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