->is_dentry)
1052                 d_delete(dentry);
1053         ceph_mdsc_put_request(req);
1054 out:
1055         return err;
1056 }
1057
1058 static int ceph_rename(struct inode *old_dir, struct dentry *old_dentry,
1059                        struct inode *new_dir, struct dentry *new_dentry,
1060                        unsigned int flags)
1061 {
1062         struct ceph_fs_client *fsc = ceph_sb_to_client(old_dir->i_sb);
1063         struct ceph_mds_client *mdsc = fsc->mdsc;
1064         struct ceph_mds_request *req;
1065         int op = CEPH_MDS_OP_RENAME;
1066         int err;
1067
1068         if (flags)
1069                 return -EINVAL;
1070
1071         if (ceph_snap(old_dir) != ceph_snap(new_dir))
1072                 return -EXDEV;
1073         if (ceph_snap(old_dir) != CEPH_NOSNAP) {
1074                 if (old_dir == new_dir && ceph_snap(old_dir) == CEPH_SNAPDIR)
1075                         op = CEPH_MDS_OP_RENAMESNAP;
1076                 else
1077                         return -EROFS;
1078         }
1079         dout("rename dir %p dentry %p to dir %p dentry %p\n",
1080              old_dir, old_dentry, new_dir, new_dentry);
1081         req = ceph_mdsc_create_request(mdsc, op, USE_AUTH_MDS);
1082         if (IS_ERR(req))
1083                 return PTR_ERR(req);
1084         ihold(old_dir);
1085         req->r_dentry = dget(new_dentry);
1086         req->r_num_caps = 2;
1087         req->r_old_dentry = dget(old_dentry);
1088         req->r_old_dentry_dir = old_dir;
1089         req->r_parent = new_dir;
1090         set_bit(CEPH_MDS_R_PARENT_LOCKED, &req->r_req_flags);
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 = READ_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, err;
1255                 u32 mask;
1256
1257                 if (flags & LOOKUP_RCU)
1258                         return -ECHILD;
1259
1260                 op = ceph_snap(dir) == CEPH_SNAPDIR ?
1261                         CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
1262                 req = ceph_mdsc_create_request(mdsc, op, USE_ANY_MDS);
1263                 if (!IS_ERR(req)) {
1264                         req->r_dentry = dget(dentry);
1265                         req->r_num_caps = 2;
1266                         req->r_parent = dir;
1267
1268                         mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
1269                         if (ceph_security_xattr_wanted(dir))
1270                                 mask |= CEPH_CAP_XATTR_SHARED;
1271                         req->r_args.getattr.mask = cpu_to_le32(mask);
1272
1273                         err = ceph_mdsc_do_request(mdsc, NULL, req);
1274                         switch (err) {
1275                         case 0:
1276                                 if (d_really_is_positive(dentry) &&
1277                                     d_inode(dentry) == req->r_target_inode)
1278                                         valid = 1;
1279                                 break;
1280                         case -ENOENT:
1281                                 if (d_really_is_negative(dentry))
1282                                         valid = 1;
1283                                 /* Fallthrough */
1284                         default:
1285                                 break;
1286                         }
1287                         ceph_mdsc_put_request(req);
1288                         dout("d_revalidate %p lookup result=%d\n",
1289                              dentry, err);
1290                 }
1291         }
1292
1293         dout("d_revalidate %p %s\n", dentry, valid ? "valid" : "invalid");
1294         if (valid) {
1295                 ceph_dentry_lru_touch(dentry);
1296         } else {
1297                 ceph_dir_clear_complete(dir);
1298         }
1299
1300         if (!(flags & LOOKUP_RCU))
1301                 dput(parent);
1302         return valid;
1303 }
1304
1305 /*
1306  * Release our ceph_dentry_info.
1307  */
1308 static void ceph_d_release(struct dentry *dentry)
1309 {
1310         struct ceph_dentry_info *di = ceph_dentry(dentry);
1311
1312         dout("d_release %p\n", dentry);
1313         ceph_dentry_lru_del(dentry);
1314
1315         spin_lock(&dentry->d_lock);
1316         dentry->d_fsdata = NULL;
1317         spin_unlock(&dentry->d_lock);
1318
1319         if (di->lease_session)
1320                 ceph_put_mds_session(di->lease_session);
1321         kmem_cache_free(ceph_dentry_cachep, di);
1322 }
1323
1324 /*
1325  * When the VFS prunes a dentry from the cache, we need to clear the
1326  * complete flag on the parent directory.
1327  *
1328  * Called under dentry->d_lock.
1329  */
1330 static void ceph_d_prune(struct dentry *dentry)
1331 {
1332         dout("ceph_d_prune %p\n", dentry);
1333
1334         /* do we have a valid parent? */
1335         if (IS_ROOT(dentry))
1336                 return;
1337
1338         /* if we are not hashed, we don't affect dir's completeness */
1339         if (d_unhashed(dentry))
1340                 return;
1341
1342         if (ceph_snap(d_inode(dentry->d_parent)) == CEPH_SNAPDIR)
1343                 return;
1344
1345         /*
1346          * we hold d_lock, so d_parent is stable, and d_fsdata is never
1347          * cleared until d_release
1348          */
1349         ceph_dir_clear_complete(d_inode(dentry->d_parent));
1350 }
1351
1352 /*
1353  * read() on a dir.  This weird interface hack only works if mounted
1354  * with '-o dirstat'.
1355  */
1356 static ssize_t ceph_read_dir(struct file *file, char __user *buf, size_t size,
1357                              loff_t *ppos)
1358 {
1359         struct ceph_file_info *cf = file->private_data;
1360         struct inode *inode = file_inode(file);
1361         struct ceph_inode_info *ci = ceph_inode(inode);
1362         int left;
1363         const int bufsize = 1024;
1364
1365         if (!ceph_test_mount_opt(ceph_sb_to_client(inode->i_sb), DIRSTAT))
1366                 return -EISDIR;
1367
1368         if (!cf->dir_info) {
1369                 cf->dir_info = kmalloc(bufsize, GFP_KERNEL);
1370                 if (!cf->dir_info)
1371                         return -ENOMEM;
1372                 cf->dir_info_len =
1373                         snprintf(cf->dir_info, bufsize,
1374                                 "entries:   %20lld\n"
1375                                 " files:    %20lld\n"
1376                                 " subdirs:  %20lld\n"
1377                                 "rentries:  %20lld\n"
1378                                 " rfiles:   %20lld\n"
1379                                 " rsubdirs: %20lld\n"
1380                                 "rbytes:    %20lld\n"
1381                                 "rctime:    %10ld.%09ld\n",
1382                                 ci->i_files + ci->i_subdirs,
1383                                 ci->i_files,
1384                                 ci->i_subdirs,
1385                                 ci->i_rfiles + ci->i_rsubdirs,
1386                                 ci->i_rfiles,
1387                                 ci->i_rsubdirs,
1388                                 ci->i_rbytes,
1389                                 (long)ci->i_rctime.tv_sec,
1390                                 (long)ci->i_rctime.tv_nsec);
1391         }
1392
1393         if (*ppos >= cf->dir_info_len)
1394                 return 0;
1395         size = min_t(unsigned, size, cf->dir_info_len-*ppos);
1396         left = copy_to_user(buf, cf->dir_info + *ppos, size);
1397         if (left == size)
1398                 return -EFAULT;
1399         *ppos += (size - left);
1400         return size - left;
1401 }
1402
1403 /*
1404  * We maintain a private dentry LRU.
1405  *
1406  * FIXME: this needs to be changed to a per-mds lru to be useful.
1407  */
1408 void ceph_dentry_lru_add(struct dentry *dn)
1409 {
1410         struct ceph_dentry_info *di = ceph_dentry(dn);
1411         struct ceph_mds_client *mdsc;
1412
1413         dout("dentry_lru_add %p %p '%pd'\n", di, dn, dn);
1414         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1415         spin_lock(&mdsc->dentry_lru_lock);
1416         list_add_tail(&di->lru, &mdsc->dentry_lru);
1417         mdsc->num_dentry++;
1418         spin_unlock(&mdsc->dentry_lru_lock);
1419 }
1420
1421 void ceph_dentry_lru_touch(struct dentry *dn)
1422 {
1423         struct ceph_dentry_info *di = ceph_dentry(dn);
1424         struct ceph_mds_client *mdsc;
1425
1426         dout("dentry_lru_touch %p %p '%pd' (offset %lld)\n", di, dn, dn,
1427              di->offset);
1428         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1429         spin_lock(&mdsc->dentry_lru_lock);
1430         list_move_tail(&di->lru, &mdsc->dentry_lru);
1431         spin_unlock(&mdsc->dentry_lru_lock);
1432 }
1433
1434 void ceph_dentry_lru_del(struct dentry *dn)
1435 {
1436         struct ceph_dentry_info *di = ceph_dentry(dn);
1437         struct ceph_mds_client *mdsc;
1438
1439         dout("dentry_lru_del %p %p '%pd'\n", di, dn, dn);
1440         mdsc = ceph_sb_to_client(dn->d_sb)->mdsc;
1441         spin_lock(&mdsc->dentry_lru_lock);
1442         list_del_init(&di->lru);
1443         mdsc->num_dentry--;
1444         spin_unlock(&mdsc->dentry_lru_lock);
1445 }
1446
1447 /*
1448  * Return name hash for a given dentry.  This is dependent on
1449  * the parent directory's hash function.
1450  */
1451 unsigned ceph_dentry_hash(struct inode *dir, struct dentry *dn)
1452 {
1453         struct ceph_inode_info *dci = ceph_inode(dir);
1454
1455         switch (dci->i_dir_layout.dl_dir_hash) {
1456         case 0: /* for backward compat */
1457         case CEPH_STR_HASH_LINUX:
1458                 return dn->d_name.hash;
1459
1460         default:
1461                 return ceph_str_hash(dci->i_dir_layout.dl_dir_hash,
1462                                      dn->d_name.name, dn->d_name.len);
1463         }
1464 }
1465
1466 const struct file_operations ceph_dir_fops = {
1467         .read = ceph_read_dir,
1468         .iterate = ceph_readdir,
1469         .llseek = ceph_dir_llseek,
1470         .open = ceph_open,
1471         .release = ceph_release,
1472         .unlocked_ioctl = ceph_ioctl,
1473         .fsync = ceph_fsync,
1474 };
1475
1476 const struct file_operations ceph_snapdir_fops = {
1477         .iterate = ceph_readdir,
1478         .llseek = ceph_dir_llseek,
1479         .open = ceph_open,
1480         .release = ceph_release,
1481 };
1482
1483 const struct inode_operations ceph_dir_iops = {
1484         .lookup = ceph_lookup,
1485         .permission = ceph_permission,
1486         .getattr = ceph_getattr,
1487         .setattr = ceph_setattr,
1488         .listxattr = ceph_listxattr,
1489         .get_acl = ceph_get_acl,
1490         .set_acl = ceph_set_acl,
1491         .mknod = ceph_mknod,
1492         .symlink = ceph_symlink,
1493         .mkdir = ceph_mkdir,
1494         .link = ceph_link,
1495         .unlink = ceph_unlink,
1496         .rmdir = ceph_unlink,
1497         .rename = ceph_rename,
1498         .create = ceph_create,
1499         .atomic_open = ceph_atomic_open,
1500 };
1501
1502 const struct inode_operations ceph_snapdir_iops = {
1503         .lookup = ceph_lookup,
1504         .permission = ceph_permission,
1505         .getattr = ceph_getattr,
1506         .mkdir = ceph_mkdir,
1507         .rmdir = ceph_unlink,
1508         .rename = ceph_rename,
1509 };
1510
1511 const struct dentry_operations ceph_dentry_ops = {
1512         .d_revalidate = ceph_d_revalidate,
1513         .d_release = ceph_d_release,
1514         .d_prune = ceph_d_prune,
1515         .d_init = ceph_d_init,
1516 };